1 /*
2 * Copyright (c) 2001, 2026, 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 "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/parse.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/runtime.hpp"
55 #include "opto/subtypenode.hpp"
56 #include "runtime/arguments.hpp"
57 #include "runtime/deoptimization.hpp"
58 #include "runtime/sharedRuntime.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "utilities/bitMap.inline.hpp"
61 #include "utilities/growableArray.hpp"
62 #include "utilities/powerOfTwo.hpp"
63
64 //----------------------------GraphKit-----------------------------------------
65 // Main utility constructor.
66 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
67 : Phase(Phase::Parser),
68 _env(C->env()),
69 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
70 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
71 {
72 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
73 _exceptions = jvms->map()->next_exception();
74 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
75 set_jvms(jvms);
76 #ifdef ASSERT
77 if (_gvn.is_IterGVN() != nullptr) {
78 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
79 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
80 _worklist_size = _gvn.C->igvn_worklist()->size();
81 }
82 #endif
83 }
84
85 // Private constructor for parser.
86 GraphKit::GraphKit()
87 : Phase(Phase::Parser),
88 _env(C->env()),
89 _gvn(*C->initial_gvn()),
90 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
91 {
92 _exceptions = nullptr;
93 set_map(nullptr);
94 DEBUG_ONLY(_sp = -99);
95 DEBUG_ONLY(set_bci(-99));
96 }
97
98 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
99 : Phase(Phase::Parser),
100 _env(C->env()),
101 _gvn(igvn),
102 _exceptions(nullptr),
103 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
104 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
105 assert(sft->next_exception() == nullptr, "must not have a pending exception");
106 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
107 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
108 for (uint i = 0; i < sft->req(); i++) {
109 cloned_map->init_req(i, sft->in(i));
110 }
111 igvn.record_for_igvn(cloned_map);
112 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
113 current->set_map(cloned_map);
114 }
115 set_jvms(cloned_jvms);
116 set_all_memory(reset_memory());
117 }
118
119 //---------------------------clean_stack---------------------------------------
120 // Clear away rubbish from the stack area of the JVM state.
121 // This destroys any arguments that may be waiting on the stack.
122 void GraphKit::clean_stack(int from_sp) {
123 SafePointNode* map = this->map();
124 JVMState* jvms = this->jvms();
125 int stk_size = jvms->stk_size();
126 int stkoff = jvms->stkoff();
127 Node* top = this->top();
128 for (int i = from_sp; i < stk_size; i++) {
129 if (map->in(stkoff + i) != top) {
130 map->set_req(stkoff + i, top);
131 }
132 }
133 }
134
135
136 //--------------------------------sync_jvms-----------------------------------
137 // Make sure our current jvms agrees with our parse state.
138 JVMState* GraphKit::sync_jvms() const {
139 JVMState* jvms = this->jvms();
140 jvms->set_bci(bci()); // Record the new bci in the JVMState
141 jvms->set_sp(sp()); // Record the new sp in the JVMState
142 assert(jvms_in_sync(), "jvms is now in sync");
143 return jvms;
144 }
145
146 //--------------------------------sync_jvms_for_reexecute---------------------
147 // Make sure our current jvms agrees with our parse state. This version
148 // uses the reexecute_sp for reexecuting bytecodes.
149 JVMState* GraphKit::sync_jvms_for_reexecute() {
150 JVMState* jvms = this->jvms();
151 jvms->set_bci(bci()); // Record the new bci in the JVMState
152 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
153 return jvms;
154 }
155
156 #ifdef ASSERT
157 bool GraphKit::jvms_in_sync() const {
158 Parse* parse = is_Parse();
159 if (parse == nullptr) {
160 if (bci() != jvms()->bci()) return false;
161 if (sp() != (int)jvms()->sp()) return false;
162 return true;
163 }
164 if (jvms()->method() != parse->method()) return false;
165 if (jvms()->bci() != parse->bci()) return false;
166 int jvms_sp = jvms()->sp();
167 if (jvms_sp != parse->sp()) return false;
168 int jvms_depth = jvms()->depth();
169 if (jvms_depth != parse->depth()) return false;
170 return true;
171 }
172
173 // Local helper checks for special internal merge points
174 // used to accumulate and merge exception states.
175 // They are marked by the region's in(0) edge being the map itself.
176 // Such merge points must never "escape" into the parser at large,
177 // until they have been handed to gvn.transform.
178 static bool is_hidden_merge(Node* reg) {
179 if (reg == nullptr) return false;
180 if (reg->is_Phi()) {
181 reg = reg->in(0);
182 if (reg == nullptr) return false;
183 }
184 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
185 }
186
187 void GraphKit::verify_map() const {
188 if (map() == nullptr) return; // null map is OK
189 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
190 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
191 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
192 }
193
194 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
195 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
196 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
197 }
198 #endif
199
200 //---------------------------stop_and_kill_map---------------------------------
201 // Set _map to null, signalling a stop to further bytecode execution.
202 // First smash the current map's control to a constant, to mark it dead.
203 void GraphKit::stop_and_kill_map() {
204 SafePointNode* dead_map = stop();
205 if (dead_map != nullptr) {
206 dead_map->disconnect_inputs(C); // Mark the map as killed.
207 assert(dead_map->is_killed(), "must be so marked");
208 }
209 }
210
211
212 //--------------------------------stopped--------------------------------------
213 // Tell if _map is null, or control is top.
214 bool GraphKit::stopped() {
215 if (map() == nullptr) return true;
216 else if (control() == top()) return true;
217 else return false;
218 }
219
220
221 //-----------------------------has_exception_handler----------------------------------
222 // Tell if this method or any caller method has exception handlers.
223 bool GraphKit::has_exception_handler() {
224 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
225 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
226 return true;
227 }
228 }
229 return false;
230 }
231
232 //------------------------------save_ex_oop------------------------------------
233 // Save an exception without blowing stack contents or other JVM state.
234 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
235 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
236 ex_map->add_req(ex_oop);
237 DEBUG_ONLY(verify_exception_state(ex_map));
238 }
239
240 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
241 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
242 Node* ex_oop = ex_map->in(ex_map->req()-1);
243 if (clear_it) ex_map->del_req(ex_map->req()-1);
244 return ex_oop;
245 }
246
247 //-----------------------------saved_ex_oop------------------------------------
248 // Recover a saved exception from its map.
249 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
250 return common_saved_ex_oop(ex_map, false);
251 }
252
253 //--------------------------clear_saved_ex_oop---------------------------------
254 // Erase a previously saved exception from its map.
255 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
256 return common_saved_ex_oop(ex_map, true);
257 }
258
259 #ifdef ASSERT
260 //---------------------------has_saved_ex_oop----------------------------------
261 // Erase a previously saved exception from its map.
262 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
263 return ex_map->req() == ex_map->jvms()->endoff()+1;
264 }
265 #endif
266
267 //-------------------------make_exception_state--------------------------------
268 // Turn the current JVM state into an exception state, appending the ex_oop.
269 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
270 sync_jvms();
271 SafePointNode* ex_map = stop(); // do not manipulate this map any more
272 set_saved_ex_oop(ex_map, ex_oop);
273 return ex_map;
274 }
275
276
277 //--------------------------add_exception_state--------------------------------
278 // Add an exception to my list of exceptions.
279 void GraphKit::add_exception_state(SafePointNode* ex_map) {
280 if (ex_map == nullptr || ex_map->control() == top()) {
281 return;
282 }
283 #ifdef ASSERT
284 verify_exception_state(ex_map);
285 if (has_exceptions()) {
286 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
287 }
288 #endif
289
290 // If there is already an exception of exactly this type, merge with it.
291 // In particular, null-checks and other low-level exceptions common up here.
292 Node* ex_oop = saved_ex_oop(ex_map);
293 const Type* ex_type = _gvn.type(ex_oop);
294 if (ex_oop == top()) {
295 // No action needed.
296 return;
297 }
298 assert(ex_type->isa_instptr(), "exception must be an instance");
299 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
300 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
301 // We check sp also because call bytecodes can generate exceptions
302 // both before and after arguments are popped!
303 if (ex_type2 == ex_type
304 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
305 combine_exception_states(ex_map, e2);
306 return;
307 }
308 }
309
310 // No pre-existing exception of the same type. Chain it on the list.
311 push_exception_state(ex_map);
312 }
313
314 //-----------------------add_exception_states_from-----------------------------
315 void GraphKit::add_exception_states_from(JVMState* jvms) {
316 SafePointNode* ex_map = jvms->map()->next_exception();
317 if (ex_map != nullptr) {
318 jvms->map()->set_next_exception(nullptr);
319 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
320 next_map = ex_map->next_exception();
321 ex_map->set_next_exception(nullptr);
322 add_exception_state(ex_map);
323 }
324 }
325 }
326
327 //-----------------------transfer_exceptions_into_jvms-------------------------
328 JVMState* GraphKit::transfer_exceptions_into_jvms() {
329 if (map() == nullptr) {
330 // We need a JVMS to carry the exceptions, but the map has gone away.
331 // Create a scratch JVMS, cloned from any of the exception states...
332 if (has_exceptions()) {
333 _map = _exceptions;
334 _map = clone_map();
335 _map->set_next_exception(nullptr);
336 clear_saved_ex_oop(_map);
337 DEBUG_ONLY(verify_map());
338 } else {
339 // ...or created from scratch
340 JVMState* jvms = new (C) JVMState(_method, nullptr);
341 jvms->set_bci(_bci);
342 jvms->set_sp(_sp);
343 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
344 set_jvms(jvms);
345 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
346 set_all_memory(top());
347 while (map()->req() < jvms->endoff()) map()->add_req(top());
348 }
349 // (This is a kludge, in case you didn't notice.)
350 set_control(top());
351 }
352 JVMState* jvms = sync_jvms();
353 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
354 jvms->map()->set_next_exception(_exceptions);
355 _exceptions = nullptr; // done with this set of exceptions
356 return jvms;
357 }
358
359 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
360 assert(is_hidden_merge(dstphi), "must be a special merge node");
361 assert(is_hidden_merge(srcphi), "must be a special merge node");
362 uint limit = srcphi->req();
363 for (uint i = PhiNode::Input; i < limit; i++) {
364 dstphi->add_req(srcphi->in(i));
365 }
366 }
367 static inline void add_one_req(Node* dstphi, Node* src) {
368 assert(is_hidden_merge(dstphi), "must be a special merge node");
369 assert(!is_hidden_merge(src), "must not be a special merge node");
370 dstphi->add_req(src);
371 }
372
373 //-----------------------combine_exception_states------------------------------
374 // This helper function combines exception states by building phis on a
375 // specially marked state-merging region. These regions and phis are
376 // untransformed, and can build up gradually. The region is marked by
377 // having a control input of its exception map, rather than null. Such
378 // regions do not appear except in this function, and in use_exception_state.
379 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
380 if (failing_internal()) {
381 return; // dying anyway...
382 }
383 JVMState* ex_jvms = ex_map->_jvms;
384 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
385 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
386 // TODO 8325632 Re-enable
387 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
388 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
389 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
390 assert(ex_map->req() == phi_map->req(), "matching maps");
391 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
392 Node* hidden_merge_mark = root();
393 Node* region = phi_map->control();
394 MergeMemNode* phi_mem = phi_map->merged_memory();
395 MergeMemNode* ex_mem = ex_map->merged_memory();
396 if (region->in(0) != hidden_merge_mark) {
397 // The control input is not (yet) a specially-marked region in phi_map.
398 // Make it so, and build some phis.
399 region = new RegionNode(2);
400 _gvn.set_type(region, Type::CONTROL);
401 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
402 region->init_req(1, phi_map->control());
403 phi_map->set_control(region);
404 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
405 record_for_igvn(io_phi);
406 _gvn.set_type(io_phi, Type::ABIO);
407 phi_map->set_i_o(io_phi);
408 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
409 Node* m = mms.memory();
410 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
411 record_for_igvn(m_phi);
412 _gvn.set_type(m_phi, Type::MEMORY);
413 mms.set_memory(m_phi);
414 }
415 }
416
417 // Either or both of phi_map and ex_map might already be converted into phis.
418 Node* ex_control = ex_map->control();
419 // if there is special marking on ex_map also, we add multiple edges from src
420 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
421 // how wide was the destination phi_map, originally?
422 uint orig_width = region->req();
423
424 if (add_multiple) {
425 add_n_reqs(region, ex_control);
426 add_n_reqs(phi_map->i_o(), ex_map->i_o());
427 } else {
428 // ex_map has no merges, so we just add single edges everywhere
429 add_one_req(region, ex_control);
430 add_one_req(phi_map->i_o(), ex_map->i_o());
431 }
432 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
433 if (mms.is_empty()) {
434 // get a copy of the base memory, and patch some inputs into it
435 const TypePtr* adr_type = mms.adr_type(C);
436 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
437 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
438 mms.set_memory(phi);
439 // Prepare to append interesting stuff onto the newly sliced phi:
440 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
441 }
442 // Append stuff from ex_map:
443 if (add_multiple) {
444 add_n_reqs(mms.memory(), mms.memory2());
445 } else {
446 add_one_req(mms.memory(), mms.memory2());
447 }
448 }
449 uint limit = ex_map->req();
450 for (uint i = TypeFunc::Parms; i < limit; i++) {
451 // Skip everything in the JVMS after tos. (The ex_oop follows.)
452 if (i == tos) i = ex_jvms->monoff();
453 Node* src = ex_map->in(i);
454 Node* dst = phi_map->in(i);
455 if (src != dst) {
456 PhiNode* phi;
457 if (dst->in(0) != region) {
458 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
459 record_for_igvn(phi);
460 _gvn.set_type(phi, phi->type());
461 phi_map->set_req(i, dst);
462 // Prepare to append interesting stuff onto the new phi:
463 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
464 } else {
465 assert(dst->is_Phi(), "nobody else uses a hidden region");
466 phi = dst->as_Phi();
467 }
468 if (add_multiple && src->in(0) == ex_control) {
469 // Both are phis.
470 add_n_reqs(dst, src);
471 } else {
472 while (dst->req() < region->req()) add_one_req(dst, src);
473 }
474 const Type* srctype = _gvn.type(src);
475 if (phi->type() != srctype) {
476 const Type* dsttype = phi->type()->meet_speculative(srctype);
477 if (phi->type() != dsttype) {
478 phi->set_type(dsttype);
479 _gvn.set_type(phi, dsttype);
480 }
481 }
482 }
483 }
484 phi_map->merge_replaced_nodes_with(ex_map);
485 }
486
487 //--------------------------use_exception_state--------------------------------
488 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
489 if (failing_internal()) { stop(); return top(); }
490 Node* region = phi_map->control();
491 Node* hidden_merge_mark = root();
492 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
493 Node* ex_oop = clear_saved_ex_oop(phi_map);
494 if (region->in(0) == hidden_merge_mark) {
495 // Special marking for internal ex-states. Process the phis now.
496 region->set_req(0, region); // now it's an ordinary region
497 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
498 // Note: Setting the jvms also sets the bci and sp.
499 set_control(_gvn.transform(region));
500 uint tos = jvms()->stkoff() + sp();
501 for (uint i = 1; i < tos; i++) {
502 Node* x = phi_map->in(i);
503 if (x->in(0) == region) {
504 assert(x->is_Phi(), "expected a special phi");
505 phi_map->set_req(i, _gvn.transform(x));
506 }
507 }
508 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
509 Node* x = mms.memory();
510 if (x->in(0) == region) {
511 assert(x->is_Phi(), "nobody else uses a hidden region");
512 mms.set_memory(_gvn.transform(x));
513 }
514 }
515 if (ex_oop->in(0) == region) {
516 assert(ex_oop->is_Phi(), "expected a special phi");
517 ex_oop = _gvn.transform(ex_oop);
518 }
519 } else {
520 set_jvms(phi_map->jvms());
521 }
522
523 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
524 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
525 return ex_oop;
526 }
527
528 //---------------------------------java_bc-------------------------------------
529 Bytecodes::Code GraphKit::java_bc() const {
530 ciMethod* method = this->method();
531 int bci = this->bci();
532 if (method != nullptr && bci != InvocationEntryBci)
533 return method->java_code_at_bci(bci);
534 else
535 return Bytecodes::_illegal;
536 }
537
538 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
539 bool must_throw) {
540 // if the exception capability is set, then we will generate code
541 // to check the JavaThread.should_post_on_exceptions flag to see
542 // if we actually need to report exception events (for this
543 // thread). If we don't need to report exception events, we will
544 // take the normal fast path provided by add_exception_events. If
545 // exception event reporting is enabled for this thread, we will
546 // take the uncommon_trap in the BuildCutout below.
547
548 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
549 Node* jthread = _gvn.transform(new ThreadLocalNode());
550 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
551 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
552
553 // Test the should_post_on_exceptions_flag vs. 0
554 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
555 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
556
557 // Branch to slow_path if should_post_on_exceptions_flag was true
558 { BuildCutout unless(this, tst, PROB_MAX);
559 // Do not try anything fancy if we're notifying the VM on every throw.
560 // Cf. case Bytecodes::_athrow in parse2.cpp.
561 uncommon_trap(reason, Deoptimization::Action_none,
562 (ciKlass*)nullptr, (char*)nullptr, must_throw);
563 }
564
565 }
566
567 //------------------------------builtin_throw----------------------------------
568 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
569 builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true);
570 }
571
572 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason,
573 ciInstance* ex_obj,
574 bool allow_too_many_traps) {
575 // If this throw happens frequently, an uncommon trap might cause
576 // a performance pothole. If there is a local exception handler,
577 // and if this particular bytecode appears to be deoptimizing often,
578 // let us handle the throw inline, with a preconstructed instance.
579 // Note: If the deopt count has blown up, the uncommon trap
580 // runtime is going to flush this nmethod, not matter what.
581 if (is_builtin_throw_hot(reason)) {
582 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
583 // If the throw is local, we use a pre-existing instance and
584 // punt on the backtrace. This would lead to a missing backtrace
585 // (a repeat of 4292742) if the backtrace object is ever asked
586 // for its backtrace.
587 // Fixing this remaining case of 4292742 requires some flavor of
588 // escape analysis. Leave that for the future.
589 if (env()->jvmti_can_post_on_exceptions()) {
590 // check if we must post exception events, take uncommon trap if so
591 uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/);
592 // here if should_post_on_exceptions is false
593 // continue on with the normal codegen
594 }
595
596 // Cheat with a preallocated exception object.
597 if (C->log() != nullptr)
598 C->log()->elem("hot_throw preallocated='1' reason='%s'",
599 Deoptimization::trap_reason_name(reason));
600 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
601 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
602
603 // Clear the detail message of the preallocated exception object.
604 // Weblogic sometimes mutates the detail message of exceptions
605 // using reflection.
606 int offset = java_lang_Throwable::get_detailMessage_offset();
607 const TypePtr* adr_typ = ex_con->add_offset(offset);
608
609 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
610 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
611 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
612
613 if (!method()->has_exception_handlers()) {
614 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
615 // This prevents issues with exception handling and late inlining.
616 set_sp(0);
617 clean_stack(0);
618 }
619
620 add_exception_state(make_exception_state(ex_node));
621 return;
622 } else if (builtin_throw_too_many_traps(reason, ex_obj)) {
623 // We cannot afford to take too many traps here. Suffer in the interpreter instead.
624 assert(allow_too_many_traps, "not allowed");
625 if (C->log() != nullptr) {
626 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
627 Deoptimization::trap_reason_name(reason),
628 C->trap_count(reason));
629 }
630 uncommon_trap(reason, Deoptimization::Action_none,
631 (ciKlass*) nullptr, (char*) nullptr,
632 true /*must_throw*/);
633 return;
634 }
635 }
636
637 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
638 // It won't be much cheaper than bailing to the interp., since we'll
639 // have to pass up all the debug-info, and the runtime will have to
640 // create the stack trace.
