1 /*
2 * Copyright (c) 2001, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/parse.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/runtime.hpp"
55 #include "opto/subtypenode.hpp"
56 #include "runtime/arguments.hpp"
57 #include "runtime/deoptimization.hpp"
58 #include "runtime/sharedRuntime.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "utilities/bitMap.inline.hpp"
61 #include "utilities/growableArray.hpp"
62 #include "utilities/powerOfTwo.hpp"
63
64 //----------------------------GraphKit-----------------------------------------
65 // Main utility constructor.
66 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
67 : Phase(Phase::Parser),
68 _env(C->env()),
69 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
70 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
71 {
72 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
73 _exceptions = jvms->map()->next_exception();
74 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
75 set_jvms(jvms);
76 #ifdef ASSERT
77 if (_gvn.is_IterGVN() != nullptr) {
78 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
79 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
80 _worklist_size = _gvn.C->igvn_worklist()->size();
81 }
82 #endif
83 }
84
85 // Private constructor for parser.
86 GraphKit::GraphKit()
87 : Phase(Phase::Parser),
88 _env(C->env()),
89 _gvn(*C->initial_gvn()),
90 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
91 {
92 _exceptions = nullptr;
93 set_map(nullptr);
94 DEBUG_ONLY(_sp = -99);
95 DEBUG_ONLY(set_bci(-99));
96 }
97
98 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
99 : Phase(Phase::Parser),
100 _env(C->env()),
101 _gvn(igvn),
102 _exceptions(nullptr),
103 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
104 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
105 assert(sft->next_exception() == nullptr, "must not have a pending exception");
106 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
107 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
108 for (uint i = 0; i < sft->req(); i++) {
109 cloned_map->init_req(i, sft->in(i));
110 }
111 igvn.record_for_igvn(cloned_map);
112 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
113 current->set_map(cloned_map);
114 }
115 set_jvms(cloned_jvms);
116 set_all_memory(reset_memory());
117 }
118
119 //---------------------------clean_stack---------------------------------------
120 // Clear away rubbish from the stack area of the JVM state.
121 // This destroys any arguments that may be waiting on the stack.
122 void GraphKit::clean_stack(int from_sp) {
123 SafePointNode* map = this->map();
124 JVMState* jvms = this->jvms();
125 int stk_size = jvms->stk_size();
126 int stkoff = jvms->stkoff();
127 Node* top = this->top();
128 for (int i = from_sp; i < stk_size; i++) {
129 if (map->in(stkoff + i) != top) {
130 map->set_req(stkoff + i, top);
131 }
132 }
133 }
134
135
136 //--------------------------------sync_jvms-----------------------------------
137 // Make sure our current jvms agrees with our parse state.
138 JVMState* GraphKit::sync_jvms() const {
139 JVMState* jvms = this->jvms();
140 jvms->set_bci(bci()); // Record the new bci in the JVMState
141 jvms->set_sp(sp()); // Record the new sp in the JVMState
142 assert(jvms_in_sync(), "jvms is now in sync");
143 return jvms;
144 }
145
146 //--------------------------------sync_jvms_for_reexecute---------------------
147 // Make sure our current jvms agrees with our parse state. This version
148 // uses the reexecute_sp for reexecuting bytecodes.
149 JVMState* GraphKit::sync_jvms_for_reexecute() {
150 JVMState* jvms = this->jvms();
151 jvms->set_bci(bci()); // Record the new bci in the JVMState
152 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
153 return jvms;
154 }
155
156 #ifdef ASSERT
157 bool GraphKit::jvms_in_sync() const {
158 Parse* parse = is_Parse();
159 if (parse == nullptr) {
160 if (bci() != jvms()->bci()) return false;
161 if (sp() != (int)jvms()->sp()) return false;
162 return true;
163 }
164 if (jvms()->method() != parse->method()) return false;
165 if (jvms()->bci() != parse->bci()) return false;
166 int jvms_sp = jvms()->sp();
167 if (jvms_sp != parse->sp()) return false;
168 int jvms_depth = jvms()->depth();
169 if (jvms_depth != parse->depth()) return false;
170 return true;
171 }
172
173 // Local helper checks for special internal merge points
174 // used to accumulate and merge exception states.
175 // They are marked by the region's in(0) edge being the map itself.
176 // Such merge points must never "escape" into the parser at large,
177 // until they have been handed to gvn.transform.
178 static bool is_hidden_merge(Node* reg) {
179 if (reg == nullptr) return false;
180 if (reg->is_Phi()) {
181 reg = reg->in(0);
182 if (reg == nullptr) return false;
183 }
184 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
185 }
186
187 void GraphKit::verify_map() const {
188 if (map() == nullptr) return; // null map is OK
189 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
190 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
191 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
192 }
193
194 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
195 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
196 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
197 }
198 #endif
199
200 //---------------------------stop_and_kill_map---------------------------------
201 // Set _map to null, signalling a stop to further bytecode execution.
202 // First smash the current map's control to a constant, to mark it dead.
203 void GraphKit::stop_and_kill_map() {
204 SafePointNode* dead_map = stop();
205 if (dead_map != nullptr) {
206 dead_map->disconnect_inputs(C); // Mark the map as killed.
207 assert(dead_map->is_killed(), "must be so marked");
208 }
209 }
210
211
212 //--------------------------------stopped--------------------------------------
213 // Tell if _map is null, or control is top.
214 bool GraphKit::stopped() {
215 if (map() == nullptr) return true;
216 else if (control() == top()) return true;
217 else return false;
218 }
219
220
221 //-----------------------------has_exception_handler----------------------------------
222 // Tell if this method or any caller method has exception handlers.
223 bool GraphKit::has_exception_handler() {
224 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
225 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
226 return true;
227 }
228 }
229 return false;
230 }
231
232 //------------------------------save_ex_oop------------------------------------
233 // Save an exception without blowing stack contents or other JVM state.
234 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
235 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
236 ex_map->add_req(ex_oop);
237 DEBUG_ONLY(verify_exception_state(ex_map));
238 }
239
240 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
241 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
242 Node* ex_oop = ex_map->in(ex_map->req()-1);
243 if (clear_it) ex_map->del_req(ex_map->req()-1);
244 return ex_oop;
245 }
246
247 //-----------------------------saved_ex_oop------------------------------------
248 // Recover a saved exception from its map.
249 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
250 return common_saved_ex_oop(ex_map, false);
251 }
252
253 //--------------------------clear_saved_ex_oop---------------------------------
254 // Erase a previously saved exception from its map.
255 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
256 return common_saved_ex_oop(ex_map, true);
257 }
258
259 #ifdef ASSERT
260 //---------------------------has_saved_ex_oop----------------------------------
261 // Erase a previously saved exception from its map.
262 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
263 return ex_map->req() == ex_map->jvms()->endoff()+1;
264 }
265 #endif
266
267 //-------------------------make_exception_state--------------------------------
268 // Turn the current JVM state into an exception state, appending the ex_oop.
269 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
270 sync_jvms();
271 SafePointNode* ex_map = stop(); // do not manipulate this map any more
272 set_saved_ex_oop(ex_map, ex_oop);
273 return ex_map;
274 }
275
276
277 //--------------------------add_exception_state--------------------------------
278 // Add an exception to my list of exceptions.
279 void GraphKit::add_exception_state(SafePointNode* ex_map) {
280 if (ex_map == nullptr || ex_map->control() == top()) {
281 return;
282 }
283 #ifdef ASSERT
284 verify_exception_state(ex_map);
285 if (has_exceptions()) {
286 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
287 }
288 #endif
289
290 // If there is already an exception of exactly this type, merge with it.
291 // In particular, null-checks and other low-level exceptions common up here.
292 Node* ex_oop = saved_ex_oop(ex_map);
293 const Type* ex_type = _gvn.type(ex_oop);
294 if (ex_oop == top()) {
295 // No action needed.
296 return;
297 }
298 assert(ex_type->isa_instptr(), "exception must be an instance");
299 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
300 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
301 // We check sp also because call bytecodes can generate exceptions
302 // both before and after arguments are popped!
303 if (ex_type2 == ex_type
304 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
305 combine_exception_states(ex_map, e2);
306 return;
307 }
308 }
309
310 // No pre-existing exception of the same type. Chain it on the list.
311 push_exception_state(ex_map);
312 }
313
314 //-----------------------add_exception_states_from-----------------------------
315 void GraphKit::add_exception_states_from(JVMState* jvms) {
316 SafePointNode* ex_map = jvms->map()->next_exception();
317 if (ex_map != nullptr) {
318 jvms->map()->set_next_exception(nullptr);
319 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
320 next_map = ex_map->next_exception();
321 ex_map->set_next_exception(nullptr);
322 add_exception_state(ex_map);
323 }
324 }
325 }
326
327 //-----------------------transfer_exceptions_into_jvms-------------------------
328 JVMState* GraphKit::transfer_exceptions_into_jvms() {
329 if (map() == nullptr) {
330 // We need a JVMS to carry the exceptions, but the map has gone away.
331 // Create a scratch JVMS, cloned from any of the exception states...
332 if (has_exceptions()) {
333 _map = _exceptions;
334 _map = clone_map();
335 _map->set_next_exception(nullptr);
336 clear_saved_ex_oop(_map);
337 DEBUG_ONLY(verify_map());
338 } else {
339 // ...or created from scratch
340 JVMState* jvms = new (C) JVMState(_method, nullptr);
341 jvms->set_bci(_bci);
342 jvms->set_sp(_sp);
343 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
344 set_jvms(jvms);
345 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
346 set_all_memory(top());
347 while (map()->req() < jvms->endoff()) map()->add_req(top());
348 }
349 // (This is a kludge, in case you didn't notice.)
350 set_control(top());
351 }
352 JVMState* jvms = sync_jvms();
353 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
354 jvms->map()->set_next_exception(_exceptions);
355 _exceptions = nullptr; // done with this set of exceptions
356 return jvms;
357 }
358
359 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
360 assert(is_hidden_merge(dstphi), "must be a special merge node");
361 assert(is_hidden_merge(srcphi), "must be a special merge node");
362 uint limit = srcphi->req();
363 for (uint i = PhiNode::Input; i < limit; i++) {
364 dstphi->add_req(srcphi->in(i));
365 }
366 }
367 static inline void add_one_req(Node* dstphi, Node* src) {
368 assert(is_hidden_merge(dstphi), "must be a special merge node");
369 assert(!is_hidden_merge(src), "must not be a special merge node");
370 dstphi->add_req(src);
371 }
372
373 //-----------------------combine_exception_states------------------------------
374 // This helper function combines exception states by building phis on a
375 // specially marked state-merging region. These regions and phis are
376 // untransformed, and can build up gradually. The region is marked by
377 // having a control input of its exception map, rather than null. Such
378 // regions do not appear except in this function, and in use_exception_state.
379 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
380 if (failing_internal()) {
381 return; // dying anyway...
382 }
383 JVMState* ex_jvms = ex_map->_jvms;
384 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
385 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
386 // TODO 8325632 Re-enable
387 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
388 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
389 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
390 assert(ex_map->req() == phi_map->req(), "matching maps");
391 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
392 Node* hidden_merge_mark = root();
393 Node* region = phi_map->control();
394 MergeMemNode* phi_mem = phi_map->merged_memory();
395 MergeMemNode* ex_mem = ex_map->merged_memory();
396 if (region->in(0) != hidden_merge_mark) {
397 // The control input is not (yet) a specially-marked region in phi_map.
398 // Make it so, and build some phis.
399 region = new RegionNode(2);
400 _gvn.set_type(region, Type::CONTROL);
401 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
402 region->init_req(1, phi_map->control());
403 phi_map->set_control(region);
404 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
405 record_for_igvn(io_phi);
406 _gvn.set_type(io_phi, Type::ABIO);
407 phi_map->set_i_o(io_phi);
408 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
409 Node* m = mms.memory();
410 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
411 record_for_igvn(m_phi);
412 _gvn.set_type(m_phi, Type::MEMORY);
413 mms.set_memory(m_phi);
414 }
415 }
416
417 // Either or both of phi_map and ex_map might already be converted into phis.
418 Node* ex_control = ex_map->control();
419 // if there is special marking on ex_map also, we add multiple edges from src
420 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
421 // how wide was the destination phi_map, originally?
422 uint orig_width = region->req();
423
424 if (add_multiple) {
425 add_n_reqs(region, ex_control);
426 add_n_reqs(phi_map->i_o(), ex_map->i_o());
427 } else {
428 // ex_map has no merges, so we just add single edges everywhere
429 add_one_req(region, ex_control);
430 add_one_req(phi_map->i_o(), ex_map->i_o());
431 }
432 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
433 if (mms.is_empty()) {
434 // get a copy of the base memory, and patch some inputs into it
435 const TypePtr* adr_type = mms.adr_type(C);
436 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
437 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
438 mms.set_memory(phi);
439 // Prepare to append interesting stuff onto the newly sliced phi:
440 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
441 }
442 // Append stuff from ex_map:
443 if (add_multiple) {
444 add_n_reqs(mms.memory(), mms.memory2());
445 } else {
446 add_one_req(mms.memory(), mms.memory2());
447 }
448 }
449 uint limit = ex_map->req();
450 for (uint i = TypeFunc::Parms; i < limit; i++) {
451 // Skip everything in the JVMS after tos. (The ex_oop follows.)
452 if (i == tos) i = ex_jvms->monoff();
453 Node* src = ex_map->in(i);
454 Node* dst = phi_map->in(i);
455 if (src != dst) {
456 PhiNode* phi;
457 if (dst->in(0) != region) {
458 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
459 record_for_igvn(phi);
460 _gvn.set_type(phi, phi->type());
461 phi_map->set_req(i, dst);
462 // Prepare to append interesting stuff onto the new phi:
463 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
464 } else {
465 assert(dst->is_Phi(), "nobody else uses a hidden region");
466 phi = dst->as_Phi();
467 }
468 if (add_multiple && src->in(0) == ex_control) {
469 // Both are phis.
470 add_n_reqs(dst, src);
471 } else {
472 while (dst->req() < region->req()) add_one_req(dst, src);
473 }
474 const Type* srctype = _gvn.type(src);
475 if (phi->type() != srctype) {
476 const Type* dsttype = phi->type()->meet_speculative(srctype);
477 if (phi->type() != dsttype) {
478 phi->set_type(dsttype);
479 _gvn.set_type(phi, dsttype);
480 }
481 }
482 }
483 }
484 phi_map->merge_replaced_nodes_with(ex_map);
485 }
486
487 //--------------------------use_exception_state--------------------------------
488 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
489 if (failing_internal()) { stop(); return top(); }
490 Node* region = phi_map->control();
491 Node* hidden_merge_mark = root();
492 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
493 Node* ex_oop = clear_saved_ex_oop(phi_map);
494 if (region->in(0) == hidden_merge_mark) {
495 // Special marking for internal ex-states. Process the phis now.
496 region->set_req(0, region); // now it's an ordinary region
497 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
498 // Note: Setting the jvms also sets the bci and sp.
499 set_control(_gvn.transform(region));
500 uint tos = jvms()->stkoff() + sp();
501 for (uint i = 1; i < tos; i++) {
502 Node* x = phi_map->in(i);
503 if (x->in(0) == region) {
504 assert(x->is_Phi(), "expected a special phi");
505 phi_map->set_req(i, _gvn.transform(x));
506 }
507 }
508 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
509 Node* x = mms.memory();
510 if (x->in(0) == region) {
511 assert(x->is_Phi(), "nobody else uses a hidden region");
512 mms.set_memory(_gvn.transform(x));
513 }
514 }
515 if (ex_oop->in(0) == region) {
516 assert(ex_oop->is_Phi(), "expected a special phi");
517 ex_oop = _gvn.transform(ex_oop);
518 }
519 } else {
520 set_jvms(phi_map->jvms());
521 }
522
523 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
524 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
525 return ex_oop;
526 }
527
528 //---------------------------------java_bc-------------------------------------
529 Bytecodes::Code GraphKit::java_bc() const {
530 ciMethod* method = this->method();
531 int bci = this->bci();
532 if (method != nullptr && bci != InvocationEntryBci)
533 return method->java_code_at_bci(bci);
534 else
535 return Bytecodes::_illegal;
536 }
537
538 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
539 bool must_throw) {
540 // if the exception capability is set, then we will generate code
541 // to check the JavaThread.should_post_on_exceptions flag to see
542 // if we actually need to report exception events (for this
543 // thread). If we don't need to report exception events, we will
544 // take the normal fast path provided by add_exception_events. If
545 // exception event reporting is enabled for this thread, we will
546 // take the uncommon_trap in the BuildCutout below.
547
548 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
549 Node* jthread = _gvn.transform(new ThreadLocalNode());
550 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
551 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
552
553 // Test the should_post_on_exceptions_flag vs. 0
554 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
555 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
556
557 // Branch to slow_path if should_post_on_exceptions_flag was true
558 { BuildCutout unless(this, tst, PROB_MAX);
559 // Do not try anything fancy if we're notifying the VM on every throw.
560 // Cf. case Bytecodes::_athrow in parse2.cpp.
561 uncommon_trap(reason, Deoptimization::Action_none,
562 (ciKlass*)nullptr, (char*)nullptr, must_throw);
563 }
564
565 }
566
567 //------------------------------builtin_throw----------------------------------
568 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
569 builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true);
570 }
571
572 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason,
573 ciInstance* ex_obj,
574 bool allow_too_many_traps) {
575 // If this throw happens frequently, an uncommon trap might cause
576 // a performance pothole. If there is a local exception handler,
577 // and if this particular bytecode appears to be deoptimizing often,
578 // let us handle the throw inline, with a preconstructed instance.
579 // Note: If the deopt count has blown up, the uncommon trap
580 // runtime is going to flush this nmethod, not matter what.
581 if (is_builtin_throw_hot(reason)) {
582 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
583 // If the throw is local, we use a pre-existing instance and
584 // punt on the backtrace. This would lead to a missing backtrace
585 // (a repeat of 4292742) if the backtrace object is ever asked
586 // for its backtrace.
587 // Fixing this remaining case of 4292742 requires some flavor of
588 // escape analysis. Leave that for the future.
589 if (env()->jvmti_can_post_on_exceptions()) {
590 // check if we must post exception events, take uncommon trap if so
591 uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/);
592 // here if should_post_on_exceptions is false
593 // continue on with the normal codegen
594 }
595
596 // Cheat with a preallocated exception object.
597 if (C->log() != nullptr)
598 C->log()->elem("hot_throw preallocated='1' reason='%s'",
599 Deoptimization::trap_reason_name(reason));
600 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
601 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
602
603 // Clear the detail message of the preallocated exception object.
604 // Weblogic sometimes mutates the detail message of exceptions
605 // using reflection.
606 int offset = java_lang_Throwable::get_detailMessage_offset();
607 const TypePtr* adr_typ = ex_con->add_offset(offset);
608
609 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
610 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
611 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
612
613 if (!method()->has_exception_handlers()) {
614 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
615 // This prevents issues with exception handling and late inlining.
616 set_sp(0);
617 clean_stack(0);
618 }
619
620 add_exception_state(make_exception_state(ex_node));
621 return;
622 } else if (builtin_throw_too_many_traps(reason, ex_obj)) {
623 // We cannot afford to take too many traps here. Suffer in the interpreter instead.
624 assert(allow_too_many_traps, "not allowed");
625 if (C->log() != nullptr) {
626 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
627 Deoptimization::trap_reason_name(reason),
628 C->trap_count(reason));
629 }
630 uncommon_trap(reason, Deoptimization::Action_none,
631 (ciKlass*) nullptr, (char*) nullptr,
632 true /*must_throw*/);
633 return;
634 }
635 }
636
637 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
638 // It won't be much cheaper than bailing to the interp., since we'll
639 // have to pass up all the debug-info, and the runtime will have to
640 // create the stack trace.
