1 /*
2 * Copyright (c) 2001, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/opcodes.hpp"
53 #include "opto/parse.hpp"
54 #include "opto/reachability.hpp"
55 #include "opto/rootnode.hpp"
56 #include "opto/runtime.hpp"
57 #include "opto/subtypenode.hpp"
58 #include "opto/type.hpp"
59 #include "runtime/arguments.hpp"
60 #include "runtime/deoptimization.hpp"
61 #include "runtime/sharedRuntime.hpp"
62 #include "runtime/stubRoutines.hpp"
63 #include "utilities/bitMap.inline.hpp"
64 #include "utilities/growableArray.hpp"
65 #include "utilities/powerOfTwo.hpp"
66
67 //----------------------------GraphKit-----------------------------------------
68 // Main utility constructor.
69 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
70 : Phase(Phase::Parser),
71 _env(C->env()),
72 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
73 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
74 {
75 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
76 _exceptions = jvms->map()->next_exception();
77 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
78 set_jvms(jvms);
79 #ifdef ASSERT
80 if (_gvn.is_IterGVN() != nullptr) {
81 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
82 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
83 _worklist_size = _gvn.C->igvn_worklist()->size();
84 }
85 #endif
86 }
87
88 // Private constructor for parser.
89 GraphKit::GraphKit()
90 : Phase(Phase::Parser),
91 _env(C->env()),
92 _gvn(*C->initial_gvn()),
93 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
94 {
95 _exceptions = nullptr;
96 set_map(nullptr);
97 DEBUG_ONLY(_sp = -99);
98 DEBUG_ONLY(set_bci(-99));
99 }
100
101 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
102 : Phase(Phase::Parser),
103 _env(C->env()),
104 _gvn(igvn),
105 _exceptions(nullptr),
106 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
107 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
108 assert(sft->next_exception() == nullptr, "must not have a pending exception");
109 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
110 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
111 for (uint i = 0; i < sft->req(); i++) {
112 cloned_map->init_req(i, sft->in(i));
113 }
114 igvn.record_for_igvn(cloned_map);
115 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
116 current->set_map(cloned_map);
117 }
118 set_jvms(cloned_jvms);
119 set_all_memory(cloned_map->memory());
120 }
121
122 //---------------------------clean_stack---------------------------------------
123 // Clear away rubbish from the stack area of the JVM state.
124 // This destroys any arguments that may be waiting on the stack.
125 void GraphKit::clean_stack(int from_sp) {
126 SafePointNode* map = this->map();
127 JVMState* jvms = this->jvms();
128 int stk_size = jvms->stk_size();
129 int stkoff = jvms->stkoff();
130 Node* top = this->top();
131 for (int i = from_sp; i < stk_size; i++) {
132 if (map->in(stkoff + i) != top) {
133 map->set_req(stkoff + i, top);
134 }
135 }
136 }
137
138
139 //--------------------------------sync_jvms-----------------------------------
140 // Make sure our current jvms agrees with our parse state.
141 JVMState* GraphKit::sync_jvms() const {
142 JVMState* jvms = this->jvms();
143 jvms->set_bci(bci()); // Record the new bci in the JVMState
144 jvms->set_sp(sp()); // Record the new sp in the JVMState
145 assert(jvms_in_sync(), "jvms is now in sync");
146 return jvms;
147 }
148
149 //--------------------------------sync_jvms_for_reexecute---------------------
150 // Make sure our current jvms agrees with our parse state. This version
151 // uses the reexecute_sp for reexecuting bytecodes.
152 JVMState* GraphKit::sync_jvms_for_reexecute() {
153 JVMState* jvms = this->jvms();
154 jvms->set_bci(bci()); // Record the new bci in the JVMState
155 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
156 return jvms;
157 }
158
159 #ifdef ASSERT
160 bool GraphKit::jvms_in_sync() const {
161 Parse* parse = is_Parse();
162 if (parse == nullptr) {
163 if (bci() != jvms()->bci()) return false;
164 if (sp() != (int)jvms()->sp()) return false;
165 return true;
166 }
167 if (jvms()->method() != parse->method()) return false;
168 if (jvms()->bci() != parse->bci()) return false;
169 int jvms_sp = jvms()->sp();
170 if (jvms_sp != parse->sp()) return false;
171 int jvms_depth = jvms()->depth();
172 if (jvms_depth != parse->depth()) return false;
173 return true;
174 }
175
176 // Local helper checks for special internal merge points
177 // used to accumulate and merge exception states.
178 // They are marked by the region's in(0) edge being the map itself.
179 // Such merge points must never "escape" into the parser at large,
180 // until they have been handed to gvn.transform.
181 static bool is_hidden_merge(Node* reg) {
182 if (reg == nullptr) return false;
183 if (reg->is_Phi()) {
184 reg = reg->in(0);
185 if (reg == nullptr) return false;
186 }
187 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
188 }
189
190 void GraphKit::verify_map() const {
191 if (map() == nullptr) return; // null map is OK
192 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
193 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
194 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
195 }
196
197 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
198 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
199 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
200 }
201 #endif
202
203 //---------------------------stop_and_kill_map---------------------------------
204 // Set _map to null, signalling a stop to further bytecode execution.
205 // First smash the current map's control to a constant, to mark it dead.
206 void GraphKit::stop_and_kill_map() {
207 SafePointNode* dead_map = stop();
208 if (dead_map != nullptr) {
209 dead_map->disconnect_inputs(C); // Mark the map as killed.
210 assert(dead_map->is_killed(), "must be so marked");
211 }
212 }
213
214
215 //--------------------------------stopped--------------------------------------
216 // Tell if _map is null, or control is top.
217 bool GraphKit::stopped() {
218 if (map() == nullptr) return true;
219 else if (control() == top()) return true;
220 else return false;
221 }
222
223
224 //-----------------------------has_exception_handler----------------------------------
225 // Tell if this method or any caller method has exception handlers.
226 bool GraphKit::has_exception_handler() {
227 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
228 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
229 return true;
230 }
231 }
232 return false;
233 }
234
235 //------------------------------save_ex_oop------------------------------------
236 // Save an exception without blowing stack contents or other JVM state.
237 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
238 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
239 ex_map->add_req(ex_oop);
240 DEBUG_ONLY(verify_exception_state(ex_map));
241 }
242
243 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
244 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
245 Node* ex_oop = ex_map->in(ex_map->req()-1);
246 if (clear_it) ex_map->del_req(ex_map->req()-1);
247 return ex_oop;
248 }
249
250 //-----------------------------saved_ex_oop------------------------------------
251 // Recover a saved exception from its map.
252 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
253 return common_saved_ex_oop(ex_map, false);
254 }
255
256 //--------------------------clear_saved_ex_oop---------------------------------
257 // Erase a previously saved exception from its map.
258 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
259 return common_saved_ex_oop(ex_map, true);
260 }
261
262 #ifdef ASSERT
263 //---------------------------has_saved_ex_oop----------------------------------
264 // Erase a previously saved exception from its map.
265 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
266 return ex_map->req() == ex_map->jvms()->endoff()+1;
267 }
268 #endif
269
270 //-------------------------make_exception_state--------------------------------
271 // Turn the current JVM state into an exception state, appending the ex_oop.
272 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
273 sync_jvms();
274 SafePointNode* ex_map = stop(); // do not manipulate this map any more
275 set_saved_ex_oop(ex_map, ex_oop);
276 return ex_map;
277 }
278
279
280 //--------------------------add_exception_state--------------------------------
281 // Add an exception to my list of exceptions.
282 void GraphKit::add_exception_state(SafePointNode* ex_map) {
283 if (ex_map == nullptr || ex_map->control() == top()) {
284 return;
285 }
286 #ifdef ASSERT
287 verify_exception_state(ex_map);
288 if (has_exceptions()) {
289 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
290 }
291 #endif
292
293 // If there is already an exception of exactly this type, merge with it.
294 // In particular, null-checks and other low-level exceptions common up here.
295 Node* ex_oop = saved_ex_oop(ex_map);
296 const Type* ex_type = _gvn.type(ex_oop);
297 if (ex_oop == top()) {
298 // No action needed.
299 return;
300 }
301 assert(ex_type->isa_instptr(), "exception must be an instance");
302 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
303 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
304 // We check sp also because call bytecodes can generate exceptions
305 // both before and after arguments are popped!
306 if (ex_type2 == ex_type
307 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
308 combine_exception_states(ex_map, e2);
309 return;
310 }
311 }
312
313 // No pre-existing exception of the same type. Chain it on the list.
314 push_exception_state(ex_map);
315 }
316
317 //-----------------------add_exception_states_from-----------------------------
318 void GraphKit::add_exception_states_from(JVMState* jvms) {
319 SafePointNode* ex_map = jvms->map()->next_exception();
320 if (ex_map != nullptr) {
321 jvms->map()->set_next_exception(nullptr);
322 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
323 next_map = ex_map->next_exception();
324 ex_map->set_next_exception(nullptr);
325 add_exception_state(ex_map);
326 }
327 }
328 }
329
330 //-----------------------transfer_exceptions_into_jvms-------------------------
331 JVMState* GraphKit::transfer_exceptions_into_jvms() {
332 if (map() == nullptr) {
333 // We need a JVMS to carry the exceptions, but the map has gone away.
334 // Create a scratch JVMS, cloned from any of the exception states...
335 if (has_exceptions()) {
336 _map = _exceptions;
337 _map = clone_map();
338 _map->set_next_exception(nullptr);
339 clear_saved_ex_oop(_map);
340 DEBUG_ONLY(verify_map());
341 } else {
342 // ...or created from scratch
343 JVMState* jvms = new (C) JVMState(_method, nullptr);
344 jvms->set_bci(_bci);
345 jvms->set_sp(_sp);
346 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
347 set_jvms(jvms);
348 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
349 set_all_memory(top());
350 while (map()->req() < jvms->endoff()) map()->add_req(top());
351 }
352 // (This is a kludge, in case you didn't notice.)
353 set_control(top());
354 }
355 JVMState* jvms = sync_jvms();
356 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
357 jvms->map()->set_next_exception(_exceptions);
358 _exceptions = nullptr; // done with this set of exceptions
359 return jvms;
360 }
361
362 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
363 assert(is_hidden_merge(dstphi), "must be a special merge node");
364 assert(is_hidden_merge(srcphi), "must be a special merge node");
365 uint limit = srcphi->req();
366 for (uint i = PhiNode::Input; i < limit; i++) {
367 dstphi->add_req(srcphi->in(i));
368 }
369 }
370 static inline void add_one_req(Node* dstphi, Node* src) {
371 assert(is_hidden_merge(dstphi), "must be a special merge node");
372 assert(!is_hidden_merge(src), "must not be a special merge node");
373 dstphi->add_req(src);
374 }
375
376 //-----------------------combine_exception_states------------------------------
377 // This helper function combines exception states by building phis on a
378 // specially marked state-merging region. These regions and phis are
379 // untransformed, and can build up gradually. The region is marked by
380 // having a control input of its exception map, rather than null. Such
381 // regions do not appear except in this function, and in use_exception_state.
382 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
383 if (failing_internal()) {
384 return; // dying anyway...
385 }
386 JVMState* ex_jvms = ex_map->_jvms;
387 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
388 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
389 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
390 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
391 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
392 assert(ex_map->req() == phi_map->req(), "matching maps");
393 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
394 Node* hidden_merge_mark = root();
395 Node* region = phi_map->control();
396 MergeMemNode* phi_mem = phi_map->merged_memory();
397 MergeMemNode* ex_mem = ex_map->merged_memory();
398 if (region->in(0) != hidden_merge_mark) {
399 // The control input is not (yet) a specially-marked region in phi_map.
400 // Make it so, and build some phis.
401 region = new RegionNode(2);
402 _gvn.set_type(region, Type::CONTROL);
403 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
404 region->init_req(1, phi_map->control());
405 phi_map->set_control(region);
406 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
407 record_for_igvn(io_phi);
408 _gvn.set_type(io_phi, Type::ABIO);
409 phi_map->set_i_o(io_phi);
410 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
411 Node* m = mms.memory();
412 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
413 record_for_igvn(m_phi);
414 _gvn.set_type(m_phi, Type::MEMORY);
415 mms.set_memory(m_phi);
416 }
417 }
418
419 // Either or both of phi_map and ex_map might already be converted into phis.
420 Node* ex_control = ex_map->control();
421 // if there is special marking on ex_map also, we add multiple edges from src
422 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
423 // how wide was the destination phi_map, originally?
424 uint orig_width = region->req();
425
426 if (add_multiple) {
427 add_n_reqs(region, ex_control);
428 add_n_reqs(phi_map->i_o(), ex_map->i_o());
429 } else {
430 // ex_map has no merges, so we just add single edges everywhere
431 add_one_req(region, ex_control);
432 add_one_req(phi_map->i_o(), ex_map->i_o());
433 }
434 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
435 if (mms.is_empty()) {
436 // get a copy of the base memory, and patch some inputs into it
437 const TypePtr* adr_type = mms.adr_type(C);
438 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
439 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
440 mms.set_memory(phi);
441 // Prepare to append interesting stuff onto the newly sliced phi:
442 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
443 }
444 // Append stuff from ex_map:
445 if (add_multiple) {
446 add_n_reqs(mms.memory(), mms.memory2());
447 } else {
448 add_one_req(mms.memory(), mms.memory2());
449 }
450 }
451 uint limit = ex_map->req();
452 for (uint i = TypeFunc::Parms; i < limit; i++) {
453 // Skip everything in the JVMS after tos. (The ex_oop follows.)
454 if (i == tos) i = ex_jvms->monoff();
455 Node* src = ex_map->in(i);
456 Node* dst = phi_map->in(i);
457 if (src != dst) {
458 PhiNode* phi;
459 if (dst->in(0) != region) {
460 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
461 record_for_igvn(phi);
462 _gvn.set_type(phi, phi->type());
463 phi_map->set_req(i, dst);
464 // Prepare to append interesting stuff onto the new phi:
465 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
466 } else {
467 assert(dst->is_Phi(), "nobody else uses a hidden region");
468 phi = dst->as_Phi();
469 }
470 if (add_multiple && src->in(0) == ex_control) {
471 // Both are phis.
472 add_n_reqs(dst, src);
473 } else {
474 while (dst->req() < region->req()) add_one_req(dst, src);
475 }
476 const Type* srctype = _gvn.type(src);
477 if (phi->type() != srctype) {
478 const Type* dsttype = phi->type()->meet_speculative(srctype);
479 if (phi->type() != dsttype) {
480 phi->set_type(dsttype);
481 _gvn.set_type(phi, dsttype);
482 }
483 }
484 }
485 }
486 phi_map->merge_replaced_nodes_with(ex_map);
487 }
488
489 //--------------------------use_exception_state--------------------------------
490 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
491 if (failing_internal()) { stop(); return top(); }
492 Node* region = phi_map->control();
493 Node* hidden_merge_mark = root();
494 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
495 Node* ex_oop = clear_saved_ex_oop(phi_map);
496 if (region->in(0) == hidden_merge_mark) {
497 // Special marking for internal ex-states. Process the phis now.
498 region->set_req(0, region); // now it's an ordinary region
499 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
500 // Note: Setting the jvms also sets the bci and sp.
501 set_control(_gvn.transform(region));
502 uint tos = jvms()->stkoff() + sp();
503 for (uint i = 1; i < tos; i++) {
504 Node* x = phi_map->in(i);
505 if (x->in(0) == region) {
506 assert(x->is_Phi(), "expected a special phi");
507 phi_map->set_req(i, _gvn.transform(x));
508 }
509 }
510 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
511 Node* x = mms.memory();
512 if (x->in(0) == region) {
513 assert(x->is_Phi(), "nobody else uses a hidden region");
514 mms.set_memory(_gvn.transform(x));
515 }
516 }
517 if (ex_oop->in(0) == region) {
518 assert(ex_oop->is_Phi(), "expected a special phi");
519 ex_oop = _gvn.transform(ex_oop);
520 }
521 } else {
522 set_jvms(phi_map->jvms());
523 }
524
525 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
526 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
527 return ex_oop;
528 }
529
530 //---------------------------------java_bc-------------------------------------
531 Bytecodes::Code GraphKit::java_bc() const {
532 ciMethod* method = this->method();
533 int bci = this->bci();
534 if (method != nullptr && bci != InvocationEntryBci)
535 return method->java_code_at_bci(bci);
536 else
537 return Bytecodes::_illegal;
538 }
539
540 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
541 bool must_throw) {
542 // if the exception capability is set, then we will generate code
543 // to check the JavaThread.should_post_on_exceptions flag to see
544 // if we actually need to report exception events (for this
545 // thread). If we don't need to report exception events, we will
546 // take the normal fast path provided by add_exception_events. If
547 // exception event reporting is enabled for this thread, we will
548 // take the uncommon_trap in the BuildCutout below.
549
550 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
551 Node* jthread = _gvn.transform(new ThreadLocalNode());
552 Node* adr = off_heap_plus_addr(jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
553 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
554
555 // Test the should_post_on_exceptions_flag vs. 0
556 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
557 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
558
559 // Branch to slow_path if should_post_on_exceptions_flag was true
560 { BuildCutout unless(this, tst, PROB_MAX);
561 // Do not try anything fancy if we're notifying the VM on every throw.
562 // Cf. case Bytecodes::_athrow in parse2.cpp.
563 uncommon_trap(reason, Deoptimization::Action_none,
564 (ciKlass*)nullptr, (char*)nullptr, must_throw);
565 }
566
567 }
568
569 //------------------------------builtin_throw----------------------------------
570 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
571 builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true);
572 }
573
574 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason,
575 ciInstance* ex_obj,
576 bool allow_too_many_traps) {
577 // If this throw happens frequently, an uncommon trap might cause
578 // a performance pothole. If there is a local exception handler,
579 // and if this particular bytecode appears to be deoptimizing often,
580 // let us handle the throw inline, with a preconstructed instance.
581 // Note: If the deopt count has blown up, the uncommon trap
582 // runtime is going to flush this nmethod, not matter what.
583 if (is_builtin_throw_hot(reason)) {
584 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
585 // If the throw is local, we use a pre-existing instance and
586 // punt on the backtrace. This would lead to a missing backtrace
587 // (a repeat of 4292742) if the backtrace object is ever asked
588 // for its backtrace.
589 // Fixing this remaining case of 4292742 requires some flavor of
590 // escape analysis. Leave that for the future.
591 if (env()->jvmti_can_post_on_exceptions()) {
592 // check if we must post exception events, take uncommon trap if so
593 uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/);
594 // here if should_post_on_exceptions is false
595 // continue on with the normal codegen
596 }
597
598 // Cheat with a preallocated exception object.
599 if (C->log() != nullptr)
600 C->log()->elem("hot_throw preallocated='1' reason='%s'",
601 Deoptimization::trap_reason_name(reason));
602 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
603 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
604
605 // Clear the detail message of the preallocated exception object.
606 // Weblogic sometimes mutates the detail message of exceptions
607 // using reflection.
608 int offset = java_lang_Throwable::get_detailMessage_offset();
609 const TypePtr* adr_typ = ex_con->add_offset(offset);
610
611 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
612 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
613 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
614
615 if (!method()->has_exception_handlers()) {
616 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
617 // This prevents issues with exception handling and late inlining.
618 set_sp(0);
619 clean_stack(0);
620 }
621
622 add_exception_state(make_exception_state(ex_node));
623 return;
624 } else if (builtin_throw_too_many_traps(reason, ex_obj)) {
625 // We cannot afford to take too many traps here. Suffer in the interpreter instead.
626 assert(allow_too_many_traps, "not allowed");
627 if (C->log() != nullptr) {
628 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
629 Deoptimization::trap_reason_name(reason),
630 C->trap_count(reason));
631 }
632 uncommon_trap(reason, Deoptimization::Action_none,
633 (ciKlass*) nullptr, (char*) nullptr,
634 true /*must_throw*/);
635 return;
636 }
637 }
638
639 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
640 // It won't be much cheaper than bailing to the interp., since we'll
641 // have to pass up all the debug-info, and the runtime will have to
642 // create the stack trace.
643
644 // Usual case: Bail to interpreter.
645 // Reserve the right to recompile if we haven't seen anything yet.
646
647 // "must_throw" prunes the JVM state to include only the stack, if there
648 // are no local exception handlers. This should cut down on register
649 // allocation time and code size, by drastically reducing the number
650 // of in-edges on the call to the uncommon trap.
651 uncommon_trap(reason, Deoptimization::Action_maybe_recompile,
652 (ciKlass*) nullptr, (char*) nullptr,
653 true /*must_throw*/);
654 }
655
656 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) {
657 // If this particular condition has not yet happened at this
658 // bytecode, then use the uncommon trap mechanism, and allow for
659 // a future recompilation if several traps occur here.