641
642 // Usual case: Bail to interpreter.
643 // Reserve the right to recompile if we haven't seen anything yet.
644
645 // "must_throw" prunes the JVM state to include only the stack, if there
646 // are no local exception handlers. This should cut down on register
647 // allocation time and code size, by drastically reducing the number
648 // of in-edges on the call to the uncommon trap.
649 uncommon_trap(reason, Deoptimization::Action_maybe_recompile,
650 (ciKlass*) nullptr, (char*) nullptr,
651 true /*must_throw*/);
652 }
653
654 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) {
655 // If this particular condition has not yet happened at this
656 // bytecode, then use the uncommon trap mechanism, and allow for
657 // a future recompilation if several traps occur here.
658 // If the throw is hot, try to use a more complicated inline mechanism
659 // which keeps execution inside the compiled code.
660 if (ProfileTraps) {
661 if (too_many_traps(reason)) {
662 return true;
663 }
664 // (If there is no MDO at all, assume it is early in
665 // execution, and that any deopts are part of the
666 // startup transient, and don't need to be remembered.)
667
668 // Also, if there is a local exception handler, treat all throws
669 // as hot if there has been at least one in this method.
670 if (C->trap_count(reason) != 0 &&
671 method()->method_data()->trap_count(reason) != 0 &&
672 has_exception_handler()) {
673 return true;
674 }
675 }
676 return false;
677 }
678
679 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason,
680 ciInstance* ex_obj) {
681 if (is_builtin_throw_hot(reason)) {
682 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
683 return false; // no traps; throws preallocated exception instead
684 }
685 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
686 if (method()->method_data()->trap_recompiled_at(bci(), m) ||
687 C->too_many_traps(reason)) {
688 return true;
689 }
690 }
691 return false;
692 }
693
694 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const {
695 // Preallocated exception objects to use when we don't need the backtrace.
696 switch (reason) {
697 case Deoptimization::Reason_null_check:
698 return env()->NullPointerException_instance();
699 case Deoptimization::Reason_div0_check:
700 return env()->ArithmeticException_instance();
701 case Deoptimization::Reason_range_check:
702 return env()->ArrayIndexOutOfBoundsException_instance();
703 case Deoptimization::Reason_class_check:
704 return env()->ClassCastException_instance();
705 case Deoptimization::Reason_array_check:
706 return env()->ArrayStoreException_instance();
707 default:
708 return nullptr;
709 }
710 }
711
712 //----------------------------PreserveJVMState---------------------------------
713 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
714 DEBUG_ONLY(kit->verify_map());
715 _kit = kit;
716 _map = kit->map(); // preserve the map
717 _sp = kit->sp();
718 kit->set_map(clone_map ? kit->clone_map() : nullptr);
719 #ifdef ASSERT
720 _bci = kit->bci();
721 Parse* parser = kit->is_Parse();
722 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
723 _block = block;
724 #endif
725 }
726 PreserveJVMState::~PreserveJVMState() {
727 GraphKit* kit = _kit;
728 #ifdef ASSERT
729 assert(kit->bci() == _bci, "bci must not shift");
730 Parse* parser = kit->is_Parse();
731 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
732 assert(block == _block, "block must not shift");
733 #endif
734 kit->set_map(_map);
735 kit->set_sp(_sp);
736 }
737
738
739 //-----------------------------BuildCutout-------------------------------------
740 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
741 : PreserveJVMState(kit)
742 {
743 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
744 SafePointNode* outer_map = _map; // preserved map is caller's
745 SafePointNode* inner_map = kit->map();
746 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
747 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
748 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
749 }
750 BuildCutout::~BuildCutout() {
751 GraphKit* kit = _kit;
752 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
753 }
754
755 //---------------------------PreserveReexecuteState----------------------------
756 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
757 assert(!kit->stopped(), "must call stopped() before");
758 _kit = kit;
759 _sp = kit->sp();
760 _reexecute = kit->jvms()->_reexecute;
761 }
762 PreserveReexecuteState::~PreserveReexecuteState() {
763 if (_kit->stopped()) return;
764 _kit->jvms()->_reexecute = _reexecute;
765 _kit->set_sp(_sp);
766 }
767
768 //------------------------------clone_map--------------------------------------
769 // Implementation of PreserveJVMState
770 //
771 // Only clone_map(...) here. If this function is only used in the
772 // PreserveJVMState class we may want to get rid of this extra
773 // function eventually and do it all there.
774
775 SafePointNode* GraphKit::clone_map() {
776 if (map() == nullptr) return nullptr;
777
778 // Clone the memory edge first
779 Node* mem = MergeMemNode::make(map()->memory());
780 gvn().set_type_bottom(mem);
781
782 SafePointNode *clonemap = (SafePointNode*)map()->clone();
783 JVMState* jvms = this->jvms();
784 JVMState* clonejvms = jvms->clone_shallow(C);
785 clonemap->set_memory(mem);
786 clonemap->set_jvms(clonejvms);
787 clonejvms->set_map(clonemap);
788 record_for_igvn(clonemap);
789 gvn().set_type_bottom(clonemap);
790 return clonemap;
791 }
792
793 //-----------------------------destruct_map_clone------------------------------
794 //
795 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
796 // to destruct/free/delete in the exact opposite order as clone_map().
797 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
798 if (sfp == nullptr) return;
799
800 Node* mem = sfp->memory();
801 JVMState* jvms = sfp->jvms();
802
803 if (jvms != nullptr) {
804 delete jvms;
805 }
806
807 remove_for_igvn(sfp);
808 gvn().clear_type(sfp);
809 sfp->destruct(&_gvn);
810
811 if (mem != nullptr) {
812 gvn().clear_type(mem);
813 mem->destruct(&_gvn);
814 }
815 }
816
817 //-----------------------------set_map_clone-----------------------------------
818 void GraphKit::set_map_clone(SafePointNode* m) {
819 _map = m;
820 _map = clone_map();
821 _map->set_next_exception(nullptr);
822 DEBUG_ONLY(verify_map());
823 }
824
825
826 //----------------------------kill_dead_locals---------------------------------
827 // Detect any locals which are known to be dead, and force them to top.
828 void GraphKit::kill_dead_locals() {
829 // Consult the liveness information for the locals. If any
830 // of them are unused, then they can be replaced by top(). This
831 // should help register allocation time and cut down on the size
832 // of the deoptimization information.
833
834 // This call is made from many of the bytecode handling
835 // subroutines called from the Big Switch in do_one_bytecode.
836 // Every bytecode which might include a slow path is responsible
837 // for killing its dead locals. The more consistent we
838 // are about killing deads, the fewer useless phis will be
839 // constructed for them at various merge points.
840
841 // bci can be -1 (InvocationEntryBci). We return the entry
842 // liveness for the method.
843
844 if (method() == nullptr || method()->code_size() == 0) {
845 // We are building a graph for a call to a native method.
846 // All locals are live.
847 return;
848 }
849
850 ResourceMark rm;
851
852 // Consult the liveness information for the locals. If any
853 // of them are unused, then they can be replaced by top(). This
854 // should help register allocation time and cut down on the size
855 // of the deoptimization information.
856 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
857
858 int len = (int)live_locals.size();
859 assert(len <= jvms()->loc_size(), "too many live locals");
860 for (int local = 0; local < len; local++) {
861 if (!live_locals.at(local)) {
862 set_local(local, top());
863 }
864 }
865 }
866
867 #ifdef ASSERT
868 //-------------------------dead_locals_are_killed------------------------------
869 // Return true if all dead locals are set to top in the map.
870 // Used to assert "clean" debug info at various points.
871 bool GraphKit::dead_locals_are_killed() {
872 if (method() == nullptr || method()->code_size() == 0) {
873 // No locals need to be dead, so all is as it should be.
874 return true;
875 }
876
877 // Make sure somebody called kill_dead_locals upstream.
878 ResourceMark rm;
879 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
880 if (jvms->loc_size() == 0) continue; // no locals to consult
881 SafePointNode* map = jvms->map();
882 ciMethod* method = jvms->method();
883 int bci = jvms->bci();
884 if (jvms == this->jvms()) {
885 bci = this->bci(); // it might not yet be synched
886 }
887 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
888 int len = (int)live_locals.size();
889 if (!live_locals.is_valid() || len == 0)
890 // This method is trivial, or is poisoned by a breakpoint.
891 return true;
892 assert(len == jvms->loc_size(), "live map consistent with locals map");
893 for (int local = 0; local < len; local++) {
894 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
895 if (PrintMiscellaneous && (Verbose || WizardMode)) {
896 tty->print_cr("Zombie local %d: ", local);
897 jvms->dump();
898 }
899 return false;
900 }
901 }
902 }
903 return true;
904 }
905
906 #endif //ASSERT
907
908 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
909 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
910 ciMethod* cur_method = jvms->method();
911 int cur_bci = jvms->bci();
912 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
913 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
914 return Interpreter::bytecode_should_reexecute(code) ||
915 (is_anewarray && (code == Bytecodes::_multianewarray));
916 // Reexecute _multianewarray bytecode which was replaced with
917 // sequence of [a]newarray. See Parse::do_multianewarray().
918 //
919 // Note: interpreter should not have it set since this optimization
920 // is limited by dimensions and guarded by flag so in some cases
921 // multianewarray() runtime calls will be generated and
922 // the bytecode should not be reexecutes (stack will not be reset).
923 } else {
924 return false;
925 }
926 }
927
928 // Helper function for adding JVMState and debug information to node
929 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
930 // Add the safepoint edges to the call (or other safepoint).
931
932 // Make sure dead locals are set to top. This
933 // should help register allocation time and cut down on the size
934 // of the deoptimization information.
935 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
936
937 // Walk the inline list to fill in the correct set of JVMState's
938 // Also fill in the associated edges for each JVMState.
939
940 // If the bytecode needs to be reexecuted we need to put
941 // the arguments back on the stack.
942 const bool should_reexecute = jvms()->should_reexecute();
943 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
944
945 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
946 // undefined if the bci is different. This is normal for Parse but it
947 // should not happen for LibraryCallKit because only one bci is processed.
948 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
949 "in LibraryCallKit the reexecute bit should not change");
950
951 // If we are guaranteed to throw, we can prune everything but the
952 // input to the current bytecode.
953 bool can_prune_locals = false;
954 uint stack_slots_not_pruned = 0;
955 int inputs = 0, depth = 0;
956 if (must_throw) {
957 assert(method() == youngest_jvms->method(), "sanity");
958 if (compute_stack_effects(inputs, depth)) {
959 can_prune_locals = true;
960 stack_slots_not_pruned = inputs;
961 }
962 }
963
964 if (env()->should_retain_local_variables()) {
965 // At any safepoint, this method can get breakpointed, which would
966 // then require an immediate deoptimization.
967 can_prune_locals = false; // do not prune locals
968 stack_slots_not_pruned = 0;
969 }
970
971 // do not scribble on the input jvms
972 JVMState* out_jvms = youngest_jvms->clone_deep(C);
973 call->set_jvms(out_jvms); // Start jvms list for call node
974
975 // For a known set of bytecodes, the interpreter should reexecute them if
976 // deoptimization happens. We set the reexecute state for them here
977 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
978 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
979 #ifdef ASSERT
980 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
981 assert(method() == youngest_jvms->method(), "sanity");
982 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
983 // TODO 8371125
984 // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
985 #endif // ASSERT
986 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
987 }
988
989 // Presize the call:
990 DEBUG_ONLY(uint non_debug_edges = call->req());
991 call->add_req_batch(top(), youngest_jvms->debug_depth());
992 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
993
994 // Set up edges so that the call looks like this:
995 // Call [state:] ctl io mem fptr retadr
996 // [parms:] parm0 ... parmN
997 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
998 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
999 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1000 // Note that caller debug info precedes callee debug info.
1001
1002 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1003 uint debug_ptr = call->req();
1004
1005 // Loop over the map input edges associated with jvms, add them
1006 // to the call node, & reset all offsets to match call node array.
1007
1008 JVMState* callee_jvms = nullptr;
1009 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1010 uint debug_end = debug_ptr;
1011 uint debug_start = debug_ptr - in_jvms->debug_size();
1012 debug_ptr = debug_start; // back up the ptr
1013
1014 uint p = debug_start; // walks forward in [debug_start, debug_end)
1015 uint j, k, l;
1016 SafePointNode* in_map = in_jvms->map();
1017 out_jvms->set_map(call);
1018
1019 if (can_prune_locals) {
1020 assert(in_jvms->method() == out_jvms->method(), "sanity");
1021 // If the current throw can reach an exception handler in this JVMS,
1022 // then we must keep everything live that can reach that handler.
1023 // As a quick and dirty approximation, we look for any handlers at all.
1024 if (in_jvms->method()->has_exception_handlers()) {
1025 can_prune_locals = false;
1026 }
1027 }
1028
1029 // Add the Locals
1030 k = in_jvms->locoff();
1031 l = in_jvms->loc_size();
1032 out_jvms->set_locoff(p);
1033 if (!can_prune_locals) {
1034 for (j = 0; j < l; j++) {
1035 call->set_req(p++, in_map->in(k + j));
1036 }
1037 } else {
1038 p += l; // already set to top above by add_req_batch
1039 }
1040
1041 // Add the Expression Stack
1042 k = in_jvms->stkoff();
1043 l = in_jvms->sp();
1044 out_jvms->set_stkoff(p);
1045 if (!can_prune_locals) {
1046 for (j = 0; j < l; j++) {
1047 call->set_req(p++, in_map->in(k + j));
1048 }
1049 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1050 // Divide stack into {S0,...,S1}, where S0 is set to top.
1051 uint s1 = stack_slots_not_pruned;
1052 stack_slots_not_pruned = 0; // for next iteration
1053 if (s1 > l) s1 = l;
1054 uint s0 = l - s1;
1055 p += s0; // skip the tops preinstalled by add_req_batch
1056 for (j = s0; j < l; j++)
1057 call->set_req(p++, in_map->in(k+j));
1058 } else {
1059 p += l; // already set to top above by add_req_batch
1060 }
1061
1062 // Add the Monitors
1063 k = in_jvms->monoff();
1064 l = in_jvms->mon_size();
1065 out_jvms->set_monoff(p);
1066 for (j = 0; j < l; j++)
1067 call->set_req(p++, in_map->in(k+j));
1068
1069 // Copy any scalar object fields.
1070 k = in_jvms->scloff();
1071 l = in_jvms->scl_size();
1072 out_jvms->set_scloff(p);
1073 for (j = 0; j < l; j++)
1074 call->set_req(p++, in_map->in(k+j));
1075
1076 // Finish the new jvms.
1077 out_jvms->set_endoff(p);
1078
1079 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1080 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1081 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1082 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1083 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1084 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1085
1086 // Update the two tail pointers in parallel.
1087 callee_jvms = out_jvms;
1088 out_jvms = out_jvms->caller();
1089 in_jvms = in_jvms->caller();
1090 }
1091
1092 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1093
1094 // Test the correctness of JVMState::debug_xxx accessors:
1095 assert(call->jvms()->debug_start() == non_debug_edges, "");
1096 assert(call->jvms()->debug_end() == call->req(), "");
1097 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1098 }
1099
1100 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1101 Bytecodes::Code code = java_bc();
1102 if (code == Bytecodes::_wide) {
1103 code = method()->java_code_at_bci(bci() + 1);
1104 }
1105
1106 if (code != Bytecodes::_illegal) {
1107 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1108 }
1109
1110 auto rsize = [&]() {
1111 assert(code != Bytecodes::_illegal, "code is illegal!");
1112 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1113 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1114 };
1115
1116 switch (code) {
1117 case Bytecodes::_illegal:
1118 return false;
1119
1120 case Bytecodes::_ldc:
1121 case Bytecodes::_ldc_w:
1122 case Bytecodes::_ldc2_w:
1123 inputs = 0;
1124 break;
1125
1126 case Bytecodes::_dup: inputs = 1; break;
1127 case Bytecodes::_dup_x1: inputs = 2; break;
1128 case Bytecodes::_dup_x2: inputs = 3; break;
1129 case Bytecodes::_dup2: inputs = 2; break;
1130 case Bytecodes::_dup2_x1: inputs = 3; break;
1131 case Bytecodes::_dup2_x2: inputs = 4; break;
1132 case Bytecodes::_swap: inputs = 2; break;
1133 case Bytecodes::_arraylength: inputs = 1; break;
1134
1135 case Bytecodes::_getstatic:
1136 case Bytecodes::_putstatic:
1137 case Bytecodes::_getfield:
1138 case Bytecodes::_putfield:
1139 {
1140 bool ignored_will_link;
1141 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1142 int size = field->type()->size();
1143 bool is_get = (depth >= 0), is_static = (depth & 1);
1144 inputs = (is_static ? 0 : 1);
1145 if (is_get) {
1146 depth = size - inputs;
1147 } else {
1148 inputs += size; // putxxx pops the value from the stack
1149 depth = - inputs;
1150 }
1151 }
1152 break;
1153
1154 case Bytecodes::_invokevirtual:
1155 case Bytecodes::_invokespecial:
1156 case Bytecodes::_invokestatic:
1157 case Bytecodes::_invokedynamic:
1158 case Bytecodes::_invokeinterface:
1159 {
1160 bool ignored_will_link;
1161 ciSignature* declared_signature = nullptr;
1162 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1163 assert(declared_signature != nullptr, "cannot be null");
1164 inputs = declared_signature->arg_size_for_bc(code);
1165 int size = declared_signature->return_type()->size();
1166 depth = size - inputs;
1167 }
1168 break;
1169
1170 case Bytecodes::_multianewarray:
1171 {
1172 ciBytecodeStream iter(method());
1173 iter.reset_to_bci(bci());
1174 iter.next();
1175 inputs = iter.get_dimensions();
1176 assert(rsize() == 1, "");
1177 depth = 1 - inputs;
1178 }
1179 break;
1180
1181 case Bytecodes::_ireturn:
1182 case Bytecodes::_lreturn:
1183 case Bytecodes::_freturn:
1184 case Bytecodes::_dreturn:
1185 case Bytecodes::_areturn:
1186 assert(rsize() == -depth, "");
1187 inputs = -depth;
1188 break;
1189
1190 case Bytecodes::_jsr:
1191 case Bytecodes::_jsr_w:
1192 inputs = 0;
1193 depth = 1; // S.B. depth=1, not zero
1194 break;
1195
1196 default:
1197 // bytecode produces a typed result
1198 inputs = rsize() - depth;
1199 assert(inputs >= 0, "");
1200 break;
1201 }
1202
1203 #ifdef ASSERT
1204 // spot check
1205 int outputs = depth + inputs;
1206 assert(outputs >= 0, "sanity");
1207 switch (code) {
1208 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1209 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1210 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1211 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1212 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1213 default: break;
1214 }
1215 #endif //ASSERT
1216
1217 return true;
1218 }
1219
1220
1221
1222 //------------------------------basic_plus_adr---------------------------------
1223 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1224 // short-circuit a common case
1225 if (offset == intcon(0)) return ptr;
1226 return _gvn.transform( new AddPNode(base, ptr, offset) );
1227 }
1228
1229 Node* GraphKit::ConvI2L(Node* offset) {
1230 // short-circuit a common case
1231 jint offset_con = find_int_con(offset, Type::OffsetBot);
1232 if (offset_con != Type::OffsetBot) {
1233 return longcon((jlong) offset_con);
1234 }
1235 return _gvn.transform( new ConvI2LNode(offset));
1236 }
1237
1238 Node* GraphKit::ConvI2UL(Node* offset) {
1239 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1240 if (offset_con != (juint) Type::OffsetBot) {
1241 return longcon((julong) offset_con);
1242 }
1243 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1244 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1245 return _gvn.transform( new AndLNode(conv, mask) );
1246 }
1247
1248 Node* GraphKit::ConvL2I(Node* offset) {
1249 // short-circuit a common case
1250 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1251 if (offset_con != (jlong)Type::OffsetBot) {
1252 return intcon((int) offset_con);
1253 }
1254 return _gvn.transform( new ConvL2INode(offset));
1255 }
1256
1257 //-------------------------load_object_klass-----------------------------------
1258 Node* GraphKit::load_object_klass(Node* obj) {
1259 // Special-case a fresh allocation to avoid building nodes:
1260 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1261 if (akls != nullptr) return akls;
1262 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1263 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1264 }
1265
1266 //-------------------------load_array_length-----------------------------------
1267 Node* GraphKit::load_array_length(Node* array) {
1268 // Special-case a fresh allocation to avoid building nodes:
1269 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1270 Node *alen;
1271 if (alloc == nullptr) {
1272 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1273 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1274 } else {
1275 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1276 }
1277 return alen;
1278 }
1279
1280 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1281 const TypeOopPtr* oop_type,
1282 bool replace_length_in_map) {
1283 Node* length = alloc->Ideal_length();
1284 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1285 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1286 if (ccast != length) {
1287 // do not transform ccast here, it might convert to top node for
1288 // negative array length and break assumptions in parsing stage.