641
642 // Usual case: Bail to interpreter.
643 // Reserve the right to recompile if we haven't seen anything yet.
644
645 // "must_throw" prunes the JVM state to include only the stack, if there
646 // are no local exception handlers. This should cut down on register
647 // allocation time and code size, by drastically reducing the number
648 // of in-edges on the call to the uncommon trap.
649 uncommon_trap(reason, Deoptimization::Action_maybe_recompile,
650 (ciKlass*) nullptr, (char*) nullptr,
651 true /*must_throw*/);
652 }
653
654 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) {
655 // If this particular condition has not yet happened at this
656 // bytecode, then use the uncommon trap mechanism, and allow for
657 // a future recompilation if several traps occur here.
658 // If the throw is hot, try to use a more complicated inline mechanism
659 // which keeps execution inside the compiled code.
660 if (ProfileTraps) {
661 if (too_many_traps(reason)) {
662 return true;
663 }
664 // (If there is no MDO at all, assume it is early in
665 // execution, and that any deopts are part of the
666 // startup transient, and don't need to be remembered.)
667
668 // Also, if there is a local exception handler, treat all throws
669 // as hot if there has been at least one in this method.
670 if (C->trap_count(reason) != 0 &&
671 method()->method_data()->trap_count(reason) != 0 &&
672 has_exception_handler()) {
673 return true;
674 }
675 }
676 return false;
677 }
678
679 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason,
680 ciInstance* ex_obj) {
681 if (is_builtin_throw_hot(reason)) {
682 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
683 return false; // no traps; throws preallocated exception instead
684 }
685 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
686 if (method()->method_data()->trap_recompiled_at(bci(), m) ||
687 C->too_many_traps(reason)) {
688 return true;
689 }
690 }
691 return false;
692 }
693
694 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const {
695 // Preallocated exception objects to use when we don't need the backtrace.
696 switch (reason) {
697 case Deoptimization::Reason_null_check:
698 return env()->NullPointerException_instance();
699 case Deoptimization::Reason_div0_check:
700 return env()->ArithmeticException_instance();
701 case Deoptimization::Reason_range_check:
702 return env()->ArrayIndexOutOfBoundsException_instance();
703 case Deoptimization::Reason_class_check:
704 return env()->ClassCastException_instance();
705 case Deoptimization::Reason_array_check:
706 return env()->ArrayStoreException_instance();
707 default:
708 return nullptr;
709 }
710 }
711
712 //----------------------------PreserveJVMState---------------------------------
713 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
714 DEBUG_ONLY(kit->verify_map());
715 _kit = kit;
716 _map = kit->map(); // preserve the map
717 _sp = kit->sp();
718 kit->set_map(clone_map ? kit->clone_map() : nullptr);
719 #ifdef ASSERT
720 _bci = kit->bci();
721 Parse* parser = kit->is_Parse();
722 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
723 _block = block;
724 #endif
725 }
726 PreserveJVMState::~PreserveJVMState() {
727 GraphKit* kit = _kit;
728 #ifdef ASSERT
729 assert(kit->bci() == _bci, "bci must not shift");
730 Parse* parser = kit->is_Parse();
731 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
732 assert(block == _block, "block must not shift");
733 #endif
734 kit->set_map(_map);
735 kit->set_sp(_sp);
736 }
737
738
739 //-----------------------------BuildCutout-------------------------------------
740 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
741 : PreserveJVMState(kit)
742 {
743 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
744 SafePointNode* outer_map = _map; // preserved map is caller's
745 SafePointNode* inner_map = kit->map();
746 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
747 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
748 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
749 }
750 BuildCutout::~BuildCutout() {
751 GraphKit* kit = _kit;
752 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
753 }
754
755 //---------------------------PreserveReexecuteState----------------------------
756 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
757 assert(!kit->stopped(), "must call stopped() before");
758 _kit = kit;
759 _sp = kit->sp();
760 _reexecute = kit->jvms()->_reexecute;
761 }
762 PreserveReexecuteState::~PreserveReexecuteState() {
763 if (_kit->stopped()) return;
764 _kit->jvms()->_reexecute = _reexecute;
765 _kit->set_sp(_sp);
766 }
767
768 //------------------------------clone_map--------------------------------------
769 // Implementation of PreserveJVMState
770 //
771 // Only clone_map(...) here. If this function is only used in the
772 // PreserveJVMState class we may want to get rid of this extra
773 // function eventually and do it all there.
774
775 SafePointNode* GraphKit::clone_map() {
776 if (map() == nullptr) return nullptr;
777
778 // Clone the memory edge first
779 Node* mem = MergeMemNode::make(map()->memory());
780 gvn().set_type_bottom(mem);
781
782 SafePointNode *clonemap = (SafePointNode*)map()->clone();
783 JVMState* jvms = this->jvms();
784 JVMState* clonejvms = jvms->clone_shallow(C);
785 clonemap->set_memory(mem);
786 clonemap->set_jvms(clonejvms);
787 clonejvms->set_map(clonemap);
788 record_for_igvn(clonemap);
789 gvn().set_type_bottom(clonemap);
790 return clonemap;
791 }
792
793 //-----------------------------destruct_map_clone------------------------------
794 //
795 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
796 // to destruct/free/delete in the exact opposite order as clone_map().
797 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
798 if (sfp == nullptr) return;
799
800 Node* mem = sfp->memory();
801 JVMState* jvms = sfp->jvms();
802
803 if (jvms != nullptr) {
804 delete jvms;
805 }
806
807 remove_for_igvn(sfp);
808 gvn().clear_type(sfp);
809 sfp->destruct(&_gvn);
810
811 if (mem != nullptr) {
812 gvn().clear_type(mem);
813 mem->destruct(&_gvn);
814 }
815 }
816
817 //-----------------------------set_map_clone-----------------------------------
818 void GraphKit::set_map_clone(SafePointNode* m) {
819 _map = m;
820 _map = clone_map();
821 _map->set_next_exception(nullptr);
822 DEBUG_ONLY(verify_map());
823 }
824
825
826 //----------------------------kill_dead_locals---------------------------------
827 // Detect any locals which are known to be dead, and force them to top.
828 void GraphKit::kill_dead_locals() {
829 // Consult the liveness information for the locals. If any
830 // of them are unused, then they can be replaced by top(). This
831 // should help register allocation time and cut down on the size
832 // of the deoptimization information.
833
834 // This call is made from many of the bytecode handling
835 // subroutines called from the Big Switch in do_one_bytecode.
836 // Every bytecode which might include a slow path is responsible
837 // for killing its dead locals. The more consistent we
838 // are about killing deads, the fewer useless phis will be
839 // constructed for them at various merge points.
840
841 // bci can be -1 (InvocationEntryBci). We return the entry
842 // liveness for the method.
843
844 if (method() == nullptr || method()->code_size() == 0) {
845 // We are building a graph for a call to a native method.
846 // All locals are live.
847 return;
848 }
849
850 ResourceMark rm;
851
852 // Consult the liveness information for the locals. If any
853 // of them are unused, then they can be replaced by top(). This
854 // should help register allocation time and cut down on the size
855 // of the deoptimization information.
856 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
857
858 int len = (int)live_locals.size();
859 assert(len <= jvms()->loc_size(), "too many live locals");
860 for (int local = 0; local < len; local++) {
861 if (!live_locals.at(local)) {
862 set_local(local, top());
863 }
864 }
865 }
866
867 #ifdef ASSERT
868 //-------------------------dead_locals_are_killed------------------------------
869 // Return true if all dead locals are set to top in the map.
870 // Used to assert "clean" debug info at various points.
871 bool GraphKit::dead_locals_are_killed() {
872 if (method() == nullptr || method()->code_size() == 0) {
873 // No locals need to be dead, so all is as it should be.
874 return true;
875 }
876
877 // Make sure somebody called kill_dead_locals upstream.
878 ResourceMark rm;
879 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
880 if (jvms->loc_size() == 0) continue; // no locals to consult
881 SafePointNode* map = jvms->map();
882 ciMethod* method = jvms->method();
883 int bci = jvms->bci();
884 if (jvms == this->jvms()) {
885 bci = this->bci(); // it might not yet be synched
886 }
887 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
888 int len = (int)live_locals.size();
889 if (!live_locals.is_valid() || len == 0)
890 // This method is trivial, or is poisoned by a breakpoint.
891 return true;
892 assert(len == jvms->loc_size(), "live map consistent with locals map");
893 for (int local = 0; local < len; local++) {
894 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
895 if (PrintMiscellaneous && (Verbose || WizardMode)) {
896 tty->print_cr("Zombie local %d: ", local);
897 jvms->dump();
898 }
899 return false;
900 }
901 }
902 }
903 return true;
904 }
905
906 #endif //ASSERT
907
908 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
909 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
910 ciMethod* cur_method = jvms->method();
911 int cur_bci = jvms->bci();
912 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
913 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
914 return Interpreter::bytecode_should_reexecute(code) ||
915 (is_anewarray && (code == Bytecodes::_multianewarray));
916 // Reexecute _multianewarray bytecode which was replaced with
917 // sequence of [a]newarray. See Parse::do_multianewarray().
918 //
919 // Note: interpreter should not have it set since this optimization
920 // is limited by dimensions and guarded by flag so in some cases
921 // multianewarray() runtime calls will be generated and
922 // the bytecode should not be reexecutes (stack will not be reset).
923 } else {
924 return false;
925 }
926 }
927
928 // Helper function for adding JVMState and debug information to node
929 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
930 // Add the safepoint edges to the call (or other safepoint).
931
932 // Make sure dead locals are set to top. This
933 // should help register allocation time and cut down on the size
934 // of the deoptimization information.
935 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
936
937 // Walk the inline list to fill in the correct set of JVMState's
938 // Also fill in the associated edges for each JVMState.
939
940 // If the bytecode needs to be reexecuted we need to put
941 // the arguments back on the stack.
942 const bool should_reexecute = jvms()->should_reexecute();
943 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
944
945 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
946 // undefined if the bci is different. This is normal for Parse but it
947 // should not happen for LibraryCallKit because only one bci is processed.
948 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
949 "in LibraryCallKit the reexecute bit should not change");
950
951 // If we are guaranteed to throw, we can prune everything but the
952 // input to the current bytecode.
953 bool can_prune_locals = false;
954 uint stack_slots_not_pruned = 0;
955 int inputs = 0, depth = 0;
956 if (must_throw) {
957 assert(method() == youngest_jvms->method(), "sanity");
958 if (compute_stack_effects(inputs, depth)) {
959 can_prune_locals = true;
960 stack_slots_not_pruned = inputs;
961 }
962 }
963
964 if (env()->should_retain_local_variables()) {
965 // At any safepoint, this method can get breakpointed, which would
966 // then require an immediate deoptimization.
967 can_prune_locals = false; // do not prune locals
968 stack_slots_not_pruned = 0;
969 }
970
971 // do not scribble on the input jvms
972 JVMState* out_jvms = youngest_jvms->clone_deep(C);
973 call->set_jvms(out_jvms); // Start jvms list for call node
974
975 // For a known set of bytecodes, the interpreter should reexecute them if
976 // deoptimization happens. We set the reexecute state for them here
977 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
978 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
979 #ifdef ASSERT
980 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
981 assert(method() == youngest_jvms->method(), "sanity");
982 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
983 // TODO 8371125
984 // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
985 #endif // ASSERT
986 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
987 }
988
989 // Presize the call:
990 DEBUG_ONLY(uint non_debug_edges = call->req());
991 call->add_req_batch(top(), youngest_jvms->debug_depth());
992 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
993
994 // Set up edges so that the call looks like this:
995 // Call [state:] ctl io mem fptr retadr
996 // [parms:] parm0 ... parmN
997 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
998 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
999 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1000 // Note that caller debug info precedes callee debug info.
1001
1002 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1003 uint debug_ptr = call->req();
1004
1005 // Loop over the map input edges associated with jvms, add them
1006 // to the call node, & reset all offsets to match call node array.
1007
1008 JVMState* callee_jvms = nullptr;
1009 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1010 uint debug_end = debug_ptr;
1011 uint debug_start = debug_ptr - in_jvms->debug_size();
1012 debug_ptr = debug_start; // back up the ptr
1013
1014 uint p = debug_start; // walks forward in [debug_start, debug_end)
1015 uint j, k, l;
1016 SafePointNode* in_map = in_jvms->map();
1017 out_jvms->set_map(call);
1018
1019 if (can_prune_locals) {
1020 assert(in_jvms->method() == out_jvms->method(), "sanity");
1021 // If the current throw can reach an exception handler in this JVMS,
1022 // then we must keep everything live that can reach that handler.
1023 // As a quick and dirty approximation, we look for any handlers at all.
1024 if (in_jvms->method()->has_exception_handlers()) {
1025 can_prune_locals = false;
1026 }
1027 }
1028
1029 // Add the Locals
1030 k = in_jvms->locoff();
1031 l = in_jvms->loc_size();
1032 out_jvms->set_locoff(p);
1033 if (!can_prune_locals) {
1034 for (j = 0; j < l; j++) {
1035 call->set_req(p++, in_map->in(k + j));
1036 }
1037 } else {
1038 p += l; // already set to top above by add_req_batch
1039 }
1040
1041 // Add the Expression Stack
1042 k = in_jvms->stkoff();
1043 l = in_jvms->sp();
1044 out_jvms->set_stkoff(p);
1045 if (!can_prune_locals) {
1046 for (j = 0; j < l; j++) {
1047 call->set_req(p++, in_map->in(k + j));
1048 }
1049 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1050 // Divide stack into {S0,...,S1}, where S0 is set to top.
1051 uint s1 = stack_slots_not_pruned;
1052 stack_slots_not_pruned = 0; // for next iteration
1053 if (s1 > l) s1 = l;
1054 uint s0 = l - s1;
1055 p += s0; // skip the tops preinstalled by add_req_batch
1056 for (j = s0; j < l; j++)
1057 call->set_req(p++, in_map->in(k+j));
1058 } else {
1059 p += l; // already set to top above by add_req_batch
1060 }
1061
1062 // Add the Monitors
1063 k = in_jvms->monoff();
1064 l = in_jvms->mon_size();
1065 out_jvms->set_monoff(p);
1066 for (j = 0; j < l; j++)
1067 call->set_req(p++, in_map->in(k+j));
1068
1069 // Copy any scalar object fields.
1070 k = in_jvms->scloff();
1071 l = in_jvms->scl_size();
1072 out_jvms->set_scloff(p);
1073 for (j = 0; j < l; j++)
1074 call->set_req(p++, in_map->in(k+j));
1075
1076 // Finish the new jvms.
1077 out_jvms->set_endoff(p);
1078
1079 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1080 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1081 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1082 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1083 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1084 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1085
1086 // Update the two tail pointers in parallel.
1087 callee_jvms = out_jvms;
1088 out_jvms = out_jvms->caller();
1089 in_jvms = in_jvms->caller();
1090 }
1091
1092 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1093
1094 // Test the correctness of JVMState::debug_xxx accessors:
1095 assert(call->jvms()->debug_start() == non_debug_edges, "");
1096 assert(call->jvms()->debug_end() == call->req(), "");
1097 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1098 }
1099
1100 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1101 Bytecodes::Code code = java_bc();
1102 if (code == Bytecodes::_wide) {
1103 code = method()->java_code_at_bci(bci() + 1);
1104 }
1105
1106 if (code != Bytecodes::_illegal) {
1107 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1108 }
1109
1110 auto rsize = [&]() {
1111 assert(code != Bytecodes::_illegal, "code is illegal!");
1112 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1113 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1114 };
1115
1116 switch (code) {
1117 case Bytecodes::_illegal:
1118 return false;
1119
1120 case Bytecodes::_ldc:
1121 case Bytecodes::_ldc_w:
1122 case Bytecodes::_ldc2_w:
1123 inputs = 0;
1124 break;
1125
1126 case Bytecodes::_dup: inputs = 1; break;
1127 case Bytecodes::_dup_x1: inputs = 2; break;
1128 case Bytecodes::_dup_x2: inputs = 3; break;
1129 case Bytecodes::_dup2: inputs = 2; break;
1130 case Bytecodes::_dup2_x1: inputs = 3; break;
1131 case Bytecodes::_dup2_x2: inputs = 4; break;
1132 case Bytecodes::_swap: inputs = 2; break;
1133 case Bytecodes::_arraylength: inputs = 1; break;
1134
1135 case Bytecodes::_getstatic:
1136 case Bytecodes::_putstatic:
1137 case Bytecodes::_getfield:
1138 case Bytecodes::_putfield:
1139 {
1140 bool ignored_will_link;
1141 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1142 int size = field->type()->size();
1143 bool is_get = (depth >= 0), is_static = (depth & 1);
1144 inputs = (is_static ? 0 : 1);
1145 if (is_get) {
1146 depth = size - inputs;
1147 } else {
1148 inputs += size; // putxxx pops the value from the stack
1149 depth = - inputs;
1150 }
1151 }
1152 break;
1153
1154 case Bytecodes::_invokevirtual:
1155 case Bytecodes::_invokespecial:
1156 case Bytecodes::_invokestatic:
1157 case Bytecodes::_invokedynamic:
1158 case Bytecodes::_invokeinterface:
1159 {
1160 bool ignored_will_link;
1161 ciSignature* declared_signature = nullptr;
1162 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1163 assert(declared_signature != nullptr, "cannot be null");
1164 inputs = declared_signature->arg_size_for_bc(code);
1165 int size = declared_signature->return_type()->size();
1166 depth = size - inputs;
1167 }
1168 break;
1169
1170 case Bytecodes::_multianewarray:
1171 {
1172 ciBytecodeStream iter(method());
1173 iter.reset_to_bci(bci());
1174 iter.next();
1175 inputs = iter.get_dimensions();
1176 assert(rsize() == 1, "");
1177 depth = 1 - inputs;
1178 }
1179 break;
1180
1181 case Bytecodes::_ireturn:
1182 case Bytecodes::_lreturn:
1183 case Bytecodes::_freturn:
1184 case Bytecodes::_dreturn:
1185 case Bytecodes::_areturn:
1186 assert(rsize() == -depth, "");
1187 inputs = -depth;
1188 break;
1189
1190 case Bytecodes::_jsr:
1191 case Bytecodes::_jsr_w:
1192 inputs = 0;
1193 depth = 1; // S.B. depth=1, not zero
1194 break;
1195
1196 default:
1197 // bytecode produces a typed result
1198 inputs = rsize() - depth;
1199 assert(inputs >= 0, "");
1200 break;
1201 }
1202
1203 #ifdef ASSERT
1204 // spot check
1205 int outputs = depth + inputs;
1206 assert(outputs >= 0, "sanity");
1207 switch (code) {
1208 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1209 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1210 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1211 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1212 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1213 default: break;
1214 }
1215 #endif //ASSERT
1216
1217 return true;
1218 }
1219
1220
1221
1222 //------------------------------basic_plus_adr---------------------------------
1223 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1224 // short-circuit a common case
1225 if (offset == MakeConX(0)) {
1226 return ptr;
1227 }
1228 #ifdef ASSERT
1229 // Both 32-bit and 64-bit zeros should have been handled by the previous `if`
1230 // statement, so if we see either 32-bit or 64-bit zeros here, then we have a
1231 // problem.