660 // If the throw is hot, try to use a more complicated inline mechanism
661 // which keeps execution inside the compiled code.
662 if (ProfileTraps) {
663 if (too_many_traps(reason)) {
664 return true;
665 }
666 // (If there is no MDO at all, assume it is early in
667 // execution, and that any deopts are part of the
668 // startup transient, and don't need to be remembered.)
669
670 // Also, if there is a local exception handler, treat all throws
671 // as hot if there has been at least one in this method.
672 if (C->trap_count(reason) != 0 &&
673 method()->method_data()->trap_count(reason) != 0 &&
674 has_exception_handler()) {
675 return true;
676 }
677 }
678 return false;
679 }
680
681 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason,
682 ciInstance* ex_obj) {
683 if (is_builtin_throw_hot(reason)) {
684 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
685 return false; // no traps; throws preallocated exception instead
686 }
687 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
688 if (method()->method_data()->trap_recompiled_at(bci(), m) ||
689 C->too_many_traps(reason)) {
690 return true;
691 }
692 }
693 return false;
694 }
695
696 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const {
697 // Preallocated exception objects to use when we don't need the backtrace.
698 switch (reason) {
699 case Deoptimization::Reason_null_check:
700 return env()->NullPointerException_instance();
701 case Deoptimization::Reason_div0_check:
702 return env()->ArithmeticException_instance();
703 case Deoptimization::Reason_range_check:
704 return env()->ArrayIndexOutOfBoundsException_instance();
705 case Deoptimization::Reason_class_check:
706 return env()->ClassCastException_instance();
707 case Deoptimization::Reason_array_check:
708 return env()->ArrayStoreException_instance();
709 default:
710 return nullptr;
711 }
712 }
713
714 GraphKit::SavedState::SavedState(GraphKit* kit) :
715 _kit(kit),
716 _sp(kit->sp()),
717 _jvms(kit->jvms()),
718 _map(kit->clone_map()),
719 _discarded(false)
720 {
721 for (DUIterator_Fast imax, i = kit->control()->fast_outs(imax); i < imax; i++) {
722 Node* out = kit->control()->fast_out(i);
723 if (out->is_CFG()) {
724 _ctrl_succ.push(out);
725 }
726 }
727 }
728
729 GraphKit::SavedState::~SavedState() {
730 if (_discarded) {
731 _kit->destruct_map_clone(_map);
732 return;
733 }
734 _kit->jvms()->set_map(_map);
735 _kit->jvms()->set_sp(_sp);
736 _map->set_jvms(_kit->jvms());
737 _kit->set_map(_map);
738 _kit->set_sp(_sp);
739 for (DUIterator_Fast imax, i = _kit->control()->fast_outs(imax); i < imax; i++) {
740 Node* out = _kit->control()->fast_out(i);
741 if (out->is_CFG() && out->in(0) == _kit->control() && out != _kit->map() && !_ctrl_succ.member(out)) {
742 _kit->_gvn.hash_delete(out);
743 out->set_req(0, _kit->C->top());
744 _kit->C->record_for_igvn(out);
745 --i; --imax;
746 _kit->_gvn.hash_find_insert(out);
747 }
748 }
749 }
750
751 void GraphKit::SavedState::discard() {
752 _discarded = true;
753 }
754
755
756 //----------------------------PreserveJVMState---------------------------------
757 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
758 DEBUG_ONLY(kit->verify_map());
759 _kit = kit;
760 _map = kit->map(); // preserve the map
761 _sp = kit->sp();
762 kit->set_map(clone_map ? kit->clone_map() : nullptr);
763 #ifdef ASSERT
764 _bci = kit->bci();
765 Parse* parser = kit->is_Parse();
766 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
767 _block = block;
768 #endif
769 }
770 PreserveJVMState::~PreserveJVMState() {
771 GraphKit* kit = _kit;
772 #ifdef ASSERT
773 assert(kit->bci() == _bci, "bci must not shift");
774 Parse* parser = kit->is_Parse();
775 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
776 assert(block == _block, "block must not shift");
777 #endif
778 kit->set_map(_map);
779 kit->set_sp(_sp);
780 }
781
782
783 //-----------------------------BuildCutout-------------------------------------
784 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
785 : PreserveJVMState(kit)
786 {
787 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
788 SafePointNode* outer_map = _map; // preserved map is caller's
789 SafePointNode* inner_map = kit->map();
790 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
791 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
792 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
793 }
794 BuildCutout::~BuildCutout() {
795 GraphKit* kit = _kit;
796 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
797 }
798
799 //---------------------------PreserveReexecuteState----------------------------
800 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
801 assert(!kit->stopped(), "must call stopped() before");
802 _kit = kit;
803 _sp = kit->sp();
804 _reexecute = kit->jvms()->_reexecute;
805 }
806 PreserveReexecuteState::~PreserveReexecuteState() {
807 if (_kit->stopped()) return;
808 _kit->jvms()->_reexecute = _reexecute;
809 _kit->set_sp(_sp);
810 }
811
812 //------------------------------clone_map--------------------------------------
813 // Implementation of PreserveJVMState
814 //
815 // Only clone_map(...) here. If this function is only used in the
816 // PreserveJVMState class we may want to get rid of this extra
817 // function eventually and do it all there.
818
819 SafePointNode* GraphKit::clone_map() {
820 if (map() == nullptr) return nullptr;
821
822 // Clone the memory edge first
823 Node* mem = MergeMemNode::make(map()->memory());
824 gvn().set_type_bottom(mem);
825
826 SafePointNode *clonemap = (SafePointNode*)map()->clone();
827 JVMState* jvms = this->jvms();
828 JVMState* clonejvms = jvms->clone_shallow(C);
829 clonemap->set_memory(mem);
830 clonemap->set_jvms(clonejvms);
831 clonejvms->set_map(clonemap);
832 record_for_igvn(clonemap);
833 gvn().set_type_bottom(clonemap);
834 return clonemap;
835 }
836
837 //-----------------------------destruct_map_clone------------------------------
838 //
839 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
840 // to destruct/free/delete in the exact opposite order as clone_map().
841 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
842 if (sfp == nullptr) return;
843
844 Node* mem = sfp->memory();
845 JVMState* jvms = sfp->jvms();
846
847 if (jvms != nullptr) {
848 delete jvms;
849 }
850
851 remove_for_igvn(sfp);
852 gvn().clear_type(sfp);
853 sfp->destruct(&_gvn);
854
855 if (mem != nullptr) {
856 gvn().clear_type(mem);
857 mem->destruct(&_gvn);
858 }
859 }
860
861 //-----------------------------set_map_clone-----------------------------------
862 void GraphKit::set_map_clone(SafePointNode* m) {
863 _map = m;
864 _map = clone_map();
865 _map->set_next_exception(nullptr);
866 DEBUG_ONLY(verify_map());
867 }
868
869
870 //----------------------------kill_dead_locals---------------------------------
871 // Detect any locals which are known to be dead, and force them to top.
872 void GraphKit::kill_dead_locals() {
873 // Consult the liveness information for the locals. If any
874 // of them are unused, then they can be replaced by top(). This
875 // should help register allocation time and cut down on the size
876 // of the deoptimization information.
877
878 // This call is made from many of the bytecode handling
879 // subroutines called from the Big Switch in do_one_bytecode.
880 // Every bytecode which might include a slow path is responsible
881 // for killing its dead locals. The more consistent we
882 // are about killing deads, the fewer useless phis will be
883 // constructed for them at various merge points.
884
885 // bci can be -1 (InvocationEntryBci). We return the entry
886 // liveness for the method.
887
888 if (method() == nullptr || method()->code_size() == 0) {
889 // We are building a graph for a call to a native method.
890 // All locals are live.
891 return;
892 }
893
894 ResourceMark rm;
895
896 // Consult the liveness information for the locals. If any
897 // of them are unused, then they can be replaced by top(). This
898 // should help register allocation time and cut down on the size
899 // of the deoptimization information.
900 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
901
902 int len = (int)live_locals.size();
903 assert(len <= jvms()->loc_size(), "too many live locals");
904 for (int local = 0; local < len; local++) {
905 if (!live_locals.at(local)) {
906 set_local(local, top());
907 }
908 }
909 }
910
911 #ifdef ASSERT
912 //-------------------------dead_locals_are_killed------------------------------
913 // Return true if all dead locals are set to top in the map.
914 // Used to assert "clean" debug info at various points.
915 bool GraphKit::dead_locals_are_killed() {
916 if (method() == nullptr || method()->code_size() == 0) {
917 // No locals need to be dead, so all is as it should be.
918 return true;
919 }
920
921 // Make sure somebody called kill_dead_locals upstream.
922 ResourceMark rm;
923 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
924 if (jvms->loc_size() == 0) continue; // no locals to consult
925 SafePointNode* map = jvms->map();
926 ciMethod* method = jvms->method();
927 int bci = jvms->bci();
928 if (jvms == this->jvms()) {
929 bci = this->bci(); // it might not yet be synched
930 }
931 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
932 int len = (int)live_locals.size();
933 if (!live_locals.is_valid() || len == 0)
934 // This method is trivial, or is poisoned by a breakpoint.
935 return true;
936 assert(len == jvms->loc_size(), "live map consistent with locals map");
937 for (int local = 0; local < len; local++) {
938 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
939 if (PrintMiscellaneous && (Verbose || WizardMode)) {
940 tty->print_cr("Zombie local %d: ", local);
941 jvms->dump();
942 }
943 return false;
944 }
945 }
946 }
947 return true;
948 }
949
950 #endif //ASSERT
951
952 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
953 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
954 ciMethod* cur_method = jvms->method();
955 int cur_bci = jvms->bci();
956 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
957 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
958 return Interpreter::bytecode_should_reexecute(code) ||
959 (is_anewarray && (code == Bytecodes::_multianewarray));
960 // Reexecute _multianewarray bytecode which was replaced with
961 // sequence of [a]newarray. See Parse::do_multianewarray().
962 //
963 // Note: interpreter should not have it set since this optimization
964 // is limited by dimensions and guarded by flag so in some cases
965 // multianewarray() runtime calls will be generated and
966 // the bytecode should not be reexecutes (stack will not be reset).
967 } else {
968 return false;
969 }
970 }
971
972 // Helper function for adding JVMState and debug information to node
973 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
974 // Add the safepoint edges to the call (or other safepoint).
975
976 // Make sure dead locals are set to top. This
977 // should help register allocation time and cut down on the size
978 // of the deoptimization information.
979 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
980
981 // Walk the inline list to fill in the correct set of JVMState's
982 // Also fill in the associated edges for each JVMState.
983
984 // If the bytecode needs to be reexecuted we need to put
985 // the arguments back on the stack.
986 const bool should_reexecute = jvms()->should_reexecute();
987 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
988
989 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
990 // undefined if the bci is different. This is normal for Parse but it
991 // should not happen for LibraryCallKit because only one bci is processed.
992 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
993 "in LibraryCallKit the reexecute bit should not change");
994
995 // If we are guaranteed to throw, we can prune everything but the
996 // input to the current bytecode.
997 bool can_prune_locals = false;
998 uint stack_slots_not_pruned = 0;
999 int inputs = 0, depth = 0;
1000 if (must_throw) {
1001 assert(method() == youngest_jvms->method(), "sanity");
1002 if (compute_stack_effects(inputs, depth)) {
1003 can_prune_locals = true;
1004 stack_slots_not_pruned = inputs;
1005 }
1006 }
1007
1008 if (env()->should_retain_local_variables()) {
1009 // At any safepoint, this method can get breakpointed, which would
1010 // then require an immediate deoptimization.
1011 can_prune_locals = false; // do not prune locals
1012 stack_slots_not_pruned = 0;
1013 }
1014
1015 // do not scribble on the input jvms
1016 JVMState* out_jvms = youngest_jvms->clone_deep(C);
1017 call->set_jvms(out_jvms); // Start jvms list for call node
1018
1019 // For a known set of bytecodes, the interpreter should reexecute them if
1020 // deoptimization happens. We set the reexecute state for them here
1021 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
1022 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
1023 #ifdef ASSERT
1024 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
1025 assert(method() == youngest_jvms->method(), "sanity");
1026 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
1027 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
1028 #endif // ASSERT
1029 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
1030 }
1031
1032 // Presize the call:
1033 DEBUG_ONLY(uint non_debug_edges = call->req());
1034 call->add_req_batch(top(), youngest_jvms->debug_depth());
1035 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
1036
1037 // Set up edges so that the call looks like this:
1038 // Call [state:] ctl io mem fptr retadr
1039 // [parms:] parm0 ... parmN
1040 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1041 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
1042 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1043 // Note that caller debug info precedes callee debug info.
1044
1045 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1046 uint debug_ptr = call->req();
1047
1048 // Loop over the map input edges associated with jvms, add them
1049 // to the call node, & reset all offsets to match call node array.
1050 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1051 uint debug_end = debug_ptr;
1052 uint debug_start = debug_ptr - in_jvms->debug_size();
1053 debug_ptr = debug_start; // back up the ptr
1054
1055 uint p = debug_start; // walks forward in [debug_start, debug_end)
1056 uint j, k, l;
1057 SafePointNode* in_map = in_jvms->map();
1058 out_jvms->set_map(call);
1059
1060 if (can_prune_locals) {
1061 assert(in_jvms->method() == out_jvms->method(), "sanity");
1062 // If the current throw can reach an exception handler in this JVMS,
1063 // then we must keep everything live that can reach that handler.
1064 // As a quick and dirty approximation, we look for any handlers at all.
1065 if (in_jvms->method()->has_exception_handlers()) {
1066 can_prune_locals = false;
1067 }
1068 }
1069
1070 // Add the Locals
1071 k = in_jvms->locoff();
1072 l = in_jvms->loc_size();
1073 out_jvms->set_locoff(p);
1074 if (!can_prune_locals) {
1075 for (j = 0; j < l; j++) {
1076 call->set_req(p++, in_map->in(k + j));
1077 }
1078 } else {
1079 p += l; // already set to top above by add_req_batch
1080 }
1081
1082 // Add the Expression Stack
1083 k = in_jvms->stkoff();
1084 l = in_jvms->sp();
1085 out_jvms->set_stkoff(p);
1086 if (!can_prune_locals) {
1087 for (j = 0; j < l; j++) {
1088 call->set_req(p++, in_map->in(k + j));
1089 }
1090 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1091 // Divide stack into {S0,...,S1}, where S0 is set to top.
1092 uint s1 = stack_slots_not_pruned;
1093 stack_slots_not_pruned = 0; // for next iteration
1094 if (s1 > l) s1 = l;
1095 uint s0 = l - s1;
1096 p += s0; // skip the tops preinstalled by add_req_batch
1097 for (j = s0; j < l; j++)
1098 call->set_req(p++, in_map->in(k+j));
1099 } else {
1100 p += l; // already set to top above by add_req_batch
1101 }
1102
1103 // Add the Monitors
1104 k = in_jvms->monoff();
1105 l = in_jvms->mon_size();
1106 out_jvms->set_monoff(p);
1107 for (j = 0; j < l; j++)
1108 call->set_req(p++, in_map->in(k+j));
1109
1110 // Copy any scalar object fields.
1111 k = in_jvms->scloff();
1112 l = in_jvms->scl_size();
1113 out_jvms->set_scloff(p);
1114 for (j = 0; j < l; j++)
1115 call->set_req(p++, in_map->in(k+j));
1116
1117 // Finish the new jvms.
1118 out_jvms->set_endoff(p);
1119
1120 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1121 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1122 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1123 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1124 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1125 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1126
1127 // Update the two tail pointers in parallel.
1128 out_jvms = out_jvms->caller();
1129 in_jvms = in_jvms->caller();
1130 }
1131
1132 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1133
1134 // Test the correctness of JVMState::debug_xxx accessors:
1135 assert(call->jvms()->debug_start() == non_debug_edges, "");
1136 assert(call->jvms()->debug_end() == call->req(), "");
1137 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1138 }
1139
1140 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1141 Bytecodes::Code code = java_bc();
1142 if (code == Bytecodes::_wide) {
1143 code = method()->java_code_at_bci(bci() + 1);
1144 }
1145
1146 if (code != Bytecodes::_illegal) {
1147 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1148 }
1149
1150 auto rsize = [&]() {
1151 assert(code != Bytecodes::_illegal, "code is illegal!");
1152 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1153 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1154 };
1155
1156 switch (code) {
1157 case Bytecodes::_illegal:
1158 return false;
1159
1160 case Bytecodes::_ldc:
1161 case Bytecodes::_ldc_w:
1162 case Bytecodes::_ldc2_w:
1163 inputs = 0;
1164 break;
1165
1166 case Bytecodes::_dup: inputs = 1; break;
1167 case Bytecodes::_dup_x1: inputs = 2; break;
1168 case Bytecodes::_dup_x2: inputs = 3; break;
1169 case Bytecodes::_dup2: inputs = 2; break;
1170 case Bytecodes::_dup2_x1: inputs = 3; break;
1171 case Bytecodes::_dup2_x2: inputs = 4; break;
1172 case Bytecodes::_swap: inputs = 2; break;
1173 case Bytecodes::_arraylength: inputs = 1; break;
1174
1175 case Bytecodes::_getstatic:
1176 case Bytecodes::_putstatic:
1177 case Bytecodes::_getfield:
1178 case Bytecodes::_putfield:
1179 {
1180 bool ignored_will_link;
1181 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1182 int size = field->type()->size();
1183 bool is_get = (depth >= 0), is_static = (depth & 1);
1184 inputs = (is_static ? 0 : 1);
1185 if (is_get) {
1186 depth = size - inputs;
1187 } else {
1188 inputs += size; // putxxx pops the value from the stack
1189 depth = - inputs;
1190 }
1191 }
1192 break;
1193
1194 case Bytecodes::_invokevirtual:
1195 case Bytecodes::_invokespecial:
1196 case Bytecodes::_invokestatic:
1197 case Bytecodes::_invokedynamic:
1198 case Bytecodes::_invokeinterface:
1199 {
1200 bool ignored_will_link;
1201 ciSignature* declared_signature = nullptr;
1202 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1203 assert(declared_signature != nullptr, "cannot be null");
1204 inputs = declared_signature->arg_size_for_bc(code);
1205 int size = declared_signature->return_type()->size();
1206 depth = size - inputs;
1207 }
1208 break;
1209
1210 case Bytecodes::_multianewarray:
1211 {
1212 ciBytecodeStream iter(method());
1213 iter.reset_to_bci(bci());
1214 iter.next();
1215 inputs = iter.get_dimensions();
1216 assert(rsize() == 1, "");
1217 depth = 1 - inputs;
1218 }
1219 break;
1220
1221 case Bytecodes::_ireturn:
1222 case Bytecodes::_lreturn:
1223 case Bytecodes::_freturn:
1224 case Bytecodes::_dreturn:
1225 case Bytecodes::_areturn:
1226 assert(rsize() == -depth, "");
1227 inputs = -depth;
1228 break;
1229
1230 case Bytecodes::_jsr:
1231 case Bytecodes::_jsr_w:
1232 inputs = 0;
1233 depth = 1; // S.B. depth=1, not zero
1234 break;
1235
1236 default:
1237 // bytecode produces a typed result
1238 inputs = rsize() - depth;
1239 assert(inputs >= 0, "");
1240 break;
1241 }
1242
1243 #ifdef ASSERT
1244 // spot check
1245 int outputs = depth + inputs;
1246 assert(outputs >= 0, "sanity");
1247 switch (code) {
1248 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1249 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1250 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1251 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1252 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1253 default: break;
1254 }
1255 #endif //ASSERT
1256
1257 return true;
1258 }
1259
1260
1261
1262 //------------------------------basic_plus_adr---------------------------------
1263 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1264 // short-circuit a common case
1265 if (offset == MakeConX(0)) {
1266 return ptr;
1267 }
1268 #ifdef ASSERT
1269 // Both 32-bit and 64-bit zeros should have been handled by the previous `if`
1270 // statement, so if we see either 32-bit or 64-bit zeros here, then we have a
1271 // problem.
