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 = off_heap_plus_addr(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 GraphKit::SavedState::SavedState(GraphKit* kit) :
713 _kit(kit),
714 _sp(kit->sp()),
715 _jvms(kit->jvms()),
716 _map(kit->clone_map()),
717 _discarded(false)
718 {
719 for (DUIterator_Fast imax, i = kit->control()->fast_outs(imax); i < imax; i++) {
720 Node* out = kit->control()->fast_out(i);
721 if (out->is_CFG()) {
722 _ctrl_succ.push(out);
723 }
724 }
725 }
726
727 GraphKit::SavedState::~SavedState() {
728 if (_discarded) {
729 _kit->destruct_map_clone(_map);
730 return;
731 }
732 _kit->jvms()->set_map(_map);
733 _kit->jvms()->set_sp(_sp);
734 _map->set_jvms(_kit->jvms());
735 _kit->set_map(_map);
736 _kit->set_sp(_sp);
737 for (DUIterator_Fast imax, i = _kit->control()->fast_outs(imax); i < imax; i++) {
738 Node* out = _kit->control()->fast_out(i);
739 if (out->is_CFG() && out->in(0) == _kit->control() && out != _kit->map() && !_ctrl_succ.member(out)) {
740 _kit->_gvn.hash_delete(out);
741 out->set_req(0, _kit->C->top());
742 _kit->C->record_for_igvn(out);
743 --i; --imax;
744 _kit->_gvn.hash_find_insert(out);
745 }
746 }
747 }
748
749 void GraphKit::SavedState::discard() {
750 _discarded = true;
751 }
752
753
754 //----------------------------PreserveJVMState---------------------------------
755 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
756 DEBUG_ONLY(kit->verify_map());
757 _kit = kit;
758 _map = kit->map(); // preserve the map
759 _sp = kit->sp();
760 kit->set_map(clone_map ? kit->clone_map() : nullptr);
761 #ifdef ASSERT
762 _bci = kit->bci();
763 Parse* parser = kit->is_Parse();
764 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
765 _block = block;
766 #endif
767 }
768 PreserveJVMState::~PreserveJVMState() {
769 GraphKit* kit = _kit;
770 #ifdef ASSERT
771 assert(kit->bci() == _bci, "bci must not shift");
772 Parse* parser = kit->is_Parse();
773 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
774 assert(block == _block, "block must not shift");
775 #endif
776 kit->set_map(_map);
777 kit->set_sp(_sp);
778 }
779
780
781 //-----------------------------BuildCutout-------------------------------------
782 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
783 : PreserveJVMState(kit)
784 {
785 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
786 SafePointNode* outer_map = _map; // preserved map is caller's
787 SafePointNode* inner_map = kit->map();
788 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
789 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
790 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
791 }
792 BuildCutout::~BuildCutout() {
793 GraphKit* kit = _kit;
794 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
795 }
796
797 //---------------------------PreserveReexecuteState----------------------------
798 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
799 assert(!kit->stopped(), "must call stopped() before");
800 _kit = kit;
801 _sp = kit->sp();
802 _reexecute = kit->jvms()->_reexecute;
803 }
804 PreserveReexecuteState::~PreserveReexecuteState() {
805 if (_kit->stopped()) return;
806 _kit->jvms()->_reexecute = _reexecute;
807 _kit->set_sp(_sp);
808 }
809
810 //------------------------------clone_map--------------------------------------
811 // Implementation of PreserveJVMState
812 //
813 // Only clone_map(...) here. If this function is only used in the
814 // PreserveJVMState class we may want to get rid of this extra
815 // function eventually and do it all there.
816
817 SafePointNode* GraphKit::clone_map() {
818 if (map() == nullptr) return nullptr;
819
820 // Clone the memory edge first
821 Node* mem = MergeMemNode::make(map()->memory());
822 gvn().set_type_bottom(mem);
823
824 SafePointNode *clonemap = (SafePointNode*)map()->clone();
825 JVMState* jvms = this->jvms();
826 JVMState* clonejvms = jvms->clone_shallow(C);
827 clonemap->set_memory(mem);
828 clonemap->set_jvms(clonejvms);
829 clonejvms->set_map(clonemap);
830 record_for_igvn(clonemap);
831 gvn().set_type_bottom(clonemap);
832 return clonemap;
833 }
834
835 //-----------------------------destruct_map_clone------------------------------
836 //
837 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
838 // to destruct/free/delete in the exact opposite order as clone_map().
839 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
840 if (sfp == nullptr) return;
841
842 Node* mem = sfp->memory();
843 JVMState* jvms = sfp->jvms();
844
845 if (jvms != nullptr) {
846 delete jvms;
847 }
848
849 remove_for_igvn(sfp);
850 gvn().clear_type(sfp);
851 sfp->destruct(&_gvn);
852
853 if (mem != nullptr) {
854 gvn().clear_type(mem);
855 mem->destruct(&_gvn);
856 }
857 }
858
859 //-----------------------------set_map_clone-----------------------------------
860 void GraphKit::set_map_clone(SafePointNode* m) {
861 _map = m;
862 _map = clone_map();
863 _map->set_next_exception(nullptr);
864 DEBUG_ONLY(verify_map());
865 }
866
867
868 //----------------------------kill_dead_locals---------------------------------
869 // Detect any locals which are known to be dead, and force them to top.
870 void GraphKit::kill_dead_locals() {
871 // Consult the liveness information for the locals. If any
872 // of them are unused, then they can be replaced by top(). This
873 // should help register allocation time and cut down on the size
874 // of the deoptimization information.
875
876 // This call is made from many of the bytecode handling
877 // subroutines called from the Big Switch in do_one_bytecode.
878 // Every bytecode which might include a slow path is responsible
879 // for killing its dead locals. The more consistent we
880 // are about killing deads, the fewer useless phis will be
881 // constructed for them at various merge points.
882
883 // bci can be -1 (InvocationEntryBci). We return the entry
884 // liveness for the method.
885
886 if (method() == nullptr || method()->code_size() == 0) {
887 // We are building a graph for a call to a native method.
888 // All locals are live.
889 return;
890 }
891
892 ResourceMark rm;
893
894 // Consult the liveness information for the locals. If any
895 // of them are unused, then they can be replaced by top(). This
896 // should help register allocation time and cut down on the size
897 // of the deoptimization information.
898 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
899
900 int len = (int)live_locals.size();
901 assert(len <= jvms()->loc_size(), "too many live locals");
902 for (int local = 0; local < len; local++) {
903 if (!live_locals.at(local)) {
904 set_local(local, top());
905 }
906 }
907 }
908
909 #ifdef ASSERT
910 //-------------------------dead_locals_are_killed------------------------------
911 // Return true if all dead locals are set to top in the map.
912 // Used to assert "clean" debug info at various points.
913 bool GraphKit::dead_locals_are_killed() {
914 if (method() == nullptr || method()->code_size() == 0) {
915 // No locals need to be dead, so all is as it should be.
916 return true;
917 }
918
919 // Make sure somebody called kill_dead_locals upstream.
920 ResourceMark rm;
921 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
922 if (jvms->loc_size() == 0) continue; // no locals to consult
923 SafePointNode* map = jvms->map();
924 ciMethod* method = jvms->method();
925 int bci = jvms->bci();
926 if (jvms == this->jvms()) {
927 bci = this->bci(); // it might not yet be synched
928 }
929 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
930 int len = (int)live_locals.size();
931 if (!live_locals.is_valid() || len == 0)
932 // This method is trivial, or is poisoned by a breakpoint.
933 return true;
934 assert(len == jvms->loc_size(), "live map consistent with locals map");
935 for (int local = 0; local < len; local++) {
936 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
937 if (PrintMiscellaneous && (Verbose || WizardMode)) {
938 tty->print_cr("Zombie local %d: ", local);
939 jvms->dump();
940 }
941 return false;
942 }
943 }
944 }
945 return true;
946 }
947
948 #endif //ASSERT
949
950 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
951 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
952 ciMethod* cur_method = jvms->method();
953 int cur_bci = jvms->bci();
954 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
955 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
956 return Interpreter::bytecode_should_reexecute(code) ||
957 (is_anewarray && (code == Bytecodes::_multianewarray));
958 // Reexecute _multianewarray bytecode which was replaced with
959 // sequence of [a]newarray. See Parse::do_multianewarray().
960 //
961 // Note: interpreter should not have it set since this optimization
962 // is limited by dimensions and guarded by flag so in some cases
963 // multianewarray() runtime calls will be generated and
964 // the bytecode should not be reexecutes (stack will not be reset).
965 } else {
966 return false;
967 }
968 }
969
970 // Helper function for adding JVMState and debug information to node
971 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
972 // Add the safepoint edges to the call (or other safepoint).
973
974 // Make sure dead locals are set to top. This
975 // should help register allocation time and cut down on the size
976 // of the deoptimization information.
977 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
978
979 // Walk the inline list to fill in the correct set of JVMState's
980 // Also fill in the associated edges for each JVMState.
981
982 // If the bytecode needs to be reexecuted we need to put
983 // the arguments back on the stack.
984 const bool should_reexecute = jvms()->should_reexecute();
985 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
986
987 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
988 // undefined if the bci is different. This is normal for Parse but it
989 // should not happen for LibraryCallKit because only one bci is processed.
990 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
991 "in LibraryCallKit the reexecute bit should not change");
992
993 // If we are guaranteed to throw, we can prune everything but the
994 // input to the current bytecode.
995 bool can_prune_locals = false;
996 uint stack_slots_not_pruned = 0;
997 int inputs = 0, depth = 0;
998 if (must_throw) {
999 assert(method() == youngest_jvms->method(), "sanity");
1000 if (compute_stack_effects(inputs, depth)) {
1001 can_prune_locals = true;
1002 stack_slots_not_pruned = inputs;
1003 }
1004 }
1005
1006 if (env()->should_retain_local_variables()) {
1007 // At any safepoint, this method can get breakpointed, which would
1008 // then require an immediate deoptimization.
1009 can_prune_locals = false; // do not prune locals
1010 stack_slots_not_pruned = 0;
1011 }
1012
1013 // do not scribble on the input jvms
1014 JVMState* out_jvms = youngest_jvms->clone_deep(C);
1015 call->set_jvms(out_jvms); // Start jvms list for call node
1016
1017 // For a known set of bytecodes, the interpreter should reexecute them if
1018 // deoptimization happens. We set the reexecute state for them here
1019 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
1020 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
1021 #ifdef ASSERT
1022 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
1023 assert(method() == youngest_jvms->method(), "sanity");
1024 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
1025 // TODO 8371125
1026 // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
1027 #endif // ASSERT
1028 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
1029 }
1030
1031 // Presize the call:
1032 DEBUG_ONLY(uint non_debug_edges = call->req());
1033 call->add_req_batch(top(), youngest_jvms->debug_depth());
1034 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
1035
1036 // Set up edges so that the call looks like this:
1037 // Call [state:] ctl io mem fptr retadr
1038 // [parms:] parm0 ... parmN
1039 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1040 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
1041 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1042 // Note that caller debug info precedes callee debug info.
1043
1044 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1045 uint debug_ptr = call->req();
1046
1047 // Loop over the map input edges associated with jvms, add them
1048 // to the call node, & reset all offsets to match call node array.
1049
1050 JVMState* callee_jvms = nullptr;
1051 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1052 uint debug_end = debug_ptr;
1053 uint debug_start = debug_ptr - in_jvms->debug_size();
1054 debug_ptr = debug_start; // back up the ptr
1055
1056 uint p = debug_start; // walks forward in [debug_start, debug_end)
1057 uint j, k, l;
1058 SafePointNode* in_map = in_jvms->map();
1059 out_jvms->set_map(call);
1060
1061 if (can_prune_locals) {
1062 assert(in_jvms->method() == out_jvms->method(), "sanity");
1063 // If the current throw can reach an exception handler in this JVMS,
1064 // then we must keep everything live that can reach that handler.
1065 // As a quick and dirty approximation, we look for any handlers at all.
1066 if (in_jvms->method()->has_exception_handlers()) {
1067 can_prune_locals = false;
1068 }
1069 }
1070
1071 // Add the Locals
1072 k = in_jvms->locoff();
1073 l = in_jvms->loc_size();
1074 out_jvms->set_locoff(p);
1075 if (!can_prune_locals) {
1076 for (j = 0; j < l; j++) {
1077 call->set_req(p++, in_map->in(k + j));
1078 }
1079 } else {
1080 p += l; // already set to top above by add_req_batch
1081 }
1082
1083 // Add the Expression Stack
1084 k = in_jvms->stkoff();
1085 l = in_jvms->sp();
1086 out_jvms->set_stkoff(p);
1087 if (!can_prune_locals) {
1088 for (j = 0; j < l; j++) {
1089 call->set_req(p++, in_map->in(k + j));
1090 }
1091 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1092 // Divide stack into {S0,...,S1}, where S0 is set to top.
1093 uint s1 = stack_slots_not_pruned;
1094 stack_slots_not_pruned = 0; // for next iteration
1095 if (s1 > l) s1 = l;
1096 uint s0 = l - s1;
1097 p += s0; // skip the tops preinstalled by add_req_batch
1098 for (j = s0; j < l; j++)
1099 call->set_req(p++, in_map->in(k+j));
1100 } else {
1101 p += l; // already set to top above by add_req_batch
1102 }
1103
1104 // Add the Monitors
1105 k = in_jvms->monoff();
1106 l = in_jvms->mon_size();
1107 out_jvms->set_monoff(p);
1108 for (j = 0; j < l; j++)
1109 call->set_req(p++, in_map->in(k+j));
1110
1111 // Copy any scalar object fields.
1112 k = in_jvms->scloff();
1113 l = in_jvms->scl_size();
1114 out_jvms->set_scloff(p);
1115 for (j = 0; j < l; j++)
1116 call->set_req(p++, in_map->in(k+j));
1117
1118 // Finish the new jvms.
1119 out_jvms->set_endoff(p);
1120
1121 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1122 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1123 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1124 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1125 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1126 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1127
1128 // Update the two tail pointers in parallel.
1129 callee_jvms = out_jvms;
1130 out_jvms = out_jvms->caller();
1131 in_jvms = in_jvms->caller();
1132 }
1133
1134 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1135
1136 // Test the correctness of JVMState::debug_xxx accessors:
1137 assert(call->jvms()->debug_start() == non_debug_edges, "");
1138 assert(call->jvms()->debug_end() == call->req(), "");
1139 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1140 }
1141
1142 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1143 Bytecodes::Code code = java_bc();
1144 if (code == Bytecodes::_wide) {
1145 code = method()->java_code_at_bci(bci() + 1);
1146 }
1147
1148 if (code != Bytecodes::_illegal) {
1149 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1150 }
1151
1152 auto rsize = [&]() {
1153 assert(code != Bytecodes::_illegal, "code is illegal!");
1154 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1155 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1156 };
1157
1158 switch (code) {
1159 case Bytecodes::_illegal:
1160 return false;
1161
1162 case Bytecodes::_ldc:
1163 case Bytecodes::_ldc_w:
1164 case Bytecodes::_ldc2_w:
1165 inputs = 0;
1166 break;
1167
1168 case Bytecodes::_dup: inputs = 1; break;
1169 case Bytecodes::_dup_x1: inputs = 2; break;
1170 case Bytecodes::_dup_x2: inputs = 3; break;
1171 case Bytecodes::_dup2: inputs = 2; break;
1172 case Bytecodes::_dup2_x1: inputs = 3; break;
1173 case Bytecodes::_dup2_x2: inputs = 4; break;
1174 case Bytecodes::_swap: inputs = 2; break;
1175 case Bytecodes::_arraylength: inputs = 1; break;
1176
1177 case Bytecodes::_getstatic:
1178 case Bytecodes::_putstatic:
1179 case Bytecodes::_getfield:
1180 case Bytecodes::_putfield:
1181 {
1182 bool ignored_will_link;
1183 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1184 int size = field->type()->size();
1185 bool is_get = (depth >= 0), is_static = (depth & 1);
1186 inputs = (is_static ? 0 : 1);
1187 if (is_get) {
1188 depth = size - inputs;
1189 } else {
1190 inputs += size; // putxxx pops the value from the stack
1191 depth = - inputs;
1192 }
1193 }
1194 break;
1195
1196 case Bytecodes::_invokevirtual:
1197 case Bytecodes::_invokespecial:
1198 case Bytecodes::_invokestatic:
1199 case Bytecodes::_invokedynamic:
1200 case Bytecodes::_invokeinterface:
1201 {
1202 bool ignored_will_link;
1203 ciSignature* declared_signature = nullptr;
1204 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1205 assert(declared_signature != nullptr, "cannot be null");
1206 inputs = declared_signature->arg_size_for_bc(code);
1207 int size = declared_signature->return_type()->size();
1208 depth = size - inputs;
1209 }
1210 break;
1211
1212 case Bytecodes::_multianewarray:
1213 {
1214 ciBytecodeStream iter(method());
1215 iter.reset_to_bci(bci());
1216 iter.next();
1217 inputs = iter.get_dimensions();
1218 assert(rsize() == 1, "");
1219 depth = 1 - inputs;
1220 }
1221 break;
1222
1223 case Bytecodes::_ireturn:
1224 case Bytecodes::_lreturn:
1225 case Bytecodes::_freturn:
1226 case Bytecodes::_dreturn:
1227 case Bytecodes::_areturn:
1228 assert(rsize() == -depth, "");
1229 inputs = -depth;
1230 break;
1231
1232 case Bytecodes::_jsr:
1233 case Bytecodes::_jsr_w:
1234 inputs = 0;
1235 depth = 1; // S.B. depth=1, not zero
1236 break;
1237
1238 default:
1239 // bytecode produces a typed result
1240 inputs = rsize() - depth;
1241 assert(inputs >= 0, "");
1242 break;
1243 }
1244
1245 #ifdef ASSERT
1246 // spot check
1247 int outputs = depth + inputs;
1248 assert(outputs >= 0, "sanity");
1249 switch (code) {
1250 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1251 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1252 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1253 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1254 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1255 default: break;
1256 }
1257 #endif //ASSERT
1258
1259 return true;
1260 }
1261
1262
1263
1264 //------------------------------basic_plus_adr---------------------------------
1265 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1266 // short-circuit a common case
1267 if (offset == MakeConX(0)) {
1268 return ptr;
1269 }
1270 #ifdef ASSERT
1271 // Both 32-bit and 64-bit zeros should have been handled by the previous `if`
1272 // statement, so if we see either 32-bit or 64-bit zeros here, then we have a
1273 // problem.
