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 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
362 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
363 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
364 assert(ex_map->req() == phi_map->req(), "matching maps");
365 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
366 Node* hidden_merge_mark = root();
367 Node* region = phi_map->control();
368 MergeMemNode* phi_mem = phi_map->merged_memory();
369 MergeMemNode* ex_mem = ex_map->merged_memory();
370 if (region->in(0) != hidden_merge_mark) {
371 // The control input is not (yet) a specially-marked region in phi_map.
372 // Make it so, and build some phis.
373 region = new RegionNode(2);
374 _gvn.set_type(region, Type::CONTROL);
375 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
376 region->init_req(1, phi_map->control());
377 phi_map->set_control(region);
378 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
379 record_for_igvn(io_phi);
380 _gvn.set_type(io_phi, Type::ABIO);
381 phi_map->set_i_o(io_phi);
382 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
383 Node* m = mms.memory();
384 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
385 record_for_igvn(m_phi);
386 _gvn.set_type(m_phi, Type::MEMORY);
387 mms.set_memory(m_phi);
388 }
389 }
390
391 // Either or both of phi_map and ex_map might already be converted into phis.
392 Node* ex_control = ex_map->control();
393 // if there is special marking on ex_map also, we add multiple edges from src
394 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
395 // how wide was the destination phi_map, originally?
396 uint orig_width = region->req();
397
398 if (add_multiple) {
399 add_n_reqs(region, ex_control);
400 add_n_reqs(phi_map->i_o(), ex_map->i_o());
401 } else {
402 // ex_map has no merges, so we just add single edges everywhere
403 add_one_req(region, ex_control);
404 add_one_req(phi_map->i_o(), ex_map->i_o());
405 }
406 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
407 if (mms.is_empty()) {
408 // get a copy of the base memory, and patch some inputs into it
409 const TypePtr* adr_type = mms.adr_type(C);
410 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
411 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
412 mms.set_memory(phi);
413 // Prepare to append interesting stuff onto the newly sliced phi:
414 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
415 }
416 // Append stuff from ex_map:
417 if (add_multiple) {
418 add_n_reqs(mms.memory(), mms.memory2());
419 } else {
420 add_one_req(mms.memory(), mms.memory2());
421 }
422 }
423 uint limit = ex_map->req();
424 for (uint i = TypeFunc::Parms; i < limit; i++) {
425 // Skip everything in the JVMS after tos. (The ex_oop follows.)
426 if (i == tos) i = ex_jvms->monoff();
427 Node* src = ex_map->in(i);
428 Node* dst = phi_map->in(i);
429 if (src != dst) {
430 PhiNode* phi;
431 if (dst->in(0) != region) {
432 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
433 record_for_igvn(phi);
434 _gvn.set_type(phi, phi->type());
435 phi_map->set_req(i, dst);
436 // Prepare to append interesting stuff onto the new phi:
437 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
438 } else {
439 assert(dst->is_Phi(), "nobody else uses a hidden region");
440 phi = dst->as_Phi();
441 }
442 if (add_multiple && src->in(0) == ex_control) {
443 // Both are phis.
444 add_n_reqs(dst, src);
445 } else {
446 while (dst->req() < region->req()) add_one_req(dst, src);
447 }
448 const Type* srctype = _gvn.type(src);
449 if (phi->type() != srctype) {
450 const Type* dsttype = phi->type()->meet_speculative(srctype);
451 if (phi->type() != dsttype) {
452 phi->set_type(dsttype);
453 _gvn.set_type(phi, dsttype);
454 }
455 }
456 }
457 }
458 phi_map->merge_replaced_nodes_with(ex_map);
459 }
460
461 //--------------------------use_exception_state--------------------------------
462 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
463 if (failing_internal()) { stop(); return top(); }
464 Node* region = phi_map->control();
465 Node* hidden_merge_mark = root();
466 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
467 Node* ex_oop = clear_saved_ex_oop(phi_map);
468 if (region->in(0) == hidden_merge_mark) {
469 // Special marking for internal ex-states. Process the phis now.
470 region->set_req(0, region); // now it's an ordinary region
471 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
472 // Note: Setting the jvms also sets the bci and sp.
473 set_control(_gvn.transform(region));
474 uint tos = jvms()->stkoff() + sp();
475 for (uint i = 1; i < tos; i++) {
476 Node* x = phi_map->in(i);
477 if (x->in(0) == region) {
478 assert(x->is_Phi(), "expected a special phi");
479 phi_map->set_req(i, _gvn.transform(x));
480 }
481 }
482 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
483 Node* x = mms.memory();
484 if (x->in(0) == region) {
485 assert(x->is_Phi(), "nobody else uses a hidden region");
486 mms.set_memory(_gvn.transform(x));
487 }
488 }
489 if (ex_oop->in(0) == region) {
490 assert(ex_oop->is_Phi(), "expected a special phi");
491 ex_oop = _gvn.transform(ex_oop);
492 }
493 } else {
494 set_jvms(phi_map->jvms());
495 }
496
497 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
498 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
499 return ex_oop;
500 }
501
502 //---------------------------------java_bc-------------------------------------
503 Bytecodes::Code GraphKit::java_bc() const {
504 ciMethod* method = this->method();
505 int bci = this->bci();
506 if (method != nullptr && bci != InvocationEntryBci)
507 return method->java_code_at_bci(bci);
508 else
509 return Bytecodes::_illegal;
510 }
511
512 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
513 bool must_throw) {
514 // if the exception capability is set, then we will generate code
515 // to check the JavaThread.should_post_on_exceptions flag to see
516 // if we actually need to report exception events (for this
517 // thread). If we don't need to report exception events, we will
518 // take the normal fast path provided by add_exception_events. If
519 // exception event reporting is enabled for this thread, we will
520 // take the uncommon_trap in the BuildCutout below.
521
522 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
523 Node* jthread = _gvn.transform(new ThreadLocalNode());
524 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
525 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
526
527 // Test the should_post_on_exceptions_flag vs. 0
528 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
529 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
530
531 // Branch to slow_path if should_post_on_exceptions_flag was true
532 { BuildCutout unless(this, tst, PROB_MAX);
533 // Do not try anything fancy if we're notifying the VM on every throw.
534 // Cf. case Bytecodes::_athrow in parse2.cpp.
535 uncommon_trap(reason, Deoptimization::Action_none,
536 (ciKlass*)nullptr, (char*)nullptr, must_throw);
537 }
538
539 }
540
541 //------------------------------builtin_throw----------------------------------
542 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
543 bool must_throw = true;
544
545 // If this particular condition has not yet happened at this
546 // bytecode, then use the uncommon trap mechanism, and allow for
547 // a future recompilation if several traps occur here.
548 // If the throw is hot, try to use a more complicated inline mechanism
549 // which keeps execution inside the compiled code.
550 bool treat_throw_as_hot = false;
551 ciMethodData* md = method()->method_data();
552
553 if (ProfileTraps) {
554 if (too_many_traps(reason)) {
555 treat_throw_as_hot = true;
556 }
557 // (If there is no MDO at all, assume it is early in
558 // execution, and that any deopts are part of the
559 // startup transient, and don't need to be remembered.)
560
561 // Also, if there is a local exception handler, treat all throws
562 // as hot if there has been at least one in this method.
563 if (C->trap_count(reason) != 0
564 && method()->method_data()->trap_count(reason) != 0
565 && has_exception_handler()) {
566 treat_throw_as_hot = true;
567 }
568 }
569
570 // If this throw happens frequently, an uncommon trap might cause
571 // a performance pothole. If there is a local exception handler,
572 // and if this particular bytecode appears to be deoptimizing often,
573 // let us handle the throw inline, with a preconstructed instance.
574 // Note: If the deopt count has blown up, the uncommon trap
575 // runtime is going to flush this nmethod, not matter what.
576 if (treat_throw_as_hot && method()->can_omit_stack_trace()) {
577 // If the throw is local, we use a pre-existing instance and
578 // punt on the backtrace. This would lead to a missing backtrace
579 // (a repeat of 4292742) if the backtrace object is ever asked
580 // for its backtrace.
581 // Fixing this remaining case of 4292742 requires some flavor of
582 // escape analysis. Leave that for the future.
583 ciInstance* ex_obj = nullptr;
584 switch (reason) {
585 case Deoptimization::Reason_null_check:
586 ex_obj = env()->NullPointerException_instance();
587 break;
588 case Deoptimization::Reason_div0_check:
589 ex_obj = env()->ArithmeticException_instance();
590 break;
591 case Deoptimization::Reason_range_check:
592 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
593 break;
594 case Deoptimization::Reason_class_check:
595 ex_obj = env()->ClassCastException_instance();
596 break;
597 case Deoptimization::Reason_array_check:
598 ex_obj = env()->ArrayStoreException_instance();
599 break;
600 default:
601 break;
602 }
603 // If we have a preconstructed exception object, use it.
604 if (ex_obj != nullptr) {
605 if (env()->jvmti_can_post_on_exceptions()) {
606 // check if we must post exception events, take uncommon trap if so
607 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
608 // here if should_post_on_exceptions is false
609 // continue on with the normal codegen
610 }
611
612 // Cheat with a preallocated exception object.
613 if (C->log() != nullptr)
614 C->log()->elem("hot_throw preallocated='1' reason='%s'",
615 Deoptimization::trap_reason_name(reason));
616 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
617 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
618
619 // Clear the detail message of the preallocated exception object.
620 // Weblogic sometimes mutates the detail message of exceptions
621 // using reflection.
622 int offset = java_lang_Throwable::get_detailMessage_offset();
623 const TypePtr* adr_typ = ex_con->add_offset(offset);
624
625 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
626 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
627 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
628
629 if (!method()->has_exception_handlers()) {
630 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
631 // This prevents issues with exception handling and late inlining.
632 set_sp(0);
633 clean_stack(0);
634 }
635
636 add_exception_state(make_exception_state(ex_node));
637 return;
638 }
639 }
640
641 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
642 // It won't be much cheaper than bailing to the interp., since we'll
643 // have to pass up all the debug-info, and the runtime will have to
644 // create the stack trace.
645
646 // Usual case: Bail to interpreter.
647 // Reserve the right to recompile if we haven't seen anything yet.
648
649 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
650 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
651 if (treat_throw_as_hot
652 && (method()->method_data()->trap_recompiled_at(bci(), m)
653 || C->too_many_traps(reason))) {
654 // We cannot afford to take more traps here. Suffer in the interpreter.
655 if (C->log() != nullptr)
656 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
657 Deoptimization::trap_reason_name(reason),
658 C->trap_count(reason));
659 action = Deoptimization::Action_none;
660 }
661
662 // "must_throw" prunes the JVM state to include only the stack, if there
663 // are no local exception handlers. This should cut down on register
664 // allocation time and code size, by drastically reducing the number
665 // of in-edges on the call to the uncommon trap.
666
667 uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw);
668 }
669
670
671 //----------------------------PreserveJVMState---------------------------------
672 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
673 debug_only(kit->verify_map());
674 _kit = kit;
675 _map = kit->map(); // preserve the map
676 _sp = kit->sp();
677 kit->set_map(clone_map ? kit->clone_map() : nullptr);
678 #ifdef ASSERT
679 _bci = kit->bci();
680 Parse* parser = kit->is_Parse();
681 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
682 _block = block;
683 #endif
684 }
685 PreserveJVMState::~PreserveJVMState() {
686 GraphKit* kit = _kit;
687 #ifdef ASSERT
688 assert(kit->bci() == _bci, "bci must not shift");
689 Parse* parser = kit->is_Parse();
690 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
691 assert(block == _block, "block must not shift");
692 #endif
693 kit->set_map(_map);
694 kit->set_sp(_sp);
695 }
696
697
698 //-----------------------------BuildCutout-------------------------------------
699 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
700 : PreserveJVMState(kit)
701 {
702 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
703 SafePointNode* outer_map = _map; // preserved map is caller's
704 SafePointNode* inner_map = kit->map();
705 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
706 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
707 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
708 }
709 BuildCutout::~BuildCutout() {
710 GraphKit* kit = _kit;
711 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
712 }
713
714 //---------------------------PreserveReexecuteState----------------------------
715 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
716 assert(!kit->stopped(), "must call stopped() before");
717 _kit = kit;
718 _sp = kit->sp();
719 _reexecute = kit->jvms()->_reexecute;
720 }
721 PreserveReexecuteState::~PreserveReexecuteState() {
722 if (_kit->stopped()) return;
723 _kit->jvms()->_reexecute = _reexecute;
724 _kit->set_sp(_sp);
725 }
726
727 //------------------------------clone_map--------------------------------------
728 // Implementation of PreserveJVMState
729 //
730 // Only clone_map(...) here. If this function is only used in the
731 // PreserveJVMState class we may want to get rid of this extra
732 // function eventually and do it all there.
733
734 SafePointNode* GraphKit::clone_map() {
735 if (map() == nullptr) return nullptr;
736
737 // Clone the memory edge first
738 Node* mem = MergeMemNode::make(map()->memory());
739 gvn().set_type_bottom(mem);
740
741 SafePointNode *clonemap = (SafePointNode*)map()->clone();
742 JVMState* jvms = this->jvms();
743 JVMState* clonejvms = jvms->clone_shallow(C);
744 clonemap->set_memory(mem);
745 clonemap->set_jvms(clonejvms);
746 clonejvms->set_map(clonemap);
747 record_for_igvn(clonemap);
748 gvn().set_type_bottom(clonemap);
749 return clonemap;
750 }
751
752 //-----------------------------destruct_map_clone------------------------------
753 //
754 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
755 // to destruct/free/delete in the exact opposite order as clone_map().
756 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
757 if (sfp == nullptr) return;
758
759 Node* mem = sfp->memory();
760 JVMState* jvms = sfp->jvms();
761
762 if (jvms != nullptr) {
763 delete jvms;
764 }
765
766 remove_for_igvn(sfp);
767 gvn().clear_type(sfp);
768 sfp->destruct(&_gvn);
769
770 if (mem != nullptr) {
771 gvn().clear_type(mem);
772 mem->destruct(&_gvn);
773 }
774 }
775
776 //-----------------------------set_map_clone-----------------------------------
777 void GraphKit::set_map_clone(SafePointNode* m) {
778 _map = m;
779 _map = clone_map();
780 _map->set_next_exception(nullptr);
781 debug_only(verify_map());
782 }
783
784
785 //----------------------------kill_dead_locals---------------------------------
786 // Detect any locals which are known to be dead, and force them to top.
787 void GraphKit::kill_dead_locals() {
788 // Consult the liveness information for the locals. If any
789 // of them are unused, then they can be replaced by top(). This
790 // should help register allocation time and cut down on the size
791 // of the deoptimization information.
792
793 // This call is made from many of the bytecode handling
794 // subroutines called from the Big Switch in do_one_bytecode.
795 // Every bytecode which might include a slow path is responsible
796 // for killing its dead locals. The more consistent we
797 // are about killing deads, the fewer useless phis will be
798 // constructed for them at various merge points.
799
800 // bci can be -1 (InvocationEntryBci). We return the entry
801 // liveness for the method.
802
803 if (method() == nullptr || method()->code_size() == 0) {
804 // We are building a graph for a call to a native method.
805 // All locals are live.
806 return;
807 }
808
809 ResourceMark rm;
810
811 // Consult the liveness information for the locals. If any
812 // of them are unused, then they can be replaced by top(). This
813 // should help register allocation time and cut down on the size
814 // of the deoptimization information.
815 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
816
817 int len = (int)live_locals.size();
818 assert(len <= jvms()->loc_size(), "too many live locals");
819 for (int local = 0; local < len; local++) {
820 if (!live_locals.at(local)) {
821 set_local(local, top());
822 }
823 }
824 }
825
826 #ifdef ASSERT
827 //-------------------------dead_locals_are_killed------------------------------
828 // Return true if all dead locals are set to top in the map.
829 // Used to assert "clean" debug info at various points.
830 bool GraphKit::dead_locals_are_killed() {
831 if (method() == nullptr || method()->code_size() == 0) {
832 // No locals need to be dead, so all is as it should be.
833 return true;
834 }
835
836 // Make sure somebody called kill_dead_locals upstream.
837 ResourceMark rm;
838 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
839 if (jvms->loc_size() == 0) continue; // no locals to consult
840 SafePointNode* map = jvms->map();
841 ciMethod* method = jvms->method();
842 int bci = jvms->bci();
843 if (jvms == this->jvms()) {
844 bci = this->bci(); // it might not yet be synched
845 }
846 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
847 int len = (int)live_locals.size();
848 if (!live_locals.is_valid() || len == 0)
849 // This method is trivial, or is poisoned by a breakpoint.
850 return true;
851 assert(len == jvms->loc_size(), "live map consistent with locals map");
852 for (int local = 0; local < len; local++) {
853 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
854 if (PrintMiscellaneous && (Verbose || WizardMode)) {
855 tty->print_cr("Zombie local %d: ", local);
856 jvms->dump();
857 }
858 return false;
859 }
860 }
861 }
862 return true;
863 }
864
865 #endif //ASSERT
866
867 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
868 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
869 ciMethod* cur_method = jvms->method();
870 int cur_bci = jvms->bci();
871 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
872 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
873 return Interpreter::bytecode_should_reexecute(code) ||
874 (is_anewarray && (code == Bytecodes::_multianewarray));
875 // Reexecute _multianewarray bytecode which was replaced with
876 // sequence of [a]newarray. See Parse::do_multianewarray().
877 //
878 // Note: interpreter should not have it set since this optimization
879 // is limited by dimensions and guarded by flag so in some cases
880 // multianewarray() runtime calls will be generated and
881 // the bytecode should not be reexecutes (stack will not be reset).
882 } else {
883 return false;
884 }
885 }
886
887 // Helper function for adding JVMState and debug information to node
888 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
889 // Add the safepoint edges to the call (or other safepoint).
890
891 // Make sure dead locals are set to top. This
892 // should help register allocation time and cut down on the size
893 // of the deoptimization information.