1272 if (offset->is_Con()) {
1273 const Type* t = offset->bottom_type();
1274 bool is_zero_int = t->isa_int() && t->is_int()->get_con() == 0;
1275 bool is_zero_long = t->isa_long() && t->is_long()->get_con() == 0;
1276 assert(!is_zero_int && !is_zero_long,
1277 "Unexpected zero offset - should have matched MakeConX(0)");
1278 }
1279 #endif
1280 return _gvn.transform(AddPNode::make_with_base(base, ptr, offset));
1281 }
1282
1283 Node* GraphKit::ConvI2L(Node* offset) {
1284 // short-circuit a common case
1285 jint offset_con = find_int_con(offset, Type::OffsetBot);
1286 if (offset_con != Type::OffsetBot) {
1287 return longcon((jlong) offset_con);
1288 }
1289 return _gvn.transform( new ConvI2LNode(offset));
1290 }
1291
1292 Node* GraphKit::ConvI2UL(Node* offset) {
1293 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1294 if (offset_con != (juint) Type::OffsetBot) {
1295 return longcon((julong) offset_con);
1296 }
1297 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1298 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1299 return _gvn.transform( new AndLNode(conv, mask) );
1300 }
1301
1302 Node* GraphKit::ConvL2I(Node* offset) {
1303 // short-circuit a common case
1304 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1305 if (offset_con != (jlong)Type::OffsetBot) {
1306 return intcon((int) offset_con);
1307 }
1308 return _gvn.transform( new ConvL2INode(offset));
1309 }
1310
1311 //-------------------------load_object_klass-----------------------------------
1312 Node* GraphKit::load_object_klass(Node* obj) {
1313 // Special-case a fresh allocation to avoid building nodes:
1314 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1315 if (akls != nullptr) return akls;
1316 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1317 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1318 }
1319
1320 //-------------------------load_array_length-----------------------------------
1321 Node* GraphKit::load_array_length(Node* array) {
1322 // Special-case a fresh allocation to avoid building nodes:
1323 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1324 Node *alen;
1325 if (alloc == nullptr) {
1326 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1327 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1328 } else {
1329 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1330 }
1331 return alen;
1332 }
1333
1334 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1335 const TypeOopPtr* oop_type,
1336 bool replace_length_in_map) {
1337 Node* length = alloc->Ideal_length();
1338 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1339 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1340 if (ccast != length) {
1341 // do not transform ccast here, it might convert to top node for
1342 // negative array length and break assumptions in parsing stage.
1343 _gvn.set_type_bottom(ccast);
1344 record_for_igvn(ccast);
1345 if (replace_length_in_map) {
1346 replace_in_map(length, ccast);
1347 }
1348 return ccast;
1349 }
1350 }
1351 return length;
1352 }
1353
1354 //------------------------------do_null_check----------------------------------
1355 // Helper function to do a null pointer check. Returned value is
1356 // the incoming address with null casted away. You are allowed to use the
1357 // not-null value only if you are control dependent on the test.
1358 #ifndef PRODUCT
1359 extern uint explicit_null_checks_inserted,
1360 explicit_null_checks_elided;
1361 #endif
1362 Node* GraphKit::null_check_common(Node* value, BasicType type,
1363 // optional arguments for variations:
1364 bool assert_null,
1365 Node* *null_control,
1366 bool speculative,
1367 bool null_marker_check) {
1368 assert(!assert_null || null_control == nullptr, "not both at once");
1369 if (stopped()) return top();
1370 NOT_PRODUCT(explicit_null_checks_inserted++);
1371
1372 if (value->is_InlineType()) {
1373 // Null checking a scalarized but nullable inline type. Check the null marker
1374 // input instead of the oop input to avoid keeping buffer allocations alive.
1375 null_check_common(value->as_InlineType()->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1376 if (stopped()) {
1377 return top();
1378 }
1379 if (assert_null) {
1380 // TODO 8350865 Scalarize here (this leads to failures with TestLWorld::test45)
1381 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1382 // replace_in_map(value, vtptr);
1383 // return vtptr;
1384 replace_in_map(value, null());
1385 return null();
1386 }
1387 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1388 return cast_not_null(value, do_replace_in_map);
1389 }
1390
1391 // Construct null check
1392 Node *chk = nullptr;
1393 switch(type) {
1394 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1395 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1396 case T_ARRAY : // fall through
1397 type = T_OBJECT; // simplify further tests
1398 case T_OBJECT : {
1399 const Type *t = _gvn.type( value );
1400
1401 const TypeOopPtr* tp = t->isa_oopptr();
1402 if (tp != nullptr && !tp->is_loaded()
1403 // Only for do_null_check, not any of its siblings:
1404 && !assert_null && null_control == nullptr) {
1405 // Usually, any field access or invocation on an unloaded oop type
1406 // will simply fail to link, since the statically linked class is
1407 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1408 // the static class is loaded but the sharper oop type is not.
1409 // Rather than checking for this obscure case in lots of places,
1410 // we simply observe that a null check on an unloaded class
1411 // will always be followed by a nonsense operation, so we
1412 // can just issue the uncommon trap here.
1413 // Our access to the unloaded class will only be correct
1414 // after it has been loaded and initialized, which requires
1415 // a trip through the interpreter.
1416 ciKlass* klass = tp->unloaded_klass();
1417 #ifndef PRODUCT
1418 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1419 #endif
1420 uncommon_trap(Deoptimization::Reason_unloaded,
1421 Deoptimization::Action_reinterpret,
1422 klass, "!loaded");
1423 return top();
1424 }
1425
1426 if (assert_null) {
1427 // See if the type is contained in NULL_PTR.
1428 // If so, then the value is already null.
1429 if (t->higher_equal(TypePtr::NULL_PTR)) {
1430 NOT_PRODUCT(explicit_null_checks_elided++);
1431 return value; // Elided null assert quickly!
1432 }
1433 } else {
1434 // See if mixing in the null pointer changes type.
1435 // If so, then the null pointer was not allowed in the original
1436 // type. In other words, "value" was not-null.
1437 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1438 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1439 NOT_PRODUCT(explicit_null_checks_elided++);
1440 return value; // Elided null check quickly!
1441 }
1442 }
1443 chk = new CmpPNode( value, null() );
1444 break;
1445 }
1446
1447 default:
1448 fatal("unexpected type: %s", type2name(type));
1449 }
1450 assert(chk != nullptr, "sanity check");
1451 chk = _gvn.transform(chk);
1452
1453 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1454 BoolNode *btst = new BoolNode( chk, btest);
1455 Node *tst = _gvn.transform( btst );
1456
1457 //-----------
1458 // if peephole optimizations occurred, a prior test existed.
1459 // If a prior test existed, maybe it dominates as we can avoid this test.
1460 if (tst != btst && type == T_OBJECT) {
1461 // At this point we want to scan up the CFG to see if we can
1462 // find an identical test (and so avoid this test altogether).
1463 Node *cfg = control();
1464 int depth = 0;
1465 while( depth < 16 ) { // Limit search depth for speed
1466 if( cfg->Opcode() == Op_IfTrue &&
1467 cfg->in(0)->in(1) == tst ) {
1468 // Found prior test. Use "cast_not_null" to construct an identical
1469 // CastPP (and hence hash to) as already exists for the prior test.
1470 // Return that casted value.
1471 if (assert_null) {
1472 replace_in_map(value, null());
1473 return null(); // do not issue the redundant test
1474 }
1475 Node *oldcontrol = control();
1476 set_control(cfg);
1477 Node *res = cast_not_null(value);
1478 set_control(oldcontrol);
1479 NOT_PRODUCT(explicit_null_checks_elided++);
1480 return res;
1481 }
1482 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1483 if (cfg == nullptr) break; // Quit at region nodes
1484 depth++;
1485 }
1486 }
1487
1488 //-----------
1489 // Branch to failure if null
1490 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1491 Deoptimization::DeoptReason reason;
1492 if (assert_null) {
1493 reason = Deoptimization::reason_null_assert(speculative);
1494 } else if (type == T_OBJECT || null_marker_check) {
1495 reason = Deoptimization::reason_null_check(speculative);
1496 } else {
1497 reason = Deoptimization::Reason_div0_check;
1498 }
1499 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1500 // ciMethodData::has_trap_at will return a conservative -1 if any
1501 // must-be-null assertion has failed. This could cause performance
1502 // problems for a method after its first do_null_assert failure.
1503 // Consider using 'Reason_class_check' instead?
1504
1505 // To cause an implicit null check, we set the not-null probability
1506 // to the maximum (PROB_MAX). For an explicit check the probability
1507 // is set to a smaller value.
1508 if (null_control != nullptr || too_many_traps(reason)) {
1509 // probability is less likely
1510 ok_prob = PROB_LIKELY_MAG(3);
1511 } else if (!assert_null &&
1512 (ImplicitNullCheckThreshold > 0) &&
1513 method() != nullptr &&
1514 (method()->method_data()->trap_count(reason)
1515 >= (uint)ImplicitNullCheckThreshold)) {
1516 ok_prob = PROB_LIKELY_MAG(3);
1517 }
1518
1519 if (null_control != nullptr) {
1520 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1521 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1522 set_control( _gvn.transform( new IfTrueNode(iff)));
1523 #ifndef PRODUCT
1524 if (null_true == top()) {
1525 explicit_null_checks_elided++;
1526 }
1527 #endif
1528 (*null_control) = null_true;
1529 } else {
1530 BuildCutout unless(this, tst, ok_prob);
1531 // Check for optimizer eliding test at parse time
1532 if (stopped()) {
1533 // Failure not possible; do not bother making uncommon trap.
1534 NOT_PRODUCT(explicit_null_checks_elided++);
1535 } else if (assert_null) {
1536 uncommon_trap(reason,
1537 Deoptimization::Action_make_not_entrant,
1538 nullptr, "assert_null");
1539 } else {
1540 replace_in_map(value, zerocon(type));
1541 builtin_throw(reason);
1542 }
1543 }
1544
1545 // Must throw exception, fall-thru not possible?
1546 if (stopped()) {
1547 return top(); // No result
1548 }
1549
1550 if (assert_null) {
1551 // Cast obj to null on this path.
1552 replace_in_map(value, zerocon(type));
1553 return zerocon(type);
1554 }
1555
1556 // Cast obj to not-null on this path, if there is no null_control.
1557 // (If there is a null_control, a non-null value may come back to haunt us.)
1558 if (type == T_OBJECT) {
1559 Node* cast = cast_not_null(value, false);
1560 if (null_control == nullptr || (*null_control) == top())
1561 replace_in_map(value, cast);
1562 value = cast;
1563 }
1564
1565 return value;
1566 }
1567
1568 //------------------------------cast_not_null----------------------------------
1569 // Cast obj to not-null on this path
1570 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1571 const Type* t = _gvn.type(obj);
1572 const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL);
1573 if (t == t_not_null) {
1574 return obj;
1575 }
1576
1577 Node* cast = new CastPPNode(control(), obj, t_not_null);
1578 cast = _gvn.transform(cast);
1579
1580 // Scan for instances of 'obj' in the current JVM mapping.
1581 // These instances are known to be not-null after the test.
1582 if (do_replace_in_map) {
1583 replace_in_map(obj, cast);
1584 }
1585 return cast;
1586 }
1587
1588 // Sometimes in intrinsics, we implicitly know an object is not null
1589 // (there's no actual null check) so we can cast it to not null. In
1590 // the course of optimizations, the input to the cast can become null.
1591 // In that case that data path will die and we need the control path
1592 // to become dead as well to keep the graph consistent. So we have to
1593 // add a check for null for which one branch can't be taken. It uses
1594 // an OpaqueConstantBool node that will cause the check to be removed after loop
1595 // opts so the test goes away and the compiled code doesn't execute a
1596 // useless check.
1597 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1598 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1599 return value;
1600 }
1601 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1602 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1603 Node* opaq = _gvn.transform(new OpaqueConstantBoolNode(C, tst, true));
1604 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1605 _gvn.set_type(iff, iff->Value(&_gvn));
1606 if (!tst->is_Con()) {
1607 record_for_igvn(iff);
1608 }
1609 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1610 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1611 halt(if_f, frame, "unexpected null in intrinsic");
1612 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1613 set_control(if_t);
1614 return cast_not_null(value, do_replace_in_map);
1615 }
1616
1617
1618 //--------------------------replace_in_map-------------------------------------
1619 void GraphKit::replace_in_map(Node* old, Node* neww) {
1620 if (old == neww) {
1621 return;
1622 }
1623
1624 map()->replace_edge(old, neww);
1625
1626 // Note: This operation potentially replaces any edge
1627 // on the map. This includes locals, stack, and monitors
1628 // of the current (innermost) JVM state.
1629
1630 // don't let inconsistent types from profiling escape this
1631 // method
1632
1633 const Type* told = _gvn.type(old);
1634 const Type* tnew = _gvn.type(neww);
1635
1636 if (!tnew->higher_equal(told)) {
1637 return;
1638 }
1639
1640 map()->record_replaced_node(old, neww);
1641 }
1642
1643
1644 //=============================================================================
1645 //--------------------------------memory---------------------------------------
1646 Node* GraphKit::memory(uint alias_idx) {
1647 MergeMemNode* mem = merged_memory();
1648 Node* p = mem->memory_at(alias_idx);
1649 assert(p != mem->empty_memory(), "empty");
1650 _gvn.set_type(p, Type::MEMORY); // must be mapped
1651 return p;
1652 }
1653
1654 //-----------------------------reset_memory------------------------------------
1655 Node* GraphKit::reset_memory() {
1656 Node* mem = map()->memory();
1657 // do not use this node for any more parsing!
1658 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1659 return _gvn.transform( mem );
1660 }
1661
1662 //------------------------------set_all_memory---------------------------------
1663 void GraphKit::set_all_memory(Node* newmem) {
1664 // The 2 cases are semantically equivalent
1665 MergeMemNode* mergemem;
1666 if (_gvn.is_IterGVN()) {
1667 // During IGVN, create a more predictable pattern so it is easier to verify that the GraphKit
1668 // does not modify memory
1669 mergemem = MergeMemNode::make(C->top());
1670 mergemem->set_base_memory(newmem);
1671 } else {
1672 // During parsing, be a little more aggressive so that GVN can fold accesses more easily
1673 mergemem = MergeMemNode::make(newmem);
1674 }
1675 _gvn.set_type_bottom(mergemem);
1676 record_for_igvn(mergemem);
1677 map()->set_memory(mergemem);
1678 }
1679
1680 //------------------------------set_all_memory_call----------------------------
1681 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1682 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1683 set_all_memory(newmem);
1684 }
1685
1686 //=============================================================================
1687 //
1688 // parser factory methods for MemNodes
1689 //
1690 // These are layered on top of the factory methods in LoadNode and StoreNode,
1691 // and integrate with the parser's memory state and _gvn engine.
1692 //
1693
1694 // factory methods in "int adr_idx"
1695 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1696 MemNode::MemOrd mo,
1697 LoadNode::ControlDependency control_dependency,
1698 bool require_atomic_access,
1699 bool unaligned,
1700 bool mismatched,
1701 bool unsafe,
1702 uint8_t barrier_data) {
1703 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1704 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1705 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1706 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1707 Node* mem = memory(adr_idx);
1708 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1709 ld = _gvn.transform(ld);
1710
1711 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1712 // Improve graph before escape analysis and boxing elimination.
1713 record_for_igvn(ld);
1714 if (ld->is_DecodeN()) {
1715 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1716 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1717 // a Phi). Recording such cases is still perfectly sound, but may be
1718 // unnecessary and result in some minor IGVN overhead.
1719 record_for_igvn(ld->in(1));
1720 }
1721 }
1722 return ld;
1723 }
1724
1725 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1726 MemNode::MemOrd mo,
1727 bool require_atomic_access,
1728 bool unaligned,
1729 bool mismatched,
1730 bool unsafe,
1731 int barrier_data) {
1732 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1733 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1734 const TypePtr* adr_type = nullptr;
1735 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1736 Node *mem = memory(adr_idx);
1737 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1738 if (unaligned) {
1739 st->as_Store()->set_unaligned_access();
1740 }
1741 if (mismatched) {
1742 st->as_Store()->set_mismatched_access();
1743 }
1744 if (unsafe) {
1745 st->as_Store()->set_unsafe_access();
1746 }
1747 st->as_Store()->set_barrier_data(barrier_data);
1748 st = _gvn.transform(st);
1749 set_memory(st, adr_idx);
1750 // Back-to-back stores can only remove intermediate store with DU info
1751 // so push on worklist for optimizer.
1752 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1753 record_for_igvn(st);
1754
1755 return st;
1756 }
1757
1758 Node* GraphKit::access_store_at(Node* obj,
1759 Node* adr,
1760 const TypePtr* adr_type,
1761 Node* val,
1762 const Type* val_type,
1763 BasicType bt,
1764 DecoratorSet decorators,
1765 bool safe_for_replace,
1766 const InlineTypeNode* vt) {
1767 // Transformation of a value which could be null pointer (CastPP #null)
1768 // could be delayed during Parse (for example, in adjust_map_after_if()).
1769 // Execute transformation here to avoid barrier generation in such case.
1770 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1771 val = _gvn.makecon(TypePtr::NULL_PTR);
1772 }
1773
1774 if (stopped()) {
1775 return top(); // Dead path ?
1776 }
1777
1778 assert(val != nullptr, "not dead path");
1779 if (val->is_InlineType()) {
1780 // Store to non-flat field. Buffer the inline type and make sure
1781 // the store is re-executed if the allocation triggers deoptimization.
1782 PreserveReexecuteState preexecs(this);
1783 jvms()->set_should_reexecute(true);
1784 val = val->as_InlineType()->buffer(this, safe_for_replace);
1785 }
1786
1787 C2AccessValuePtr addr(adr, adr_type);
1788 C2AccessValue value(val, val_type);
1789 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1790 if (access.is_raw()) {
1791 return _barrier_set->BarrierSetC2::store_at(access, value);
1792 } else {
1793 return _barrier_set->store_at(access, value);
1794 }
1795 }
1796
1797 Node* GraphKit::access_load_at(Node* obj, // containing obj
1798 Node* adr, // actual address to store val at
1799 const TypePtr* adr_type,
1800 const Type* val_type,
1801 BasicType bt,
1802 DecoratorSet decorators,
1803 Node* ctl) {
1804 if (stopped()) {
1805 return top(); // Dead path ?
1806 }
1807
1808 SavedState old_state(this);
1809 C2AccessValuePtr addr(adr, adr_type);
1810 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1811 Node* load;
1812 if (access.is_raw()) {
1813 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1814 } else {
1815 load = _barrier_set->load_at(access, val_type);
1816 }
1817
1818 // Restore the previous state only if the load got folded to a constant
1819 // and we can discard any barriers that might have been added.
1820 if (load == nullptr || !load->is_Con()) {
1821 old_state.discard();
1822 }
1823 return load;
1824 }
1825
1826 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1827 const Type* val_type,
1828 BasicType bt,
1829 DecoratorSet decorators) {
1830 if (stopped()) {
1831 return top(); // Dead path ?
1832 }
1833
1834 SavedState old_state(this);
1835 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1836 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1837 Node* load;
1838 if (access.is_raw()) {
1839 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1840 } else {
1841 load = _barrier_set->load_at(access, val_type);
1842 }
1843
1844 // Restore the previous state only if the load got folded to a constant
1845 // and we can discard any barriers that might have been added.
1846 if (load == nullptr || !load->is_Con()) {
1847 old_state.discard();
1848 }
1849 return load;
1850 }
1851
1852 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1853 Node* adr,
1854 const TypePtr* adr_type,
1855 int alias_idx,
1856 Node* expected_val,
1857 Node* new_val,
1858 const Type* value_type,
1859 BasicType bt,
1860 DecoratorSet decorators) {
1861 C2AccessValuePtr addr(adr, adr_type);
1862 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1863 bt, obj, addr, alias_idx);
1864 if (access.is_raw()) {
1865 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1866 } else {
1867 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1868 }
1869 }
1870
1871 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1872 Node* adr,
1873 const TypePtr* adr_type,
1874 int alias_idx,
1875 Node* expected_val,
1876 Node* new_val,
1877 const Type* value_type,
1878 BasicType bt,
1879 DecoratorSet decorators) {
1880 C2AccessValuePtr addr(adr, adr_type);
1881 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1882 bt, obj, addr, alias_idx);
1883 if (access.is_raw()) {
1884 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1885 } else {
1886 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1887 }
1888 }
1889
1890 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1891 Node* adr,
1892 const TypePtr* adr_type,
1893 int alias_idx,
1894 Node* new_val,
1895 const Type* value_type,
1896 BasicType bt,
1897 DecoratorSet decorators) {
1898 C2AccessValuePtr addr(adr, adr_type);
1899 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1900 bt, obj, addr, alias_idx);
1901 if (access.is_raw()) {
1902 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1903 } else {
1904 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1905 }
1906 }
1907
1908 Node* GraphKit::access_atomic_add_at(Node* obj,
1909 Node* adr,
1910 const TypePtr* adr_type,
1911 int alias_idx,
1912 Node* new_val,
1913 const Type* value_type,
1914 BasicType bt,
1915 DecoratorSet decorators) {
1916 C2AccessValuePtr addr(adr, adr_type);
1917 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1918 if (access.is_raw()) {
1919 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1920 } else {
1921 return _barrier_set->atomic_add_at(access, new_val, value_type);
1922 }
1923 }
1924
1925 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1926 return _barrier_set->clone(this, src, dst, size, is_array);
1927 }
1928
1929 //-------------------------array_element_address-------------------------
1930 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1931 const TypeInt* sizetype, Node* ctrl) {
1932 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1933 uint shift;
1934 uint header;
1935 if (arytype->is_flat() && arytype->klass_is_exact()) {
1936 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1937 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1938 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1939 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1940 // though we don't need the address node in this case and throw it away again.