1274 if (offset->is_Con()) {
1275 const Type* t = offset->bottom_type();
1276 bool is_zero_int = t->isa_int() && t->is_int()->get_con() == 0;
1277 bool is_zero_long = t->isa_long() && t->is_long()->get_con() == 0;
1278 assert(!is_zero_int && !is_zero_long,
1279 "Unexpected zero offset - should have matched MakeConX(0)");
1280 }
1281 #endif
1282 return _gvn.transform(AddPNode::make_with_base(base, ptr, offset));
1283 }
1284
1285 Node* GraphKit::ConvI2L(Node* offset) {
1286 // short-circuit a common case
1287 jint offset_con = find_int_con(offset, Type::OffsetBot);
1288 if (offset_con != Type::OffsetBot) {
1289 return longcon((jlong) offset_con);
1290 }
1291 return _gvn.transform( new ConvI2LNode(offset));
1292 }
1293
1294 Node* GraphKit::ConvI2UL(Node* offset) {
1295 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1296 if (offset_con != (juint) Type::OffsetBot) {
1297 return longcon((julong) offset_con);
1298 }
1299 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1300 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1301 return _gvn.transform( new AndLNode(conv, mask) );
1302 }
1303
1304 Node* GraphKit::ConvL2I(Node* offset) {
1305 // short-circuit a common case
1306 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1307 if (offset_con != (jlong)Type::OffsetBot) {
1308 return intcon((int) offset_con);
1309 }
1310 return _gvn.transform( new ConvL2INode(offset));
1311 }
1312
1313 //-------------------------load_object_klass-----------------------------------
1314 Node* GraphKit::load_object_klass(Node* obj) {
1315 // Special-case a fresh allocation to avoid building nodes:
1316 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1317 if (akls != nullptr) return akls;
1318 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1319 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1320 }
1321
1322 //-------------------------load_array_length-----------------------------------
1323 Node* GraphKit::load_array_length(Node* array) {
1324 // Special-case a fresh allocation to avoid building nodes:
1325 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1326 Node *alen;
1327 if (alloc == nullptr) {
1328 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1329 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1330 } else {
1331 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1332 }
1333 return alen;
1334 }
1335
1336 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1337 const TypeOopPtr* oop_type,
1338 bool replace_length_in_map) {
1339 Node* length = alloc->Ideal_length();
1340 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1341 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1342 if (ccast != length) {
1343 // do not transform ccast here, it might convert to top node for
1344 // negative array length and break assumptions in parsing stage.
1345 _gvn.set_type_bottom(ccast);
1346 record_for_igvn(ccast);
1347 if (replace_length_in_map) {
1348 replace_in_map(length, ccast);
1349 }
1350 return ccast;
1351 }
1352 }
1353 return length;
1354 }
1355
1356 //------------------------------do_null_check----------------------------------
1357 // Helper function to do a null pointer check. Returned value is
1358 // the incoming address with null casted away. You are allowed to use the
1359 // not-null value only if you are control dependent on the test.
1360 #ifndef PRODUCT
1361 extern uint explicit_null_checks_inserted,
1362 explicit_null_checks_elided;
1363 #endif
1364 Node* GraphKit::null_check_common(Node* value, BasicType type,
1365 // optional arguments for variations:
1366 bool assert_null,
1367 Node* *null_control,
1368 bool speculative,
1369 bool null_marker_check) {
1370 assert(!assert_null || null_control == nullptr, "not both at once");
1371 if (stopped()) return top();
1372 NOT_PRODUCT(explicit_null_checks_inserted++);
1373
1374 if (value->is_InlineType()) {
1375 // Null checking a scalarized but nullable inline type. Check the null marker
1376 // input instead of the oop input to avoid keeping buffer allocations alive.
1377 InlineTypeNode* vtptr = value->as_InlineType();
1378 while (vtptr->get_oop()->is_InlineType()) {
1379 vtptr = vtptr->get_oop()->as_InlineType();
1380 }
1381 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1382 if (stopped()) {
1383 return top();
1384 }
1385 if (assert_null) {
1386 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1387 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1388 // replace_in_map(value, vtptr);
1389 // return vtptr;
1390 replace_in_map(value, null());
1391 return null();
1392 }
1393 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1394 return cast_not_null(value, do_replace_in_map);
1395 }
1396
1397 // Construct null check
1398 Node *chk = nullptr;
1399 switch(type) {
1400 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1401 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1402 case T_ARRAY : // fall through
1403 type = T_OBJECT; // simplify further tests
1404 case T_OBJECT : {
1405 const Type *t = _gvn.type( value );
1406
1407 const TypeOopPtr* tp = t->isa_oopptr();
1408 if (tp != nullptr && !tp->is_loaded()
1409 // Only for do_null_check, not any of its siblings:
1410 && !assert_null && null_control == nullptr) {
1411 // Usually, any field access or invocation on an unloaded oop type
1412 // will simply fail to link, since the statically linked class is
1413 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1414 // the static class is loaded but the sharper oop type is not.
1415 // Rather than checking for this obscure case in lots of places,
1416 // we simply observe that a null check on an unloaded class
1417 // will always be followed by a nonsense operation, so we
1418 // can just issue the uncommon trap here.
1419 // Our access to the unloaded class will only be correct
1420 // after it has been loaded and initialized, which requires
1421 // a trip through the interpreter.
1422 ciKlass* klass = tp->unloaded_klass();
1423 #ifndef PRODUCT
1424 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1425 #endif
1426 uncommon_trap(Deoptimization::Reason_unloaded,
1427 Deoptimization::Action_reinterpret,
1428 klass, "!loaded");
1429 return top();
1430 }
1431
1432 if (assert_null) {
1433 // See if the type is contained in NULL_PTR.
1434 // If so, then the value is already null.
1435 if (t->higher_equal(TypePtr::NULL_PTR)) {
1436 NOT_PRODUCT(explicit_null_checks_elided++);
1437 return value; // Elided null assert quickly!
1438 }
1439 } else {
1440 // See if mixing in the null pointer changes type.
1441 // If so, then the null pointer was not allowed in the original
1442 // type. In other words, "value" was not-null.
1443 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1444 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1445 NOT_PRODUCT(explicit_null_checks_elided++);
1446 return value; // Elided null check quickly!
1447 }
1448 }
1449 chk = new CmpPNode( value, null() );
1450 break;
1451 }
1452
1453 default:
1454 fatal("unexpected type: %s", type2name(type));
1455 }
1456 assert(chk != nullptr, "sanity check");
1457 chk = _gvn.transform(chk);
1458
1459 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1460 BoolNode *btst = new BoolNode( chk, btest);
1461 Node *tst = _gvn.transform( btst );
1462
1463 //-----------
1464 // if peephole optimizations occurred, a prior test existed.
1465 // If a prior test existed, maybe it dominates as we can avoid this test.
1466 if (tst != btst && type == T_OBJECT) {
1467 // At this point we want to scan up the CFG to see if we can
1468 // find an identical test (and so avoid this test altogether).
1469 Node *cfg = control();
1470 int depth = 0;
1471 while( depth < 16 ) { // Limit search depth for speed
1472 if( cfg->Opcode() == Op_IfTrue &&
1473 cfg->in(0)->in(1) == tst ) {
1474 // Found prior test. Use "cast_not_null" to construct an identical
1475 // CastPP (and hence hash to) as already exists for the prior test.
1476 // Return that casted value.
1477 if (assert_null) {
1478 replace_in_map(value, null());
1479 return null(); // do not issue the redundant test
1480 }
1481 Node *oldcontrol = control();
1482 set_control(cfg);
1483 Node *res = cast_not_null(value);
1484 set_control(oldcontrol);
1485 NOT_PRODUCT(explicit_null_checks_elided++);
1486 return res;
1487 }
1488 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1489 if (cfg == nullptr) break; // Quit at region nodes
1490 depth++;
1491 }
1492 }
1493
1494 //-----------
1495 // Branch to failure if null
1496 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1497 Deoptimization::DeoptReason reason;
1498 if (assert_null) {
1499 reason = Deoptimization::reason_null_assert(speculative);
1500 } else if (type == T_OBJECT || null_marker_check) {
1501 reason = Deoptimization::reason_null_check(speculative);
1502 } else {
1503 reason = Deoptimization::Reason_div0_check;
1504 }
1505 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1506 // ciMethodData::has_trap_at will return a conservative -1 if any
1507 // must-be-null assertion has failed. This could cause performance
1508 // problems for a method after its first do_null_assert failure.
1509 // Consider using 'Reason_class_check' instead?
1510
1511 // To cause an implicit null check, we set the not-null probability
1512 // to the maximum (PROB_MAX). For an explicit check the probability
1513 // is set to a smaller value.
1514 if (null_control != nullptr || too_many_traps(reason)) {
1515 // probability is less likely
1516 ok_prob = PROB_LIKELY_MAG(3);
1517 } else if (!assert_null &&
1518 (ImplicitNullCheckThreshold > 0) &&
1519 method() != nullptr &&
1520 (method()->method_data()->trap_count(reason)
1521 >= (uint)ImplicitNullCheckThreshold)) {
1522 ok_prob = PROB_LIKELY_MAG(3);
1523 }
1524
1525 if (null_control != nullptr) {
1526 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1527 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1528 set_control( _gvn.transform( new IfTrueNode(iff)));
1529 #ifndef PRODUCT
1530 if (null_true == top()) {
1531 explicit_null_checks_elided++;
1532 }
1533 #endif
1534 (*null_control) = null_true;
1535 } else {
1536 BuildCutout unless(this, tst, ok_prob);
1537 // Check for optimizer eliding test at parse time
1538 if (stopped()) {
1539 // Failure not possible; do not bother making uncommon trap.
1540 NOT_PRODUCT(explicit_null_checks_elided++);
1541 } else if (assert_null) {
1542 uncommon_trap(reason,
1543 Deoptimization::Action_make_not_entrant,
1544 nullptr, "assert_null");
1545 } else {
1546 replace_in_map(value, zerocon(type));
1547 builtin_throw(reason);
1548 }
1549 }
1550
1551 // Must throw exception, fall-thru not possible?
1552 if (stopped()) {
1553 return top(); // No result
1554 }
1555
1556 if (assert_null) {
1557 // Cast obj to null on this path.
1558 replace_in_map(value, zerocon(type));
1559 return zerocon(type);
1560 }
1561
1562 // Cast obj to not-null on this path, if there is no null_control.
1563 // (If there is a null_control, a non-null value may come back to haunt us.)
1564 if (type == T_OBJECT) {
1565 Node* cast = cast_not_null(value, false);
1566 if (null_control == nullptr || (*null_control) == top())
1567 replace_in_map(value, cast);
1568 value = cast;
1569 }
1570
1571 return value;
1572 }
1573
1574 //------------------------------cast_not_null----------------------------------
1575 // Cast obj to not-null on this path
1576 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1577 if (obj->is_InlineType()) {
1578 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1579 vt->as_InlineType()->set_null_marker(_gvn);
1580 vt = _gvn.transform(vt);
1581 if (do_replace_in_map) {
1582 replace_in_map(obj, vt);
1583 }
1584 return vt;
1585 }
1586 const Type *t = _gvn.type(obj);
1587 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1588 // Object is already not-null?
1589 if( t == t_not_null ) return obj;
1590
1591 Node* cast = new CastPPNode(control(), obj,t_not_null);
1592 cast = _gvn.transform( cast );
1593
1594 // Scan for instances of 'obj' in the current JVM mapping.
1595 // These instances are known to be not-null after the test.
1596 if (do_replace_in_map)
1597 replace_in_map(obj, cast);
1598
1599 return cast; // Return casted value
1600 }
1601
1602 // Sometimes in intrinsics, we implicitly know an object is not null
1603 // (there's no actual null check) so we can cast it to not null. In
1604 // the course of optimizations, the input to the cast can become null.
1605 // In that case that data path will die and we need the control path
1606 // to become dead as well to keep the graph consistent. So we have to
1607 // add a check for null for which one branch can't be taken. It uses
1608 // an OpaqueConstantBool node that will cause the check to be removed after loop
1609 // opts so the test goes away and the compiled code doesn't execute a
1610 // useless check.
1611 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1612 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1613 return value;
1614 }
1615 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1616 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1617 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1618 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1619 _gvn.set_type(iff, iff->Value(&_gvn));
1620 if (!tst->is_Con()) {
1621 record_for_igvn(iff);
1622 }
1623 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1624 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1625 halt(if_f, frame, "unexpected null in intrinsic");
1626 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1627 set_control(if_t);
1628 return cast_not_null(value, do_replace_in_map);
1629 }
1630
1631
1632 //--------------------------replace_in_map-------------------------------------
1633 void GraphKit::replace_in_map(Node* old, Node* neww) {
1634 if (old == neww) {
1635 return;
1636 }
1637
1638 map()->replace_edge(old, neww);
1639
1640 // Note: This operation potentially replaces any edge
1641 // on the map. This includes locals, stack, and monitors
1642 // of the current (innermost) JVM state.
1643
1644 // don't let inconsistent types from profiling escape this
1645 // method
1646
1647 const Type* told = _gvn.type(old);
1648 const Type* tnew = _gvn.type(neww);
1649
1650 if (!tnew->higher_equal(told)) {
1651 return;
1652 }
1653
1654 map()->record_replaced_node(old, neww);
1655 }
1656
1657
1658 //=============================================================================
1659 //--------------------------------memory---------------------------------------
1660 Node* GraphKit::memory(uint alias_idx) {
1661 MergeMemNode* mem = merged_memory();
1662 Node* p = mem->memory_at(alias_idx);
1663 assert(p != mem->empty_memory(), "empty");
1664 _gvn.set_type(p, Type::MEMORY); // must be mapped
1665 return p;
1666 }
1667
1668 //-----------------------------reset_memory------------------------------------
1669 Node* GraphKit::reset_memory() {
1670 Node* mem = map()->memory();
1671 // do not use this node for any more parsing!
1672 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1673 return _gvn.transform( mem );
1674 }
1675
1676 //------------------------------set_all_memory---------------------------------
1677 void GraphKit::set_all_memory(Node* newmem) {
1678 Node* mergemem = MergeMemNode::make(newmem);
1679 gvn().set_type_bottom(mergemem);
1680 if (_gvn.is_IterGVN() != nullptr) {
1681 record_for_igvn(mergemem);
1682 }
1683 map()->set_memory(mergemem);
1684 }
1685
1686 //------------------------------set_all_memory_call----------------------------
1687 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1688 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1689 set_all_memory(newmem);
1690 }
1691
1692 //=============================================================================
1693 //
1694 // parser factory methods for MemNodes
1695 //
1696 // These are layered on top of the factory methods in LoadNode and StoreNode,
1697 // and integrate with the parser's memory state and _gvn engine.
1698 //
1699
1700 // factory methods in "int adr_idx"
1701 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1702 MemNode::MemOrd mo,
1703 LoadNode::ControlDependency control_dependency,
1704 bool require_atomic_access,
1705 bool unaligned,
1706 bool mismatched,
1707 bool unsafe,
1708 uint8_t barrier_data) {
1709 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1710 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1711 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1712 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1713 Node* mem = memory(adr_idx);
1714 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1715 ld = _gvn.transform(ld);
1716
1717 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1718 // Improve graph before escape analysis and boxing elimination.
1719 record_for_igvn(ld);
1720 if (ld->is_DecodeN()) {
1721 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1722 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1723 // a Phi). Recording such cases is still perfectly sound, but may be
1724 // unnecessary and result in some minor IGVN overhead.
1725 record_for_igvn(ld->in(1));
1726 }
1727 }
1728 return ld;
1729 }
1730
1731 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1732 MemNode::MemOrd mo,
1733 bool require_atomic_access,
1734 bool unaligned,
1735 bool mismatched,
1736 bool unsafe,
1737 int barrier_data) {
1738 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1739 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1740 const TypePtr* adr_type = nullptr;
1741 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1742 Node *mem = memory(adr_idx);
1743 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1744 if (unaligned) {
1745 st->as_Store()->set_unaligned_access();
1746 }
1747 if (mismatched) {
1748 st->as_Store()->set_mismatched_access();
1749 }
1750 if (unsafe) {
1751 st->as_Store()->set_unsafe_access();
1752 }
1753 st->as_Store()->set_barrier_data(barrier_data);
1754 st = _gvn.transform(st);
1755 set_memory(st, adr_idx);
1756 // Back-to-back stores can only remove intermediate store with DU info
1757 // so push on worklist for optimizer.