1289 _gvn.set_type_bottom(ccast);
1290 record_for_igvn(ccast);
1291 if (replace_length_in_map) {
1292 replace_in_map(length, ccast);
1293 }
1294 return ccast;
1295 }
1296 }
1297 return length;
1298 }
1299
1300 //------------------------------do_null_check----------------------------------
1301 // Helper function to do a null pointer check. Returned value is
1302 // the incoming address with null casted away. You are allowed to use the
1303 // not-null value only if you are control dependent on the test.
1304 #ifndef PRODUCT
1305 extern uint explicit_null_checks_inserted,
1306 explicit_null_checks_elided;
1307 #endif
1308 Node* GraphKit::null_check_common(Node* value, BasicType type,
1309 // optional arguments for variations:
1310 bool assert_null,
1311 Node* *null_control,
1312 bool speculative,
1313 bool null_marker_check) {
1314 assert(!assert_null || null_control == nullptr, "not both at once");
1315 if (stopped()) return top();
1316 NOT_PRODUCT(explicit_null_checks_inserted++);
1317
1318 if (value->is_InlineType()) {
1319 // Null checking a scalarized but nullable inline type. Check the null marker
1320 // input instead of the oop input to avoid keeping buffer allocations alive.
1321 InlineTypeNode* vtptr = value->as_InlineType();
1322 while (vtptr->get_oop()->is_InlineType()) {
1323 vtptr = vtptr->get_oop()->as_InlineType();
1324 }
1325 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1326 if (stopped()) {
1327 return top();
1328 }
1329 if (assert_null) {
1330 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1331 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1332 // replace_in_map(value, vtptr);
1333 // return vtptr;
1334 replace_in_map(value, null());
1335 return null();
1336 }
1337 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1338 return cast_not_null(value, do_replace_in_map);
1339 }
1340
1341 // Construct null check
1342 Node *chk = nullptr;
1343 switch(type) {
1344 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1345 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1346 case T_ARRAY : // fall through
1347 type = T_OBJECT; // simplify further tests
1348 case T_OBJECT : {
1349 const Type *t = _gvn.type( value );
1350
1351 const TypeOopPtr* tp = t->isa_oopptr();
1352 if (tp != nullptr && !tp->is_loaded()
1353 // Only for do_null_check, not any of its siblings:
1354 && !assert_null && null_control == nullptr) {
1355 // Usually, any field access or invocation on an unloaded oop type
1356 // will simply fail to link, since the statically linked class is
1357 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1358 // the static class is loaded but the sharper oop type is not.
1359 // Rather than checking for this obscure case in lots of places,
1360 // we simply observe that a null check on an unloaded class
1361 // will always be followed by a nonsense operation, so we
1362 // can just issue the uncommon trap here.
1363 // Our access to the unloaded class will only be correct
1364 // after it has been loaded and initialized, which requires
1365 // a trip through the interpreter.
1366 ciKlass* klass = tp->unloaded_klass();
1367 #ifndef PRODUCT
1368 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1369 #endif
1370 uncommon_trap(Deoptimization::Reason_unloaded,
1371 Deoptimization::Action_reinterpret,
1372 klass, "!loaded");
1373 return top();
1374 }
1375
1376 if (assert_null) {
1377 // See if the type is contained in NULL_PTR.
1378 // If so, then the value is already null.
1379 if (t->higher_equal(TypePtr::NULL_PTR)) {
1380 NOT_PRODUCT(explicit_null_checks_elided++);
1381 return value; // Elided null assert quickly!
1382 }
1383 } else {
1384 // See if mixing in the null pointer changes type.
1385 // If so, then the null pointer was not allowed in the original
1386 // type. In other words, "value" was not-null.
1387 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1388 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1389 NOT_PRODUCT(explicit_null_checks_elided++);
1390 return value; // Elided null check quickly!
1391 }
1392 }
1393 chk = new CmpPNode( value, null() );
1394 break;
1395 }
1396
1397 default:
1398 fatal("unexpected type: %s", type2name(type));
1399 }
1400 assert(chk != nullptr, "sanity check");
1401 chk = _gvn.transform(chk);
1402
1403 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1404 BoolNode *btst = new BoolNode( chk, btest);
1405 Node *tst = _gvn.transform( btst );
1406
1407 //-----------
1408 // if peephole optimizations occurred, a prior test existed.
1409 // If a prior test existed, maybe it dominates as we can avoid this test.
1410 if (tst != btst && type == T_OBJECT) {
1411 // At this point we want to scan up the CFG to see if we can
1412 // find an identical test (and so avoid this test altogether).
1413 Node *cfg = control();
1414 int depth = 0;
1415 while( depth < 16 ) { // Limit search depth for speed
1416 if( cfg->Opcode() == Op_IfTrue &&
1417 cfg->in(0)->in(1) == tst ) {
1418 // Found prior test. Use "cast_not_null" to construct an identical
1419 // CastPP (and hence hash to) as already exists for the prior test.
1420 // Return that casted value.
1421 if (assert_null) {
1422 replace_in_map(value, null());
1423 return null(); // do not issue the redundant test
1424 }
1425 Node *oldcontrol = control();
1426 set_control(cfg);
1427 Node *res = cast_not_null(value);
1428 set_control(oldcontrol);
1429 NOT_PRODUCT(explicit_null_checks_elided++);
1430 return res;
1431 }
1432 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1433 if (cfg == nullptr) break; // Quit at region nodes
1434 depth++;
1435 }
1436 }
1437
1438 //-----------
1439 // Branch to failure if null
1440 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1441 Deoptimization::DeoptReason reason;
1442 if (assert_null) {
1443 reason = Deoptimization::reason_null_assert(speculative);
1444 } else if (type == T_OBJECT || null_marker_check) {
1445 reason = Deoptimization::reason_null_check(speculative);
1446 } else {
1447 reason = Deoptimization::Reason_div0_check;
1448 }
1449 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1450 // ciMethodData::has_trap_at will return a conservative -1 if any
1451 // must-be-null assertion has failed. This could cause performance
1452 // problems for a method after its first do_null_assert failure.
1453 // Consider using 'Reason_class_check' instead?
1454
1455 // To cause an implicit null check, we set the not-null probability
1456 // to the maximum (PROB_MAX). For an explicit check the probability
1457 // is set to a smaller value.
1458 if (null_control != nullptr || too_many_traps(reason)) {
1459 // probability is less likely
1460 ok_prob = PROB_LIKELY_MAG(3);
1461 } else if (!assert_null &&
1462 (ImplicitNullCheckThreshold > 0) &&
1463 method() != nullptr &&
1464 (method()->method_data()->trap_count(reason)
1465 >= (uint)ImplicitNullCheckThreshold)) {
1466 ok_prob = PROB_LIKELY_MAG(3);
1467 }
1468
1469 if (null_control != nullptr) {
1470 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1471 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1472 set_control( _gvn.transform( new IfTrueNode(iff)));
1473 #ifndef PRODUCT
1474 if (null_true == top()) {
1475 explicit_null_checks_elided++;
1476 }
1477 #endif
1478 (*null_control) = null_true;
1479 } else {
1480 BuildCutout unless(this, tst, ok_prob);
1481 // Check for optimizer eliding test at parse time
1482 if (stopped()) {
1483 // Failure not possible; do not bother making uncommon trap.
1484 NOT_PRODUCT(explicit_null_checks_elided++);
1485 } else if (assert_null) {
1486 uncommon_trap(reason,
1487 Deoptimization::Action_make_not_entrant,
1488 nullptr, "assert_null");
1489 } else {
1490 replace_in_map(value, zerocon(type));
1491 builtin_throw(reason);
1492 }
1493 }
1494
1495 // Must throw exception, fall-thru not possible?
1496 if (stopped()) {
1497 return top(); // No result
1498 }
1499
1500 if (assert_null) {
1501 // Cast obj to null on this path.
1502 replace_in_map(value, zerocon(type));
1503 return zerocon(type);
1504 }
1505
1506 // Cast obj to not-null on this path, if there is no null_control.
1507 // (If there is a null_control, a non-null value may come back to haunt us.)
1508 if (type == T_OBJECT) {
1509 Node* cast = cast_not_null(value, false);
1510 if (null_control == nullptr || (*null_control) == top())
1511 replace_in_map(value, cast);
1512 value = cast;
1513 }
1514
1515 return value;
1516 }
1517
1518 //------------------------------cast_not_null----------------------------------
1519 // Cast obj to not-null on this path
1520 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1521 if (obj->is_InlineType()) {
1522 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1523 vt->as_InlineType()->set_null_marker(_gvn);
1524 vt = _gvn.transform(vt);
1525 if (do_replace_in_map) {
1526 replace_in_map(obj, vt);
1527 }
1528 return vt;
1529 }
1530 const Type *t = _gvn.type(obj);
1531 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1532 // Object is already not-null?
1533 if( t == t_not_null ) return obj;
1534
1535 Node* cast = new CastPPNode(control(), obj,t_not_null);
1536 cast = _gvn.transform( cast );
1537
1538 // Scan for instances of 'obj' in the current JVM mapping.
1539 // These instances are known to be not-null after the test.
1540 if (do_replace_in_map)
1541 replace_in_map(obj, cast);
1542
1543 return cast; // Return casted value
1544 }
1545
1546 // Sometimes in intrinsics, we implicitly know an object is not null
1547 // (there's no actual null check) so we can cast it to not null. In
1548 // the course of optimizations, the input to the cast can become null.
1549 // In that case that data path will die and we need the control path
1550 // to become dead as well to keep the graph consistent. So we have to
1551 // add a check for null for which one branch can't be taken. It uses
1552 // an OpaqueConstantBool node that will cause the check to be removed after loop
1553 // opts so the test goes away and the compiled code doesn't execute a
1554 // useless check.
1555 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1556 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1557 return value;
1558 }
1559 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1560 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1561 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1562 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1563 _gvn.set_type(iff, iff->Value(&_gvn));
1564 if (!tst->is_Con()) {
1565 record_for_igvn(iff);
1566 }
1567 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1568 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1569 halt(if_f, frame, "unexpected null in intrinsic");
1570 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1571 set_control(if_t);
1572 return cast_not_null(value, do_replace_in_map);
1573 }
1574
1575
1576 //--------------------------replace_in_map-------------------------------------
1577 void GraphKit::replace_in_map(Node* old, Node* neww) {
1578 if (old == neww) {
1579 return;
1580 }
1581
1582 map()->replace_edge(old, neww);
1583
1584 // Note: This operation potentially replaces any edge
1585 // on the map. This includes locals, stack, and monitors
1586 // of the current (innermost) JVM state.
1587
1588 // don't let inconsistent types from profiling escape this
1589 // method
1590
1591 const Type* told = _gvn.type(old);
1592 const Type* tnew = _gvn.type(neww);
1593
1594 if (!tnew->higher_equal(told)) {
1595 return;
1596 }
1597
1598 map()->record_replaced_node(old, neww);
1599 }
1600
1601
1602 //=============================================================================
1603 //--------------------------------memory---------------------------------------
1604 Node* GraphKit::memory(uint alias_idx) {
1605 MergeMemNode* mem = merged_memory();
1606 Node* p = mem->memory_at(alias_idx);
1607 assert(p != mem->empty_memory(), "empty");
1608 _gvn.set_type(p, Type::MEMORY); // must be mapped
1609 return p;
1610 }
1611
1612 //-----------------------------reset_memory------------------------------------
1613 Node* GraphKit::reset_memory() {
1614 Node* mem = map()->memory();
1615 // do not use this node for any more parsing!
1616 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1617 return _gvn.transform( mem );
1618 }
1619
1620 //------------------------------set_all_memory---------------------------------
1621 void GraphKit::set_all_memory(Node* newmem) {
1622 Node* mergemem = MergeMemNode::make(newmem);
1623 gvn().set_type_bottom(mergemem);
1624 if (_gvn.is_IterGVN() != nullptr) {
1625 record_for_igvn(mergemem);
1626 }
1627 map()->set_memory(mergemem);
1628 }
1629
1630 //------------------------------set_all_memory_call----------------------------
1631 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1632 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1633 set_all_memory(newmem);
1634 }
1635
1636 //=============================================================================
1637 //
1638 // parser factory methods for MemNodes
1639 //
1640 // These are layered on top of the factory methods in LoadNode and StoreNode,
1641 // and integrate with the parser's memory state and _gvn engine.
1642 //
1643
1644 // factory methods in "int adr_idx"
1645 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1646 MemNode::MemOrd mo,
1647 LoadNode::ControlDependency control_dependency,
1648 bool require_atomic_access,
1649 bool unaligned,
1650 bool mismatched,
1651 bool unsafe,
1652 uint8_t barrier_data) {
1653 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1654 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1655 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1656 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1657 Node* mem = memory(adr_idx);
1658 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1659 ld = _gvn.transform(ld);
1660
1661 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1662 // Improve graph before escape analysis and boxing elimination.
1663 record_for_igvn(ld);
1664 if (ld->is_DecodeN()) {
1665 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1666 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1667 // a Phi). Recording such cases is still perfectly sound, but may be
1668 // unnecessary and result in some minor IGVN overhead.
1669 record_for_igvn(ld->in(1));
1670 }
1671 }
1672 return ld;
1673 }
1674
1675 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1676 MemNode::MemOrd mo,
1677 bool require_atomic_access,
1678 bool unaligned,
1679 bool mismatched,
1680 bool unsafe,
1681 int barrier_data) {
1682 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1683 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1684 const TypePtr* adr_type = nullptr;
1685 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1686 Node *mem = memory(adr_idx);
1687 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1688 if (unaligned) {
1689 st->as_Store()->set_unaligned_access();
1690 }
1691 if (mismatched) {
1692 st->as_Store()->set_mismatched_access();
1693 }
1694 if (unsafe) {
1695 st->as_Store()->set_unsafe_access();
1696 }
1697 st->as_Store()->set_barrier_data(barrier_data);
1698 st = _gvn.transform(st);
1699 set_memory(st, adr_idx);
1700 // Back-to-back stores can only remove intermediate store with DU info
1701 // so push on worklist for optimizer.
1702 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1703 record_for_igvn(st);
1704
1705 return st;
1706 }
1707
1708 Node* GraphKit::access_store_at(Node* obj,
1709 Node* adr,
1710 const TypePtr* adr_type,
1711 Node* val,
1712 const Type* val_type,
1713 BasicType bt,
1714 DecoratorSet decorators,
1715 bool safe_for_replace,
1716 const InlineTypeNode* vt) {
1717 // Transformation of a value which could be null pointer (CastPP #null)
1718 // could be delayed during Parse (for example, in adjust_map_after_if()).
1719 // Execute transformation here to avoid barrier generation in such case.
1720 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1721 val = _gvn.makecon(TypePtr::NULL_PTR);
1722 }
1723
1724 if (stopped()) {
1725 return top(); // Dead path ?
1726 }
1727
1728 assert(val != nullptr, "not dead path");
1729 if (val->is_InlineType()) {
1730 // Store to non-flat field. Buffer the inline type and make sure
1731 // the store is re-executed if the allocation triggers deoptimization.
1732 PreserveReexecuteState preexecs(this);
1733 jvms()->set_should_reexecute(true);
1734 val = val->as_InlineType()->buffer(this, safe_for_replace);
1735 }
1736
1737 C2AccessValuePtr addr(adr, adr_type);
1738 C2AccessValue value(val, val_type);
1739 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1740 if (access.is_raw()) {
1741 return _barrier_set->BarrierSetC2::store_at(access, value);
1742 } else {
1743 return _barrier_set->store_at(access, value);
1744 }
1745 }
1746
1747 Node* GraphKit::access_load_at(Node* obj, // containing obj
1748 Node* adr, // actual address to store val at
1749 const TypePtr* adr_type,
1750 const Type* val_type,
1751 BasicType bt,
1752 DecoratorSet decorators,
1753 Node* ctl) {
1754 if (stopped()) {
1755 return top(); // Dead path ?
1756 }
1757
1758 C2AccessValuePtr addr(adr, adr_type);
1759 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1762 } else {
1763 return _barrier_set->load_at(access, val_type);
1764 }
1765 }
1766
1767 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1768 const Type* val_type,
1769 BasicType bt,
1770 DecoratorSet decorators) {
1771 if (stopped()) {
1772 return top(); // Dead path ?
1773 }
1774
1775 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1776 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1777 if (access.is_raw()) {
1778 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1779 } else {
1780 return _barrier_set->load_at(access, val_type);
1781 }
1782 }
1783
1784 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1785 Node* adr,
1786 const TypePtr* adr_type,
1787 int alias_idx,
1788 Node* expected_val,
1789 Node* new_val,
1790 const Type* value_type,
1791 BasicType bt,
1792 DecoratorSet decorators) {
1793 C2AccessValuePtr addr(adr, adr_type);
1794 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1795 bt, obj, addr, alias_idx);
1796 if (access.is_raw()) {
1797 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1798 } else {
1799 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1800 }
1801 }
1802
1803 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1804 Node* adr,
1805 const TypePtr* adr_type,
1806 int alias_idx,
1807 Node* expected_val,
1808 Node* new_val,
1809 const Type* value_type,
1810 BasicType bt,
1811 DecoratorSet decorators) {
1812 C2AccessValuePtr addr(adr, adr_type);
1813 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1814 bt, obj, addr, alias_idx);
1815 if (access.is_raw()) {
1816 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1817 } else {
1818 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1819 }
1820 }
1821
1822 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1823 Node* adr,
1824 const TypePtr* adr_type,
1825 int alias_idx,
1826 Node* new_val,
1827 const Type* value_type,
1828 BasicType bt,
1829 DecoratorSet decorators) {
1830 C2AccessValuePtr addr(adr, adr_type);
1831 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1832 bt, obj, addr, alias_idx);
1833 if (access.is_raw()) {
1834 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1835 } else {
1836 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1837 }
1838 }
1839
1840 Node* GraphKit::access_atomic_add_at(Node* obj,
1841 Node* adr,
1842 const TypePtr* adr_type,
1843 int alias_idx,
1844 Node* new_val,
1845 const Type* value_type,
1846 BasicType bt,
1847 DecoratorSet decorators) {
1848 C2AccessValuePtr addr(adr, adr_type);
1849 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1850 if (access.is_raw()) {
1851 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1852 } else {
1853 return _barrier_set->atomic_add_at(access, new_val, value_type);
1854 }
1855 }
1856
1857 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1858 return _barrier_set->clone(this, src, dst, size, is_array);
1859 }
1860
1861 //-------------------------array_element_address-------------------------
1862 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1863 const TypeInt* sizetype, Node* ctrl) {
1864 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1865 uint shift;
1866 uint header;
1867 if (arytype->is_flat() && arytype->klass_is_exact()) {
1868 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1869 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1870 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1871 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1872 // though we don't need the address node in this case and throw it away again.