894 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
895
896 // Walk the inline list to fill in the correct set of JVMState's
897 // Also fill in the associated edges for each JVMState.
898
899 // If the bytecode needs to be reexecuted we need to put
900 // the arguments back on the stack.
901 const bool should_reexecute = jvms()->should_reexecute();
902 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
903
904 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
905 // undefined if the bci is different. This is normal for Parse but it
906 // should not happen for LibraryCallKit because only one bci is processed.
907 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
908 "in LibraryCallKit the reexecute bit should not change");
909
910 // If we are guaranteed to throw, we can prune everything but the
911 // input to the current bytecode.
912 bool can_prune_locals = false;
913 uint stack_slots_not_pruned = 0;
914 int inputs = 0, depth = 0;
915 if (must_throw) {
916 assert(method() == youngest_jvms->method(), "sanity");
917 if (compute_stack_effects(inputs, depth)) {
918 can_prune_locals = true;
919 stack_slots_not_pruned = inputs;
920 }
921 }
922
923 if (env()->should_retain_local_variables()) {
924 // At any safepoint, this method can get breakpointed, which would
925 // then require an immediate deoptimization.
926 can_prune_locals = false; // do not prune locals
927 stack_slots_not_pruned = 0;
928 }
929
930 // do not scribble on the input jvms
931 JVMState* out_jvms = youngest_jvms->clone_deep(C);
932 call->set_jvms(out_jvms); // Start jvms list for call node
933
934 // For a known set of bytecodes, the interpreter should reexecute them if
935 // deoptimization happens. We set the reexecute state for them here
936 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
937 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
938 #ifdef ASSERT
939 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
940 assert(method() == youngest_jvms->method(), "sanity");
941 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
942 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
943 #endif // ASSERT
944 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
945 }
946
947 // Presize the call:
948 DEBUG_ONLY(uint non_debug_edges = call->req());
949 call->add_req_batch(top(), youngest_jvms->debug_depth());
950 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
951
952 // Set up edges so that the call looks like this:
953 // Call [state:] ctl io mem fptr retadr
954 // [parms:] parm0 ... parmN
955 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
956 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
957 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // Note that caller debug info precedes callee debug info.
959
960 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
961 uint debug_ptr = call->req();
962
963 // Loop over the map input edges associated with jvms, add them
964 // to the call node, & reset all offsets to match call node array.
965
966 JVMState* callee_jvms = nullptr;
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++) {
993 Node* val = in_map->in(k + j);
994 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
995 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
996 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
997 val->bottom_type()->is_inlinetypeptr()) {
998 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
999 }
1000 call->set_req(p++, val);
1001 }
1002 } else {
1003 p += l; // already set to top above by add_req_batch
1004 }
1005
1006 // Add the Expression Stack
1007 k = in_jvms->stkoff();
1008 l = in_jvms->sp();
1009 out_jvms->set_stkoff(p);
1010 if (!can_prune_locals) {
1011 for (j = 0; j < l; j++) {
1012 Node* val = in_map->in(k + j);
1013 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
1014 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
1015 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
1016 val->bottom_type()->is_inlinetypeptr()) {
1017 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
1018 }
1019 call->set_req(p++, val);
1020 }
1021 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1022 // Divide stack into {S0,...,S1}, where S0 is set to top.
1023 uint s1 = stack_slots_not_pruned;
1024 stack_slots_not_pruned = 0; // for next iteration
1025 if (s1 > l) s1 = l;
1026 uint s0 = l - s1;
1027 p += s0; // skip the tops preinstalled by add_req_batch
1028 for (j = s0; j < l; j++)
1029 call->set_req(p++, in_map->in(k+j));
1030 } else {
1031 p += l; // already set to top above by add_req_batch
1032 }
1033
1034 // Add the Monitors
1035 k = in_jvms->monoff();
1036 l = in_jvms->mon_size();
1037 out_jvms->set_monoff(p);
1038 for (j = 0; j < l; j++)
1039 call->set_req(p++, in_map->in(k+j));
1040
1041 // Copy any scalar object fields.
1042 k = in_jvms->scloff();
1043 l = in_jvms->scl_size();
1044 out_jvms->set_scloff(p);
1045 for (j = 0; j < l; j++)
1046 call->set_req(p++, in_map->in(k+j));
1047
1048 // Finish the new jvms.
1049 out_jvms->set_endoff(p);
1050
1051 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1052 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1053 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1054 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1055 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1056 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1057
1058 // Update the two tail pointers in parallel.
1059 callee_jvms = out_jvms;
1060 out_jvms = out_jvms->caller();
1061 in_jvms = in_jvms->caller();
1062 }
1063
1064 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1065
1066 // Test the correctness of JVMState::debug_xxx accessors:
1067 assert(call->jvms()->debug_start() == non_debug_edges, "");
1068 assert(call->jvms()->debug_end() == call->req(), "");
1069 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1070 }
1071
1072 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1073 Bytecodes::Code code = java_bc();
1074 if (code == Bytecodes::_wide) {
1075 code = method()->java_code_at_bci(bci() + 1);
1076 }
1077
1078 if (code != Bytecodes::_illegal) {
1079 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1080 }
1081
1082 auto rsize = [&]() {
1083 assert(code != Bytecodes::_illegal, "code is illegal!");
1084 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1085 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1086 };
1087
1088 switch (code) {
1089 case Bytecodes::_illegal:
1090 return false;
1091
1092 case Bytecodes::_ldc:
1093 case Bytecodes::_ldc_w:
1094 case Bytecodes::_ldc2_w:
1095 inputs = 0;
1096 break;
1097
1098 case Bytecodes::_dup: inputs = 1; break;
1099 case Bytecodes::_dup_x1: inputs = 2; break;
1100 case Bytecodes::_dup_x2: inputs = 3; break;
1101 case Bytecodes::_dup2: inputs = 2; break;
1102 case Bytecodes::_dup2_x1: inputs = 3; break;
1103 case Bytecodes::_dup2_x2: inputs = 4; break;
1104 case Bytecodes::_swap: inputs = 2; break;
1105 case Bytecodes::_arraylength: inputs = 1; break;
1106
1107 case Bytecodes::_getstatic:
1108 case Bytecodes::_putstatic:
1109 case Bytecodes::_getfield:
1110 case Bytecodes::_putfield:
1111 {
1112 bool ignored_will_link;
1113 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1114 int size = field->type()->size();
1115 bool is_get = (depth >= 0), is_static = (depth & 1);
1116 inputs = (is_static ? 0 : 1);
1117 if (is_get) {
1118 depth = size - inputs;
1119 } else {
1120 inputs += size; // putxxx pops the value from the stack
1121 depth = - inputs;
1122 }
1123 }
1124 break;
1125
1126 case Bytecodes::_invokevirtual:
1127 case Bytecodes::_invokespecial:
1128 case Bytecodes::_invokestatic:
1129 case Bytecodes::_invokedynamic:
1130 case Bytecodes::_invokeinterface:
1131 {
1132 bool ignored_will_link;
1133 ciSignature* declared_signature = nullptr;
1134 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1135 assert(declared_signature != nullptr, "cannot be null");
1136 inputs = declared_signature->arg_size_for_bc(code);
1137 int size = declared_signature->return_type()->size();
1138 depth = size - inputs;
1139 }
1140 break;
1141
1142 case Bytecodes::_multianewarray:
1143 {
1144 ciBytecodeStream iter(method());
1145 iter.reset_to_bci(bci());
1146 iter.next();
1147 inputs = iter.get_dimensions();
1148 assert(rsize() == 1, "");
1149 depth = 1 - inputs;
1150 }
1151 break;
1152
1153 case Bytecodes::_ireturn:
1154 case Bytecodes::_lreturn:
1155 case Bytecodes::_freturn:
1156 case Bytecodes::_dreturn:
1157 case Bytecodes::_areturn:
1158 assert(rsize() == -depth, "");
1159 inputs = -depth;
1160 break;
1161
1162 case Bytecodes::_jsr:
1163 case Bytecodes::_jsr_w:
1164 inputs = 0;
1165 depth = 1; // S.B. depth=1, not zero
1166 break;
1167
1168 default:
1169 // bytecode produces a typed result
1170 inputs = rsize() - depth;
1171 assert(inputs >= 0, "");
1172 break;
1173 }
1174
1175 #ifdef ASSERT
1176 // spot check
1177 int outputs = depth + inputs;
1178 assert(outputs >= 0, "sanity");
1179 switch (code) {
1180 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1181 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1182 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1183 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1184 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1185 default: break;
1186 }
1187 #endif //ASSERT
1188
1189 return true;
1190 }
1191
1192
1193
1194 //------------------------------basic_plus_adr---------------------------------
1195 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1196 // short-circuit a common case
1197 if (offset == intcon(0)) return ptr;
1198 return _gvn.transform( new AddPNode(base, ptr, offset) );
1199 }
1200
1201 Node* GraphKit::ConvI2L(Node* offset) {
1202 // short-circuit a common case
1203 jint offset_con = find_int_con(offset, Type::OffsetBot);
1204 if (offset_con != Type::OffsetBot) {
1205 return longcon((jlong) offset_con);
1206 }
1207 return _gvn.transform( new ConvI2LNode(offset));
1208 }
1209
1210 Node* GraphKit::ConvI2UL(Node* offset) {
1211 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1212 if (offset_con != (juint) Type::OffsetBot) {
1213 return longcon((julong) offset_con);
1214 }
1215 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1216 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1217 return _gvn.transform( new AndLNode(conv, mask) );
1218 }
1219
1220 Node* GraphKit::ConvL2I(Node* offset) {
1221 // short-circuit a common case
1222 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1223 if (offset_con != (jlong)Type::OffsetBot) {
1224 return intcon((int) offset_con);
1225 }
1226 return _gvn.transform( new ConvL2INode(offset));
1227 }
1228
1229 //-------------------------load_object_klass-----------------------------------
1230 Node* GraphKit::load_object_klass(Node* obj) {
1231 // Special-case a fresh allocation to avoid building nodes:
1232 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1233 if (akls != nullptr) return akls;
1234 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1235 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1236 }
1237
1238 //-------------------------load_array_length-----------------------------------
1239 Node* GraphKit::load_array_length(Node* array) {
1240 // Special-case a fresh allocation to avoid building nodes:
1241 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1242 Node *alen;
1243 if (alloc == nullptr) {
1244 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1245 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1246 } else {
1247 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1248 }
1249 return alen;
1250 }
1251
1252 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1253 const TypeOopPtr* oop_type,
1254 bool replace_length_in_map) {
1255 Node* length = alloc->Ideal_length();
1256 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1257 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1258 if (ccast != length) {
1259 // do not transform ccast here, it might convert to top node for
1260 // negative array length and break assumptions in parsing stage.
1261 _gvn.set_type_bottom(ccast);
1262 record_for_igvn(ccast);
1263 if (replace_length_in_map) {
1264 replace_in_map(length, ccast);
1265 }
1266 return ccast;
1267 }
1268 }
1269 return length;
1270 }
1271
1272 //------------------------------do_null_check----------------------------------
1273 // Helper function to do a null pointer check. Returned value is
1274 // the incoming address with null casted away. You are allowed to use the
1275 // not-null value only if you are control dependent on the test.
1276 #ifndef PRODUCT
1277 extern uint explicit_null_checks_inserted,
1278 explicit_null_checks_elided;
1279 #endif
1280 Node* GraphKit::null_check_common(Node* value, BasicType type,
1281 // optional arguments for variations:
1282 bool assert_null,
1283 Node* *null_control,
1284 bool speculative,
1285 bool is_init_check) {
1286 assert(!assert_null || null_control == nullptr, "not both at once");
1287 if (stopped()) return top();
1288 NOT_PRODUCT(explicit_null_checks_inserted++);
1289
1290 if (value->is_InlineType()) {
1291 // Null checking a scalarized but nullable inline type. Check the IsInit
1292 // input instead of the oop input to avoid keeping buffer allocations alive.
1293 InlineTypeNode* vtptr = value->as_InlineType();
1294 while (vtptr->get_oop()->is_InlineType()) {
1295 vtptr = vtptr->get_oop()->as_InlineType();
1296 }
1297 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1298 if (stopped()) {
1299 return top();
1300 }
1301 if (assert_null) {
1302 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1303 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1304 // replace_in_map(value, vtptr);
1305 // return vtptr;
1306 replace_in_map(value, null());
1307 return null();
1308 }
1309 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1310 return cast_not_null(value, do_replace_in_map);
1311 }
1312
1313 // Construct null check
1314 Node *chk = nullptr;
1315 switch(type) {
1316 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1317 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1318 case T_ARRAY : // fall through
1319 type = T_OBJECT; // simplify further tests
1320 case T_OBJECT : {
1321 const Type *t = _gvn.type( value );
1322
1323 const TypeOopPtr* tp = t->isa_oopptr();
1324 if (tp != nullptr && !tp->is_loaded()
1325 // Only for do_null_check, not any of its siblings:
1326 && !assert_null && null_control == nullptr) {
1327 // Usually, any field access or invocation on an unloaded oop type
1328 // will simply fail to link, since the statically linked class is
1329 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1330 // the static class is loaded but the sharper oop type is not.
1331 // Rather than checking for this obscure case in lots of places,
1332 // we simply observe that a null check on an unloaded class
1333 // will always be followed by a nonsense operation, so we
1334 // can just issue the uncommon trap here.
1335 // Our access to the unloaded class will only be correct
1336 // after it has been loaded and initialized, which requires
1337 // a trip through the interpreter.
1338 ciKlass* klass = tp->unloaded_klass();
1339 #ifndef PRODUCT
1340 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1341 #endif
1342 uncommon_trap(Deoptimization::Reason_unloaded,
1343 Deoptimization::Action_reinterpret,
1344 klass, "!loaded");
1345 return top();
1346 }
1347
1348 if (assert_null) {
1349 // See if the type is contained in NULL_PTR.
1350 // If so, then the value is already null.
1351 if (t->higher_equal(TypePtr::NULL_PTR)) {
1352 NOT_PRODUCT(explicit_null_checks_elided++);
1353 return value; // Elided null assert quickly!
1354 }
1355 } else {
1356 // See if mixing in the null pointer changes type.
1357 // If so, then the null pointer was not allowed in the original
1358 // type. In other words, "value" was not-null.
1359 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1360 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1361 NOT_PRODUCT(explicit_null_checks_elided++);
1362 return value; // Elided null check quickly!
1363 }
1364 }
1365 chk = new CmpPNode( value, null() );
1366 break;
1367 }
1368
1369 default:
1370 fatal("unexpected type: %s", type2name(type));
1371 }
1372 assert(chk != nullptr, "sanity check");
1373 chk = _gvn.transform(chk);
1374
1375 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1376 BoolNode *btst = new BoolNode( chk, btest);
1377 Node *tst = _gvn.transform( btst );
1378
1379 //-----------
1380 // if peephole optimizations occurred, a prior test existed.
1381 // If a prior test existed, maybe it dominates as we can avoid this test.
1382 if (tst != btst && type == T_OBJECT) {
1383 // At this point we want to scan up the CFG to see if we can
1384 // find an identical test (and so avoid this test altogether).
1385 Node *cfg = control();
1386 int depth = 0;
1387 while( depth < 16 ) { // Limit search depth for speed
1388 if( cfg->Opcode() == Op_IfTrue &&
1389 cfg->in(0)->in(1) == tst ) {
1390 // Found prior test. Use "cast_not_null" to construct an identical
1391 // CastPP (and hence hash to) as already exists for the prior test.
1392 // Return that casted value.
1393 if (assert_null) {
1394 replace_in_map(value, null());
1395 return null(); // do not issue the redundant test
1396 }
1397 Node *oldcontrol = control();
1398 set_control(cfg);
1399 Node *res = cast_not_null(value);
1400 set_control(oldcontrol);
1401 NOT_PRODUCT(explicit_null_checks_elided++);
1402 return res;
1403 }
1404 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1405 if (cfg == nullptr) break; // Quit at region nodes
1406 depth++;
1407 }
1408 }
1409
1410 //-----------
1411 // Branch to failure if null
1412 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1413 Deoptimization::DeoptReason reason;
1414 if (assert_null) {
1415 reason = Deoptimization::reason_null_assert(speculative);
1416 } else if (type == T_OBJECT || is_init_check) {
1417 reason = Deoptimization::reason_null_check(speculative);
1418 } else {
1419 reason = Deoptimization::Reason_div0_check;
1420 }
1421 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1422 // ciMethodData::has_trap_at will return a conservative -1 if any
1423 // must-be-null assertion has failed. This could cause performance
1424 // problems for a method after its first do_null_assert failure.
1425 // Consider using 'Reason_class_check' instead?
1426
1427 // To cause an implicit null check, we set the not-null probability
1428 // to the maximum (PROB_MAX). For an explicit check the probability
1429 // is set to a smaller value.
1430 if (null_control != nullptr || too_many_traps(reason)) {
1431 // probability is less likely
1432 ok_prob = PROB_LIKELY_MAG(3);
1433 } else if (!assert_null &&
1434 (ImplicitNullCheckThreshold > 0) &&
1435 method() != nullptr &&
1436 (method()->method_data()->trap_count(reason)
1437 >= (uint)ImplicitNullCheckThreshold)) {
1438 ok_prob = PROB_LIKELY_MAG(3);
1439 }
1440
1441 if (null_control != nullptr) {
1442 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1443 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1444 set_control( _gvn.transform( new IfTrueNode(iff)));
1445 #ifndef PRODUCT
1446 if (null_true == top()) {
1447 explicit_null_checks_elided++;
1448 }
1449 #endif
1450 (*null_control) = null_true;
1451 } else {
1452 BuildCutout unless(this, tst, ok_prob);
1453 // Check for optimizer eliding test at parse time
1454 if (stopped()) {
1455 // Failure not possible; do not bother making uncommon trap.