1758 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1759 record_for_igvn(st);
1760
1761 return st;
1762 }
1763
1764 Node* GraphKit::access_store_at(Node* obj,
1765 Node* adr,
1766 const TypePtr* adr_type,
1767 Node* val,
1768 const Type* val_type,
1769 BasicType bt,
1770 DecoratorSet decorators,
1771 bool safe_for_replace,
1772 const InlineTypeNode* vt) {
1773 // Transformation of a value which could be null pointer (CastPP #null)
1774 // could be delayed during Parse (for example, in adjust_map_after_if()).
1775 // Execute transformation here to avoid barrier generation in such case.
1776 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1777 val = _gvn.makecon(TypePtr::NULL_PTR);
1778 }
1779
1780 if (stopped()) {
1781 return top(); // Dead path ?
1782 }
1783
1784 assert(val != nullptr, "not dead path");
1785 if (val->is_InlineType()) {
1786 // Store to non-flat field. Buffer the inline type and make sure
1787 // the store is re-executed if the allocation triggers deoptimization.
1788 PreserveReexecuteState preexecs(this);
1789 jvms()->set_should_reexecute(true);
1790 val = val->as_InlineType()->buffer(this, safe_for_replace);
1791 }
1792
1793 C2AccessValuePtr addr(adr, adr_type);
1794 C2AccessValue value(val, val_type);
1795 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1796 if (access.is_raw()) {
1797 return _barrier_set->BarrierSetC2::store_at(access, value);
1798 } else {
1799 return _barrier_set->store_at(access, value);
1800 }
1801 }
1802
1803 Node* GraphKit::access_load_at(Node* obj, // containing obj
1804 Node* adr, // actual address to store val at
1805 const TypePtr* adr_type,
1806 const Type* val_type,
1807 BasicType bt,
1808 DecoratorSet decorators,
1809 Node* ctl) {
1810 if (stopped()) {
1811 return top(); // Dead path ?
1812 }
1813
1814 SavedState old_state(this);
1815 C2AccessValuePtr addr(adr, adr_type);
1816 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1817 Node* load;
1818 if (access.is_raw()) {
1819 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1820 } else {
1821 load = _barrier_set->load_at(access, val_type);
1822 }
1823
1824 // Restore the previous state only if the load got folded to a constant
1825 // and we can discard any barriers that might have been added.
1826 if (load == nullptr || !load->is_Con()) {
1827 old_state.discard();
1828 }
1829 return load;
1830 }
1831
1832 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1833 const Type* val_type,
1834 BasicType bt,
1835 DecoratorSet decorators) {
1836 if (stopped()) {
1837 return top(); // Dead path ?
1838 }
1839
1840 SavedState old_state(this);
1841 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1842 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1843 Node* load;
1844 if (access.is_raw()) {
1845 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1846 } else {
1847 load = _barrier_set->load_at(access, val_type);
1848 }
1849
1850 // Restore the previous state only if the load got folded to a constant
1851 // and we can discard any barriers that might have been added.
1852 if (load == nullptr || !load->is_Con()) {
1853 old_state.discard();
1854 }
1855 return load;
1856 }
1857
1858 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1859 Node* adr,
1860 const TypePtr* adr_type,
1861 int alias_idx,
1862 Node* expected_val,
1863 Node* new_val,
1864 const Type* value_type,
1865 BasicType bt,
1866 DecoratorSet decorators) {
1867 C2AccessValuePtr addr(adr, adr_type);
1868 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1869 bt, obj, addr, alias_idx);
1870 if (access.is_raw()) {
1871 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1872 } else {
1873 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1874 }
1875 }
1876
1877 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1878 Node* adr,
1879 const TypePtr* adr_type,
1880 int alias_idx,
1881 Node* expected_val,
1882 Node* new_val,
1883 const Type* value_type,
1884 BasicType bt,
1885 DecoratorSet decorators) {
1886 C2AccessValuePtr addr(adr, adr_type);
1887 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1888 bt, obj, addr, alias_idx);
1889 if (access.is_raw()) {
1890 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1891 } else {
1892 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1893 }
1894 }
1895
1896 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1897 Node* adr,
1898 const TypePtr* adr_type,
1899 int alias_idx,
1900 Node* new_val,
1901 const Type* value_type,
1902 BasicType bt,
1903 DecoratorSet decorators) {
1904 C2AccessValuePtr addr(adr, adr_type);
1905 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1906 bt, obj, addr, alias_idx);
1907 if (access.is_raw()) {
1908 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1909 } else {
1910 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1911 }
1912 }
1913
1914 Node* GraphKit::access_atomic_add_at(Node* obj,
1915 Node* adr,
1916 const TypePtr* adr_type,
1917 int alias_idx,
1918 Node* new_val,
1919 const Type* value_type,
1920 BasicType bt,
1921 DecoratorSet decorators) {
1922 C2AccessValuePtr addr(adr, adr_type);
1923 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1924 if (access.is_raw()) {
1925 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1926 } else {
1927 return _barrier_set->atomic_add_at(access, new_val, value_type);
1928 }
1929 }
1930
1931 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1932 return _barrier_set->clone(this, src, dst, size, is_array);
1933 }
1934
1935 //-------------------------array_element_address-------------------------
1936 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1937 const TypeInt* sizetype, Node* ctrl) {
1938 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1939 uint shift;
1940 uint header;
1941 if (arytype->is_flat() && arytype->klass_is_exact()) {
1942 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1943 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1944 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1945 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1946 // though we don't need the address node in this case and throw it away again.
1947 shift = arytype->flat_log_elem_size();
1948 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1949 } else {
1950 shift = exact_log2(type2aelembytes(elembt));
1951 header = arrayOopDesc::base_offset_in_bytes(elembt);
1952 }
1953
1954 // short-circuit a common case (saves lots of confusing waste motion)
1955 jint idx_con = find_int_con(idx, -1);
1956 if (idx_con >= 0) {
1957 intptr_t offset = header + ((intptr_t)idx_con << shift);
1958 return basic_plus_adr(ary, offset);
1959 }
1960
1961 // must be correct type for alignment purposes
1962 Node* base = basic_plus_adr(ary, header);
1963 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1964 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1965 return basic_plus_adr(ary, base, scale);
1966 }
1967
1968 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1969 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1970 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1971 return cast_to_flat_array_exact(array, elem_vk, true, false);
1972 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1973 return cast_to_flat_array_exact(array, elem_vk, true, true);
1974 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1975 return cast_to_flat_array_exact(array, elem_vk, false, true);
1976 }
1977
1978 bool is_null_free = false;
1979 if (!elem_vk->has_nullable_atomic_layout()) {
1980 // Element does not have a nullable flat layout, cannot be nullable
1981 is_null_free = true;
1982 }
1983
1984 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1985 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1986 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1987 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1988 }
1989
1990 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1991 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1992 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1993 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1994 assert(arytype->klass_is_exact(), "inconsistency");
1995 assert(arytype->is_flat(), "inconsistency");
1996 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1997 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1998 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1999 }
2000
2001 //-------------------------load_array_element-------------------------
2002 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
2003 const Type* elemtype = arytype->elem();
2004 BasicType elembt = elemtype->array_element_basic_type();
2005 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
2006 if (elembt == T_NARROWOOP) {
2007 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
2008 }
2009 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2010 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2011 return ld;
2012 }
2013
2014 //-------------------------set_arguments_for_java_call-------------------------
2015 // Arguments (pre-popped from the stack) are taken from the JVMS.
2016 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2017 PreserveReexecuteState preexecs(this);
2018 if (Arguments::is_valhalla_enabled()) {
2019 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2020 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2021 jvms()->set_should_reexecute(true);
2022 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2023 inc_sp(arg_size);
2024 }
2025 // Add the call arguments
2026 const TypeTuple* domain = call->tf()->domain_sig();
2027 uint nargs = domain->cnt();
2028 int arg_num = 0;
2029 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2030 uint arg_idx = i - TypeFunc::Parms;
2031 Node* arg = argument(arg_idx);
2032 const Type* t = domain->field_at(i);
2033 // TODO 8284443 A static call to a mismatched method should still be scalarized
2034 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2035 // We don't pass inline type arguments by reference but instead pass each field of the inline type
2036 if (!arg->is_InlineType()) {
2037 // There are 2 cases in which the argument has not been scalarized
2038 if (_gvn.type(arg)->is_zero_type()) {
2039 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2040 } else {
2041 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2042 // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2043 // type must be an exact value class. The method is devirtualized, and replaced with
2044 // a direct call with a scalarized receiver instead.
2045 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2046 assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2047 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2048 }
2049 }
2050 InlineTypeNode* vt = arg->as_InlineType();
2051 vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2052 // If an inline type argument is passed as fields, attach the Method* to the call site
2053 // to be able to access the extended signature later via attached_method_before_pc().
2054 // For example, see CompiledMethod::preserve_callee_argument_oops().
2055 call->set_override_symbolic_info(true);
2056 // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2057 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2058 C->dependencies()->assert_mismatch_calling_convention(call->method());
2059 arg_num++;
2060 continue;
2061 } else if (arg->is_InlineType()) {
2062 // Pass inline type argument via oop to callee
2063 arg = arg->as_InlineType()->buffer(this, true);
2064 }
2065 if (t != Type::HALF) {
2066 arg_num++;
2067 }
2068 call->init_req(idx++, arg);
2069 }
2070 }
2071
2072 //---------------------------set_edges_for_java_call---------------------------
2073 // Connect a newly created call into the current JVMS.
2074 // A return value node (if any) is returned from set_edges_for_java_call.
2075 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2076
2077 // Add the predefined inputs:
2078 call->init_req( TypeFunc::Control, control() );
2079 call->init_req( TypeFunc::I_O , i_o() );
2080 call->init_req( TypeFunc::Memory , reset_memory() );
2081 call->init_req( TypeFunc::FramePtr, frameptr() );
2082 call->init_req( TypeFunc::ReturnAdr, top() );
2083
2084 add_safepoint_edges(call, must_throw);
2085
2086 Node* xcall = _gvn.transform(call);
2087
2088 if (xcall == top()) {
2089 set_control(top());
2090 return;
2091 }
2092 assert(xcall == call, "call identity is stable");
2093
2094 // Re-use the current map to produce the result.
2095
2096 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2097 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2098 set_all_memory_call(xcall, separate_io_proj);
2099
2100 //return xcall; // no need, caller already has it
2101 }
2102
2103 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2104 if (stopped()) return top(); // maybe the call folded up?
2105
2106 // Note: Since any out-of-line call can produce an exception,
2107 // we always insert an I_O projection from the call into the result.
2108
2109 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2110
2111 if (separate_io_proj) {
2112 // The caller requested separate projections be used by the fall
2113 // through and exceptional paths, so replace the projections for
2114 // the fall through path.
2115 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2116 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2117 }
2118
2119 // Capture the return value, if any.
2120 Node* ret;
2121 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2122 ret = top();
2123 } else if (call->tf()->returns_inline_type_as_fields()) {
2124 // Return of multiple values (inline type fields): we create a
2125 // InlineType node, each field is a projection from the call.
2126 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2127 uint base_input = TypeFunc::Parms;
2128 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2129 } else {
2130 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2131 ciType* t = call->method()->return_type();
2132 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2133 // The return type is unloaded but the callee might later be C2 compiled and then return
2134 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2135 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2136 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2137 IdealKit ideal(this);
2138 IdealVariable res(ideal);
2139 ideal.declarations_done();
2140 // Change return type of call to scalarized return
2141 const TypeFunc* tf = call->_tf;
2142 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2143 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2144 call->_tf = new_tf;
2145 _gvn.set_type(call, call->Value(&_gvn));
2146 _gvn.set_type(ret, ret->Value(&_gvn));
2147 // Don't add store to buffer call if we are strength reducing
2148 if (!C->strength_reduction()) {
2149 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2150 // Return value is null
2151 ideal.set(res, makecon(TypePtr::NULL_PTR));
2152 } ideal.else_(); {
2153 // Return value is non-null
2154 sync_kit(ideal);
2155
2156 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2157 OptoRuntime::store_inline_type_fields_Type(),
2158 StubRoutines::store_inline_type_fields_to_buf(),
2159 nullptr, TypePtr::BOTTOM, ret);
2160
2161 // We don't know how many values are returned. This assumes the
2162 // worst case, that all available registers are used.
2163 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2164 if (domain->field_at(i) == Type::HALF) {
2165 store_to_buf_call->init_req(i, top());
2166 continue;
2167 }
2168 Node* proj =_gvn.transform(new ProjNode(call, i));
2169 store_to_buf_call->init_req(i, proj);
2170 }
2171 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2172
2173 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2174 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2175 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2176
2177 ideal.set(res, buf);
2178 ideal.sync_kit(this);
2179 } ideal.end_if();
2180 } else {
2181 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2182 Node* proj =_gvn.transform(new ProjNode(call, i));
2183 }
2184 ideal.set(res, ret);
2185 }
2186 sync_kit(ideal);
2187 ret = _gvn.transform(ideal.value(res));
2188 } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2189 // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2190 // result of merging several call paths. If one of them is made of an actual call node that
2191 // returns an oop, we need to call make_from_oop here as well because we want InlineType
2192 // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2193 ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2194 }
2195 }
2196
2197 return ret;
2198 }
2199
2200 //--------------------set_predefined_input_for_runtime_call--------------------
2201 // Reading and setting the memory state is way conservative here.
2202 // The real problem is that I am not doing real Type analysis on memory,
2203 // so I cannot distinguish card mark stores from other stores. Across a GC
2204 // point the Store Barrier and the card mark memory has to agree. I cannot
2205 // have a card mark store and its barrier split across the GC point from
2206 // either above or below. Here I get that to happen by reading ALL of memory.
2207 // A better answer would be to separate out card marks from other memory.
2208 // For now, return the input memory state, so that it can be reused
2209 // after the call, if this call has restricted memory effects.
2210 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2211 // Set fixed predefined input arguments
2212 call->init_req(TypeFunc::Control, control());
2213 call->init_req(TypeFunc::I_O, top()); // does no i/o
2214 call->init_req(TypeFunc::ReturnAdr, top());
2215 if (call->is_CallLeafPure()) {
2216 call->init_req(TypeFunc::Memory, top());
2217 call->init_req(TypeFunc::FramePtr, top());
2218 return nullptr;
2219 } else {
2220 Node* memory = reset_memory();
2221 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2222 call->init_req(TypeFunc::Memory, m); // may gc ptrs
2223 call->init_req(TypeFunc::FramePtr, frameptr());
2224 return memory;
2225 }
2226 }
2227
2228 //-------------------set_predefined_output_for_runtime_call--------------------
2229 // Set control and memory (not i_o) from the call.
2230 // If keep_mem is not null, use it for the output state,
2231 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2232 // If hook_mem is null, this call produces no memory effects at all.
2233 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2234 // then only that memory slice is taken from the call.
2235 // In the last case, we must put an appropriate memory barrier before
2236 // the call, so as to create the correct anti-dependencies on loads
2237 // preceding the call.
2238 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2239 Node* keep_mem,
2240 const TypePtr* hook_mem) {
2241 // no i/o
2242 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2243 if (call->is_CallLeafPure()) {
2244 // Pure function have only control (for now) and data output, in particular
2245 // they don't touch the memory, so we don't want a memory proj that is set after.
2246 return;
2247 }
2248 if (keep_mem) {
2249 // First clone the existing memory state
2250 set_all_memory(keep_mem);
2251 if (hook_mem != nullptr) {
2252 // Make memory for the call
2253 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2254 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2255 // We also use hook_mem to extract specific effects from arraycopy stubs.
2256 set_memory(mem, hook_mem);
2257 }
2258 // ...else the call has NO memory effects.
2259
2260 // Make sure the call advertises its memory effects precisely.
2261 // This lets us build accurate anti-dependences in gcm.cpp.
2262 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2263 "call node must be constructed correctly");
2264 } else {
2265 assert(hook_mem == nullptr, "");
2266 // This is not a "slow path" call; all memory comes from the call.
2267 set_all_memory_call(call);
2268 }
2269 }
2270
2271 // Keep track of MergeMems feeding into other MergeMems
2272 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2273 if (!mem->is_MergeMem()) {
2274 return;
2275 }
2276 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2277 Node* use = i.get();
2278 if (use->is_MergeMem()) {
2279 wl.push(use);
2280 }
2281 }
2282 }
2283
2284 // Replace the call with the current state of the kit.