1873 shift = arytype->flat_log_elem_size();
1874 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1875 } else {
1876 shift = exact_log2(type2aelembytes(elembt));
1877 header = arrayOopDesc::base_offset_in_bytes(elembt);
1878 }
1879
1880 // short-circuit a common case (saves lots of confusing waste motion)
1881 jint idx_con = find_int_con(idx, -1);
1882 if (idx_con >= 0) {
1883 intptr_t offset = header + ((intptr_t)idx_con << shift);
1884 return basic_plus_adr(ary, offset);
1885 }
1886
1887 // must be correct type for alignment purposes
1888 Node* base = basic_plus_adr(ary, header);
1889 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1890 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1891 return basic_plus_adr(ary, base, scale);
1892 }
1893
1894 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1895 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1896 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1897 return cast_to_flat_array_exact(array, elem_vk, true, false);
1898 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1899 return cast_to_flat_array_exact(array, elem_vk, true, true);
1900 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1901 return cast_to_flat_array_exact(array, elem_vk, false, true);
1902 }
1903
1904 bool is_null_free = false;
1905 if (!elem_vk->has_nullable_atomic_layout()) {
1906 // Element does not have a nullable flat layout, cannot be nullable
1907 is_null_free = true;
1908 }
1909
1910 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1911 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1912 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1913 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1914 }
1915
1916 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1917 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1918 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1919 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1920 assert(arytype->klass_is_exact(), "inconsistency");
1921 assert(arytype->is_flat(), "inconsistency");
1922 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1923 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1924 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1925 }
1926
1927 //-------------------------load_array_element-------------------------
1928 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1929 const Type* elemtype = arytype->elem();
1930 BasicType elembt = elemtype->array_element_basic_type();
1931 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1932 if (elembt == T_NARROWOOP) {
1933 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1934 }
1935 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1936 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1937 return ld;
1938 }
1939
1940 //-------------------------set_arguments_for_java_call-------------------------
1941 // Arguments (pre-popped from the stack) are taken from the JVMS.
1942 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1943 PreserveReexecuteState preexecs(this);
1944 if (Arguments::is_valhalla_enabled()) {
1945 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1946 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1947 jvms()->set_should_reexecute(true);
1948 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1949 inc_sp(arg_size);
1950 }
1951 // Add the call arguments
1952 const TypeTuple* domain = call->tf()->domain_sig();
1953 uint nargs = domain->cnt();
1954 int arg_num = 0;
1955 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1956 uint arg_idx = i - TypeFunc::Parms;
1957 Node* arg = argument(arg_idx);
1958 const Type* t = domain->field_at(i);
1959 // TODO 8284443 A static call to a mismatched method should still be scalarized
1960 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1961 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1962 if (!arg->is_InlineType()) {
1963 // There are 2 cases in which the argument has not been scalarized
1964 if (_gvn.type(arg)->is_zero_type()) {
1965 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
1966 } else {
1967 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
1968 // receiver is a non-scalarized oop. Later on, IGVN reveals that the receiver must be a
1969 // value object. The method is devirtualized, and replaced with a direct call with a
1970 // scalarized receiver instead.
1971 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
1972 assert(C->inlining_incrementally() || C->strength_reduction(), "must be during devirtualization of calls");
1973 assert(!is_Parse(), "must be during devirtualization of calls");
1974 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1975 }
1976 }
1977 InlineTypeNode* vt = arg->as_InlineType();
1978 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1979 // If an inline type argument is passed as fields, attach the Method* to the call site
1980 // to be able to access the extended signature later via attached_method_before_pc().
1981 // For example, see CompiledMethod::preserve_callee_argument_oops().
1982 call->set_override_symbolic_info(true);
1983 // Register an calling convention dependency on the callee method to make sure that this method is deoptimized and
1984 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1985 C->dependencies()->assert_mismatch_calling_convention(call->method());
1986 arg_num++;
1987 continue;
1988 } else if (arg->is_InlineType()) {
1989 // Pass inline type argument via oop to callee
1990 arg = arg->as_InlineType()->buffer(this, true);
1991 }
1992 if (t != Type::HALF) {
1993 arg_num++;
1994 }
1995 call->init_req(idx++, arg);
1996 }
1997 }
1998
1999 //---------------------------set_edges_for_java_call---------------------------
2000 // Connect a newly created call into the current JVMS.
2001 // A return value node (if any) is returned from set_edges_for_java_call.
2002 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2003
2004 // Add the predefined inputs:
2005 call->init_req( TypeFunc::Control, control() );
2006 call->init_req( TypeFunc::I_O , i_o() );
2007 call->init_req( TypeFunc::Memory , reset_memory() );
2008 call->init_req( TypeFunc::FramePtr, frameptr() );
2009 call->init_req( TypeFunc::ReturnAdr, top() );
2010
2011 add_safepoint_edges(call, must_throw);
2012
2013 Node* xcall = _gvn.transform(call);
2014
2015 if (xcall == top()) {
2016 set_control(top());
2017 return;
2018 }
2019 assert(xcall == call, "call identity is stable");
2020
2021 // Re-use the current map to produce the result.
2022
2023 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2024 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2025 set_all_memory_call(xcall, separate_io_proj);
2026
2027 //return xcall; // no need, caller already has it
2028 }
2029
2030 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2031 if (stopped()) return top(); // maybe the call folded up?
2032
2033 // Note: Since any out-of-line call can produce an exception,
2034 // we always insert an I_O projection from the call into the result.
2035
2036 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2037
2038 if (separate_io_proj) {
2039 // The caller requested separate projections be used by the fall
2040 // through and exceptional paths, so replace the projections for
2041 // the fall through path.
2042 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2043 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2044 }
2045
2046 // Capture the return value, if any.
2047 Node* ret;
2048 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2049 ret = top();
2050 } else if (call->tf()->returns_inline_type_as_fields()) {
2051 // Return of multiple values (inline type fields): we create a
2052 // InlineType node, each field is a projection from the call.
2053 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2054 uint base_input = TypeFunc::Parms;
2055 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2056 } else {
2057 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2058 ciType* t = call->method()->return_type();
2059 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2060 // The return type is unloaded but the callee might later be C2 compiled and then return
2061 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2062 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2063 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2064 IdealKit ideal(this);
2065 IdealVariable res(ideal);
2066 ideal.declarations_done();
2067 // Change return type of call to scalarized return
2068 const TypeFunc* tf = call->_tf;
2069 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2070 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2071 call->_tf = new_tf;
2072 _gvn.set_type(call, call->Value(&_gvn));
2073 _gvn.set_type(ret, ret->Value(&_gvn));
2074 // Don't add store to buffer call if we are strength reducing
2075 if (!C->strength_reduction()) {
2076 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2077 // Return value is null
2078 ideal.set(res, makecon(TypePtr::NULL_PTR));
2079 } ideal.else_(); {
2080 // Return value is non-null
2081 sync_kit(ideal);
2082
2083 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2084 OptoRuntime::store_inline_type_fields_Type(),
2085 StubRoutines::store_inline_type_fields_to_buf(),
2086 nullptr, TypePtr::BOTTOM, ret);
2087
2088 // We don't know how many values are returned. This assumes the
2089 // worst case, that all available registers are used.
2090 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2091 if (domain->field_at(i) == Type::HALF) {
2092 store_to_buf_call->init_req(i, top());
2093 continue;
2094 }
2095 Node* proj =_gvn.transform(new ProjNode(call, i));
2096 store_to_buf_call->init_req(i, proj);
2097 }
2098 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2099
2100 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2101 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2102 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2103
2104 ideal.set(res, buf);
2105 ideal.sync_kit(this);
2106 } ideal.end_if();
2107 } else {
2108 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2109 Node* proj =_gvn.transform(new ProjNode(call, i));
2110 }
2111 ideal.set(res, ret);
2112 }
2113 sync_kit(ideal);
2114 ret = _gvn.transform(ideal.value(res));
2115 }
2116 if (t->is_klass()) {
2117 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2118 if (type->is_inlinetypeptr()) {
2119 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2120 }
2121 }
2122 }
2123
2124 return ret;
2125 }
2126
2127 //--------------------set_predefined_input_for_runtime_call--------------------
2128 // Reading and setting the memory state is way conservative here.
2129 // The real problem is that I am not doing real Type analysis on memory,
2130 // so I cannot distinguish card mark stores from other stores. Across a GC
2131 // point the Store Barrier and the card mark memory has to agree. I cannot
2132 // have a card mark store and its barrier split across the GC point from
2133 // either above or below. Here I get that to happen by reading ALL of memory.
2134 // A better answer would be to separate out card marks from other memory.
2135 // For now, return the input memory state, so that it can be reused
2136 // after the call, if this call has restricted memory effects.
2137 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2138 // Set fixed predefined input arguments
2139 call->init_req(TypeFunc::Control, control());
2140 call->init_req(TypeFunc::I_O, top()); // does no i/o
2141 call->init_req(TypeFunc::ReturnAdr, top());
2142 if (call->is_CallLeafPure()) {
2143 call->init_req(TypeFunc::Memory, top());
2144 call->init_req(TypeFunc::FramePtr, top());
2145 return nullptr;
2146 } else {
2147 Node* memory = reset_memory();
2148 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2149 call->init_req(TypeFunc::Memory, m); // may gc ptrs
2150 call->init_req(TypeFunc::FramePtr, frameptr());
2151 return memory;
2152 }
2153 }
2154
2155 //-------------------set_predefined_output_for_runtime_call--------------------
2156 // Set control and memory (not i_o) from the call.
2157 // If keep_mem is not null, use it for the output state,
2158 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2159 // If hook_mem is null, this call produces no memory effects at all.
2160 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2161 // then only that memory slice is taken from the call.
2162 // In the last case, we must put an appropriate memory barrier before
2163 // the call, so as to create the correct anti-dependencies on loads
2164 // preceding the call.
2165 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2166 Node* keep_mem,
2167 const TypePtr* hook_mem) {
2168 // no i/o
2169 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2170 if (call->is_CallLeafPure()) {
2171 // Pure function have only control (for now) and data output, in particular
2172 // they don't touch the memory, so we don't want a memory proj that is set after.
2173 return;
2174 }
2175 if (keep_mem) {
2176 // First clone the existing memory state
2177 set_all_memory(keep_mem);
2178 if (hook_mem != nullptr) {
2179 // Make memory for the call
2180 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2181 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2182 // We also use hook_mem to extract specific effects from arraycopy stubs.
2183 set_memory(mem, hook_mem);
2184 }
2185 // ...else the call has NO memory effects.
2186
2187 // Make sure the call advertises its memory effects precisely.
2188 // This lets us build accurate anti-dependences in gcm.cpp.
2189 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2190 "call node must be constructed correctly");
2191 } else {
2192 assert(hook_mem == nullptr, "");
2193 // This is not a "slow path" call; all memory comes from the call.
2194 set_all_memory_call(call);
2195 }
2196 }
2197
2198 // Keep track of MergeMems feeding into other MergeMems
2199 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2200 if (!mem->is_MergeMem()) {
2201 return;
2202 }
2203 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2204 Node* use = i.get();
2205 if (use->is_MergeMem()) {
2206 wl.push(use);
2207 }
2208 }
2209 }
2210
2211 // Replace the call with the current state of the kit.
2212 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2213 JVMState* ejvms = nullptr;
2214 if (has_exceptions()) {
2215 ejvms = transfer_exceptions_into_jvms();
2216 }
2217
2218 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2219 ReplacedNodes replaced_nodes_exception;
2220 Node* ex_ctl = top();
2221
2222 SafePointNode* final_state = stop();
2223
2224 // Find all the needed outputs of this call
2225 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2226
2227 Unique_Node_List wl;
2228 Node* init_mem = call->in(TypeFunc::Memory);
2229 Node* final_mem = final_state->in(TypeFunc::Memory);
2230 Node* final_ctl = final_state->in(TypeFunc::Control);
2231 Node* final_io = final_state->in(TypeFunc::I_O);
2232
2233 // Replace all the old call edges with the edges from the inlining result
2234 if (callprojs->fallthrough_catchproj != nullptr) {
2235 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2236 }
2237 if (callprojs->fallthrough_memproj != nullptr) {
2238 if (final_mem->is_MergeMem()) {
2239 // Parser's exits MergeMem was not transformed but may be optimized
2240 final_mem = _gvn.transform(final_mem);
2241 }
2242 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2243 add_mergemem_users_to_worklist(wl, final_mem);
2244 }
2245 if (callprojs->fallthrough_ioproj != nullptr) {
2246 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2247 }
2248
2249 // Replace the result with the new result if it exists and is used
2250 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2251 // If the inlined code is dead, the result projections for an inline type returned as
2252 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2253 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2254 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2255 "unexpected number of results");
2256 // If we are doing strength reduction and the return type is not loaded we
2257 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2258 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2259 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2260 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2261 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2262 }
2263 } else {
2264 C->gvn_replace_by(callprojs->resproj[0], result);
2265 }
2266 }
2267
2268 if (ejvms == nullptr) {
2269 // No exception edges to simply kill off those paths
2270 if (callprojs->catchall_catchproj != nullptr) {
2271 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2272 }
2273 if (callprojs->catchall_memproj != nullptr) {
2274 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2275 }
2276 if (callprojs->catchall_ioproj != nullptr) {
2277 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2278 }
2279 // Replace the old exception object with top
2280 if (callprojs->exobj != nullptr) {
2281 C->gvn_replace_by(callprojs->exobj, C->top());
2282 }
2283 } else {
2284 GraphKit ekit(ejvms);
2285
2286 // Load my combined exception state into the kit, with all phis transformed:
2287 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2288 replaced_nodes_exception = ex_map->replaced_nodes();
2289
2290 Node* ex_oop = ekit.use_exception_state(ex_map);
2291
2292 if (callprojs->catchall_catchproj != nullptr) {
2293 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2294 ex_ctl = ekit.control();
2295 }
2296 if (callprojs->catchall_memproj != nullptr) {
2297 Node* ex_mem = ekit.reset_memory();
2298 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2299 add_mergemem_users_to_worklist(wl, ex_mem);
2300 }
2301 if (callprojs->catchall_ioproj != nullptr) {
2302 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2303 }
2304
2305 // Replace the old exception object with the newly created one
2306 if (callprojs->exobj != nullptr) {
2307 C->gvn_replace_by(callprojs->exobj, ex_oop);
2308 }
2309 }
2310
2311 // Disconnect the call from the graph
2312 call->disconnect_inputs(C);
2313 C->gvn_replace_by(call, C->top());
2314
2315 // Clean up any MergeMems that feed other MergeMems since the
2316 // optimizer doesn't like that.
2317 while (wl.size() > 0) {
2318 _gvn.transform(wl.pop());
2319 }
2320
2321 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2322 replaced_nodes.apply(C, final_ctl);
2323 }
2324 if (!ex_ctl->is_top() && do_replaced_nodes) {
2325 replaced_nodes_exception.apply(C, ex_ctl);
2326 }
2327 }
2328
2329
2330 //------------------------------increment_counter------------------------------
2331 // for statistics: increment a VM counter by 1
2332
2333 void GraphKit::increment_counter(address counter_addr) {
2334 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2335 increment_counter(adr1);
2336 }
2337
2338 void GraphKit::increment_counter(Node* counter_addr) {
2339 Node* ctrl = control();
2340 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2341 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2342 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2343 }
2344
2345 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2346 Node* halt = new HaltNode(ctrl, frameptr, reason
2347 PRODUCT_ONLY(COMMA generate_code_in_product));
2348 halt = _gvn.transform(halt);
2349 root()->add_req(halt);
2350 if (_gvn.is_IterGVN() != nullptr) {
2351 record_for_igvn(root());
2352 }
2353 }
2354
2355 //------------------------------uncommon_trap----------------------------------
2356 // Bail out to the interpreter in mid-method. Implemented by calling the
2357 // uncommon_trap blob. This helper function inserts a runtime call with the
2358 // right debug info.
2359 Node* GraphKit::uncommon_trap(int trap_request,
2360 ciKlass* klass, const char* comment,
2361 bool must_throw,
2362 bool keep_exact_action) {
2363 if (failing_internal()) {
2364 stop();
2365 }
2366 if (stopped()) return nullptr; // trap reachable?
2367
2368 // Note: If ProfileTraps is true, and if a deopt. actually
2369 // occurs here, the runtime will make sure an MDO exists. There is
2370 // no need to call method()->ensure_method_data() at this point.
2371
2372 // Set the stack pointer to the right value for reexecution:
2373 set_sp(reexecute_sp());
2374
2375 #ifdef ASSERT
2376 if (!must_throw) {
2377 // Make sure the stack has at least enough depth to execute
2378 // the current bytecode.
2379 int inputs, ignored_depth;
2380 if (compute_stack_effects(inputs, ignored_depth)) {
2381 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2382 Bytecodes::name(java_bc()), sp(), inputs);
2383 }
2384 }
2385 #endif
2386
2387 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2388 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2389
2390 switch (action) {
2391 case Deoptimization::Action_maybe_recompile:
2392 case Deoptimization::Action_reinterpret:
2393 // Temporary fix for 6529811 to allow virtual calls to be sure they
2394 // get the chance to go from mono->bi->mega
2395 if (!keep_exact_action &&
2396 Deoptimization::trap_request_index(trap_request) < 0 &&
2397 too_many_recompiles(reason)) {
2398 // This BCI is causing too many recompilations.
2399 if (C->log() != nullptr) {
2400 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2401 Deoptimization::trap_reason_name(reason),
2402 Deoptimization::trap_action_name(action));
2403 }
2404 action = Deoptimization::Action_none;
2405 trap_request = Deoptimization::make_trap_request(reason, action);
2406 } else {
2407 C->set_trap_can_recompile(true);
2408 }
2409 break;
2410 case Deoptimization::Action_make_not_entrant:
2411 C->set_trap_can_recompile(true);
2412 break;
2413 case Deoptimization::Action_none:
2414 case Deoptimization::Action_make_not_compilable:
2415 break;
2416 default:
2417 #ifdef ASSERT
2418 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2419 #endif
2420 break;
2421 }
2422
2423 if (TraceOptoParse) {
2424 char buf[100];
2425 tty->print_cr("Uncommon trap %s at bci:%d",
2426 Deoptimization::format_trap_request(buf, sizeof(buf),
2427 trap_request), bci());
2428 }
2429
2430 CompileLog* log = C->log();
2431 if (log != nullptr) {
2432 int kid = (klass == nullptr)? -1: log->identify(klass);
2433 log->begin_elem("uncommon_trap bci='%d'", bci());
2434 char buf[100];
2435 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2436 trap_request));
2437 if (kid >= 0) log->print(" klass='%d'", kid);
2438 if (comment != nullptr) log->print(" comment='%s'", comment);
2439 log->end_elem();
2440 }
2441
2442 // Make sure any guarding test views this path as very unlikely
2443 Node *i0 = control()->in(0);
2444 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2445 IfNode *iff = i0->as_If();
2446 float f = iff->_prob; // Get prob
2447 if (control()->Opcode() == Op_IfTrue) {
2448 if (f > PROB_UNLIKELY_MAG(4))
2449 iff->_prob = PROB_MIN;
2450 } else {
2451 if (f < PROB_LIKELY_MAG(4))
2452 iff->_prob = PROB_MAX;
2453 }
2454 }
2455
2456 // Clear out dead values from the debug info.