1941 shift = arytype->flat_log_elem_size();
1942 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1943 } else {
1944 shift = exact_log2(type2aelembytes(elembt));
1945 header = arrayOopDesc::base_offset_in_bytes(elembt);
1946 }
1947
1948 // short-circuit a common case (saves lots of confusing waste motion)
1949 jint idx_con = find_int_con(idx, -1);
1950 if (idx_con >= 0) {
1951 intptr_t offset = header + ((intptr_t)idx_con << shift);
1952 return basic_plus_adr(ary, offset);
1953 }
1954
1955 // must be correct type for alignment purposes
1956 Node* base = basic_plus_adr(ary, header);
1957 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1958 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1959 return basic_plus_adr(ary, base, scale);
1960 }
1961
1962 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1963 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1964 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1965 return cast_to_flat_array_exact(array, elem_vk, true, false);
1966 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1967 return cast_to_flat_array_exact(array, elem_vk, true, true);
1968 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1969 return cast_to_flat_array_exact(array, elem_vk, false, true);
1970 }
1971
1972 bool is_null_free = false;
1973 if (!elem_vk->has_nullable_atomic_layout()) {
1974 // Element does not have a nullable flat layout, cannot be nullable
1975 is_null_free = true;
1976 }
1977
1978 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1979 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1980 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1981 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1982 }
1983
1984 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1985 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1986 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1987 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1988 assert(arytype->klass_is_exact(), "inconsistency");
1989 assert(arytype->is_flat(), "inconsistency");
1990 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1991 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1992 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1993 }
1994
1995 //-------------------------load_array_element-------------------------
1996 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1997 const Type* elemtype = arytype->elem();
1998 BasicType elembt = elemtype->array_element_basic_type();
1999 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
2000 if (elembt == T_NARROWOOP) {
2001 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
2002 }
2003 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2004 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2005 return ld;
2006 }
2007
2008 //-------------------------set_arguments_for_java_call-------------------------
2009 // Arguments (pre-popped from the stack) are taken from the JVMS.
2010 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2011 PreserveReexecuteState preexecs(this);
2012 if (Arguments::is_valhalla_enabled()) {
2013 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2014 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2015 jvms()->set_should_reexecute(true);
2016 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2017 inc_sp(arg_size);
2018 }
2019 // Add the call arguments
2020 const TypeTuple* domain = call->tf()->domain_sig();
2021 uint nargs = domain->cnt();
2022 int arg_num = 0;
2023 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2024 uint arg_idx = i - TypeFunc::Parms;
2025 Node* arg = argument(arg_idx);
2026 const Type* t = domain->field_at(i);
2027 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2028 // We don't pass inline type arguments by reference but instead pass each field of the inline type
2029 if (!arg->is_InlineType()) {
2030 // There are 2 cases in which the argument has not been scalarized
2031 if (_gvn.type(arg)->is_zero_type()) {
2032 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2033 } else {
2034 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2035 // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2036 // type must be an exact value class. The method is devirtualized, and replaced with
2037 // a direct call with a scalarized receiver instead.
2038 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2039 assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2040 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2041 }
2042 }
2043 InlineTypeNode* vt = arg->as_InlineType();
2044 vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2045 // If an inline type argument is passed as fields, attach the Method* to the call site
2046 // to be able to access the extended signature later via attached_method_before_pc().
2047 // For example, see CompiledMethod::preserve_callee_argument_oops().
2048 call->set_override_symbolic_info(true);
2049 // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2050 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2051 C->dependencies()->assert_mismatch_calling_convention(call->method());
2052 arg_num++;
2053 continue;
2054 } else if (arg->is_InlineType()) {
2055 // Pass inline type argument via oop to callee
2056 arg = arg->as_InlineType()->buffer(this, true);
2057 }
2058 if (t != Type::HALF) {
2059 arg_num++;
2060 }
2061 call->init_req(idx++, arg);
2062 }
2063 }
2064
2065 //---------------------------set_edges_for_java_call---------------------------
2066 // Connect a newly created call into the current JVMS.
2067 // A return value node (if any) is returned from set_edges_for_java_call.
2068 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2069
2070 // Add the predefined inputs:
2071 call->init_req( TypeFunc::Control, control() );
2072 call->init_req( TypeFunc::I_O , i_o() );
2073 call->init_req( TypeFunc::Memory , reset_memory() );
2074 call->init_req( TypeFunc::FramePtr, frameptr() );
2075 call->init_req( TypeFunc::ReturnAdr, top() );
2076
2077 add_safepoint_edges(call, must_throw);
2078
2079 Node* xcall = _gvn.transform(call);
2080
2081 if (xcall == top()) {
2082 set_control(top());
2083 return;
2084 }
2085 assert(xcall == call, "call identity is stable");
2086
2087 // Re-use the current map to produce the result.
2088
2089 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2090 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2091 set_all_memory_call(xcall, separate_io_proj);
2092
2093 //return xcall; // no need, caller already has it
2094 }
2095
2096 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2097 if (stopped()) return top(); // maybe the call folded up?
2098
2099 // Note: Since any out-of-line call can produce an exception,
2100 // we always insert an I_O projection from the call into the result.
2101
2102 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2103
2104 if (separate_io_proj) {
2105 // The caller requested separate projections be used by the fall
2106 // through and exceptional paths, so replace the projections for
2107 // the fall through path.
2108 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2109 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2110 }
2111
2112 // Capture the return value, if any.
2113 Node* ret;
2114 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2115 ret = top();
2116 } else if (call->tf()->returns_inline_type_as_fields()) {
2117 // Return of multiple values (inline type fields): we create a
2118 // InlineType node, each field is a projection from the call.
2119 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2120 uint base_input = TypeFunc::Parms;
2121 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2122 // If we run out of registers to store the null marker, we need to reserve an extra
2123 // slot to store it on the stack. Unfortunately, we only know if stack slot is needed
2124 // when matching the call (see Matcher::return_values_mask), so we are conservative here.
2125 C->set_needs_nm_slot(true);
2126 } else {
2127 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2128 ciType* t = call->method()->return_type();
2129 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2130 // The return type is unloaded but the callee might later be C2 compiled and then return
2131 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2132 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2133 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2134 IdealKit ideal(this);
2135 IdealVariable res(ideal);
2136 ideal.declarations_done();
2137 // Change return type of call to scalarized return
2138 const TypeFunc* tf = call->_tf;
2139 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2140 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2141 call->_tf = new_tf;
2142 _gvn.set_type(call, call->Value(&_gvn));
2143 _gvn.set_type(ret, ret->Value(&_gvn));
2144 // Don't add store to buffer call if we are strength reducing
2145 if (!C->strength_reduction()) {
2146 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2147 // Return value is null
2148 ideal.set(res, makecon(TypePtr::NULL_PTR));
2149 } ideal.else_(); {
2150 // Return value is non-null
2151 sync_kit(ideal);
2152
2153 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2154 OptoRuntime::store_inline_type_fields_Type(),
2155 StubRoutines::store_inline_type_fields_to_buf(),
2156 nullptr, TypePtr::BOTTOM, ret);
2157
2158 // We don't know how many values are returned. This assumes the
2159 // worst case, that all available registers are used.
2160 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2161 if (domain->field_at(i) == Type::HALF) {
2162 store_to_buf_call->init_req(i, top());
2163 continue;
2164 }
2165 Node* proj =_gvn.transform(new ProjNode(call, i));
2166 store_to_buf_call->init_req(i, proj);
2167 }
2168 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2169
2170 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2171 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2172 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2173
2174 ideal.set(res, buf);
2175 ideal.sync_kit(this);
2176 } ideal.end_if();
2177 } else {
2178 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2179 // Will be rewired later in replace_call().
2180 _gvn.transform(new ProjNode(call, i));
2181 }
2182 ideal.set(res, ret);
2183 }
2184 sync_kit(ideal);
2185 ret = _gvn.transform(ideal.value(res));
2186 } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2187 // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2188 // result of merging several call paths. If one of them is made of an actual call node that
2189 // returns an oop, we need to call make_from_oop here as well because we want InlineType
2190 // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2191 ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2192 }
2193 }
2194
2195 return ret;
2196 }
2197
2198 //--------------------set_predefined_input_for_runtime_call--------------------
2199 // Reading and setting the memory state is way conservative here.
2200 // The real problem is that I am not doing real Type analysis on memory,
2201 // so I cannot distinguish card mark stores from other stores. Across a GC
2202 // point the Store Barrier and the card mark memory has to agree. I cannot
2203 // have a card mark store and its barrier split across the GC point from
2204 // either above or below. Here I get that to happen by reading ALL of memory.
2205 // A better answer would be to separate out card marks from other memory.
2206 // For now, return the input memory state, so that it can be reused
2207 // after the call, if this call has restricted memory effects.
2208 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2209 // Set fixed predefined input arguments
2210 call->init_req(TypeFunc::Control, control());
2211 call->init_req(TypeFunc::I_O, top()); // does no i/o
2212 call->init_req(TypeFunc::ReturnAdr, top());
2213 if (call->is_CallLeafPure()) {
2214 call->init_req(TypeFunc::Memory, top());
2215 call->init_req(TypeFunc::FramePtr, top());
2216 return nullptr;
2217 } else {
2218 Node* memory = reset_memory();
2219 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2220 call->init_req(TypeFunc::Memory, m); // may gc ptrs
2221 call->init_req(TypeFunc::FramePtr, frameptr());
2222 return memory;
2223 }
2224 }
2225
2226 //-------------------set_predefined_output_for_runtime_call--------------------
2227 // Set control and memory (not i_o) from the call.
2228 // If keep_mem is not null, use it for the output state,
2229 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2230 // If hook_mem is null, this call produces no memory effects at all.
2231 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2232 // then only that memory slice is taken from the call.
2233 // In the last case, we must put an appropriate memory barrier before
2234 // the call, so as to create the correct anti-dependencies on loads
2235 // preceding the call.
2236 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2237 Node* keep_mem,
2238 const TypePtr* hook_mem) {
2239 // no i/o
2240 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2241 if (call->is_CallLeafPure()) {
2242 // Pure function have only control (for now) and data output, in particular
2243 // they don't touch the memory, so we don't want a memory proj that is set after.
2244 return;
2245 }
2246 if (keep_mem) {
2247 // First clone the existing memory state
2248 set_all_memory(keep_mem);
2249 if (hook_mem != nullptr) {
2250 // Make memory for the call
2251 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2252 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2253 // We also use hook_mem to extract specific effects from arraycopy stubs.
2254 set_memory(mem, hook_mem);
2255 }
2256 // ...else the call has NO memory effects.
2257
2258 // Make sure the call advertises its memory effects precisely.
2259 // This lets us build accurate anti-dependences in gcm.cpp.
2260 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2261 "call node must be constructed correctly");
2262 } else {
2263 assert(hook_mem == nullptr, "");
2264 // This is not a "slow path" call; all memory comes from the call.
2265 set_all_memory_call(call);
2266 }
2267 }
2268
2269 // Keep track of MergeMems feeding into other MergeMems
2270 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2271 if (!mem->is_MergeMem()) {
2272 return;
2273 }
2274 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2275 Node* use = i.get();
2276 if (use->is_MergeMem()) {
2277 wl.push(use);
2278 }
2279 }
2280 }
2281
2282 // Replace the call with the current state of the kit.
2283 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2284 JVMState* ejvms = nullptr;
2285 if (has_exceptions()) {
2286 ejvms = transfer_exceptions_into_jvms();
2287 }
2288
2289 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2290 ReplacedNodes replaced_nodes_exception;
2291 Node* ex_ctl = top();
2292
2293 SafePointNode* final_state = stop();
2294
2295 // Find all the needed outputs of this call
2296 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2297
2298 Unique_Node_List wl;
2299 Node* final_mem = final_state->in(TypeFunc::Memory);
2300 Node* final_ctl = final_state->in(TypeFunc::Control);
2301 Node* final_io = final_state->in(TypeFunc::I_O);
2302
2303 // Replace all the old call edges with the edges from the inlining result
2304 if (callprojs->fallthrough_catchproj != nullptr) {
2305 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2306 }
2307 if (callprojs->fallthrough_memproj != nullptr) {
2308 if (final_mem->is_MergeMem()) {
2309 // Parser's exits MergeMem was not transformed but may be optimized
2310 final_mem = _gvn.transform(final_mem);
2311 }
2312 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2313 add_mergemem_users_to_worklist(wl, final_mem);
2314 }
2315 if (callprojs->fallthrough_ioproj != nullptr) {
2316 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2317 }
2318
2319 // Replace the result with the new result if it exists and is used
2320 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2321 // If the inlined code is dead, the result projections for an inline type returned as
2322 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2323 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2324 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2325 "unexpected number of results");
2326 // If we are doing strength reduction and the return type is not loaded we
2327 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2328 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2329 CallNode* new_call = result->in(0)->as_Call();
2330 assert(new_call->proj_out_or_null(TypeFunc::Parms) == result, "the first data projection should be result");
2331 for (uint i = 0; i < callprojs->nb_resproj; i++) {
2332 if (callprojs->resproj[i] != nullptr) {
2333 Node* new_proj = new_call->proj_out_or_null(TypeFunc::Parms + i);
2334 assert(new_proj != nullptr, "projection should be available");
2335 C->gvn_replace_by(callprojs->resproj[i], new_proj);
2336 }
2337 }
2338 } else {
2339 C->gvn_replace_by(callprojs->resproj[0], result);
2340 }
2341 }
2342
2343 if (ejvms == nullptr) {
2344 // No exception edges to simply kill off those paths
2345 if (callprojs->catchall_catchproj != nullptr) {
2346 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2347 }
2348 if (callprojs->catchall_memproj != nullptr) {
2349 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2350 }
2351 if (callprojs->catchall_ioproj != nullptr) {
2352 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2353 }
2354 // Replace the old exception object with top
2355 if (callprojs->exobj != nullptr) {
2356 C->gvn_replace_by(callprojs->exobj, C->top());
2357 }
2358 } else {
2359 GraphKit ekit(ejvms);
2360
2361 // Load my combined exception state into the kit, with all phis transformed:
2362 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2363 replaced_nodes_exception = ex_map->replaced_nodes();
2364
2365 Node* ex_oop = ekit.use_exception_state(ex_map);
2366
2367 if (callprojs->catchall_catchproj != nullptr) {
2368 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2369 ex_ctl = ekit.control();
2370 }
2371 if (callprojs->catchall_memproj != nullptr) {
2372 Node* ex_mem = ekit.reset_memory();
2373 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2374 add_mergemem_users_to_worklist(wl, ex_mem);
2375 }
2376 if (callprojs->catchall_ioproj != nullptr) {
2377 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2378 }
2379
2380 // Replace the old exception object with the newly created one
2381 if (callprojs->exobj != nullptr) {
2382 C->gvn_replace_by(callprojs->exobj, ex_oop);
2383 }
2384 }
2385
2386 // Disconnect the call from the graph
2387 call->disconnect_inputs(C);
2388 C->gvn_replace_by(call, C->top());
2389
2390 // Clean up any MergeMems that feed other MergeMems since the
2391 // optimizer doesn't like that.
2392 while (wl.size() > 0) {
2393 _gvn.transform(wl.pop());
2394 }
2395
2396 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2397 replaced_nodes.apply(C, final_ctl);
2398 }
2399 if (!ex_ctl->is_top() && do_replaced_nodes) {
2400 replaced_nodes_exception.apply(C, ex_ctl);
2401 }
2402 }
2403
2404
2405 //------------------------------increment_counter------------------------------
2406 // for statistics: increment a VM counter by 1
2407
2408 void GraphKit::increment_counter(address counter_addr) {
2409 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2410 increment_counter(adr1);
2411 }
2412
2413 void GraphKit::increment_counter(Node* counter_addr) {
2414 Node* ctrl = control();
2415 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2416 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2417 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2418 }
2419
2420 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2421 Node* halt = new HaltNode(ctrl, frameptr, reason
2422 PRODUCT_ONLY(COMMA generate_code_in_product));
2423 halt = _gvn.transform(halt);
2424 root()->add_req(halt);
2425 if (_gvn.is_IterGVN() != nullptr) {
2426 record_for_igvn(root());
2427 }
2428 }
2429
2430 //------------------------------uncommon_trap----------------------------------
2431 // Bail out to the interpreter in mid-method. Implemented by calling the
2432 // uncommon_trap blob. This helper function inserts a runtime call with the
2433 // right debug info.
2434 Node* GraphKit::uncommon_trap(int trap_request,
2435 ciKlass* klass, const char* comment,
2436 bool must_throw,
2437 bool keep_exact_action) {
2438 if (failing_internal()) {
2439 stop();
2440 }
2441 if (stopped()) return nullptr; // trap reachable?
2442
2443 // Note: If ProfileTraps is true, and if a deopt. actually
2444 // occurs here, the runtime will make sure an MDO exists. There is
2445 // no need to call method()->ensure_method_data() at this point.
2446
2447 // Set the stack pointer to the right value for reexecution:
2448 set_sp(reexecute_sp());
2449
2450 #ifdef ASSERT
2451 if (!must_throw) {
2452 // Make sure the stack has at least enough depth to execute
2453 // the current bytecode.
2454 int inputs, ignored_depth;
2455 if (compute_stack_effects(inputs, ignored_depth)) {
2456 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2457 Bytecodes::name(java_bc()), sp(), inputs);
2458 }
2459 }
2460 #endif
2461
2462 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2463 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2464
2465 switch (action) {
2466 case Deoptimization::Action_maybe_recompile:
2467 case Deoptimization::Action_reinterpret:
2468 // Temporary fix for 6529811 to allow virtual calls to be sure they
2469 // get the chance to go from mono->bi->mega
2470 if (!keep_exact_action &&
2471 Deoptimization::trap_request_index(trap_request) < 0 &&
2472 too_many_recompiles(reason)) {
2473 // This BCI is causing too many recompilations.
2474 if (C->log() != nullptr) {
2475 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2476 Deoptimization::trap_reason_name(reason),
2477 Deoptimization::trap_action_name(action));
2478 }
2479 action = Deoptimization::Action_none;
2480 trap_request = Deoptimization::make_trap_request(reason, action);
2481 } else {
2482 C->set_trap_can_recompile(true);
2483 }
2484 break;
2485 case Deoptimization::Action_make_not_entrant:
2486 C->set_trap_can_recompile(true);
2487 break;
2488 case Deoptimization::Action_none:
2489 case Deoptimization::Action_make_not_compilable:
2490 break;
2491 default:
2492 #ifdef ASSERT
2493 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2494 #endif
2495 break;
2496 }
2497
2498 if (TraceOptoParse) {
2499 char buf[100];
2500 tty->print_cr("Uncommon trap %s at bci:%d",
2501 Deoptimization::format_trap_request(buf, sizeof(buf),
2502 trap_request), bci());
2503 }
2504
2505 CompileLog* log = C->log();
2506 if (log != nullptr) {
2507 int kid = (klass == nullptr)? -1: log->identify(klass);
2508 log->begin_elem("uncommon_trap bci='%d'", bci());
2509 char buf[100];
2510 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2511 trap_request));
2512 if (kid >= 0) log->print(" klass='%d'", kid);
2513 if (comment != nullptr) log->print(" comment='%s'", comment);
2514 log->end_elem();
2515 }
2516
2517 // Make sure any guarding test views this path as very unlikely
2518 Node *i0 = control()->in(0);
2519 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2520 IfNode *iff = i0->as_If();
2521 float f = iff->_prob; // Get prob
2522 if (control()->Opcode() == Op_IfTrue) {
2523 if (f > PROB_UNLIKELY_MAG(4))
2524 iff->_prob = PROB_MIN;
2525 } else {
2526 if (f < PROB_LIKELY_MAG(4))
2527 iff->_prob = PROB_MAX;
2528 }
2529 }
2530
2531 // Clear out dead values from the debug info.