1456 NOT_PRODUCT(explicit_null_checks_elided++);
1457 } else if (assert_null) {
1458 uncommon_trap(reason,
1459 Deoptimization::Action_make_not_entrant,
1460 nullptr, "assert_null");
1461 } else {
1462 replace_in_map(value, zerocon(type));
1463 builtin_throw(reason);
1464 }
1465 }
1466
1467 // Must throw exception, fall-thru not possible?
1468 if (stopped()) {
1469 return top(); // No result
1470 }
1471
1472 if (assert_null) {
1473 // Cast obj to null on this path.
1474 replace_in_map(value, zerocon(type));
1475 return zerocon(type);
1476 }
1477
1478 // Cast obj to not-null on this path, if there is no null_control.
1479 // (If there is a null_control, a non-null value may come back to haunt us.)
1480 if (type == T_OBJECT) {
1481 Node* cast = cast_not_null(value, false);
1482 if (null_control == nullptr || (*null_control) == top())
1483 replace_in_map(value, cast);
1484 value = cast;
1485 }
1486
1487 return value;
1488 }
1489
1490 //------------------------------cast_not_null----------------------------------
1491 // Cast obj to not-null on this path
1492 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1493 if (obj->is_InlineType()) {
1494 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1495 vt->as_InlineType()->set_is_init(_gvn);
1496 vt = _gvn.transform(vt);
1497 if (do_replace_in_map) {
1498 replace_in_map(obj, vt);
1499 }
1500 return vt;
1501 }
1502 const Type *t = _gvn.type(obj);
1503 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1504 // Object is already not-null?
1505 if( t == t_not_null ) return obj;
1506
1507 Node* cast = new CastPPNode(control(), obj,t_not_null);
1508 cast = _gvn.transform( cast );
1509
1510 // Scan for instances of 'obj' in the current JVM mapping.
1511 // These instances are known to be not-null after the test.
1512 if (do_replace_in_map)
1513 replace_in_map(obj, cast);
1514
1515 return cast; // Return casted value
1516 }
1517
1518 // Sometimes in intrinsics, we implicitly know an object is not null
1519 // (there's no actual null check) so we can cast it to not null. In
1520 // the course of optimizations, the input to the cast can become null.
1521 // In that case that data path will die and we need the control path
1522 // to become dead as well to keep the graph consistent. So we have to
1523 // add a check for null for which one branch can't be taken. It uses
1524 // an OpaqueNotNull node that will cause the check to be removed after loop
1525 // opts so the test goes away and the compiled code doesn't execute a
1526 // useless check.
1527 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1528 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1529 return value;
1530 }
1531 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1532 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1533 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1534 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1535 _gvn.set_type(iff, iff->Value(&_gvn));
1536 if (!tst->is_Con()) {
1537 record_for_igvn(iff);
1538 }
1539 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1540 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1541 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1542 C->root()->add_req(halt);
1543 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1544 set_control(if_t);
1545 return cast_not_null(value, do_replace_in_map);
1546 }
1547
1548
1549 //--------------------------replace_in_map-------------------------------------
1550 void GraphKit::replace_in_map(Node* old, Node* neww) {
1551 if (old == neww) {
1552 return;
1553 }
1554
1555 map()->replace_edge(old, neww);
1556
1557 // Note: This operation potentially replaces any edge
1558 // on the map. This includes locals, stack, and monitors
1559 // of the current (innermost) JVM state.
1560
1561 // don't let inconsistent types from profiling escape this
1562 // method
1563
1564 const Type* told = _gvn.type(old);
1565 const Type* tnew = _gvn.type(neww);
1566
1567 if (!tnew->higher_equal(told)) {
1568 return;
1569 }
1570
1571 map()->record_replaced_node(old, neww);
1572 }
1573
1574
1575 //=============================================================================
1576 //--------------------------------memory---------------------------------------
1577 Node* GraphKit::memory(uint alias_idx) {
1578 MergeMemNode* mem = merged_memory();
1579 Node* p = mem->memory_at(alias_idx);
1580 assert(p != mem->empty_memory(), "empty");
1581 _gvn.set_type(p, Type::MEMORY); // must be mapped
1582 return p;
1583 }
1584
1585 //-----------------------------reset_memory------------------------------------
1586 Node* GraphKit::reset_memory() {
1587 Node* mem = map()->memory();
1588 // do not use this node for any more parsing!
1589 debug_only( map()->set_memory((Node*)nullptr) );
1590 return _gvn.transform( mem );
1591 }
1592
1593 //------------------------------set_all_memory---------------------------------
1594 void GraphKit::set_all_memory(Node* newmem) {
1595 Node* mergemem = MergeMemNode::make(newmem);
1596 gvn().set_type_bottom(mergemem);
1597 map()->set_memory(mergemem);
1598 }
1599
1600 //------------------------------set_all_memory_call----------------------------
1601 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1602 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1603 set_all_memory(newmem);
1604 }
1605
1606 //=============================================================================
1607 //
1608 // parser factory methods for MemNodes
1609 //
1610 // These are layered on top of the factory methods in LoadNode and StoreNode,
1611 // and integrate with the parser's memory state and _gvn engine.
1612 //
1613
1614 // factory methods in "int adr_idx"
1615 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1616 MemNode::MemOrd mo,
1617 LoadNode::ControlDependency control_dependency,
1618 bool require_atomic_access,
1619 bool unaligned,
1620 bool mismatched,
1621 bool unsafe,
1622 uint8_t barrier_data) {
1623 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1624 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1625 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1626 debug_only(adr_type = C->get_adr_type(adr_idx));
1627 Node* mem = memory(adr_idx);
1628 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1629 ld = _gvn.transform(ld);
1630
1631 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1632 // Improve graph before escape analysis and boxing elimination.
1633 record_for_igvn(ld);
1634 if (ld->is_DecodeN()) {
1635 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1636 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1637 // a Phi). Recording such cases is still perfectly sound, but may be
1638 // unnecessary and result in some minor IGVN overhead.
1639 record_for_igvn(ld->in(1));
1640 }
1641 }
1642 return ld;
1643 }
1644
1645 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1646 MemNode::MemOrd mo,
1647 bool require_atomic_access,
1648 bool unaligned,
1649 bool mismatched,
1650 bool unsafe,
1651 int barrier_data) {
1652 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1653 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1654 const TypePtr* adr_type = nullptr;
1655 debug_only(adr_type = C->get_adr_type(adr_idx));
1656 Node *mem = memory(adr_idx);
1657 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1658 if (unaligned) {
1659 st->as_Store()->set_unaligned_access();
1660 }
1661 if (mismatched) {
1662 st->as_Store()->set_mismatched_access();
1663 }
1664 if (unsafe) {
1665 st->as_Store()->set_unsafe_access();
1666 }
1667 st->as_Store()->set_barrier_data(barrier_data);
1668 st = _gvn.transform(st);
1669 set_memory(st, adr_idx);
1670 // Back-to-back stores can only remove intermediate store with DU info
1671 // so push on worklist for optimizer.
1672 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1673 record_for_igvn(st);
1674
1675 return st;
1676 }
1677
1678 Node* GraphKit::access_store_at(Node* obj,
1679 Node* adr,
1680 const TypePtr* adr_type,
1681 Node* val,
1682 const Type* val_type,
1683 BasicType bt,
1684 DecoratorSet decorators,
1685 bool safe_for_replace,
1686 const InlineTypeNode* vt) {
1687 // Transformation of a value which could be null pointer (CastPP #null)
1688 // could be delayed during Parse (for example, in adjust_map_after_if()).
1689 // Execute transformation here to avoid barrier generation in such case.
1690 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1691 val = _gvn.makecon(TypePtr::NULL_PTR);
1692 }
1693
1694 if (stopped()) {
1695 return top(); // Dead path ?
1696 }
1697
1698 assert(val != nullptr, "not dead path");
1699 if (val->is_InlineType()) {
1700 // Store to non-flat field. Buffer the inline type and make sure
1701 // the store is re-executed if the allocation triggers deoptimization.
1702 PreserveReexecuteState preexecs(this);
1703 jvms()->set_should_reexecute(true);
1704 val = val->as_InlineType()->buffer(this, safe_for_replace);
1705 }
1706
1707 C2AccessValuePtr addr(adr, adr_type);
1708 C2AccessValue value(val, val_type);
1709 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1710 if (access.is_raw()) {
1711 return _barrier_set->BarrierSetC2::store_at(access, value);
1712 } else {
1713 return _barrier_set->store_at(access, value);
1714 }
1715 }
1716
1717 Node* GraphKit::access_load_at(Node* obj, // containing obj
1718 Node* adr, // actual address to store val at
1719 const TypePtr* adr_type,
1720 const Type* val_type,
1721 BasicType bt,
1722 DecoratorSet decorators,
1723 Node* ctl) {
1724 if (stopped()) {
1725 return top(); // Dead path ?
1726 }
1727
1728 C2AccessValuePtr addr(adr, adr_type);
1729 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1730 if (access.is_raw()) {
1731 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1732 } else {
1733 return _barrier_set->load_at(access, val_type);
1734 }
1735 }
1736
1737 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1738 const Type* val_type,
1739 BasicType bt,
1740 DecoratorSet decorators) {
1741 if (stopped()) {
1742 return top(); // Dead path ?
1743 }
1744
1745 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1746 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1747 if (access.is_raw()) {
1748 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1749 } else {
1750 return _barrier_set->load_at(access, val_type);
1751 }
1752 }
1753
1754 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1755 Node* adr,
1756 const TypePtr* adr_type,
1757 int alias_idx,
1758 Node* expected_val,
1759 Node* new_val,
1760 const Type* value_type,
1761 BasicType bt,
1762 DecoratorSet decorators) {
1763 C2AccessValuePtr addr(adr, adr_type);
1764 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1765 bt, obj, addr, alias_idx);
1766 if (access.is_raw()) {
1767 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1768 } else {
1769 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1770 }
1771 }
1772
1773 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1774 Node* adr,
1775 const TypePtr* adr_type,
1776 int alias_idx,
1777 Node* expected_val,
1778 Node* new_val,
1779 const Type* value_type,
1780 BasicType bt,
1781 DecoratorSet decorators) {
1782 C2AccessValuePtr addr(adr, adr_type);
1783 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1784 bt, obj, addr, alias_idx);
1785 if (access.is_raw()) {
1786 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1787 } else {
1788 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1789 }
1790 }
1791
1792 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1793 Node* adr,
1794 const TypePtr* adr_type,
1795 int alias_idx,
1796 Node* new_val,
1797 const Type* value_type,
1798 BasicType bt,
1799 DecoratorSet decorators) {
1800 C2AccessValuePtr addr(adr, adr_type);
1801 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1802 bt, obj, addr, alias_idx);
1803 if (access.is_raw()) {
1804 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1805 } else {
1806 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1807 }
1808 }
1809
1810 Node* GraphKit::access_atomic_add_at(Node* obj,
1811 Node* adr,
1812 const TypePtr* adr_type,
1813 int alias_idx,
1814 Node* new_val,
1815 const Type* value_type,
1816 BasicType bt,
1817 DecoratorSet decorators) {
1818 C2AccessValuePtr addr(adr, adr_type);
1819 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1820 if (access.is_raw()) {
1821 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1822 } else {
1823 return _barrier_set->atomic_add_at(access, new_val, value_type);
1824 }
1825 }
1826
1827 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1828 return _barrier_set->clone(this, src, dst, size, is_array);
1829 }
1830
1831 //-------------------------array_element_address-------------------------
1832 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1833 const TypeInt* sizetype, Node* ctrl) {
1834 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1835 uint shift;
1836 if (arytype->is_flat() && arytype->klass_is_exact()) {
1837 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1838 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1839 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1840 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1841 // though we don't need the address node in this case and throw it away again.
1842 shift = arytype->flat_log_elem_size();
1843 } else {
1844 shift = exact_log2(type2aelembytes(elembt));
1845 }
1846 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1847
1848 // short-circuit a common case (saves lots of confusing waste motion)
1849 jint idx_con = find_int_con(idx, -1);
1850 if (idx_con >= 0) {
1851 intptr_t offset = header + ((intptr_t)idx_con << shift);
1852 return basic_plus_adr(ary, offset);
1853 }
1854
1855 // must be correct type for alignment purposes
1856 Node* base = basic_plus_adr(ary, header);
1857 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1858 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1859 return basic_plus_adr(ary, base, scale);
1860 }
1861
1862 Node* GraphKit::flat_array_element_address(Node*& array, Node* idx, ciInlineKlass* vk, bool is_null_free,
1863 bool is_not_null_free, bool is_atomic) {
1864 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* flat */ true, is_null_free, is_atomic);
1865 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1866 arytype = arytype->cast_to_exactness(true);
1867 arytype = arytype->cast_to_not_null_free(is_not_null_free);
1868 array = _gvn.transform(new CheckCastPPNode(control(), array, arytype));
1869 return array_element_address(array, idx, T_FLAT_ELEMENT, arytype->size(), control());
1870 }
1871
1872 //-------------------------load_array_element-------------------------
1873 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1874 const Type* elemtype = arytype->elem();
1875 BasicType elembt = elemtype->array_element_basic_type();
1876 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1877 if (elembt == T_NARROWOOP) {
1878 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1879 }
1880 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1881 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1882 return ld;
1883 }
1884
1885 //-------------------------set_arguments_for_java_call-------------------------
1886 // Arguments (pre-popped from the stack) are taken from the JVMS.
1887 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1888 PreserveReexecuteState preexecs(this);
1889 if (EnableValhalla) {
1890 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1891 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1892 jvms()->set_should_reexecute(true);
1893 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1894 inc_sp(arg_size);
1895 }
1896 // Add the call arguments
1897 const TypeTuple* domain = call->tf()->domain_sig();
1898 uint nargs = domain->cnt();
1899 int arg_num = 0;
1900 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1901 Node* arg = argument(i-TypeFunc::Parms);
1902 const Type* t = domain->field_at(i);
1903 // TODO 8284443 A static call to a mismatched method should still be scalarized
1904 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1905 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1906 if (!arg->is_InlineType()) {
1907 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1908 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1909 }
1910 InlineTypeNode* vt = arg->as_InlineType();
1911 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1912 // If an inline type argument is passed as fields, attach the Method* to the call site
1913 // to be able to access the extended signature later via attached_method_before_pc().
1914 // For example, see CompiledMethod::preserve_callee_argument_oops().
1915 call->set_override_symbolic_info(true);
1916 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1917 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1918 C->dependencies()->assert_evol_method(call->method());
1919 arg_num++;
1920 continue;
1921 } else if (arg->is_InlineType()) {
1922 // Pass inline type argument via oop to callee
1923 InlineTypeNode* inline_type = arg->as_InlineType();
1924 const ciMethod* method = call->method();
1925 ciInstanceKlass* holder = method->holder();
1926 const bool is_receiver = (i == TypeFunc::Parms);
1927 const bool is_abstract_or_object_klass_constructor = method->is_object_constructor() &&
1928 (holder->is_abstract() || holder->is_java_lang_Object());
1929 const bool is_larval_receiver_on_super_constructor = is_receiver && is_abstract_or_object_klass_constructor;
1930 bool must_init_buffer = true;
1931 // We always need to buffer inline types when they are escaping. However, we can skip the actual initialization
1932 // of the buffer if the inline type is a larval because we are going to update the buffer anyway which requires
1933 // us to create a new one. But there is one special case where we are still required to initialize the buffer:
1934 // When we have a larval receiver invoked on an abstract (value class) constructor or the Object constructor (that
1935 // is not going to be inlined). After this call, the larval is completely initialized and thus not a larval anymore.
1936 // We therefore need to force an initialization of the buffer to not lose all the field writes so far in case the
1937 // buffer needs to be used (e.g. to read from when deoptimizing at runtime) or further updated in abstract super
1938 // value class constructors which could have more fields to be initialized. Note that we do not need to
1939 // initialize the buffer when invoking another constructor in the same class on a larval receiver because we
1940 // have not initialized any fields, yet (this is done completely by the other constructor call).
1941 if (inline_type->is_larval() && !is_larval_receiver_on_super_constructor) {
1942 must_init_buffer = false;
1943 }
1944 arg = inline_type->buffer(this, true, must_init_buffer);
1945 }
1946 if (t != Type::HALF) {
1947 arg_num++;
1948 }
1949 call->init_req(idx++, arg);
1950 }
1951 }
1952
1953 //---------------------------set_edges_for_java_call---------------------------
1954 // Connect a newly created call into the current JVMS.
1955 // A return value node (if any) is returned from set_edges_for_java_call.
1956 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1957
1958 // Add the predefined inputs:
1959 call->init_req( TypeFunc::Control, control() );
1960 call->init_req( TypeFunc::I_O , i_o() );
1961 call->init_req( TypeFunc::Memory , reset_memory() );
1962 call->init_req( TypeFunc::FramePtr, frameptr() );
1963 call->init_req( TypeFunc::ReturnAdr, top() );
1964
1965 add_safepoint_edges(call, must_throw);
1966
1967 Node* xcall = _gvn.transform(call);
1968
1969 if (xcall == top()) {
1970 set_control(top());
1971 return;
1972 }
1973 assert(xcall == call, "call identity is stable");
1974
1975 // Re-use the current map to produce the result.
1976
1977 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1978 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1979 set_all_memory_call(xcall, separate_io_proj);
1980
1981 //return xcall; // no need, caller already has it
1982 }
1983
1984 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1985 if (stopped()) return top(); // maybe the call folded up?
1986
1987 // Note: Since any out-of-line call can produce an exception,
1988 // we always insert an I_O projection from the call into the result.
1989
1990 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1991
1992 if (separate_io_proj) {
1993 // The caller requested separate projections be used by the fall
1994 // through and exceptional paths, so replace the projections for
1995 // the fall through path.
1996 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1997 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1998 }
1999
2000 // Capture the return value, if any.
2001 Node* ret;
2002 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2003 ret = top();
2004 } else if (call->tf()->returns_inline_type_as_fields()) {
2005 // Return of multiple values (inline type fields): we create a
2006 // InlineType node, each field is a projection from the call.