1232 if (offset->is_Con()) {
1233 const Type* t = offset->bottom_type();
1234 bool is_zero_int = t->isa_int() && t->is_int()->get_con() == 0;
1235 bool is_zero_long = t->isa_long() && t->is_long()->get_con() == 0;
1236 assert(!is_zero_int && !is_zero_long,
1237 "Unexpected zero offset - should have matched MakeConX(0)");
1238 }
1239 #endif
1240 return _gvn.transform( new AddPNode(base, ptr, offset) );
1241 }
1242
1243 Node* GraphKit::ConvI2L(Node* offset) {
1244 // short-circuit a common case
1245 jint offset_con = find_int_con(offset, Type::OffsetBot);
1246 if (offset_con != Type::OffsetBot) {
1247 return longcon((jlong) offset_con);
1248 }
1249 return _gvn.transform( new ConvI2LNode(offset));
1250 }
1251
1252 Node* GraphKit::ConvI2UL(Node* offset) {
1253 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1254 if (offset_con != (juint) Type::OffsetBot) {
1255 return longcon((julong) offset_con);
1256 }
1257 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1258 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1259 return _gvn.transform( new AndLNode(conv, mask) );
1260 }
1261
1262 Node* GraphKit::ConvL2I(Node* offset) {
1263 // short-circuit a common case
1264 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1265 if (offset_con != (jlong)Type::OffsetBot) {
1266 return intcon((int) offset_con);
1267 }
1268 return _gvn.transform( new ConvL2INode(offset));
1269 }
1270
1271 //-------------------------load_object_klass-----------------------------------
1272 Node* GraphKit::load_object_klass(Node* obj) {
1273 // Special-case a fresh allocation to avoid building nodes:
1274 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1275 if (akls != nullptr) return akls;
1276 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1277 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1278 }
1279
1280 //-------------------------load_array_length-----------------------------------
1281 Node* GraphKit::load_array_length(Node* array) {
1282 // Special-case a fresh allocation to avoid building nodes:
1283 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1284 Node *alen;
1285 if (alloc == nullptr) {
1286 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1287 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1288 } else {
1289 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1290 }
1291 return alen;
1292 }
1293
1294 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1295 const TypeOopPtr* oop_type,
1296 bool replace_length_in_map) {
1297 Node* length = alloc->Ideal_length();
1298 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1299 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1300 if (ccast != length) {
1301 // do not transform ccast here, it might convert to top node for
1302 // negative array length and break assumptions in parsing stage.
1303 _gvn.set_type_bottom(ccast);
1304 record_for_igvn(ccast);
1305 if (replace_length_in_map) {
1306 replace_in_map(length, ccast);
1307 }
1308 return ccast;
1309 }
1310 }
1311 return length;
1312 }
1313
1314 //------------------------------do_null_check----------------------------------
1315 // Helper function to do a null pointer check. Returned value is
1316 // the incoming address with null casted away. You are allowed to use the
1317 // not-null value only if you are control dependent on the test.
1318 #ifndef PRODUCT
1319 extern uint explicit_null_checks_inserted,
1320 explicit_null_checks_elided;
1321 #endif
1322 Node* GraphKit::null_check_common(Node* value, BasicType type,
1323 // optional arguments for variations:
1324 bool assert_null,
1325 Node* *null_control,
1326 bool speculative,
1327 bool null_marker_check) {
1328 assert(!assert_null || null_control == nullptr, "not both at once");
1329 if (stopped()) return top();
1330 NOT_PRODUCT(explicit_null_checks_inserted++);
1331
1332 if (value->is_InlineType()) {
1333 // Null checking a scalarized but nullable inline type. Check the null marker
1334 // input instead of the oop input to avoid keeping buffer allocations alive.
1335 InlineTypeNode* vtptr = value->as_InlineType();
1336 while (vtptr->get_oop()->is_InlineType()) {
1337 vtptr = vtptr->get_oop()->as_InlineType();
1338 }
1339 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1340 if (stopped()) {
1341 return top();
1342 }
1343 if (assert_null) {
1344 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1345 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1346 // replace_in_map(value, vtptr);
1347 // return vtptr;
1348 replace_in_map(value, null());
1349 return null();
1350 }
1351 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1352 return cast_not_null(value, do_replace_in_map);
1353 }
1354
1355 // Construct null check
1356 Node *chk = nullptr;
1357 switch(type) {
1358 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1359 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1360 case T_ARRAY : // fall through
1361 type = T_OBJECT; // simplify further tests
1362 case T_OBJECT : {
1363 const Type *t = _gvn.type( value );
1364
1365 const TypeOopPtr* tp = t->isa_oopptr();
1366 if (tp != nullptr && !tp->is_loaded()
1367 // Only for do_null_check, not any of its siblings:
1368 && !assert_null && null_control == nullptr) {
1369 // Usually, any field access or invocation on an unloaded oop type
1370 // will simply fail to link, since the statically linked class is
1371 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1372 // the static class is loaded but the sharper oop type is not.
1373 // Rather than checking for this obscure case in lots of places,
1374 // we simply observe that a null check on an unloaded class
1375 // will always be followed by a nonsense operation, so we
1376 // can just issue the uncommon trap here.
1377 // Our access to the unloaded class will only be correct
1378 // after it has been loaded and initialized, which requires
1379 // a trip through the interpreter.
1380 ciKlass* klass = tp->unloaded_klass();
1381 #ifndef PRODUCT
1382 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1383 #endif
1384 uncommon_trap(Deoptimization::Reason_unloaded,
1385 Deoptimization::Action_reinterpret,
1386 klass, "!loaded");
1387 return top();
1388 }
1389
1390 if (assert_null) {
1391 // See if the type is contained in NULL_PTR.
1392 // If so, then the value is already null.
1393 if (t->higher_equal(TypePtr::NULL_PTR)) {
1394 NOT_PRODUCT(explicit_null_checks_elided++);
1395 return value; // Elided null assert quickly!
1396 }
1397 } else {
1398 // See if mixing in the null pointer changes type.
1399 // If so, then the null pointer was not allowed in the original
1400 // type. In other words, "value" was not-null.
1401 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1402 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1403 NOT_PRODUCT(explicit_null_checks_elided++);
1404 return value; // Elided null check quickly!
1405 }
1406 }
1407 chk = new CmpPNode( value, null() );
1408 break;
1409 }
1410
1411 default:
1412 fatal("unexpected type: %s", type2name(type));
1413 }
1414 assert(chk != nullptr, "sanity check");
1415 chk = _gvn.transform(chk);
1416
1417 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1418 BoolNode *btst = new BoolNode( chk, btest);
1419 Node *tst = _gvn.transform( btst );
1420
1421 //-----------
1422 // if peephole optimizations occurred, a prior test existed.
1423 // If a prior test existed, maybe it dominates as we can avoid this test.
1424 if (tst != btst && type == T_OBJECT) {
1425 // At this point we want to scan up the CFG to see if we can
1426 // find an identical test (and so avoid this test altogether).
1427 Node *cfg = control();
1428 int depth = 0;
1429 while( depth < 16 ) { // Limit search depth for speed
1430 if( cfg->Opcode() == Op_IfTrue &&
1431 cfg->in(0)->in(1) == tst ) {
1432 // Found prior test. Use "cast_not_null" to construct an identical
1433 // CastPP (and hence hash to) as already exists for the prior test.
1434 // Return that casted value.
1435 if (assert_null) {
1436 replace_in_map(value, null());
1437 return null(); // do not issue the redundant test
1438 }
1439 Node *oldcontrol = control();
1440 set_control(cfg);
1441 Node *res = cast_not_null(value);
1442 set_control(oldcontrol);
1443 NOT_PRODUCT(explicit_null_checks_elided++);
1444 return res;
1445 }
1446 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1447 if (cfg == nullptr) break; // Quit at region nodes
1448 depth++;
1449 }
1450 }
1451
1452 //-----------
1453 // Branch to failure if null
1454 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1455 Deoptimization::DeoptReason reason;
1456 if (assert_null) {
1457 reason = Deoptimization::reason_null_assert(speculative);
1458 } else if (type == T_OBJECT || null_marker_check) {
1459 reason = Deoptimization::reason_null_check(speculative);
1460 } else {
1461 reason = Deoptimization::Reason_div0_check;
1462 }
1463 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1464 // ciMethodData::has_trap_at will return a conservative -1 if any
1465 // must-be-null assertion has failed. This could cause performance
1466 // problems for a method after its first do_null_assert failure.
1467 // Consider using 'Reason_class_check' instead?
1468
1469 // To cause an implicit null check, we set the not-null probability
1470 // to the maximum (PROB_MAX). For an explicit check the probability
1471 // is set to a smaller value.
1472 if (null_control != nullptr || too_many_traps(reason)) {
1473 // probability is less likely
1474 ok_prob = PROB_LIKELY_MAG(3);
1475 } else if (!assert_null &&
1476 (ImplicitNullCheckThreshold > 0) &&
1477 method() != nullptr &&
1478 (method()->method_data()->trap_count(reason)
1479 >= (uint)ImplicitNullCheckThreshold)) {
1480 ok_prob = PROB_LIKELY_MAG(3);
1481 }
1482
1483 if (null_control != nullptr) {
1484 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1485 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1486 set_control( _gvn.transform( new IfTrueNode(iff)));
1487 #ifndef PRODUCT
1488 if (null_true == top()) {
1489 explicit_null_checks_elided++;
1490 }
1491 #endif
1492 (*null_control) = null_true;
1493 } else {
1494 BuildCutout unless(this, tst, ok_prob);
1495 // Check for optimizer eliding test at parse time
1496 if (stopped()) {
1497 // Failure not possible; do not bother making uncommon trap.
1498 NOT_PRODUCT(explicit_null_checks_elided++);
1499 } else if (assert_null) {
1500 uncommon_trap(reason,
1501 Deoptimization::Action_make_not_entrant,
1502 nullptr, "assert_null");
1503 } else {
1504 replace_in_map(value, zerocon(type));
1505 builtin_throw(reason);
1506 }
1507 }
1508
1509 // Must throw exception, fall-thru not possible?
1510 if (stopped()) {
1511 return top(); // No result
1512 }
1513
1514 if (assert_null) {
1515 // Cast obj to null on this path.
1516 replace_in_map(value, zerocon(type));
1517 return zerocon(type);
1518 }
1519
1520 // Cast obj to not-null on this path, if there is no null_control.
1521 // (If there is a null_control, a non-null value may come back to haunt us.)
1522 if (type == T_OBJECT) {
1523 Node* cast = cast_not_null(value, false);
1524 if (null_control == nullptr || (*null_control) == top())
1525 replace_in_map(value, cast);
1526 value = cast;
1527 }
1528
1529 return value;
1530 }
1531
1532 //------------------------------cast_not_null----------------------------------
1533 // Cast obj to not-null on this path
1534 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1535 if (obj->is_InlineType()) {
1536 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1537 vt->as_InlineType()->set_null_marker(_gvn);
1538 vt = _gvn.transform(vt);
1539 if (do_replace_in_map) {
1540 replace_in_map(obj, vt);
1541 }
1542 return vt;
1543 }
1544 const Type *t = _gvn.type(obj);
1545 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1546 // Object is already not-null?
1547 if( t == t_not_null ) return obj;
1548
1549 Node* cast = new CastPPNode(control(), obj,t_not_null);
1550 cast = _gvn.transform( cast );
1551
1552 // Scan for instances of 'obj' in the current JVM mapping.
1553 // These instances are known to be not-null after the test.
1554 if (do_replace_in_map)
1555 replace_in_map(obj, cast);
1556
1557 return cast; // Return casted value
1558 }
1559
1560 // Sometimes in intrinsics, we implicitly know an object is not null
1561 // (there's no actual null check) so we can cast it to not null. In
1562 // the course of optimizations, the input to the cast can become null.
1563 // In that case that data path will die and we need the control path
1564 // to become dead as well to keep the graph consistent. So we have to
1565 // add a check for null for which one branch can't be taken. It uses
1566 // an OpaqueConstantBool node that will cause the check to be removed after loop
1567 // opts so the test goes away and the compiled code doesn't execute a
1568 // useless check.
1569 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1570 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1571 return value;
1572 }
1573 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1574 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1575 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1576 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1577 _gvn.set_type(iff, iff->Value(&_gvn));
1578 if (!tst->is_Con()) {
1579 record_for_igvn(iff);
1580 }
1581 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1582 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1583 halt(if_f, frame, "unexpected null in intrinsic");
1584 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1585 set_control(if_t);
1586 return cast_not_null(value, do_replace_in_map);
1587 }
1588
1589
1590 //--------------------------replace_in_map-------------------------------------
1591 void GraphKit::replace_in_map(Node* old, Node* neww) {
1592 if (old == neww) {
1593 return;
1594 }
1595
1596 map()->replace_edge(old, neww);
1597
1598 // Note: This operation potentially replaces any edge
1599 // on the map. This includes locals, stack, and monitors
1600 // of the current (innermost) JVM state.
1601
1602 // don't let inconsistent types from profiling escape this
1603 // method
1604
1605 const Type* told = _gvn.type(old);
1606 const Type* tnew = _gvn.type(neww);
1607
1608 if (!tnew->higher_equal(told)) {
1609 return;
1610 }
1611
1612 map()->record_replaced_node(old, neww);
1613 }
1614
1615
1616 //=============================================================================
1617 //--------------------------------memory---------------------------------------
1618 Node* GraphKit::memory(uint alias_idx) {
1619 MergeMemNode* mem = merged_memory();
1620 Node* p = mem->memory_at(alias_idx);
1621 assert(p != mem->empty_memory(), "empty");
1622 _gvn.set_type(p, Type::MEMORY); // must be mapped
1623 return p;
1624 }
1625
1626 //-----------------------------reset_memory------------------------------------
1627 Node* GraphKit::reset_memory() {
1628 Node* mem = map()->memory();
1629 // do not use this node for any more parsing!
1630 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1631 return _gvn.transform( mem );
1632 }
1633
1634 //------------------------------set_all_memory---------------------------------
1635 void GraphKit::set_all_memory(Node* newmem) {
1636 Node* mergemem = MergeMemNode::make(newmem);
1637 gvn().set_type_bottom(mergemem);
1638 if (_gvn.is_IterGVN() != nullptr) {
1639 record_for_igvn(mergemem);
1640 }
1641 map()->set_memory(mergemem);
1642 }
1643
1644 //------------------------------set_all_memory_call----------------------------
1645 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1646 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1647 set_all_memory(newmem);
1648 }
1649
1650 //=============================================================================
1651 //
1652 // parser factory methods for MemNodes
1653 //
1654 // These are layered on top of the factory methods in LoadNode and StoreNode,
1655 // and integrate with the parser's memory state and _gvn engine.
1656 //
1657
1658 // factory methods in "int adr_idx"
1659 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1660 MemNode::MemOrd mo,
1661 LoadNode::ControlDependency control_dependency,
1662 bool require_atomic_access,
1663 bool unaligned,
1664 bool mismatched,
1665 bool unsafe,
1666 uint8_t barrier_data) {
1667 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1668 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1669 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1670 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1671 Node* mem = memory(adr_idx);
1672 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1673 ld = _gvn.transform(ld);
1674
1675 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1676 // Improve graph before escape analysis and boxing elimination.
1677 record_for_igvn(ld);
1678 if (ld->is_DecodeN()) {
1679 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1680 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1681 // a Phi). Recording such cases is still perfectly sound, but may be
1682 // unnecessary and result in some minor IGVN overhead.
1683 record_for_igvn(ld->in(1));
1684 }
1685 }
1686 return ld;
1687 }
1688
1689 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1690 MemNode::MemOrd mo,
1691 bool require_atomic_access,
1692 bool unaligned,
1693 bool mismatched,
1694 bool unsafe,
1695 int barrier_data) {
1696 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1697 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1698 const TypePtr* adr_type = nullptr;
1699 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1700 Node *mem = memory(adr_idx);
1701 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1702 if (unaligned) {
1703 st->as_Store()->set_unaligned_access();
1704 }
1705 if (mismatched) {
1706 st->as_Store()->set_mismatched_access();
1707 }
1708 if (unsafe) {
1709 st->as_Store()->set_unsafe_access();
1710 }
1711 st->as_Store()->set_barrier_data(barrier_data);
1712 st = _gvn.transform(st);
1713 set_memory(st, adr_idx);
1714 // Back-to-back stores can only remove intermediate store with DU info
1715 // so push on worklist for optimizer.
1716 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1717 record_for_igvn(st);
1718
1719 return st;
1720 }
1721
1722 Node* GraphKit::access_store_at(Node* obj,
1723 Node* adr,
1724 const TypePtr* adr_type,
1725 Node* val,
1726 const Type* val_type,
1727 BasicType bt,
1728 DecoratorSet decorators,
1729 bool safe_for_replace,
1730 const InlineTypeNode* vt) {
1731 // Transformation of a value which could be null pointer (CastPP #null)
1732 // could be delayed during Parse (for example, in adjust_map_after_if()).
1733 // Execute transformation here to avoid barrier generation in such case.
1734 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1735 val = _gvn.makecon(TypePtr::NULL_PTR);
1736 }
1737
1738 if (stopped()) {
1739 return top(); // Dead path ?
1740 }
1741
1742 assert(val != nullptr, "not dead path");
1743 if (val->is_InlineType()) {
1744 // Store to non-flat field. Buffer the inline type and make sure
1745 // the store is re-executed if the allocation triggers deoptimization.
1746 PreserveReexecuteState preexecs(this);
1747 jvms()->set_should_reexecute(true);
1748 val = val->as_InlineType()->buffer(this, safe_for_replace);
1749 }
1750
1751 C2AccessValuePtr addr(adr, adr_type);
1752 C2AccessValue value(val, val_type);
1753 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1754 if (access.is_raw()) {
1755 return _barrier_set->BarrierSetC2::store_at(access, value);
1756 } else {
1757 return _barrier_set->store_at(access, value);
1758 }
1759 }
1760
1761 Node* GraphKit::access_load_at(Node* obj, // containing obj
1762 Node* adr, // actual address to store val at
1763 const TypePtr* adr_type,
1764 const Type* val_type,
1765 BasicType bt,
1766 DecoratorSet decorators,
1767 Node* ctl) {
1768 if (stopped()) {
1769 return top(); // Dead path ?
1770 }
1771
1772 C2AccessValuePtr addr(adr, adr_type);
1773 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1774 if (access.is_raw()) {
1775 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1776 } else {
1777 return _barrier_set->load_at(access, val_type);
1778 }
1779 }
1780
1781 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1782 const Type* val_type,
1783 BasicType bt,
1784 DecoratorSet decorators) {
1785 if (stopped()) {
1786 return top(); // Dead path ?
1787 }
1788
1789 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1790 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1791 if (access.is_raw()) {
1792 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1793 } else {
1794 return _barrier_set->load_at(access, val_type);
1795 }
1796 }
1797
1798 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1799 Node* adr,
1800 const TypePtr* adr_type,
1801 int alias_idx,
1802 Node* expected_val,
1803 Node* new_val,
1804 const Type* value_type,
1805 BasicType bt,
1806 DecoratorSet decorators) {
1807 C2AccessValuePtr addr(adr, adr_type);
1808 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1809 bt, obj, addr, alias_idx);
1810 if (access.is_raw()) {
1811 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1812 } else {
1813 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1814 }
1815 }
1816
1817 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1818 Node* adr,
1819 const TypePtr* adr_type,
1820 int alias_idx,
1821 Node* expected_val,
1822 Node* new_val,
1823 const Type* value_type,
1824 BasicType bt,
1825 DecoratorSet decorators) {
1826 C2AccessValuePtr addr(adr, adr_type);
1827 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1828 bt, obj, addr, alias_idx);
1829 if (access.is_raw()) {
1830 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1831 } else {
1832 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1833 }
1834 }
1835
1836 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1837 Node* adr,
1838 const TypePtr* adr_type,
1839 int alias_idx,
1840 Node* new_val,
1841 const Type* value_type,
1842 BasicType bt,
1843 DecoratorSet decorators) {
1844 C2AccessValuePtr addr(adr, adr_type);
1845 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1846 bt, obj, addr, alias_idx);
1847 if (access.is_raw()) {
1848 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1849 } else {
1850 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1851 }
1852 }
1853
1854 Node* GraphKit::access_atomic_add_at(Node* obj,
1855 Node* adr,
1856 const TypePtr* adr_type,
1857 int alias_idx,
1858 Node* new_val,
1859 const Type* value_type,
1860 BasicType bt,
1861 DecoratorSet decorators) {
1862 C2AccessValuePtr addr(adr, adr_type);
1863 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1864 if (access.is_raw()) {
1865 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1866 } else {
1867 return _barrier_set->atomic_add_at(access, new_val, value_type);
1868 }
1869 }
1870
1871 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1872 return _barrier_set->clone(this, src, dst, size, is_array);
1873 }
1874
1875 //-------------------------array_element_address-------------------------
1876 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1877 const TypeInt* sizetype, Node* ctrl) {
1878 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1879 uint shift;
1880 uint header;
1881 if (arytype->is_flat() && arytype->klass_is_exact()) {
1882 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1883 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1884 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1885 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1886 // though we don't need the address node in this case and throw it away again.