2457 kill_dead_locals();
2458
2459 // Now insert the uncommon trap subroutine call
2460 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2461 const TypePtr* no_memory_effects = nullptr;
2462 // Pass the index of the class to be loaded
2463 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2464 (must_throw ? RC_MUST_THROW : 0),
2465 OptoRuntime::uncommon_trap_Type(),
2466 call_addr, "uncommon_trap", no_memory_effects,
2467 intcon(trap_request));
2468 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2469 "must extract request correctly from the graph");
2470 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2471
2472 call->set_req(TypeFunc::ReturnAdr, returnadr());
2473 // The debug info is the only real input to this call.
2474
2475 // Halt-and-catch fire here. The above call should never return!
2476 // We only emit code for the HaltNode in debug, which is enough for
2477 // verifying correctness. In product, we don't want to emit it so
2478 // that we can save on code space. HaltNode often get folded because
2479 // the compiler can prove that the unreachable path is dead. But we
2480 // cannot generally expect that for uncommon traps, which are often
2481 // reachable and occasionally taken.
2482 halt(control(), frameptr(),
2483 "uncommon trap returned which should never happen",
2484 false /* don't emit code in product */);
2485 stop_and_kill_map();
2486 return call;
2487 }
2488
2489
2490 //--------------------------just_allocated_object------------------------------
2491 // Report the object that was just allocated.
2492 // It must be the case that there are no intervening safepoints.
2493 // We use this to determine if an object is so "fresh" that
2494 // it does not require card marks.
2495 Node* GraphKit::just_allocated_object(Node* current_control) {
2496 Node* ctrl = current_control;
2497 // Object::<init> is invoked after allocation, most of invoke nodes
2498 // will be reduced, but a region node is kept in parse time, we check
2499 // the pattern and skip the region node if it degraded to a copy.
2500 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2501 ctrl->as_Region()->is_copy()) {
2502 ctrl = ctrl->as_Region()->is_copy();
2503 }
2504 if (C->recent_alloc_ctl() == ctrl) {
2505 return C->recent_alloc_obj();
2506 }
2507 return nullptr;
2508 }
2509
2510
2511 /**
2512 * Record profiling data exact_kls for Node n with the type system so
2513 * that it can propagate it (speculation)
2514 *
2515 * @param n node that the type applies to
2516 * @param exact_kls type from profiling
2517 * @param maybe_null did profiling see null?
2518 *
2519 * @return node with improved type
2520 */
2521 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2522 const Type* current_type = _gvn.type(n);
2523 assert(UseTypeSpeculation, "type speculation must be on");
2524
2525 const TypePtr* speculative = current_type->speculative();
2526
2527 // Should the klass from the profile be recorded in the speculative type?
2528 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2529 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2530 const TypeOopPtr* xtype = tklass->as_instance_type();
2531 assert(xtype->klass_is_exact(), "Should be exact");
2532 // Any reason to believe n is not null (from this profiling or a previous one)?
2533 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2534 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2535 // record the new speculative type's depth
2536 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2537 speculative = speculative->with_inline_depth(jvms()->depth());
2538 } else if (current_type->would_improve_ptr(ptr_kind)) {
2539 // Profiling report that null was never seen so we can change the
2540 // speculative type to non null ptr.
2541 if (ptr_kind == ProfileAlwaysNull) {
2542 speculative = TypePtr::NULL_PTR;
2543 } else {
2544 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2545 const TypePtr* ptr = TypePtr::NOTNULL;
2546 if (speculative != nullptr) {
2547 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2548 } else {
2549 speculative = ptr;
2550 }
2551 }
2552 }
2553
2554 if (speculative != current_type->speculative()) {
2555 // Build a type with a speculative type (what we think we know
2556 // about the type but will need a guard when we use it)
2557 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2558 // We're changing the type, we need a new CheckCast node to carry
2559 // the new type. The new type depends on the control: what
2560 // profiling tells us is only valid from here as far as we can
2561 // tell.
2562 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2563 cast = _gvn.transform(cast);
2564 replace_in_map(n, cast);
2565 n = cast;
2566 }
2567
2568 return n;
2569 }
2570
2571 /**
2572 * Record profiling data from receiver profiling at an invoke with the
2573 * type system so that it can propagate it (speculation)
2574 *
2575 * @param n receiver node
2576 *
2577 * @return node with improved type
2578 */
2579 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2580 if (!UseTypeSpeculation) {
2581 return n;
2582 }
2583 ciKlass* exact_kls = profile_has_unique_klass();
2584 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2585 if ((java_bc() == Bytecodes::_checkcast ||
2586 java_bc() == Bytecodes::_instanceof ||
2587 java_bc() == Bytecodes::_aastore) &&
2588 method()->method_data()->is_mature()) {
2589 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2590 if (data != nullptr) {
2591 if (java_bc() == Bytecodes::_aastore) {
2592 ciKlass* array_type = nullptr;
2593 ciKlass* element_type = nullptr;
2594 ProfilePtrKind element_ptr = ProfileMaybeNull;
2595 bool flat_array = true;
2596 bool null_free_array = true;
2597 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2598 exact_kls = element_type;
2599 ptr_kind = element_ptr;
2600 } else {
2601 if (!data->as_BitData()->null_seen()) {
2602 ptr_kind = ProfileNeverNull;
2603 } else {
2604 if (TypeProfileCasts) {
2605 assert(data->is_ReceiverTypeData(), "bad profile data type");
2606 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2607 uint i = 0;
2608 for (; i < call->row_limit(); i++) {
2609 ciKlass* receiver = call->receiver(i);
2610 if (receiver != nullptr) {
2611 break;
2612 }
2613 }
2614 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2615 }
2616 }
2617 }
2618 }
2619 }
2620 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2621 }
2622
2623 /**
2624 * Record profiling data from argument profiling at an invoke with the
2625 * type system so that it can propagate it (speculation)
2626 *
2627 * @param dest_method target method for the call
2628 * @param bc what invoke bytecode is this?
2629 */
2630 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2631 if (!UseTypeSpeculation) {
2632 return;
2633 }
2634 const TypeFunc* tf = TypeFunc::make(dest_method);
2635 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2636 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2637 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2638 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2639 if (is_reference_type(targ->basic_type())) {
2640 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2641 ciKlass* better_type = nullptr;
2642 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2643 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2644 }
2645 i++;
2646 }
2647 }
2648 }
2649
2650 /**
2651 * Record profiling data from parameter profiling at an invoke with
2652 * the type system so that it can propagate it (speculation)
2653 */
2654 void GraphKit::record_profiled_parameters_for_speculation() {
2655 if (!UseTypeSpeculation) {
2656 return;
2657 }
2658 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2659 if (_gvn.type(local(i))->isa_oopptr()) {
2660 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2661 ciKlass* better_type = nullptr;
2662 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2663 record_profile_for_speculation(local(i), better_type, ptr_kind);
2664 }
2665 j++;
2666 }
2667 }
2668 }
2669
2670 /**
2671 * Record profiling data from return value profiling at an invoke with
2672 * the type system so that it can propagate it (speculation)
2673 */
2674 void GraphKit::record_profiled_return_for_speculation() {
2675 if (!UseTypeSpeculation) {
2676 return;
2677 }
2678 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2679 ciKlass* better_type = nullptr;
2680 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2681 // If profiling reports a single type for the return value,
2682 // feed it to the type system so it can propagate it as a
2683 // speculative type
2684 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2685 }
2686 }
2687
2688
2689 //=============================================================================
2690 // Generate a fast path/slow path idiom. Graph looks like:
2691 // [foo] indicates that 'foo' is a parameter
2692 //
2693 // [in] null
2694 // \ /
2695 // CmpP
2696 // Bool ne
2697 // If
2698 // / \
2699 // True False-<2>
2700 // / |
2701 // / cast_not_null
2702 // Load | | ^
2703 // [fast_test] | |
2704 // gvn to opt_test | |
2705 // / \ | <1>
2706 // True False |
2707 // | \\ |
2708 // [slow_call] \[fast_result]
2709 // Ctl Val \ \
2710 // | \ \
2711 // Catch <1> \ \
2712 // / \ ^ \ \
2713 // Ex No_Ex | \ \
2714 // | \ \ | \ <2> \
2715 // ... \ [slow_res] | | \ [null_result]
2716 // \ \--+--+--- | |
2717 // \ | / \ | /
2718 // --------Region Phi
2719 //
2720 //=============================================================================
2721 // Code is structured as a series of driver functions all called 'do_XXX' that
2722 // call a set of helper functions. Helper functions first, then drivers.
2723
2724 //------------------------------null_check_oop---------------------------------
2725 // Null check oop. Set null-path control into Region in slot 3.
2726 // Make a cast-not-nullness use the other not-null control. Return cast.
2727 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2728 bool never_see_null,
2729 bool safe_for_replace,
2730 bool speculative) {
2731 // Initial null check taken path
2732 (*null_control) = top();
2733 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2734
2735 // Generate uncommon_trap:
2736 if (never_see_null && (*null_control) != top()) {
2737 // If we see an unexpected null at a check-cast we record it and force a
2738 // recompile; the offending check-cast will be compiled to handle nulls.
2739 // If we see more than one offending BCI, then all checkcasts in the
2740 // method will be compiled to handle nulls.
2741 PreserveJVMState pjvms(this);
2742 set_control(*null_control);
2743 replace_in_map(value, null());
2744 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2745 uncommon_trap(reason,
2746 Deoptimization::Action_make_not_entrant);
2747 (*null_control) = top(); // null path is dead
2748 }
2749 if ((*null_control) == top() && safe_for_replace) {
2750 replace_in_map(value, cast);
2751 }
2752
2753 // Cast away null-ness on the result
2754 return cast;
2755 }
2756
2757 //------------------------------opt_iff----------------------------------------
2758 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2759 // Return slow-path control.
2760 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2761 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2762
2763 // Fast path taken; set region slot 2
2764 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2765 region->init_req(2,fast_taken); // Capture fast-control
2766
2767 // Fast path not-taken, i.e. slow path
2768 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2769 return slow_taken;
2770 }
2771
2772 //-----------------------------make_runtime_call-------------------------------
2773 Node* GraphKit::make_runtime_call(int flags,
2774 const TypeFunc* call_type, address call_addr,
2775 const char* call_name,
2776 const TypePtr* adr_type,
2777 // The following parms are all optional.
2778 // The first null ends the list.
2779 Node* parm0, Node* parm1,
2780 Node* parm2, Node* parm3,
2781 Node* parm4, Node* parm5,
2782 Node* parm6, Node* parm7) {
2783 assert(call_addr != nullptr, "must not call null targets");
2784
2785 // Slow-path call
2786 bool is_leaf = !(flags & RC_NO_LEAF);
2787 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2788 if (call_name == nullptr) {
2789 assert(!is_leaf, "must supply name for leaf");
2790 call_name = OptoRuntime::stub_name(call_addr);
2791 }
2792 CallNode* call;
2793 if (!is_leaf) {
2794 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2795 } else if (flags & RC_NO_FP) {
2796 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2797 } else if (flags & RC_VECTOR){
2798 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2799 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2800 } else if (flags & RC_PURE) {
2801 assert(adr_type == nullptr, "pure call does not touch memory");
2802 call = new CallLeafPureNode(call_type, call_addr, call_name);
2803 } else {
2804 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2805 }
2806
2807 // The following is similar to set_edges_for_java_call,
2808 // except that the memory effects of the call are restricted to AliasIdxRaw.
2809
2810 // Slow path call has no side-effects, uses few values
2811 bool wide_in = !(flags & RC_NARROW_MEM);
2812 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2813
2814 Node* prev_mem = nullptr;
2815 if (wide_in) {
2816 prev_mem = set_predefined_input_for_runtime_call(call);
2817 } else {
2818 assert(!wide_out, "narrow in => narrow out");
2819 Node* narrow_mem = memory(adr_type);
2820 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2821 }
2822
2823 // Hook each parm in order. Stop looking at the first null.
2824 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2825 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2826 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2827 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2828 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2829 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2830 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2831 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2832 /* close each nested if ===> */ } } } } } } } }
2833 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2834
2835 if (!is_leaf) {
2836 // Non-leaves can block and take safepoints:
2837 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2838 }
2839 // Non-leaves can throw exceptions:
2840 if (has_io) {
2841 call->set_req(TypeFunc::I_O, i_o());
2842 }
2843
2844 if (flags & RC_UNCOMMON) {
2845 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2846 // (An "if" probability corresponds roughly to an unconditional count.
2847 // Sort of.)
2848 call->set_cnt(PROB_UNLIKELY_MAG(4));
2849 }
2850
2851 Node* c = _gvn.transform(call);
2852 assert(c == call, "cannot disappear");
2853
2854 if (wide_out) {
2855 // Slow path call has full side-effects.
2856 set_predefined_output_for_runtime_call(call);
2857 } else {
2858 // Slow path call has few side-effects, and/or sets few values.
2859 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2860 }
2861
2862 if (has_io) {
2863 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2864 }
2865 return call;
2866
2867 }
2868
2869 // i2b
2870 Node* GraphKit::sign_extend_byte(Node* in) {
2871 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2872 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2873 }
2874
2875 // i2s
2876 Node* GraphKit::sign_extend_short(Node* in) {
2877 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2878 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2879 }
2880
2881
2882 //------------------------------merge_memory-----------------------------------
2883 // Merge memory from one path into the current memory state.
2884 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2885 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2886 Node* old_slice = mms.force_memory();
2887 Node* new_slice = mms.memory2();
2888 if (old_slice != new_slice) {
2889 PhiNode* phi;
2890 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2891 if (mms.is_empty()) {
2892 // clone base memory Phi's inputs for this memory slice
2893 assert(old_slice == mms.base_memory(), "sanity");
2894 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2895 _gvn.set_type(phi, Type::MEMORY);
2896 for (uint i = 1; i < phi->req(); i++) {
2897 phi->init_req(i, old_slice->in(i));
2898 }
2899 } else {
2900 phi = old_slice->as_Phi(); // Phi was generated already
2901 }
2902 } else {
2903 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2904 _gvn.set_type(phi, Type::MEMORY);
2905 }
2906 phi->set_req(new_path, new_slice);
2907 mms.set_memory(phi);
2908 }
2909 }
2910 }
2911
2912 //------------------------------make_slow_call_ex------------------------------
2913 // Make the exception handler hookups for the slow call
2914 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2915 if (stopped()) return;
2916
2917 // Make a catch node with just two handlers: fall-through and catch-all
2918 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2919 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2920 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2921 _gvn.set_type_bottom(norm);
2922 C->record_for_igvn(norm);
2923 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2924
2925 { PreserveJVMState pjvms(this);
2926 set_control(excp);
2927 set_i_o(i_o);
2928
2929 if (excp != top()) {
2930 if (deoptimize) {
2931 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2932 uncommon_trap(Deoptimization::Reason_unhandled,
2933 Deoptimization::Action_none);
2934 } else {
2935 // Create an exception state also.
2936 // Use an exact type if the caller has a specific exception.
2937 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2938 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2939 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2940 }
2941 }
2942 }
2943
2944 // Get the no-exception control from the CatchNode.
2945 set_control(norm);
2946 }
2947
2948 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2949 Node* cmp = nullptr;
2950 switch(bt) {
2951 case T_INT: cmp = new CmpINode(in1, in2); break;
2952 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2953 default: fatal("unexpected comparison type %s", type2name(bt));
2954 }
2955 cmp = gvn.transform(cmp);
2956 Node* bol = gvn.transform(new BoolNode(cmp, test));
2957 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2958 gvn.transform(iff);
2959 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2960 return iff;
2961 }
2962
2963 //-------------------------------gen_subtype_check-----------------------------
2964 // Generate a subtyping check. Takes as input the subtype and supertype.
2965 // Returns 2 values: sets the default control() to the true path and returns
2966 // the false path. Only reads invariant memory; sets no (visible) memory.
2967 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2968 // but that's not exposed to the optimizer. This call also doesn't take in an
2969 // Object; if you wish to check an Object you need to load the Object's class
2970 // prior to coming here.
2971 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2972 ciMethod* method, int bci) {
2973 Compile* C = gvn.C;
2974 if ((*ctrl)->is_top()) {
2975 return C->top();
2976 }
2977
2978 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2979 // For a direct pointer comparison, we need the refined array klass pointer
2980 Node* vm_superklass = superklass;
2981 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
2982 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
2983 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
2984 }
2985
2986 // Fast check for identical types, perhaps identical constants.
2987 // The types can even be identical non-constants, in cases
2988 // involving Array.newInstance, Object.clone, etc.
2989 if (subklass == superklass)
2990 return C->top(); // false path is dead; no test needed.
2991
2992 if (gvn.type(superklass)->singleton()) {
2993 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2994 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2995
2996 // In the common case of an exact superklass, try to fold up the
2997 // test before generating code. You may ask, why not just generate
2998 // the code and then let it fold up? The answer is that the generated
2999 // code will necessarily include null checks, which do not always
3000 // completely fold away. If they are also needless, then they turn
3001 // into a performance loss. Example:
3002 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3003 // Here, the type of 'fa' is often exact, so the store check
3004 // of fa[1]=x will fold up, without testing the nullness of x.
3005 //
3006 // At macro expansion, we would have already folded the SubTypeCheckNode
3007 // being expanded here because we always perform the static sub type
3008 // check in SubTypeCheckNode::sub() regardless of whether
3009 // StressReflectiveCode is set or not. We can therefore skip this
3010 // static check when StressReflectiveCode is on.
3011 switch (C->static_subtype_check(superk, subk)) {
3012 case Compile::SSC_always_false:
3013 {
3014 Node* always_fail = *ctrl;
3015 *ctrl = gvn.C->top();
3016 return always_fail;
3017 }
3018 case Compile::SSC_always_true:
3019 return C->top();
3020 case Compile::SSC_easy_test:
3021 {
3022 // Just do a direct pointer compare and be done.
3023 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3024 *ctrl = gvn.transform(new IfTrueNode(iff));
3025 return gvn.transform(new IfFalseNode(iff));
3026 }
3027 case Compile::SSC_full_test:
3028 break;
3029 default:
3030 ShouldNotReachHere();
3031 }
3032 }
3033
3034 // %%% Possible further optimization: Even if the superklass is not exact,
3035 // if the subklass is the unique subtype of the superklass, the check
3036 // will always succeed. We could leave a dependency behind to ensure this.
3037
3038 // First load the super-klass's check-offset
3039 Node *p1 = gvn.transform(new AddPNode(C->top(), superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3040 Node* m = C->immutable_memory();
3041 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3042 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3043 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3044 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
3045 bool might_be_cache = (chk_off_con == cacheoff_con);
3046
3047 // Load from the sub-klass's super-class display list, or a 1-word cache of
3048 // the secondary superclass list, or a failing value with a sentinel offset
3049 // if the super-klass is an interface or exceptionally deep in the Java
3050 // hierarchy and we have to scan the secondary superclass list the hard way.
3051 // Worst-case type is a little odd: null is allowed as a result (usually
3052 // klass loads can never produce a null).
3053 Node *chk_off_X = chk_off;
3054 #ifdef _LP64
3055 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
3056 #endif
3057 Node* p2 = gvn.transform(new AddPNode(C->top(), subklass, chk_off_X));
3058 // For some types like interfaces the following loadKlass is from a 1-word
3059 // cache which is mutable so can't use immutable memory. Other
3060 // types load from the super-class display table which is immutable.
