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