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