2007 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2008 uint base_input = TypeFunc::Parms;
2009 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2010 } else {
2011 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2012 ciType* t = call->method()->return_type();
2013 if (t->is_klass()) {
2014 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2015 if (type->is_inlinetypeptr()) {
2016 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2017 }
2018 }
2019 }
2020
2021 // We just called the constructor on a value type receiver. Reload it from the buffer
2022 ciMethod* method = call->method();
2023 if (method->is_object_constructor() && !method->holder()->is_java_lang_Object()) {
2024 InlineTypeNode* inline_type_receiver = call->in(TypeFunc::Parms)->isa_InlineType();
2025 if (inline_type_receiver != nullptr) {
2026 assert(inline_type_receiver->is_larval(), "must be larval");
2027 assert(inline_type_receiver->is_allocated(&gvn()), "larval must be buffered");
2028 InlineTypeNode* reloaded = InlineTypeNode::make_from_oop(this, inline_type_receiver->get_oop(),
2029 inline_type_receiver->bottom_type()->inline_klass());
2030 assert(!reloaded->is_larval(), "should not be larval anymore");
2031 replace_in_map(inline_type_receiver, reloaded);
2032 }
2033 }
2034
2035 return ret;
2036 }
2037
2038 //--------------------set_predefined_input_for_runtime_call--------------------
2039 // Reading and setting the memory state is way conservative here.
2040 // The real problem is that I am not doing real Type analysis on memory,
2041 // so I cannot distinguish card mark stores from other stores. Across a GC
2042 // point the Store Barrier and the card mark memory has to agree. I cannot
2043 // have a card mark store and its barrier split across the GC point from
2044 // either above or below. Here I get that to happen by reading ALL of memory.
2045 // A better answer would be to separate out card marks from other memory.
2046 // For now, return the input memory state, so that it can be reused
2047 // after the call, if this call has restricted memory effects.
2048 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2049 // Set fixed predefined input arguments
2050 Node* memory = reset_memory();
2051 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2052 call->init_req( TypeFunc::Control, control() );
2053 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2054 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2055 call->init_req( TypeFunc::FramePtr, frameptr() );
2056 call->init_req( TypeFunc::ReturnAdr, top() );
2057 return memory;
2058 }
2059
2060 //-------------------set_predefined_output_for_runtime_call--------------------
2061 // Set control and memory (not i_o) from the call.
2062 // If keep_mem is not null, use it for the output state,
2063 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2064 // If hook_mem is null, this call produces no memory effects at all.
2065 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2066 // then only that memory slice is taken from the call.
2067 // In the last case, we must put an appropriate memory barrier before
2068 // the call, so as to create the correct anti-dependencies on loads
2069 // preceding the call.
2070 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2071 Node* keep_mem,
2072 const TypePtr* hook_mem) {
2073 // no i/o
2074 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2075 if (keep_mem) {
2076 // First clone the existing memory state
2077 set_all_memory(keep_mem);
2078 if (hook_mem != nullptr) {
2079 // Make memory for the call
2080 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2081 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2082 // We also use hook_mem to extract specific effects from arraycopy stubs.
2083 set_memory(mem, hook_mem);
2084 }
2085 // ...else the call has NO memory effects.
2086
2087 // Make sure the call advertises its memory effects precisely.
2088 // This lets us build accurate anti-dependences in gcm.cpp.
2089 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2090 "call node must be constructed correctly");
2091 } else {
2092 assert(hook_mem == nullptr, "");
2093 // This is not a "slow path" call; all memory comes from the call.
2094 set_all_memory_call(call);
2095 }
2096 }
2097
2098 // Keep track of MergeMems feeding into other MergeMems
2099 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2100 if (!mem->is_MergeMem()) {
2101 return;
2102 }
2103 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2104 Node* use = i.get();
2105 if (use->is_MergeMem()) {
2106 wl.push(use);
2107 }
2108 }
2109 }
2110
2111 // Replace the call with the current state of the kit.
2112 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2113 JVMState* ejvms = nullptr;
2114 if (has_exceptions()) {
2115 ejvms = transfer_exceptions_into_jvms();
2116 }
2117
2118 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2119 ReplacedNodes replaced_nodes_exception;
2120 Node* ex_ctl = top();
2121
2122 SafePointNode* final_state = stop();
2123
2124 // Find all the needed outputs of this call
2125 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2126
2127 Unique_Node_List wl;
2128 Node* init_mem = call->in(TypeFunc::Memory);
2129 Node* final_mem = final_state->in(TypeFunc::Memory);
2130 Node* final_ctl = final_state->in(TypeFunc::Control);
2131 Node* final_io = final_state->in(TypeFunc::I_O);
2132
2133 // Replace all the old call edges with the edges from the inlining result
2134 if (callprojs->fallthrough_catchproj != nullptr) {
2135 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2136 }
2137 if (callprojs->fallthrough_memproj != nullptr) {
2138 if (final_mem->is_MergeMem()) {
2139 // Parser's exits MergeMem was not transformed but may be optimized
2140 final_mem = _gvn.transform(final_mem);
2141 }
2142 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2143 add_mergemem_users_to_worklist(wl, final_mem);
2144 }
2145 if (callprojs->fallthrough_ioproj != nullptr) {
2146 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2147 }
2148
2149 // Replace the result with the new result if it exists and is used
2150 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2151 // If the inlined code is dead, the result projections for an inline type returned as
2152 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2153 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2154 "unexpected number of results");
2155 C->gvn_replace_by(callprojs->resproj[0], result);
2156 }
2157
2158 if (ejvms == nullptr) {
2159 // No exception edges to simply kill off those paths
2160 if (callprojs->catchall_catchproj != nullptr) {
2161 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2162 }
2163 if (callprojs->catchall_memproj != nullptr) {
2164 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2165 }
2166 if (callprojs->catchall_ioproj != nullptr) {
2167 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2168 }
2169 // Replace the old exception object with top
2170 if (callprojs->exobj != nullptr) {
2171 C->gvn_replace_by(callprojs->exobj, C->top());
2172 }
2173 } else {
2174 GraphKit ekit(ejvms);
2175
2176 // Load my combined exception state into the kit, with all phis transformed:
2177 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2178 replaced_nodes_exception = ex_map->replaced_nodes();
2179
2180 Node* ex_oop = ekit.use_exception_state(ex_map);
2181
2182 if (callprojs->catchall_catchproj != nullptr) {
2183 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2184 ex_ctl = ekit.control();
2185 }
2186 if (callprojs->catchall_memproj != nullptr) {
2187 Node* ex_mem = ekit.reset_memory();
2188 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2189 add_mergemem_users_to_worklist(wl, ex_mem);
2190 }
2191 if (callprojs->catchall_ioproj != nullptr) {
2192 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2193 }
2194
2195 // Replace the old exception object with the newly created one
2196 if (callprojs->exobj != nullptr) {
2197 C->gvn_replace_by(callprojs->exobj, ex_oop);
2198 }
2199 }
2200
2201 // Disconnect the call from the graph
2202 call->disconnect_inputs(C);
2203 C->gvn_replace_by(call, C->top());
2204
2205 // Clean up any MergeMems that feed other MergeMems since the
2206 // optimizer doesn't like that.
2207 while (wl.size() > 0) {
2208 _gvn.transform(wl.pop());
2209 }
2210
2211 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2212 replaced_nodes.apply(C, final_ctl);
2213 }
2214 if (!ex_ctl->is_top() && do_replaced_nodes) {
2215 replaced_nodes_exception.apply(C, ex_ctl);
2216 }
2217 }
2218
2219
2220 //------------------------------increment_counter------------------------------
2221 // for statistics: increment a VM counter by 1
2222
2223 void GraphKit::increment_counter(address counter_addr) {
2224 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2225 increment_counter(adr1);
2226 }
2227
2228 void GraphKit::increment_counter(Node* counter_addr) {
2229 Node* ctrl = control();
2230 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2231 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2232 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2233 }
2234
2235
2236 //------------------------------uncommon_trap----------------------------------
2237 // Bail out to the interpreter in mid-method. Implemented by calling the
2238 // uncommon_trap blob. This helper function inserts a runtime call with the
2239 // right debug info.
2240 Node* GraphKit::uncommon_trap(int trap_request,
2241 ciKlass* klass, const char* comment,
2242 bool must_throw,
2243 bool keep_exact_action) {
2244 if (failing_internal()) {
2245 stop();
2246 }
2247 if (stopped()) return nullptr; // trap reachable?
2248
2249 // Note: If ProfileTraps is true, and if a deopt. actually
2250 // occurs here, the runtime will make sure an MDO exists. There is
2251 // no need to call method()->ensure_method_data() at this point.
2252
2253 // Set the stack pointer to the right value for reexecution:
2254 set_sp(reexecute_sp());
2255
2256 #ifdef ASSERT
2257 if (!must_throw) {
2258 // Make sure the stack has at least enough depth to execute
2259 // the current bytecode.
2260 int inputs, ignored_depth;
2261 if (compute_stack_effects(inputs, ignored_depth)) {
2262 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2263 Bytecodes::name(java_bc()), sp(), inputs);
2264 }
2265 }
2266 #endif
2267
2268 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2269 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2270
2271 switch (action) {
2272 case Deoptimization::Action_maybe_recompile:
2273 case Deoptimization::Action_reinterpret:
2274 // Temporary fix for 6529811 to allow virtual calls to be sure they
2275 // get the chance to go from mono->bi->mega
2276 if (!keep_exact_action &&
2277 Deoptimization::trap_request_index(trap_request) < 0 &&
2278 too_many_recompiles(reason)) {
2279 // This BCI is causing too many recompilations.
2280 if (C->log() != nullptr) {
2281 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2282 Deoptimization::trap_reason_name(reason),
2283 Deoptimization::trap_action_name(action));
2284 }
2285 action = Deoptimization::Action_none;
2286 trap_request = Deoptimization::make_trap_request(reason, action);
2287 } else {
2288 C->set_trap_can_recompile(true);
2289 }
2290 break;
2291 case Deoptimization::Action_make_not_entrant:
2292 C->set_trap_can_recompile(true);
2293 break;
2294 case Deoptimization::Action_none:
2295 case Deoptimization::Action_make_not_compilable:
2296 break;
2297 default:
2298 #ifdef ASSERT
2299 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2300 #endif
2301 break;
2302 }
2303
2304 if (TraceOptoParse) {
2305 char buf[100];
2306 tty->print_cr("Uncommon trap %s at bci:%d",
2307 Deoptimization::format_trap_request(buf, sizeof(buf),
2308 trap_request), bci());
2309 }
2310
2311 CompileLog* log = C->log();
2312 if (log != nullptr) {
2313 int kid = (klass == nullptr)? -1: log->identify(klass);
2314 log->begin_elem("uncommon_trap bci='%d'", bci());
2315 char buf[100];
2316 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2317 trap_request));
2318 if (kid >= 0) log->print(" klass='%d'", kid);
2319 if (comment != nullptr) log->print(" comment='%s'", comment);
2320 log->end_elem();
2321 }
2322
2323 // Make sure any guarding test views this path as very unlikely
2324 Node *i0 = control()->in(0);
2325 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2326 IfNode *iff = i0->as_If();
2327 float f = iff->_prob; // Get prob
2328 if (control()->Opcode() == Op_IfTrue) {
2329 if (f > PROB_UNLIKELY_MAG(4))
2330 iff->_prob = PROB_MIN;
2331 } else {
2332 if (f < PROB_LIKELY_MAG(4))
2333 iff->_prob = PROB_MAX;
2334 }
2335 }
2336
2337 // Clear out dead values from the debug info.
2338 kill_dead_locals();
2339
2340 // Now insert the uncommon trap subroutine call
2341 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2342 const TypePtr* no_memory_effects = nullptr;
2343 // Pass the index of the class to be loaded
2344 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2345 (must_throw ? RC_MUST_THROW : 0),
2346 OptoRuntime::uncommon_trap_Type(),
2347 call_addr, "uncommon_trap", no_memory_effects,
2348 intcon(trap_request));
2349 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2350 "must extract request correctly from the graph");
2351 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2352
2353 call->set_req(TypeFunc::ReturnAdr, returnadr());
2354 // The debug info is the only real input to this call.
2355
2356 // Halt-and-catch fire here. The above call should never return!
2357 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2358 PRODUCT_ONLY(COMMA /*reachable*/false));
2359 _gvn.set_type_bottom(halt);
2360 root()->add_req(halt);
2361
2362 stop_and_kill_map();
2363 return call;
2364 }
2365
2366
2367 //--------------------------just_allocated_object------------------------------
2368 // Report the object that was just allocated.
2369 // It must be the case that there are no intervening safepoints.
2370 // We use this to determine if an object is so "fresh" that
2371 // it does not require card marks.
2372 Node* GraphKit::just_allocated_object(Node* current_control) {
2373 Node* ctrl = current_control;
2374 // Object::<init> is invoked after allocation, most of invoke nodes
2375 // will be reduced, but a region node is kept in parse time, we check
2376 // the pattern and skip the region node if it degraded to a copy.
2377 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2378 ctrl->as_Region()->is_copy()) {
2379 ctrl = ctrl->as_Region()->is_copy();
2380 }
2381 if (C->recent_alloc_ctl() == ctrl) {
2382 return C->recent_alloc_obj();
2383 }
2384 return nullptr;
2385 }
2386
2387
2388 /**
2389 * Record profiling data exact_kls for Node n with the type system so
2390 * that it can propagate it (speculation)
2391 *
2392 * @param n node that the type applies to
2393 * @param exact_kls type from profiling
2394 * @param maybe_null did profiling see null?
2395 *
2396 * @return node with improved type
2397 */
2398 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2399 const Type* current_type = _gvn.type(n);
2400 assert(UseTypeSpeculation, "type speculation must be on");
2401
2402 const TypePtr* speculative = current_type->speculative();
2403
2404 // Should the klass from the profile be recorded in the speculative type?
2405 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2406 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2407 const TypeOopPtr* xtype = tklass->as_instance_type();
2408 assert(xtype->klass_is_exact(), "Should be exact");
2409 // Any reason to believe n is not null (from this profiling or a previous one)?
2410 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2411 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2412 // record the new speculative type's depth
2413 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2414 speculative = speculative->with_inline_depth(jvms()->depth());
2415 } else if (current_type->would_improve_ptr(ptr_kind)) {
2416 // Profiling report that null was never seen so we can change the
2417 // speculative type to non null ptr.
2418 if (ptr_kind == ProfileAlwaysNull) {
2419 speculative = TypePtr::NULL_PTR;
2420 } else {
2421 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2422 const TypePtr* ptr = TypePtr::NOTNULL;
2423 if (speculative != nullptr) {
2424 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2425 } else {
2426 speculative = ptr;
2427 }
2428 }
2429 }
2430
2431 if (speculative != current_type->speculative()) {
2432 // Build a type with a speculative type (what we think we know
2433 // about the type but will need a guard when we use it)
2434 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2435 // We're changing the type, we need a new CheckCast node to carry
2436 // the new type. The new type depends on the control: what
2437 // profiling tells us is only valid from here as far as we can
2438 // tell.
2439 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2440 cast = _gvn.transform(cast);
2441 replace_in_map(n, cast);
2442 n = cast;
2443 }
2444
2445 return n;
2446 }
2447
2448 /**
2449 * Record profiling data from receiver profiling at an invoke with the
2450 * type system so that it can propagate it (speculation)
2451 *
2452 * @param n receiver node
2453 *
2454 * @return node with improved type
2455 */
2456 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2457 if (!UseTypeSpeculation) {
2458 return n;
2459 }
2460 ciKlass* exact_kls = profile_has_unique_klass();
2461 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2462 if ((java_bc() == Bytecodes::_checkcast ||
2463 java_bc() == Bytecodes::_instanceof ||
2464 java_bc() == Bytecodes::_aastore) &&
2465 method()->method_data()->is_mature()) {
2466 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2467 if (data != nullptr) {
2468 if (java_bc() == Bytecodes::_aastore) {
2469 ciKlass* array_type = nullptr;
2470 ciKlass* element_type = nullptr;
2471 ProfilePtrKind element_ptr = ProfileMaybeNull;
2472 bool flat_array = true;
2473 bool null_free_array = true;
2474 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2475 exact_kls = element_type;
2476 ptr_kind = element_ptr;
2477 } else {
2478 if (!data->as_BitData()->null_seen()) {
2479 ptr_kind = ProfileNeverNull;
2480 } else {
2481 assert(data->is_ReceiverTypeData(), "bad profile data type");
2482 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2483 uint i = 0;
2484 for (; i < call->row_limit(); i++) {
2485 ciKlass* receiver = call->receiver(i);
2486 if (receiver != nullptr) {
2487 break;
2488 }
2489 }
2490 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2491 }
2492 }
2493 }
2494 }
2495 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2496 }
2497
2498 /**
2499 * Record profiling data from argument profiling at an invoke with the
2500 * type system so that it can propagate it (speculation)
2501 *
2502 * @param dest_method target method for the call
2503 * @param bc what invoke bytecode is this?