1887 shift = arytype->flat_log_elem_size();
1888 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1889 } else {
1890 shift = exact_log2(type2aelembytes(elembt));
1891 header = arrayOopDesc::base_offset_in_bytes(elembt);
1892 }
1893
1894 // short-circuit a common case (saves lots of confusing waste motion)
1895 jint idx_con = find_int_con(idx, -1);
1896 if (idx_con >= 0) {
1897 intptr_t offset = header + ((intptr_t)idx_con << shift);
1898 return basic_plus_adr(ary, offset);
1899 }
1900
1901 // must be correct type for alignment purposes
1902 Node* base = basic_plus_adr(ary, header);
1903 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1904 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1905 return basic_plus_adr(ary, base, scale);
1906 }
1907
1908 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1909 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1910 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1911 return cast_to_flat_array_exact(array, elem_vk, true, false);
1912 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1913 return cast_to_flat_array_exact(array, elem_vk, true, true);
1914 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1915 return cast_to_flat_array_exact(array, elem_vk, false, true);
1916 }
1917
1918 bool is_null_free = false;
1919 if (!elem_vk->has_nullable_atomic_layout()) {
1920 // Element does not have a nullable flat layout, cannot be nullable
1921 is_null_free = true;
1922 }
1923
1924 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1925 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1926 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1927 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1928 }
1929
1930 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1931 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1932 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1933 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1934 assert(arytype->klass_is_exact(), "inconsistency");
1935 assert(arytype->is_flat(), "inconsistency");
1936 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1937 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1938 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1939 }
1940
1941 //-------------------------load_array_element-------------------------
1942 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1943 const Type* elemtype = arytype->elem();
1944 BasicType elembt = elemtype->array_element_basic_type();
1945 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1946 if (elembt == T_NARROWOOP) {
1947 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1948 }
1949 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1950 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1951 return ld;
1952 }
1953
1954 //-------------------------set_arguments_for_java_call-------------------------
1955 // Arguments (pre-popped from the stack) are taken from the JVMS.
1956 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1957 PreserveReexecuteState preexecs(this);
1958 if (Arguments::is_valhalla_enabled()) {
1959 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1960 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1961 jvms()->set_should_reexecute(true);
1962 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1963 inc_sp(arg_size);
1964 }
1965 // Add the call arguments
1966 const TypeTuple* domain = call->tf()->domain_sig();
1967 uint nargs = domain->cnt();
1968 int arg_num = 0;
1969 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1970 uint arg_idx = i - TypeFunc::Parms;
1971 Node* arg = argument(arg_idx);
1972 const Type* t = domain->field_at(i);
1973 // TODO 8284443 A static call to a mismatched method should still be scalarized
1974 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1975 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1976 if (!arg->is_InlineType()) {
1977 // There are 2 cases in which the argument has not been scalarized
1978 if (_gvn.type(arg)->is_zero_type()) {
1979 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
1980 } else {
1981 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
1982 // receiver is a non-scalarized oop. Later on, IGVN reveals that the receiver must be a
1983 // value object. The method is devirtualized, and replaced with a direct call with a
1984 // scalarized receiver instead.
1985 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
1986 assert(C->inlining_incrementally() || C->strength_reduction(), "must be during devirtualization of calls");
1987 assert(!is_Parse(), "must be during devirtualization of calls");
1988 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1989 }
1990 }
1991 InlineTypeNode* vt = arg->as_InlineType();
1992 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1993 // If an inline type argument is passed as fields, attach the Method* to the call site
1994 // to be able to access the extended signature later via attached_method_before_pc().
1995 // For example, see CompiledMethod::preserve_callee_argument_oops().
1996 call->set_override_symbolic_info(true);
1997 // Register an calling convention dependency on the callee method to make sure that this method is deoptimized and
1998 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1999 C->dependencies()->assert_mismatch_calling_convention(call->method());
2000 arg_num++;
2001 continue;
2002 } else if (arg->is_InlineType()) {
2003 // Pass inline type argument via oop to callee
2004 arg = arg->as_InlineType()->buffer(this, true);
2005 }
2006 if (t != Type::HALF) {
2007 arg_num++;
2008 }
2009 call->init_req(idx++, arg);
2010 }
2011 }
2012
2013 //---------------------------set_edges_for_java_call---------------------------
2014 // Connect a newly created call into the current JVMS.
2015 // A return value node (if any) is returned from set_edges_for_java_call.
2016 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2017
2018 // Add the predefined inputs:
2019 call->init_req( TypeFunc::Control, control() );
2020 call->init_req( TypeFunc::I_O , i_o() );
2021 call->init_req( TypeFunc::Memory , reset_memory() );
2022 call->init_req( TypeFunc::FramePtr, frameptr() );
2023 call->init_req( TypeFunc::ReturnAdr, top() );
2024
2025 add_safepoint_edges(call, must_throw);
2026
2027 Node* xcall = _gvn.transform(call);
2028
2029 if (xcall == top()) {
2030 set_control(top());
2031 return;
2032 }
2033 assert(xcall == call, "call identity is stable");
2034
2035 // Re-use the current map to produce the result.
2036
2037 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2038 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2039 set_all_memory_call(xcall, separate_io_proj);
2040
2041 //return xcall; // no need, caller already has it
2042 }
2043
2044 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2045 if (stopped()) return top(); // maybe the call folded up?
2046
2047 // Note: Since any out-of-line call can produce an exception,
2048 // we always insert an I_O projection from the call into the result.
2049
2050 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2051
2052 if (separate_io_proj) {
2053 // The caller requested separate projections be used by the fall
2054 // through and exceptional paths, so replace the projections for
2055 // the fall through path.
2056 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2057 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2058 }
2059
2060 // Capture the return value, if any.
2061 Node* ret;
2062 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2063 ret = top();
2064 } else if (call->tf()->returns_inline_type_as_fields()) {
2065 // Return of multiple values (inline type fields): we create a
2066 // InlineType node, each field is a projection from the call.
2067 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2068 uint base_input = TypeFunc::Parms;
2069 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2070 } else {
2071 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2072 ciType* t = call->method()->return_type();
2073 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2074 // The return type is unloaded but the callee might later be C2 compiled and then return
2075 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2076 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2077 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2078 IdealKit ideal(this);
2079 IdealVariable res(ideal);
2080 ideal.declarations_done();
2081 // Change return type of call to scalarized return
2082 const TypeFunc* tf = call->_tf;
2083 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2084 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2085 call->_tf = new_tf;
2086 _gvn.set_type(call, call->Value(&_gvn));
2087 _gvn.set_type(ret, ret->Value(&_gvn));
2088 // Don't add store to buffer call if we are strength reducing
2089 if (!C->strength_reduction()) {
2090 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2091 // Return value is null
2092 ideal.set(res, makecon(TypePtr::NULL_PTR));
2093 } ideal.else_(); {
2094 // Return value is non-null
2095 sync_kit(ideal);
2096
2097 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2098 OptoRuntime::store_inline_type_fields_Type(),
2099 StubRoutines::store_inline_type_fields_to_buf(),
2100 nullptr, TypePtr::BOTTOM, ret);
2101
2102 // We don't know how many values are returned. This assumes the
2103 // worst case, that all available registers are used.
2104 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2105 if (domain->field_at(i) == Type::HALF) {
2106 store_to_buf_call->init_req(i, top());
2107 continue;
2108 }
2109 Node* proj =_gvn.transform(new ProjNode(call, i));
2110 store_to_buf_call->init_req(i, proj);
2111 }
2112 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2113
2114 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2115 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2116 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2117
2118 ideal.set(res, buf);
2119 ideal.sync_kit(this);
2120 } ideal.end_if();
2121 } else {
2122 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2123 Node* proj =_gvn.transform(new ProjNode(call, i));
2124 }
2125 ideal.set(res, ret);
2126 }
2127 sync_kit(ideal);
2128 ret = _gvn.transform(ideal.value(res));
2129 }
2130 }
2131
2132 return ret;
2133 }
2134
2135 //--------------------set_predefined_input_for_runtime_call--------------------
2136 // Reading and setting the memory state is way conservative here.
2137 // The real problem is that I am not doing real Type analysis on memory,
2138 // so I cannot distinguish card mark stores from other stores. Across a GC
2139 // point the Store Barrier and the card mark memory has to agree. I cannot
2140 // have a card mark store and its barrier split across the GC point from
2141 // either above or below. Here I get that to happen by reading ALL of memory.
2142 // A better answer would be to separate out card marks from other memory.
2143 // For now, return the input memory state, so that it can be reused
2144 // after the call, if this call has restricted memory effects.
2145 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2146 // Set fixed predefined input arguments
2147 call->init_req(TypeFunc::Control, control());
2148 call->init_req(TypeFunc::I_O, top()); // does no i/o
2149 call->init_req(TypeFunc::ReturnAdr, top());
2150 if (call->is_CallLeafPure()) {
2151 call->init_req(TypeFunc::Memory, top());
2152 call->init_req(TypeFunc::FramePtr, top());
2153 return nullptr;
2154 } else {
2155 Node* memory = reset_memory();
2156 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2157 call->init_req(TypeFunc::Memory, m); // may gc ptrs
2158 call->init_req(TypeFunc::FramePtr, frameptr());
2159 return memory;
2160 }
2161 }
2162
2163 //-------------------set_predefined_output_for_runtime_call--------------------
2164 // Set control and memory (not i_o) from the call.
2165 // If keep_mem is not null, use it for the output state,
2166 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2167 // If hook_mem is null, this call produces no memory effects at all.
2168 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2169 // then only that memory slice is taken from the call.
2170 // In the last case, we must put an appropriate memory barrier before
2171 // the call, so as to create the correct anti-dependencies on loads
2172 // preceding the call.
2173 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2174 Node* keep_mem,
2175 const TypePtr* hook_mem) {
2176 // no i/o
2177 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2178 if (call->is_CallLeafPure()) {
2179 // Pure function have only control (for now) and data output, in particular
2180 // they don't touch the memory, so we don't want a memory proj that is set after.
2181 return;
2182 }
2183 if (keep_mem) {
2184 // First clone the existing memory state
2185 set_all_memory(keep_mem);
2186 if (hook_mem != nullptr) {
2187 // Make memory for the call
2188 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2189 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2190 // We also use hook_mem to extract specific effects from arraycopy stubs.
2191 set_memory(mem, hook_mem);
2192 }
2193 // ...else the call has NO memory effects.
2194
2195 // Make sure the call advertises its memory effects precisely.
2196 // This lets us build accurate anti-dependences in gcm.cpp.
2197 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2198 "call node must be constructed correctly");
2199 } else {
2200 assert(hook_mem == nullptr, "");
2201 // This is not a "slow path" call; all memory comes from the call.
2202 set_all_memory_call(call);
2203 }
2204 }
2205
2206 // Keep track of MergeMems feeding into other MergeMems
2207 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2208 if (!mem->is_MergeMem()) {
2209 return;
2210 }
2211 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2212 Node* use = i.get();
2213 if (use->is_MergeMem()) {
2214 wl.push(use);
2215 }
2216 }
2217 }
2218
2219 // Replace the call with the current state of the kit.
2220 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2221 JVMState* ejvms = nullptr;
2222 if (has_exceptions()) {
2223 ejvms = transfer_exceptions_into_jvms();
2224 }
2225
2226 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2227 ReplacedNodes replaced_nodes_exception;
2228 Node* ex_ctl = top();
2229
2230 SafePointNode* final_state = stop();
2231
2232 // Find all the needed outputs of this call
2233 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2234
2235 Unique_Node_List wl;
2236 Node* init_mem = call->in(TypeFunc::Memory);
2237 Node* final_mem = final_state->in(TypeFunc::Memory);
2238 Node* final_ctl = final_state->in(TypeFunc::Control);
2239 Node* final_io = final_state->in(TypeFunc::I_O);
2240
2241 // Replace all the old call edges with the edges from the inlining result
2242 if (callprojs->fallthrough_catchproj != nullptr) {
2243 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2244 }
2245 if (callprojs->fallthrough_memproj != nullptr) {
2246 if (final_mem->is_MergeMem()) {
2247 // Parser's exits MergeMem was not transformed but may be optimized
2248 final_mem = _gvn.transform(final_mem);
2249 }
2250 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2251 add_mergemem_users_to_worklist(wl, final_mem);
2252 }
2253 if (callprojs->fallthrough_ioproj != nullptr) {
2254 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2255 }
2256
2257 // Replace the result with the new result if it exists and is used
2258 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2259 // If the inlined code is dead, the result projections for an inline type returned as
2260 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2261 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2262 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2263 "unexpected number of results");
2264 // If we are doing strength reduction and the return type is not loaded we
2265 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2266 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2267 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2268 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2269 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2270 }
2271 } else {
2272 C->gvn_replace_by(callprojs->resproj[0], result);
2273 }
2274 }
2275
2276 if (ejvms == nullptr) {
2277 // No exception edges to simply kill off those paths
2278 if (callprojs->catchall_catchproj != nullptr) {
2279 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2280 }
2281 if (callprojs->catchall_memproj != nullptr) {
2282 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2283 }
2284 if (callprojs->catchall_ioproj != nullptr) {
2285 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2286 }
2287 // Replace the old exception object with top
2288 if (callprojs->exobj != nullptr) {
2289 C->gvn_replace_by(callprojs->exobj, C->top());
2290 }
2291 } else {
2292 GraphKit ekit(ejvms);
2293
2294 // Load my combined exception state into the kit, with all phis transformed:
2295 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2296 replaced_nodes_exception = ex_map->replaced_nodes();
2297
2298 Node* ex_oop = ekit.use_exception_state(ex_map);
2299
2300 if (callprojs->catchall_catchproj != nullptr) {
2301 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2302 ex_ctl = ekit.control();
2303 }
2304 if (callprojs->catchall_memproj != nullptr) {
2305 Node* ex_mem = ekit.reset_memory();
2306 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2307 add_mergemem_users_to_worklist(wl, ex_mem);
2308 }
2309 if (callprojs->catchall_ioproj != nullptr) {
2310 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2311 }
2312
2313 // Replace the old exception object with the newly created one
2314 if (callprojs->exobj != nullptr) {
2315 C->gvn_replace_by(callprojs->exobj, ex_oop);
2316 }
2317 }
2318
2319 // Disconnect the call from the graph
2320 call->disconnect_inputs(C);
2321 C->gvn_replace_by(call, C->top());
2322
2323 // Clean up any MergeMems that feed other MergeMems since the
2324 // optimizer doesn't like that.
2325 while (wl.size() > 0) {
2326 _gvn.transform(wl.pop());
2327 }
2328
2329 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2330 replaced_nodes.apply(C, final_ctl);
2331 }
2332 if (!ex_ctl->is_top() && do_replaced_nodes) {
2333 replaced_nodes_exception.apply(C, ex_ctl);
2334 }
2335 }
2336
2337
2338 //------------------------------increment_counter------------------------------
2339 // for statistics: increment a VM counter by 1
2340
2341 void GraphKit::increment_counter(address counter_addr) {
2342 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2343 increment_counter(adr1);
2344 }
2345
2346 void GraphKit::increment_counter(Node* counter_addr) {
2347 Node* ctrl = control();
2348 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2349 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2350 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2351 }
2352
2353 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2354 Node* halt = new HaltNode(ctrl, frameptr, reason
2355 PRODUCT_ONLY(COMMA generate_code_in_product));
2356 halt = _gvn.transform(halt);
2357 root()->add_req(halt);
2358 if (_gvn.is_IterGVN() != nullptr) {
2359 record_for_igvn(root());
2360 }
2361 }
2362
2363 //------------------------------uncommon_trap----------------------------------
2364 // Bail out to the interpreter in mid-method. Implemented by calling the
2365 // uncommon_trap blob. This helper function inserts a runtime call with the
2366 // right debug info.
2367 Node* GraphKit::uncommon_trap(int trap_request,
2368 ciKlass* klass, const char* comment,
2369 bool must_throw,
2370 bool keep_exact_action) {
2371 if (failing_internal()) {
2372 stop();
2373 }
2374 if (stopped()) return nullptr; // trap reachable?
2375
2376 // Note: If ProfileTraps is true, and if a deopt. actually
2377 // occurs here, the runtime will make sure an MDO exists. There is
2378 // no need to call method()->ensure_method_data() at this point.
2379
2380 // Set the stack pointer to the right value for reexecution:
2381 set_sp(reexecute_sp());
2382
2383 #ifdef ASSERT
2384 if (!must_throw) {
2385 // Make sure the stack has at least enough depth to execute
2386 // the current bytecode.
2387 int inputs, ignored_depth;
2388 if (compute_stack_effects(inputs, ignored_depth)) {
2389 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2390 Bytecodes::name(java_bc()), sp(), inputs);
2391 }
2392 }
2393 #endif
2394
2395 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2396 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2397
2398 switch (action) {
2399 case Deoptimization::Action_maybe_recompile:
2400 case Deoptimization::Action_reinterpret:
2401 // Temporary fix for 6529811 to allow virtual calls to be sure they
2402 // get the chance to go from mono->bi->mega
2403 if (!keep_exact_action &&
2404 Deoptimization::trap_request_index(trap_request) < 0 &&
2405 too_many_recompiles(reason)) {
2406 // This BCI is causing too many recompilations.
2407 if (C->log() != nullptr) {
2408 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2409 Deoptimization::trap_reason_name(reason),
2410 Deoptimization::trap_action_name(action));
2411 }
2412 action = Deoptimization::Action_none;
2413 trap_request = Deoptimization::make_trap_request(reason, action);
2414 } else {
2415 C->set_trap_can_recompile(true);
2416 }
2417 break;
2418 case Deoptimization::Action_make_not_entrant:
2419 C->set_trap_can_recompile(true);
2420 break;
2421 case Deoptimization::Action_none:
2422 case Deoptimization::Action_make_not_compilable:
2423 break;
2424 default:
2425 #ifdef ASSERT
2426 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2427 #endif
2428 break;
2429 }
2430
2431 if (TraceOptoParse) {
2432 char buf[100];
2433 tty->print_cr("Uncommon trap %s at bci:%d",
2434 Deoptimization::format_trap_request(buf, sizeof(buf),
2435 trap_request), bci());
2436 }
2437
2438 CompileLog* log = C->log();
2439 if (log != nullptr) {
2440 int kid = (klass == nullptr)? -1: log->identify(klass);
2441 log->begin_elem("uncommon_trap bci='%d'", bci());
2442 char buf[100];
2443 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2444 trap_request));
2445 if (kid >= 0) log->print(" klass='%d'", kid);
2446 if (comment != nullptr) log->print(" comment='%s'", comment);
2447 log->end_elem();
2448 }
2449
2450 // Make sure any guarding test views this path as very unlikely
2451 Node *i0 = control()->in(0);
2452 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2453 IfNode *iff = i0->as_If();
2454 float f = iff->_prob; // Get prob
2455 if (control()->Opcode() == Op_IfTrue) {
2456 if (f > PROB_UNLIKELY_MAG(4))
2457 iff->_prob = PROB_MIN;
2458 } else {
2459 if (f < PROB_LIKELY_MAG(4))
2460 iff->_prob = PROB_MAX;
2461 }
2462 }
2463
2464 // Clear out dead values from the debug info.