2285 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2286 JVMState* ejvms = nullptr;
2287 if (has_exceptions()) {
2288 ejvms = transfer_exceptions_into_jvms();
2289 }
2290
2291 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2292 ReplacedNodes replaced_nodes_exception;
2293 Node* ex_ctl = top();
2294
2295 SafePointNode* final_state = stop();
2296
2297 // Find all the needed outputs of this call
2298 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2299
2300 Unique_Node_List wl;
2301 Node* init_mem = call->in(TypeFunc::Memory);
2302 Node* final_mem = final_state->in(TypeFunc::Memory);
2303 Node* final_ctl = final_state->in(TypeFunc::Control);
2304 Node* final_io = final_state->in(TypeFunc::I_O);
2305
2306 // Replace all the old call edges with the edges from the inlining result
2307 if (callprojs->fallthrough_catchproj != nullptr) {
2308 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2309 }
2310 if (callprojs->fallthrough_memproj != nullptr) {
2311 if (final_mem->is_MergeMem()) {
2312 // Parser's exits MergeMem was not transformed but may be optimized
2313 final_mem = _gvn.transform(final_mem);
2314 }
2315 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2316 add_mergemem_users_to_worklist(wl, final_mem);
2317 }
2318 if (callprojs->fallthrough_ioproj != nullptr) {
2319 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2320 }
2321
2322 // Replace the result with the new result if it exists and is used
2323 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2324 // If the inlined code is dead, the result projections for an inline type returned as
2325 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2326 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2327 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2328 "unexpected number of results");
2329 // If we are doing strength reduction and the return type is not loaded we
2330 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2331 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2332 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2333 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2334 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2335 }
2336 } else {
2337 C->gvn_replace_by(callprojs->resproj[0], result);
2338 }
2339 }
2340
2341 if (ejvms == nullptr) {
2342 // No exception edges to simply kill off those paths
2343 if (callprojs->catchall_catchproj != nullptr) {
2344 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2345 }
2346 if (callprojs->catchall_memproj != nullptr) {
2347 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2348 }
2349 if (callprojs->catchall_ioproj != nullptr) {
2350 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2351 }
2352 // Replace the old exception object with top
2353 if (callprojs->exobj != nullptr) {
2354 C->gvn_replace_by(callprojs->exobj, C->top());
2355 }
2356 } else {
2357 GraphKit ekit(ejvms);
2358
2359 // Load my combined exception state into the kit, with all phis transformed:
2360 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2361 replaced_nodes_exception = ex_map->replaced_nodes();
2362
2363 Node* ex_oop = ekit.use_exception_state(ex_map);
2364
2365 if (callprojs->catchall_catchproj != nullptr) {
2366 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2367 ex_ctl = ekit.control();
2368 }
2369 if (callprojs->catchall_memproj != nullptr) {
2370 Node* ex_mem = ekit.reset_memory();
2371 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2372 add_mergemem_users_to_worklist(wl, ex_mem);
2373 }
2374 if (callprojs->catchall_ioproj != nullptr) {
2375 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2376 }
2377
2378 // Replace the old exception object with the newly created one
2379 if (callprojs->exobj != nullptr) {
2380 C->gvn_replace_by(callprojs->exobj, ex_oop);
2381 }
2382 }
2383
2384 // Disconnect the call from the graph
2385 call->disconnect_inputs(C);
2386 C->gvn_replace_by(call, C->top());
2387
2388 // Clean up any MergeMems that feed other MergeMems since the
2389 // optimizer doesn't like that.
2390 while (wl.size() > 0) {
2391 _gvn.transform(wl.pop());
2392 }
2393
2394 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2395 replaced_nodes.apply(C, final_ctl);
2396 }
2397 if (!ex_ctl->is_top() && do_replaced_nodes) {
2398 replaced_nodes_exception.apply(C, ex_ctl);
2399 }
2400 }
2401
2402
2403 //------------------------------increment_counter------------------------------
2404 // for statistics: increment a VM counter by 1
2405
2406 void GraphKit::increment_counter(address counter_addr) {
2407 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2408 increment_counter(adr1);
2409 }
2410
2411 void GraphKit::increment_counter(Node* counter_addr) {
2412 Node* ctrl = control();
2413 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2414 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2415 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2416 }
2417
2418 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2419 Node* halt = new HaltNode(ctrl, frameptr, reason
2420 PRODUCT_ONLY(COMMA generate_code_in_product));
2421 halt = _gvn.transform(halt);
2422 root()->add_req(halt);
2423 if (_gvn.is_IterGVN() != nullptr) {
2424 record_for_igvn(root());
2425 }
2426 }
2427
2428 //------------------------------uncommon_trap----------------------------------
2429 // Bail out to the interpreter in mid-method. Implemented by calling the
2430 // uncommon_trap blob. This helper function inserts a runtime call with the
2431 // right debug info.
2432 Node* GraphKit::uncommon_trap(int trap_request,
2433 ciKlass* klass, const char* comment,
2434 bool must_throw,
2435 bool keep_exact_action) {
2436 if (failing_internal()) {
2437 stop();
2438 }
2439 if (stopped()) return nullptr; // trap reachable?
2440
2441 // Note: If ProfileTraps is true, and if a deopt. actually
2442 // occurs here, the runtime will make sure an MDO exists. There is
2443 // no need to call method()->ensure_method_data() at this point.
2444
2445 // Set the stack pointer to the right value for reexecution:
2446 set_sp(reexecute_sp());
2447
2448 #ifdef ASSERT
2449 if (!must_throw) {
2450 // Make sure the stack has at least enough depth to execute
2451 // the current bytecode.
2452 int inputs, ignored_depth;
2453 if (compute_stack_effects(inputs, ignored_depth)) {
2454 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2455 Bytecodes::name(java_bc()), sp(), inputs);
2456 }
2457 }
2458 #endif
2459
2460 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2461 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2462
2463 switch (action) {
2464 case Deoptimization::Action_maybe_recompile:
2465 case Deoptimization::Action_reinterpret:
2466 // Temporary fix for 6529811 to allow virtual calls to be sure they
2467 // get the chance to go from mono->bi->mega
2468 if (!keep_exact_action &&
2469 Deoptimization::trap_request_index(trap_request) < 0 &&
2470 too_many_recompiles(reason)) {
2471 // This BCI is causing too many recompilations.
2472 if (C->log() != nullptr) {
2473 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2474 Deoptimization::trap_reason_name(reason),
2475 Deoptimization::trap_action_name(action));
2476 }
2477 action = Deoptimization::Action_none;
2478 trap_request = Deoptimization::make_trap_request(reason, action);
2479 } else {
2480 C->set_trap_can_recompile(true);
2481 }
2482 break;
2483 case Deoptimization::Action_make_not_entrant:
2484 C->set_trap_can_recompile(true);
2485 break;
2486 case Deoptimization::Action_none:
2487 case Deoptimization::Action_make_not_compilable:
2488 break;
2489 default:
2490 #ifdef ASSERT
2491 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2492 #endif
2493 break;
2494 }
2495
2496 if (TraceOptoParse) {
2497 char buf[100];
2498 tty->print_cr("Uncommon trap %s at bci:%d",
2499 Deoptimization::format_trap_request(buf, sizeof(buf),
2500 trap_request), bci());
2501 }
2502
2503 CompileLog* log = C->log();
2504 if (log != nullptr) {
2505 int kid = (klass == nullptr)? -1: log->identify(klass);
2506 log->begin_elem("uncommon_trap bci='%d'", bci());
2507 char buf[100];
2508 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2509 trap_request));
2510 if (kid >= 0) log->print(" klass='%d'", kid);
2511 if (comment != nullptr) log->print(" comment='%s'", comment);
2512 log->end_elem();
2513 }
2514
2515 // Make sure any guarding test views this path as very unlikely
2516 Node *i0 = control()->in(0);
2517 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2518 IfNode *iff = i0->as_If();
2519 float f = iff->_prob; // Get prob
2520 if (control()->Opcode() == Op_IfTrue) {
2521 if (f > PROB_UNLIKELY_MAG(4))
2522 iff->_prob = PROB_MIN;
2523 } else {
2524 if (f < PROB_LIKELY_MAG(4))
2525 iff->_prob = PROB_MAX;
2526 }
2527 }
2528
2529 // Clear out dead values from the debug info.
2530 kill_dead_locals();
2531
2532 // Now insert the uncommon trap subroutine call
2533 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2534 const TypePtr* no_memory_effects = nullptr;
2535 // Pass the index of the class to be loaded
2536 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2537 (must_throw ? RC_MUST_THROW : 0),
2538 OptoRuntime::uncommon_trap_Type(),
2539 call_addr, "uncommon_trap", no_memory_effects,
2540 intcon(trap_request));
2541 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2542 "must extract request correctly from the graph");
2543 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2544
2545 call->set_req(TypeFunc::ReturnAdr, returnadr());
2546 // The debug info is the only real input to this call.
2547
2548 // Halt-and-catch fire here. The above call should never return!
2549 // We only emit code for the HaltNode in debug, which is enough for
2550 // verifying correctness. In product, we don't want to emit it so
2551 // that we can save on code space. HaltNode often get folded because
2552 // the compiler can prove that the unreachable path is dead. But we
2553 // cannot generally expect that for uncommon traps, which are often
2554 // reachable and occasionally taken.
2555 halt(control(), frameptr(),
2556 "uncommon trap returned which should never happen",
2557 false /* don't emit code in product */);
2558 stop_and_kill_map();
2559 return call;
2560 }
2561
2562
2563 //--------------------------just_allocated_object------------------------------
2564 // Report the object that was just allocated.
2565 // It must be the case that there are no intervening safepoints.
2566 // We use this to determine if an object is so "fresh" that
2567 // it does not require card marks.
2568 Node* GraphKit::just_allocated_object(Node* current_control) {
2569 Node* ctrl = current_control;
2570 // Object::<init> is invoked after allocation, most of invoke nodes
2571 // will be reduced, but a region node is kept in parse time, we check
2572 // the pattern and skip the region node if it degraded to a copy.
2573 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2574 ctrl->as_Region()->is_copy()) {
2575 ctrl = ctrl->as_Region()->is_copy();
2576 }
2577 if (C->recent_alloc_ctl() == ctrl) {
2578 return C->recent_alloc_obj();
2579 }
2580 return nullptr;
2581 }
2582
2583
2584 /**
2585 * Record profiling data exact_kls for Node n with the type system so
2586 * that it can propagate it (speculation)
2587 *
2588 * @param n node that the type applies to
2589 * @param exact_kls type from profiling
2590 * @param maybe_null did profiling see null?
2591 *
2592 * @return node with improved type
2593 */
2594 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2595 const Type* current_type = _gvn.type(n);
2596 assert(UseTypeSpeculation, "type speculation must be on");
2597
2598 const TypePtr* speculative = current_type->speculative();
2599
2600 // Should the klass from the profile be recorded in the speculative type?
2601 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2602 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2603 const TypeOopPtr* xtype = tklass->as_instance_type();
2604 assert(xtype->klass_is_exact(), "Should be exact");
2605 // Any reason to believe n is not null (from this profiling or a previous one)?
2606 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2607 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2608 // record the new speculative type's depth
2609 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2610 speculative = speculative->with_inline_depth(jvms()->depth());
2611 } else if (current_type->would_improve_ptr(ptr_kind)) {
2612 // Profiling report that null was never seen so we can change the
2613 // speculative type to non null ptr.
2614 if (ptr_kind == ProfileAlwaysNull) {
2615 speculative = TypePtr::NULL_PTR;
2616 } else {
2617 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2618 const TypePtr* ptr = TypePtr::NOTNULL;
2619 if (speculative != nullptr) {
2620 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2621 } else {
2622 speculative = ptr;
2623 }
2624 }
2625 }
2626
2627 if (speculative != current_type->speculative()) {
2628 // Build a type with a speculative type (what we think we know
2629 // about the type but will need a guard when we use it)
2630 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2631 // We're changing the type, we need a new CheckCast node to carry
2632 // the new type. The new type depends on the control: what
2633 // profiling tells us is only valid from here as far as we can
2634 // tell.
2635 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2636 cast = _gvn.transform(cast);
2637 replace_in_map(n, cast);
2638 n = cast;
2639 }
2640
2641 return n;
2642 }
2643
2644 /**
2645 * Record profiling data from receiver profiling at an invoke with the
2646 * type system so that it can propagate it (speculation)
2647 *
2648 * @param n receiver node
2649 *
2650 * @return node with improved type
2651 */
2652 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2653 if (!UseTypeSpeculation) {
2654 return n;
2655 }
2656 ciKlass* exact_kls = profile_has_unique_klass();
2657 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2658 if ((java_bc() == Bytecodes::_checkcast ||
2659 java_bc() == Bytecodes::_instanceof ||
2660 java_bc() == Bytecodes::_aastore) &&
2661 method()->method_data()->is_mature()) {
2662 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2663 if (data != nullptr) {
2664 if (java_bc() == Bytecodes::_aastore) {
2665 ciKlass* array_type = nullptr;
2666 ciKlass* element_type = nullptr;
2667 ProfilePtrKind element_ptr = ProfileMaybeNull;
2668 bool flat_array = true;
2669 bool null_free_array = true;
2670 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2671 exact_kls = element_type;
2672 ptr_kind = element_ptr;
2673 } else {
2674 if (!data->as_BitData()->null_seen()) {
2675 ptr_kind = ProfileNeverNull;
2676 } else {
2677 if (TypeProfileCasts) {
2678 assert(data->is_ReceiverTypeData(), "bad profile data type");
2679 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2680 uint i = 0;
2681 for (; i < call->row_limit(); i++) {
2682 ciKlass* receiver = call->receiver(i);
2683 if (receiver != nullptr) {
2684 break;
2685 }
2686 }
2687 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2688 }
2689 }
2690 }
2691 }
2692 }
2693 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2694 }
2695
2696 /**
2697 * Record profiling data from argument profiling at an invoke with the
2698 * type system so that it can propagate it (speculation)
2699 *
2700 * @param dest_method target method for the call
2701 * @param bc what invoke bytecode is this?
2702 */
2703 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2704 if (!UseTypeSpeculation) {
2705 return;
2706 }
2707 const TypeFunc* tf = TypeFunc::make(dest_method);
2708 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2709 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2710 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2711 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2712 if (is_reference_type(targ->basic_type())) {
2713 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2714 ciKlass* better_type = nullptr;
2715 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2716 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2717 }
2718 i++;
2719 }
2720 }
2721 }
2722
2723 /**
2724 * Record profiling data from parameter profiling at an invoke with
2725 * the type system so that it can propagate it (speculation)
2726 */
2727 void GraphKit::record_profiled_parameters_for_speculation() {
2728 if (!UseTypeSpeculation) {
2729 return;
2730 }
2731 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2732 if (_gvn.type(local(i))->isa_oopptr()) {
2733 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2734 ciKlass* better_type = nullptr;
2735 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2736 record_profile_for_speculation(local(i), better_type, ptr_kind);
2737 }
2738 j++;
2739 }
2740 }
2741 }
2742
2743 /**
2744 * Record profiling data from return value profiling at an invoke with
2745 * the type system so that it can propagate it (speculation)
2746 */
2747 void GraphKit::record_profiled_return_for_speculation() {
2748 if (!UseTypeSpeculation) {
2749 return;
2750 }
2751 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2752 ciKlass* better_type = nullptr;
2753 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2754 // If profiling reports a single type for the return value,
2755 // feed it to the type system so it can propagate it as a
2756 // speculative type
2757 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2758 }
2759 }
2760
2761
2762 //=============================================================================
2763 // Generate a fast path/slow path idiom. Graph looks like:
2764 // [foo] indicates that 'foo' is a parameter
2765 //
2766 // [in] null
2767 // \ /
2768 // CmpP
2769 // Bool ne
2770 // If
2771 // / \
2772 // True False-<2>
2773 // / |
2774 // / cast_not_null
2775 // Load | | ^
2776 // [fast_test] | |
2777 // gvn to opt_test | |
2778 // / \ | <1>
2779 // True False |
2780 // | \\ |
2781 // [slow_call] \[fast_result]
2782 // Ctl Val \ \
2783 // | \ \
2784 // Catch <1> \ \
2785 // / \ ^ \ \
2786 // Ex No_Ex | \ \
2787 // | \ \ | \ <2> \
2788 // ... \ [slow_res] | | \ [null_result]
2789 // \ \--+--+--- | |
2790 // \ | / \ | /
2791 // --------Region Phi
2792 //
2793 //=============================================================================
2794 // Code is structured as a series of driver functions all called 'do_XXX' that
2795 // call a set of helper functions. Helper functions first, then drivers.
2796
2797 //------------------------------null_check_oop---------------------------------
2798 // Null check oop. Set null-path control into Region in slot 3.
2799 // Make a cast-not-nullness use the other not-null control. Return cast.
2800 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2801 bool never_see_null,
2802 bool safe_for_replace,
2803 bool speculative) {
2804 // Initial null check taken path
2805 (*null_control) = top();
2806 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2807
2808 // Generate uncommon_trap:
2809 if (never_see_null && (*null_control) != top()) {
2810 // If we see an unexpected null at a check-cast we record it and force a
2811 // recompile; the offending check-cast will be compiled to handle nulls.
2812 // If we see more than one offending BCI, then all checkcasts in the
2813 // method will be compiled to handle nulls.
2814 PreserveJVMState pjvms(this);
2815 set_control(*null_control);
2816 replace_in_map(value, null());
2817 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2818 uncommon_trap(reason,
2819 Deoptimization::Action_make_not_entrant);
2820 (*null_control) = top(); // null path is dead
2821 }
2822 if ((*null_control) == top() && safe_for_replace) {
2823 replace_in_map(value, cast);
2824 }
2825
2826 // Cast away null-ness on the result
2827 return cast;
2828 }
2829
2830 //------------------------------opt_iff----------------------------------------
2831 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2832 // Return slow-path control.