2532 kill_dead_locals();
2533
2534 // Now insert the uncommon trap subroutine call
2535 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2536 const TypePtr* no_memory_effects = nullptr;
2537 // Pass the index of the class to be loaded
2538 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2539 (must_throw ? RC_MUST_THROW : 0),
2540 OptoRuntime::uncommon_trap_Type(),
2541 call_addr, "uncommon_trap", no_memory_effects,
2542 intcon(trap_request));
2543 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2544 "must extract request correctly from the graph");
2545 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2546
2547 call->set_req(TypeFunc::ReturnAdr, returnadr());
2548 // The debug info is the only real input to this call.
2549
2550 // Halt-and-catch fire here. The above call should never return!
2551 // We only emit code for the HaltNode in debug, which is enough for
2552 // verifying correctness. In product, we don't want to emit it so
2553 // that we can save on code space. HaltNode often get folded because
2554 // the compiler can prove that the unreachable path is dead. But we
2555 // cannot generally expect that for uncommon traps, which are often
2556 // reachable and occasionally taken.
2557 halt(control(), frameptr(),
2558 "uncommon trap returned which should never happen",
2559 false /* don't emit code in product */);
2560 stop_and_kill_map();
2561 return call;
2562 }
2563
2564
2565 //--------------------------just_allocated_object------------------------------
2566 // Report the object that was just allocated.
2567 // It must be the case that there are no intervening safepoints.
2568 // We use this to determine if an object is so "fresh" that
2569 // it does not require card marks.
2570 Node* GraphKit::just_allocated_object(Node* current_control) {
2571 Node* ctrl = current_control;
2572 // Object::<init> is invoked after allocation, most of invoke nodes
2573 // will be reduced, but a region node is kept in parse time, we check
2574 // the pattern and skip the region node if it degraded to a copy.
2575 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2576 ctrl->as_Region()->is_copy()) {
2577 ctrl = ctrl->as_Region()->is_copy();
2578 }
2579 if (C->recent_alloc_ctl() == ctrl) {
2580 return C->recent_alloc_obj();
2581 }
2582 return nullptr;
2583 }
2584
2585
2586 /**
2587 * Record profiling data exact_kls for Node n with the type system so
2588 * that it can propagate it (speculation)
2589 *
2590 * @param n node that the type applies to
2591 * @param exact_kls type from profiling
2592 * @param maybe_null did profiling see null?
2593 *
2594 * @return node with improved type
2595 */
2596 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2597 const Type* current_type = _gvn.type(n);
2598 assert(UseTypeSpeculation, "type speculation must be on");
2599
2600 const TypePtr* speculative = current_type->speculative();
2601
2602 // Should the klass from the profile be recorded in the speculative type?
2603 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2604 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2605 const TypeOopPtr* xtype = tklass->as_instance_type();
2606 assert(xtype->klass_is_exact(), "Should be exact");
2607 // Any reason to believe n is not null (from this profiling or a previous one)?
2608 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2609 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2610 // record the new speculative type's depth
2611 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2612 speculative = speculative->with_inline_depth(jvms()->depth());
2613 } else if (current_type->would_improve_ptr(ptr_kind)) {
2614 // Profiling report that null was never seen so we can change the
2615 // speculative type to non null ptr.
2616 if (ptr_kind == ProfileAlwaysNull) {
2617 speculative = TypePtr::NULL_PTR;
2618 } else {
2619 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2620 const TypePtr* ptr = TypePtr::NOTNULL;
2621 if (speculative != nullptr) {
2622 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2623 } else {
2624 speculative = ptr;
2625 }
2626 }
2627 }
2628
2629 if (speculative != current_type->speculative()) {
2630 // Build a type with a speculative type (what we think we know
2631 // about the type but will need a guard when we use it)
2632 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2633 // We're changing the type, we need a new CheckCast node to carry
2634 // the new type. The new type depends on the control: what
2635 // profiling tells us is only valid from here as far as we can
2636 // tell.
2637 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2638 cast = _gvn.transform(cast);
2639 replace_in_map(n, cast);
2640 n = cast;
2641 }
2642
2643 return n;
2644 }
2645
2646 /**
2647 * Record profiling data from receiver profiling at an invoke with the
2648 * type system so that it can propagate it (speculation)
2649 *
2650 * @param n receiver node
2651 *
2652 * @return node with improved type
2653 */
2654 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2655 if (!UseTypeSpeculation) {
2656 return n;
2657 }
2658 ciKlass* exact_kls = profile_has_unique_klass();
2659 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2660 if ((java_bc() == Bytecodes::_checkcast ||
2661 java_bc() == Bytecodes::_instanceof ||
2662 java_bc() == Bytecodes::_aastore) &&
2663 method()->method_data()->is_mature()) {
2664 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2665 if (data != nullptr) {
2666 if (java_bc() == Bytecodes::_aastore) {
2667 ciKlass* array_type = nullptr;
2668 ciKlass* element_type = nullptr;
2669 ProfilePtrKind element_ptr = ProfileMaybeNull;
2670 bool flat_array = true;
2671 bool null_free_array = true;
2672 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2673 exact_kls = element_type;
2674 ptr_kind = element_ptr;
2675 } else {
2676 if (!data->as_BitData()->null_seen()) {
2677 ptr_kind = ProfileNeverNull;
2678 } else {
2679 if (TypeProfileCasts) {
2680 assert(data->is_ReceiverTypeData(), "bad profile data type");
2681 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2682 uint i = 0;
2683 for (; i < call->row_limit(); i++) {
2684 ciKlass* receiver = call->receiver(i);
2685 if (receiver != nullptr) {
2686 break;
2687 }
2688 }
2689 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2690 }
2691 }
2692 }
2693 }
2694 }
2695 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2696 }
2697
2698 /**
2699 * Record profiling data from argument profiling at an invoke with the
2700 * type system so that it can propagate it (speculation)
2701 *
2702 * @param dest_method target method for the call
2703 * @param bc what invoke bytecode is this?
2704 */
2705 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2706 if (!UseTypeSpeculation) {
2707 return;
2708 }
2709 const TypeFunc* tf = TypeFunc::make(dest_method);
2710 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2711 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2712 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2713 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2714 if (is_reference_type(targ->basic_type())) {
2715 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2716 ciKlass* better_type = nullptr;
2717 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2718 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2719 }
2720 i++;
2721 }
2722 }
2723 }
2724
2725 /**
2726 * Record profiling data from parameter profiling at an invoke with
2727 * the type system so that it can propagate it (speculation)
2728 */
2729 void GraphKit::record_profiled_parameters_for_speculation() {
2730 if (!UseTypeSpeculation) {
2731 return;
2732 }
2733 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2734 if (_gvn.type(local(i))->isa_oopptr()) {
2735 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2736 ciKlass* better_type = nullptr;
2737 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2738 record_profile_for_speculation(local(i), better_type, ptr_kind);
2739 }
2740 j++;
2741 }
2742 }
2743 }
2744
2745 /**
2746 * Record profiling data from return value profiling at an invoke with
2747 * the type system so that it can propagate it (speculation)
2748 */
2749 void GraphKit::record_profiled_return_for_speculation() {
2750 if (!UseTypeSpeculation) {
2751 return;
2752 }
2753 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2754 ciKlass* better_type = nullptr;
2755 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2756 // If profiling reports a single type for the return value,
2757 // feed it to the type system so it can propagate it as a
2758 // speculative type
2759 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2760 }
2761 }
2762
2763
2764 //=============================================================================
2765 // Generate a fast path/slow path idiom. Graph looks like:
2766 // [foo] indicates that 'foo' is a parameter
2767 //
2768 // [in] null
2769 // \ /
2770 // CmpP
2771 // Bool ne
2772 // If
2773 // / \
2774 // True False-<2>
2775 // / |
2776 // / cast_not_null
2777 // Load | | ^
2778 // [fast_test] | |
2779 // gvn to opt_test | |
2780 // / \ | <1>
2781 // True False |
2782 // | \\ |
2783 // [slow_call] \[fast_result]
2784 // Ctl Val \ \
2785 // | \ \
2786 // Catch <1> \ \
2787 // / \ ^ \ \
2788 // Ex No_Ex | \ \
2789 // | \ \ | \ <2> \
2790 // ... \ [slow_res] | | \ [null_result]
2791 // \ \--+--+--- | |
2792 // \ | / \ | /
2793 // --------Region Phi
2794 //
2795 //=============================================================================
2796 // Code is structured as a series of driver functions all called 'do_XXX' that
2797 // call a set of helper functions. Helper functions first, then drivers.
2798
2799 //------------------------------null_check_oop---------------------------------
2800 // Null check oop. Set null-path control into Region in slot 3.
2801 // Make a cast-not-nullness use the other not-null control. Return cast.
2802 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2803 bool never_see_null,
2804 bool safe_for_replace,
2805 bool speculative) {
2806 // Initial null check taken path
2807 (*null_control) = top();
2808 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2809
2810 // Generate uncommon_trap:
2811 if (never_see_null && (*null_control) != top()) {
2812 // If we see an unexpected null at a check-cast we record it and force a
2813 // recompile; the offending check-cast will be compiled to handle nulls.
2814 // If we see more than one offending BCI, then all checkcasts in the
2815 // method will be compiled to handle nulls.
2816 PreserveJVMState pjvms(this);
2817 set_control(*null_control);
2818 replace_in_map(value, null());
2819 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2820 uncommon_trap(reason,
2821 Deoptimization::Action_make_not_entrant);
2822 (*null_control) = top(); // null path is dead
2823 }
2824 if ((*null_control) == top() && safe_for_replace) {
2825 replace_in_map(value, cast);
2826 }
2827
2828 // Cast away null-ness on the result
2829 return cast;
2830 }
2831
2832 //------------------------------opt_iff----------------------------------------
2833 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2834 // Return slow-path control.
2835 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2836 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2837
2838 // Fast path taken; set region slot 2
2839 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2840 region->init_req(2,fast_taken); // Capture fast-control
2841
2842 // Fast path not-taken, i.e. slow path
2843 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2844 return slow_taken;
2845 }
2846
2847 //-----------------------------make_runtime_call-------------------------------
2848 Node* GraphKit::make_runtime_call(int flags,
2849 const TypeFunc* call_type, address call_addr,
2850 const char* call_name,
2851 const TypePtr* adr_type,
2852 // The following parms are all optional.
2853 // The first null ends the list.
2854 Node* parm0, Node* parm1,
2855 Node* parm2, Node* parm3,
2856 Node* parm4, Node* parm5,
2857 Node* parm6, Node* parm7) {
2858 assert(call_addr != nullptr, "must not call null targets");
2859
2860 // Slow-path call
2861 bool is_leaf = !(flags & RC_NO_LEAF);
2862 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2863 if (call_name == nullptr) {
2864 assert(!is_leaf, "must supply name for leaf");
2865 call_name = OptoRuntime::stub_name(call_addr);
2866 }
2867 CallNode* call;
2868 if (!is_leaf) {
2869 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2870 } else if (flags & RC_NO_FP) {
2871 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2872 } else if (flags & RC_VECTOR){
2873 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2874 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2875 } else if (flags & RC_PURE) {
2876 assert(adr_type == nullptr, "pure call does not touch memory");
2877 call = new CallLeafPureNode(call_type, call_addr, call_name);
2878 } else {
2879 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2880 }
2881
2882 // The following is similar to set_edges_for_java_call,
2883 // except that the memory effects of the call are restricted to AliasIdxRaw.
2884
2885 // Slow path call has no side-effects, uses few values
2886 bool wide_in = !(flags & RC_NARROW_MEM);
2887 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2888
2889 Node* prev_mem = nullptr;
2890 if (wide_in) {
2891 prev_mem = set_predefined_input_for_runtime_call(call);
2892 } else {
2893 assert(!wide_out, "narrow in => narrow out");
2894 Node* narrow_mem = memory(adr_type);
2895 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2896 }
2897
2898 // Hook each parm in order. Stop looking at the first null.
2899 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2900 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2901 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2902 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2903 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2904 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2905 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2906 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2907 /* close each nested if ===> */ } } } } } } } }
2908 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2909
2910 if (!is_leaf) {
2911 // Non-leaves can block and take safepoints:
2912 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2913 }
2914 // Non-leaves can throw exceptions:
2915 if (has_io) {
2916 call->set_req(TypeFunc::I_O, i_o());
2917 }
2918
2919 if (flags & RC_UNCOMMON) {
2920 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2921 // (An "if" probability corresponds roughly to an unconditional count.
2922 // Sort of.)
2923 call->set_cnt(PROB_UNLIKELY_MAG(4));
2924 }
2925
2926 Node* c = _gvn.transform(call);
2927 assert(c == call, "cannot disappear");
2928
2929 if (wide_out) {
2930 // Slow path call has full side-effects.
2931 set_predefined_output_for_runtime_call(call);
2932 } else {
2933 // Slow path call has few side-effects, and/or sets few values.
2934 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2935 }
2936
2937 if (has_io) {
2938 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2939 }
2940 return call;
2941
2942 }
2943
2944 // i2b
2945 Node* GraphKit::sign_extend_byte(Node* in) {
2946 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2947 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2948 }
2949
2950 // i2s
2951 Node* GraphKit::sign_extend_short(Node* in) {
2952 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2953 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2954 }
2955
2956
2957 //------------------------------merge_memory-----------------------------------
2958 // Merge memory from one path into the current memory state.
2959 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2960 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2961 Node* old_slice = mms.force_memory();
2962 Node* new_slice = mms.memory2();
2963 if (old_slice != new_slice) {
2964 PhiNode* phi;
2965 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2966 if (mms.is_empty()) {
2967 // clone base memory Phi's inputs for this memory slice
2968 assert(old_slice == mms.base_memory(), "sanity");
2969 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2970 _gvn.set_type(phi, Type::MEMORY);
2971 for (uint i = 1; i < phi->req(); i++) {
2972 phi->init_req(i, old_slice->in(i));
2973 }
2974 } else {
2975 phi = old_slice->as_Phi(); // Phi was generated already
2976 }
2977 } else {
2978 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2979 _gvn.set_type(phi, Type::MEMORY);
2980 }
2981 phi->set_req(new_path, new_slice);
2982 mms.set_memory(phi);
2983 }
2984 }
2985 }
2986
2987 //------------------------------make_slow_call_ex------------------------------
2988 // Make the exception handler hookups for the slow call
2989 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2990 if (stopped()) return;
2991
2992 // Make a catch node with just two handlers: fall-through and catch-all
2993 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2994 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2995 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2996 _gvn.set_type_bottom(norm);
2997 C->record_for_igvn(norm);
2998 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2999
3000 { PreserveJVMState pjvms(this);
3001 set_control(excp);
3002 set_i_o(i_o);
3003
3004 if (excp != top()) {
3005 if (deoptimize) {
3006 // Deoptimize if an exception is caught. Don't construct exception state in this case.
3007 uncommon_trap(Deoptimization::Reason_unhandled,
3008 Deoptimization::Action_none);
3009 } else {
3010 // Create an exception state also.
3011 // Use an exact type if the caller has a specific exception.
3012 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
3013 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
3014 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
3015 }
3016 }
3017 }
3018
3019 // Get the no-exception control from the CatchNode.
3020 set_control(norm);
3021 }
3022
3023 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
3024 Node* cmp = nullptr;
3025 switch(bt) {
3026 case T_INT: cmp = new CmpINode(in1, in2); break;
3027 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
3028 default: fatal("unexpected comparison type %s", type2name(bt));
3029 }
3030 cmp = gvn.transform(cmp);
3031 Node* bol = gvn.transform(new BoolNode(cmp, test));
3032 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
3033 gvn.transform(iff);
3034 if (!bol->is_Con()) gvn.record_for_igvn(iff);
3035 return iff;
3036 }
3037
3038 //-------------------------------gen_subtype_check-----------------------------
3039 // Generate a subtyping check. Takes as input the subtype and supertype.
3040 // Returns 2 values: sets the default control() to the true path and returns
3041 // the false path. Only reads invariant memory; sets no (visible) memory.
3042 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3043 // but that's not exposed to the optimizer. This call also doesn't take in an
3044 // Object; if you wish to check an Object you need to load the Object's class
3045 // prior to coming here.
3046 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3047 ciMethod* method, int bci) {
3048 Compile* C = gvn.C;
3049 if ((*ctrl)->is_top()) {
3050 return C->top();
3051 }
3052
3053 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3054 // For a direct pointer comparison, we need the refined array klass pointer
3055 Node* vm_superklass = superklass;
3056 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3057 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3058 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3059 }
3060
3061 // Fast check for identical types, perhaps identical constants.
3062 // The types can even be identical non-constants, in cases
3063 // involving Array.newInstance, Object.clone, etc.
3064 if (subklass == superklass)
3065 return C->top(); // false path is dead; no test needed.
3066
3067 if (gvn.type(superklass)->singleton()) {
3068 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3069 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3070
3071 // In the common case of an exact superklass, try to fold up the
3072 // test before generating code. You may ask, why not just generate
3073 // the code and then let it fold up? The answer is that the generated
3074 // code will necessarily include null checks, which do not always
3075 // completely fold away. If they are also needless, then they turn
3076 // into a performance loss. Example:
3077 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3078 // Here, the type of 'fa' is often exact, so the store check
3079 // of fa[1]=x will fold up, without testing the nullness of x.
3080 //
3081 // At macro expansion, we would have already folded the SubTypeCheckNode
3082 // being expanded here because we always perform the static sub type
3083 // check in SubTypeCheckNode::sub() regardless of whether
3084 // StressReflectiveCode is set or not. We can therefore skip this
3085 // static check when StressReflectiveCode is on.
3086 switch (C->static_subtype_check(superk, subk)) {
3087 case Compile::SSC_always_false:
3088 {
3089 Node* always_fail = *ctrl;
3090 *ctrl = gvn.C->top();
3091 return always_fail;
3092 }
3093 case Compile::SSC_always_true:
3094 return C->top();
3095 case Compile::SSC_easy_test:
3096 {
3097 // Just do a direct pointer compare and be done.
3098 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3099 *ctrl = gvn.transform(new IfTrueNode(iff));
3100 return gvn.transform(new IfFalseNode(iff));
3101 }
3102 case Compile::SSC_full_test:
3103 break;
3104 default:
3105 ShouldNotReachHere();
3106 }
3107 }
3108
3109 // %%% Possible further optimization: Even if the superklass is not exact,
3110 // if the subklass is the unique subtype of the superklass, the check
3111 // will always succeed. We could leave a dependency behind to ensure this.
3112
3113 // First load the super-klass's check-offset
3114 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3115 Node* m = C->immutable_memory();
3116 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3117 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3118 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3119 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
3120 bool might_be_cache = (chk_off_con == cacheoff_con);
3121
3122 // Load from the sub-klass's super-class display list, or a 1-word cache of
3123 // the secondary superclass list, or a failing value with a sentinel offset
3124 // if the super-klass is an interface or exceptionally deep in the Java
3125 // hierarchy and we have to scan the secondary superclass list the hard way.
3126 // Worst-case type is a little odd: null is allowed as a result (usually
3127 // klass loads can never produce a null).
3128 Node *chk_off_X = chk_off;
3129 #ifdef _LP64
3130 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
3131 #endif
3132 Node* p2 = gvn.transform(AddPNode::make_off_heap(subklass, chk_off_X));
3133 // For some types like interfaces the following loadKlass is from a 1-word
3134 // cache which is mutable so can't use immutable memory. Other
3135 // types load from the super-class display table which is immutable.
3136 Node* kmem = C->immutable_memory();
3137 // secondary_super_cache is not immutable but can be treated as such because:
3138 // - no ideal node writes to it in a way that could cause an
3139 // incorrect/missed optimization of the following Load.
3140 // - it's a cache so, worse case, not reading the latest value
3141 // wouldn't cause incorrect execution
3142 if (might_be_cache && mem != nullptr) {
3143 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
3144 }
3145 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3146
3147 // Compile speed common case: ARE a subtype and we canNOT fail
3148 if (superklass == nkls) {
3149 return C->top(); // false path is dead; no test needed.