3061 Node *kmem = C->immutable_memory();
3062 // secondary_super_cache is not immutable but can be treated as such because:
3063 // - no ideal node writes to it in a way that could cause an
3064 // incorrect/missed optimization of the following Load.
3065 // - it's a cache so, worse case, not reading the latest value
3066 // wouldn't cause incorrect execution
3067 if (might_be_cache && mem != nullptr) {
3068 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
3069 }
3070 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3071
3072 // Compile speed common case: ARE a subtype and we canNOT fail
3073 if (superklass == nkls) {
3074 return C->top(); // false path is dead; no test needed.
3075 }
3076
3077 // Gather the various success & failures here
3078 RegionNode* r_not_subtype = new RegionNode(3);
3079 gvn.record_for_igvn(r_not_subtype);
3080 RegionNode* r_ok_subtype = new RegionNode(4);
3081 gvn.record_for_igvn(r_ok_subtype);
3082
3083 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3084 // SubTypeCheck node
3085 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3086 ciCallProfile profile = method->call_profile_at_bci(bci);
3087 float total_prob = 0;
3088 for (int i = 0; profile.has_receiver(i); ++i) {
3089 float prob = profile.receiver_prob(i);
3090 total_prob += prob;
3091 }
3092 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3093 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3094 for (int i = 0; profile.has_receiver(i); ++i) {
3095 ciKlass* klass = profile.receiver(i);
3096 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3097 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3098 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3099 continue;
3100 }
3101 if (klass_t->isa_aryklassptr()) {
3102 // For a direct pointer comparison, we need the refined array klass pointer
3103 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3104 }
3105 float prob = profile.receiver_prob(i);
3106 ConNode* klass_node = gvn.makecon(klass_t);
3107 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3108 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3109
3110 if (result == Compile::SSC_always_true) {
3111 r_ok_subtype->add_req(iftrue);
3112 } else {
3113 assert(result == Compile::SSC_always_false, "");
3114 r_not_subtype->add_req(iftrue);
3115 }
3116 *ctrl = gvn.transform(new IfFalseNode(iff));
3117 }
3118 }
3119 }
3120
3121 // See if we get an immediate positive hit. Happens roughly 83% of the
3122 // time. Test to see if the value loaded just previously from the subklass
3123 // is exactly the superklass.
3124 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3125 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3126 *ctrl = gvn.transform(new IfFalseNode(iff1));
3127
3128 // Compile speed common case: Check for being deterministic right now. If
3129 // chk_off is a constant and not equal to cacheoff then we are NOT a
3130 // subklass. In this case we need exactly the 1 test above and we can
3131 // return those results immediately.
3132 if (!might_be_cache) {
3133 Node* not_subtype_ctrl = *ctrl;
3134 *ctrl = iftrue1; // We need exactly the 1 test above
3135 PhaseIterGVN* igvn = gvn.is_IterGVN();
3136 if (igvn != nullptr) {
3137 igvn->remove_globally_dead_node(r_ok_subtype);
3138 igvn->remove_globally_dead_node(r_not_subtype);
3139 }
3140 return not_subtype_ctrl;
3141 }
3142
3143 r_ok_subtype->init_req(1, iftrue1);
3144
3145 // Check for immediate negative hit. Happens roughly 11% of the time (which
3146 // is roughly 63% of the remaining cases). Test to see if the loaded
3147 // check-offset points into the subklass display list or the 1-element
3148 // cache. If it points to the display (and NOT the cache) and the display
3149 // missed then it's not a subtype.
3150 Node *cacheoff = gvn.intcon(cacheoff_con);
3151 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3152 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3153 *ctrl = gvn.transform(new IfFalseNode(iff2));
3154
3155 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3156 // No performance impact (too rare) but allows sharing of secondary arrays
3157 // which has some footprint reduction.
3158 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3159 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3160 *ctrl = gvn.transform(new IfFalseNode(iff3));
3161
3162 // -- Roads not taken here: --
3163 // We could also have chosen to perform the self-check at the beginning
3164 // of this code sequence, as the assembler does. This would not pay off
3165 // the same way, since the optimizer, unlike the assembler, can perform
3166 // static type analysis to fold away many successful self-checks.
3167 // Non-foldable self checks work better here in second position, because
3168 // the initial primary superclass check subsumes a self-check for most
3169 // types. An exception would be a secondary type like array-of-interface,
3170 // which does not appear in its own primary supertype display.
3171 // Finally, we could have chosen to move the self-check into the
3172 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3173 // dependent manner. But it is worthwhile to have the check here,
3174 // where it can be perhaps be optimized. The cost in code space is
3175 // small (register compare, branch).
3176
3177 // Now do a linear scan of the secondary super-klass array. Again, no real
3178 // performance impact (too rare) but it's gotta be done.
3179 // Since the code is rarely used, there is no penalty for moving it
3180 // out of line, and it can only improve I-cache density.
3181 // The decision to inline or out-of-line this final check is platform
3182 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3183 Node* psc = gvn.transform(
3184 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3185
3186 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3187 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3188 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3189
3190 // Return false path; set default control to true path.
3191 *ctrl = gvn.transform(r_ok_subtype);
3192 return gvn.transform(r_not_subtype);
3193 }
3194
3195 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3196 const Type* sub_t = _gvn.type(obj_or_subklass);
3197 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3198 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3199 obj_or_subklass = makecon(sub_t);
3200 }
3201 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3202 if (expand_subtype_check) {
3203 MergeMemNode* mem = merged_memory();
3204 Node* ctrl = control();
3205 Node* subklass = obj_or_subklass;
3206 if (!sub_t->isa_klassptr()) {
3207 subklass = load_object_klass(obj_or_subklass);
3208 }
3209
3210 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3211 set_control(ctrl);
3212 return n;
3213 }
3214
3215 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3216 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3217 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3218 set_control(_gvn.transform(new IfTrueNode(iff)));
3219 return _gvn.transform(new IfFalseNode(iff));
3220 }
3221
3222 // Profile-driven exact type check:
3223 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3224 float prob, Node* *casted_receiver) {
3225 assert(!klass->is_interface(), "no exact type check on interfaces");
3226 Node* fail = top();
3227 const Type* rec_t = _gvn.type(receiver);
3228 if (rec_t->is_inlinetypeptr()) {
3229 if (klass->equals(rec_t->inline_klass())) {
3230 (*casted_receiver) = receiver; // Always passes
3231 } else {
3232 (*casted_receiver) = top(); // Always fails
3233 fail = control();
3234 set_control(top());
3235 }
3236 return fail;
3237 }
3238 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3239 if (tklass->isa_aryklassptr()) {
3240 // For a direct pointer comparison, we need the refined array klass pointer
3241 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3242 }
3243 Node* recv_klass = load_object_klass(receiver);
3244 fail = type_check(recv_klass, tklass, prob);
3245
3246 if (!stopped()) {
3247 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3248 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3249 assert(recv_xtype->klass_is_exact(), "");
3250
3251 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3252 // Subsume downstream occurrences of receiver with a cast to
3253 // recv_xtype, since now we know what the type will be.
3254 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3255 Node* res = _gvn.transform(cast);
3256 if (recv_xtype->is_inlinetypeptr()) {
3257 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3258 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3259 }
3260 (*casted_receiver) = res;
3261 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3262 // (User must make the replace_in_map call.)
3263 }
3264 }
3265
3266 return fail;
3267 }
3268
3269 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3270 float prob) {
3271 Node* want_klass = makecon(tklass);
3272 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3273 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3274 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3275 set_control(_gvn.transform(new IfTrueNode (iff)));
3276 Node* fail = _gvn.transform(new IfFalseNode(iff));
3277 return fail;
3278 }
3279
3280 //------------------------------subtype_check_receiver-------------------------
3281 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3282 Node** casted_receiver) {
3283 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3284 Node* want_klass = makecon(tklass);
3285
3286 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3287
3288 // Ignore interface type information until interface types are properly tracked.
3289 if (!stopped() && !klass->is_interface()) {
3290 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3291 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3292 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3293 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3294 if (recv_type->is_inlinetypeptr()) {
3295 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3296 }
3297 (*casted_receiver) = cast;
3298 }
3299 }
3300
3301 return slow_ctl;
3302 }
3303
3304 //------------------------------seems_never_null-------------------------------
3305 // Use null_seen information if it is available from the profile.
3306 // If we see an unexpected null at a type check we record it and force a
3307 // recompile; the offending check will be recompiled to handle nulls.
3308 // If we see several offending BCIs, then all checks in the
3309 // method will be recompiled.
3310 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3311 speculating = !_gvn.type(obj)->speculative_maybe_null();
3312 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3313 if (UncommonNullCast // Cutout for this technique
3314 && obj != null() // And not the -Xcomp stupid case?
3315 && !too_many_traps(reason)
3316 ) {
3317 if (speculating) {
3318 return true;
3319 }
3320 if (data == nullptr)
3321 // Edge case: no mature data. Be optimistic here.
3322 return true;
3323 // If the profile has not seen a null, assume it won't happen.
3324 assert(java_bc() == Bytecodes::_checkcast ||
3325 java_bc() == Bytecodes::_instanceof ||
3326 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3327 return !data->as_BitData()->null_seen();
3328 }
3329 speculating = false;
3330 return false;
3331 }
3332
3333 void GraphKit::guard_klass_being_initialized(Node* klass) {
3334 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3335 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3336 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3337 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3338 T_BYTE, MemNode::acquire);
3339 init_state = _gvn.transform(init_state);
3340
3341 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3342
3343 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3344 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3345
3346 { BuildCutout unless(this, tst, PROB_MAX);
3347 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3348 }
3349 }
3350
3351 void GraphKit::guard_init_thread(Node* klass) {
3352 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3353 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3354
3355 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3356 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3357 T_ADDRESS, MemNode::unordered);
3358 init_thread = _gvn.transform(init_thread);
3359
3360 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3361
3362 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3363 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3364
3365 { BuildCutout unless(this, tst, PROB_MAX);
3366 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3367 }
3368 }
3369
3370 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3371 if (ik->is_being_initialized()) {
3372 if (C->needs_clinit_barrier(ik, context)) {
3373 Node* klass = makecon(TypeKlassPtr::make(ik));
3374 guard_klass_being_initialized(klass);
3375 guard_init_thread(klass);
3376 insert_mem_bar(Op_MemBarCPUOrder);
3377 }
3378 } else if (ik->is_initialized()) {
3379 return; // no barrier needed
3380 } else {
3381 uncommon_trap(Deoptimization::Reason_uninitialized,
3382 Deoptimization::Action_reinterpret,
3383 nullptr);
3384 }
3385 }
3386
3387 //------------------------maybe_cast_profiled_receiver-------------------------
3388 // If the profile has seen exactly one type, narrow to exactly that type.
3389 // Subsequent type checks will always fold up.
3390 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3391 const TypeKlassPtr* require_klass,
3392 ciKlass* spec_klass,
3393 bool safe_for_replace) {
3394 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3395
3396 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3397
3398 // Make sure we haven't already deoptimized from this tactic.
3399 if (too_many_traps_or_recompiles(reason))
3400 return nullptr;
3401
3402 // (No, this isn't a call, but it's enough like a virtual call
3403 // to use the same ciMethod accessor to get the profile info...)
3404 // If we have a speculative type use it instead of profiling (which
3405 // may not help us)
3406 ciKlass* exact_kls = spec_klass;
3407 if (exact_kls == nullptr) {
3408 if (java_bc() == Bytecodes::_aastore) {
3409 ciKlass* array_type = nullptr;
3410 ciKlass* element_type = nullptr;
3411 ProfilePtrKind element_ptr = ProfileMaybeNull;
3412 bool flat_array = true;
3413 bool null_free_array = true;
3414 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3415 exact_kls = element_type;
3416 } else {
3417 exact_kls = profile_has_unique_klass();
3418 }
3419 }
3420 if (exact_kls != nullptr) {// no cast failures here
3421 if (require_klass == nullptr ||
3422 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3423 // If we narrow the type to match what the type profile sees or
3424 // the speculative type, we can then remove the rest of the
3425 // cast.
3426 // This is a win, even if the exact_kls is very specific,
3427 // because downstream operations, such as method calls,
3428 // will often benefit from the sharper type.
3429 Node* exact_obj = not_null_obj; // will get updated in place...
3430 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3431 &exact_obj);
3432 { PreserveJVMState pjvms(this);
3433 set_control(slow_ctl);
3434 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3435 }
3436 if (safe_for_replace) {
3437 replace_in_map(not_null_obj, exact_obj);
3438 }
3439 return exact_obj;
3440 }
3441 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3442 }
3443
3444 return nullptr;
3445 }
3446
3447 /**
3448 * Cast obj to type and emit guard unless we had too many traps here
3449 * already
3450 *
3451 * @param obj node being casted
3452 * @param type type to cast the node to
3453 * @param not_null true if we know node cannot be null
3454 */
3455 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3456 ciKlass* type,
3457 bool not_null) {
3458 if (stopped()) {
3459 return obj;
3460 }
3461
3462 // type is null if profiling tells us this object is always null
3463 if (type != nullptr) {
3464 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3465 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3466
3467 if (!too_many_traps_or_recompiles(null_reason) &&
3468 !too_many_traps_or_recompiles(class_reason)) {
3469 Node* not_null_obj = nullptr;
3470 // not_null is true if we know the object is not null and
3471 // there's no need for a null check
3472 if (!not_null) {
3473 Node* null_ctl = top();
3474 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3475 assert(null_ctl->is_top(), "no null control here");
3476 } else {
3477 not_null_obj = obj;
3478 }
3479
3480 Node* exact_obj = not_null_obj;
3481 ciKlass* exact_kls = type;
3482 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3483 &exact_obj);
3484 {
3485 PreserveJVMState pjvms(this);
3486 set_control(slow_ctl);
3487 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3488 }
3489 replace_in_map(not_null_obj, exact_obj);
3490 obj = exact_obj;
3491 }
3492 } else {
3493 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3494 Node* exact_obj = null_assert(obj);
3495 replace_in_map(obj, exact_obj);
3496 obj = exact_obj;
3497 }
3498 }
3499 return obj;
3500 }
3501
3502 //-------------------------------gen_instanceof--------------------------------
3503 // Generate an instance-of idiom. Used by both the instance-of bytecode
3504 // and the reflective instance-of call.
3505 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3506 kill_dead_locals(); // Benefit all the uncommon traps
3507 assert( !stopped(), "dead parse path should be checked in callers" );
3508 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3509 "must check for not-null not-dead klass in callers");
3510
3511 // Make the merge point
3512 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3513 RegionNode* region = new RegionNode(PATH_LIMIT);
3514 Node* phi = new PhiNode(region, TypeInt::BOOL);
3515 C->set_has_split_ifs(true); // Has chance for split-if optimization
3516
3517 ciProfileData* data = nullptr;
3518 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3519 data = method()->method_data()->bci_to_data(bci());
3520 }
3521 bool speculative_not_null = false;
3522 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3523 && seems_never_null(obj, data, speculative_not_null));
3524
3525 // Null check; get casted pointer; set region slot 3
3526 Node* null_ctl = top();
3527 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3528
3529 // If not_null_obj is dead, only null-path is taken
3530 if (stopped()) { // Doing instance-of on a null?
3531 set_control(null_ctl);
3532 return intcon(0);
3533 }
3534 region->init_req(_null_path, null_ctl);
3535 phi ->init_req(_null_path, intcon(0)); // Set null path value
3536 if (null_ctl == top()) {
3537 // Do this eagerly, so that pattern matches like is_diamond_phi
3538 // will work even during parsing.
3539 assert(_null_path == PATH_LIMIT-1, "delete last");
3540 region->del_req(_null_path);
3541 phi ->del_req(_null_path);
3542 }
3543
3544 // Do we know the type check always succeed?
3545 bool known_statically = false;
3546 if (_gvn.type(superklass)->singleton()) {
3547 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3548 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3549 if (subk != nullptr && subk->is_loaded()) {
3550 int static_res = C->static_subtype_check(superk, subk);
3551 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3552 }
3553 }
3554
3555 if (!known_statically) {
3556 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3557 // We may not have profiling here or it may not help us. If we
3558 // have a speculative type use it to perform an exact cast.
3559 ciKlass* spec_obj_type = obj_type->speculative_type();
3560 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3561 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3562 if (stopped()) { // Profile disagrees with this path.
3563 set_control(null_ctl); // Null is the only remaining possibility.
3564 return intcon(0);
3565 }
3566 if (cast_obj != nullptr) {
3567 not_null_obj = cast_obj;
3568 }
3569 }
3570 }
3571
3572 // Generate the subtype check
3573 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3574
3575 // Plug in the success path to the general merge in slot 1.
3576 region->init_req(_obj_path, control());
3577 phi ->init_req(_obj_path, intcon(1));
3578
3579 // Plug in the failing path to the general merge in slot 2.
3580 region->init_req(_fail_path, not_subtype_ctrl);
3581 phi ->init_req(_fail_path, intcon(0));
3582
3583 // Return final merged results
3584 set_control( _gvn.transform(region) );
3585 record_for_igvn(region);
3586
3587 // If we know the type check always succeeds then we don't use the
3588 // profiling data at this bytecode. Don't lose it, feed it to the
3589 // type system as a speculative type.
3590 if (safe_for_replace) {
3591 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3592 replace_in_map(obj, casted_obj);
3593 }
3594
3595 return _gvn.transform(phi);
3596 }
3597
3598 //-------------------------------gen_checkcast---------------------------------
3599 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3600 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3601 // uncommon-trap paths work. Adjust stack after this call.
3602 // If failure_control is supplied and not null, it is filled in with
3603 // the control edge for the cast failure. Otherwise, an appropriate
3604 // uncommon trap or exception is thrown.
3605 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3606 kill_dead_locals(); // Benefit all the uncommon traps
3607 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3608 const Type* obj_type = _gvn.type(obj);
3609
3610 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3611 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3612 bool safe_for_replace = (failure_control == nullptr);
3613 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3614
3615 // Fast cutout: Check the case that the cast is vacuously true.
3616 // This detects the common cases where the test will short-circuit
3617 // away completely. We do this before we perform the null check,
3618 // because if the test is going to turn into zero code, we don't
3619 // want a residual null check left around. (Causes a slowdown,
3620 // for example, in some objArray manipulations, such as a[i]=a[j].)
3621 if (improved_klass_ptr_type->singleton()) {
3622 const TypeKlassPtr* kptr = nullptr;
3623 if (obj_type->isa_oop_ptr()) {
3624 kptr = obj_type->is_oopptr()->as_klass_type();
3625 } else if (obj->is_InlineType()) {
3626 ciInlineKlass* vk = obj_type->inline_klass();
3627 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3628 }
3629
3630 if (kptr != nullptr) {
3631 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3632 case Compile::SSC_always_true:
3633 // If we know the type check always succeed then we don't use
3634 // the profiling data at this bytecode. Don't lose it, feed it
3635 // to the type system as a speculative type.
3636 obj = record_profiled_receiver_for_speculation(obj);
3637 if (null_free) {
3638 assert(safe_for_replace, "must be");
3639 obj = null_check(obj);
3640 }
3641 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3642 return obj;
3643 case Compile::SSC_always_false:
3644 if (null_free) {
3645 assert(safe_for_replace, "must be");
3646 obj = null_check(obj);
3647 }
3648 // It needs a null check because a null will *pass* the cast check.
3649 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3650 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3651 Deoptimization::DeoptReason reason = is_aastore ?