2833 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2834 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2835
2836 // Fast path taken; set region slot 2
2837 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2838 region->init_req(2,fast_taken); // Capture fast-control
2839
2840 // Fast path not-taken, i.e. slow path
2841 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2842 return slow_taken;
2843 }
2844
2845 //-----------------------------make_runtime_call-------------------------------
2846 Node* GraphKit::make_runtime_call(int flags,
2847 const TypeFunc* call_type, address call_addr,
2848 const char* call_name,
2849 const TypePtr* adr_type,
2850 // The following parms are all optional.
2851 // The first null ends the list.
2852 Node* parm0, Node* parm1,
2853 Node* parm2, Node* parm3,
2854 Node* parm4, Node* parm5,
2855 Node* parm6, Node* parm7) {
2856 assert(call_addr != nullptr, "must not call null targets");
2857
2858 // Slow-path call
2859 bool is_leaf = !(flags & RC_NO_LEAF);
2860 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2861 if (call_name == nullptr) {
2862 assert(!is_leaf, "must supply name for leaf");
2863 call_name = OptoRuntime::stub_name(call_addr);
2864 }
2865 CallNode* call;
2866 if (!is_leaf) {
2867 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2868 } else if (flags & RC_NO_FP) {
2869 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2870 } else if (flags & RC_VECTOR){
2871 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2872 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2873 } else if (flags & RC_PURE) {
2874 assert(adr_type == nullptr, "pure call does not touch memory");
2875 call = new CallLeafPureNode(call_type, call_addr, call_name);
2876 } else {
2877 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2878 }
2879
2880 // The following is similar to set_edges_for_java_call,
2881 // except that the memory effects of the call are restricted to AliasIdxRaw.
2882
2883 // Slow path call has no side-effects, uses few values
2884 bool wide_in = !(flags & RC_NARROW_MEM);
2885 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2886
2887 Node* prev_mem = nullptr;
2888 if (wide_in) {
2889 prev_mem = set_predefined_input_for_runtime_call(call);
2890 } else {
2891 assert(!wide_out, "narrow in => narrow out");
2892 Node* narrow_mem = memory(adr_type);
2893 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2894 }
2895
2896 // Hook each parm in order. Stop looking at the first null.
2897 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2898 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2899 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2900 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2901 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2902 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2903 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2904 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2905 /* close each nested if ===> */ } } } } } } } }
2906 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2907
2908 if (!is_leaf) {
2909 // Non-leaves can block and take safepoints:
2910 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2911 }
2912 // Non-leaves can throw exceptions:
2913 if (has_io) {
2914 call->set_req(TypeFunc::I_O, i_o());
2915 }
2916
2917 if (flags & RC_UNCOMMON) {
2918 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2919 // (An "if" probability corresponds roughly to an unconditional count.
2920 // Sort of.)
2921 call->set_cnt(PROB_UNLIKELY_MAG(4));
2922 }
2923
2924 Node* c = _gvn.transform(call);
2925 assert(c == call, "cannot disappear");
2926
2927 if (wide_out) {
2928 // Slow path call has full side-effects.
2929 set_predefined_output_for_runtime_call(call);
2930 } else {
2931 // Slow path call has few side-effects, and/or sets few values.
2932 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2933 }
2934
2935 if (has_io) {
2936 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2937 }
2938 return call;
2939
2940 }
2941
2942 // i2b
2943 Node* GraphKit::sign_extend_byte(Node* in) {
2944 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2945 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2946 }
2947
2948 // i2s
2949 Node* GraphKit::sign_extend_short(Node* in) {
2950 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2951 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2952 }
2953
2954
2955 //------------------------------merge_memory-----------------------------------
2956 // Merge memory from one path into the current memory state.
2957 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2958 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2959 Node* old_slice = mms.force_memory();
2960 Node* new_slice = mms.memory2();
2961 if (old_slice != new_slice) {
2962 PhiNode* phi;
2963 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2964 if (mms.is_empty()) {
2965 // clone base memory Phi's inputs for this memory slice
2966 assert(old_slice == mms.base_memory(), "sanity");
2967 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2968 _gvn.set_type(phi, Type::MEMORY);
2969 for (uint i = 1; i < phi->req(); i++) {
2970 phi->init_req(i, old_slice->in(i));
2971 }
2972 } else {
2973 phi = old_slice->as_Phi(); // Phi was generated already
2974 }
2975 } else {
2976 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2977 _gvn.set_type(phi, Type::MEMORY);
2978 }
2979 phi->set_req(new_path, new_slice);
2980 mms.set_memory(phi);
2981 }
2982 }
2983 }
2984
2985 //------------------------------make_slow_call_ex------------------------------
2986 // Make the exception handler hookups for the slow call
2987 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2988 if (stopped()) return;
2989
2990 // Make a catch node with just two handlers: fall-through and catch-all
2991 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2992 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2993 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2994 _gvn.set_type_bottom(norm);
2995 C->record_for_igvn(norm);
2996 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2997
2998 { PreserveJVMState pjvms(this);
2999 set_control(excp);
3000 set_i_o(i_o);
3001
3002 if (excp != top()) {
3003 if (deoptimize) {
3004 // Deoptimize if an exception is caught. Don't construct exception state in this case.
3005 uncommon_trap(Deoptimization::Reason_unhandled,
3006 Deoptimization::Action_none);
3007 } else {
3008 // Create an exception state also.
3009 // Use an exact type if the caller has a specific exception.
3010 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
3011 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
3012 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
3013 }
3014 }
3015 }
3016
3017 // Get the no-exception control from the CatchNode.
3018 set_control(norm);
3019 }
3020
3021 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
3022 Node* cmp = nullptr;
3023 switch(bt) {
3024 case T_INT: cmp = new CmpINode(in1, in2); break;
3025 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
3026 default: fatal("unexpected comparison type %s", type2name(bt));
3027 }
3028 cmp = gvn.transform(cmp);
3029 Node* bol = gvn.transform(new BoolNode(cmp, test));
3030 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
3031 gvn.transform(iff);
3032 if (!bol->is_Con()) gvn.record_for_igvn(iff);
3033 return iff;
3034 }
3035
3036 //-------------------------------gen_subtype_check-----------------------------
3037 // Generate a subtyping check. Takes as input the subtype and supertype.
3038 // Returns 2 values: sets the default control() to the true path and returns
3039 // the false path. Only reads invariant memory; sets no (visible) memory.
3040 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3041 // but that's not exposed to the optimizer. This call also doesn't take in an
3042 // Object; if you wish to check an Object you need to load the Object's class
3043 // prior to coming here.
3044 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3045 ciMethod* method, int bci) {
3046 Compile* C = gvn.C;
3047 if ((*ctrl)->is_top()) {
3048 return C->top();
3049 }
3050
3051 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3052 // For a direct pointer comparison, we need the refined array klass pointer
3053 Node* vm_superklass = superklass;
3054 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3055 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3056 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3057 }
3058
3059 // Fast check for identical types, perhaps identical constants.
3060 // The types can even be identical non-constants, in cases
3061 // involving Array.newInstance, Object.clone, etc.
3062 if (subklass == superklass)
3063 return C->top(); // false path is dead; no test needed.
3064
3065 if (gvn.type(superklass)->singleton()) {
3066 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3067 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3068
3069 // In the common case of an exact superklass, try to fold up the
3070 // test before generating code. You may ask, why not just generate
3071 // the code and then let it fold up? The answer is that the generated
3072 // code will necessarily include null checks, which do not always
3073 // completely fold away. If they are also needless, then they turn
3074 // into a performance loss. Example:
3075 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3076 // Here, the type of 'fa' is often exact, so the store check
3077 // of fa[1]=x will fold up, without testing the nullness of x.
3078 //
3079 // At macro expansion, we would have already folded the SubTypeCheckNode
3080 // being expanded here because we always perform the static sub type
3081 // check in SubTypeCheckNode::sub() regardless of whether
3082 // StressReflectiveCode is set or not. We can therefore skip this
3083 // static check when StressReflectiveCode is on.
3084 switch (C->static_subtype_check(superk, subk)) {
3085 case Compile::SSC_always_false:
3086 {
3087 Node* always_fail = *ctrl;
3088 *ctrl = gvn.C->top();
3089 return always_fail;
3090 }
3091 case Compile::SSC_always_true:
3092 return C->top();
3093 case Compile::SSC_easy_test:
3094 {
3095 // Just do a direct pointer compare and be done.
3096 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3097 *ctrl = gvn.transform(new IfTrueNode(iff));
3098 return gvn.transform(new IfFalseNode(iff));
3099 }
3100 case Compile::SSC_full_test:
3101 break;
3102 default:
3103 ShouldNotReachHere();
3104 }
3105 }
3106
3107 // %%% Possible further optimization: Even if the superklass is not exact,
3108 // if the subklass is the unique subtype of the superklass, the check
3109 // will always succeed. We could leave a dependency behind to ensure this.
3110
3111 // First load the super-klass's check-offset
3112 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3113 Node* m = C->immutable_memory();
3114 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3115 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3116 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3117 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
3118 bool might_be_cache = (chk_off_con == cacheoff_con);
3119
3120 // Load from the sub-klass's super-class display list, or a 1-word cache of
3121 // the secondary superclass list, or a failing value with a sentinel offset
3122 // if the super-klass is an interface or exceptionally deep in the Java
3123 // hierarchy and we have to scan the secondary superclass list the hard way.
3124 // Worst-case type is a little odd: null is allowed as a result (usually
3125 // klass loads can never produce a null).
3126 Node *chk_off_X = chk_off;
3127 #ifdef _LP64
3128 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
3129 #endif
3130 Node* p2 = gvn.transform(AddPNode::make_off_heap(subklass, chk_off_X));
3131 // For some types like interfaces the following loadKlass is from a 1-word
3132 // cache which is mutable so can't use immutable memory. Other
3133 // types load from the super-class display table which is immutable.
3134 Node* kmem = C->immutable_memory();
3135 // secondary_super_cache is not immutable but can be treated as such because:
3136 // - no ideal node writes to it in a way that could cause an
3137 // incorrect/missed optimization of the following Load.
3138 // - it's a cache so, worse case, not reading the latest value
3139 // wouldn't cause incorrect execution
3140 if (might_be_cache && mem != nullptr) {
3141 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
3142 }
3143 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3144
3145 // Compile speed common case: ARE a subtype and we canNOT fail
3146 if (superklass == nkls) {
3147 return C->top(); // false path is dead; no test needed.
3148 }
3149
3150 // Gather the various success & failures here
3151 RegionNode* r_not_subtype = new RegionNode(3);
3152 gvn.record_for_igvn(r_not_subtype);
3153 RegionNode* r_ok_subtype = new RegionNode(4);
3154 gvn.record_for_igvn(r_ok_subtype);
3155
3156 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3157 // SubTypeCheck node
3158 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3159 ciCallProfile profile = method->call_profile_at_bci(bci);
3160 float total_prob = 0;
3161 for (int i = 0; profile.has_receiver(i); ++i) {
3162 float prob = profile.receiver_prob(i);
3163 total_prob += prob;
3164 }
3165 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3166 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3167 for (int i = 0; profile.has_receiver(i); ++i) {
3168 ciKlass* klass = profile.receiver(i);
3169 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3170 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3171 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3172 continue;
3173 }
3174 if (klass_t->isa_aryklassptr()) {
3175 // For a direct pointer comparison, we need the refined array klass pointer
3176 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3177 }
3178 float prob = profile.receiver_prob(i);
3179 ConNode* klass_node = gvn.makecon(klass_t);
3180 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3181 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3182
3183 if (result == Compile::SSC_always_true) {
3184 r_ok_subtype->add_req(iftrue);
3185 } else {
3186 assert(result == Compile::SSC_always_false, "");
3187 r_not_subtype->add_req(iftrue);
3188 }
3189 *ctrl = gvn.transform(new IfFalseNode(iff));
3190 }
3191 }
3192 }
3193
3194 // See if we get an immediate positive hit. Happens roughly 83% of the
3195 // time. Test to see if the value loaded just previously from the subklass
3196 // is exactly the superklass.
3197 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3198 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3199 *ctrl = gvn.transform(new IfFalseNode(iff1));
3200
3201 // Compile speed common case: Check for being deterministic right now. If
3202 // chk_off is a constant and not equal to cacheoff then we are NOT a
3203 // subklass. In this case we need exactly the 1 test above and we can
3204 // return those results immediately.
3205 if (!might_be_cache) {
3206 Node* not_subtype_ctrl = *ctrl;
3207 *ctrl = iftrue1; // We need exactly the 1 test above
3208 PhaseIterGVN* igvn = gvn.is_IterGVN();
3209 if (igvn != nullptr) {
3210 igvn->remove_globally_dead_node(r_ok_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3211 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3212 }
3213 return not_subtype_ctrl;
3214 }
3215
3216 r_ok_subtype->init_req(1, iftrue1);
3217
3218 // Check for immediate negative hit. Happens roughly 11% of the time (which
3219 // is roughly 63% of the remaining cases). Test to see if the loaded
3220 // check-offset points into the subklass display list or the 1-element
3221 // cache. If it points to the display (and NOT the cache) and the display
3222 // missed then it's not a subtype.
3223 Node *cacheoff = gvn.intcon(cacheoff_con);
3224 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3225 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3226 *ctrl = gvn.transform(new IfFalseNode(iff2));
3227
3228 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3229 // No performance impact (too rare) but allows sharing of secondary arrays
3230 // which has some footprint reduction.
3231 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3232 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3233 *ctrl = gvn.transform(new IfFalseNode(iff3));
3234
3235 // -- Roads not taken here: --
3236 // We could also have chosen to perform the self-check at the beginning
3237 // of this code sequence, as the assembler does. This would not pay off
3238 // the same way, since the optimizer, unlike the assembler, can perform
3239 // static type analysis to fold away many successful self-checks.
3240 // Non-foldable self checks work better here in second position, because
3241 // the initial primary superclass check subsumes a self-check for most
3242 // types. An exception would be a secondary type like array-of-interface,
3243 // which does not appear in its own primary supertype display.
3244 // Finally, we could have chosen to move the self-check into the
3245 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3246 // dependent manner. But it is worthwhile to have the check here,
3247 // where it can be perhaps be optimized. The cost in code space is
3248 // small (register compare, branch).
3249
3250 // Now do a linear scan of the secondary super-klass array. Again, no real
3251 // performance impact (too rare) but it's gotta be done.
3252 // Since the code is rarely used, there is no penalty for moving it
3253 // out of line, and it can only improve I-cache density.
3254 // The decision to inline or out-of-line this final check is platform
3255 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3256 Node* psc = gvn.transform(
3257 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3258
3259 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3260 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3261 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3262
3263 // Return false path; set default control to true path.
3264 *ctrl = gvn.transform(r_ok_subtype);
3265 return gvn.transform(r_not_subtype);
3266 }
3267
3268 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3269 const Type* sub_t = _gvn.type(obj_or_subklass);
3270 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3271 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3272 obj_or_subklass = makecon(sub_t);
3273 }
3274 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3275 if (expand_subtype_check) {
3276 MergeMemNode* mem = merged_memory();
3277 Node* ctrl = control();
3278 Node* subklass = obj_or_subklass;
3279 if (!sub_t->isa_klassptr()) {
3280 subklass = load_object_klass(obj_or_subklass);
3281 }
3282
3283 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3284 set_control(ctrl);
3285 return n;
3286 }
3287
3288 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3289 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3290 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3291 set_control(_gvn.transform(new IfTrueNode(iff)));
3292 return _gvn.transform(new IfFalseNode(iff));
3293 }
3294
3295 // Profile-driven exact type check:
3296 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3297 float prob, Node* *casted_receiver) {
3298 assert(!klass->is_interface(), "no exact type check on interfaces");
3299 Node* fail = top();
3300 const Type* rec_t = _gvn.type(receiver);
3301 if (rec_t->is_inlinetypeptr()) {
3302 if (klass->equals(rec_t->inline_klass())) {
3303 (*casted_receiver) = receiver; // Always passes
3304 } else {
3305 (*casted_receiver) = top(); // Always fails
3306 fail = control();
3307 set_control(top());
3308 }
3309 return fail;
3310 }
3311 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3312 if (tklass->isa_aryklassptr()) {
3313 // For a direct pointer comparison, we need the refined array klass pointer
3314 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3315 }
3316 Node* recv_klass = load_object_klass(receiver);
3317 fail = type_check(recv_klass, tklass, prob);
3318
3319 if (!stopped()) {
3320 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3321 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3322 assert(recv_xtype->klass_is_exact(), "");
3323
3324 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3325 // Subsume downstream occurrences of receiver with a cast to
3326 // recv_xtype, since now we know what the type will be.
3327 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3328 Node* res = _gvn.transform(cast);
3329 if (recv_xtype->is_inlinetypeptr()) {
3330 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3331 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3332 }
3333 (*casted_receiver) = res;
3334 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3335 // (User must make the replace_in_map call.)
3336 }
3337 }
3338
3339 return fail;
3340 }
3341
3342 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3343 float prob) {
3344 Node* want_klass = makecon(tklass);
3345 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3346 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3347 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3348 set_control(_gvn.transform(new IfTrueNode (iff)));
3349 Node* fail = _gvn.transform(new IfFalseNode(iff));
3350 return fail;
3351 }
3352
3353 //------------------------------subtype_check_receiver-------------------------
3354 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3355 Node** casted_receiver) {
3356 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3357 Node* want_klass = makecon(tklass);
3358
3359 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3360
3361 // Ignore interface type information until interface types are properly tracked.
