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