2465 kill_dead_locals();
2466
2467 // Now insert the uncommon trap subroutine call
2468 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2469 const TypePtr* no_memory_effects = nullptr;
2470 // Pass the index of the class to be loaded
2471 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2472 (must_throw ? RC_MUST_THROW : 0),
2473 OptoRuntime::uncommon_trap_Type(),
2474 call_addr, "uncommon_trap", no_memory_effects,
2475 intcon(trap_request));
2476 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2477 "must extract request correctly from the graph");
2478 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2479
2480 call->set_req(TypeFunc::ReturnAdr, returnadr());
2481 // The debug info is the only real input to this call.
2482
2483 // Halt-and-catch fire here. The above call should never return!
2484 // We only emit code for the HaltNode in debug, which is enough for
2485 // verifying correctness. In product, we don't want to emit it so
2486 // that we can save on code space. HaltNode often get folded because
2487 // the compiler can prove that the unreachable path is dead. But we
2488 // cannot generally expect that for uncommon traps, which are often
2489 // reachable and occasionally taken.
2490 halt(control(), frameptr(),
2491 "uncommon trap returned which should never happen",
2492 false /* don't emit code in product */);
2493 stop_and_kill_map();
2494 return call;
2495 }
2496
2497
2498 //--------------------------just_allocated_object------------------------------
2499 // Report the object that was just allocated.
2500 // It must be the case that there are no intervening safepoints.
2501 // We use this to determine if an object is so "fresh" that
2502 // it does not require card marks.
2503 Node* GraphKit::just_allocated_object(Node* current_control) {
2504 Node* ctrl = current_control;
2505 // Object::<init> is invoked after allocation, most of invoke nodes
2506 // will be reduced, but a region node is kept in parse time, we check
2507 // the pattern and skip the region node if it degraded to a copy.
2508 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2509 ctrl->as_Region()->is_copy()) {
2510 ctrl = ctrl->as_Region()->is_copy();
2511 }
2512 if (C->recent_alloc_ctl() == ctrl) {
2513 return C->recent_alloc_obj();
2514 }
2515 return nullptr;
2516 }
2517
2518
2519 /**
2520 * Record profiling data exact_kls for Node n with the type system so
2521 * that it can propagate it (speculation)
2522 *
2523 * @param n node that the type applies to
2524 * @param exact_kls type from profiling
2525 * @param maybe_null did profiling see null?
2526 *
2527 * @return node with improved type
2528 */
2529 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2530 const Type* current_type = _gvn.type(n);
2531 assert(UseTypeSpeculation, "type speculation must be on");
2532
2533 const TypePtr* speculative = current_type->speculative();
2534
2535 // Should the klass from the profile be recorded in the speculative type?
2536 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2537 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2538 const TypeOopPtr* xtype = tklass->as_instance_type();
2539 assert(xtype->klass_is_exact(), "Should be exact");
2540 // Any reason to believe n is not null (from this profiling or a previous one)?
2541 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2542 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2543 // record the new speculative type's depth
2544 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2545 speculative = speculative->with_inline_depth(jvms()->depth());
2546 } else if (current_type->would_improve_ptr(ptr_kind)) {
2547 // Profiling report that null was never seen so we can change the
2548 // speculative type to non null ptr.
2549 if (ptr_kind == ProfileAlwaysNull) {
2550 speculative = TypePtr::NULL_PTR;
2551 } else {
2552 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2553 const TypePtr* ptr = TypePtr::NOTNULL;
2554 if (speculative != nullptr) {
2555 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2556 } else {
2557 speculative = ptr;
2558 }
2559 }
2560 }
2561
2562 if (speculative != current_type->speculative()) {
2563 // Build a type with a speculative type (what we think we know
2564 // about the type but will need a guard when we use it)
2565 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2566 // We're changing the type, we need a new CheckCast node to carry
2567 // the new type. The new type depends on the control: what
2568 // profiling tells us is only valid from here as far as we can
2569 // tell.
2570 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2571 cast = _gvn.transform(cast);
2572 replace_in_map(n, cast);
2573 n = cast;
2574 }
2575
2576 return n;
2577 }
2578
2579 /**
2580 * Record profiling data from receiver profiling at an invoke with the
2581 * type system so that it can propagate it (speculation)
2582 *
2583 * @param n receiver node
2584 *
2585 * @return node with improved type
2586 */
2587 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2588 if (!UseTypeSpeculation) {
2589 return n;
2590 }
2591 ciKlass* exact_kls = profile_has_unique_klass();
2592 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2593 if ((java_bc() == Bytecodes::_checkcast ||
2594 java_bc() == Bytecodes::_instanceof ||
2595 java_bc() == Bytecodes::_aastore) &&
2596 method()->method_data()->is_mature()) {
2597 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2598 if (data != nullptr) {
2599 if (java_bc() == Bytecodes::_aastore) {
2600 ciKlass* array_type = nullptr;
2601 ciKlass* element_type = nullptr;
2602 ProfilePtrKind element_ptr = ProfileMaybeNull;
2603 bool flat_array = true;
2604 bool null_free_array = true;
2605 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2606 exact_kls = element_type;
2607 ptr_kind = element_ptr;
2608 } else {
2609 if (!data->as_BitData()->null_seen()) {
2610 ptr_kind = ProfileNeverNull;
2611 } else {
2612 if (TypeProfileCasts) {
2613 assert(data->is_ReceiverTypeData(), "bad profile data type");
2614 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2615 uint i = 0;
2616 for (; i < call->row_limit(); i++) {
2617 ciKlass* receiver = call->receiver(i);
2618 if (receiver != nullptr) {
2619 break;
2620 }
2621 }
2622 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2623 }
2624 }
2625 }
2626 }
2627 }
2628 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2629 }
2630
2631 /**
2632 * Record profiling data from argument profiling at an invoke with the
2633 * type system so that it can propagate it (speculation)
2634 *
2635 * @param dest_method target method for the call
2636 * @param bc what invoke bytecode is this?
2637 */
2638 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2639 if (!UseTypeSpeculation) {
2640 return;
2641 }
2642 const TypeFunc* tf = TypeFunc::make(dest_method);
2643 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2644 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2645 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2646 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2647 if (is_reference_type(targ->basic_type())) {
2648 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2649 ciKlass* better_type = nullptr;
2650 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2651 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2652 }
2653 i++;
2654 }
2655 }
2656 }
2657
2658 /**
2659 * Record profiling data from parameter profiling at an invoke with
2660 * the type system so that it can propagate it (speculation)
2661 */
2662 void GraphKit::record_profiled_parameters_for_speculation() {
2663 if (!UseTypeSpeculation) {
2664 return;
2665 }
2666 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2667 if (_gvn.type(local(i))->isa_oopptr()) {
2668 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2669 ciKlass* better_type = nullptr;
2670 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2671 record_profile_for_speculation(local(i), better_type, ptr_kind);
2672 }
2673 j++;
2674 }
2675 }
2676 }
2677
2678 /**
2679 * Record profiling data from return value profiling at an invoke with
2680 * the type system so that it can propagate it (speculation)
2681 */
2682 void GraphKit::record_profiled_return_for_speculation() {
2683 if (!UseTypeSpeculation) {
2684 return;
2685 }
2686 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2687 ciKlass* better_type = nullptr;
2688 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2689 // If profiling reports a single type for the return value,
2690 // feed it to the type system so it can propagate it as a
2691 // speculative type
2692 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2693 }
2694 }
2695
2696
2697 //=============================================================================
2698 // Generate a fast path/slow path idiom. Graph looks like:
2699 // [foo] indicates that 'foo' is a parameter
2700 //
2701 // [in] null
2702 // \ /
2703 // CmpP
2704 // Bool ne
2705 // If
2706 // / \
2707 // True False-<2>
2708 // / |
2709 // / cast_not_null
2710 // Load | | ^
2711 // [fast_test] | |
2712 // gvn to opt_test | |
2713 // / \ | <1>
2714 // True False |
2715 // | \\ |
2716 // [slow_call] \[fast_result]
2717 // Ctl Val \ \
2718 // | \ \
2719 // Catch <1> \ \
2720 // / \ ^ \ \
2721 // Ex No_Ex | \ \
2722 // | \ \ | \ <2> \
2723 // ... \ [slow_res] | | \ [null_result]
2724 // \ \--+--+--- | |
2725 // \ | / \ | /
2726 // --------Region Phi
2727 //
2728 //=============================================================================
2729 // Code is structured as a series of driver functions all called 'do_XXX' that
2730 // call a set of helper functions. Helper functions first, then drivers.
2731
2732 //------------------------------null_check_oop---------------------------------
2733 // Null check oop. Set null-path control into Region in slot 3.
2734 // Make a cast-not-nullness use the other not-null control. Return cast.
2735 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2736 bool never_see_null,
2737 bool safe_for_replace,
2738 bool speculative) {
2739 // Initial null check taken path
2740 (*null_control) = top();
2741 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2742
2743 // Generate uncommon_trap:
2744 if (never_see_null && (*null_control) != top()) {
2745 // If we see an unexpected null at a check-cast we record it and force a
2746 // recompile; the offending check-cast will be compiled to handle nulls.
2747 // If we see more than one offending BCI, then all checkcasts in the
2748 // method will be compiled to handle nulls.
2749 PreserveJVMState pjvms(this);
2750 set_control(*null_control);
2751 replace_in_map(value, null());
2752 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2753 uncommon_trap(reason,
2754 Deoptimization::Action_make_not_entrant);
2755 (*null_control) = top(); // null path is dead
2756 }
2757 if ((*null_control) == top() && safe_for_replace) {
2758 replace_in_map(value, cast);
2759 }
2760
2761 // Cast away null-ness on the result
2762 return cast;
2763 }
2764
2765 //------------------------------opt_iff----------------------------------------
2766 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2767 // Return slow-path control.
2768 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2769 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2770
2771 // Fast path taken; set region slot 2
2772 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2773 region->init_req(2,fast_taken); // Capture fast-control
2774
2775 // Fast path not-taken, i.e. slow path
2776 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2777 return slow_taken;
2778 }
2779
2780 //-----------------------------make_runtime_call-------------------------------
2781 Node* GraphKit::make_runtime_call(int flags,
2782 const TypeFunc* call_type, address call_addr,
2783 const char* call_name,
2784 const TypePtr* adr_type,
2785 // The following parms are all optional.
2786 // The first null ends the list.
2787 Node* parm0, Node* parm1,
2788 Node* parm2, Node* parm3,
2789 Node* parm4, Node* parm5,
2790 Node* parm6, Node* parm7) {
2791 assert(call_addr != nullptr, "must not call null targets");
2792
2793 // Slow-path call
2794 bool is_leaf = !(flags & RC_NO_LEAF);
2795 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2796 if (call_name == nullptr) {
2797 assert(!is_leaf, "must supply name for leaf");
2798 call_name = OptoRuntime::stub_name(call_addr);
2799 }
2800 CallNode* call;
2801 if (!is_leaf) {
2802 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2803 } else if (flags & RC_NO_FP) {
2804 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2805 } else if (flags & RC_VECTOR){
2806 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2807 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2808 } else if (flags & RC_PURE) {
2809 assert(adr_type == nullptr, "pure call does not touch memory");
2810 call = new CallLeafPureNode(call_type, call_addr, call_name);
2811 } else {
2812 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2813 }
2814
2815 // The following is similar to set_edges_for_java_call,
2816 // except that the memory effects of the call are restricted to AliasIdxRaw.
2817
2818 // Slow path call has no side-effects, uses few values
2819 bool wide_in = !(flags & RC_NARROW_MEM);
2820 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2821
2822 Node* prev_mem = nullptr;
2823 if (wide_in) {
2824 prev_mem = set_predefined_input_for_runtime_call(call);
2825 } else {
2826 assert(!wide_out, "narrow in => narrow out");
2827 Node* narrow_mem = memory(adr_type);
2828 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2829 }
2830
2831 // Hook each parm in order. Stop looking at the first null.
2832 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2833 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2834 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2835 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2836 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2837 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2838 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2839 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2840 /* close each nested if ===> */ } } } } } } } }
2841 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2842
2843 if (!is_leaf) {
2844 // Non-leaves can block and take safepoints:
2845 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2846 }
2847 // Non-leaves can throw exceptions:
2848 if (has_io) {
2849 call->set_req(TypeFunc::I_O, i_o());
2850 }
2851
2852 if (flags & RC_UNCOMMON) {
2853 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2854 // (An "if" probability corresponds roughly to an unconditional count.
2855 // Sort of.)
2856 call->set_cnt(PROB_UNLIKELY_MAG(4));
2857 }
2858
2859 Node* c = _gvn.transform(call);
2860 assert(c == call, "cannot disappear");
2861
2862 if (wide_out) {
2863 // Slow path call has full side-effects.
2864 set_predefined_output_for_runtime_call(call);
2865 } else {
2866 // Slow path call has few side-effects, and/or sets few values.
2867 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2868 }
2869
2870 if (has_io) {
2871 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2872 }
2873 return call;
2874
2875 }
2876
2877 // i2b
2878 Node* GraphKit::sign_extend_byte(Node* in) {
2879 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2880 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2881 }
2882
2883 // i2s
2884 Node* GraphKit::sign_extend_short(Node* in) {
2885 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2886 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2887 }
2888
2889
2890 //------------------------------merge_memory-----------------------------------
2891 // Merge memory from one path into the current memory state.
2892 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2893 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2894 Node* old_slice = mms.force_memory();
2895 Node* new_slice = mms.memory2();
2896 if (old_slice != new_slice) {
2897 PhiNode* phi;
2898 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2899 if (mms.is_empty()) {
2900 // clone base memory Phi's inputs for this memory slice
2901 assert(old_slice == mms.base_memory(), "sanity");
2902 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2903 _gvn.set_type(phi, Type::MEMORY);
2904 for (uint i = 1; i < phi->req(); i++) {
2905 phi->init_req(i, old_slice->in(i));
2906 }
2907 } else {
2908 phi = old_slice->as_Phi(); // Phi was generated already
2909 }
2910 } else {
2911 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2912 _gvn.set_type(phi, Type::MEMORY);
2913 }
2914 phi->set_req(new_path, new_slice);
2915 mms.set_memory(phi);
2916 }
2917 }
2918 }
2919
2920 //------------------------------make_slow_call_ex------------------------------
2921 // Make the exception handler hookups for the slow call
2922 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2923 if (stopped()) return;
2924
2925 // Make a catch node with just two handlers: fall-through and catch-all
2926 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2927 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2928 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2929 _gvn.set_type_bottom(norm);
2930 C->record_for_igvn(norm);
2931 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2932
2933 { PreserveJVMState pjvms(this);
2934 set_control(excp);
2935 set_i_o(i_o);
2936
2937 if (excp != top()) {
2938 if (deoptimize) {
2939 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2940 uncommon_trap(Deoptimization::Reason_unhandled,
2941 Deoptimization::Action_none);
2942 } else {
2943 // Create an exception state also.
2944 // Use an exact type if the caller has a specific exception.
2945 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2946 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2947 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2948 }
2949 }
2950 }
2951
2952 // Get the no-exception control from the CatchNode.
2953 set_control(norm);
2954 }
2955
2956 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2957 Node* cmp = nullptr;
2958 switch(bt) {
2959 case T_INT: cmp = new CmpINode(in1, in2); break;
2960 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2961 default: fatal("unexpected comparison type %s", type2name(bt));
2962 }
2963 cmp = gvn.transform(cmp);
2964 Node* bol = gvn.transform(new BoolNode(cmp, test));
2965 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2966 gvn.transform(iff);
2967 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2968 return iff;
2969 }
2970
2971 //-------------------------------gen_subtype_check-----------------------------
2972 // Generate a subtyping check. Takes as input the subtype and supertype.
2973 // Returns 2 values: sets the default control() to the true path and returns
2974 // the false path. Only reads invariant memory; sets no (visible) memory.
2975 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2976 // but that's not exposed to the optimizer. This call also doesn't take in an
2977 // Object; if you wish to check an Object you need to load the Object's class
2978 // prior to coming here.
2979 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2980 ciMethod* method, int bci) {
2981 Compile* C = gvn.C;
2982 if ((*ctrl)->is_top()) {
2983 return C->top();
2984 }
2985
2986 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2987 // For a direct pointer comparison, we need the refined array klass pointer
2988 Node* vm_superklass = superklass;
2989 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
2990 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
2991 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
2992 }
2993
2994 // Fast check for identical types, perhaps identical constants.
2995 // The types can even be identical non-constants, in cases
2996 // involving Array.newInstance, Object.clone, etc.
2997 if (subklass == superklass)
2998 return C->top(); // false path is dead; no test needed.
2999
3000 if (gvn.type(superklass)->singleton()) {
3001 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3002 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3003
3004 // In the common case of an exact superklass, try to fold up the
3005 // test before generating code. You may ask, why not just generate
3006 // the code and then let it fold up? The answer is that the generated
3007 // code will necessarily include null checks, which do not always
3008 // completely fold away. If they are also needless, then they turn
3009 // into a performance loss. Example:
3010 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3011 // Here, the type of 'fa' is often exact, so the store check
3012 // of fa[1]=x will fold up, without testing the nullness of x.
3013 //
3014 // At macro expansion, we would have already folded the SubTypeCheckNode
3015 // being expanded here because we always perform the static sub type
3016 // check in SubTypeCheckNode::sub() regardless of whether
3017 // StressReflectiveCode is set or not. We can therefore skip this
3018 // static check when StressReflectiveCode is on.
3019 switch (C->static_subtype_check(superk, subk)) {
3020 case Compile::SSC_always_false:
3021 {
3022 Node* always_fail = *ctrl;
3023 *ctrl = gvn.C->top();
3024 return always_fail;
3025 }
3026 case Compile::SSC_always_true:
3027 return C->top();
3028 case Compile::SSC_easy_test:
3029 {
3030 // Just do a direct pointer compare and be done.
3031 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3032 *ctrl = gvn.transform(new IfTrueNode(iff));
3033 return gvn.transform(new IfFalseNode(iff));
3034 }
3035 case Compile::SSC_full_test:
3036 break;
3037 default:
3038 ShouldNotReachHere();
3039 }
3040 }
3041
3042 // %%% Possible further optimization: Even if the superklass is not exact,
3043 // if the subklass is the unique subtype of the superklass, the check
3044 // will always succeed. We could leave a dependency behind to ensure this.
3045
3046 // First load the super-klass's check-offset
3047 Node *p1 = gvn.transform(new AddPNode(C->top(), superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3048 Node* m = C->immutable_memory();
3049 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3050 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3051 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3052 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
3053 bool might_be_cache = (chk_off_con == cacheoff_con);
3054
3055 // Load from the sub-klass's super-class display list, or a 1-word cache of
3056 // the secondary superclass list, or a failing value with a sentinel offset
3057 // if the super-klass is an interface or exceptionally deep in the Java
3058 // hierarchy and we have to scan the secondary superclass list the hard way.
3059 // Worst-case type is a little odd: null is allowed as a result (usually
3060 // klass loads can never produce a null).