3362 if (!stopped() && !klass->is_interface()) {
3363 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3364 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3365 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3366 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3367 if (recv_type->is_inlinetypeptr()) {
3368 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3369 }
3370 (*casted_receiver) = cast;
3371 }
3372 }
3373
3374 return slow_ctl;
3375 }
3376
3377 //------------------------------seems_never_null-------------------------------
3378 // Use null_seen information if it is available from the profile.
3379 // If we see an unexpected null at a type check we record it and force a
3380 // recompile; the offending check will be recompiled to handle nulls.
3381 // If we see several offending BCIs, then all checks in the
3382 // method will be recompiled.
3383 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3384 speculating = !_gvn.type(obj)->speculative_maybe_null();
3385 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3386 if (UncommonNullCast // Cutout for this technique
3387 && obj != null() // And not the -Xcomp stupid case?
3388 && !too_many_traps(reason)
3389 ) {
3390 if (speculating) {
3391 return true;
3392 }
3393 if (data == nullptr)
3394 // Edge case: no mature data. Be optimistic here.
3395 return true;
3396 // If the profile has not seen a null, assume it won't happen.
3397 assert(java_bc() == Bytecodes::_checkcast ||
3398 java_bc() == Bytecodes::_instanceof ||
3399 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3400 return !data->as_BitData()->null_seen();
3401 }
3402 speculating = false;
3403 return false;
3404 }
3405
3406 void GraphKit::guard_klass_being_initialized(Node* klass) {
3407 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3408 Node* adr = off_heap_plus_addr(klass, init_state_off);
3409 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3410 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3411 T_BYTE, MemNode::acquire);
3412 init_state = _gvn.transform(init_state);
3413
3414 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3415
3416 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3417 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3418
3419 { BuildCutout unless(this, tst, PROB_MAX);
3420 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3421 }
3422 }
3423
3424 void GraphKit::guard_init_thread(Node* klass) {
3425 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3426 Node* adr = off_heap_plus_addr(klass, init_thread_off);
3427
3428 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3429 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3430 T_ADDRESS, MemNode::unordered);
3431 init_thread = _gvn.transform(init_thread);
3432
3433 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3434
3435 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3436 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3437
3438 { BuildCutout unless(this, tst, PROB_MAX);
3439 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3440 }
3441 }
3442
3443 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3444 if (ik->is_being_initialized()) {
3445 if (C->needs_clinit_barrier(ik, context)) {
3446 Node* klass = makecon(TypeKlassPtr::make(ik));
3447 guard_klass_being_initialized(klass);
3448 guard_init_thread(klass);
3449 insert_mem_bar(Op_MemBarCPUOrder);
3450 }
3451 } else if (ik->is_initialized()) {
3452 return; // no barrier needed
3453 } else {
3454 uncommon_trap(Deoptimization::Reason_uninitialized,
3455 Deoptimization::Action_reinterpret,
3456 nullptr);
3457 }
3458 }
3459
3460 //------------------------maybe_cast_profiled_receiver-------------------------
3461 // If the profile has seen exactly one type, narrow to exactly that type.
3462 // Subsequent type checks will always fold up.
3463 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3464 const TypeKlassPtr* require_klass,
3465 ciKlass* spec_klass,
3466 bool safe_for_replace) {
3467 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3468
3469 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3470
3471 // Make sure we haven't already deoptimized from this tactic.
3472 if (too_many_traps_or_recompiles(reason))
3473 return nullptr;
3474
3475 // (No, this isn't a call, but it's enough like a virtual call
3476 // to use the same ciMethod accessor to get the profile info...)
3477 // If we have a speculative type use it instead of profiling (which
3478 // may not help us)
3479 ciKlass* exact_kls = spec_klass;
3480 if (exact_kls == nullptr) {
3481 if (java_bc() == Bytecodes::_aastore) {
3482 ciKlass* array_type = nullptr;
3483 ciKlass* element_type = nullptr;
3484 ProfilePtrKind element_ptr = ProfileMaybeNull;
3485 bool flat_array = true;
3486 bool null_free_array = true;
3487 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3488 exact_kls = element_type;
3489 } else {
3490 exact_kls = profile_has_unique_klass();
3491 }
3492 }
3493 if (exact_kls != nullptr) {// no cast failures here
3494 if (require_klass == nullptr ||
3495 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3496 // If we narrow the type to match what the type profile sees or
3497 // the speculative type, we can then remove the rest of the
3498 // cast.
3499 // This is a win, even if the exact_kls is very specific,
3500 // because downstream operations, such as method calls,
3501 // will often benefit from the sharper type.
3502 Node* exact_obj = not_null_obj; // will get updated in place...
3503 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3504 &exact_obj);
3505 { PreserveJVMState pjvms(this);
3506 set_control(slow_ctl);
3507 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3508 }
3509 if (safe_for_replace) {
3510 replace_in_map(not_null_obj, exact_obj);
3511 }
3512 return exact_obj;
3513 }
3514 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3515 }
3516
3517 return nullptr;
3518 }
3519
3520 /**
3521 * Cast obj to type and emit guard unless we had too many traps here
3522 * already
3523 *
3524 * @param obj node being casted
3525 * @param type type to cast the node to
3526 * @param not_null true if we know node cannot be null
3527 */
3528 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3529 ciKlass* type,
3530 bool not_null) {
3531 if (stopped()) {
3532 return obj;
3533 }
3534
3535 // type is null if profiling tells us this object is always null
3536 if (type != nullptr) {
3537 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3538 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3539
3540 if (!too_many_traps_or_recompiles(null_reason) &&
3541 !too_many_traps_or_recompiles(class_reason)) {
3542 Node* not_null_obj = nullptr;
3543 // not_null is true if we know the object is not null and
3544 // there's no need for a null check
3545 if (!not_null) {
3546 Node* null_ctl = top();
3547 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3548 assert(null_ctl->is_top(), "no null control here");
3549 } else {
3550 not_null_obj = obj;
3551 }
3552
3553 Node* exact_obj = not_null_obj;
3554 ciKlass* exact_kls = type;
3555 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3556 &exact_obj);
3557 {
3558 PreserveJVMState pjvms(this);
3559 set_control(slow_ctl);
3560 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3561 }
3562 replace_in_map(not_null_obj, exact_obj);
3563 obj = exact_obj;
3564 }
3565 } else {
3566 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3567 Node* exact_obj = null_assert(obj);
3568 replace_in_map(obj, exact_obj);
3569 obj = exact_obj;
3570 }
3571 }
3572 return obj;
3573 }
3574
3575 //-------------------------------gen_instanceof--------------------------------
3576 // Generate an instance-of idiom. Used by both the instance-of bytecode
3577 // and the reflective instance-of call.
3578 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3579 kill_dead_locals(); // Benefit all the uncommon traps
3580 assert( !stopped(), "dead parse path should be checked in callers" );
3581 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3582 "must check for not-null not-dead klass in callers");
3583
3584 // Make the merge point
3585 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3586 RegionNode* region = new RegionNode(PATH_LIMIT);
3587 Node* phi = new PhiNode(region, TypeInt::BOOL);
3588 C->set_has_split_ifs(true); // Has chance for split-if optimization
3589
3590 ciProfileData* data = nullptr;
3591 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3592 data = method()->method_data()->bci_to_data(bci());
3593 }
3594 bool speculative_not_null = false;
3595 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3596 && seems_never_null(obj, data, speculative_not_null));
3597
3598 // Null check; get casted pointer; set region slot 3
3599 Node* null_ctl = top();
3600 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3601
3602 // If not_null_obj is dead, only null-path is taken
3603 if (stopped()) { // Doing instance-of on a null?
3604 set_control(null_ctl);
3605 return intcon(0);
3606 }
3607 region->init_req(_null_path, null_ctl);
3608 phi ->init_req(_null_path, intcon(0)); // Set null path value
3609 if (null_ctl == top()) {
3610 // Do this eagerly, so that pattern matches like is_diamond_phi
3611 // will work even during parsing.
3612 assert(_null_path == PATH_LIMIT-1, "delete last");
3613 region->del_req(_null_path);
3614 phi ->del_req(_null_path);
3615 }
3616
3617 // Do we know the type check always succeed?
3618 bool known_statically = false;
3619 if (_gvn.type(superklass)->singleton()) {
3620 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3621 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3622 if (subk != nullptr && subk->is_loaded()) {
3623 int static_res = C->static_subtype_check(superk, subk);
3624 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3625 }
3626 }
3627
3628 if (!known_statically) {
3629 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3630 // We may not have profiling here or it may not help us. If we
3631 // have a speculative type use it to perform an exact cast.
3632 ciKlass* spec_obj_type = obj_type->speculative_type();
3633 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3634 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3635 if (stopped()) { // Profile disagrees with this path.
3636 set_control(null_ctl); // Null is the only remaining possibility.
3637 return intcon(0);
3638 }
3639 if (cast_obj != nullptr) {
3640 not_null_obj = cast_obj;
3641 }
3642 }
3643 }
3644
3645 // Generate the subtype check
3646 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3647
3648 // Plug in the success path to the general merge in slot 1.
3649 region->init_req(_obj_path, control());
3650 phi ->init_req(_obj_path, intcon(1));
3651
3652 // Plug in the failing path to the general merge in slot 2.
3653 region->init_req(_fail_path, not_subtype_ctrl);
3654 phi ->init_req(_fail_path, intcon(0));
3655
3656 // Return final merged results
3657 set_control( _gvn.transform(region) );
3658 record_for_igvn(region);
3659
3660 // If we know the type check always succeeds then we don't use the
3661 // profiling data at this bytecode. Don't lose it, feed it to the
3662 // type system as a speculative type.
3663 if (safe_for_replace) {
3664 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3665 replace_in_map(obj, casted_obj);
3666 }
3667
3668 return _gvn.transform(phi);
3669 }
3670
3671 //-------------------------------gen_checkcast---------------------------------
3672 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3673 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3674 // uncommon-trap paths work. Adjust stack after this call.
3675 // If failure_control is supplied and not null, it is filled in with
3676 // the control edge for the cast failure. Otherwise, an appropriate
3677 // uncommon trap or exception is thrown.
3678 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3679 // when the current map for the success path is updated with information only present
3680 // on the success path and not the cast failure path. The newly cloned map should then be
3681 // used to emit the uncommon trap in the caller.
3682 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3683 assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3684 "failure_control must be set when new_failure_map is used");
3685 kill_dead_locals(); // Benefit all the uncommon traps
3686 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3687 const Type* obj_type = _gvn.type(obj);
3688
3689 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3690 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3691 bool safe_for_replace = (failure_control == nullptr);
3692 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3693
3694 // Fast cutout: Check the case that the cast is vacuously true.
3695 // This detects the common cases where the test will short-circuit
3696 // away completely. We do this before we perform the null check,
3697 // because if the test is going to turn into zero code, we don't
3698 // want a residual null check left around. (Causes a slowdown,
3699 // for example, in some objArray manipulations, such as a[i]=a[j].)
3700 if (improved_klass_ptr_type->singleton()) {
3701 const TypeKlassPtr* kptr = nullptr;
3702 if (obj_type->isa_oop_ptr()) {
3703 kptr = obj_type->is_oopptr()->as_klass_type();
3704 } else if (obj->is_InlineType()) {
3705 ciInlineKlass* vk = obj_type->inline_klass();
3706 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3707 }
3708
3709 if (kptr != nullptr) {
3710 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3711 case Compile::SSC_always_true:
3712 // If we know the type check always succeed then we don't use
3713 // the profiling data at this bytecode. Don't lose it, feed it
3714 // to the type system as a speculative type.
3715 obj = record_profiled_receiver_for_speculation(obj);
3716 if (null_free) {
3717 assert(safe_for_replace, "must be");
3718 obj = null_check(obj);
3719 }
3720 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3721 return obj;
3722 case Compile::SSC_always_false:
3723 if (null_free) {
3724 assert(safe_for_replace, "must be");
3725 obj = null_check(obj);
3726 }
3727 // It needs a null check because a null will *pass* the cast check.
3728 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3729 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3730 Deoptimization::DeoptReason reason = is_aastore ?
3731 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3732 builtin_throw(reason);
3733 return top();
3734 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3735 return null_assert(obj);
3736 }
3737 break; // Fall through to full check
3738 default:
3739 break;
3740 }
3741 }
3742 }
3743
3744 ciProfileData* data = nullptr;
3745 if (failure_control == nullptr) { // use MDO in regular case only
3746 assert(java_bc() == Bytecodes::_aastore ||
3747 java_bc() == Bytecodes::_checkcast,
3748 "interpreter profiles type checks only for these BCs");
3749 if (method()->method_data()->is_mature()) {
3750 data = method()->method_data()->bci_to_data(bci());
3751 }
3752 }
3753
3754 // Make the merge point
3755 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3756 RegionNode* region = new RegionNode(PATH_LIMIT);
3757 Node* phi = new PhiNode(region, toop);
3758 _gvn.set_type(region, Type::CONTROL);
3759 _gvn.set_type(phi, toop);
3760
3761 C->set_has_split_ifs(true); // Has chance for split-if optimization
3762
3763 // Use null-cast information if it is available
3764 bool speculative_not_null = false;
3765 bool never_see_null = ((failure_control == nullptr) // regular case only
3766 && seems_never_null(obj, data, speculative_not_null));
3767
3768 if (obj->is_InlineType()) {
3769 // Re-execute if buffering during triggers deoptimization
3770 PreserveReexecuteState preexecs(this);
3771 jvms()->set_should_reexecute(true);
3772 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3773 }
3774
3775 // Null check; get casted pointer; set region slot 3
3776 Node* null_ctl = top();
3777 Node* not_null_obj = nullptr;
3778 if (null_free) {
3779 assert(safe_for_replace, "must be");
3780 not_null_obj = null_check(obj);
3781 } else {
3782 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3783 }
3784
3785 // If not_null_obj is dead, only null-path is taken
3786 if (stopped()) { // Doing instance-of on a null?
3787 set_control(null_ctl);
3788 if (toop->is_inlinetypeptr()) {
3789 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3790 }
3791 return null();
3792 }
3793 region->init_req(_null_path, null_ctl);
3794 phi ->init_req(_null_path, null()); // Set null path value
3795 if (null_ctl == top()) {
3796 // Do this eagerly, so that pattern matches like is_diamond_phi
3797 // will work even during parsing.
3798 assert(_null_path == PATH_LIMIT-1, "delete last");
3799 region->del_req(_null_path);
3800 phi ->del_req(_null_path);
3801 }
3802
3803 Node* cast_obj = nullptr;
3804 if (improved_klass_ptr_type->klass_is_exact()) {
3805 // The following optimization tries to statically cast the speculative type of the object
3806 // (for example obtained during profiling) to the type of the superklass and then do a
3807 // dynamic check that the type of the object is what we expect. To work correctly
3808 // for checkcast and aastore the type of superklass should be exact.
3809 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3810 // We may not have profiling here or it may not help us. If we have
3811 // a speculative type use it to perform an exact cast.
3812 ciKlass* spec_obj_type = obj_type->speculative_type();
3813 if (spec_obj_type != nullptr || data != nullptr) {
3814 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3815 if (cast_obj != nullptr) {
3816 if (failure_control != nullptr) // failure is now impossible
3817 (*failure_control) = top();
3818 // adjust the type of the phi to the exact klass:
3819 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3820 }
3821 }
3822 }
3823
3824 if (cast_obj == nullptr) {
3825 // Generate the subtype check
3826 Node* improved_superklass = superklass;
3827 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3828 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3829 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3830 // Additionally, the benefit would only be minor in non-constant cases.
3831 improved_superklass = makecon(improved_klass_ptr_type);
3832 }
3833 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3834 // Plug in success path into the merge
3835 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3836 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3837 if (failure_control == nullptr) {
3838 if (not_subtype_ctrl != top()) { // If failure is possible
3839 PreserveJVMState pjvms(this);
3840 set_control(not_subtype_ctrl);
3841 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3842 Deoptimization::DeoptReason reason = is_aastore ?
3843 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3844 builtin_throw(reason);
3845 }
3846 } else {
3847 (*failure_control) = not_subtype_ctrl;
3848 }
3849 }
3850
3851 region->init_req(_obj_path, control());
3852 phi ->init_req(_obj_path, cast_obj);
3853
3854 // A merge of null or Casted-NotNull obj
3855 Node* res = _gvn.transform(phi);
3856
3857 // Note I do NOT always 'replace_in_map(obj,result)' here.
3858 // if( tk->klass()->can_be_primary_super() )
3859 // This means that if I successfully store an Object into an array-of-String
3860 // I 'forget' that the Object is really now known to be a String. I have to
3861 // do this because we don't have true union types for interfaces - if I store
3862 // a Baz into an array-of-Interface and then tell the optimizer it's an
3863 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3864 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3865 // replace_in_map( obj, res );
3866
3867 // Return final merged results
3868 set_control( _gvn.transform(region) );
3869 record_for_igvn(region);
3870
3871 bool not_inline = !toop->can_be_inline_type();
3872 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3873 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3874 // Check if obj has been loaded from an array
3875 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3876 Node* array = nullptr;
3877 if (obj->isa_Load()) {
3878 Node* address = obj->in(MemNode::Address);
3879 if (address->isa_AddP()) {
3880 array = address->as_AddP()->in(AddPNode::Base);
3881 }
3882 } else if (obj->is_Phi()) {
3883 Node* region = obj->in(0);
3884 // TODO make this more robust (see JDK-8231346)
3885 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3886 IfNode* iff = region->in(2)->in(0)->isa_If();
3887 if (iff != nullptr) {
3888 iff->is_flat_array_check(&_gvn, &array);
3889 }
3890 }
3891 }
3892 if (array != nullptr) {
3893 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3894 if (ary_t != nullptr) {
3895 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3896 // Casting array element to a non-inline-type, mark array as not null-free.
