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