2504 */
2505 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2506 if (!UseTypeSpeculation) {
2507 return;
2508 }
2509 const TypeFunc* tf = TypeFunc::make(dest_method);
2510 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2511 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2512 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2513 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2514 if (is_reference_type(targ->basic_type())) {
2515 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2516 ciKlass* better_type = nullptr;
2517 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2518 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2519 }
2520 i++;
2521 }
2522 }
2523 }
2524
2525 /**
2526 * Record profiling data from parameter profiling at an invoke with
2527 * the type system so that it can propagate it (speculation)
2528 */
2529 void GraphKit::record_profiled_parameters_for_speculation() {
2530 if (!UseTypeSpeculation) {
2531 return;
2532 }
2533 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2534 if (_gvn.type(local(i))->isa_oopptr()) {
2535 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2536 ciKlass* better_type = nullptr;
2537 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2538 record_profile_for_speculation(local(i), better_type, ptr_kind);
2539 }
2540 j++;
2541 }
2542 }
2543 }
2544
2545 /**
2546 * Record profiling data from return value profiling at an invoke with
2547 * the type system so that it can propagate it (speculation)
2548 */
2549 void GraphKit::record_profiled_return_for_speculation() {
2550 if (!UseTypeSpeculation) {
2551 return;
2552 }
2553 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2554 ciKlass* better_type = nullptr;
2555 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2556 // If profiling reports a single type for the return value,
2557 // feed it to the type system so it can propagate it as a
2558 // speculative type
2559 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2560 }
2561 }
2562
2563
2564 //=============================================================================
2565 // Generate a fast path/slow path idiom. Graph looks like:
2566 // [foo] indicates that 'foo' is a parameter
2567 //
2568 // [in] null
2569 // \ /
2570 // CmpP
2571 // Bool ne
2572 // If
2573 // / \
2574 // True False-<2>
2575 // / |
2576 // / cast_not_null
2577 // Load | | ^
2578 // [fast_test] | |
2579 // gvn to opt_test | |
2580 // / \ | <1>
2581 // True False |
2582 // | \\ |
2583 // [slow_call] \[fast_result]
2584 // Ctl Val \ \
2585 // | \ \
2586 // Catch <1> \ \
2587 // / \ ^ \ \
2588 // Ex No_Ex | \ \
2589 // | \ \ | \ <2> \
2590 // ... \ [slow_res] | | \ [null_result]
2591 // \ \--+--+--- | |
2592 // \ | / \ | /
2593 // --------Region Phi
2594 //
2595 //=============================================================================
2596 // Code is structured as a series of driver functions all called 'do_XXX' that
2597 // call a set of helper functions. Helper functions first, then drivers.
2598
2599 //------------------------------null_check_oop---------------------------------
2600 // Null check oop. Set null-path control into Region in slot 3.
2601 // Make a cast-not-nullness use the other not-null control. Return cast.
2602 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2603 bool never_see_null,
2604 bool safe_for_replace,
2605 bool speculative) {
2606 // Initial null check taken path
2607 (*null_control) = top();
2608 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2609
2610 // Generate uncommon_trap:
2611 if (never_see_null && (*null_control) != top()) {
2612 // If we see an unexpected null at a check-cast we record it and force a
2613 // recompile; the offending check-cast will be compiled to handle nulls.
2614 // If we see more than one offending BCI, then all checkcasts in the
2615 // method will be compiled to handle nulls.
2616 PreserveJVMState pjvms(this);
2617 set_control(*null_control);
2618 replace_in_map(value, null());
2619 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2620 uncommon_trap(reason,
2621 Deoptimization::Action_make_not_entrant);
2622 (*null_control) = top(); // null path is dead
2623 }
2624 if ((*null_control) == top() && safe_for_replace) {
2625 replace_in_map(value, cast);
2626 }
2627
2628 // Cast away null-ness on the result
2629 return cast;
2630 }
2631
2632 //------------------------------opt_iff----------------------------------------
2633 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2634 // Return slow-path control.
2635 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2636 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2637
2638 // Fast path taken; set region slot 2
2639 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2640 region->init_req(2,fast_taken); // Capture fast-control
2641
2642 // Fast path not-taken, i.e. slow path
2643 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2644 return slow_taken;
2645 }
2646
2647 //-----------------------------make_runtime_call-------------------------------
2648 Node* GraphKit::make_runtime_call(int flags,
2649 const TypeFunc* call_type, address call_addr,
2650 const char* call_name,
2651 const TypePtr* adr_type,
2652 // The following parms are all optional.
2653 // The first null ends the list.
2654 Node* parm0, Node* parm1,
2655 Node* parm2, Node* parm3,
2656 Node* parm4, Node* parm5,
2657 Node* parm6, Node* parm7) {
2658 assert(call_addr != nullptr, "must not call null targets");
2659
2660 // Slow-path call
2661 bool is_leaf = !(flags & RC_NO_LEAF);
2662 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2663 if (call_name == nullptr) {
2664 assert(!is_leaf, "must supply name for leaf");
2665 call_name = OptoRuntime::stub_name(call_addr);
2666 }
2667 CallNode* call;
2668 if (!is_leaf) {
2669 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2670 } else if (flags & RC_NO_FP) {
2671 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2672 } else if (flags & RC_VECTOR){
2673 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2674 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2675 } else {
2676 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2677 }
2678
2679 // The following is similar to set_edges_for_java_call,
2680 // except that the memory effects of the call are restricted to AliasIdxRaw.
2681
2682 // Slow path call has no side-effects, uses few values
2683 bool wide_in = !(flags & RC_NARROW_MEM);
2684 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2685
2686 Node* prev_mem = nullptr;
2687 if (wide_in) {
2688 prev_mem = set_predefined_input_for_runtime_call(call);
2689 } else {
2690 assert(!wide_out, "narrow in => narrow out");
2691 Node* narrow_mem = memory(adr_type);
2692 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2693 }
2694
2695 // Hook each parm in order. Stop looking at the first null.
2696 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2697 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2698 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2699 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2700 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2701 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2702 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2703 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2704 /* close each nested if ===> */ } } } } } } } }
2705 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2706
2707 if (!is_leaf) {
2708 // Non-leaves can block and take safepoints:
2709 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2710 }
2711 // Non-leaves can throw exceptions:
2712 if (has_io) {
2713 call->set_req(TypeFunc::I_O, i_o());
2714 }
2715
2716 if (flags & RC_UNCOMMON) {
2717 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2718 // (An "if" probability corresponds roughly to an unconditional count.
2719 // Sort of.)
2720 call->set_cnt(PROB_UNLIKELY_MAG(4));
2721 }
2722
2723 Node* c = _gvn.transform(call);
2724 assert(c == call, "cannot disappear");
2725
2726 if (wide_out) {
2727 // Slow path call has full side-effects.
2728 set_predefined_output_for_runtime_call(call);
2729 } else {
2730 // Slow path call has few side-effects, and/or sets few values.
2731 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2732 }
2733
2734 if (has_io) {
2735 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2736 }
2737 return call;
2738
2739 }
2740
2741 // i2b
2742 Node* GraphKit::sign_extend_byte(Node* in) {
2743 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2744 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2745 }
2746
2747 // i2s
2748 Node* GraphKit::sign_extend_short(Node* in) {
2749 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2750 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2751 }
2752
2753
2754 //------------------------------merge_memory-----------------------------------
2755 // Merge memory from one path into the current memory state.
2756 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2757 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2758 Node* old_slice = mms.force_memory();
2759 Node* new_slice = mms.memory2();
2760 if (old_slice != new_slice) {
2761 PhiNode* phi;
2762 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2763 if (mms.is_empty()) {
2764 // clone base memory Phi's inputs for this memory slice
2765 assert(old_slice == mms.base_memory(), "sanity");
2766 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2767 _gvn.set_type(phi, Type::MEMORY);
2768 for (uint i = 1; i < phi->req(); i++) {
2769 phi->init_req(i, old_slice->in(i));
2770 }
2771 } else {
2772 phi = old_slice->as_Phi(); // Phi was generated already
2773 }
2774 } else {
2775 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2776 _gvn.set_type(phi, Type::MEMORY);
2777 }
2778 phi->set_req(new_path, new_slice);
2779 mms.set_memory(phi);
2780 }
2781 }
2782 }
2783
2784 //------------------------------make_slow_call_ex------------------------------
2785 // Make the exception handler hookups for the slow call
2786 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2787 if (stopped()) return;
2788
2789 // Make a catch node with just two handlers: fall-through and catch-all
2790 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2791 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2792 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2793 _gvn.set_type_bottom(norm);
2794 C->record_for_igvn(norm);
2795 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2796
2797 { PreserveJVMState pjvms(this);
2798 set_control(excp);
2799 set_i_o(i_o);
2800
2801 if (excp != top()) {
2802 if (deoptimize) {
2803 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2804 uncommon_trap(Deoptimization::Reason_unhandled,
2805 Deoptimization::Action_none);
2806 } else {
2807 // Create an exception state also.
2808 // Use an exact type if the caller has a specific exception.
2809 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2810 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2811 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2812 }
2813 }
2814 }
2815
2816 // Get the no-exception control from the CatchNode.
2817 set_control(norm);
2818 }
2819
2820 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2821 Node* cmp = nullptr;
2822 switch(bt) {
2823 case T_INT: cmp = new CmpINode(in1, in2); break;
2824 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2825 default: fatal("unexpected comparison type %s", type2name(bt));
2826 }
2827 cmp = gvn.transform(cmp);
2828 Node* bol = gvn.transform(new BoolNode(cmp, test));
2829 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2830 gvn.transform(iff);
2831 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2832 return iff;
2833 }
2834
2835 //-------------------------------gen_subtype_check-----------------------------
2836 // Generate a subtyping check. Takes as input the subtype and supertype.
2837 // Returns 2 values: sets the default control() to the true path and returns
2838 // the false path. Only reads invariant memory; sets no (visible) memory.
2839 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2840 // but that's not exposed to the optimizer. This call also doesn't take in an
2841 // Object; if you wish to check an Object you need to load the Object's class
2842 // prior to coming here.
2843 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2844 ciMethod* method, int bci) {
2845 Compile* C = gvn.C;
2846 if ((*ctrl)->is_top()) {
2847 return C->top();
2848 }
2849
2850 // Fast check for identical types, perhaps identical constants.
2851 // The types can even be identical non-constants, in cases
2852 // involving Array.newInstance, Object.clone, etc.
2853 if (subklass == superklass)
2854 return C->top(); // false path is dead; no test needed.
2855
2856 if (gvn.type(superklass)->singleton()) {
2857 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2858 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2859
2860 // In the common case of an exact superklass, try to fold up the
2861 // test before generating code. You may ask, why not just generate
2862 // the code and then let it fold up? The answer is that the generated
2863 // code will necessarily include null checks, which do not always
2864 // completely fold away. If they are also needless, then they turn
2865 // into a performance loss. Example:
2866 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2867 // Here, the type of 'fa' is often exact, so the store check
2868 // of fa[1]=x will fold up, without testing the nullness of x.
2869 //
2870 // At macro expansion, we would have already folded the SubTypeCheckNode
2871 // being expanded here because we always perform the static sub type
2872 // check in SubTypeCheckNode::sub() regardless of whether
2873 // StressReflectiveCode is set or not. We can therefore skip this
2874 // static check when StressReflectiveCode is on.
2875 switch (C->static_subtype_check(superk, subk)) {
2876 case Compile::SSC_always_false:
2877 {
2878 Node* always_fail = *ctrl;
2879 *ctrl = gvn.C->top();
2880 return always_fail;
2881 }
2882 case Compile::SSC_always_true:
2883 return C->top();
2884 case Compile::SSC_easy_test:
2885 {
2886 // Just do a direct pointer compare and be done.
2887 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2888 *ctrl = gvn.transform(new IfTrueNode(iff));
2889 return gvn.transform(new IfFalseNode(iff));
2890 }
2891 case Compile::SSC_full_test:
2892 break;
2893 default:
2894 ShouldNotReachHere();
2895 }
2896 }
2897
2898 // %%% Possible further optimization: Even if the superklass is not exact,
2899 // if the subklass is the unique subtype of the superklass, the check
2900 // will always succeed. We could leave a dependency behind to ensure this.
2901
2902 // First load the super-klass's check-offset
2903 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2904 Node* m = C->immutable_memory();
2905 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2906 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2907 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2908 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2909 bool might_be_cache = (chk_off_con == cacheoff_con);
2910
2911 // Load from the sub-klass's super-class display list, or a 1-word cache of
2912 // the secondary superclass list, or a failing value with a sentinel offset
2913 // if the super-klass is an interface or exceptionally deep in the Java
2914 // hierarchy and we have to scan the secondary superclass list the hard way.
2915 // Worst-case type is a little odd: null is allowed as a result (usually
2916 // klass loads can never produce a null).
2917 Node *chk_off_X = chk_off;
2918 #ifdef _LP64
2919 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2920 #endif
2921 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2922 // For some types like interfaces the following loadKlass is from a 1-word
2923 // cache which is mutable so can't use immutable memory. Other
2924 // types load from the super-class display table which is immutable.
2925 Node *kmem = C->immutable_memory();
2926 // secondary_super_cache is not immutable but can be treated as such because:
2927 // - no ideal node writes to it in a way that could cause an
2928 // incorrect/missed optimization of the following Load.
2929 // - it's a cache so, worse case, not reading the latest value
2930 // wouldn't cause incorrect execution
2931 if (might_be_cache && mem != nullptr) {
2932 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2933 }
2934 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2935
2936 // Compile speed common case: ARE a subtype and we canNOT fail
2937 if (superklass == nkls) {
2938 return C->top(); // false path is dead; no test needed.
2939 }
2940
2941 // Gather the various success & failures here
2942 RegionNode* r_not_subtype = new RegionNode(3);
2943 gvn.record_for_igvn(r_not_subtype);
2944 RegionNode* r_ok_subtype = new RegionNode(4);
2945 gvn.record_for_igvn(r_ok_subtype);
2946
2947 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2948 // SubTypeCheck node
2949 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2950 ciCallProfile profile = method->call_profile_at_bci(bci);
2951 float total_prob = 0;
2952 for (int i = 0; profile.has_receiver(i); ++i) {
2953 float prob = profile.receiver_prob(i);
2954 total_prob += prob;
2955 }
2956 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2957 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2958 for (int i = 0; profile.has_receiver(i); ++i) {
2959 ciKlass* klass = profile.receiver(i);
2960 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2961 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2962 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2963 continue;
2964 }
2965 float prob = profile.receiver_prob(i);
2966 ConNode* klass_node = gvn.makecon(klass_t);
2967 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2968 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2969
2970 if (result == Compile::SSC_always_true) {
2971 r_ok_subtype->add_req(iftrue);
2972 } else {
2973 assert(result == Compile::SSC_always_false, "");
2974 r_not_subtype->add_req(iftrue);
2975 }
2976 *ctrl = gvn.transform(new IfFalseNode(iff));
2977 }
2978 }
2979 }
2980
2981 // See if we get an immediate positive hit. Happens roughly 83% of the
2982 // time. Test to see if the value loaded just previously from the subklass
2983 // is exactly the superklass.
2984 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2985 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2986 *ctrl = gvn.transform(new IfFalseNode(iff1));
2987
2988 // Compile speed common case: Check for being deterministic right now. If
2989 // chk_off is a constant and not equal to cacheoff then we are NOT a
2990 // subklass. In this case we need exactly the 1 test above and we can
2991 // return those results immediately.
2992 if (!might_be_cache) {
2993 Node* not_subtype_ctrl = *ctrl;
2994 *ctrl = iftrue1; // We need exactly the 1 test above
2995 PhaseIterGVN* igvn = gvn.is_IterGVN();
2996 if (igvn != nullptr) {
2997 igvn->remove_globally_dead_node(r_ok_subtype);
2998 igvn->remove_globally_dead_node(r_not_subtype);
2999 }
3000 return not_subtype_ctrl;
3001 }
3002
3003 r_ok_subtype->init_req(1, iftrue1);
3004
3005 // Check for immediate negative hit. Happens roughly 11% of the time (which
3006 // is roughly 63% of the remaining cases). Test to see if the loaded
3007 // check-offset points into the subklass display list or the 1-element
3008 // cache. If it points to the display (and NOT the cache) and the display
3009 // missed then it's not a subtype.
3010 Node *cacheoff = gvn.intcon(cacheoff_con);
3011 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3012 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3013 *ctrl = gvn.transform(new IfFalseNode(iff2));
3014
3015 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3016 // No performance impact (too rare) but allows sharing of secondary arrays
3017 // which has some footprint reduction.
3018 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3019 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3020 *ctrl = gvn.transform(new IfFalseNode(iff3));
3021
3022 // -- Roads not taken here: --
3023 // We could also have chosen to perform the self-check at the beginning
3024 // of this code sequence, as the assembler does. This would not pay off
3025 // the same way, since the optimizer, unlike the assembler, can perform
3026 // static type analysis to fold away many successful self-checks.
3027 // Non-foldable self checks work better here in second position, because
3028 // the initial primary superclass check subsumes a self-check for most
3029 // types. An exception would be a secondary type like array-of-interface,
3030 // which does not appear in its own primary supertype display.
3031 // Finally, we could have chosen to move the self-check into the
3032 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3033 // dependent manner. But it is worthwhile to have the check here,
3034 // where it can be perhaps be optimized. The cost in code space is
3035 // small (register compare, branch).
3036
3037 // Now do a linear scan of the secondary super-klass array. Again, no real
3038 // performance impact (too rare) but it's gotta be done.
3039 // Since the code is rarely used, there is no penalty for moving it
3040 // out of line, and it can only improve I-cache density.
3041 // The decision to inline or out-of-line this final check is platform
3042 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3043 Node* psc = gvn.transform(
3044 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3045
3046 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3047 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3048 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3049
3050 // Return false path; set default control to true path.
3051 *ctrl = gvn.transform(r_ok_subtype);
3052 return gvn.transform(r_not_subtype);
3053 }
3054
3055 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3056 const Type* sub_t = _gvn.type(obj_or_subklass);
3057 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3058 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3059 obj_or_subklass = makecon(sub_t);
3060 }
3061 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3062 if (expand_subtype_check) {
3063 MergeMemNode* mem = merged_memory();
3064 Node* ctrl = control();
3065 Node* subklass = obj_or_subklass;
3066 if (!sub_t->isa_klassptr()) {
3067 subklass = load_object_klass(obj_or_subklass);
3068 }
3069
3070 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3071 set_control(ctrl);
3072 return n;
3073 }
3074
3075 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3076 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3077 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3078 set_control(_gvn.transform(new IfTrueNode(iff)));
3079 return _gvn.transform(new IfFalseNode(iff));
3080 }
3081
3082 // Profile-driven exact type check:
3083 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3084 float prob, Node* *casted_receiver) {
3085 assert(!klass->is_interface(), "no exact type check on interfaces");
3086 Node* fail = top();
3087 const Type* rec_t = _gvn.type(receiver);
3088 if (rec_t->is_inlinetypeptr()) {
3089 if (klass->equals(rec_t->inline_klass())) {
3090 (*casted_receiver) = receiver; // Always passes
3091 } else {
3092 (*casted_receiver) = top(); // Always fails
3093 fail = control();
3094 set_control(top());
3095 }
3096 return fail;
3097 }
3098 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3099 Node* recv_klass = load_object_klass(receiver);
3100 fail = type_check(recv_klass, tklass, prob);
3101
3102 if (!stopped()) {
3103 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3104 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3105 assert(recv_xtype->klass_is_exact(), "");
3106
3107 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3108 // Subsume downstream occurrences of receiver with a cast to
3109 // recv_xtype, since now we know what the type will be.