3897 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3898 if (new_cast_failure_map != nullptr) {
3899 // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3900 // create the cast failure path in the caller without wrongly making the cast node live there.
3901 *new_cast_failure_map = clone_map();
3902 }
3903 replace_in_map(array, cast);
3904 array = cast;
3905 }
3906 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3907 // Casting array element to a non-flat-in-array type, mark array as not flat.
3908 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3909 if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3910 // Same as above.
3911 *new_cast_failure_map = clone_map();
3912 }
3913 replace_in_map(array, cast);
3914 array = cast;
3915 }
3916 }
3917 }
3918 }
3919
3920 if (!stopped() && !res->is_InlineType()) {
3921 res = record_profiled_receiver_for_speculation(res);
3922 if (toop->is_inlinetypeptr() && !maybe_larval) {
3923 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3924 res = vt;
3925 if (safe_for_replace) {
3926 replace_in_map(obj, vt);
3927 replace_in_map(not_null_obj, vt);
3928 replace_in_map(res, vt);
3929 }
3930 }
3931 }
3932 return res;
3933 }
3934
3935 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3936 // Load markword
3937 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3938 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3939 if (check_lock && !UseCompactObjectHeaders) {
3940 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3941 // Check if obj is locked
3942 Node* locked_bit = MakeConX(markWord::unlocked_value);
3943 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3944 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3945 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3946 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3947 _gvn.transform(iff);
3948 Node* locked_region = new RegionNode(3);
3949 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3950
3951 // Unlocked: Use bits from mark word
3952 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3953 mark_phi->init_req(1, mark);
3954
3955 // Locked: Load prototype header from klass
3956 set_control(_gvn.transform(new IfFalseNode(iff)));
3957 // Make loads control dependent to make sure they are only executed if array is locked
3958 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3959 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3960 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3961 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));
3962
3963 locked_region->init_req(2, control());
3964 mark_phi->init_req(2, proto);
3965 set_control(_gvn.transform(locked_region));
3966 record_for_igvn(locked_region);
3967
3968 mark = mark_phi;
3969 }
3970
3971 // Now check if mark word bits are set
3972 Node* mask = MakeConX(mask_val);
3973 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3974 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3975 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3976 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3977 }
3978
3979 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3980 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3981 }
3982
3983 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3984 // We can't use immutable memory here because the mark word is mutable.
3985 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3986 // check is moved out of loops (mainly to enable loop unswitching).
3987 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3988 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3989 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3990 }
3991
3992 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3993 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3994 }
3995
3996 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3997 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3998
3999 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
4000 if (!vk->has_null_free_non_atomic_layout()) {
4001 return intcon(1); // Always atomic
4002 } else if (!vk->has_null_free_atomic_layout()) {
4003 return intcon(0); // Never atomic
4004 }
4005
4006 Node* array_klass = load_object_klass(array);
4007 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4008 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4009 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4010 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
4011 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4012 }
4013
4014 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4015 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4016 RegionNode* region = new RegionNode(3);
4017 Node* null_ctl = top();
4018 null_check_oop(val, &null_ctl);
4019 if (null_ctl != top()) {
4020 PreserveJVMState pjvms(this);
4021 set_control(null_ctl);
4022 {
4023 // Deoptimize if null-free array
4024 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4025 inc_sp(nargs);
4026 uncommon_trap(Deoptimization::Reason_null_check,
4027 Deoptimization::Action_none);
4028 }
4029 region->init_req(1, control());
4030 }
4031 region->init_req(2, control());
4032 set_control(_gvn.transform(region));
4033 record_for_igvn(region);
4034 if (_gvn.type(val) == TypePtr::NULL_PTR) {
4035 // Since we were just successfully storing null, the array can't be null free.
4036 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4037 ary_t = ary_t->cast_to_not_null_free();
4038 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4039 if (safe_for_replace) {
4040 replace_in_map(ary, cast);
4041 }
4042 ary = cast;
4043 }
4044 return ary;
4045 }
4046
4047 //------------------------------next_monitor-----------------------------------
4048 // What number should be given to the next monitor?
4049 int GraphKit::next_monitor() {
4050 int current = jvms()->monitor_depth()* C->sync_stack_slots();
4051 int next = current + C->sync_stack_slots();
4052 // Keep the toplevel high water mark current:
4053 if (C->fixed_slots() < next) C->set_fixed_slots(next);
4054 return current;
4055 }
4056
4057 //------------------------------insert_mem_bar---------------------------------
4058 // Memory barrier to avoid floating things around
4059 // The membar serves as a pinch point between both control and all memory slices.
4060 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4061 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4062 mb->init_req(TypeFunc::Control, control());
4063 mb->init_req(TypeFunc::Memory, reset_memory());
4064 Node* membar = _gvn.transform(mb);
4065 record_for_igvn(membar);
4066 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4067 set_all_memory_call(membar);
4068 return membar;
4069 }
4070
4071 //-------------------------insert_mem_bar_volatile----------------------------
4072 // Memory barrier to avoid floating things around
4073 // The membar serves as a pinch point between both control and memory(alias_idx).
4074 // If you want to make a pinch point on all memory slices, do not use this
4075 // function (even with AliasIdxBot); use insert_mem_bar() instead.
4076 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
4077 // When Parse::do_put_xxx updates a volatile field, it appends a series
4078 // of MemBarVolatile nodes, one for *each* volatile field alias category.
4079 // The first membar is on the same memory slice as the field store opcode.
4080 // This forces the membar to follow the store. (Bug 6500685 broke this.)
4081 // All the other membars (for other volatile slices, including AliasIdxBot,
4082 // which stands for all unknown volatile slices) are control-dependent
4083 // on the first membar. This prevents later volatile loads or stores
4084 // from sliding up past the just-emitted store.
4085
4086 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
4087 mb->set_req(TypeFunc::Control,control());
4088 if (alias_idx == Compile::AliasIdxBot) {
4089 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
4090 } else {
4091 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
4092 mb->set_req(TypeFunc::Memory, memory(alias_idx));
4093 }
4094 Node* membar = _gvn.transform(mb);
4095 record_for_igvn(membar);
4096 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4097 if (alias_idx == Compile::AliasIdxBot) {
4098 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
4099 } else {
4100 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
4101 }
4102 return membar;
4103 }
4104
4105 //------------------------------shared_lock------------------------------------
4106 // Emit locking code.
4107 FastLockNode* GraphKit::shared_lock(Node* obj) {
4108 // bci is either a monitorenter bc or InvocationEntryBci
4109 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4110 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4111
4112 if (stopped()) // Dead monitor?
4113 return nullptr;
4114
4115 assert(dead_locals_are_killed(), "should kill locals before sync. point");
4116
4117 // Box the stack location
4118 Node* box = new BoxLockNode(next_monitor());
4119 // Check for bailout after new BoxLockNode
4120 if (failing()) { return nullptr; }
4121 box = _gvn.transform(box);
4122 Node* mem = reset_memory();
4123
4124 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
4125
4126 // Add monitor to debug info for the slow path. If we block inside the
4127 // slow path and de-opt, we need the monitor hanging around
4128 map()->push_monitor( flock );
4129
4130 const TypeFunc *tf = LockNode::lock_type();
4131 LockNode *lock = new LockNode(C, tf);
4132
4133 lock->init_req( TypeFunc::Control, control() );
4134 lock->init_req( TypeFunc::Memory , mem );
4135 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4136 lock->init_req( TypeFunc::FramePtr, frameptr() );
4137 lock->init_req( TypeFunc::ReturnAdr, top() );
4138
4139 lock->init_req(TypeFunc::Parms + 0, obj);
4140 lock->init_req(TypeFunc::Parms + 1, box);
4141 lock->init_req(TypeFunc::Parms + 2, flock);
4142 add_safepoint_edges(lock);
4143
4144 lock = _gvn.transform( lock )->as_Lock();
4145
4146 // lock has no side-effects, sets few values
4147 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
4148
4149 insert_mem_bar(Op_MemBarAcquireLock);
4150
4151 // Add this to the worklist so that the lock can be eliminated
4152 record_for_igvn(lock);
4153
4154 #ifndef PRODUCT
4155 if (PrintLockStatistics) {
4156 // Update the counter for this lock. Don't bother using an atomic
4157 // operation since we don't require absolute accuracy.
4158 lock->create_lock_counter(map()->jvms());
4159 increment_counter(lock->counter()->addr());
4160 }
4161 #endif
4162
4163 return flock;
4164 }
4165
4166
4167 //------------------------------shared_unlock----------------------------------
4168 // Emit unlocking code.
4169 void GraphKit::shared_unlock(Node* box, Node* obj) {
4170 // bci is either a monitorenter bc or InvocationEntryBci
4171 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4172 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4173
4174 if (stopped()) { // Dead monitor?
4175 map()->pop_monitor(); // Kill monitor from debug info
4176 return;
4177 }
4178 assert(!obj->is_InlineType(), "should not unlock on inline type");
4179
4180 // Memory barrier to avoid floating things down past the locked region
4181 insert_mem_bar(Op_MemBarReleaseLock);
4182
4183 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4184 UnlockNode *unlock = new UnlockNode(C, tf);
4185 #ifdef ASSERT
4186 unlock->set_dbg_jvms(sync_jvms());
4187 #endif
4188 uint raw_idx = Compile::AliasIdxRaw;
4189 unlock->init_req( TypeFunc::Control, control() );
4190 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4191 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4192 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4193 unlock->init_req( TypeFunc::ReturnAdr, top() );
4194
4195 unlock->init_req(TypeFunc::Parms + 0, obj);
4196 unlock->init_req(TypeFunc::Parms + 1, box);
4197 unlock = _gvn.transform(unlock)->as_Unlock();
4198
4199 Node* mem = reset_memory();
4200
4201 // unlock has no side-effects, sets few values
4202 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4203
4204 // Kill monitor from debug info
4205 map()->pop_monitor( );
4206 }
4207
4208 //-------------------------------get_layout_helper-----------------------------
4209 // If the given klass is a constant or known to be an array,
4210 // fetch the constant layout helper value into constant_value
4211 // and return null. Otherwise, load the non-constant
4212 // layout helper value, and return the node which represents it.
4213 // This two-faced routine is useful because allocation sites
4214 // almost always feature constant types.
4215 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4216 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4217 if (!StressReflectiveCode && klass_t != nullptr) {
4218 bool xklass = klass_t->klass_is_exact();
4219 bool can_be_flat = false;
4220 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4221 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4222 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4223 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4224 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4225 }
4226 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4227 jint lhelper;
4228 if (klass_t->is_flat()) {
4229 lhelper = ary_type->flat_layout_helper();
4230 } else if (klass_t->isa_aryklassptr()) {
4231 BasicType elem = ary_type->elem()->array_element_basic_type();
4232 if (is_reference_type(elem, true)) {
4233 elem = T_OBJECT;
4234 }
4235 lhelper = Klass::array_layout_helper(elem);
4236 } else {
4237 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4238 }
4239 if (lhelper != Klass::_lh_neutral_value) {
4240 constant_value = lhelper;
4241 return (Node*) nullptr;
4242 }
4243 }
4244 }
4245 constant_value = Klass::_lh_neutral_value; // put in a known value
4246 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4247 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4248 }
4249
4250 // We just put in an allocate/initialize with a big raw-memory effect.
4251 // Hook selected additional alias categories on the initialization.
4252 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4253 MergeMemNode* init_in_merge,
4254 Node* init_out_raw) {
4255 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4256 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4257
4258 Node* prevmem = kit.memory(alias_idx);
4259 init_in_merge->set_memory_at(alias_idx, prevmem);
4260 if (init_out_raw != nullptr) {
4261 kit.set_memory(init_out_raw, alias_idx);
4262 }
4263 }
4264
4265 //---------------------------set_output_for_allocation-------------------------
4266 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4267 const TypeOopPtr* oop_type,
4268 bool deoptimize_on_exception) {
4269 int rawidx = Compile::AliasIdxRaw;
4270 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4271 add_safepoint_edges(alloc);
4272 Node* allocx = _gvn.transform(alloc);
4273 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4274 // create memory projection for i_o
4275 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4276 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4277
4278 // create a memory projection as for the normal control path
4279 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4280 set_memory(malloc, rawidx);
4281
4282 // a normal slow-call doesn't change i_o, but an allocation does
4283 // we create a separate i_o projection for the normal control path
4284 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4285 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4286
4287 // put in an initialization barrier
4288 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4289 rawoop)->as_Initialize();
4290 assert(alloc->initialization() == init, "2-way macro link must work");
4291 assert(init ->allocation() == alloc, "2-way macro link must work");
4292 {
4293 // Extract memory strands which may participate in the new object's
4294 // initialization, and source them from the new InitializeNode.
4295 // This will allow us to observe initializations when they occur,
4296 // and link them properly (as a group) to the InitializeNode.
4297 assert(init->in(InitializeNode::Memory) == malloc, "");
4298 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4299 init->set_req(InitializeNode::Memory, minit_in);
4300 record_for_igvn(minit_in); // fold it up later, if possible
4301 _gvn.set_type(minit_in, Type::MEMORY);
4302 Node* minit_out = memory(rawidx);
4303 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4304 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4305 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4306 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4307 // multiple projections as a result.
4308 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4309 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4310 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4311 if (oop_type->isa_aryptr()) {
4312 // Initially all flat array accesses share a single slice
4313 // but that changes after parsing. Prepare the memory graph so
4314 // it can optimize flat array accesses properly once they
4315 // don't share a single slice.
4316 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4317 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4318 int elemidx = C->get_alias_index(telemref);
4319 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4320 if (alias_adr_type->is_flat()) {
4321 C->set_flat_accesses();
4322 }
4323 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4324 } else if (oop_type->isa_instptr()) {
4325 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4326 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4327 ciField* field = ik->nonstatic_field_at(i);
4328 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4329 continue; // do not bother to track really large numbers of fields
4330 // Find (or create) the alias category for this field:
4331 int fieldidx = C->alias_type(field)->index();
4332 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4333 }
4334 }
4335 }
4336
4337 // Cast raw oop to the real thing...
4338 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4339 javaoop = _gvn.transform(javaoop);
4340 C->set_recent_alloc(control(), javaoop);
4341 assert(just_allocated_object(control()) == javaoop, "just allocated");
4342
4343 #ifdef ASSERT
4344 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4345 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4346 "Ideal_allocation works");
4347 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4348 "Ideal_allocation works");
4349 if (alloc->is_AllocateArray()) {
4350 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4351 "Ideal_allocation works");
4352 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4353 "Ideal_allocation works");
4354 } else {
4355 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4356 }
4357 }
4358 #endif //ASSERT
4359
4360 return javaoop;
4361 }
4362
4363 //---------------------------new_instance--------------------------------------
4364 // This routine takes a klass_node which may be constant (for a static type)
4365 // or may be non-constant (for reflective code). It will work equally well
4366 // for either, and the graph will fold nicely if the optimizer later reduces
4367 // the type to a constant.
4368 // The optional arguments are for specialized use by intrinsics:
4369 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4370 // - If 'return_size_val', report the total object size to the caller.
4371 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4372 Node* GraphKit::new_instance(Node* klass_node,
4373 Node* extra_slow_test,
4374 Node* *return_size_val,
4375 bool deoptimize_on_exception,
4376 InlineTypeNode* inline_type_node) {
4377 // Compute size in doublewords
4378 // The size is always an integral number of doublewords, represented
4379 // as a positive bytewise size stored in the klass's layout_helper.
4380 // The layout_helper also encodes (in a low bit) the need for a slow path.
4381 jint layout_con = Klass::_lh_neutral_value;
4382 Node* layout_val = get_layout_helper(klass_node, layout_con);
4383 bool layout_is_con = (layout_val == nullptr);
4384
4385 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4386 // Generate the initial go-slow test. It's either ALWAYS (return a
4387 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4388 // case) a computed value derived from the layout_helper.
4389 Node* initial_slow_test = nullptr;
4390 if (layout_is_con) {
4391 assert(!StressReflectiveCode, "stress mode does not use these paths");
4392 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4393 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4394 } else { // reflective case
4395 // This reflective path is used by Unsafe.allocateInstance.
4396 // (It may be stress-tested by specifying StressReflectiveCode.)
4397 // Basically, we want to get into the VM is there's an illegal argument.
4398 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4399 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4400 if (extra_slow_test != intcon(0)) {
4401 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4402 }
4403 // (Macro-expander will further convert this to a Bool, if necessary.)
4404 }
4405
4406 // Find the size in bytes. This is easy; it's the layout_helper.
4407 // The size value must be valid even if the slow path is taken.
4408 Node* size = nullptr;
4409 if (layout_is_con) {
4410 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4411 } else { // reflective case
4412 // This reflective path is used by clone and Unsafe.allocateInstance.
4413 size = ConvI2X(layout_val);
4414
4415 // Clear the low bits to extract layout_helper_size_in_bytes:
4416 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4417 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4418 size = _gvn.transform( new AndXNode(size, mask) );
4419 }
4420 if (return_size_val != nullptr) {
4421 (*return_size_val) = size;
4422 }
4423
4424 // This is a precise notnull oop of the klass.