3061 Node *chk_off_X = chk_off;
3062 #ifdef _LP64
3063 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
3064 #endif
3065 Node* p2 = gvn.transform(new AddPNode(C->top(), subklass, chk_off_X));
3066 // For some types like interfaces the following loadKlass is from a 1-word
3067 // cache which is mutable so can't use immutable memory. Other
3068 // types load from the super-class display table which is immutable.
3069 Node *kmem = C->immutable_memory();
3070 // secondary_super_cache is not immutable but can be treated as such because:
3071 // - no ideal node writes to it in a way that could cause an
3072 // incorrect/missed optimization of the following Load.
3073 // - it's a cache so, worse case, not reading the latest value
3074 // wouldn't cause incorrect execution
3075 if (might_be_cache && mem != nullptr) {
3076 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
3077 }
3078 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3079
3080 // Compile speed common case: ARE a subtype and we canNOT fail
3081 if (superklass == nkls) {
3082 return C->top(); // false path is dead; no test needed.
3083 }
3084
3085 // Gather the various success & failures here
3086 RegionNode* r_not_subtype = new RegionNode(3);
3087 gvn.record_for_igvn(r_not_subtype);
3088 RegionNode* r_ok_subtype = new RegionNode(4);
3089 gvn.record_for_igvn(r_ok_subtype);
3090
3091 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3092 // SubTypeCheck node
3093 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3094 ciCallProfile profile = method->call_profile_at_bci(bci);
3095 float total_prob = 0;
3096 for (int i = 0; profile.has_receiver(i); ++i) {
3097 float prob = profile.receiver_prob(i);
3098 total_prob += prob;
3099 }
3100 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3101 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3102 for (int i = 0; profile.has_receiver(i); ++i) {
3103 ciKlass* klass = profile.receiver(i);
3104 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3105 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3106 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3107 continue;
3108 }
3109 if (klass_t->isa_aryklassptr()) {
3110 // For a direct pointer comparison, we need the refined array klass pointer
3111 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3112 }
3113 float prob = profile.receiver_prob(i);
3114 ConNode* klass_node = gvn.makecon(klass_t);
3115 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3116 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3117
3118 if (result == Compile::SSC_always_true) {
3119 r_ok_subtype->add_req(iftrue);
3120 } else {
3121 assert(result == Compile::SSC_always_false, "");
3122 r_not_subtype->add_req(iftrue);
3123 }
3124 *ctrl = gvn.transform(new IfFalseNode(iff));
3125 }
3126 }
3127 }
3128
3129 // See if we get an immediate positive hit. Happens roughly 83% of the
3130 // time. Test to see if the value loaded just previously from the subklass
3131 // is exactly the superklass.
3132 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3133 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3134 *ctrl = gvn.transform(new IfFalseNode(iff1));
3135
3136 // Compile speed common case: Check for being deterministic right now. If
3137 // chk_off is a constant and not equal to cacheoff then we are NOT a
3138 // subklass. In this case we need exactly the 1 test above and we can
3139 // return those results immediately.
3140 if (!might_be_cache) {
3141 Node* not_subtype_ctrl = *ctrl;
3142 *ctrl = iftrue1; // We need exactly the 1 test above
3143 PhaseIterGVN* igvn = gvn.is_IterGVN();
3144 if (igvn != nullptr) {
3145 igvn->remove_globally_dead_node(r_ok_subtype);
3146 igvn->remove_globally_dead_node(r_not_subtype);
3147 }
3148 return not_subtype_ctrl;
3149 }
3150
3151 r_ok_subtype->init_req(1, iftrue1);
3152
3153 // Check for immediate negative hit. Happens roughly 11% of the time (which
3154 // is roughly 63% of the remaining cases). Test to see if the loaded
3155 // check-offset points into the subklass display list or the 1-element
3156 // cache. If it points to the display (and NOT the cache) and the display
3157 // missed then it's not a subtype.
3158 Node *cacheoff = gvn.intcon(cacheoff_con);
3159 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3160 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3161 *ctrl = gvn.transform(new IfFalseNode(iff2));
3162
3163 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3164 // No performance impact (too rare) but allows sharing of secondary arrays
3165 // which has some footprint reduction.
3166 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3167 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3168 *ctrl = gvn.transform(new IfFalseNode(iff3));
3169
3170 // -- Roads not taken here: --
3171 // We could also have chosen to perform the self-check at the beginning
3172 // of this code sequence, as the assembler does. This would not pay off
3173 // the same way, since the optimizer, unlike the assembler, can perform
3174 // static type analysis to fold away many successful self-checks.
3175 // Non-foldable self checks work better here in second position, because
3176 // the initial primary superclass check subsumes a self-check for most
3177 // types. An exception would be a secondary type like array-of-interface,
3178 // which does not appear in its own primary supertype display.
3179 // Finally, we could have chosen to move the self-check into the
3180 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3181 // dependent manner. But it is worthwhile to have the check here,
3182 // where it can be perhaps be optimized. The cost in code space is
3183 // small (register compare, branch).
3184
3185 // Now do a linear scan of the secondary super-klass array. Again, no real
3186 // performance impact (too rare) but it's gotta be done.
3187 // Since the code is rarely used, there is no penalty for moving it
3188 // out of line, and it can only improve I-cache density.
3189 // The decision to inline or out-of-line this final check is platform
3190 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3191 Node* psc = gvn.transform(
3192 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3193
3194 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3195 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3196 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3197
3198 // Return false path; set default control to true path.
3199 *ctrl = gvn.transform(r_ok_subtype);
3200 return gvn.transform(r_not_subtype);
3201 }
3202
3203 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3204 const Type* sub_t = _gvn.type(obj_or_subklass);
3205 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3206 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3207 obj_or_subklass = makecon(sub_t);
3208 }
3209 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3210 if (expand_subtype_check) {
3211 MergeMemNode* mem = merged_memory();
3212 Node* ctrl = control();
3213 Node* subklass = obj_or_subklass;
3214 if (!sub_t->isa_klassptr()) {
3215 subklass = load_object_klass(obj_or_subklass);
3216 }
3217
3218 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3219 set_control(ctrl);
3220 return n;
3221 }
3222
3223 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3224 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3225 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3226 set_control(_gvn.transform(new IfTrueNode(iff)));
3227 return _gvn.transform(new IfFalseNode(iff));
3228 }
3229
3230 // Profile-driven exact type check:
3231 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3232 float prob, Node* *casted_receiver) {
3233 assert(!klass->is_interface(), "no exact type check on interfaces");
3234 Node* fail = top();
3235 const Type* rec_t = _gvn.type(receiver);
3236 if (rec_t->is_inlinetypeptr()) {
3237 if (klass->equals(rec_t->inline_klass())) {
3238 (*casted_receiver) = receiver; // Always passes
3239 } else {
3240 (*casted_receiver) = top(); // Always fails
3241 fail = control();
3242 set_control(top());
3243 }
3244 return fail;
3245 }
3246 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3247 if (tklass->isa_aryklassptr()) {
3248 // For a direct pointer comparison, we need the refined array klass pointer
3249 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3250 }
3251 Node* recv_klass = load_object_klass(receiver);
3252 fail = type_check(recv_klass, tklass, prob);
3253
3254 if (!stopped()) {
3255 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3256 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3257 assert(recv_xtype->klass_is_exact(), "");
3258
3259 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3260 // Subsume downstream occurrences of receiver with a cast to
3261 // recv_xtype, since now we know what the type will be.
3262 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3263 Node* res = _gvn.transform(cast);
3264 if (recv_xtype->is_inlinetypeptr()) {
3265 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3266 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3267 }
3268 (*casted_receiver) = res;
3269 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3270 // (User must make the replace_in_map call.)
3271 }
3272 }
3273
3274 return fail;
3275 }
3276
3277 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3278 float prob) {
3279 Node* want_klass = makecon(tklass);
3280 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3281 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3282 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3283 set_control(_gvn.transform(new IfTrueNode (iff)));
3284 Node* fail = _gvn.transform(new IfFalseNode(iff));
3285 return fail;
3286 }
3287
3288 //------------------------------subtype_check_receiver-------------------------
3289 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3290 Node** casted_receiver) {
3291 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3292 Node* want_klass = makecon(tklass);
3293
3294 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3295
3296 // Ignore interface type information until interface types are properly tracked.
3297 if (!stopped() && !klass->is_interface()) {
3298 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3299 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3300 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3301 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3302 if (recv_type->is_inlinetypeptr()) {
3303 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3304 }
3305 (*casted_receiver) = cast;
3306 }
3307 }
3308
3309 return slow_ctl;
3310 }
3311
3312 //------------------------------seems_never_null-------------------------------
3313 // Use null_seen information if it is available from the profile.
3314 // If we see an unexpected null at a type check we record it and force a
3315 // recompile; the offending check will be recompiled to handle nulls.
3316 // If we see several offending BCIs, then all checks in the
3317 // method will be recompiled.
3318 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3319 speculating = !_gvn.type(obj)->speculative_maybe_null();
3320 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3321 if (UncommonNullCast // Cutout for this technique
3322 && obj != null() // And not the -Xcomp stupid case?
3323 && !too_many_traps(reason)
3324 ) {
3325 if (speculating) {
3326 return true;
3327 }
3328 if (data == nullptr)
3329 // Edge case: no mature data. Be optimistic here.
3330 return true;
3331 // If the profile has not seen a null, assume it won't happen.
3332 assert(java_bc() == Bytecodes::_checkcast ||
3333 java_bc() == Bytecodes::_instanceof ||
3334 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3335 return !data->as_BitData()->null_seen();
3336 }
3337 speculating = false;
3338 return false;
3339 }
3340
3341 void GraphKit::guard_klass_being_initialized(Node* klass) {
3342 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3343 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3344 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3345 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3346 T_BYTE, MemNode::acquire);
3347 init_state = _gvn.transform(init_state);
3348
3349 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3350
3351 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3352 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3353
3354 { BuildCutout unless(this, tst, PROB_MAX);
3355 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3356 }
3357 }
3358
3359 void GraphKit::guard_init_thread(Node* klass) {
3360 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3361 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3362
3363 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3364 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3365 T_ADDRESS, MemNode::unordered);
3366 init_thread = _gvn.transform(init_thread);
3367
3368 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3369
3370 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3371 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3372
3373 { BuildCutout unless(this, tst, PROB_MAX);
3374 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3375 }
3376 }
3377
3378 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3379 if (ik->is_being_initialized()) {
3380 if (C->needs_clinit_barrier(ik, context)) {
3381 Node* klass = makecon(TypeKlassPtr::make(ik));
3382 guard_klass_being_initialized(klass);
3383 guard_init_thread(klass);
3384 insert_mem_bar(Op_MemBarCPUOrder);
3385 }
3386 } else if (ik->is_initialized()) {
3387 return; // no barrier needed
3388 } else {
3389 uncommon_trap(Deoptimization::Reason_uninitialized,
3390 Deoptimization::Action_reinterpret,
3391 nullptr);
3392 }
3393 }
3394
3395 //------------------------maybe_cast_profiled_receiver-------------------------
3396 // If the profile has seen exactly one type, narrow to exactly that type.
3397 // Subsequent type checks will always fold up.
3398 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3399 const TypeKlassPtr* require_klass,
3400 ciKlass* spec_klass,
3401 bool safe_for_replace) {
3402 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3403
3404 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3405
3406 // Make sure we haven't already deoptimized from this tactic.
3407 if (too_many_traps_or_recompiles(reason))
3408 return nullptr;
3409
3410 // (No, this isn't a call, but it's enough like a virtual call
3411 // to use the same ciMethod accessor to get the profile info...)
3412 // If we have a speculative type use it instead of profiling (which
3413 // may not help us)
3414 ciKlass* exact_kls = spec_klass;
3415 if (exact_kls == nullptr) {
3416 if (java_bc() == Bytecodes::_aastore) {
3417 ciKlass* array_type = nullptr;
3418 ciKlass* element_type = nullptr;
3419 ProfilePtrKind element_ptr = ProfileMaybeNull;
3420 bool flat_array = true;
3421 bool null_free_array = true;
3422 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3423 exact_kls = element_type;
3424 } else {
3425 exact_kls = profile_has_unique_klass();
3426 }
3427 }
3428 if (exact_kls != nullptr) {// no cast failures here
3429 if (require_klass == nullptr ||
3430 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3431 // If we narrow the type to match what the type profile sees or
3432 // the speculative type, we can then remove the rest of the
3433 // cast.
3434 // This is a win, even if the exact_kls is very specific,
3435 // because downstream operations, such as method calls,
3436 // will often benefit from the sharper type.
3437 Node* exact_obj = not_null_obj; // will get updated in place...
3438 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3439 &exact_obj);
3440 { PreserveJVMState pjvms(this);
3441 set_control(slow_ctl);
3442 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3443 }
3444 if (safe_for_replace) {
3445 replace_in_map(not_null_obj, exact_obj);
3446 }
3447 return exact_obj;
3448 }
3449 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3450 }
3451
3452 return nullptr;
3453 }
3454
3455 /**
3456 * Cast obj to type and emit guard unless we had too many traps here
3457 * already
3458 *
3459 * @param obj node being casted
3460 * @param type type to cast the node to
3461 * @param not_null true if we know node cannot be null
3462 */
3463 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3464 ciKlass* type,
3465 bool not_null) {
3466 if (stopped()) {
3467 return obj;
3468 }
3469
3470 // type is null if profiling tells us this object is always null
3471 if (type != nullptr) {
3472 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3473 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3474
3475 if (!too_many_traps_or_recompiles(null_reason) &&
3476 !too_many_traps_or_recompiles(class_reason)) {
3477 Node* not_null_obj = nullptr;
3478 // not_null is true if we know the object is not null and
3479 // there's no need for a null check
3480 if (!not_null) {
3481 Node* null_ctl = top();
3482 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3483 assert(null_ctl->is_top(), "no null control here");
3484 } else {
3485 not_null_obj = obj;
3486 }
3487
3488 Node* exact_obj = not_null_obj;
3489 ciKlass* exact_kls = type;
3490 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3491 &exact_obj);
3492 {
3493 PreserveJVMState pjvms(this);
3494 set_control(slow_ctl);
3495 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3496 }
3497 replace_in_map(not_null_obj, exact_obj);
3498 obj = exact_obj;
3499 }
3500 } else {
3501 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3502 Node* exact_obj = null_assert(obj);
3503 replace_in_map(obj, exact_obj);
3504 obj = exact_obj;
3505 }
3506 }
3507 return obj;
3508 }
3509
3510 //-------------------------------gen_instanceof--------------------------------
3511 // Generate an instance-of idiom. Used by both the instance-of bytecode
3512 // and the reflective instance-of call.
3513 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3514 kill_dead_locals(); // Benefit all the uncommon traps
3515 assert( !stopped(), "dead parse path should be checked in callers" );
3516 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3517 "must check for not-null not-dead klass in callers");
3518
3519 // Make the merge point
3520 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3521 RegionNode* region = new RegionNode(PATH_LIMIT);
3522 Node* phi = new PhiNode(region, TypeInt::BOOL);
3523 C->set_has_split_ifs(true); // Has chance for split-if optimization
3524
3525 ciProfileData* data = nullptr;
3526 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3527 data = method()->method_data()->bci_to_data(bci());
3528 }
3529 bool speculative_not_null = false;
3530 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3531 && seems_never_null(obj, data, speculative_not_null));
3532
3533 // Null check; get casted pointer; set region slot 3
3534 Node* null_ctl = top();
3535 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3536
3537 // If not_null_obj is dead, only null-path is taken
3538 if (stopped()) { // Doing instance-of on a null?
3539 set_control(null_ctl);
3540 return intcon(0);
3541 }
3542 region->init_req(_null_path, null_ctl);
3543 phi ->init_req(_null_path, intcon(0)); // Set null path value
3544 if (null_ctl == top()) {
3545 // Do this eagerly, so that pattern matches like is_diamond_phi
3546 // will work even during parsing.
3547 assert(_null_path == PATH_LIMIT-1, "delete last");
3548 region->del_req(_null_path);
3549 phi ->del_req(_null_path);
3550 }
3551
3552 // Do we know the type check always succeed?
3553 bool known_statically = false;
3554 if (_gvn.type(superklass)->singleton()) {
3555 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3556 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3557 if (subk != nullptr && subk->is_loaded()) {
3558 int static_res = C->static_subtype_check(superk, subk);
3559 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3560 }
3561 }
3562
3563 if (!known_statically) {
3564 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3565 // We may not have profiling here or it may not help us. If we
3566 // have a speculative type use it to perform an exact cast.
3567 ciKlass* spec_obj_type = obj_type->speculative_type();
3568 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3569 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3570 if (stopped()) { // Profile disagrees with this path.
3571 set_control(null_ctl); // Null is the only remaining possibility.
3572 return intcon(0);
3573 }
3574 if (cast_obj != nullptr) {
3575 not_null_obj = cast_obj;
3576 }
3577 }
3578 }
3579
3580 // Generate the subtype check
3581 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3582
3583 // Plug in the success path to the general merge in slot 1.
3584 region->init_req(_obj_path, control());
3585 phi ->init_req(_obj_path, intcon(1));
3586
3587 // Plug in the failing path to the general merge in slot 2.
3588 region->init_req(_fail_path, not_subtype_ctrl);
3589 phi ->init_req(_fail_path, intcon(0));
3590
3591 // Return final merged results
3592 set_control( _gvn.transform(region) );
3593 record_for_igvn(region);
3594
3595 // If we know the type check always succeeds then we don't use the
3596 // profiling data at this bytecode. Don't lose it, feed it to the
3597 // type system as a speculative type.
3598 if (safe_for_replace) {
3599 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3600 replace_in_map(obj, casted_obj);
3601 }
3602
3603 return _gvn.transform(phi);
3604 }
3605
3606 //-------------------------------gen_checkcast---------------------------------
3607 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3608 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3609 // uncommon-trap paths work. Adjust stack after this call.
3610 // If failure_control is supplied and not null, it is filled in with
3611 // the control edge for the cast failure. Otherwise, an appropriate
3612 // uncommon trap or exception is thrown.
3613 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3614 kill_dead_locals(); // Benefit all the uncommon traps
3615 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3616 const Type* obj_type = _gvn.type(obj);
3617
3618 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3619 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3620 bool safe_for_replace = (failure_control == nullptr);
3621 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3622
3623 // Fast cutout: Check the case that the cast is vacuously true.
3624 // This detects the common cases where the test will short-circuit
3625 // away completely. We do this before we perform the null check,
3626 // because if the test is going to turn into zero code, we don't
3627 // want a residual null check left around. (Causes a slowdown,
3628 // for example, in some objArray manipulations, such as a[i]=a[j].)
3629 if (improved_klass_ptr_type->singleton()) {
3630 const TypeKlassPtr* kptr = nullptr;
3631 if (obj_type->isa_oop_ptr()) {
3632 kptr = obj_type->is_oopptr()->as_klass_type();
3633 } else if (obj->is_InlineType()) {
3634 ciInlineKlass* vk = obj_type->inline_klass();
3635 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3636 }
3637
3638 if (kptr != nullptr) {
3639 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3640 case Compile::SSC_always_true:
3641 // If we know the type check always succeed then we don't use
3642 // the profiling data at this bytecode. Don't lose it, feed it
3643 // to the type system as a speculative type.
3644 obj = record_profiled_receiver_for_speculation(obj);
3645 if (null_free) {
3646 assert(safe_for_replace, "must be");
3647 obj = null_check(obj);
3648 }
3649 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3650 return obj;
3651 case Compile::SSC_always_false:
3652 if (null_free) {
3653 assert(safe_for_replace, "must be");
3654 obj = null_check(obj);
3655 }
3656 // It needs a null check because a null will *pass* the cast check.
3657 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3658 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3659 Deoptimization::DeoptReason reason = is_aastore ?