3110 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3111 Node* res = _gvn.transform(cast);
3112 if (recv_xtype->is_inlinetypeptr()) {
3113 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3114 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3115 }
3116 (*casted_receiver) = res;
3117 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3118 // (User must make the replace_in_map call.)
3119 }
3120 }
3121
3122 return fail;
3123 }
3124
3125 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3126 float prob) {
3127 Node* want_klass = makecon(tklass);
3128 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3129 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3130 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3131 set_control(_gvn.transform(new IfTrueNode (iff)));
3132 Node* fail = _gvn.transform(new IfFalseNode(iff));
3133 return fail;
3134 }
3135
3136 //------------------------------subtype_check_receiver-------------------------
3137 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3138 Node** casted_receiver) {
3139 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3140 Node* want_klass = makecon(tklass);
3141
3142 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3143
3144 // Ignore interface type information until interface types are properly tracked.
3145 if (!stopped() && !klass->is_interface()) {
3146 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3147 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3148 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3149 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3150 if (recv_type->is_inlinetypeptr()) {
3151 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3152 }
3153 (*casted_receiver) = cast;
3154 }
3155 }
3156
3157 return slow_ctl;
3158 }
3159
3160 //------------------------------seems_never_null-------------------------------
3161 // Use null_seen information if it is available from the profile.
3162 // If we see an unexpected null at a type check we record it and force a
3163 // recompile; the offending check will be recompiled to handle nulls.
3164 // If we see several offending BCIs, then all checks in the
3165 // method will be recompiled.
3166 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3167 speculating = !_gvn.type(obj)->speculative_maybe_null();
3168 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3169 if (UncommonNullCast // Cutout for this technique
3170 && obj != null() // And not the -Xcomp stupid case?
3171 && !too_many_traps(reason)
3172 ) {
3173 if (speculating) {
3174 return true;
3175 }
3176 if (data == nullptr)
3177 // Edge case: no mature data. Be optimistic here.
3178 return true;
3179 // If the profile has not seen a null, assume it won't happen.
3180 assert(java_bc() == Bytecodes::_checkcast ||
3181 java_bc() == Bytecodes::_instanceof ||
3182 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3183 return !data->as_BitData()->null_seen();
3184 }
3185 speculating = false;
3186 return false;
3187 }
3188
3189 void GraphKit::guard_klass_being_initialized(Node* klass) {
3190 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3191 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3192 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3193 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3194 T_BYTE, MemNode::acquire);
3195 init_state = _gvn.transform(init_state);
3196
3197 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3198
3199 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3200 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3201
3202 { BuildCutout unless(this, tst, PROB_MAX);
3203 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3204 }
3205 }
3206
3207 void GraphKit::guard_init_thread(Node* klass) {
3208 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3209 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3210
3211 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3212 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3213 T_ADDRESS, MemNode::unordered);
3214 init_thread = _gvn.transform(init_thread);
3215
3216 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3217
3218 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3219 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3220
3221 { BuildCutout unless(this, tst, PROB_MAX);
3222 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3223 }
3224 }
3225
3226 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3227 if (ik->is_being_initialized()) {
3228 if (C->needs_clinit_barrier(ik, context)) {
3229 Node* klass = makecon(TypeKlassPtr::make(ik));
3230 guard_klass_being_initialized(klass);
3231 guard_init_thread(klass);
3232 insert_mem_bar(Op_MemBarCPUOrder);
3233 }
3234 } else if (ik->is_initialized()) {
3235 return; // no barrier needed
3236 } else {
3237 uncommon_trap(Deoptimization::Reason_uninitialized,
3238 Deoptimization::Action_reinterpret,
3239 nullptr);
3240 }
3241 }
3242
3243 //------------------------maybe_cast_profiled_receiver-------------------------
3244 // If the profile has seen exactly one type, narrow to exactly that type.
3245 // Subsequent type checks will always fold up.
3246 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3247 const TypeKlassPtr* require_klass,
3248 ciKlass* spec_klass,
3249 bool safe_for_replace) {
3250 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3251
3252 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3253
3254 // Make sure we haven't already deoptimized from this tactic.
3255 if (too_many_traps_or_recompiles(reason))
3256 return nullptr;
3257
3258 // (No, this isn't a call, but it's enough like a virtual call
3259 // to use the same ciMethod accessor to get the profile info...)
3260 // If we have a speculative type use it instead of profiling (which
3261 // may not help us)
3262 ciKlass* exact_kls = spec_klass;
3263 if (exact_kls == nullptr) {
3264 if (java_bc() == Bytecodes::_aastore) {
3265 ciKlass* array_type = nullptr;
3266 ciKlass* element_type = nullptr;
3267 ProfilePtrKind element_ptr = ProfileMaybeNull;
3268 bool flat_array = true;
3269 bool null_free_array = true;
3270 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3271 exact_kls = element_type;
3272 } else {
3273 exact_kls = profile_has_unique_klass();
3274 }
3275 }
3276 if (exact_kls != nullptr) {// no cast failures here
3277 if (require_klass == nullptr ||
3278 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3279 // If we narrow the type to match what the type profile sees or
3280 // the speculative type, we can then remove the rest of the
3281 // cast.
3282 // This is a win, even if the exact_kls is very specific,
3283 // because downstream operations, such as method calls,
3284 // will often benefit from the sharper type.
3285 Node* exact_obj = not_null_obj; // will get updated in place...
3286 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3287 &exact_obj);
3288 { PreserveJVMState pjvms(this);
3289 set_control(slow_ctl);
3290 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3291 }
3292 if (safe_for_replace) {
3293 replace_in_map(not_null_obj, exact_obj);
3294 }
3295 return exact_obj;
3296 }
3297 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3298 }
3299
3300 return nullptr;
3301 }
3302
3303 /**
3304 * Cast obj to type and emit guard unless we had too many traps here
3305 * already
3306 *
3307 * @param obj node being casted
3308 * @param type type to cast the node to
3309 * @param not_null true if we know node cannot be null
3310 */
3311 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3312 ciKlass* type,
3313 bool not_null) {
3314 if (stopped()) {
3315 return obj;
3316 }
3317
3318 // type is null if profiling tells us this object is always null
3319 if (type != nullptr) {
3320 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3321 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3322
3323 if (!too_many_traps_or_recompiles(null_reason) &&
3324 !too_many_traps_or_recompiles(class_reason)) {
3325 Node* not_null_obj = nullptr;
3326 // not_null is true if we know the object is not null and
3327 // there's no need for a null check
3328 if (!not_null) {
3329 Node* null_ctl = top();
3330 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3331 assert(null_ctl->is_top(), "no null control here");
3332 } else {
3333 not_null_obj = obj;
3334 }
3335
3336 Node* exact_obj = not_null_obj;
3337 ciKlass* exact_kls = type;
3338 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3339 &exact_obj);
3340 {
3341 PreserveJVMState pjvms(this);
3342 set_control(slow_ctl);
3343 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3344 }
3345 replace_in_map(not_null_obj, exact_obj);
3346 obj = exact_obj;
3347 }
3348 } else {
3349 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3350 Node* exact_obj = null_assert(obj);
3351 replace_in_map(obj, exact_obj);
3352 obj = exact_obj;
3353 }
3354 }
3355 return obj;
3356 }
3357
3358 //-------------------------------gen_instanceof--------------------------------
3359 // Generate an instance-of idiom. Used by both the instance-of bytecode
3360 // and the reflective instance-of call.
3361 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3362 kill_dead_locals(); // Benefit all the uncommon traps
3363 assert( !stopped(), "dead parse path should be checked in callers" );
3364 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3365 "must check for not-null not-dead klass in callers");
3366
3367 // Make the merge point
3368 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3369 RegionNode* region = new RegionNode(PATH_LIMIT);
3370 Node* phi = new PhiNode(region, TypeInt::BOOL);
3371 C->set_has_split_ifs(true); // Has chance for split-if optimization
3372
3373 ciProfileData* data = nullptr;
3374 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3375 data = method()->method_data()->bci_to_data(bci());
3376 }
3377 bool speculative_not_null = false;
3378 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3379 && seems_never_null(obj, data, speculative_not_null));
3380
3381 // Null check; get casted pointer; set region slot 3
3382 Node* null_ctl = top();
3383 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3384
3385 // If not_null_obj is dead, only null-path is taken
3386 if (stopped()) { // Doing instance-of on a null?
3387 set_control(null_ctl);
3388 return intcon(0);
3389 }
3390 region->init_req(_null_path, null_ctl);
3391 phi ->init_req(_null_path, intcon(0)); // Set null path value
3392 if (null_ctl == top()) {
3393 // Do this eagerly, so that pattern matches like is_diamond_phi
3394 // will work even during parsing.
3395 assert(_null_path == PATH_LIMIT-1, "delete last");
3396 region->del_req(_null_path);
3397 phi ->del_req(_null_path);
3398 }
3399
3400 // Do we know the type check always succeed?
3401 bool known_statically = false;
3402 if (_gvn.type(superklass)->singleton()) {
3403 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3404 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3405 if (subk != nullptr && subk->is_loaded()) {
3406 int static_res = C->static_subtype_check(superk, subk);
3407 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3408 }
3409 }
3410
3411 if (!known_statically) {
3412 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3413 // We may not have profiling here or it may not help us. If we
3414 // have a speculative type use it to perform an exact cast.
3415 ciKlass* spec_obj_type = obj_type->speculative_type();
3416 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3417 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3418 if (stopped()) { // Profile disagrees with this path.
3419 set_control(null_ctl); // Null is the only remaining possibility.
3420 return intcon(0);
3421 }
3422 if (cast_obj != nullptr) {
3423 not_null_obj = cast_obj;
3424 }
3425 }
3426 }
3427
3428 // Generate the subtype check
3429 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3430
3431 // Plug in the success path to the general merge in slot 1.
3432 region->init_req(_obj_path, control());
3433 phi ->init_req(_obj_path, intcon(1));
3434
3435 // Plug in the failing path to the general merge in slot 2.
3436 region->init_req(_fail_path, not_subtype_ctrl);
3437 phi ->init_req(_fail_path, intcon(0));
3438
3439 // Return final merged results
3440 set_control( _gvn.transform(region) );
3441 record_for_igvn(region);
3442
3443 // If we know the type check always succeeds then we don't use the
3444 // profiling data at this bytecode. Don't lose it, feed it to the
3445 // type system as a speculative type.
3446 if (safe_for_replace) {
3447 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3448 replace_in_map(obj, casted_obj);
3449 }
3450
3451 return _gvn.transform(phi);
3452 }
3453
3454 //-------------------------------gen_checkcast---------------------------------
3455 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3456 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3457 // uncommon-trap paths work. Adjust stack after this call.
3458 // If failure_control is supplied and not null, it is filled in with
3459 // the control edge for the cast failure. Otherwise, an appropriate
3460 // uncommon trap or exception is thrown.
3461 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free) {
3462 kill_dead_locals(); // Benefit all the uncommon traps
3463 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3464 const Type* obj_type = _gvn.type(obj);
3465 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)) {
3466 // Special case: larval inline objects must not be scalarized. They are also generally not
3467 // allowed to participate in most operations except as the first operand of putfield, or as an
3468 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3469 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3470 // value objects in all other places. This special case comes from the limitation of the Java
3471 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3472 // value type. We must do this first because C->static_subtype_check may do nothing when
3473 // StressReflectiveCode is set.
3474 return obj;
3475 }
3476
3477 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3478 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3479 bool safe_for_replace = (failure_control == nullptr);
3480 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3481
3482 // Fast cutout: Check the case that the cast is vacuously true.
3483 // This detects the common cases where the test will short-circuit
3484 // away completely. We do this before we perform the null check,
3485 // because if the test is going to turn into zero code, we don't
3486 // want a residual null check left around. (Causes a slowdown,
3487 // for example, in some objArray manipulations, such as a[i]=a[j].)
3488 if (improved_klass_ptr_type->singleton()) {
3489 const TypeKlassPtr* kptr = nullptr;
3490 if (obj_type->isa_oop_ptr()) {
3491 kptr = obj_type->is_oopptr()->as_klass_type();
3492 } else if (obj->is_InlineType()) {
3493 ciInlineKlass* vk = obj_type->inline_klass();
3494 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3495 }
3496
3497 if (kptr != nullptr) {
3498 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3499 case Compile::SSC_always_true:
3500 // If we know the type check always succeed then we don't use
3501 // the profiling data at this bytecode. Don't lose it, feed it
3502 // to the type system as a speculative type.
3503 obj = record_profiled_receiver_for_speculation(obj);
3504 if (null_free) {
3505 assert(safe_for_replace, "must be");
3506 obj = null_check(obj);
3507 }
3508 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3509 return obj;
3510 case Compile::SSC_always_false:
3511 if (null_free) {
3512 assert(safe_for_replace, "must be");
3513 obj = null_check(obj);
3514 }
3515 // It needs a null check because a null will *pass* the cast check.
3516 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3517 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3518 Deoptimization::DeoptReason reason = is_aastore ?
3519 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3520 builtin_throw(reason);
3521 return top();
3522 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3523 return null_assert(obj);
3524 }
3525 break; // Fall through to full check
3526 default:
3527 break;
3528 }
3529 }
3530 }
3531
3532 ciProfileData* data = nullptr;
3533 if (failure_control == nullptr) { // use MDO in regular case only
3534 assert(java_bc() == Bytecodes::_aastore ||
3535 java_bc() == Bytecodes::_checkcast,
3536 "interpreter profiles type checks only for these BCs");
3537 if (method()->method_data()->is_mature()) {
3538 data = method()->method_data()->bci_to_data(bci());
3539 }
3540 }
3541
3542 // Make the merge point
3543 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3544 RegionNode* region = new RegionNode(PATH_LIMIT);
3545 Node* phi = new PhiNode(region, toop);
3546 _gvn.set_type(region, Type::CONTROL);
3547 _gvn.set_type(phi, toop);
3548
3549 C->set_has_split_ifs(true); // Has chance for split-if optimization
3550
3551 // Use null-cast information if it is available
3552 bool speculative_not_null = false;
3553 bool never_see_null = ((failure_control == nullptr) // regular case only
3554 && seems_never_null(obj, data, speculative_not_null));
3555
3556 if (obj->is_InlineType()) {
3557 // Re-execute if buffering during triggers deoptimization
3558 PreserveReexecuteState preexecs(this);
3559 jvms()->set_should_reexecute(true);
3560 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3561 }
3562
3563 // Null check; get casted pointer; set region slot 3
3564 Node* null_ctl = top();
3565 Node* not_null_obj = nullptr;
3566 if (null_free) {
3567 assert(safe_for_replace, "must be");
3568 not_null_obj = null_check(obj);
3569 } else {
3570 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3571 }
3572
3573 // If not_null_obj is dead, only null-path is taken
3574 if (stopped()) { // Doing instance-of on a null?
3575 set_control(null_ctl);
3576 if (toop->is_inlinetypeptr()) {
3577 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3578 }
3579 return null();
3580 }
3581 region->init_req(_null_path, null_ctl);
3582 phi ->init_req(_null_path, null()); // Set null path value
3583 if (null_ctl == top()) {
3584 // Do this eagerly, so that pattern matches like is_diamond_phi
3585 // will work even during parsing.
3586 assert(_null_path == PATH_LIMIT-1, "delete last");
3587 region->del_req(_null_path);
3588 phi ->del_req(_null_path);
3589 }
3590
3591 Node* cast_obj = nullptr;
3592 if (improved_klass_ptr_type->klass_is_exact()) {
3593 // The following optimization tries to statically cast the speculative type of the object
3594 // (for example obtained during profiling) to the type of the superklass and then do a
3595 // dynamic check that the type of the object is what we expect. To work correctly
3596 // for checkcast and aastore the type of superklass should be exact.
3597 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3598 // We may not have profiling here or it may not help us. If we have
3599 // a speculative type use it to perform an exact cast.
3600 ciKlass* spec_obj_type = obj_type->speculative_type();
3601 if (spec_obj_type != nullptr || data != nullptr) {
3602 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3603 if (cast_obj != nullptr) {
3604 if (failure_control != nullptr) // failure is now impossible
3605 (*failure_control) = top();
3606 // adjust the type of the phi to the exact klass:
3607 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3608 }
3609 }
3610 }
3611
3612 if (cast_obj == nullptr) {
3613 // Generate the subtype check
3614 Node* improved_superklass = superklass;
3615 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3616 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3617 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3618 // Additionally, the benefit would only be minor in non-constant cases.
3619 improved_superklass = makecon(improved_klass_ptr_type);
3620 }
3621 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3622 // Plug in success path into the merge
3623 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3624 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3625 if (failure_control == nullptr) {
3626 if (not_subtype_ctrl != top()) { // If failure is possible
3627 PreserveJVMState pjvms(this);
3628 set_control(not_subtype_ctrl);
3629 Node* obj_klass = nullptr;
3630 if (not_null_obj->is_InlineType()) {
3631 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3632 } else {
3633 obj_klass = load_object_klass(not_null_obj);
3634 }
3635 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3636 Deoptimization::DeoptReason reason = is_aastore ?