4425 // (Actually, it need not be precise if this is a reflective allocation.)
4426 // It's what we cast the result to.
4427 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4428 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4429 const TypeOopPtr* oop_type = tklass->as_instance_type();
4430
4431 // Now generate allocation code
4432
4433 // The entire memory state is needed for slow path of the allocation
4434 // since GC and deoptimization can happen.
4435 Node *mem = reset_memory();
4436 set_all_memory(mem); // Create new memory state
4437
4438 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4439 control(), mem, i_o(),
4440 size, klass_node,
4441 initial_slow_test, inline_type_node);
4442
4443 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4444 }
4445
4446 //-------------------------------new_array-------------------------------------
4447 // helper for newarray and anewarray
4448 // The 'length' parameter is (obviously) the length of the array.
4449 // The optional arguments are for specialized use by intrinsics:
4450 // - If 'return_size_val', report the non-padded array size (sum of header size
4451 // and array body) to the caller.
4452 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4453 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4454 Node* length, // number of array elements
4455 int nargs, // number of arguments to push back for uncommon trap
4456 Node* *return_size_val,
4457 bool deoptimize_on_exception,
4458 Node* init_val) {
4459 jint layout_con = Klass::_lh_neutral_value;
4460 Node* layout_val = get_layout_helper(klass_node, layout_con);
4461 bool layout_is_con = (layout_val == nullptr);
4462
4463 if (!layout_is_con && !StressReflectiveCode &&
4464 !too_many_traps(Deoptimization::Reason_class_check)) {
4465 // This is a reflective array creation site.
4466 // Optimistically assume that it is a subtype of Object[],
4467 // so that we can fold up all the address arithmetic.
4468 layout_con = Klass::array_layout_helper(T_OBJECT);
4469 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4470 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4471 { BuildCutout unless(this, bol_lh, PROB_MAX);
4472 inc_sp(nargs);
4473 uncommon_trap(Deoptimization::Reason_class_check,
4474 Deoptimization::Action_maybe_recompile);
4475 }
4476 layout_val = nullptr;
4477 layout_is_con = true;
4478 }
4479
4480 // Generate the initial go-slow test. Make sure we do not overflow
4481 // if length is huge (near 2Gig) or negative! We do not need
4482 // exact double-words here, just a close approximation of needed
4483 // double-words. We can't add any offset or rounding bits, lest we
4484 // take a size -1 of bytes and make it positive. Use an unsigned
4485 // compare, so negative sizes look hugely positive.
4486 int fast_size_limit = FastAllocateSizeLimit;
4487 if (layout_is_con) {
4488 assert(!StressReflectiveCode, "stress mode does not use these paths");
4489 // Increase the size limit if we have exact knowledge of array type.
4490 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4491 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4492 }
4493
4494 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4495 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4496
4497 // --- Size Computation ---
4498 // array_size = round_to_heap(array_header + (length << elem_shift));
4499 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4500 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4501 // The rounding mask is strength-reduced, if possible.
4502 int round_mask = MinObjAlignmentInBytes - 1;
4503 Node* header_size = nullptr;
4504 // (T_BYTE has the weakest alignment and size restrictions...)
4505 if (layout_is_con) {
4506 int hsize = Klass::layout_helper_header_size(layout_con);
4507 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4508 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4509 if ((round_mask & ~right_n_bits(eshift)) == 0)
4510 round_mask = 0; // strength-reduce it if it goes away completely
4511 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4512 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4513 assert(header_size_min <= hsize, "generic minimum is smallest");
4514 header_size = intcon(hsize);
4515 } else {
4516 Node* hss = intcon(Klass::_lh_header_size_shift);
4517 Node* hsm = intcon(Klass::_lh_header_size_mask);
4518 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4519 header_size = _gvn.transform(new AndINode(header_size, hsm));
4520 }
4521
4522 Node* elem_shift = nullptr;
4523 if (layout_is_con) {
4524 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4525 if (eshift != 0)
4526 elem_shift = intcon(eshift);
4527 } else {
4528 // There is no need to mask or shift this value.
4529 // The semantics of LShiftINode include an implicit mask to 0x1F.
4530 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4531 elem_shift = layout_val;
4532 }
4533
4534 // Transition to native address size for all offset calculations:
4535 Node* lengthx = ConvI2X(length);
4536 Node* headerx = ConvI2X(header_size);
4537 #ifdef _LP64
4538 { const TypeInt* tilen = _gvn.find_int_type(length);
4539 if (tilen != nullptr && tilen->_lo < 0) {
4540 // Add a manual constraint to a positive range. Cf. array_element_address.
4541 jint size_max = fast_size_limit;
4542 if (size_max > tilen->_hi && tilen->_hi >= 0) {
4543 size_max = tilen->_hi;
4544 }
4545 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4546
4547 // Only do a narrow I2L conversion if the range check passed.
4548 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4549 _gvn.transform(iff);
4550 RegionNode* region = new RegionNode(3);
4551 _gvn.set_type(region, Type::CONTROL);
4552 lengthx = new PhiNode(region, TypeLong::LONG);
4553 _gvn.set_type(lengthx, TypeLong::LONG);
4554
4555 // Range check passed. Use ConvI2L node with narrow type.
4556 Node* passed = IfFalse(iff);
4557 region->init_req(1, passed);
4558 // Make I2L conversion control dependent to prevent it from
4559 // floating above the range check during loop optimizations.
4560 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4561
4562 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4563 region->init_req(2, IfTrue(iff));
4564 lengthx->init_req(2, ConvI2X(length));
4565
4566 set_control(region);
4567 record_for_igvn(region);
4568 record_for_igvn(lengthx);
4569 }
4570 }
4571 #endif
4572
4573 // Combine header size and body size for the array copy part, then align (if
4574 // necessary) for the allocation part. This computation cannot overflow,
4575 // because it is used only in two places, one where the length is sharply
4576 // limited, and the other after a successful allocation.
4577 Node* abody = lengthx;
4578 if (elem_shift != nullptr) {
4579 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4580 }
4581 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4582
4583 if (return_size_val != nullptr) {
4584 // This is the size
4585 (*return_size_val) = non_rounded_size;
4586 }
4587
4588 Node* size = non_rounded_size;
4589 if (round_mask != 0) {
4590 Node* mask1 = MakeConX(round_mask);
4591 size = _gvn.transform(new AddXNode(size, mask1));
4592 Node* mask2 = MakeConX(~round_mask);
4593 size = _gvn.transform(new AndXNode(size, mask2));
4594 }
4595 // else if round_mask == 0, the size computation is self-rounding
4596
4597 // Now generate allocation code
4598
4599 // The entire memory state is needed for slow path of the allocation
4600 // since GC and deoptimization can happen.
4601 Node *mem = reset_memory();
4602 set_all_memory(mem); // Create new memory state
4603
4604 if (initial_slow_test->is_Bool()) {
4605 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4606 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4607 }
4608
4609 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4610 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4611
4612 Node* raw_init_value = nullptr;
4613 if (init_val != nullptr) {
4614 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4615 if (ary_type->is_flat()) {
4616 initial_slow_test = intcon(1);
4617 }
4618
4619 if (UseCompressedOops) {
4620 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4621 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4622 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4623 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4624 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4625 } else {
4626 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4627 }
4628 }
4629
4630 Node* valid_length_test = _gvn.intcon(1);
4631 if (ary_type->isa_aryptr()) {
4632 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4633 jint max = TypeAryPtr::max_array_length(bt);
4634 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4635 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4636 }
4637
4638 // Create the AllocateArrayNode and its result projections
4639 AllocateArrayNode* alloc
4640 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4641 control(), mem, i_o(),
4642 size, klass_node,
4643 initial_slow_test,
4644 length, valid_length_test,
4645 init_val, raw_init_value);
4646 // Cast to correct type. Note that the klass_node may be constant or not,
4647 // and in the latter case the actual array type will be inexact also.
4648 // (This happens via a non-constant argument to inline_native_newArray.)
4649 // In any case, the value of klass_node provides the desired array type.
4650 const TypeInt* length_type = _gvn.find_int_type(length);
4651 if (ary_type->isa_aryptr() && length_type != nullptr) {
4652 // Try to get a better type than POS for the size
4653 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4654 }
4655
4656 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4657
4658 array_ideal_length(alloc, ary_type, true);
4659 return javaoop;
4660 }
4661
4662 // The following "Ideal_foo" functions are placed here because they recognize
4663 // the graph shapes created by the functions immediately above.
4664
4665 //---------------------------Ideal_allocation----------------------------------
4666 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4667 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4668 if (ptr == nullptr) { // reduce dumb test in callers
4669 return nullptr;
4670 }
4671
4672 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4673 ptr = bs->step_over_gc_barrier(ptr);
4674
4675 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4676 ptr = ptr->in(1);
4677 if (ptr == nullptr) return nullptr;
4678 }
4679 // Return null for allocations with several casts:
4680 // j.l.reflect.Array.newInstance(jobject, jint)
4681 // Object.clone()
4682 // to keep more precise type from last cast.
4683 if (ptr->is_Proj()) {
4684 Node* allo = ptr->in(0);
4685 if (allo != nullptr && allo->is_Allocate()) {
4686 return allo->as_Allocate();
4687 }
4688 }
4689 // Report failure to match.
4690 return nullptr;
4691 }
4692
4693 // Fancy version which also strips off an offset (and reports it to caller).
4694 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4695 intptr_t& offset) {
4696 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4697 if (base == nullptr) return nullptr;
4698 return Ideal_allocation(base);
4699 }
4700
4701 // Trace Initialize <- Proj[Parm] <- Allocate
4702 AllocateNode* InitializeNode::allocation() {
4703 Node* rawoop = in(InitializeNode::RawAddress);
4704 if (rawoop->is_Proj()) {
4705 Node* alloc = rawoop->in(0);
4706 if (alloc->is_Allocate()) {
4707 return alloc->as_Allocate();
4708 }
4709 }
4710 return nullptr;
4711 }
4712
4713 // Trace Allocate -> Proj[Parm] -> Initialize
4714 InitializeNode* AllocateNode::initialization() {
4715 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4716 if (rawoop == nullptr) return nullptr;
4717 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4718 Node* init = rawoop->fast_out(i);
4719 if (init->is_Initialize()) {
4720 assert(init->as_Initialize()->allocation() == this, "2-way link");
4721 return init->as_Initialize();
4722 }
4723 }
4724 return nullptr;
4725 }
4726
4727 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4728 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4729 // Too many traps seen?
4730 if (too_many_traps(reason)) {
4731 #ifdef ASSERT
4732 if (TraceLoopPredicate) {
4733 int tc = C->trap_count(reason);
4734 tty->print("too many traps=%s tcount=%d in ",
4735 Deoptimization::trap_reason_name(reason), tc);
4736 method()->print(); // which method has too many predicate traps
4737 tty->cr();
4738 }
4739 #endif
4740 // We cannot afford to take more traps here,
4741 // do not generate Parse Predicate.
4742 return;
4743 }
4744
4745 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4746 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4747 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4748 {
4749 PreserveJVMState pjvms(this);
4750 set_control(if_false);
4751 inc_sp(nargs);
4752 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4753 }
4754 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4755 set_control(if_true);
4756 }
4757
4758 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4759 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4760 void GraphKit::add_parse_predicates(int nargs) {
4761 if (ShortRunningLongLoop) {
4762 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4763 // walking up from the loop.
4764 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4765 }
4766 if (UseLoopPredicate) {
4767 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4768 if (UseProfiledLoopPredicate) {
4769 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4770 }
4771 }
4772 if (UseAutoVectorizationPredicate) {
4773 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4774 }
4775 // Loop Limit Check Predicate should be near the loop.
4776 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4777 }
4778
4779 void GraphKit::sync_kit(IdealKit& ideal) {
4780 reset_memory();
4781 set_all_memory(ideal.merged_memory());
4782 set_i_o(ideal.i_o());
4783 set_control(ideal.ctrl());
4784 }
4785
4786 void GraphKit::final_sync(IdealKit& ideal) {
4787 // Final sync IdealKit and graphKit.
4788 sync_kit(ideal);
4789 }
4790
4791 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4792 Node* len = load_array_length(load_String_value(str, set_ctrl));
4793 Node* coder = load_String_coder(str, set_ctrl);
4794 // Divide length by 2 if coder is UTF16
4795 return _gvn.transform(new RShiftINode(len, coder));
4796 }
4797
4798 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4799 int value_offset = java_lang_String::value_offset();
4800 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4801 false, nullptr, Type::Offset(0));
4802 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4803 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4804 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4805 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4806 Node* p = basic_plus_adr(str, str, value_offset);
4807 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4808 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4809 return load;
4810 }
4811
4812 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4813 if (!CompactStrings) {
4814 return intcon(java_lang_String::CODER_UTF16);
4815 }
4816 int coder_offset = java_lang_String::coder_offset();
4817 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4818 false, nullptr, Type::Offset(0));
4819 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4820
4821 Node* p = basic_plus_adr(str, str, coder_offset);
4822 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4823 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4824 return load;
4825 }
4826
4827 void GraphKit::store_String_value(Node* str, Node* value) {
4828 int value_offset = java_lang_String::value_offset();
4829 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4830 false, nullptr, Type::Offset(0));
4831 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4832
4833 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4834 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4835 }
4836
4837 void GraphKit::store_String_coder(Node* str, Node* value) {
4838 int coder_offset = java_lang_String::coder_offset();
4839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4840 false, nullptr, Type::Offset(0));
4841 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4842
4843 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4844 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4845 }
4846
4847 // Capture src and dst memory state with a MergeMemNode
4848 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4849 if (src_type == dst_type) {
4850 // Types are equal, we don't need a MergeMemNode
4851 return memory(src_type);
4852 }
4853 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4854 record_for_igvn(merge); // fold it up later, if possible
4855 int src_idx = C->get_alias_index(src_type);
4856 int dst_idx = C->get_alias_index(dst_type);
4857 merge->set_memory_at(src_idx, memory(src_idx));
4858 merge->set_memory_at(dst_idx, memory(dst_idx));
4859 return merge;
4860 }
4861
4862 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4863 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4864 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4865 // If input and output memory types differ, capture both states to preserve
4866 // the dependency between preceding and subsequent loads/stores.
4867 // For example, the following program:
4868 // StoreB
4869 // compress_string
4870 // LoadB
4871 // has this memory graph (use->def):
4872 // LoadB -> compress_string -> CharMem
4873 // ... -> StoreB -> ByteMem
4874 // The intrinsic hides the dependency between LoadB and StoreB, causing
4875 // the load to read from memory not containing the result of the StoreB.
4876 // The correct memory graph should look like this:
4877 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4878 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4879 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4880 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4881 set_memory(res_mem, TypeAryPtr::BYTES);
4882 return str;
4883 }
4884
4885 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4886 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4887 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4888 // Capture src and dst memory (see comment in 'compress_string').
4889 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4890 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4891 set_memory(_gvn.transform(str), dst_type);
4892 }
4893
4894 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4895 /**
4896 * int i_char = start;
4897 * for (int i_byte = 0; i_byte < count; i_byte++) {
4898 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4899 * }
4900 */
4901 add_parse_predicates();
4902 C->set_has_loops(true);
4903
4904 RegionNode* head = new RegionNode(3);
4905 head->init_req(1, control());
4906 gvn().set_type(head, Type::CONTROL);
4907 record_for_igvn(head);
4908
4909 Node* i_byte = new PhiNode(head, TypeInt::INT);
4910 i_byte->init_req(1, intcon(0));
4911 gvn().set_type(i_byte, TypeInt::INT);
4912 record_for_igvn(i_byte);
4913
4914 Node* i_char = new PhiNode(head, TypeInt::INT);
4915 i_char->init_req(1, start);
4916 gvn().set_type(i_char, TypeInt::INT);
4917 record_for_igvn(i_char);
4918
4919 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4920 gvn().set_type(mem, Type::MEMORY);
4921 record_for_igvn(mem);
4922 set_control(head);
4923 set_memory(mem, TypeAryPtr::BYTES);
4924 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4925 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4926 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4927 false, true /* mismatched */);
4928
4929 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4930 head->init_req(2, IfTrue(iff));
4931 mem->init_req(2, st);
4932 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4933 i_char->init_req(2, AddI(i_char, intcon(2)));
4934
4935 set_control(IfFalse(iff));
4936 set_memory(st, TypeAryPtr::BYTES);
4937 }
4938
4939 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4940 if (!field->is_constant()) {
4941 return nullptr; // Field not marked as constant.
4942 }
4943 ciInstance* holder = nullptr;
4944 if (!field->is_static()) {
4945 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4946 if (const_oop != nullptr && const_oop->is_instance()) {
4947 holder = const_oop->as_instance();
4948 }
4949 }
4950 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4951 /*is_unsigned_load=*/false);
4952 if (con_type != nullptr) {
4953 Node* con = makecon(con_type);
4954 if (field->type()->is_inlinetype()) {
4955 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4956 } else if (con_type->is_inlinetypeptr()) {
4957 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4958 }
4959 return con;
4960 }
4961 return nullptr;
4962 }
4963
4964 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
4965 const Type* obj_type = obj->bottom_type();
4966 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4967 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4968 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4969 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4970 obj = casted_obj;
4971 }
4972 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
4973 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4974 }
4975 return obj;
4976 }