3660 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3661 builtin_throw(reason);
3662 return top();
3663 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3664 return null_assert(obj);
3665 }
3666 break; // Fall through to full check
3667 default:
3668 break;
3669 }
3670 }
3671 }
3672
3673 ciProfileData* data = nullptr;
3674 if (failure_control == nullptr) { // use MDO in regular case only
3675 assert(java_bc() == Bytecodes::_aastore ||
3676 java_bc() == Bytecodes::_checkcast,
3677 "interpreter profiles type checks only for these BCs");
3678 if (method()->method_data()->is_mature()) {
3679 data = method()->method_data()->bci_to_data(bci());
3680 }
3681 }
3682
3683 // Make the merge point
3684 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3685 RegionNode* region = new RegionNode(PATH_LIMIT);
3686 Node* phi = new PhiNode(region, toop);
3687 _gvn.set_type(region, Type::CONTROL);
3688 _gvn.set_type(phi, toop);
3689
3690 C->set_has_split_ifs(true); // Has chance for split-if optimization
3691
3692 // Use null-cast information if it is available
3693 bool speculative_not_null = false;
3694 bool never_see_null = ((failure_control == nullptr) // regular case only
3695 && seems_never_null(obj, data, speculative_not_null));
3696
3697 if (obj->is_InlineType()) {
3698 // Re-execute if buffering during triggers deoptimization
3699 PreserveReexecuteState preexecs(this);
3700 jvms()->set_should_reexecute(true);
3701 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3702 }
3703
3704 // Null check; get casted pointer; set region slot 3
3705 Node* null_ctl = top();
3706 Node* not_null_obj = nullptr;
3707 if (null_free) {
3708 assert(safe_for_replace, "must be");
3709 not_null_obj = null_check(obj);
3710 } else {
3711 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3712 }
3713
3714 // If not_null_obj is dead, only null-path is taken
3715 if (stopped()) { // Doing instance-of on a null?
3716 set_control(null_ctl);
3717 if (toop->is_inlinetypeptr()) {
3718 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3719 }
3720 return null();
3721 }
3722 region->init_req(_null_path, null_ctl);
3723 phi ->init_req(_null_path, null()); // Set null path value
3724 if (null_ctl == top()) {
3725 // Do this eagerly, so that pattern matches like is_diamond_phi
3726 // will work even during parsing.
3727 assert(_null_path == PATH_LIMIT-1, "delete last");
3728 region->del_req(_null_path);
3729 phi ->del_req(_null_path);
3730 }
3731
3732 Node* cast_obj = nullptr;
3733 if (improved_klass_ptr_type->klass_is_exact()) {
3734 // The following optimization tries to statically cast the speculative type of the object
3735 // (for example obtained during profiling) to the type of the superklass and then do a
3736 // dynamic check that the type of the object is what we expect. To work correctly
3737 // for checkcast and aastore the type of superklass should be exact.
3738 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3739 // We may not have profiling here or it may not help us. If we have
3740 // a speculative type use it to perform an exact cast.
3741 ciKlass* spec_obj_type = obj_type->speculative_type();
3742 if (spec_obj_type != nullptr || data != nullptr) {
3743 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3744 if (cast_obj != nullptr) {
3745 if (failure_control != nullptr) // failure is now impossible
3746 (*failure_control) = top();
3747 // adjust the type of the phi to the exact klass:
3748 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3749 }
3750 }
3751 }
3752
3753 if (cast_obj == nullptr) {
3754 // Generate the subtype check
3755 Node* improved_superklass = superklass;
3756 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3757 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3758 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3759 // Additionally, the benefit would only be minor in non-constant cases.
3760 improved_superklass = makecon(improved_klass_ptr_type);
3761 }
3762 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3763 // Plug in success path into the merge
3764 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3765 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3766 if (failure_control == nullptr) {
3767 if (not_subtype_ctrl != top()) { // If failure is possible
3768 PreserveJVMState pjvms(this);
3769 set_control(not_subtype_ctrl);
3770 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3771 Deoptimization::DeoptReason reason = is_aastore ?
3772 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3773 builtin_throw(reason);
3774 }
3775 } else {
3776 (*failure_control) = not_subtype_ctrl;
3777 }
3778 }
3779
3780 region->init_req(_obj_path, control());
3781 phi ->init_req(_obj_path, cast_obj);
3782
3783 // A merge of null or Casted-NotNull obj
3784 Node* res = _gvn.transform(phi);
3785
3786 // Note I do NOT always 'replace_in_map(obj,result)' here.
3787 // if( tk->klass()->can_be_primary_super() )
3788 // This means that if I successfully store an Object into an array-of-String
3789 // I 'forget' that the Object is really now known to be a String. I have to
3790 // do this because we don't have true union types for interfaces - if I store
3791 // a Baz into an array-of-Interface and then tell the optimizer it's an
3792 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3793 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3794 // replace_in_map( obj, res );
3795
3796 // Return final merged results
3797 set_control( _gvn.transform(region) );
3798 record_for_igvn(region);
3799
3800 bool not_inline = !toop->can_be_inline_type();
3801 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3802 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3803 // Check if obj has been loaded from an array
3804 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3805 Node* array = nullptr;
3806 if (obj->isa_Load()) {
3807 Node* address = obj->in(MemNode::Address);
3808 if (address->isa_AddP()) {
3809 array = address->as_AddP()->in(AddPNode::Base);
3810 }
3811 } else if (obj->is_Phi()) {
3812 Node* region = obj->in(0);
3813 // TODO make this more robust (see JDK-8231346)
3814 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3815 IfNode* iff = region->in(2)->in(0)->isa_If();
3816 if (iff != nullptr) {
3817 iff->is_flat_array_check(&_gvn, &array);
3818 }
3819 }
3820 }
3821 if (array != nullptr) {
3822 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3823 if (ary_t != nullptr) {
3824 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3825 // Casting array element to a non-inline-type, mark array as not null-free.
3826 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3827 replace_in_map(array, cast);
3828 array = cast;
3829 }
3830 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3831 // Casting array element to a non-flat-in-array type, mark array as not flat.
3832 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3833 replace_in_map(array, cast);
3834 array = cast;
3835 }
3836 }
3837 }
3838 }
3839
3840 if (!stopped() && !res->is_InlineType()) {
3841 res = record_profiled_receiver_for_speculation(res);
3842 if (toop->is_inlinetypeptr() && !maybe_larval) {
3843 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3844 res = vt;
3845 if (safe_for_replace) {
3846 replace_in_map(obj, vt);
3847 replace_in_map(not_null_obj, vt);
3848 replace_in_map(res, vt);
3849 }
3850 }
3851 }
3852 return res;
3853 }
3854
3855 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3856 // Load markword
3857 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3858 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3859 if (check_lock && !UseCompactObjectHeaders) {
3860 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3861 // Check if obj is locked
3862 Node* locked_bit = MakeConX(markWord::unlocked_value);
3863 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3864 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3865 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3866 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3867 _gvn.transform(iff);
3868 Node* locked_region = new RegionNode(3);
3869 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3870
3871 // Unlocked: Use bits from mark word
3872 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3873 mark_phi->init_req(1, mark);
3874
3875 // Locked: Load prototype header from klass
3876 set_control(_gvn.transform(new IfFalseNode(iff)));
3877 // Make loads control dependent to make sure they are only executed if array is locked
3878 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3879 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3880 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3881 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));
3882
3883 locked_region->init_req(2, control());
3884 mark_phi->init_req(2, proto);
3885 set_control(_gvn.transform(locked_region));
3886 record_for_igvn(locked_region);
3887
3888 mark = mark_phi;
3889 }
3890
3891 // Now check if mark word bits are set
3892 Node* mask = MakeConX(mask_val);
3893 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3894 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3895 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3896 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3897 }
3898
3899 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3900 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3901 }
3902
3903 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3904 // We can't use immutable memory here because the mark word is mutable.
3905 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3906 // check is moved out of loops (mainly to enable loop unswitching).
3907 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3908 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3909 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3910 }
3911
3912 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3913 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3914 }
3915
3916 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3917 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3918
3919 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3920 if (!vk->has_null_free_non_atomic_layout()) {
3921 return intcon(1); // Always atomic
3922 } else if (!vk->has_null_free_atomic_layout()) {
3923 return intcon(0); // Never atomic
3924 }
3925
3926 Node* array_klass = load_object_klass(array);
3927 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3928 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
3929 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3930 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3931 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3932 }
3933
3934 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3935 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3936 RegionNode* region = new RegionNode(3);
3937 Node* null_ctl = top();
3938 null_check_oop(val, &null_ctl);
3939 if (null_ctl != top()) {
3940 PreserveJVMState pjvms(this);
3941 set_control(null_ctl);
3942 {
3943 // Deoptimize if null-free array
3944 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3945 inc_sp(nargs);
3946 uncommon_trap(Deoptimization::Reason_null_check,
3947 Deoptimization::Action_none);
3948 }
3949 region->init_req(1, control());
3950 }
3951 region->init_req(2, control());
3952 set_control(_gvn.transform(region));
3953 record_for_igvn(region);
3954 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3955 // Since we were just successfully storing null, the array can't be null free.
3956 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3957 ary_t = ary_t->cast_to_not_null_free();
3958 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3959 if (safe_for_replace) {
3960 replace_in_map(ary, cast);
3961 }
3962 ary = cast;
3963 }
3964 return ary;
3965 }
3966
3967 //------------------------------next_monitor-----------------------------------
3968 // What number should be given to the next monitor?
3969 int GraphKit::next_monitor() {
3970 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3971 int next = current + C->sync_stack_slots();
3972 // Keep the toplevel high water mark current:
3973 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3974 return current;
3975 }
3976
3977 //------------------------------insert_mem_bar---------------------------------
3978 // Memory barrier to avoid floating things around
3979 // The membar serves as a pinch point between both control and all memory slices.
3980 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3981 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3982 mb->init_req(TypeFunc::Control, control());
3983 mb->init_req(TypeFunc::Memory, reset_memory());
3984 Node* membar = _gvn.transform(mb);
3985 record_for_igvn(membar);
3986 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3987 set_all_memory_call(membar);
3988 return membar;
3989 }
3990
3991 //-------------------------insert_mem_bar_volatile----------------------------
3992 // Memory barrier to avoid floating things around
3993 // The membar serves as a pinch point between both control and memory(alias_idx).
3994 // If you want to make a pinch point on all memory slices, do not use this
3995 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3996 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3997 // When Parse::do_put_xxx updates a volatile field, it appends a series
3998 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3999 // The first membar is on the same memory slice as the field store opcode.
4000 // This forces the membar to follow the store. (Bug 6500685 broke this.)
4001 // All the other membars (for other volatile slices, including AliasIdxBot,
4002 // which stands for all unknown volatile slices) are control-dependent
4003 // on the first membar. This prevents later volatile loads or stores
4004 // from sliding up past the just-emitted store.
4005
4006 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
4007 mb->set_req(TypeFunc::Control,control());
4008 if (alias_idx == Compile::AliasIdxBot) {
4009 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
4010 } else {
4011 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
4012 mb->set_req(TypeFunc::Memory, memory(alias_idx));
4013 }
4014 Node* membar = _gvn.transform(mb);
4015 record_for_igvn(membar);
4016 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4017 if (alias_idx == Compile::AliasIdxBot) {
4018 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
4019 } else {
4020 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
4021 }
4022 return membar;
4023 }
4024
4025 //------------------------------shared_lock------------------------------------
4026 // Emit locking code.
4027 FastLockNode* GraphKit::shared_lock(Node* obj) {
4028 // bci is either a monitorenter bc or InvocationEntryBci
4029 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4030 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4031
4032 if (stopped()) // Dead monitor?
4033 return nullptr;
4034
4035 assert(dead_locals_are_killed(), "should kill locals before sync. point");
4036
4037 // Box the stack location
4038 Node* box = new BoxLockNode(next_monitor());
4039 // Check for bailout after new BoxLockNode
4040 if (failing()) { return nullptr; }
4041 box = _gvn.transform(box);
4042 Node* mem = reset_memory();
4043
4044 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
4045
4046 // Add monitor to debug info for the slow path. If we block inside the
4047 // slow path and de-opt, we need the monitor hanging around
4048 map()->push_monitor( flock );
4049
4050 const TypeFunc *tf = LockNode::lock_type();
4051 LockNode *lock = new LockNode(C, tf);
4052
4053 lock->init_req( TypeFunc::Control, control() );
4054 lock->init_req( TypeFunc::Memory , mem );
4055 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4056 lock->init_req( TypeFunc::FramePtr, frameptr() );
4057 lock->init_req( TypeFunc::ReturnAdr, top() );
4058
4059 lock->init_req(TypeFunc::Parms + 0, obj);
4060 lock->init_req(TypeFunc::Parms + 1, box);
4061 lock->init_req(TypeFunc::Parms + 2, flock);
4062 add_safepoint_edges(lock);
4063
4064 lock = _gvn.transform( lock )->as_Lock();
4065
4066 // lock has no side-effects, sets few values
4067 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
4068
4069 insert_mem_bar(Op_MemBarAcquireLock);
4070
4071 // Add this to the worklist so that the lock can be eliminated
4072 record_for_igvn(lock);
4073
4074 #ifndef PRODUCT
4075 if (PrintLockStatistics) {
4076 // Update the counter for this lock. Don't bother using an atomic
4077 // operation since we don't require absolute accuracy.
4078 lock->create_lock_counter(map()->jvms());
4079 increment_counter(lock->counter()->addr());
4080 }
4081 #endif
4082
4083 return flock;
4084 }
4085
4086
4087 //------------------------------shared_unlock----------------------------------
4088 // Emit unlocking code.
4089 void GraphKit::shared_unlock(Node* box, Node* obj) {
4090 // bci is either a monitorenter bc or InvocationEntryBci
4091 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4092 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4093
4094 if (stopped()) { // Dead monitor?
4095 map()->pop_monitor(); // Kill monitor from debug info
4096 return;
4097 }
4098 assert(!obj->is_InlineType(), "should not unlock on inline type");
4099
4100 // Memory barrier to avoid floating things down past the locked region
4101 insert_mem_bar(Op_MemBarReleaseLock);
4102
4103 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4104 UnlockNode *unlock = new UnlockNode(C, tf);
4105 #ifdef ASSERT
4106 unlock->set_dbg_jvms(sync_jvms());
4107 #endif
4108 uint raw_idx = Compile::AliasIdxRaw;
4109 unlock->init_req( TypeFunc::Control, control() );
4110 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4111 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4112 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4113 unlock->init_req( TypeFunc::ReturnAdr, top() );
4114
4115 unlock->init_req(TypeFunc::Parms + 0, obj);
4116 unlock->init_req(TypeFunc::Parms + 1, box);
4117 unlock = _gvn.transform(unlock)->as_Unlock();
4118
4119 Node* mem = reset_memory();
4120
4121 // unlock has no side-effects, sets few values
4122 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4123
4124 // Kill monitor from debug info
4125 map()->pop_monitor( );
4126 }
4127
4128 //-------------------------------get_layout_helper-----------------------------
4129 // If the given klass is a constant or known to be an array,
4130 // fetch the constant layout helper value into constant_value
4131 // and return null. Otherwise, load the non-constant
4132 // layout helper value, and return the node which represents it.
4133 // This two-faced routine is useful because allocation sites
4134 // almost always feature constant types.
4135 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4136 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4137 if (!StressReflectiveCode && klass_t != nullptr) {
4138 bool xklass = klass_t->klass_is_exact();
4139 bool can_be_flat = false;
4140 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4141 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4142 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4143 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4144 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4145 }
4146 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4147 jint lhelper;
4148 if (klass_t->is_flat()) {
4149 lhelper = ary_type->flat_layout_helper();
4150 } else if (klass_t->isa_aryklassptr()) {
4151 BasicType elem = ary_type->elem()->array_element_basic_type();
4152 if (is_reference_type(elem, true)) {
4153 elem = T_OBJECT;
4154 }
4155 lhelper = Klass::array_layout_helper(elem);
4156 } else {
4157 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4158 }
4159 if (lhelper != Klass::_lh_neutral_value) {
4160 constant_value = lhelper;
4161 return (Node*) nullptr;
4162 }
4163 }
4164 }
4165 constant_value = Klass::_lh_neutral_value; // put in a known value
4166 Node* lhp = basic_plus_adr(top(), klass_node, in_bytes(Klass::layout_helper_offset()));
4167 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4168 }
4169
4170 // We just put in an allocate/initialize with a big raw-memory effect.
4171 // Hook selected additional alias categories on the initialization.
4172 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4173 MergeMemNode* init_in_merge,
4174 Node* init_out_raw) {
4175 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4176 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4177
4178 Node* prevmem = kit.memory(alias_idx);
4179 init_in_merge->set_memory_at(alias_idx, prevmem);
4180 if (init_out_raw != nullptr) {
4181 kit.set_memory(init_out_raw, alias_idx);
4182 }
4183 }
4184
4185 //---------------------------set_output_for_allocation-------------------------
4186 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4187 const TypeOopPtr* oop_type,
4188 bool deoptimize_on_exception) {
4189 int rawidx = Compile::AliasIdxRaw;
4190 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4191 add_safepoint_edges(alloc);
4192 Node* allocx = _gvn.transform(alloc);
4193 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4194 // create memory projection for i_o
4195 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4196 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4197
4198 // create a memory projection as for the normal control path
4199 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4200 set_memory(malloc, rawidx);
4201
4202 // a normal slow-call doesn't change i_o, but an allocation does
4203 // we create a separate i_o projection for the normal control path
4204 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4205 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4206
4207 // put in an initialization barrier
4208 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4209 rawoop)->as_Initialize();
4210 assert(alloc->initialization() == init, "2-way macro link must work");
4211 assert(init ->allocation() == alloc, "2-way macro link must work");
4212 {
4213 // Extract memory strands which may participate in the new object's
4214 // initialization, and source them from the new InitializeNode.
4215 // This will allow us to observe initializations when they occur,
4216 // and link them properly (as a group) to the InitializeNode.
4217 assert(init->in(InitializeNode::Memory) == malloc, "");
4218 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4219 init->set_req(InitializeNode::Memory, minit_in);
4220 record_for_igvn(minit_in); // fold it up later, if possible
4221 _gvn.set_type(minit_in, Type::MEMORY);
4222 Node* minit_out = memory(rawidx);
4223 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4224 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4225 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4226 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4227 // multiple projections as a result.
4228 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4229 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4230 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4231 if (oop_type->isa_aryptr()) {
4232 // Initially all flat array accesses share a single slice
4233 // but that changes after parsing. Prepare the memory graph so
4234 // it can optimize flat array accesses properly once they
4235 // don't share a single slice.
4236 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4237 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4238 int elemidx = C->get_alias_index(telemref);
4239 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4240 if (alias_adr_type->is_flat()) {
4241 C->set_flat_accesses();
4242 }
4243 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4244 } else if (oop_type->isa_instptr()) {
4245 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4246 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4247 ciField* field = ik->nonstatic_field_at(i);
4248 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4249 continue; // do not bother to track really large numbers of fields
4250 // Find (or create) the alias category for this field:
4251 int fieldidx = C->alias_type(field)->index();
4252 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4253 }
4254 }
4255 }
4256
4257 // Cast raw oop to the real thing...
4258 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4259 javaoop = _gvn.transform(javaoop);
4260 C->set_recent_alloc(control(), javaoop);
4261 assert(just_allocated_object(control()) == javaoop, "just allocated");
4262
4263 #ifdef ASSERT
4264 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4265 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4266 "Ideal_allocation works");
4267 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4268 "Ideal_allocation works");
4269 if (alloc->is_AllocateArray()) {
4270 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4271 "Ideal_allocation works");
4272 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4273 "Ideal_allocation works");
4274 } else {
4275 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4276 }
4277 }
4278 #endif //ASSERT
4279
4280 return javaoop;
4281 }
4282
4283 //---------------------------new_instance--------------------------------------
4284 // This routine takes a klass_node which may be constant (for a static type)
4285 // or may be non-constant (for reflective code). It will work equally well
4286 // for either, and the graph will fold nicely if the optimizer later reduces
4287 // the type to a constant.
