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