3637 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3638 builtin_throw(reason);
3639 }
3640 } else {
3641 (*failure_control) = not_subtype_ctrl;
3642 }
3643 }
3644
3645 region->init_req(_obj_path, control());
3646 phi ->init_req(_obj_path, cast_obj);
3647
3648 // A merge of null or Casted-NotNull obj
3649 Node* res = _gvn.transform(phi);
3650
3651 // Note I do NOT always 'replace_in_map(obj,result)' here.
3652 // if( tk->klass()->can_be_primary_super() )
3653 // This means that if I successfully store an Object into an array-of-String
3654 // I 'forget' that the Object is really now known to be a String. I have to
3655 // do this because we don't have true union types for interfaces - if I store
3656 // a Baz into an array-of-Interface and then tell the optimizer it's an
3657 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3658 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3659 // replace_in_map( obj, res );
3660
3661 // Return final merged results
3662 set_control( _gvn.transform(region) );
3663 record_for_igvn(region);
3664
3665 bool not_inline = !toop->can_be_inline_type();
3666 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3667 if (EnableValhalla && (not_inline || not_flat_in_array)) {
3668 // Check if obj has been loaded from an array
3669 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3670 Node* array = nullptr;
3671 if (obj->isa_Load()) {
3672 Node* address = obj->in(MemNode::Address);
3673 if (address->isa_AddP()) {
3674 array = address->as_AddP()->in(AddPNode::Base);
3675 }
3676 } else if (obj->is_Phi()) {
3677 Node* region = obj->in(0);
3678 // TODO make this more robust (see JDK-8231346)
3679 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3680 IfNode* iff = region->in(2)->in(0)->isa_If();
3681 if (iff != nullptr) {
3682 iff->is_flat_array_check(&_gvn, &array);
3683 }
3684 }
3685 }
3686 if (array != nullptr) {
3687 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3688 if (ary_t != nullptr) {
3689 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3690 // Casting array element to a non-inline-type, mark array as not null-free.
3691 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3692 replace_in_map(array, cast);
3693 array = cast;
3694 }
3695 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3696 // Casting array element to a non-flat-in-array type, mark array as not flat.
3697 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3698 replace_in_map(array, cast);
3699 array = cast;
3700 }
3701 }
3702 }
3703 }
3704
3705 if (!stopped() && !res->is_InlineType()) {
3706 res = record_profiled_receiver_for_speculation(res);
3707 if (toop->is_inlinetypeptr()) {
3708 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3709 res = vt;
3710 if (safe_for_replace) {
3711 replace_in_map(obj, vt);
3712 replace_in_map(not_null_obj, vt);
3713 replace_in_map(res, vt);
3714 }
3715 }
3716 }
3717 return res;
3718 }
3719
3720 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3721 // Load markword
3722 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3723 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3724 if (check_lock) {
3725 // Check if obj is locked
3726 Node* locked_bit = MakeConX(markWord::unlocked_value);
3727 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3728 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3729 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3730 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3731 _gvn.transform(iff);
3732 Node* locked_region = new RegionNode(3);
3733 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3734
3735 // Unlocked: Use bits from mark word
3736 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3737 mark_phi->init_req(1, mark);
3738
3739 // Locked: Load prototype header from klass
3740 set_control(_gvn.transform(new IfFalseNode(iff)));
3741 // Make loads control dependent to make sure they are only executed if array is locked
3742 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3743 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3744 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3745 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));
3746
3747 locked_region->init_req(2, control());
3748 mark_phi->init_req(2, proto);
3749 set_control(_gvn.transform(locked_region));
3750 record_for_igvn(locked_region);
3751
3752 mark = mark_phi;
3753 }
3754
3755 // Now check if mark word bits are set
3756 Node* mask = MakeConX(mask_val);
3757 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3758 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3759 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3760 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3761 }
3762
3763 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3764 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3765 }
3766
3767 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3768 // We can't use immutable memory here because the mark word is mutable.
3769 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3770 // check is moved out of loops (mainly to enable loop unswitching).
3771 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3772 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3773 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3774 }
3775
3776 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3777 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3778 }
3779
3780 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3781 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3782
3783 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3784 if (!vk->has_non_atomic_layout()) {
3785 return intcon(1); // Always atomic
3786 } else if (!vk->has_atomic_layout()) {
3787 return intcon(0); // Never atomic
3788 }
3789
3790 Node* array_klass = load_object_klass(array);
3791 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3792 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3793 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3794 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT)));
3795 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3796 }
3797
3798 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3799 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3800 RegionNode* region = new RegionNode(3);
3801 Node* null_ctl = top();
3802 null_check_oop(val, &null_ctl);
3803 if (null_ctl != top()) {
3804 PreserveJVMState pjvms(this);
3805 set_control(null_ctl);
3806 {
3807 // Deoptimize if null-free array
3808 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3809 inc_sp(nargs);
3810 uncommon_trap(Deoptimization::Reason_null_check,
3811 Deoptimization::Action_none);
3812 }
3813 region->init_req(1, control());
3814 }
3815 region->init_req(2, control());
3816 set_control(_gvn.transform(region));
3817 record_for_igvn(region);
3818 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3819 // Since we were just successfully storing null, the array can't be null free.
3820 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3821 ary_t = ary_t->cast_to_not_null_free();
3822 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3823 if (safe_for_replace) {
3824 replace_in_map(ary, cast);
3825 }
3826 ary = cast;
3827 }
3828 return ary;
3829 }
3830
3831 //------------------------------next_monitor-----------------------------------
3832 // What number should be given to the next monitor?
3833 int GraphKit::next_monitor() {
3834 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3835 int next = current + C->sync_stack_slots();
3836 // Keep the toplevel high water mark current:
3837 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3838 return current;
3839 }
3840
3841 //------------------------------insert_mem_bar---------------------------------
3842 // Memory barrier to avoid floating things around
3843 // The membar serves as a pinch point between both control and all memory slices.
3844 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3845 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3846 mb->init_req(TypeFunc::Control, control());
3847 mb->init_req(TypeFunc::Memory, reset_memory());
3848 Node* membar = _gvn.transform(mb);
3849 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3850 set_all_memory_call(membar);
3851 return membar;
3852 }
3853
3854 //-------------------------insert_mem_bar_volatile----------------------------
3855 // Memory barrier to avoid floating things around
3856 // The membar serves as a pinch point between both control and memory(alias_idx).
3857 // If you want to make a pinch point on all memory slices, do not use this
3858 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3859 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3860 // When Parse::do_put_xxx updates a volatile field, it appends a series
3861 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3862 // The first membar is on the same memory slice as the field store opcode.
3863 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3864 // All the other membars (for other volatile slices, including AliasIdxBot,
3865 // which stands for all unknown volatile slices) are control-dependent
3866 // on the first membar. This prevents later volatile loads or stores
3867 // from sliding up past the just-emitted store.
3868
3869 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3870 mb->set_req(TypeFunc::Control,control());
3871 if (alias_idx == Compile::AliasIdxBot) {
3872 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3873 } else {
3874 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3875 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3876 }
3877 Node* membar = _gvn.transform(mb);
3878 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3879 if (alias_idx == Compile::AliasIdxBot) {
3880 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3881 } else {
3882 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3883 }
3884 return membar;
3885 }
3886
3887 //------------------------------shared_lock------------------------------------
3888 // Emit locking code.
3889 FastLockNode* GraphKit::shared_lock(Node* obj) {
3890 // bci is either a monitorenter bc or InvocationEntryBci
3891 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3892 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3893
3894 if( !GenerateSynchronizationCode )
3895 return nullptr; // Not locking things?
3896
3897 if (stopped()) // Dead monitor?
3898 return nullptr;
3899
3900 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3901
3902 // Box the stack location
3903 Node* box = new BoxLockNode(next_monitor());
3904 // Check for bailout after new BoxLockNode
3905 if (failing()) { return nullptr; }
3906 box = _gvn.transform(box);
3907 Node* mem = reset_memory();
3908
3909 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3910
3911 // Add monitor to debug info for the slow path. If we block inside the
3912 // slow path and de-opt, we need the monitor hanging around
3913 map()->push_monitor( flock );
3914
3915 const TypeFunc *tf = LockNode::lock_type();
3916 LockNode *lock = new LockNode(C, tf);
3917
3918 lock->init_req( TypeFunc::Control, control() );
3919 lock->init_req( TypeFunc::Memory , mem );
3920 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3921 lock->init_req( TypeFunc::FramePtr, frameptr() );
3922 lock->init_req( TypeFunc::ReturnAdr, top() );
3923
3924 lock->init_req(TypeFunc::Parms + 0, obj);
3925 lock->init_req(TypeFunc::Parms + 1, box);
3926 lock->init_req(TypeFunc::Parms + 2, flock);
3927 add_safepoint_edges(lock);
3928
3929 lock = _gvn.transform( lock )->as_Lock();
3930
3931 // lock has no side-effects, sets few values
3932 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3933
3934 insert_mem_bar(Op_MemBarAcquireLock);
3935
3936 // Add this to the worklist so that the lock can be eliminated
3937 record_for_igvn(lock);
3938
3939 #ifndef PRODUCT
3940 if (PrintLockStatistics) {
3941 // Update the counter for this lock. Don't bother using an atomic
3942 // operation since we don't require absolute accuracy.
3943 lock->create_lock_counter(map()->jvms());
3944 increment_counter(lock->counter()->addr());
3945 }
3946 #endif
3947
3948 return flock;
3949 }
3950
3951
3952 //------------------------------shared_unlock----------------------------------
3953 // Emit unlocking code.
3954 void GraphKit::shared_unlock(Node* box, Node* obj) {
3955 // bci is either a monitorenter bc or InvocationEntryBci
3956 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3957 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3958
3959 if( !GenerateSynchronizationCode )
3960 return;
3961 if (stopped()) { // Dead monitor?
3962 map()->pop_monitor(); // Kill monitor from debug info
3963 return;
3964 }
3965 assert(!obj->is_InlineType(), "should not unlock on inline type");
3966
3967 // Memory barrier to avoid floating things down past the locked region
3968 insert_mem_bar(Op_MemBarReleaseLock);
3969
3970 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3971 UnlockNode *unlock = new UnlockNode(C, tf);
3972 #ifdef ASSERT
3973 unlock->set_dbg_jvms(sync_jvms());
3974 #endif
3975 uint raw_idx = Compile::AliasIdxRaw;
3976 unlock->init_req( TypeFunc::Control, control() );
3977 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3978 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3979 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3980 unlock->init_req( TypeFunc::ReturnAdr, top() );
3981
3982 unlock->init_req(TypeFunc::Parms + 0, obj);
3983 unlock->init_req(TypeFunc::Parms + 1, box);
3984 unlock = _gvn.transform(unlock)->as_Unlock();
3985
3986 Node* mem = reset_memory();
3987
3988 // unlock has no side-effects, sets few values
3989 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3990
3991 // Kill monitor from debug info
3992 map()->pop_monitor( );
3993 }
3994
3995 //-------------------------------get_layout_helper-----------------------------
3996 // If the given klass is a constant or known to be an array,
3997 // fetch the constant layout helper value into constant_value
3998 // and return null. Otherwise, load the non-constant
3999 // layout helper value, and return the node which represents it.
4000 // This two-faced routine is useful because allocation sites
4001 // almost always feature constant types.
4002 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4003 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4004 if (!StressReflectiveCode && klass_t != nullptr) {
4005 bool xklass = klass_t->klass_is_exact();
4006 bool can_be_flat = false;
4007 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4008 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4009 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4010 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4011 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4012 }
4013 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4014 jint lhelper;
4015 if (klass_t->is_flat()) {
4016 lhelper = ary_type->flat_layout_helper();
4017 } else if (klass_t->isa_aryklassptr()) {
4018 BasicType elem = ary_type->elem()->array_element_basic_type();
4019 if (is_reference_type(elem, true)) {
4020 elem = T_OBJECT;
4021 }
4022 lhelper = Klass::array_layout_helper(elem);
4023 } else {
4024 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4025 }
4026 if (lhelper != Klass::_lh_neutral_value) {
4027 constant_value = lhelper;
4028 return (Node*) nullptr;
4029 }
4030 }
4031 }
4032 constant_value = Klass::_lh_neutral_value; // put in a known value
4033 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4034 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4035 }
4036
4037 // We just put in an allocate/initialize with a big raw-memory effect.
4038 // Hook selected additional alias categories on the initialization.
4039 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4040 MergeMemNode* init_in_merge,
4041 Node* init_out_raw) {
4042 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4043 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4044
4045 Node* prevmem = kit.memory(alias_idx);
4046 init_in_merge->set_memory_at(alias_idx, prevmem);
4047 if (init_out_raw != nullptr) {
4048 kit.set_memory(init_out_raw, alias_idx);
4049 }
4050 }
4051
4052 //---------------------------set_output_for_allocation-------------------------
4053 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4054 const TypeOopPtr* oop_type,
4055 bool deoptimize_on_exception) {
4056 int rawidx = Compile::AliasIdxRaw;
4057 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4058 add_safepoint_edges(alloc);
4059 Node* allocx = _gvn.transform(alloc);
4060 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4061 // create memory projection for i_o
4062 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4063 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4064
4065 // create a memory projection as for the normal control path
4066 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4067 set_memory(malloc, rawidx);
4068
4069 // a normal slow-call doesn't change i_o, but an allocation does
4070 // we create a separate i_o projection for the normal control path
4071 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4072 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4073
4074 // put in an initialization barrier
4075 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4076 rawoop)->as_Initialize();
4077 assert(alloc->initialization() == init, "2-way macro link must work");
4078 assert(init ->allocation() == alloc, "2-way macro link must work");
4079 {
4080 // Extract memory strands which may participate in the new object's
4081 // initialization, and source them from the new InitializeNode.
4082 // This will allow us to observe initializations when they occur,
4083 // and link them properly (as a group) to the InitializeNode.
4084 assert(init->in(InitializeNode::Memory) == malloc, "");
4085 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4086 init->set_req(InitializeNode::Memory, minit_in);
4087 record_for_igvn(minit_in); // fold it up later, if possible
4088 _gvn.set_type(minit_in, Type::MEMORY);
4089 Node* minit_out = memory(rawidx);
4090 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4091 // Add an edge in the MergeMem for the header fields so an access
4092 // to one of those has correct memory state
4093 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4094 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4095 if (oop_type->isa_aryptr()) {
4096 const TypeAryPtr* arytype = oop_type->is_aryptr();
4097 if (arytype->is_flat()) {
4098 // Initially all flat array accesses share a single slice
4099 // but that changes after parsing. Prepare the memory graph so
4100 // it can optimize flat array accesses properly once they
4101 // don't share a single slice.
4102 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4103 C->set_flat_accesses_share_alias(false);
4104 ciInlineKlass* vk = arytype->elem()->inline_klass();
4105 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4106 ciField* field = vk->nonstatic_field_at(i);
4107 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4108 continue; // do not bother to track really large numbers of fields
4109 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
4110 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4111 int fieldidx = C->get_alias_index(adr_type, true);
4112 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4113 // can result in per flat array field Phis to be created which confuses the logic of
4114 // Compile::adjust_flat_array_access_aliases().
4115 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4116 }
4117 C->set_flat_accesses_share_alias(true);
4118 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4119 } else {
4120 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4121 int elemidx = C->get_alias_index(telemref);
4122 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4123 }
4124 } else if (oop_type->isa_instptr()) {
4125 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4126 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4127 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4128 ciField* field = ik->nonstatic_field_at(i);
4129 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4130 continue; // do not bother to track really large numbers of fields
4131 // Find (or create) the alias category for this field:
4132 int fieldidx = C->alias_type(field)->index();
4133 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4134 }
4135 }
4136 }
4137
4138 // Cast raw oop to the real thing...
4139 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4140 javaoop = _gvn.transform(javaoop);
4141 C->set_recent_alloc(control(), javaoop);
4142 assert(just_allocated_object(control()) == javaoop, "just allocated");
4143
4144 #ifdef ASSERT
4145 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4146 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4147 "Ideal_allocation works");
4148 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4149 "Ideal_allocation works");
4150 if (alloc->is_AllocateArray()) {
4151 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4152 "Ideal_allocation works");
4153 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4154 "Ideal_allocation works");
4155 } else {
4156 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4157 }
4158 }
4159 #endif //ASSERT
4160
4161 return javaoop;
4162 }
4163
4164 //---------------------------new_instance--------------------------------------
4165 // This routine takes a klass_node which may be constant (for a static type)
4166 // or may be non-constant (for reflective code). It will work equally well
4167 // for either, and the graph will fold nicely if the optimizer later reduces
4168 // the type to a constant.
4169 // The optional arguments are for specialized use by intrinsics:
4170 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4171 // - If 'return_size_val', report the total object size to the caller.
4172 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4173 Node* GraphKit::new_instance(Node* klass_node,
4174 Node* extra_slow_test,
4175 Node* *return_size_val,
4176 bool deoptimize_on_exception,
4177 InlineTypeNode* inline_type_node) {
4178 // Compute size in doublewords
4179 // The size is always an integral number of doublewords, represented
4180 // as a positive bytewise size stored in the klass's layout_helper.
4181 // The layout_helper also encodes (in a low bit) the need for a slow path.
4182 jint layout_con = Klass::_lh_neutral_value;
4183 Node* layout_val = get_layout_helper(klass_node, layout_con);
4184 bool layout_is_con = (layout_val == nullptr);
4185
4186 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4187 // Generate the initial go-slow test. It's either ALWAYS (return a
4188 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4189 // case) a computed value derived from the layout_helper.
4190 Node* initial_slow_test = nullptr;
4191 if (layout_is_con) {
4192 assert(!StressReflectiveCode, "stress mode does not use these paths");
4193 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4194 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4195 } else { // reflective case
4196 // This reflective path is used by Unsafe.allocateInstance.
4197 // (It may be stress-tested by specifying StressReflectiveCode.)
4198 // Basically, we want to get into the VM is there's an illegal argument.
4199 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4200 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4201 if (extra_slow_test != intcon(0)) {
4202 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4203 }
4204 // (Macro-expander will further convert this to a Bool, if necessary.)