4288 // The optional arguments are for specialized use by intrinsics:
4289 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4290 // - If 'return_size_val', report the total object size to the caller.
4291 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4292 Node* GraphKit::new_instance(Node* klass_node,
4293 Node* extra_slow_test,
4294 Node* *return_size_val,
4295 bool deoptimize_on_exception,
4296 InlineTypeNode* inline_type_node) {
4297 // Compute size in doublewords
4298 // The size is always an integral number of doublewords, represented
4299 // as a positive bytewise size stored in the klass's layout_helper.
4300 // The layout_helper also encodes (in a low bit) the need for a slow path.
4301 jint layout_con = Klass::_lh_neutral_value;
4302 Node* layout_val = get_layout_helper(klass_node, layout_con);
4303 bool layout_is_con = (layout_val == nullptr);
4304
4305 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4306 // Generate the initial go-slow test. It's either ALWAYS (return a
4307 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4308 // case) a computed value derived from the layout_helper.
4309 Node* initial_slow_test = nullptr;
4310 if (layout_is_con) {
4311 assert(!StressReflectiveCode, "stress mode does not use these paths");
4312 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4313 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4314 } else { // reflective case
4315 // This reflective path is used by Unsafe.allocateInstance.
4316 // (It may be stress-tested by specifying StressReflectiveCode.)
4317 // Basically, we want to get into the VM is there's an illegal argument.
4318 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4319 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4320 if (extra_slow_test != intcon(0)) {
4321 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4322 }
4323 // (Macro-expander will further convert this to a Bool, if necessary.)
4324 }
4325
4326 // Find the size in bytes. This is easy; it's the layout_helper.
4327 // The size value must be valid even if the slow path is taken.
4328 Node* size = nullptr;
4329 if (layout_is_con) {
4330 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4331 } else { // reflective case
4332 // This reflective path is used by clone and Unsafe.allocateInstance.
4333 size = ConvI2X(layout_val);
4334
4335 // Clear the low bits to extract layout_helper_size_in_bytes:
4336 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4337 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4338 size = _gvn.transform( new AndXNode(size, mask) );
4339 }
4340 if (return_size_val != nullptr) {
4341 (*return_size_val) = size;
4342 }
4343
4344 // This is a precise notnull oop of the klass.
4345 // (Actually, it need not be precise if this is a reflective allocation.)
4346 // It's what we cast the result to.
4347 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4348 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4349 const TypeOopPtr* oop_type = tklass->as_instance_type();
4350
4351 // Now generate allocation code
4352
4353 // The entire memory state is needed for slow path of the allocation
4354 // since GC and deoptimization can happen.
4355 Node *mem = reset_memory();
4356 set_all_memory(mem); // Create new memory state
4357
4358 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4359 control(), mem, i_o(),
4360 size, klass_node,
4361 initial_slow_test, inline_type_node);
4362
4363 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4364 }
4365
4366 //-------------------------------new_array-------------------------------------
4367 // helper for newarray and anewarray
4368 // The 'length' parameter is (obviously) the length of the array.
4369 // The optional arguments are for specialized use by intrinsics:
4370 // - If 'return_size_val', report the non-padded array size (sum of header size
4371 // and array body) to the caller.
4372 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4373 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4374 Node* length, // number of array elements
4375 int nargs, // number of arguments to push back for uncommon trap
4376 Node* *return_size_val,
4377 bool deoptimize_on_exception,
4378 Node* init_val) {
4379 jint layout_con = Klass::_lh_neutral_value;
4380 Node* layout_val = get_layout_helper(klass_node, layout_con);
4381 bool layout_is_con = (layout_val == nullptr);
4382
4383 if (!layout_is_con && !StressReflectiveCode &&
4384 !too_many_traps(Deoptimization::Reason_class_check)) {
4385 // This is a reflective array creation site.
4386 // Optimistically assume that it is a subtype of Object[],
4387 // so that we can fold up all the address arithmetic.
4388 layout_con = Klass::array_layout_helper(T_OBJECT);
4389 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4390 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4391 { BuildCutout unless(this, bol_lh, PROB_MAX);
4392 inc_sp(nargs);
4393 uncommon_trap(Deoptimization::Reason_class_check,
4394 Deoptimization::Action_maybe_recompile);
4395 }
4396 layout_val = nullptr;
4397 layout_is_con = true;
4398 }
4399
4400 // Generate the initial go-slow test. Make sure we do not overflow
4401 // if length is huge (near 2Gig) or negative! We do not need
4402 // exact double-words here, just a close approximation of needed
4403 // double-words. We can't add any offset or rounding bits, lest we
4404 // take a size -1 of bytes and make it positive. Use an unsigned
4405 // compare, so negative sizes look hugely positive.
4406 int fast_size_limit = FastAllocateSizeLimit;
4407 if (layout_is_con) {
4408 assert(!StressReflectiveCode, "stress mode does not use these paths");
4409 // Increase the size limit if we have exact knowledge of array type.
4410 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4411 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4412 }
4413
4414 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4415 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4416
4417 // --- Size Computation ---
4418 // array_size = round_to_heap(array_header + (length << elem_shift));
4419 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4420 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4421 // The rounding mask is strength-reduced, if possible.
4422 int round_mask = MinObjAlignmentInBytes - 1;
4423 Node* header_size = nullptr;
4424 // (T_BYTE has the weakest alignment and size restrictions...)
4425 if (layout_is_con) {
4426 int hsize = Klass::layout_helper_header_size(layout_con);
4427 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4428 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4429 if ((round_mask & ~right_n_bits(eshift)) == 0)
4430 round_mask = 0; // strength-reduce it if it goes away completely
4431 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4432 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4433 assert(header_size_min <= hsize, "generic minimum is smallest");
4434 header_size = intcon(hsize);
4435 } else {
4436 Node* hss = intcon(Klass::_lh_header_size_shift);
4437 Node* hsm = intcon(Klass::_lh_header_size_mask);
4438 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4439 header_size = _gvn.transform(new AndINode(header_size, hsm));
4440 }
4441
4442 Node* elem_shift = nullptr;
4443 if (layout_is_con) {
4444 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4445 if (eshift != 0)
4446 elem_shift = intcon(eshift);
4447 } else {
4448 // There is no need to mask or shift this value.
4449 // The semantics of LShiftINode include an implicit mask to 0x1F.
4450 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4451 elem_shift = layout_val;
4452 }
4453
4454 // Transition to native address size for all offset calculations:
4455 Node* lengthx = ConvI2X(length);
4456 Node* headerx = ConvI2X(header_size);
4457 #ifdef _LP64
4458 { const TypeInt* tilen = _gvn.find_int_type(length);
4459 if (tilen != nullptr && tilen->_lo < 0) {
4460 // Add a manual constraint to a positive range. Cf. array_element_address.
4461 jint size_max = fast_size_limit;
4462 if (size_max > tilen->_hi && tilen->_hi >= 0) {
4463 size_max = tilen->_hi;
4464 }
4465 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4466
4467 // Only do a narrow I2L conversion if the range check passed.
4468 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4469 _gvn.transform(iff);
4470 RegionNode* region = new RegionNode(3);
4471 _gvn.set_type(region, Type::CONTROL);
4472 lengthx = new PhiNode(region, TypeLong::LONG);
4473 _gvn.set_type(lengthx, TypeLong::LONG);
4474
4475 // Range check passed. Use ConvI2L node with narrow type.
4476 Node* passed = IfFalse(iff);
4477 region->init_req(1, passed);
4478 // Make I2L conversion control dependent to prevent it from
4479 // floating above the range check during loop optimizations.
4480 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4481
4482 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4483 region->init_req(2, IfTrue(iff));
4484 lengthx->init_req(2, ConvI2X(length));
4485
4486 set_control(region);
4487 record_for_igvn(region);
4488 record_for_igvn(lengthx);
4489 }
4490 }
4491 #endif
4492
4493 // Combine header size and body size for the array copy part, then align (if
4494 // necessary) for the allocation part. This computation cannot overflow,
4495 // because it is used only in two places, one where the length is sharply
4496 // limited, and the other after a successful allocation.
4497 Node* abody = lengthx;
4498 if (elem_shift != nullptr) {
4499 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4500 }
4501 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4502
4503 if (return_size_val != nullptr) {
4504 // This is the size
4505 (*return_size_val) = non_rounded_size;
4506 }
4507
4508 Node* size = non_rounded_size;
4509 if (round_mask != 0) {
4510 Node* mask1 = MakeConX(round_mask);
4511 size = _gvn.transform(new AddXNode(size, mask1));
4512 Node* mask2 = MakeConX(~round_mask);
4513 size = _gvn.transform(new AndXNode(size, mask2));
4514 }
4515 // else if round_mask == 0, the size computation is self-rounding
4516
4517 // Now generate allocation code
4518
4519 // The entire memory state is needed for slow path of the allocation
4520 // since GC and deoptimization can happen.
4521 Node *mem = reset_memory();
4522 set_all_memory(mem); // Create new memory state
4523
4524 if (initial_slow_test->is_Bool()) {
4525 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4526 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4527 }
4528
4529 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4530 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4531
4532 Node* raw_init_value = nullptr;
4533 if (init_val != nullptr) {
4534 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4535 if (ary_type->is_flat()) {
4536 initial_slow_test = intcon(1);
4537 }
4538
4539 if (UseCompressedOops) {
4540 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4541 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4542 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4543 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4544 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4545 } else {
4546 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4547 }
4548 }
4549
4550 Node* valid_length_test = _gvn.intcon(1);
4551 if (ary_type->isa_aryptr()) {
4552 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4553 jint max = TypeAryPtr::max_array_length(bt);
4554 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4555 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4556 }
4557
4558 // Create the AllocateArrayNode and its result projections
4559 AllocateArrayNode* alloc
4560 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4561 control(), mem, i_o(),
4562 size, klass_node,
4563 initial_slow_test,
4564 length, valid_length_test,
4565 init_val, raw_init_value);
4566 // Cast to correct type. Note that the klass_node may be constant or not,
4567 // and in the latter case the actual array type will be inexact also.
4568 // (This happens via a non-constant argument to inline_native_newArray.)
4569 // In any case, the value of klass_node provides the desired array type.
4570 const TypeInt* length_type = _gvn.find_int_type(length);
4571 if (ary_type->isa_aryptr() && length_type != nullptr) {
4572 // Try to get a better type than POS for the size
4573 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4574 }
4575
4576 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4577
4578 array_ideal_length(alloc, ary_type, true);
4579 return javaoop;
4580 }
4581
4582 // The following "Ideal_foo" functions are placed here because they recognize
4583 // the graph shapes created by the functions immediately above.
4584
4585 //---------------------------Ideal_allocation----------------------------------
4586 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4587 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4588 if (ptr == nullptr) { // reduce dumb test in callers
4589 return nullptr;
4590 }
4591
4592 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4593 ptr = bs->step_over_gc_barrier(ptr);
4594
4595 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4596 ptr = ptr->in(1);
4597 if (ptr == nullptr) return nullptr;
4598 }
4599 // Return null for allocations with several casts:
4600 // j.l.reflect.Array.newInstance(jobject, jint)
4601 // Object.clone()
4602 // to keep more precise type from last cast.
4603 if (ptr->is_Proj()) {
4604 Node* allo = ptr->in(0);
4605 if (allo != nullptr && allo->is_Allocate()) {
4606 return allo->as_Allocate();
4607 }
4608 }
4609 // Report failure to match.
4610 return nullptr;
4611 }
4612
4613 // Fancy version which also strips off an offset (and reports it to caller).
4614 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4615 intptr_t& offset) {
4616 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4617 if (base == nullptr) return nullptr;
4618 return Ideal_allocation(base);
4619 }
4620
4621 // Trace Initialize <- Proj[Parm] <- Allocate
4622 AllocateNode* InitializeNode::allocation() {
4623 Node* rawoop = in(InitializeNode::RawAddress);
4624 if (rawoop->is_Proj()) {
4625 Node* alloc = rawoop->in(0);
4626 if (alloc->is_Allocate()) {
4627 return alloc->as_Allocate();
4628 }
4629 }
4630 return nullptr;
4631 }
4632
4633 // Trace Allocate -> Proj[Parm] -> Initialize
4634 InitializeNode* AllocateNode::initialization() {
4635 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4636 if (rawoop == nullptr) return nullptr;
4637 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4638 Node* init = rawoop->fast_out(i);
4639 if (init->is_Initialize()) {
4640 assert(init->as_Initialize()->allocation() == this, "2-way link");
4641 return init->as_Initialize();
4642 }
4643 }
4644 return nullptr;
4645 }
4646
4647 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4648 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4649 // Too many traps seen?
4650 if (too_many_traps(reason)) {
4651 #ifdef ASSERT
4652 if (TraceLoopPredicate) {
4653 int tc = C->trap_count(reason);
4654 tty->print("too many traps=%s tcount=%d in ",
4655 Deoptimization::trap_reason_name(reason), tc);
4656 method()->print(); // which method has too many predicate traps
4657 tty->cr();
4658 }
4659 #endif
4660 // We cannot afford to take more traps here,
4661 // do not generate Parse Predicate.
4662 return;
4663 }
4664
4665 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4666 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4667 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4668 {
4669 PreserveJVMState pjvms(this);
4670 set_control(if_false);
4671 inc_sp(nargs);
4672 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4673 }
4674 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4675 set_control(if_true);
4676 }
4677
4678 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4679 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4680 void GraphKit::add_parse_predicates(int nargs) {
4681 if (ShortRunningLongLoop) {
4682 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4683 // walking up from the loop.
4684 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4685 }
4686 if (UseLoopPredicate) {
4687 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4688 if (UseProfiledLoopPredicate) {
4689 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4690 }
4691 }
4692 if (UseAutoVectorizationPredicate) {
4693 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4694 }
4695 // Loop Limit Check Predicate should be near the loop.
4696 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4697 }
4698
4699 void GraphKit::sync_kit(IdealKit& ideal) {
4700 reset_memory();
4701 set_all_memory(ideal.merged_memory());
4702 set_i_o(ideal.i_o());
4703 set_control(ideal.ctrl());
4704 }
4705
4706 void GraphKit::final_sync(IdealKit& ideal) {
4707 // Final sync IdealKit and graphKit.
4708 sync_kit(ideal);
4709 }
4710
4711 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4712 Node* len = load_array_length(load_String_value(str, set_ctrl));
4713 Node* coder = load_String_coder(str, set_ctrl);
4714 // Divide length by 2 if coder is UTF16
4715 return _gvn.transform(new RShiftINode(len, coder));
4716 }
4717
4718 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4719 int value_offset = java_lang_String::value_offset();
4720 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4721 false, nullptr, Type::Offset(0));
4722 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4723 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4724 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4725 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4726 Node* p = basic_plus_adr(str, str, value_offset);
4727 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4728 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4729 return load;
4730 }
4731
4732 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4733 if (!CompactStrings) {
4734 return intcon(java_lang_String::CODER_UTF16);
4735 }
4736 int coder_offset = java_lang_String::coder_offset();
4737 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4738 false, nullptr, Type::Offset(0));
4739 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4740
4741 Node* p = basic_plus_adr(str, str, coder_offset);
4742 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4743 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4744 return load;
4745 }
4746
4747 void GraphKit::store_String_value(Node* str, Node* value) {
4748 int value_offset = java_lang_String::value_offset();
4749 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4750 false, nullptr, Type::Offset(0));
4751 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4752
4753 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4754 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4755 }
4756
4757 void GraphKit::store_String_coder(Node* str, Node* value) {
4758 int coder_offset = java_lang_String::coder_offset();
4759 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4760 false, nullptr, Type::Offset(0));
4761 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4762
4763 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4764 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4765 }
4766
4767 // Capture src and dst memory state with a MergeMemNode
4768 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4769 if (src_type == dst_type) {
4770 // Types are equal, we don't need a MergeMemNode
4771 return memory(src_type);
4772 }
4773 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4774 record_for_igvn(merge); // fold it up later, if possible
4775 int src_idx = C->get_alias_index(src_type);
4776 int dst_idx = C->get_alias_index(dst_type);
4777 merge->set_memory_at(src_idx, memory(src_idx));
4778 merge->set_memory_at(dst_idx, memory(dst_idx));
4779 return merge;
4780 }
4781
4782 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4783 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4784 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4785 // If input and output memory types differ, capture both states to preserve
4786 // the dependency between preceding and subsequent loads/stores.
4787 // For example, the following program:
4788 // StoreB
4789 // compress_string
4790 // LoadB
4791 // has this memory graph (use->def):
4792 // LoadB -> compress_string -> CharMem
4793 // ... -> StoreB -> ByteMem
4794 // The intrinsic hides the dependency between LoadB and StoreB, causing
4795 // the load to read from memory not containing the result of the StoreB.
4796 // The correct memory graph should look like this:
4797 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4798 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4799 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4800 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4801 set_memory(res_mem, TypeAryPtr::BYTES);
4802 return str;
4803 }
4804
4805 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4806 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4807 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4808 // Capture src and dst memory (see comment in 'compress_string').
4809 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4810 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4811 set_memory(_gvn.transform(str), dst_type);
4812 }
4813
4814 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4815 /**
4816 * int i_char = start;
4817 * for (int i_byte = 0; i_byte < count; i_byte++) {
4818 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4819 * }
4820 */
4821 add_parse_predicates();
4822 C->set_has_loops(true);
4823
4824 RegionNode* head = new RegionNode(3);
4825 head->init_req(1, control());
4826 gvn().set_type(head, Type::CONTROL);
4827 record_for_igvn(head);
4828
4829 Node* i_byte = new PhiNode(head, TypeInt::INT);
4830 i_byte->init_req(1, intcon(0));
4831 gvn().set_type(i_byte, TypeInt::INT);
4832 record_for_igvn(i_byte);
4833
4834 Node* i_char = new PhiNode(head, TypeInt::INT);
4835 i_char->init_req(1, start);
4836 gvn().set_type(i_char, TypeInt::INT);
4837 record_for_igvn(i_char);
4838
4839 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4840 gvn().set_type(mem, Type::MEMORY);
4841 record_for_igvn(mem);
4842 set_control(head);
4843 set_memory(mem, TypeAryPtr::BYTES);
4844 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4845 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4846 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4847 false, true /* mismatched */);
4848
4849 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4850 head->init_req(2, IfTrue(iff));
4851 mem->init_req(2, st);
4852 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4853 i_char->init_req(2, AddI(i_char, intcon(2)));
4854
4855 set_control(IfFalse(iff));
4856 set_memory(st, TypeAryPtr::BYTES);
4857 }
4858
4859 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4860 if (!field->is_constant()) {
4861 return nullptr; // Field not marked as constant.
4862 }
4863 ciInstance* holder = nullptr;
4864 if (!field->is_static()) {
4865 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4866 if (const_oop != nullptr && const_oop->is_instance()) {
4867 holder = const_oop->as_instance();
4868 }
4869 }
4870 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4871 /*is_unsigned_load=*/false);
4872 if (con_type != nullptr) {
4873 Node* con = makecon(con_type);
4874 if (field->type()->is_inlinetype()) {
4875 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4876 } else if (con_type->is_inlinetypeptr()) {
4877 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4878 }
4879 return con;
4880 }
4881 return nullptr;
4882 }
4883
4884 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
4885 const Type* obj_type = obj->bottom_type();
4886 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4887 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4888 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4889 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4890 obj = casted_obj;
4891 }
4892 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
4893 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4894 }
4895 return obj;
4896 }