3150 }
3151
3152 // Gather the various success & failures here
3153 RegionNode* r_not_subtype = new RegionNode(3);
3154 gvn.record_for_igvn(r_not_subtype);
3155 RegionNode* r_ok_subtype = new RegionNode(4);
3156 gvn.record_for_igvn(r_ok_subtype);
3157
3158 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3159 // SubTypeCheck node
3160 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3161 ciCallProfile profile = method->call_profile_at_bci(bci);
3162 float total_prob = 0;
3163 for (int i = 0; profile.has_receiver(i); ++i) {
3164 float prob = profile.receiver_prob(i);
3165 total_prob += prob;
3166 }
3167 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3168 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3169 for (int i = 0; profile.has_receiver(i); ++i) {
3170 ciKlass* klass = profile.receiver(i);
3171 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3172 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3173 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3174 continue;
3175 }
3176 if (klass_t->isa_aryklassptr()) {
3177 // For a direct pointer comparison, we need the refined array klass pointer
3178 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3179 }
3180 float prob = profile.receiver_prob(i);
3181 ConNode* klass_node = gvn.makecon(klass_t);
3182 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3183 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3184
3185 if (result == Compile::SSC_always_true) {
3186 r_ok_subtype->add_req(iftrue);
3187 } else {
3188 assert(result == Compile::SSC_always_false, "");
3189 r_not_subtype->add_req(iftrue);
3190 }
3191 *ctrl = gvn.transform(new IfFalseNode(iff));
3192 }
3193 }
3194 }
3195
3196 // See if we get an immediate positive hit. Happens roughly 83% of the
3197 // time. Test to see if the value loaded just previously from the subklass
3198 // is exactly the superklass.
3199 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3200 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3201 *ctrl = gvn.transform(new IfFalseNode(iff1));
3202
3203 // Compile speed common case: Check for being deterministic right now. If
3204 // chk_off is a constant and not equal to cacheoff then we are NOT a
3205 // subklass. In this case we need exactly the 1 test above and we can
3206 // return those results immediately.
3207 if (!might_be_cache) {
3208 Node* not_subtype_ctrl = *ctrl;
3209 *ctrl = iftrue1; // We need exactly the 1 test above
3210 PhaseIterGVN* igvn = gvn.is_IterGVN();
3211 if (igvn != nullptr) {
3212 igvn->remove_globally_dead_node(r_ok_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3213 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3214 }
3215 return not_subtype_ctrl;
3216 }
3217
3218 r_ok_subtype->init_req(1, iftrue1);
3219
3220 // Check for immediate negative hit. Happens roughly 11% of the time (which
3221 // is roughly 63% of the remaining cases). Test to see if the loaded
3222 // check-offset points into the subklass display list or the 1-element
3223 // cache. If it points to the display (and NOT the cache) and the display
3224 // missed then it's not a subtype.
3225 Node *cacheoff = gvn.intcon(cacheoff_con);
3226 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3227 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3228 *ctrl = gvn.transform(new IfFalseNode(iff2));
3229
3230 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3231 // No performance impact (too rare) but allows sharing of secondary arrays
3232 // which has some footprint reduction.
3233 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3234 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3235 *ctrl = gvn.transform(new IfFalseNode(iff3));
3236
3237 // -- Roads not taken here: --
3238 // We could also have chosen to perform the self-check at the beginning
3239 // of this code sequence, as the assembler does. This would not pay off
3240 // the same way, since the optimizer, unlike the assembler, can perform
3241 // static type analysis to fold away many successful self-checks.
3242 // Non-foldable self checks work better here in second position, because
3243 // the initial primary superclass check subsumes a self-check for most
3244 // types. An exception would be a secondary type like array-of-interface,
3245 // which does not appear in its own primary supertype display.
3246 // Finally, we could have chosen to move the self-check into the
3247 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3248 // dependent manner. But it is worthwhile to have the check here,
3249 // where it can be perhaps be optimized. The cost in code space is
3250 // small (register compare, branch).
3251
3252 // Now do a linear scan of the secondary super-klass array. Again, no real
3253 // performance impact (too rare) but it's gotta be done.
3254 // Since the code is rarely used, there is no penalty for moving it
3255 // out of line, and it can only improve I-cache density.
3256 // The decision to inline or out-of-line this final check is platform
3257 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3258 Node* psc = gvn.transform(
3259 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3260
3261 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3262 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3263 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3264
3265 // Return false path; set default control to true path.
3266 *ctrl = gvn.transform(r_ok_subtype);
3267 return gvn.transform(r_not_subtype);
3268 }
3269
3270 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3271 const Type* sub_t = _gvn.type(obj_or_subklass);
3272 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3273 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3274 obj_or_subklass = makecon(sub_t);
3275 }
3276 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3277 if (expand_subtype_check) {
3278 MergeMemNode* mem = merged_memory();
3279 Node* ctrl = control();
3280 Node* subklass = obj_or_subklass;
3281 if (!sub_t->isa_klassptr()) {
3282 subklass = load_object_klass(obj_or_subklass);
3283 }
3284
3285 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3286 set_control(ctrl);
3287 return n;
3288 }
3289
3290 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3291 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3292 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3293 set_control(_gvn.transform(new IfTrueNode(iff)));
3294 return _gvn.transform(new IfFalseNode(iff));
3295 }
3296
3297 // Profile-driven exact type check:
3298 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3299 float prob, Node* *casted_receiver) {
3300 assert(!klass->is_interface(), "no exact type check on interfaces");
3301 Node* fail = top();
3302 const Type* rec_t = _gvn.type(receiver);
3303 if (rec_t->is_inlinetypeptr()) {
3304 if (klass->equals(rec_t->inline_klass())) {
3305 (*casted_receiver) = receiver; // Always passes
3306 } else {
3307 (*casted_receiver) = top(); // Always fails
3308 fail = control();
3309 set_control(top());
3310 }
3311 return fail;
3312 }
3313 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3314 if (tklass->isa_aryklassptr()) {
3315 // For a direct pointer comparison, we need the refined array klass pointer
3316 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3317 }
3318 Node* recv_klass = load_object_klass(receiver);
3319 fail = type_check(recv_klass, tklass, prob);
3320
3321 if (!stopped()) {
3322 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3323 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3324 assert(recv_xtype->klass_is_exact(), "");
3325
3326 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3327 // Subsume downstream occurrences of receiver with a cast to
3328 // recv_xtype, since now we know what the type will be.
3329 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3330 Node* res = _gvn.transform(cast);
3331 if (recv_xtype->is_inlinetypeptr()) {
3332 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3333 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3334 }
3335 (*casted_receiver) = res;
3336 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3337 // (User must make the replace_in_map call.)
3338 }
3339 }
3340
3341 return fail;
3342 }
3343
3344 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3345 float prob) {
3346 Node* want_klass = makecon(tklass);
3347 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3348 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3349 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3350 set_control(_gvn.transform(new IfTrueNode (iff)));
3351 Node* fail = _gvn.transform(new IfFalseNode(iff));
3352 return fail;
3353 }
3354
3355 //------------------------------subtype_check_receiver-------------------------
3356 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3357 Node** casted_receiver) {
3358 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3359 Node* want_klass = makecon(tklass);
3360
3361 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3362
3363 // Ignore interface type information until interface types are properly tracked.
3364 if (!stopped() && !klass->is_interface()) {
3365 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3366 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3367 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3368 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3369 if (recv_type->is_inlinetypeptr()) {
3370 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3371 }
3372 (*casted_receiver) = cast;
3373 }
3374 }
3375
3376 return slow_ctl;
3377 }
3378
3379 //------------------------------seems_never_null-------------------------------
3380 // Use null_seen information if it is available from the profile.
3381 // If we see an unexpected null at a type check we record it and force a
3382 // recompile; the offending check will be recompiled to handle nulls.
3383 // If we see several offending BCIs, then all checks in the
3384 // method will be recompiled.
3385 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3386 speculating = !_gvn.type(obj)->speculative_maybe_null();
3387 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3388 if (UncommonNullCast // Cutout for this technique
3389 && obj != null() // And not the -Xcomp stupid case?
3390 && !too_many_traps(reason)
3391 ) {
3392 if (speculating) {
3393 return true;
3394 }
3395 if (data == nullptr)
3396 // Edge case: no mature data. Be optimistic here.
3397 return true;
3398 // If the profile has not seen a null, assume it won't happen.
3399 assert(java_bc() == Bytecodes::_checkcast ||
3400 java_bc() == Bytecodes::_instanceof ||
3401 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3402 return !data->as_BitData()->null_seen();
3403 }
3404 speculating = false;
3405 return false;
3406 }
3407
3408 void GraphKit::guard_klass_being_initialized(Node* klass) {
3409 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3410 Node* adr = off_heap_plus_addr(klass, init_state_off);
3411 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3412 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3413 T_BYTE, MemNode::acquire);
3414 init_state = _gvn.transform(init_state);
3415
3416 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3417
3418 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3419 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3420
3421 { BuildCutout unless(this, tst, PROB_MAX);
3422 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3423 }
3424 }
3425
3426 void GraphKit::guard_init_thread(Node* klass) {
3427 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3428 Node* adr = off_heap_plus_addr(klass, init_thread_off);
3429
3430 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3431 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3432 T_ADDRESS, MemNode::unordered);
3433 init_thread = _gvn.transform(init_thread);
3434
3435 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3436
3437 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3438 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3439
3440 { BuildCutout unless(this, tst, PROB_MAX);
3441 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3442 }
3443 }
3444
3445 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3446 if (ik->is_being_initialized()) {
3447 if (C->needs_clinit_barrier(ik, context)) {
3448 Node* klass = makecon(TypeKlassPtr::make(ik));
3449 guard_klass_being_initialized(klass);
3450 guard_init_thread(klass);
3451 insert_mem_bar(Op_MemBarCPUOrder);
3452 }
3453 } else if (ik->is_initialized()) {
3454 return; // no barrier needed
3455 } else {
3456 uncommon_trap(Deoptimization::Reason_uninitialized,
3457 Deoptimization::Action_reinterpret,
3458 nullptr);
3459 }
3460 }
3461
3462 //------------------------maybe_cast_profiled_receiver-------------------------
3463 // If the profile has seen exactly one type, narrow to exactly that type.
3464 // Subsequent type checks will always fold up.
3465 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3466 const TypeKlassPtr* require_klass,
3467 ciKlass* spec_klass,
3468 bool safe_for_replace) {
3469 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3470
3471 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3472
3473 // Make sure we haven't already deoptimized from this tactic.
3474 if (too_many_traps_or_recompiles(reason))
3475 return nullptr;
3476
3477 // (No, this isn't a call, but it's enough like a virtual call
3478 // to use the same ciMethod accessor to get the profile info...)
3479 // If we have a speculative type use it instead of profiling (which
3480 // may not help us)
3481 ciKlass* exact_kls = spec_klass;
3482 if (exact_kls == nullptr) {
3483 if (java_bc() == Bytecodes::_aastore) {
3484 ciKlass* array_type = nullptr;
3485 ciKlass* element_type = nullptr;
3486 ProfilePtrKind element_ptr = ProfileMaybeNull;
3487 bool flat_array = true;
3488 bool null_free_array = true;
3489 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3490 exact_kls = element_type;
3491 } else {
3492 exact_kls = profile_has_unique_klass();
3493 }
3494 }
3495 if (exact_kls != nullptr) {// no cast failures here
3496 if (require_klass == nullptr ||
3497 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3498 // If we narrow the type to match what the type profile sees or
3499 // the speculative type, we can then remove the rest of the
3500 // cast.
3501 // This is a win, even if the exact_kls is very specific,
3502 // because downstream operations, such as method calls,
3503 // will often benefit from the sharper type.
3504 Node* exact_obj = not_null_obj; // will get updated in place...
3505 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3506 &exact_obj);
3507 { PreserveJVMState pjvms(this);
3508 set_control(slow_ctl);
3509 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3510 }
3511 if (safe_for_replace) {
3512 replace_in_map(not_null_obj, exact_obj);
3513 }
3514 return exact_obj;
3515 }
3516 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3517 }
3518
3519 return nullptr;
3520 }
3521
3522 /**
3523 * Cast obj to type and emit guard unless we had too many traps here
3524 * already
3525 *
3526 * @param obj node being casted
3527 * @param type type to cast the node to
3528 * @param not_null true if we know node cannot be null
3529 */
3530 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3531 ciKlass* type,
3532 bool not_null) {
3533 if (stopped()) {
3534 return obj;
3535 }
3536
3537 // type is null if profiling tells us this object is always null
3538 if (type != nullptr) {
3539 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3540 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3541
3542 if (!too_many_traps_or_recompiles(null_reason) &&
3543 !too_many_traps_or_recompiles(class_reason)) {
3544 Node* not_null_obj = nullptr;
3545 // not_null is true if we know the object is not null and
3546 // there's no need for a null check
3547 if (!not_null) {
3548 Node* null_ctl = top();
3549 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3550 assert(null_ctl->is_top(), "no null control here");
3551 } else {
3552 not_null_obj = obj;
3553 }
3554
3555 Node* exact_obj = not_null_obj;
3556 ciKlass* exact_kls = type;
3557 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3558 &exact_obj);
3559 {
3560 PreserveJVMState pjvms(this);
3561 set_control(slow_ctl);
3562 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3563 }
3564 replace_in_map(not_null_obj, exact_obj);
3565 obj = exact_obj;
3566 }
3567 } else {
3568 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3569 Node* exact_obj = null_assert(obj);
3570 replace_in_map(obj, exact_obj);
3571 obj = exact_obj;
3572 }
3573 }
3574 return obj;
3575 }
3576
3577 //-------------------------------gen_instanceof--------------------------------
3578 // Generate an instance-of idiom. Used by both the instance-of bytecode
3579 // and the reflective instance-of call.
3580 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3581 kill_dead_locals(); // Benefit all the uncommon traps
3582 assert( !stopped(), "dead parse path should be checked in callers" );
3583 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3584 "must check for not-null not-dead klass in callers");
3585
3586 // Make the merge point
3587 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3588 RegionNode* region = new RegionNode(PATH_LIMIT);
3589 Node* phi = new PhiNode(region, TypeInt::BOOL);
3590 C->set_has_split_ifs(true); // Has chance for split-if optimization
3591
3592 ciProfileData* data = nullptr;
3593 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3594 data = method()->method_data()->bci_to_data(bci());
3595 }
3596 bool speculative_not_null = false;
3597 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3598 && seems_never_null(obj, data, speculative_not_null));
3599
3600 // Null check; get casted pointer; set region slot 3
3601 Node* null_ctl = top();
3602 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3603
3604 // If not_null_obj is dead, only null-path is taken
3605 if (stopped()) { // Doing instance-of on a null?
3606 set_control(null_ctl);
3607 return intcon(0);
3608 }
3609 region->init_req(_null_path, null_ctl);
3610 phi ->init_req(_null_path, intcon(0)); // Set null path value
3611 if (null_ctl == top()) {
3612 // Do this eagerly, so that pattern matches like is_diamond_phi
3613 // will work even during parsing.
3614 assert(_null_path == PATH_LIMIT-1, "delete last");
3615 region->del_req(_null_path);
3616 phi ->del_req(_null_path);
3617 }
3618
3619 // Do we know the type check always succeed?
3620 bool known_statically = false;
3621 if (_gvn.type(superklass)->singleton()) {
3622 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3623 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3624 if (subk != nullptr && subk->is_loaded()) {
3625 int static_res = C->static_subtype_check(superk, subk);
3626 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3627 }
3628 }
3629
3630 if (!known_statically) {
3631 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3632 // We may not have profiling here or it may not help us. If we
3633 // have a speculative type use it to perform an exact cast.
3634 ciKlass* spec_obj_type = obj_type->speculative_type();
3635 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3636 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3637 if (stopped()) { // Profile disagrees with this path.
3638 set_control(null_ctl); // Null is the only remaining possibility.
3639 return intcon(0);
3640 }
3641 if (cast_obj != nullptr) {
3642 not_null_obj = cast_obj;
3643 }
3644 }
3645 }
3646
3647 // Generate the subtype check
3648 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3649
3650 // Plug in the success path to the general merge in slot 1.
3651 region->init_req(_obj_path, control());
3652 phi ->init_req(_obj_path, intcon(1));
3653
3654 // Plug in the failing path to the general merge in slot 2.
3655 region->init_req(_fail_path, not_subtype_ctrl);
3656 phi ->init_req(_fail_path, intcon(0));
3657
3658 // Return final merged results
3659 set_control( _gvn.transform(region) );
3660 record_for_igvn(region);
3661
3662 // If we know the type check always succeeds then we don't use the
3663 // profiling data at this bytecode. Don't lose it, feed it to the
3664 // type system as a speculative type.
3665 if (safe_for_replace) {
3666 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3667 replace_in_map(obj, casted_obj);
3668 }
3669
3670 return _gvn.transform(phi);
3671 }
3672
3673 //-------------------------------gen_checkcast---------------------------------
3674 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3675 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3676 // uncommon-trap paths work. Adjust stack after this call.
3677 // If failure_control is supplied and not null, it is filled in with
3678 // the control edge for the cast failure. Otherwise, an appropriate
3679 // uncommon trap or exception is thrown.
3680 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3681 // when the current map for the success path is updated with information only present
3682 // on the success path and not the cast failure path. The newly cloned map should then be
3683 // used to emit the uncommon trap in the caller.
3684 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3685 assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3686 "failure_control must be set when new_failure_map is used");
3687 kill_dead_locals(); // Benefit all the uncommon traps
3688 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3689 const Type* obj_type = _gvn.type(obj);
3690
3691 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3692 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3693 bool safe_for_replace = (failure_control == nullptr);
3694 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3695
3696 // Fast cutout: Check the case that the cast is vacuously true.
3697 // This detects the common cases where the test will short-circuit
3698 // away completely. We do this before we perform the null check,
3699 // because if the test is going to turn into zero code, we don't
3700 // want a residual null check left around. (Causes a slowdown,
3701 // for example, in some objArray manipulations, such as a[i]=a[j].)
3702 if (improved_klass_ptr_type->singleton()) {
3703 const TypeKlassPtr* kptr = nullptr;
3704 if (obj_type->isa_oop_ptr()) {
3705 kptr = obj_type->is_oopptr()->as_klass_type();
3706 } else if (obj->is_InlineType()) {
3707 ciInlineKlass* vk = obj_type->inline_klass();
3708 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3709 }
3710
3711 if (kptr != nullptr) {
3712 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3713 case Compile::SSC_always_true:
3714 // If we know the type check always succeed then we don't use
3715 // the profiling data at this bytecode. Don't lose it, feed it
3716 // to the type system as a speculative type.
3717 obj = record_profiled_receiver_for_speculation(obj);
3718 if (null_free) {
3719 assert(safe_for_replace, "must be");
3720 obj = null_check(obj);
3721 }
3722 return obj;
3723 case Compile::SSC_always_false:
3724 if (null_free) {
3725 assert(safe_for_replace, "must be");
3726 obj = null_check(obj);
3727 }
3728 // It needs a null check because a null will *pass* the cast check.
3729 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3730 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3731 Deoptimization::DeoptReason reason = is_aastore ?
3732 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3733 builtin_throw(reason);
3734 return top();
3735 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3736 return null_assert(obj);
3737 }
3738 break; // Fall through to full check
3739 default:
3740 break;
3741 }
3742 }
3743 }
3744
3745 ciProfileData* data = nullptr;
3746 if (failure_control == nullptr) { // use MDO in regular case only
3747 assert(java_bc() == Bytecodes::_aastore ||
3748 java_bc() == Bytecodes::_checkcast,
3749 "interpreter profiles type checks only for these BCs");
3750 if (method()->method_data()->is_mature()) {
3751 data = method()->method_data()->bci_to_data(bci());
3752 }
3753 }
3754
3755 // Make the merge point
3756 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3757 RegionNode* region = new RegionNode(PATH_LIMIT);
3758 Node* phi = new PhiNode(region, toop);
3759 _gvn.set_type(region, Type::CONTROL);
3760 _gvn.set_type(phi, toop);
3761
3762 C->set_has_split_ifs(true); // Has chance for split-if optimization
3763
3764 // Use null-cast information if it is available
3765 bool speculative_not_null = false;
3766 bool never_see_null = ((failure_control == nullptr) // regular case only
3767 && seems_never_null(obj, data, speculative_not_null));
3768
3769 // Null check; get casted pointer; set region slot 3
3770 Node* null_ctl = top();
3771 Node* not_null_obj = nullptr;
3772 if (null_free) {
3773 assert(safe_for_replace, "must be");
3774 not_null_obj = null_check(obj);
3775 } else {
3776 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, false /*safe_for_replace*/, speculative_not_null);
3777 }
3778
3779 // If not_null_obj is dead, only null-path is taken
3780 if (stopped()) { // Doing instance-of on a null?
3781 set_control(null_ctl);
3782 if (toop->is_inlinetypeptr()) {
3783 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3784 }
3785 return null();
3786 }
3787 region->init_req(_null_path, null_ctl);
3788 phi ->init_req(_null_path, null()); // Set null path value
3789 if (null_ctl == top()) {
3790 // Do this eagerly, so that pattern matches like is_diamond_phi
3791 // will work even during parsing.
