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