3652 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3653 builtin_throw(reason);
3654 return top();
3655 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3656 return null_assert(obj);
3657 }
3658 break; // Fall through to full check
3659 default:
3660 break;
3661 }
3662 }
3663 }
3664
3665 ciProfileData* data = nullptr;
3666 if (failure_control == nullptr) { // use MDO in regular case only
3667 assert(java_bc() == Bytecodes::_aastore ||
3668 java_bc() == Bytecodes::_checkcast,
3669 "interpreter profiles type checks only for these BCs");
3670 if (method()->method_data()->is_mature()) {
3671 data = method()->method_data()->bci_to_data(bci());
3672 }
3673 }
3674
3675 // Make the merge point
3676 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3677 RegionNode* region = new RegionNode(PATH_LIMIT);
3678 Node* phi = new PhiNode(region, toop);
3679 _gvn.set_type(region, Type::CONTROL);
3680 _gvn.set_type(phi, toop);
3681
3682 C->set_has_split_ifs(true); // Has chance for split-if optimization
3683
3684 // Use null-cast information if it is available
3685 bool speculative_not_null = false;
3686 bool never_see_null = ((failure_control == nullptr) // regular case only
3687 && seems_never_null(obj, data, speculative_not_null));
3688
3689 if (obj->is_InlineType()) {
3690 // Re-execute if buffering during triggers deoptimization
3691 PreserveReexecuteState preexecs(this);
3692 jvms()->set_should_reexecute(true);
3693 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3694 }
3695
3696 // Null check; get casted pointer; set region slot 3
3697 Node* null_ctl = top();
3698 Node* not_null_obj = nullptr;
3699 if (null_free) {
3700 assert(safe_for_replace, "must be");
3701 not_null_obj = null_check(obj);
3702 } else {
3703 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3704 }
3705
3706 // If not_null_obj is dead, only null-path is taken
3707 if (stopped()) { // Doing instance-of on a null?
3708 set_control(null_ctl);
3709 if (toop->is_inlinetypeptr()) {
3710 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3711 }
3712 return null();
3713 }
3714 region->init_req(_null_path, null_ctl);
3715 phi ->init_req(_null_path, null()); // Set null path value
3716 if (null_ctl == top()) {
3717 // Do this eagerly, so that pattern matches like is_diamond_phi
3718 // will work even during parsing.
3719 assert(_null_path == PATH_LIMIT-1, "delete last");
3720 region->del_req(_null_path);
3721 phi ->del_req(_null_path);
3722 }
3723
3724 Node* cast_obj = nullptr;
3725 if (improved_klass_ptr_type->klass_is_exact()) {
3726 // The following optimization tries to statically cast the speculative type of the object
3727 // (for example obtained during profiling) to the type of the superklass and then do a
3728 // dynamic check that the type of the object is what we expect. To work correctly
3729 // for checkcast and aastore the type of superklass should be exact.
3730 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3731 // We may not have profiling here or it may not help us. If we have
3732 // a speculative type use it to perform an exact cast.
3733 ciKlass* spec_obj_type = obj_type->speculative_type();
3734 if (spec_obj_type != nullptr || data != nullptr) {
3735 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3736 if (cast_obj != nullptr) {
3737 if (failure_control != nullptr) // failure is now impossible
3738 (*failure_control) = top();
3739 // adjust the type of the phi to the exact klass:
3740 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3741 }
3742 }
3743 }
3744
3745 if (cast_obj == nullptr) {
3746 // Generate the subtype check
3747 Node* improved_superklass = superklass;
3748 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3749 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3750 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3751 // Additionally, the benefit would only be minor in non-constant cases.
3752 improved_superklass = makecon(improved_klass_ptr_type);
3753 }
3754 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3755 // Plug in success path into the merge
3756 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3757 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3758 if (failure_control == nullptr) {
3759 if (not_subtype_ctrl != top()) { // If failure is possible
3760 PreserveJVMState pjvms(this);
3761 set_control(not_subtype_ctrl);
3762 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3763 Deoptimization::DeoptReason reason = is_aastore ?
3764 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3765 builtin_throw(reason);
3766 }
3767 } else {
3768 (*failure_control) = not_subtype_ctrl;
3769 }
3770 }
3771
3772 region->init_req(_obj_path, control());
3773 phi ->init_req(_obj_path, cast_obj);
3774
3775 // A merge of null or Casted-NotNull obj
3776 Node* res = _gvn.transform(phi);
3777
3778 // Note I do NOT always 'replace_in_map(obj,result)' here.
3779 // if( tk->klass()->can_be_primary_super() )
3780 // This means that if I successfully store an Object into an array-of-String
3781 // I 'forget' that the Object is really now known to be a String. I have to
3782 // do this because we don't have true union types for interfaces - if I store
3783 // a Baz into an array-of-Interface and then tell the optimizer it's an
3784 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3785 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3786 // replace_in_map( obj, res );
3787
3788 // Return final merged results
3789 set_control( _gvn.transform(region) );
3790 record_for_igvn(region);
3791
3792 bool not_inline = !toop->can_be_inline_type();
3793 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3794 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3795 // Check if obj has been loaded from an array
3796 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3797 Node* array = nullptr;
3798 if (obj->isa_Load()) {
3799 Node* address = obj->in(MemNode::Address);
3800 if (address->isa_AddP()) {
3801 array = address->as_AddP()->in(AddPNode::Base);
3802 }
3803 } else if (obj->is_Phi()) {
3804 Node* region = obj->in(0);
3805 // TODO make this more robust (see JDK-8231346)
3806 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3807 IfNode* iff = region->in(2)->in(0)->isa_If();
3808 if (iff != nullptr) {
3809 iff->is_flat_array_check(&_gvn, &array);
3810 }
3811 }
3812 }
3813 if (array != nullptr) {
3814 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3815 if (ary_t != nullptr) {
3816 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3817 // Casting array element to a non-inline-type, mark array as not null-free.
3818 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3819 replace_in_map(array, cast);
3820 array = cast;
3821 }
3822 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3823 // Casting array element to a non-flat-in-array type, mark array as not flat.
3824 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3825 replace_in_map(array, cast);
3826 array = cast;
3827 }
3828 }
3829 }
3830 }
3831
3832 if (!stopped() && !res->is_InlineType()) {
3833 res = record_profiled_receiver_for_speculation(res);
3834 if (toop->is_inlinetypeptr() && !maybe_larval) {
3835 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3836 res = vt;
3837 if (safe_for_replace) {
3838 replace_in_map(obj, vt);
3839 replace_in_map(not_null_obj, vt);
3840 replace_in_map(res, vt);
3841 }
3842 }
3843 }
3844 return res;
3845 }
3846
3847 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3848 // Load markword
3849 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3850 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3851 if (check_lock && !UseCompactObjectHeaders) {
3852 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3853 // Check if obj is locked
3854 Node* locked_bit = MakeConX(markWord::unlocked_value);
3855 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3856 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3857 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3858 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3859 _gvn.transform(iff);
3860 Node* locked_region = new RegionNode(3);
3861 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3862
3863 // Unlocked: Use bits from mark word
3864 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3865 mark_phi->init_req(1, mark);
3866
3867 // Locked: Load prototype header from klass
3868 set_control(_gvn.transform(new IfFalseNode(iff)));
3869 // Make loads control dependent to make sure they are only executed if array is locked
3870 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3871 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3872 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3873 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));
3874
3875 locked_region->init_req(2, control());
3876 mark_phi->init_req(2, proto);
3877 set_control(_gvn.transform(locked_region));
3878 record_for_igvn(locked_region);
3879
3880 mark = mark_phi;
3881 }
3882
3883 // Now check if mark word bits are set
3884 Node* mask = MakeConX(mask_val);
3885 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3886 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3887 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3888 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3889 }
3890
3891 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3892 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3893 }
3894
3895 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3896 // We can't use immutable memory here because the mark word is mutable.
3897 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3898 // check is moved out of loops (mainly to enable loop unswitching).
3899 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3900 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3901 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3902 }
3903
3904 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3905 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3906 }
3907
3908 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3909 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3910
3911 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3912 if (!vk->has_null_free_non_atomic_layout()) {
3913 return intcon(1); // Always atomic
3914 } else if (!vk->has_null_free_atomic_layout()) {
3915 return intcon(0); // Never atomic
3916 }
3917
3918 Node* array_klass = load_object_klass(array);
3919 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3920 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
3921 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3922 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3923 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3924 }
3925
3926 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3927 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3928 RegionNode* region = new RegionNode(3);
3929 Node* null_ctl = top();
3930 null_check_oop(val, &null_ctl);
3931 if (null_ctl != top()) {
3932 PreserveJVMState pjvms(this);
3933 set_control(null_ctl);
3934 {
3935 // Deoptimize if null-free array
3936 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3937 inc_sp(nargs);
3938 uncommon_trap(Deoptimization::Reason_null_check,
3939 Deoptimization::Action_none);
3940 }
3941 region->init_req(1, control());
3942 }
3943 region->init_req(2, control());
3944 set_control(_gvn.transform(region));
3945 record_for_igvn(region);
3946 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3947 // Since we were just successfully storing null, the array can't be null free.
3948 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3949 ary_t = ary_t->cast_to_not_null_free();
3950 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3951 if (safe_for_replace) {
3952 replace_in_map(ary, cast);
3953 }
3954 ary = cast;
3955 }
3956 return ary;
3957 }
3958
3959 //------------------------------next_monitor-----------------------------------
3960 // What number should be given to the next monitor?
3961 int GraphKit::next_monitor() {
3962 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3963 int next = current + C->sync_stack_slots();
3964 // Keep the toplevel high water mark current:
3965 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3966 return current;
3967 }
3968
3969 //------------------------------insert_mem_bar---------------------------------
3970 // Memory barrier to avoid floating things around
3971 // The membar serves as a pinch point between both control and all memory slices.
3972 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3973 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3974 mb->init_req(TypeFunc::Control, control());
3975 mb->init_req(TypeFunc::Memory, reset_memory());
3976 Node* membar = _gvn.transform(mb);
3977 record_for_igvn(membar);
3978 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3979 set_all_memory_call(membar);
3980 return membar;
3981 }
3982
3983 //-------------------------insert_mem_bar_volatile----------------------------
3984 // Memory barrier to avoid floating things around
3985 // The membar serves as a pinch point between both control and memory(alias_idx).
3986 // If you want to make a pinch point on all memory slices, do not use this
3987 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3988 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3989 // When Parse::do_put_xxx updates a volatile field, it appends a series
3990 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3991 // The first membar is on the same memory slice as the field store opcode.
3992 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3993 // All the other membars (for other volatile slices, including AliasIdxBot,
3994 // which stands for all unknown volatile slices) are control-dependent
3995 // on the first membar. This prevents later volatile loads or stores
3996 // from sliding up past the just-emitted store.
3997
3998 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3999 mb->set_req(TypeFunc::Control,control());
4000 if (alias_idx == Compile::AliasIdxBot) {
4001 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
4002 } else {
4003 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
4004 mb->set_req(TypeFunc::Memory, memory(alias_idx));
4005 }
4006 Node* membar = _gvn.transform(mb);
4007 record_for_igvn(membar);
4008 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4009 if (alias_idx == Compile::AliasIdxBot) {
4010 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
4011 } else {
4012 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
4013 }
4014 return membar;
4015 }
4016
4017 //------------------------------shared_lock------------------------------------
4018 // Emit locking code.
4019 FastLockNode* GraphKit::shared_lock(Node* obj) {
4020 // bci is either a monitorenter bc or InvocationEntryBci
4021 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4022 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4023
4024 if (stopped()) // Dead monitor?
4025 return nullptr;
4026
4027 assert(dead_locals_are_killed(), "should kill locals before sync. point");
4028
4029 // Box the stack location
4030 Node* box = new BoxLockNode(next_monitor());
4031 // Check for bailout after new BoxLockNode
4032 if (failing()) { return nullptr; }
4033 box = _gvn.transform(box);
4034 Node* mem = reset_memory();
4035
4036 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
4037
4038 // Add monitor to debug info for the slow path. If we block inside the
4039 // slow path and de-opt, we need the monitor hanging around
4040 map()->push_monitor( flock );
4041
4042 const TypeFunc *tf = LockNode::lock_type();
4043 LockNode *lock = new LockNode(C, tf);
4044
4045 lock->init_req( TypeFunc::Control, control() );
4046 lock->init_req( TypeFunc::Memory , mem );
4047 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4048 lock->init_req( TypeFunc::FramePtr, frameptr() );
4049 lock->init_req( TypeFunc::ReturnAdr, top() );
4050
4051 lock->init_req(TypeFunc::Parms + 0, obj);
4052 lock->init_req(TypeFunc::Parms + 1, box);
4053 lock->init_req(TypeFunc::Parms + 2, flock);
4054 add_safepoint_edges(lock);
4055
4056 lock = _gvn.transform( lock )->as_Lock();
4057
4058 // lock has no side-effects, sets few values
4059 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
4060
4061 insert_mem_bar(Op_MemBarAcquireLock);
4062
4063 // Add this to the worklist so that the lock can be eliminated
4064 record_for_igvn(lock);
4065
4066 #ifndef PRODUCT
4067 if (PrintLockStatistics) {
4068 // Update the counter for this lock. Don't bother using an atomic
4069 // operation since we don't require absolute accuracy.
4070 lock->create_lock_counter(map()->jvms());
4071 increment_counter(lock->counter()->addr());
4072 }
4073 #endif
4074
4075 return flock;
4076 }
4077
4078
4079 //------------------------------shared_unlock----------------------------------
4080 // Emit unlocking code.
4081 void GraphKit::shared_unlock(Node* box, Node* obj) {
4082 // bci is either a monitorenter bc or InvocationEntryBci
4083 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4084 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4085
4086 if (stopped()) { // Dead monitor?
4087 map()->pop_monitor(); // Kill monitor from debug info
4088 return;
4089 }
4090 assert(!obj->is_InlineType(), "should not unlock on inline type");
4091
4092 // Memory barrier to avoid floating things down past the locked region
4093 insert_mem_bar(Op_MemBarReleaseLock);
4094
4095 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4096 UnlockNode *unlock = new UnlockNode(C, tf);
4097 #ifdef ASSERT
4098 unlock->set_dbg_jvms(sync_jvms());
4099 #endif
4100 uint raw_idx = Compile::AliasIdxRaw;
4101 unlock->init_req( TypeFunc::Control, control() );
4102 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4103 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4104 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4105 unlock->init_req( TypeFunc::ReturnAdr, top() );
4106
4107 unlock->init_req(TypeFunc::Parms + 0, obj);
4108 unlock->init_req(TypeFunc::Parms + 1, box);
4109 unlock = _gvn.transform(unlock)->as_Unlock();
4110
4111 Node* mem = reset_memory();
4112
4113 // unlock has no side-effects, sets few values
4114 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4115
4116 // Kill monitor from debug info
4117 map()->pop_monitor( );
4118 }
4119
4120 //-------------------------------get_layout_helper-----------------------------
4121 // If the given klass is a constant or known to be an array,
4122 // fetch the constant layout helper value into constant_value
4123 // and return null. Otherwise, load the non-constant
4124 // layout helper value, and return the node which represents it.
4125 // This two-faced routine is useful because allocation sites
4126 // almost always feature constant types.
4127 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4128 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4129 if (!StressReflectiveCode && klass_t != nullptr) {
4130 bool xklass = klass_t->klass_is_exact();
4131 bool can_be_flat = false;
4132 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4133 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4134 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4135 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4136 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4137 }
4138 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4139 jint lhelper;
4140 if (klass_t->is_flat()) {
4141 lhelper = ary_type->flat_layout_helper();
4142 } else if (klass_t->isa_aryklassptr()) {
4143 BasicType elem = ary_type->elem()->array_element_basic_type();
4144 if (is_reference_type(elem, true)) {
4145 elem = T_OBJECT;
4146 }
4147 lhelper = Klass::array_layout_helper(elem);
4148 } else {
4149 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4150 }
4151 if (lhelper != Klass::_lh_neutral_value) {
4152 constant_value = lhelper;
4153 return (Node*) nullptr;
4154 }
4155 }
4156 }
4157 constant_value = Klass::_lh_neutral_value; // put in a known value
4158 Node* lhp = basic_plus_adr(top(), klass_node, in_bytes(Klass::layout_helper_offset()));
4159 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4160 }
4161
4162 // We just put in an allocate/initialize with a big raw-memory effect.
4163 // Hook selected additional alias categories on the initialization.
4164 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4165 MergeMemNode* init_in_merge,
4166 Node* init_out_raw) {
4167 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4168 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4169
4170 Node* prevmem = kit.memory(alias_idx);
4171 init_in_merge->set_memory_at(alias_idx, prevmem);
4172 if (init_out_raw != nullptr) {
4173 kit.set_memory(init_out_raw, alias_idx);
4174 }
4175 }
4176
4177 //---------------------------set_output_for_allocation-------------------------
4178 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4179 const TypeOopPtr* oop_type,
4180 bool deoptimize_on_exception) {
4181 int rawidx = Compile::AliasIdxRaw;
4182 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4183 add_safepoint_edges(alloc);
4184 Node* allocx = _gvn.transform(alloc);
4185 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4186 // create memory projection for i_o
4187 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4188 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4189
4190 // create a memory projection as for the normal control path
4191 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4192 set_memory(malloc, rawidx);
4193
4194 // a normal slow-call doesn't change i_o, but an allocation does
4195 // we create a separate i_o projection for the normal control path
4196 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4197 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4198
4199 // put in an initialization barrier
4200 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4201 rawoop)->as_Initialize();
4202 assert(alloc->initialization() == init, "2-way macro link must work");
4203 assert(init ->allocation() == alloc, "2-way macro link must work");
4204 {
4205 // Extract memory strands which may participate in the new object's
4206 // initialization, and source them from the new InitializeNode.
4207 // This will allow us to observe initializations when they occur,
4208 // and link them properly (as a group) to the InitializeNode.
4209 assert(init->in(InitializeNode::Memory) == malloc, "");
4210 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4211 init->set_req(InitializeNode::Memory, minit_in);
4212 record_for_igvn(minit_in); // fold it up later, if possible
4213 _gvn.set_type(minit_in, Type::MEMORY);
4214 Node* minit_out = memory(rawidx);
4215 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4216 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4217 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4218 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4219 // multiple projections as a result.
4220 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4221 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4222 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4223 if (oop_type->isa_aryptr()) {
4224 // Initially all flat array accesses share a single slice
4225 // but that changes after parsing. Prepare the memory graph so
4226 // it can optimize flat array accesses properly once they
4227 // don't share a single slice.
4228 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4229 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4230 int elemidx = C->get_alias_index(telemref);
4231 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4232 if (alias_adr_type->is_flat()) {
4233 C->set_flat_accesses();
4234 }
4235 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4236 } else if (oop_type->isa_instptr()) {
4237 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4238 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4239 ciField* field = ik->nonstatic_field_at(i);
4240 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4241 continue; // do not bother to track really large numbers of fields
4242 // Find (or create) the alias category for this field:
4243 int fieldidx = C->alias_type(field)->index();
4244 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4245 }
4246 }
4247 }
4248
4249 // Cast raw oop to the real thing...
4250 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4251 javaoop = _gvn.transform(javaoop);
4252 C->set_recent_alloc(control(), javaoop);
4253 assert(just_allocated_object(control()) == javaoop, "just allocated");
4254
4255 #ifdef ASSERT
4256 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4257 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4258 "Ideal_allocation works");
4259 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4260 "Ideal_allocation works");
4261 if (alloc->is_AllocateArray()) {
4262 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4263 "Ideal_allocation works");
4264 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4265 "Ideal_allocation works");
4266 } else {
4267 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4268 }
4269 }
4270 #endif //ASSERT
4271
4272 return javaoop;
4273 }
4274
4275 //---------------------------new_instance--------------------------------------
4276 // This routine takes a klass_node which may be constant (for a static type)
4277 // or may be non-constant (for reflective code). It will work equally well
4278 // for either, and the graph will fold nicely if the optimizer later reduces
4279 // the type to a constant.