3792 assert(_null_path == PATH_LIMIT-1, "delete last");
3793 region->del_req(_null_path);
3794 phi ->del_req(_null_path);
3795 }
3796
3797 Node* cast_obj = nullptr;
3798 if (improved_klass_ptr_type->klass_is_exact()) {
3799 // The following optimization tries to statically cast the speculative type of the object
3800 // (for example obtained during profiling) to the type of the superklass and then do a
3801 // dynamic check that the type of the object is what we expect. To work correctly
3802 // for checkcast and aastore the type of superklass should be exact.
3803 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3804 // We may not have profiling here or it may not help us. If we have
3805 // a speculative type use it to perform an exact cast.
3806 ciKlass* spec_obj_type = obj_type->speculative_type();
3807 if (spec_obj_type != nullptr || data != nullptr) {
3808 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, false /*safe_for_replace*/);
3809 if (cast_obj != nullptr) {
3810 if (failure_control != nullptr) // failure is now impossible
3811 (*failure_control) = top();
3812 // adjust the type of the phi to the exact klass:
3813 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3814 }
3815 }
3816 }
3817
3818 if (cast_obj == nullptr) {
3819 // Generate the subtype check
3820 Node* improved_superklass = superklass;
3821 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3822 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3823 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3824 // Additionally, the benefit would only be minor in non-constant cases.
3825 improved_superklass = makecon(improved_klass_ptr_type);
3826 }
3827 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3828 // Plug in success path into the merge
3829 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3830 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3831 if (failure_control == nullptr) {
3832 if (not_subtype_ctrl != top()) { // If failure is possible
3833 PreserveJVMState pjvms(this);
3834 set_control(not_subtype_ctrl);
3835 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3836 Deoptimization::DeoptReason reason = is_aastore ?
3837 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3838 builtin_throw(reason);
3839 }
3840 } else {
3841 (*failure_control) = not_subtype_ctrl;
3842 }
3843 }
3844
3845 region->init_req(_obj_path, control());
3846 phi ->init_req(_obj_path, cast_obj);
3847
3848 // Return final merged results
3849 set_control( _gvn.transform(region) );
3850 record_for_igvn(region);
3851
3852 bool not_inline = !toop->can_be_inline_type();
3853 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3854 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3855 // Check if obj has been loaded from an array. Keep obj unchanged for final
3856 // map replacement below.
3857 Node* array_obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3858 Node* array = nullptr;
3859 if (array_obj->isa_Load()) {
3860 Node* address = array_obj->in(MemNode::Address);
3861 if (address->isa_AddP()) {
3862 array = address->as_AddP()->in(AddPNode::Base);
3863 }
3864 } else if (array_obj->is_Phi()) {
3865 Node* region = array_obj->in(0);
3866 // TODO make this more robust (see JDK-8231346)
3867 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3868 IfNode* iff = region->in(2)->in(0)->isa_If();
3869 if (iff != nullptr) {
3870 iff->is_flat_array_check(&_gvn, &array);
3871 }
3872 }
3873 }
3874 if (array != nullptr) {
3875 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3876 if (ary_t != nullptr) {
3877 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3878 // Casting array element to a non-inline-type, mark array as not null-free.
3879 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3880 if (new_cast_failure_map != nullptr) {
3881 // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3882 // create the cast failure path in the caller without wrongly making the cast node live there.
3883 *new_cast_failure_map = clone_map();
3884 }
3885 replace_in_map(array, cast);
3886 array = cast;
3887 }
3888 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3889 // Casting array element to a non-flat-in-array type, mark array as not flat.
3890 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3891 if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3892 // Same as above.
3893 *new_cast_failure_map = clone_map();
3894 }
3895 replace_in_map(array, cast);
3896 array = cast;
3897 }
3898 }
3899 }
3900 }
3901
3902 // A merge of null or Casted-NotNull obj
3903 Node* res = _gvn.transform(phi);
3904 if (!stopped() && !res->is_InlineType()) {
3905 res = record_profiled_receiver_for_speculation(res);
3906 if (toop->is_inlinetypeptr() && !maybe_larval) {
3907 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3908 res = vt;
3909 if (safe_for_replace) {
3910 replace_in_map(obj, vt);
3911 replace_in_map(not_null_obj, vt);
3912 replace_in_map(res, vt);
3913 }
3914 } else if (safe_for_replace) {
3915 replace_in_map(obj, res);
3916 }
3917 }
3918 return res;
3919 }
3920
3921 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3922 // Load markword
3923 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3924 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3925 if (check_lock && !UseCompactObjectHeaders) {
3926 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3927 // Check if obj is locked
3928 Node* locked_bit = MakeConX(markWord::unlocked_value);
3929 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3930 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3931 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3932 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3933 _gvn.transform(iff);
3934 Node* locked_region = new RegionNode(3);
3935 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3936
3937 // Unlocked: Use bits from mark word
3938 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3939 mark_phi->init_req(1, mark);
3940
3941 // Locked: Load prototype header from klass
3942 set_control(_gvn.transform(new IfFalseNode(iff)));
3943 // Make loads control dependent to make sure they are only executed if array is locked
3944 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3945 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3946 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3947 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3948
3949 locked_region->init_req(2, control());
3950 mark_phi->init_req(2, proto);
3951 set_control(_gvn.transform(locked_region));
3952 record_for_igvn(locked_region);
3953
3954 mark = mark_phi;
3955 }
3956
3957 // Now check if mark word bits are set
3958 Node* mask = MakeConX(mask_val);
3959 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3960 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3961 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3962 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3963 }
3964
3965 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3966 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3967 }
3968
3969 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3970 // We can't use immutable memory here because the mark word is mutable.
3971 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3972 // check is moved out of loops (mainly to enable loop unswitching).
3973 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3974 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3975 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3976 }
3977
3978 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3979 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3980 }
3981
3982 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3983 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3984
3985 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3986 if (!vk->has_null_free_non_atomic_layout()) {
3987 return intcon(1); // Always atomic
3988 } else if (!vk->has_null_free_atomic_layout()) {
3989 return intcon(0); // Never atomic
3990 }
3991
3992 Node* array_klass = load_object_klass(array);
3993 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3994 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
3995 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3996 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
3997 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3998 }
3999
4000 Node* GraphKit::atomic_layout_array_test_and_get_layout_kind(Node* array, RegionNode* atomic_region) {
4001 Node* array_klass = load_object_klass(array);
4002 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4003 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4004 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4005 Node* cmp_null_free = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULL_FREE_ATOMIC_FLAT))));
4006 Node* bol_null_free = _gvn.transform(new BoolNode(cmp_null_free, BoolTest::eq));
4007 Node* cmp_nullable = _gvn.transform(new CmpINode(layout_kind, intcon(static_cast<jint>(LayoutKind::NULLABLE_ATOMIC_FLAT))));
4008 Node* bol_nullable = _gvn.transform(new BoolNode(cmp_nullable, BoolTest::eq));
4009
4010 IfNode* iff_null_free = create_and_xform_if(control(), bol_null_free, PROB_FAIR, COUNT_UNKNOWN);
4011 atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_null_free)));
4012 set_control(_gvn.transform(new IfFalseNode(iff_null_free)));
4013
4014 IfNode* iff_nullable = create_and_xform_if(control(), bol_nullable, PROB_FAIR, COUNT_UNKNOWN);
4015 atomic_region->add_req(_gvn.transform(new IfTrueNode(iff_nullable)));
4016 set_control(_gvn.transform(new IfFalseNode(iff_nullable)));
4017
4018 return layout_kind;
4019 }
4020
4021 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4022 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4023 RegionNode* region = new RegionNode(3);
4024 Node* null_ctl = top();
4025 null_check_oop(val, &null_ctl);
4026 if (null_ctl != top()) {
4027 PreserveJVMState pjvms(this);
4028 set_control(null_ctl);
4029 {
4030 // Deoptimize if null-free array
4031 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4032 inc_sp(nargs);
4033 uncommon_trap(Deoptimization::Reason_null_check,
4034 Deoptimization::Action_none);
4035 }
4036 region->init_req(1, control());
4037 }
4038 region->init_req(2, control());
4039 set_control(_gvn.transform(region));
4040 record_for_igvn(region);
4041 if (_gvn.type(val) == TypePtr::NULL_PTR) {
4042 // Since we were just successfully storing null, the array can't be null free.
4043 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4044 ary_t = ary_t->cast_to_not_null_free();
4045 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4046 if (safe_for_replace) {
4047 replace_in_map(ary, cast);
4048 }
4049 ary = cast;
4050 }
4051 return ary;
4052 }
4053
4054 //------------------------------next_monitor-----------------------------------
4055 // What number should be given to the next monitor?
4056 int GraphKit::next_monitor() {
4057 int current = jvms()->monitor_depth()* C->sync_stack_slots();
4058 int next = current + C->sync_stack_slots();
4059 // Keep the toplevel high water mark current:
4060 if (C->fixed_slots() < next) C->set_fixed_slots(next);
4061 return current;
4062 }
4063
4064 //------------------------------insert_mem_bar---------------------------------
4065 // Memory barrier to avoid floating things around
4066 // The membar serves as a pinch point between both control and all memory slices.
4067 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4068 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4069 mb->init_req(TypeFunc::Control, control());
4070 mb->init_req(TypeFunc::Memory, reset_memory());
4071 Node* membar = _gvn.transform(mb);
4072 record_for_igvn(membar);
4073 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4074 set_all_memory_call(membar);
4075 return membar;
4076 }
4077
4078 //-------------------------insert_mem_bar_volatile----------------------------
4079 // Memory barrier to avoid floating things around
4080 // The membar serves as a pinch point between both control and memory(alias_idx).
4081 // If you want to make a pinch point on all memory slices, do not use this
4082 // function (even with AliasIdxBot); use insert_mem_bar() instead.
4083 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
4084 // When Parse::do_put_xxx updates a volatile field, it appends a series
4085 // of MemBarVolatile nodes, one for *each* volatile field alias category.
4086 // The first membar is on the same memory slice as the field store opcode.
4087 // This forces the membar to follow the store. (Bug 6500685 broke this.)
4088 // All the other membars (for other volatile slices, including AliasIdxBot,
4089 // which stands for all unknown volatile slices) are control-dependent
4090 // on the first membar. This prevents later volatile loads or stores
4091 // from sliding up past the just-emitted store.
4092
4093 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
4094 mb->set_req(TypeFunc::Control,control());
4095 if (alias_idx == Compile::AliasIdxBot) {
4096 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
4097 } else {
4098 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
4099 mb->set_req(TypeFunc::Memory, memory(alias_idx));
4100 }
4101 Node* membar = _gvn.transform(mb);
4102 record_for_igvn(membar);
4103 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4104 if (alias_idx == Compile::AliasIdxBot) {
4105 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
4106 } else {
4107 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
4108 }
4109 return membar;
4110 }
4111
4112 //------------------------------insert_reachability_fence----------------------
4113 Node* GraphKit::insert_reachability_fence(Node* referent) {
4114 assert(!referent->is_top(), "");
4115 Node* rf = _gvn.transform(new ReachabilityFenceNode(C, control(), referent));
4116 set_control(rf);
4117 C->record_for_igvn(rf);
4118 return rf;
4119 }
4120
4121 //------------------------------shared_lock------------------------------------
4122 // Emit locking code.
4123 FastLockNode* GraphKit::shared_lock(Node* obj) {
4124 // bci is either a monitorenter bc or InvocationEntryBci
4125 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4126 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4127
4128 if (stopped()) // Dead monitor?
4129 return nullptr;
4130
4131 assert(dead_locals_are_killed(), "should kill locals before sync. point");
4132
4133 // Box the stack location
4134 Node* box = new BoxLockNode(next_monitor());
4135 // Check for bailout after new BoxLockNode
4136 if (failing()) { return nullptr; }
4137 box = _gvn.transform(box);
4138 Node* mem = reset_memory();
4139
4140 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
4141
4142 // Add monitor to debug info for the slow path. If we block inside the
4143 // slow path and de-opt, we need the monitor hanging around
4144 map()->push_monitor( flock );
4145
4146 const TypeFunc *tf = LockNode::lock_type();
4147 LockNode *lock = new LockNode(C, tf);
4148
4149 lock->init_req( TypeFunc::Control, control() );
4150 lock->init_req( TypeFunc::Memory , mem );
4151 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4152 lock->init_req( TypeFunc::FramePtr, frameptr() );
4153 lock->init_req( TypeFunc::ReturnAdr, top() );
4154
4155 lock->init_req(TypeFunc::Parms + 0, obj);
4156 lock->init_req(TypeFunc::Parms + 1, box);
4157 lock->init_req(TypeFunc::Parms + 2, flock);
4158 add_safepoint_edges(lock);
4159
4160 lock = _gvn.transform( lock )->as_Lock();
4161
4162 // lock has no side-effects, sets few values
4163 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
4164
4165 insert_mem_bar(Op_MemBarAcquireLock);
4166
4167 // Add this to the worklist so that the lock can be eliminated
4168 record_for_igvn(lock);
4169
4170 #ifndef PRODUCT
4171 if (PrintLockStatistics) {
4172 // Update the counter for this lock. Don't bother using an atomic
4173 // operation since we don't require absolute accuracy.
4174 lock->create_lock_counter(map()->jvms());
4175 increment_counter(lock->counter()->addr());
4176 }
4177 #endif
4178
4179 return flock;
4180 }
4181
4182
4183 //------------------------------shared_unlock----------------------------------
4184 // Emit unlocking code.
4185 void GraphKit::shared_unlock(Node* box, Node* obj) {
4186 // bci is either a monitorenter bc or InvocationEntryBci
4187 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4188 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4189
4190 if (stopped()) { // Dead monitor?
4191 map()->pop_monitor(); // Kill monitor from debug info
4192 return;
4193 }
4194 assert(!obj->is_InlineType(), "should not unlock on inline type");
4195
4196 // Memory barrier to avoid floating things down past the locked region
4197 insert_mem_bar(Op_MemBarReleaseLock);
4198
4199 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4200 UnlockNode *unlock = new UnlockNode(C, tf);
4201 #ifdef ASSERT
4202 unlock->set_dbg_jvms(sync_jvms());
4203 #endif
4204 uint raw_idx = Compile::AliasIdxRaw;
4205 unlock->init_req( TypeFunc::Control, control() );
4206 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4207 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4208 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4209 unlock->init_req( TypeFunc::ReturnAdr, top() );
4210
4211 unlock->init_req(TypeFunc::Parms + 0, obj);
4212 unlock->init_req(TypeFunc::Parms + 1, box);
4213 unlock = _gvn.transform(unlock)->as_Unlock();
4214
4215 Node* mem = reset_memory();
4216
4217 // unlock has no side-effects, sets few values
4218 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4219
4220 // Kill monitor from debug info
4221 map()->pop_monitor( );
4222 }
4223
4224 //-------------------------------get_layout_helper-----------------------------
4225 // If the given klass is a constant or known to be an array,
4226 // fetch the constant layout helper value into constant_value
4227 // and return null. Otherwise, load the non-constant
4228 // layout helper value, and return the node which represents it.
4229 // This two-faced routine is useful because allocation sites
4230 // almost always feature constant types.
4231 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4232 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4233 if (!StressReflectiveCode && klass_t != nullptr) {
4234 bool xklass = klass_t->klass_is_exact();
4235 bool can_be_flat = false;
4236 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4237 if (UseArrayFlattening && !xklass && ary_type != nullptr) {
4238 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4239 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4240 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4241 }
4242 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4243 jint lhelper;
4244 if (klass_t->is_flat()) {
4245 lhelper = ary_type->flat_layout_helper();
4246 } else if (klass_t->isa_aryklassptr()) {
4247 BasicType elem = ary_type->elem()->array_element_basic_type();
4248 if (is_reference_type(elem, true)) {
4249 elem = T_OBJECT;
4250 }
4251 lhelper = Klass::array_layout_helper(elem);
4252 } else {
4253 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4254 }
4255 if (lhelper != Klass::_lh_neutral_value) {
4256 constant_value = lhelper;
4257 return (Node*) nullptr;
4258 }
4259 }
4260 }
4261 constant_value = Klass::_lh_neutral_value; // put in a known value
4262 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4263 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4264 }
4265
4266 // We just put in an allocate/initialize with a big raw-memory effect.
4267 // Hook selected additional alias categories on the initialization.
4268 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4269 MergeMemNode* init_in_merge,
4270 Node* init_out_raw) {
4271 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4272 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4273
4274 Node* prevmem = kit.memory(alias_idx);
4275 init_in_merge->set_memory_at(alias_idx, prevmem);
4276 if (init_out_raw != nullptr) {
4277 kit.set_memory(init_out_raw, alias_idx);
4278 }
4279 }
4280
4281 //---------------------------set_output_for_allocation-------------------------
4282 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4283 const TypeOopPtr* oop_type,
4284 bool deoptimize_on_exception) {
4285 int rawidx = Compile::AliasIdxRaw;
4286 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4287 add_safepoint_edges(alloc);
4288 Node* allocx = _gvn.transform(alloc);
4289 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4290 // create memory projection for i_o
4291 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4292 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4293
4294 // create a memory projection as for the normal control path
4295 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4296 set_memory(malloc, rawidx);
4297
4298 // a normal slow-call doesn't change i_o, but an allocation does
4299 // we create a separate i_o projection for the normal control path
4300 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4301 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4302
4303 // put in an initialization barrier
4304 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4305 rawoop)->as_Initialize();
4306 assert(alloc->initialization() == init, "2-way macro link must work");
4307 assert(init ->allocation() == alloc, "2-way macro link must work");
4308 {
4309 // Extract memory strands which may participate in the new object's
4310 // initialization, and source them from the new InitializeNode.
4311 // This will allow us to observe initializations when they occur,
4312 // and link them properly (as a group) to the InitializeNode.
4313 assert(init->in(InitializeNode::Memory) == malloc, "");
4314 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4315 init->set_req(InitializeNode::Memory, minit_in);
4316 record_for_igvn(minit_in); // fold it up later, if possible
4317 _gvn.set_type(minit_in, Type::MEMORY);
4318 Node* minit_out = memory(rawidx);
4319 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4320 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4321 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4322 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4323 // multiple projections as a result.
4324 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4325 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4326 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4327 if (oop_type->isa_aryptr()) {
4328 // Initially all flat array accesses share a single slice
4329 // but that changes after parsing. Prepare the memory graph so
4330 // it can optimize flat array accesses properly once they
4331 // don't share a single slice.
4332 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4333 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4334 int elemidx = C->get_alias_index(telemref);
4335 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4336 if (alias_adr_type->is_flat()) {
4337 C->set_flat_accesses();
4338 }
4339 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4340 } else if (oop_type->isa_instptr()) {
4341 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4342 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4343 ciField* field = ik->nonstatic_field_at(i);
4344 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4345 continue; // do not bother to track really large numbers of fields
4346 // Find (or create) the alias category for this field:
4347 int fieldidx = C->alias_type(field)->index();
4348 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4349 }
4350 }
4351 }
4352
4353 // Cast raw oop to the real thing...
4354 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4355 javaoop = _gvn.transform(javaoop);
4356 C->set_recent_alloc(control(), javaoop);
4357 assert(just_allocated_object(control()) == javaoop, "just allocated");
4358
4359 #ifdef ASSERT
4360 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4361 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4362 "Ideal_allocation works");
4363 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4364 "Ideal_allocation works");
4365 if (alloc->is_AllocateArray()) {
4366 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4367 "Ideal_allocation works");
4368 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4369 "Ideal_allocation works");
4370 } else {
4371 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4372 }
4373 }
4374 #endif //ASSERT
4375
4376 return javaoop;
4377 }
4378
4379 //---------------------------new_instance--------------------------------------
4380 // This routine takes a klass_node which may be constant (for a static type)
4381 // or may be non-constant (for reflective code). It will work equally well
4382 // for either, and the graph will fold nicely if the optimizer later reduces
4383 // the type to a constant.
4384 // The optional arguments are for specialized use by intrinsics:
4385 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4386 // - If 'return_size_val', report the total object size to the caller.
4387 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4388 Node* GraphKit::new_instance(Node* klass_node,
4389 Node* extra_slow_test,
4390 Node* *return_size_val,
4391 bool deoptimize_on_exception,
4392 InlineTypeNode* inline_type_node) {
4393 // Compute size in doublewords
4394 // The size is always an integral number of doublewords, represented
4395 // as a positive bytewise size stored in the klass's layout_helper.
4396 // The layout_helper also encodes (in a low bit) the need for a slow path.