4205 }
4206
4207 // Find the size in bytes. This is easy; it's the layout_helper.
4208 // The size value must be valid even if the slow path is taken.
4209 Node* size = nullptr;
4210 if (layout_is_con) {
4211 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4212 } else { // reflective case
4213 // This reflective path is used by clone and Unsafe.allocateInstance.
4214 size = ConvI2X(layout_val);
4215
4216 // Clear the low bits to extract layout_helper_size_in_bytes:
4217 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4218 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4219 size = _gvn.transform( new AndXNode(size, mask) );
4220 }
4221 if (return_size_val != nullptr) {
4222 (*return_size_val) = size;
4223 }
4224
4225 // This is a precise notnull oop of the klass.
4226 // (Actually, it need not be precise if this is a reflective allocation.)
4227 // It's what we cast the result to.
4228 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4229 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4230 const TypeOopPtr* oop_type = tklass->as_instance_type();
4231
4232 // Now generate allocation code
4233
4234 // The entire memory state is needed for slow path of the allocation
4235 // since GC and deoptimization can happen.
4236 Node *mem = reset_memory();
4237 set_all_memory(mem); // Create new memory state
4238
4239 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4240 control(), mem, i_o(),
4241 size, klass_node,
4242 initial_slow_test, inline_type_node);
4243
4244 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4245 }
4246
4247 //-------------------------------new_array-------------------------------------
4248 // helper for newarray and anewarray
4249 // The 'length' parameter is (obviously) the length of the array.
4250 // The optional arguments are for specialized use by intrinsics:
4251 // - If 'return_size_val', report the non-padded array size (sum of header size
4252 // and array body) to the caller.
4253 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4254 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4255 Node* length, // number of array elements
4256 int nargs, // number of arguments to push back for uncommon trap
4257 Node* *return_size_val,
4258 bool deoptimize_on_exception,
4259 Node* init_val) {
4260 jint layout_con = Klass::_lh_neutral_value;
4261 Node* layout_val = get_layout_helper(klass_node, layout_con);
4262 bool layout_is_con = (layout_val == nullptr);
4263
4264 if (!layout_is_con && !StressReflectiveCode &&
4265 !too_many_traps(Deoptimization::Reason_class_check)) {
4266 // This is a reflective array creation site.
4267 // Optimistically assume that it is a subtype of Object[],
4268 // so that we can fold up all the address arithmetic.
4269 layout_con = Klass::array_layout_helper(T_OBJECT);
4270 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4271 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4272 { BuildCutout unless(this, bol_lh, PROB_MAX);
4273 inc_sp(nargs);
4274 uncommon_trap(Deoptimization::Reason_class_check,
4275 Deoptimization::Action_maybe_recompile);
4276 }
4277 layout_val = nullptr;
4278 layout_is_con = true;
4279 }
4280
4281 // Generate the initial go-slow test. Make sure we do not overflow
4282 // if length is huge (near 2Gig) or negative! We do not need
4283 // exact double-words here, just a close approximation of needed
4284 // double-words. We can't add any offset or rounding bits, lest we
4285 // take a size -1 of bytes and make it positive. Use an unsigned
4286 // compare, so negative sizes look hugely positive.
4287 int fast_size_limit = FastAllocateSizeLimit;
4288 if (layout_is_con) {
4289 assert(!StressReflectiveCode, "stress mode does not use these paths");
4290 // Increase the size limit if we have exact knowledge of array type.
4291 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4292 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4293 }
4294
4295 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4296 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4297
4298 // --- Size Computation ---
4299 // array_size = round_to_heap(array_header + (length << elem_shift));
4300 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4301 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4302 // The rounding mask is strength-reduced, if possible.
4303 int round_mask = MinObjAlignmentInBytes - 1;
4304 Node* header_size = nullptr;
4305 // (T_BYTE has the weakest alignment and size restrictions...)
4306 if (layout_is_con) {
4307 int hsize = Klass::layout_helper_header_size(layout_con);
4308 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4309 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4310 if ((round_mask & ~right_n_bits(eshift)) == 0)
4311 round_mask = 0; // strength-reduce it if it goes away completely
4312 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4313 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4314 assert(header_size_min <= hsize, "generic minimum is smallest");
4315 header_size = intcon(hsize);
4316 } else {
4317 Node* hss = intcon(Klass::_lh_header_size_shift);
4318 Node* hsm = intcon(Klass::_lh_header_size_mask);
4319 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4320 header_size = _gvn.transform(new AndINode(header_size, hsm));
4321 }
4322
4323 Node* elem_shift = nullptr;
4324 if (layout_is_con) {
4325 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4326 if (eshift != 0)
4327 elem_shift = intcon(eshift);
4328 } else {
4329 // There is no need to mask or shift this value.
4330 // The semantics of LShiftINode include an implicit mask to 0x1F.
4331 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4332 elem_shift = layout_val;
4333 }
4334
4335 // Transition to native address size for all offset calculations:
4336 Node* lengthx = ConvI2X(length);
4337 Node* headerx = ConvI2X(header_size);
4338 #ifdef _LP64
4339 { const TypeInt* tilen = _gvn.find_int_type(length);
4340 if (tilen != nullptr && tilen->_lo < 0) {
4341 // Add a manual constraint to a positive range. Cf. array_element_address.
4342 jint size_max = fast_size_limit;
4343 if (size_max > tilen->_hi) size_max = tilen->_hi;
4344 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4345
4346 // Only do a narrow I2L conversion if the range check passed.
4347 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4348 _gvn.transform(iff);
4349 RegionNode* region = new RegionNode(3);
4350 _gvn.set_type(region, Type::CONTROL);
4351 lengthx = new PhiNode(region, TypeLong::LONG);
4352 _gvn.set_type(lengthx, TypeLong::LONG);
4353
4354 // Range check passed. Use ConvI2L node with narrow type.
4355 Node* passed = IfFalse(iff);
4356 region->init_req(1, passed);
4357 // Make I2L conversion control dependent to prevent it from
4358 // floating above the range check during loop optimizations.
4359 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4360
4361 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4362 region->init_req(2, IfTrue(iff));
4363 lengthx->init_req(2, ConvI2X(length));
4364
4365 set_control(region);
4366 record_for_igvn(region);
4367 record_for_igvn(lengthx);
4368 }
4369 }
4370 #endif
4371
4372 // Combine header size and body size for the array copy part, then align (if
4373 // necessary) for the allocation part. This computation cannot overflow,
4374 // because it is used only in two places, one where the length is sharply
4375 // limited, and the other after a successful allocation.
4376 Node* abody = lengthx;
4377 if (elem_shift != nullptr) {
4378 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4379 }
4380 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4381
4382 if (return_size_val != nullptr) {
4383 // This is the size
4384 (*return_size_val) = non_rounded_size;
4385 }
4386
4387 Node* size = non_rounded_size;
4388 if (round_mask != 0) {
4389 Node* mask1 = MakeConX(round_mask);
4390 size = _gvn.transform(new AddXNode(size, mask1));
4391 Node* mask2 = MakeConX(~round_mask);
4392 size = _gvn.transform(new AndXNode(size, mask2));
4393 }
4394 // else if round_mask == 0, the size computation is self-rounding
4395
4396 // Now generate allocation code
4397
4398 // The entire memory state is needed for slow path of the allocation
4399 // since GC and deoptimization can happen.
4400 Node *mem = reset_memory();
4401 set_all_memory(mem); // Create new memory state
4402
4403 if (initial_slow_test->is_Bool()) {
4404 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4405 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4406 }
4407
4408 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4409 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4410
4411 Node* raw_init_value = nullptr;
4412 if (init_val != nullptr) {
4413 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4414 if (ary_type->is_flat()) {
4415 initial_slow_test = intcon(1);
4416 }
4417
4418 if (UseCompressedOops) {
4419 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4420 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4421 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4422 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4423 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4424 } else {
4425 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4426 }
4427 }
4428
4429 Node* valid_length_test = _gvn.intcon(1);
4430 if (ary_type->isa_aryptr()) {
4431 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4432 jint max = TypeAryPtr::max_array_length(bt);
4433 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4434 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4435 }
4436
4437 // Create the AllocateArrayNode and its result projections
4438 AllocateArrayNode* alloc
4439 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4440 control(), mem, i_o(),
4441 size, klass_node,
4442 initial_slow_test,
4443 length, valid_length_test,
4444 init_val, raw_init_value);
4445 // Cast to correct type. Note that the klass_node may be constant or not,
4446 // and in the latter case the actual array type will be inexact also.
4447 // (This happens via a non-constant argument to inline_native_newArray.)
4448 // In any case, the value of klass_node provides the desired array type.
4449 const TypeInt* length_type = _gvn.find_int_type(length);
4450 if (ary_type->isa_aryptr() && length_type != nullptr) {
4451 // Try to get a better type than POS for the size
4452 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4453 }
4454
4455 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4456
4457 array_ideal_length(alloc, ary_type, true);
4458 return javaoop;
4459 }
4460
4461 // The following "Ideal_foo" functions are placed here because they recognize
4462 // the graph shapes created by the functions immediately above.
4463
4464 //---------------------------Ideal_allocation----------------------------------
4465 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4466 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4467 if (ptr == nullptr) { // reduce dumb test in callers
4468 return nullptr;
4469 }
4470
4471 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4472 ptr = bs->step_over_gc_barrier(ptr);
4473
4474 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4475 ptr = ptr->in(1);
4476 if (ptr == nullptr) return nullptr;
4477 }
4478 // Return null for allocations with several casts:
4479 // j.l.reflect.Array.newInstance(jobject, jint)
4480 // Object.clone()
4481 // to keep more precise type from last cast.
4482 if (ptr->is_Proj()) {
4483 Node* allo = ptr->in(0);
4484 if (allo != nullptr && allo->is_Allocate()) {
4485 return allo->as_Allocate();
4486 }
4487 }
4488 // Report failure to match.
4489 return nullptr;
4490 }
4491
4492 // Fancy version which also strips off an offset (and reports it to caller).
4493 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4494 intptr_t& offset) {
4495 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4496 if (base == nullptr) return nullptr;
4497 return Ideal_allocation(base);
4498 }
4499
4500 // Trace Initialize <- Proj[Parm] <- Allocate
4501 AllocateNode* InitializeNode::allocation() {
4502 Node* rawoop = in(InitializeNode::RawAddress);
4503 if (rawoop->is_Proj()) {
4504 Node* alloc = rawoop->in(0);
4505 if (alloc->is_Allocate()) {
4506 return alloc->as_Allocate();
4507 }
4508 }
4509 return nullptr;
4510 }
4511
4512 // Trace Allocate -> Proj[Parm] -> Initialize
4513 InitializeNode* AllocateNode::initialization() {
4514 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4515 if (rawoop == nullptr) return nullptr;
4516 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4517 Node* init = rawoop->fast_out(i);
4518 if (init->is_Initialize()) {
4519 assert(init->as_Initialize()->allocation() == this, "2-way link");
4520 return init->as_Initialize();
4521 }
4522 }
4523 return nullptr;
4524 }
4525
4526 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4527 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4528 // Too many traps seen?
4529 if (too_many_traps(reason)) {
4530 #ifdef ASSERT
4531 if (TraceLoopPredicate) {
4532 int tc = C->trap_count(reason);
4533 tty->print("too many traps=%s tcount=%d in ",
4534 Deoptimization::trap_reason_name(reason), tc);
4535 method()->print(); // which method has too many predicate traps
4536 tty->cr();
4537 }
4538 #endif
4539 // We cannot afford to take more traps here,
4540 // do not generate Parse Predicate.
4541 return;
4542 }
4543
4544 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4545 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4546 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4547 {
4548 PreserveJVMState pjvms(this);
4549 set_control(if_false);
4550 inc_sp(nargs);
4551 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4552 }
4553 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4554 set_control(if_true);
4555 }
4556
4557 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4558 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4559 void GraphKit::add_parse_predicates(int nargs) {
4560 if (UseLoopPredicate) {
4561 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4562 if (UseProfiledLoopPredicate) {
4563 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4564 }
4565 }
4566 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4567 // Loop Limit Check Predicate should be near the loop.
4568 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4569 }
4570
4571 void GraphKit::sync_kit(IdealKit& ideal) {
4572 set_all_memory(ideal.merged_memory());
4573 set_i_o(ideal.i_o());
4574 set_control(ideal.ctrl());
4575 }
4576
4577 void GraphKit::final_sync(IdealKit& ideal) {
4578 // Final sync IdealKit and graphKit.
4579 sync_kit(ideal);
4580 }
4581
4582 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4583 Node* len = load_array_length(load_String_value(str, set_ctrl));
4584 Node* coder = load_String_coder(str, set_ctrl);
4585 // Divide length by 2 if coder is UTF16
4586 return _gvn.transform(new RShiftINode(len, coder));
4587 }
4588
4589 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4590 int value_offset = java_lang_String::value_offset();
4591 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4592 false, nullptr, Type::Offset(0));
4593 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4594 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4595 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4596 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4597 Node* p = basic_plus_adr(str, str, value_offset);
4598 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4599 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4600 return load;
4601 }
4602
4603 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4604 if (!CompactStrings) {
4605 return intcon(java_lang_String::CODER_UTF16);
4606 }
4607 int coder_offset = java_lang_String::coder_offset();
4608 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4609 false, nullptr, Type::Offset(0));
4610 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4611
4612 Node* p = basic_plus_adr(str, str, coder_offset);
4613 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4614 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4615 return load;
4616 }
4617
4618 void GraphKit::store_String_value(Node* str, Node* value) {
4619 int value_offset = java_lang_String::value_offset();
4620 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4621 false, nullptr, Type::Offset(0));
4622 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4623
4624 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4625 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4626 }
4627
4628 void GraphKit::store_String_coder(Node* str, Node* value) {
4629 int coder_offset = java_lang_String::coder_offset();
4630 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4631 false, nullptr, Type::Offset(0));
4632 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4633
4634 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4635 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4636 }
4637
4638 // Capture src and dst memory state with a MergeMemNode
4639 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4640 if (src_type == dst_type) {
4641 // Types are equal, we don't need a MergeMemNode
4642 return memory(src_type);
4643 }
4644 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4645 record_for_igvn(merge); // fold it up later, if possible
4646 int src_idx = C->get_alias_index(src_type);
4647 int dst_idx = C->get_alias_index(dst_type);
4648 merge->set_memory_at(src_idx, memory(src_idx));
4649 merge->set_memory_at(dst_idx, memory(dst_idx));
4650 return merge;
4651 }
4652
4653 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4654 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4655 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4656 // If input and output memory types differ, capture both states to preserve
4657 // the dependency between preceding and subsequent loads/stores.
4658 // For example, the following program:
4659 // StoreB
4660 // compress_string
4661 // LoadB
4662 // has this memory graph (use->def):
4663 // LoadB -> compress_string -> CharMem
4664 // ... -> StoreB -> ByteMem
4665 // The intrinsic hides the dependency between LoadB and StoreB, causing
4666 // the load to read from memory not containing the result of the StoreB.
4667 // The correct memory graph should look like this:
4668 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4669 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4670 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4671 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4672 set_memory(res_mem, TypeAryPtr::BYTES);
4673 return str;
4674 }
4675
4676 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4677 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4678 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4679 // Capture src and dst memory (see comment in 'compress_string').
4680 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4681 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4682 set_memory(_gvn.transform(str), dst_type);
4683 }
4684
4685 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4686 /**
4687 * int i_char = start;
4688 * for (int i_byte = 0; i_byte < count; i_byte++) {
4689 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4690 * }
4691 */
4692 add_parse_predicates();
4693 C->set_has_loops(true);
4694
4695 RegionNode* head = new RegionNode(3);
4696 head->init_req(1, control());
4697 gvn().set_type(head, Type::CONTROL);
4698 record_for_igvn(head);
4699
4700 Node* i_byte = new PhiNode(head, TypeInt::INT);
4701 i_byte->init_req(1, intcon(0));
4702 gvn().set_type(i_byte, TypeInt::INT);
4703 record_for_igvn(i_byte);
4704
4705 Node* i_char = new PhiNode(head, TypeInt::INT);
4706 i_char->init_req(1, start);
4707 gvn().set_type(i_char, TypeInt::INT);
4708 record_for_igvn(i_char);
4709
4710 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4711 gvn().set_type(mem, Type::MEMORY);
4712 record_for_igvn(mem);
4713 set_control(head);
4714 set_memory(mem, TypeAryPtr::BYTES);
4715 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4716 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4717 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4718 false, true /* mismatched */);
4719
4720 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4721 head->init_req(2, IfTrue(iff));
4722 mem->init_req(2, st);
4723 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4724 i_char->init_req(2, AddI(i_char, intcon(2)));
4725
4726 set_control(IfFalse(iff));
4727 set_memory(st, TypeAryPtr::BYTES);
4728 }
4729
4730 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4731 if (!field->is_constant()) {
4732 return nullptr; // Field not marked as constant.
4733 }
4734 ciInstance* holder = nullptr;
4735 if (!field->is_static()) {
4736 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4737 if (const_oop != nullptr && const_oop->is_instance()) {
4738 holder = const_oop->as_instance();
4739 }
4740 }
4741 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4742 /*is_unsigned_load=*/false);
4743 if (con_type != nullptr) {
4744 Node* con = makecon(con_type);
4745 if (field->type()->is_inlinetype()) {
4746 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4747 } else if (con_type->is_inlinetypeptr()) {
4748 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4749 }
4750 return con;
4751 }
4752 return nullptr;
4753 }
4754
4755 //---------------------------load_mirror_from_klass----------------------------
4756 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4757 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4758 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4759 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4760 // mirror = ((OopHandle)mirror)->resolve();
4761 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4762 }
4763
4764 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4765 const Type* obj_type = obj->bottom_type();
4766 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4767 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4768 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4769 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4770 obj = casted_obj;
4771 }
4772 if (sig_type->is_inlinetypeptr()) {
4773 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4774 }
4775 return obj;
4776 }