4280 // The optional arguments are for specialized use by intrinsics:
4281 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4282 // - If 'return_size_val', report the total object size to the caller.
4283 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4284 Node* GraphKit::new_instance(Node* klass_node,
4285 Node* extra_slow_test,
4286 Node* *return_size_val,
4287 bool deoptimize_on_exception,
4288 InlineTypeNode* inline_type_node) {
4289 // Compute size in doublewords
4290 // The size is always an integral number of doublewords, represented
4291 // as a positive bytewise size stored in the klass's layout_helper.
4292 // The layout_helper also encodes (in a low bit) the need for a slow path.
4293 jint layout_con = Klass::_lh_neutral_value;
4294 Node* layout_val = get_layout_helper(klass_node, layout_con);
4295 bool layout_is_con = (layout_val == nullptr);
4296
4297 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4298 // Generate the initial go-slow test. It's either ALWAYS (return a
4299 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4300 // case) a computed value derived from the layout_helper.
4301 Node* initial_slow_test = nullptr;
4302 if (layout_is_con) {
4303 assert(!StressReflectiveCode, "stress mode does not use these paths");
4304 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4305 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4306 } else { // reflective case
4307 // This reflective path is used by Unsafe.allocateInstance.
4308 // (It may be stress-tested by specifying StressReflectiveCode.)
4309 // Basically, we want to get into the VM is there's an illegal argument.
4310 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4311 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4312 if (extra_slow_test != intcon(0)) {
4313 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4314 }
4315 // (Macro-expander will further convert this to a Bool, if necessary.)
4316 }
4317
4318 // Find the size in bytes. This is easy; it's the layout_helper.
4319 // The size value must be valid even if the slow path is taken.
4320 Node* size = nullptr;
4321 if (layout_is_con) {
4322 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4323 } else { // reflective case
4324 // This reflective path is used by clone and Unsafe.allocateInstance.
4325 size = ConvI2X(layout_val);
4326
4327 // Clear the low bits to extract layout_helper_size_in_bytes:
4328 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4329 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4330 size = _gvn.transform( new AndXNode(size, mask) );
4331 }
4332 if (return_size_val != nullptr) {
4333 (*return_size_val) = size;
4334 }
4335
4336 // This is a precise notnull oop of the klass.
4337 // (Actually, it need not be precise if this is a reflective allocation.)
4338 // It's what we cast the result to.
4339 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4340 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4341 const TypeOopPtr* oop_type = tklass->as_instance_type();
4342
4343 // Now generate allocation code
4344
4345 // The entire memory state is needed for slow path of the allocation
4346 // since GC and deoptimization can happen.
4347 Node *mem = reset_memory();
4348 set_all_memory(mem); // Create new memory state
4349
4350 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4351 control(), mem, i_o(),
4352 size, klass_node,
4353 initial_slow_test, inline_type_node);
4354
4355 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4356 }
4357
4358 //-------------------------------new_array-------------------------------------
4359 // helper for newarray and anewarray
4360 // The 'length' parameter is (obviously) the length of the array.
4361 // The optional arguments are for specialized use by intrinsics:
4362 // - If 'return_size_val', report the non-padded array size (sum of header size
4363 // and array body) to the caller.
4364 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4365 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4366 Node* length, // number of array elements
4367 int nargs, // number of arguments to push back for uncommon trap
4368 Node* *return_size_val,
4369 bool deoptimize_on_exception,
4370 Node* init_val) {
4371 jint layout_con = Klass::_lh_neutral_value;
4372 Node* layout_val = get_layout_helper(klass_node, layout_con);
4373 bool layout_is_con = (layout_val == nullptr);
4374
4375 if (!layout_is_con && !StressReflectiveCode &&
4376 !too_many_traps(Deoptimization::Reason_class_check)) {
4377 // This is a reflective array creation site.
4378 // Optimistically assume that it is a subtype of Object[],
4379 // so that we can fold up all the address arithmetic.
4380 layout_con = Klass::array_layout_helper(T_OBJECT);
4381 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4382 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4383 { BuildCutout unless(this, bol_lh, PROB_MAX);
4384 inc_sp(nargs);
4385 uncommon_trap(Deoptimization::Reason_class_check,
4386 Deoptimization::Action_maybe_recompile);
4387 }
4388 layout_val = nullptr;
4389 layout_is_con = true;
4390 }
4391
4392 // Generate the initial go-slow test. Make sure we do not overflow
4393 // if length is huge (near 2Gig) or negative! We do not need
4394 // exact double-words here, just a close approximation of needed
4395 // double-words. We can't add any offset or rounding bits, lest we
4396 // take a size -1 of bytes and make it positive. Use an unsigned
4397 // compare, so negative sizes look hugely positive.
4398 int fast_size_limit = FastAllocateSizeLimit;
4399 if (layout_is_con) {
4400 assert(!StressReflectiveCode, "stress mode does not use these paths");
4401 // Increase the size limit if we have exact knowledge of array type.
4402 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4403 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4404 }
4405
4406 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4407 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4408
4409 // --- Size Computation ---
4410 // array_size = round_to_heap(array_header + (length << elem_shift));
4411 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4412 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4413 // The rounding mask is strength-reduced, if possible.
4414 int round_mask = MinObjAlignmentInBytes - 1;
4415 Node* header_size = nullptr;
4416 // (T_BYTE has the weakest alignment and size restrictions...)
4417 if (layout_is_con) {
4418 int hsize = Klass::layout_helper_header_size(layout_con);
4419 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4420 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4421 if ((round_mask & ~right_n_bits(eshift)) == 0)
4422 round_mask = 0; // strength-reduce it if it goes away completely
4423 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4424 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4425 assert(header_size_min <= hsize, "generic minimum is smallest");
4426 header_size = intcon(hsize);
4427 } else {
4428 Node* hss = intcon(Klass::_lh_header_size_shift);
4429 Node* hsm = intcon(Klass::_lh_header_size_mask);
4430 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4431 header_size = _gvn.transform(new AndINode(header_size, hsm));
4432 }
4433
4434 Node* elem_shift = nullptr;
4435 if (layout_is_con) {
4436 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4437 if (eshift != 0)
4438 elem_shift = intcon(eshift);
4439 } else {
4440 // There is no need to mask or shift this value.
4441 // The semantics of LShiftINode include an implicit mask to 0x1F.
4442 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4443 elem_shift = layout_val;
4444 }
4445
4446 // Transition to native address size for all offset calculations:
4447 Node* lengthx = ConvI2X(length);
4448 Node* headerx = ConvI2X(header_size);
4449 #ifdef _LP64
4450 { const TypeInt* tilen = _gvn.find_int_type(length);
4451 if (tilen != nullptr && tilen->_lo < 0) {
4452 // Add a manual constraint to a positive range. Cf. array_element_address.
4453 jint size_max = fast_size_limit;
4454 if (size_max > tilen->_hi && tilen->_hi >= 0) {
4455 size_max = tilen->_hi;
4456 }
4457 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4458
4459 // Only do a narrow I2L conversion if the range check passed.
4460 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4461 _gvn.transform(iff);
4462 RegionNode* region = new RegionNode(3);
4463 _gvn.set_type(region, Type::CONTROL);
4464 lengthx = new PhiNode(region, TypeLong::LONG);
4465 _gvn.set_type(lengthx, TypeLong::LONG);
4466
4467 // Range check passed. Use ConvI2L node with narrow type.
4468 Node* passed = IfFalse(iff);
4469 region->init_req(1, passed);
4470 // Make I2L conversion control dependent to prevent it from
4471 // floating above the range check during loop optimizations.
4472 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4473
4474 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4475 region->init_req(2, IfTrue(iff));
4476 lengthx->init_req(2, ConvI2X(length));
4477
4478 set_control(region);
4479 record_for_igvn(region);
4480 record_for_igvn(lengthx);
4481 }
4482 }
4483 #endif
4484
4485 // Combine header size and body size for the array copy part, then align (if
4486 // necessary) for the allocation part. This computation cannot overflow,
4487 // because it is used only in two places, one where the length is sharply
4488 // limited, and the other after a successful allocation.
4489 Node* abody = lengthx;
4490 if (elem_shift != nullptr) {
4491 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4492 }
4493 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4494
4495 if (return_size_val != nullptr) {
4496 // This is the size
4497 (*return_size_val) = non_rounded_size;
4498 }
4499
4500 Node* size = non_rounded_size;
4501 if (round_mask != 0) {
4502 Node* mask1 = MakeConX(round_mask);
4503 size = _gvn.transform(new AddXNode(size, mask1));
4504 Node* mask2 = MakeConX(~round_mask);
4505 size = _gvn.transform(new AndXNode(size, mask2));
4506 }
4507 // else if round_mask == 0, the size computation is self-rounding
4508
4509 // Now generate allocation code
4510
4511 // The entire memory state is needed for slow path of the allocation
4512 // since GC and deoptimization can happen.
4513 Node *mem = reset_memory();
4514 set_all_memory(mem); // Create new memory state
4515
4516 if (initial_slow_test->is_Bool()) {
4517 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4518 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4519 }
4520
4521 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4522 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4523
4524 Node* raw_init_value = nullptr;
4525 if (init_val != nullptr) {
4526 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4527 if (ary_type->is_flat()) {
4528 initial_slow_test = intcon(1);
4529 }
4530
4531 if (UseCompressedOops) {
4532 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4533 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4534 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4535 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4536 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4537 } else {
4538 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4539 }
4540 }
4541
4542 Node* valid_length_test = _gvn.intcon(1);
4543 if (ary_type->isa_aryptr()) {
4544 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4545 jint max = TypeAryPtr::max_array_length(bt);
4546 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4547 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4548 }
4549
4550 // Create the AllocateArrayNode and its result projections
4551 AllocateArrayNode* alloc
4552 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4553 control(), mem, i_o(),
4554 size, klass_node,
4555 initial_slow_test,
4556 length, valid_length_test,
4557 init_val, raw_init_value);
4558 // Cast to correct type. Note that the klass_node may be constant or not,
4559 // and in the latter case the actual array type will be inexact also.
4560 // (This happens via a non-constant argument to inline_native_newArray.)
4561 // In any case, the value of klass_node provides the desired array type.
4562 const TypeInt* length_type = _gvn.find_int_type(length);
4563 if (ary_type->isa_aryptr() && length_type != nullptr) {
4564 // Try to get a better type than POS for the size
4565 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4566 }
4567
4568 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4569
4570 array_ideal_length(alloc, ary_type, true);
4571 return javaoop;
4572 }
4573
4574 // The following "Ideal_foo" functions are placed here because they recognize
4575 // the graph shapes created by the functions immediately above.
4576
4577 //---------------------------Ideal_allocation----------------------------------
4578 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4579 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4580 if (ptr == nullptr) { // reduce dumb test in callers
4581 return nullptr;
4582 }
4583
4584 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4585 ptr = bs->step_over_gc_barrier(ptr);
4586
4587 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4588 ptr = ptr->in(1);
4589 if (ptr == nullptr) return nullptr;
4590 }
4591 // Return null for allocations with several casts:
4592 // j.l.reflect.Array.newInstance(jobject, jint)
4593 // Object.clone()
4594 // to keep more precise type from last cast.
4595 if (ptr->is_Proj()) {
4596 Node* allo = ptr->in(0);
4597 if (allo != nullptr && allo->is_Allocate()) {
4598 return allo->as_Allocate();
4599 }
4600 }
4601 // Report failure to match.
4602 return nullptr;
4603 }
4604
4605 // Fancy version which also strips off an offset (and reports it to caller).
4606 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4607 intptr_t& offset) {
4608 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4609 if (base == nullptr) return nullptr;
4610 return Ideal_allocation(base);
4611 }
4612
4613 // Trace Initialize <- Proj[Parm] <- Allocate
4614 AllocateNode* InitializeNode::allocation() {
4615 Node* rawoop = in(InitializeNode::RawAddress);
4616 if (rawoop->is_Proj()) {
4617 Node* alloc = rawoop->in(0);
4618 if (alloc->is_Allocate()) {
4619 return alloc->as_Allocate();
4620 }
4621 }
4622 return nullptr;
4623 }
4624
4625 // Trace Allocate -> Proj[Parm] -> Initialize
4626 InitializeNode* AllocateNode::initialization() {
4627 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4628 if (rawoop == nullptr) return nullptr;
4629 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4630 Node* init = rawoop->fast_out(i);
4631 if (init->is_Initialize()) {
4632 assert(init->as_Initialize()->allocation() == this, "2-way link");
4633 return init->as_Initialize();
4634 }
4635 }
4636 return nullptr;
4637 }
4638
4639 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4640 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4641 // Too many traps seen?
4642 if (too_many_traps(reason)) {
4643 #ifdef ASSERT
4644 if (TraceLoopPredicate) {
4645 int tc = C->trap_count(reason);
4646 tty->print("too many traps=%s tcount=%d in ",
4647 Deoptimization::trap_reason_name(reason), tc);
4648 method()->print(); // which method has too many predicate traps
4649 tty->cr();
4650 }
4651 #endif
4652 // We cannot afford to take more traps here,
4653 // do not generate Parse Predicate.
4654 return;
4655 }
4656
4657 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4658 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4659 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4660 {
4661 PreserveJVMState pjvms(this);
4662 set_control(if_false);
4663 inc_sp(nargs);
4664 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4665 }
4666 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4667 set_control(if_true);
4668 }
4669
4670 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4671 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4672 void GraphKit::add_parse_predicates(int nargs) {
4673 if (ShortRunningLongLoop) {
4674 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4675 // walking up from the loop.
4676 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4677 }
4678 if (UseLoopPredicate) {
4679 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4680 if (UseProfiledLoopPredicate) {
4681 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4682 }
4683 }
4684 if (UseAutoVectorizationPredicate) {
4685 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4686 }
4687 // Loop Limit Check Predicate should be near the loop.
4688 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4689 }
4690
4691 void GraphKit::sync_kit(IdealKit& ideal) {
4692 reset_memory();
4693 set_all_memory(ideal.merged_memory());
4694 set_i_o(ideal.i_o());
4695 set_control(ideal.ctrl());
4696 }
4697
4698 void GraphKit::final_sync(IdealKit& ideal) {
4699 // Final sync IdealKit and graphKit.
4700 sync_kit(ideal);
4701 }
4702
4703 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4704 Node* len = load_array_length(load_String_value(str, set_ctrl));
4705 Node* coder = load_String_coder(str, set_ctrl);
4706 // Divide length by 2 if coder is UTF16
4707 return _gvn.transform(new RShiftINode(len, coder));
4708 }
4709
4710 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4711 int value_offset = java_lang_String::value_offset();
4712 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4713 false, nullptr, Type::Offset(0));
4714 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4715 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4716 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4717 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4718 Node* p = basic_plus_adr(str, str, value_offset);
4719 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4720 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4721 return load;
4722 }
4723
4724 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4725 if (!CompactStrings) {
4726 return intcon(java_lang_String::CODER_UTF16);
4727 }
4728 int coder_offset = java_lang_String::coder_offset();
4729 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4730 false, nullptr, Type::Offset(0));
4731 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4732
4733 Node* p = basic_plus_adr(str, str, coder_offset);
4734 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4735 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4736 return load;
4737 }
4738
4739 void GraphKit::store_String_value(Node* str, Node* value) {
4740 int value_offset = java_lang_String::value_offset();
4741 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4742 false, nullptr, Type::Offset(0));
4743 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4744
4745 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4746 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4747 }
4748
4749 void GraphKit::store_String_coder(Node* str, Node* value) {
4750 int coder_offset = java_lang_String::coder_offset();
4751 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4752 false, nullptr, Type::Offset(0));
4753 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4754
4755 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4756 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4757 }
4758
4759 // Capture src and dst memory state with a MergeMemNode
4760 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4761 if (src_type == dst_type) {
4762 // Types are equal, we don't need a MergeMemNode
4763 return memory(src_type);
4764 }
4765 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4766 record_for_igvn(merge); // fold it up later, if possible
4767 int src_idx = C->get_alias_index(src_type);
4768 int dst_idx = C->get_alias_index(dst_type);
4769 merge->set_memory_at(src_idx, memory(src_idx));
4770 merge->set_memory_at(dst_idx, memory(dst_idx));
4771 return merge;
4772 }
4773
4774 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4775 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4776 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4777 // If input and output memory types differ, capture both states to preserve
4778 // the dependency between preceding and subsequent loads/stores.
4779 // For example, the following program:
4780 // StoreB
4781 // compress_string
4782 // LoadB
4783 // has this memory graph (use->def):
4784 // LoadB -> compress_string -> CharMem
4785 // ... -> StoreB -> ByteMem
4786 // The intrinsic hides the dependency between LoadB and StoreB, causing
4787 // the load to read from memory not containing the result of the StoreB.
4788 // The correct memory graph should look like this:
4789 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4790 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4791 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4792 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4793 set_memory(res_mem, TypeAryPtr::BYTES);
4794 return str;
4795 }
4796
4797 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4798 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4799 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4800 // Capture src and dst memory (see comment in 'compress_string').
4801 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4802 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4803 set_memory(_gvn.transform(str), dst_type);
4804 }
4805
4806 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4807 /**
4808 * int i_char = start;
4809 * for (int i_byte = 0; i_byte < count; i_byte++) {
4810 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4811 * }
4812 */
4813 add_parse_predicates();
4814 C->set_has_loops(true);
4815
4816 RegionNode* head = new RegionNode(3);
4817 head->init_req(1, control());
4818 gvn().set_type(head, Type::CONTROL);
4819 record_for_igvn(head);
4820
4821 Node* i_byte = new PhiNode(head, TypeInt::INT);
4822 i_byte->init_req(1, intcon(0));
4823 gvn().set_type(i_byte, TypeInt::INT);
4824 record_for_igvn(i_byte);
4825
4826 Node* i_char = new PhiNode(head, TypeInt::INT);
4827 i_char->init_req(1, start);
4828 gvn().set_type(i_char, TypeInt::INT);
4829 record_for_igvn(i_char);
4830
4831 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4832 gvn().set_type(mem, Type::MEMORY);
4833 record_for_igvn(mem);
4834 set_control(head);
4835 set_memory(mem, TypeAryPtr::BYTES);
4836 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4837 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4838 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4839 false, true /* mismatched */);
4840
4841 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4842 head->init_req(2, IfTrue(iff));
4843 mem->init_req(2, st);
4844 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4845 i_char->init_req(2, AddI(i_char, intcon(2)));
4846
4847 set_control(IfFalse(iff));
4848 set_memory(st, TypeAryPtr::BYTES);
4849 }
4850
4851 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4852 if (!field->is_constant()) {
4853 return nullptr; // Field not marked as constant.
4854 }
4855 ciInstance* holder = nullptr;
4856 if (!field->is_static()) {
4857 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4858 if (const_oop != nullptr && const_oop->is_instance()) {
4859 holder = const_oop->as_instance();
4860 }
4861 }
4862 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4863 /*is_unsigned_load=*/false);
4864 if (con_type != nullptr) {
4865 Node* con = makecon(con_type);
4866 if (field->type()->is_inlinetype()) {
4867 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4868 } else if (con_type->is_inlinetypeptr()) {
4869 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4870 }
4871 return con;
4872 }
4873 return nullptr;
4874 }
4875
4876 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4877 const Type* obj_type = obj->bottom_type();
4878 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4879 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4880 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4881 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4882 obj = casted_obj;
4883 }
4884 if (sig_type->is_inlinetypeptr()) {
4885 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4886 }
4887 return obj;
4888 }