4397 jint layout_con = Klass::_lh_neutral_value;
4398 Node* layout_val = get_layout_helper(klass_node, layout_con);
4399 bool layout_is_con = (layout_val == nullptr);
4400
4401 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4402 // Generate the initial go-slow test. It's either ALWAYS (return a
4403 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4404 // case) a computed value derived from the layout_helper.
4405 Node* initial_slow_test = nullptr;
4406 if (layout_is_con) {
4407 assert(!StressReflectiveCode, "stress mode does not use these paths");
4408 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4409 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4410 } else { // reflective case
4411 // This reflective path is used by Unsafe.allocateInstance.
4412 // (It may be stress-tested by specifying StressReflectiveCode.)
4413 // Basically, we want to get into the VM is there's an illegal argument.
4414 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4415 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4416 if (extra_slow_test != intcon(0)) {
4417 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4418 }
4419 // (Macro-expander will further convert this to a Bool, if necessary.)
4420 }
4421
4422 // Find the size in bytes. This is easy; it's the layout_helper.
4423 // The size value must be valid even if the slow path is taken.
4424 Node* size = nullptr;
4425 if (layout_is_con) {
4426 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4427 } else { // reflective case
4428 // This reflective path is used by clone and Unsafe.allocateInstance.
4429 size = ConvI2X(layout_val);
4430
4431 // Clear the low bits to extract layout_helper_size_in_bytes:
4432 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4433 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4434 size = _gvn.transform( new AndXNode(size, mask) );
4435 }
4436 if (return_size_val != nullptr) {
4437 (*return_size_val) = size;
4438 }
4439
4440 // This is a precise notnull oop of the klass.
4441 // (Actually, it need not be precise if this is a reflective allocation.)
4442 // It's what we cast the result to.
4443 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4444 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4445 const TypeOopPtr* oop_type = tklass->as_instance_type();
4446
4447 // Now generate allocation code
4448
4449 // The entire memory state is needed for slow path of the allocation
4450 // since GC and deoptimization can happen.
4451 Node *mem = reset_memory();
4452 set_all_memory(mem); // Create new memory state
4453
4454 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4455 control(), mem, i_o(),
4456 size, klass_node,
4457 initial_slow_test, inline_type_node);
4458
4459 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4460 }
4461
4462 //-------------------------------new_array-------------------------------------
4463 // helper for newarray and anewarray
4464 // The 'length' parameter is (obviously) the length of the array.
4465 // The optional arguments are for specialized use by intrinsics:
4466 // - If 'return_size_val', report the non-padded array size (sum of header size
4467 // and array body) to the caller.
4468 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4469 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4470 Node* length, // number of array elements
4471 int nargs, // number of arguments to push back for uncommon trap
4472 Node* *return_size_val,
4473 bool deoptimize_on_exception,
4474 Node* init_val) {
4475 jint layout_con = Klass::_lh_neutral_value;
4476 Node* layout_val = get_layout_helper(klass_node, layout_con);
4477 bool layout_is_con = (layout_val == nullptr);
4478
4479 if (!layout_is_con && !StressReflectiveCode &&
4480 !too_many_traps(Deoptimization::Reason_class_check)) {
4481 // This is a reflective array creation site.
4482 // Optimistically assume that it is a subtype of Object[],
4483 // so that we can fold up all the address arithmetic.
4484 layout_con = Klass::array_layout_helper(T_OBJECT);
4485 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4486 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4487 { BuildCutout unless(this, bol_lh, PROB_MAX);
4488 inc_sp(nargs);
4489 uncommon_trap(Deoptimization::Reason_class_check,
4490 Deoptimization::Action_maybe_recompile);
4491 }
4492 layout_val = nullptr;
4493 layout_is_con = true;
4494 }
4495
4496 // Generate the initial go-slow test. Make sure we do not overflow
4497 // if length is huge (near 2Gig) or negative! We do not need
4498 // exact double-words here, just a close approximation of needed
4499 // double-words. We can't add any offset or rounding bits, lest we
4500 // take a size -1 of bytes and make it positive. Use an unsigned
4501 // compare, so negative sizes look hugely positive.
4502 int fast_size_limit = FastAllocateSizeLimit;
4503 if (layout_is_con) {
4504 assert(!StressReflectiveCode, "stress mode does not use these paths");
4505 // Increase the size limit if we have exact knowledge of array type.
4506 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4507 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4508 }
4509
4510 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4511 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4512
4513 // --- Size Computation ---
4514 // array_size = round_to_heap(array_header + (length << elem_shift));
4515 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4516 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4517 // The rounding mask is strength-reduced, if possible.
4518 int round_mask = MinObjAlignmentInBytes - 1;
4519 Node* header_size = nullptr;
4520 // (T_BYTE has the weakest alignment and size restrictions...)
4521 if (layout_is_con) {
4522 int hsize = Klass::layout_helper_header_size(layout_con);
4523 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4524 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4525 if ((round_mask & ~right_n_bits(eshift)) == 0)
4526 round_mask = 0; // strength-reduce it if it goes away completely
4527 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4528 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4529 assert(header_size_min <= hsize, "generic minimum is smallest");
4530 header_size = intcon(hsize);
4531 } else {
4532 Node* hss = intcon(Klass::_lh_header_size_shift);
4533 Node* hsm = intcon(Klass::_lh_header_size_mask);
4534 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4535 header_size = _gvn.transform(new AndINode(header_size, hsm));
4536 }
4537
4538 Node* elem_shift = nullptr;
4539 if (layout_is_con) {
4540 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4541 if (eshift != 0)
4542 elem_shift = intcon(eshift);
4543 } else {
4544 // There is no need to mask or shift this value.
4545 // The semantics of LShiftINode include an implicit mask to 0x1F.
4546 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4547 elem_shift = layout_val;
4548 }
4549
4550 // Transition to native address size for all offset calculations:
4551 Node* lengthx = ConvI2X(length);
4552 Node* headerx = ConvI2X(header_size);
4553 #ifdef _LP64
4554 { const TypeInt* tilen = _gvn.find_int_type(length);
4555 if (tilen != nullptr && tilen->_lo < 0) {
4556 // Add a manual constraint to a positive range. Cf. array_element_address.
4557 jint size_max = fast_size_limit;
4558 if (size_max > tilen->_hi && tilen->_hi >= 0) {
4559 size_max = tilen->_hi;
4560 }
4561 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4562
4563 // Only do a narrow I2L conversion if the range check passed.
4564 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4565 _gvn.transform(iff);
4566 RegionNode* region = new RegionNode(3);
4567 _gvn.set_type(region, Type::CONTROL);
4568 lengthx = new PhiNode(region, TypeLong::LONG);
4569 _gvn.set_type(lengthx, TypeLong::LONG);
4570
4571 // Range check passed. Use ConvI2L node with narrow type.
4572 Node* passed = IfFalse(iff);
4573 region->init_req(1, passed);
4574 // Make I2L conversion control dependent to prevent it from
4575 // floating above the range check during loop optimizations.
4576 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4577
4578 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4579 region->init_req(2, IfTrue(iff));
4580 lengthx->init_req(2, ConvI2X(length));
4581
4582 set_control(region);
4583 record_for_igvn(region);
4584 record_for_igvn(lengthx);
4585 }
4586 }
4587 #endif
4588
4589 // Combine header size and body size for the array copy part, then align (if
4590 // necessary) for the allocation part. This computation cannot overflow,
4591 // because it is used only in two places, one where the length is sharply
4592 // limited, and the other after a successful allocation.
4593 Node* abody = lengthx;
4594 if (elem_shift != nullptr) {
4595 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4596 }
4597 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4598
4599 if (return_size_val != nullptr) {
4600 // This is the size
4601 (*return_size_val) = non_rounded_size;
4602 }
4603
4604 Node* size = non_rounded_size;
4605 if (round_mask != 0) {
4606 Node* mask1 = MakeConX(round_mask);
4607 size = _gvn.transform(new AddXNode(size, mask1));
4608 Node* mask2 = MakeConX(~round_mask);
4609 size = _gvn.transform(new AndXNode(size, mask2));
4610 }
4611 // else if round_mask == 0, the size computation is self-rounding
4612
4613 // Now generate allocation code
4614
4615 // The entire memory state is needed for slow path of the allocation
4616 // since GC and deoptimization can happen.
4617 Node *mem = reset_memory();
4618 set_all_memory(mem); // Create new memory state
4619
4620 if (initial_slow_test->is_Bool()) {
4621 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4622 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4623 }
4624
4625 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4626 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4627
4628 Node* raw_init_value = nullptr;
4629 if (init_val != nullptr) {
4630 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4631 if (ary_type->is_flat()) {
4632 initial_slow_test = intcon(1);
4633 }
4634
4635 if (UseCompressedOops) {
4636 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4637 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4638 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4639 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4640 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4641 } else {
4642 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4643 }
4644 }
4645
4646 Node* valid_length_test = _gvn.intcon(1);
4647 if (ary_type->isa_aryptr()) {
4648 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4649 jint max = TypeAryPtr::max_array_length(bt);
4650 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4651 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4652 }
4653
4654 // Create the AllocateArrayNode and its result projections
4655 AllocateArrayNode* alloc
4656 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4657 control(), mem, i_o(),
4658 size, klass_node,
4659 initial_slow_test,
4660 length, valid_length_test,
4661 init_val, raw_init_value);
4662 // Cast to correct type. Note that the klass_node may be constant or not,
4663 // and in the latter case the actual array type will be inexact also.
4664 // (This happens via a non-constant argument to inline_native_newArray.)
4665 // In any case, the value of klass_node provides the desired array type.
4666 const TypeInt* length_type = _gvn.find_int_type(length);
4667 if (ary_type->isa_aryptr() && length_type != nullptr) {
4668 // Try to get a better type than POS for the size
4669 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4670 }
4671
4672 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4673
4674 array_ideal_length(alloc, ary_type, true);
4675 return javaoop;
4676 }
4677
4678 // The following "Ideal_foo" functions are placed here because they recognize
4679 // the graph shapes created by the functions immediately above.
4680
4681 //---------------------------Ideal_allocation----------------------------------
4682 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4683 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4684 if (ptr == nullptr) { // reduce dumb test in callers
4685 return nullptr;
4686 }
4687
4688 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4689 ptr = bs->step_over_gc_barrier(ptr);
4690
4691 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4692 ptr = ptr->in(1);
4693 if (ptr == nullptr) return nullptr;
4694 }
4695 // Return null for allocations with several casts:
4696 // j.l.reflect.Array.newInstance(jobject, jint)
4697 // Object.clone()
4698 // to keep more precise type from last cast.
4699 if (ptr->is_Proj()) {
4700 Node* allo = ptr->in(0);
4701 if (allo != nullptr && allo->is_Allocate()) {
4702 return allo->as_Allocate();
4703 }
4704 }
4705 // Report failure to match.
4706 return nullptr;
4707 }
4708
4709 // Fancy version which also strips off an offset (and reports it to caller).
4710 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4711 intptr_t& offset) {
4712 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4713 if (base == nullptr) return nullptr;
4714 return Ideal_allocation(base);
4715 }
4716
4717 // Trace Initialize <- Proj[Parm] <- Allocate
4718 AllocateNode* InitializeNode::allocation() {
4719 Node* rawoop = in(InitializeNode::RawAddress);
4720 if (rawoop->is_Proj()) {
4721 Node* alloc = rawoop->in(0);
4722 if (alloc->is_Allocate()) {
4723 return alloc->as_Allocate();
4724 }
4725 }
4726 return nullptr;
4727 }
4728
4729 // Trace Allocate -> Proj[Parm] -> Initialize
4730 InitializeNode* AllocateNode::initialization() {
4731 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4732 if (rawoop == nullptr) return nullptr;
4733 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4734 Node* init = rawoop->fast_out(i);
4735 if (init->is_Initialize()) {
4736 assert(init->as_Initialize()->allocation() == this, "2-way link");
4737 return init->as_Initialize();
4738 }
4739 }
4740 return nullptr;
4741 }
4742
4743 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4744 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4745 // Too many traps seen?
4746 if (too_many_traps(reason)) {
4747 #ifdef ASSERT
4748 if (TraceLoopPredicate) {
4749 int tc = C->trap_count(reason);
4750 tty->print("too many traps=%s tcount=%d in ",
4751 Deoptimization::trap_reason_name(reason), tc);
4752 method()->print(); // which method has too many predicate traps
4753 tty->cr();
4754 }
4755 #endif
4756 // We cannot afford to take more traps here,
4757 // do not generate Parse Predicate.
4758 return;
4759 }
4760
4761 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4762 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4763 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4764 {
4765 PreserveJVMState pjvms(this);
4766 set_control(if_false);
4767 inc_sp(nargs);
4768 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4769 }
4770 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4771 set_control(if_true);
4772 }
4773
4774 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4775 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4776 void GraphKit::add_parse_predicates(int nargs) {
4777 if (ShortRunningLongLoop) {
4778 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4779 // walking up from the loop.
4780 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4781 }
4782 if (UseLoopPredicate) {
4783 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4784 if (UseProfiledLoopPredicate) {
4785 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4786 }
4787 }
4788 if (UseAutoVectorizationPredicate) {
4789 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4790 }
4791 // Loop Limit Check Predicate should be near the loop.
4792 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4793 }
4794
4795 void GraphKit::sync_kit(IdealKit& ideal) {
4796 reset_memory();
4797 set_all_memory(ideal.merged_memory());
4798 set_i_o(ideal.i_o());
4799 set_control(ideal.ctrl());
4800 }
4801
4802 void GraphKit::final_sync(IdealKit& ideal) {
4803 // Final sync IdealKit and graphKit.
4804 sync_kit(ideal);
4805 }
4806
4807 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4808 Node* len = load_array_length(load_String_value(str, set_ctrl));
4809 Node* coder = load_String_coder(str, set_ctrl);
4810 // Divide length by 2 if coder is UTF16
4811 return _gvn.transform(new RShiftINode(len, coder));
4812 }
4813
4814 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4815 int value_offset = java_lang_String::value_offset();
4816 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4817 false, nullptr, Type::Offset(0));
4818 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4819 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::BotPTR,
4820 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4821 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4822 Node* p = basic_plus_adr(str, str, value_offset);
4823 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4824 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4825 return must_be_not_null(load, true);
4826 }
4827
4828 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4829 if (!CompactStrings) {
4830 return intcon(java_lang_String::CODER_UTF16);
4831 }
4832 int coder_offset = java_lang_String::coder_offset();
4833 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4834 false, nullptr, Type::Offset(0));
4835 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4836
4837 Node* p = basic_plus_adr(str, str, coder_offset);
4838 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4839 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4840 return load;
4841 }
4842
4843 void GraphKit::store_String_value(Node* str, Node* value) {
4844 int value_offset = java_lang_String::value_offset();
4845 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4846 false, nullptr, Type::Offset(0));
4847 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4848
4849 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4850 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4851 }
4852
4853 void GraphKit::store_String_coder(Node* str, Node* value) {
4854 int coder_offset = java_lang_String::coder_offset();
4855 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4856 false, nullptr, Type::Offset(0));
4857 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4858
4859 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4860 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4861 }
4862
4863 // If input and output memory types differ, capture the whole memory to preserve
4864 // the dependency between preceding and subsequent loads/stores.
4865 // For example, the following program:
4866 // StoreB
4867 // compress_string
4868 // LoadB
4869 // has this memory graph (use->def):
4870 // LoadB -> compress_string -> CharMem
4871 // ... -> StoreB -> ByteMem
4872 // The intrinsic hides the dependency between LoadB and StoreB, causing
4873 // the load to read from memory not containing the result of the StoreB.
4874 // The correct memory graph should look like this:
4875 // LoadB -> compress_string -> MergeMem -> StoreB
4876 Node* GraphKit::capture_memory(const TypePtr*& combined_type, const TypePtr* src_type, const TypePtr* dst_type) {
4877 if (src_type == dst_type) {
4878 // Types are equal, we don't need a MergeMemNode
4879 combined_type = src_type;
4880 return memory(src_type);
4881 }
4882 Node* mem = reset_memory();
4883 set_all_memory(mem);
4884 combined_type = TypePtr::BOTTOM;
4885 return mem;
4886 }
4887
4888 // If dst_type and src_type are different, str may have an anti-dependency with another node
4889 // consuming src_type.
4890 // For example:
4891 // compress_string
4892 // StoreC
4893 // has this memory graph (use->def):
4894 // compress_string -> MergeMem -> CharMem
4895 // StoreC
4896 // The scheduler needs to ensure that compress_string is not executed after StoreC, or it will read
4897 // the wrong memory. For normal loads, the scheduler computes its anti-dependencies to ensure the
4898 // memory it reads from is not killed. Since we do not compute anti-dependencies for
4899 // StrCompressedCopyNode, manually insert a MemBar so the anti-dependency becomes use-def
4900 // dependency:
4901 // StoreC -> MemBar -> MergeMem -> compress_string -> MergeMem -> CharMem
4902 // -------------------------------->
4903 void GraphKit::memory_effect(Node* res_mem, const TypePtr* src_type, const TypePtr* dst_type) {
4904 set_memory(res_mem, dst_type);
4905 if (src_type != dst_type) {
4906 Node* all_mem = reset_memory();
4907 set_all_memory(all_mem);
4908 Node* membar = new MemBarCPUOrderNode(C, C->get_alias_index(src_type), nullptr);
4909 membar->init_req(TypeFunc::Control, control());
4910 membar->init_req(TypeFunc::Memory, all_mem);
4911 membar = _gvn.transform(membar);
4912 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
4913 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)), src_type);
4914 }
4915 }
4916
4917 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4918 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4919 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4920 const TypePtr* dst_type = TypeAryPtr::BYTES;
4921 const TypePtr* adr_type;
4922 Node* mem = capture_memory(adr_type, src_type, dst_type);
4923 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, adr_type, src, dst, count);
4924 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4925 memory_effect(res_mem, src_type, dst_type);
4926 return str;
4927 }
4928
4929 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4930 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4931 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4932 const TypePtr* src_type = TypeAryPtr::BYTES;
4933 const TypePtr* adr_type;
4934 Node* mem = capture_memory(adr_type, src_type, dst_type);
4935 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, adr_type, src, dst, count);
4936 Node* res_mem = _gvn.transform(str);
4937 memory_effect(res_mem, src_type, dst_type);
4938 }
4939
4940 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4941 /**
4942 * int i_char = start;
4943 * for (int i_byte = 0; i_byte < count; i_byte++) {
4944 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4945 * }
4946 */
4947 add_parse_predicates();
4948 C->set_has_loops(true);
4949
4950 RegionNode* head = new RegionNode(3);
4951 head->init_req(1, control());
4952 gvn().set_type(head, Type::CONTROL);
4953 record_for_igvn(head);
4954
4955 Node* i_byte = new PhiNode(head, TypeInt::INT);
4956 i_byte->init_req(1, intcon(0));
4957 gvn().set_type(i_byte, TypeInt::INT);
4958 record_for_igvn(i_byte);
4959
4960 Node* i_char = new PhiNode(head, TypeInt::INT);
4961 i_char->init_req(1, start);
4962 gvn().set_type(i_char, TypeInt::INT);
4963 record_for_igvn(i_char);
4964
4965 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4966 gvn().set_type(mem, Type::MEMORY);
4967 record_for_igvn(mem);
4968 set_control(head);
4969 set_memory(mem, TypeAryPtr::BYTES);
4970 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4971 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4972 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4973 false, true /* mismatched */);
4974
4975 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4976 head->init_req(2, IfTrue(iff));
4977 mem->init_req(2, st);
4978 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4979 i_char->init_req(2, AddI(i_char, intcon(2)));
4980
4981 set_control(IfFalse(iff));
4982 set_memory(st, TypeAryPtr::BYTES);
4983 }
4984
4985 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4986 if (!field->is_constant()) {
4987 return nullptr; // Field not marked as constant.
4988 }
4989 ciInstance* holder = nullptr;
4990 if (!field->is_static()) {
4991 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4992 if (const_oop != nullptr && const_oop->is_instance()) {
4993 holder = const_oop->as_instance();
4994 }
4995 }
4996 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4997 /*is_unsigned_load=*/false);
4998 if (con_type != nullptr) {
4999 Node* con = makecon(con_type);
5000 if (field->type()->is_inlinetype()) {
5001 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
5002 } else if (con_type->is_inlinetypeptr()) {
5003 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
5004 }
5005 return con;
5006 }
5007 return nullptr;
5008 }
5009
5010 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
5011 const Type* obj_type = obj->bottom_type();
5012 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
5013 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
5014 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
5015 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
5016 obj = casted_obj;
5017 }
5018 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
5019 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
5020 }
5021 return obj;
5022 }