1 /*
2 * Copyright (c) 2001, 2024, 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 "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "asm/register.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subtypenode.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/bitMap.inline.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/growableArray.hpp"
53
54 //----------------------------GraphKit-----------------------------------------
55 // Main utility constructor.
56 GraphKit::GraphKit(JVMState* jvms)
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn()),
60 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
61 {
62 _exceptions = jvms->map()->next_exception();
63 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
64 set_jvms(jvms);
65 }
66
67 // Private constructor for parser.
68 GraphKit::GraphKit()
69 : Phase(Phase::Parser),
70 _env(C->env()),
71 _gvn(*C->initial_gvn()),
72 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
73 {
74 _exceptions = nullptr;
75 set_map(nullptr);
76 debug_only(_sp = -99);
77 debug_only(set_bci(-99));
78 }
79
80
81
82 //---------------------------clean_stack---------------------------------------
83 // Clear away rubbish from the stack area of the JVM state.
84 // This destroys any arguments that may be waiting on the stack.
85 void GraphKit::clean_stack(int from_sp) {
86 SafePointNode* map = this->map();
87 JVMState* jvms = this->jvms();
88 int stk_size = jvms->stk_size();
89 int stkoff = jvms->stkoff();
90 Node* top = this->top();
91 for (int i = from_sp; i < stk_size; i++) {
92 if (map->in(stkoff + i) != top) {
93 map->set_req(stkoff + i, top);
94 }
95 }
96 }
97
98
99 //--------------------------------sync_jvms-----------------------------------
100 // Make sure our current jvms agrees with our parse state.
101 JVMState* GraphKit::sync_jvms() const {
102 JVMState* jvms = this->jvms();
103 jvms->set_bci(bci()); // Record the new bci in the JVMState
104 jvms->set_sp(sp()); // Record the new sp in the JVMState
105 assert(jvms_in_sync(), "jvms is now in sync");
106 return jvms;
107 }
108
109 //--------------------------------sync_jvms_for_reexecute---------------------
110 // Make sure our current jvms agrees with our parse state. This version
111 // uses the reexecute_sp for reexecuting bytecodes.
112 JVMState* GraphKit::sync_jvms_for_reexecute() {
113 JVMState* jvms = this->jvms();
114 jvms->set_bci(bci()); // Record the new bci in the JVMState
115 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
116 return jvms;
117 }
118
119 #ifdef ASSERT
120 bool GraphKit::jvms_in_sync() const {
121 Parse* parse = is_Parse();
122 if (parse == nullptr) {
123 if (bci() != jvms()->bci()) return false;
124 if (sp() != (int)jvms()->sp()) return false;
125 return true;
126 }
127 if (jvms()->method() != parse->method()) return false;
128 if (jvms()->bci() != parse->bci()) return false;
129 int jvms_sp = jvms()->sp();
130 if (jvms_sp != parse->sp()) return false;
131 int jvms_depth = jvms()->depth();
132 if (jvms_depth != parse->depth()) return false;
133 return true;
134 }
135
136 // Local helper checks for special internal merge points
137 // used to accumulate and merge exception states.
138 // They are marked by the region's in(0) edge being the map itself.
139 // Such merge points must never "escape" into the parser at large,
140 // until they have been handed to gvn.transform.
141 static bool is_hidden_merge(Node* reg) {
142 if (reg == nullptr) return false;
143 if (reg->is_Phi()) {
144 reg = reg->in(0);
145 if (reg == nullptr) return false;
146 }
147 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
148 }
149
150 void GraphKit::verify_map() const {
151 if (map() == nullptr) return; // null map is OK
152 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
153 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
154 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
155 }
156
157 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
158 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
159 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
160 }
161 #endif
162
163 //---------------------------stop_and_kill_map---------------------------------
164 // Set _map to null, signalling a stop to further bytecode execution.
165 // First smash the current map's control to a constant, to mark it dead.
166 void GraphKit::stop_and_kill_map() {
167 SafePointNode* dead_map = stop();
168 if (dead_map != nullptr) {
169 dead_map->disconnect_inputs(C); // Mark the map as killed.
170 assert(dead_map->is_killed(), "must be so marked");
171 }
172 }
173
174
175 //--------------------------------stopped--------------------------------------
176 // Tell if _map is null, or control is top.
177 bool GraphKit::stopped() {
178 if (map() == nullptr) return true;
179 else if (control() == top()) return true;
180 else return false;
181 }
182
183
184 //-----------------------------has_exception_handler----------------------------------
185 // Tell if this method or any caller method has exception handlers.
186 bool GraphKit::has_exception_handler() {
187 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
188 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
189 return true;
190 }
191 }
192 return false;
193 }
194
195 //------------------------------save_ex_oop------------------------------------
196 // Save an exception without blowing stack contents or other JVM state.
197 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
198 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
199 ex_map->add_req(ex_oop);
200 debug_only(verify_exception_state(ex_map));
201 }
202
203 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
204 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
205 Node* ex_oop = ex_map->in(ex_map->req()-1);
206 if (clear_it) ex_map->del_req(ex_map->req()-1);
207 return ex_oop;
208 }
209
210 //-----------------------------saved_ex_oop------------------------------------
211 // Recover a saved exception from its map.
212 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
213 return common_saved_ex_oop(ex_map, false);
214 }
215
216 //--------------------------clear_saved_ex_oop---------------------------------
217 // Erase a previously saved exception from its map.
218 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
219 return common_saved_ex_oop(ex_map, true);
220 }
221
222 #ifdef ASSERT
223 //---------------------------has_saved_ex_oop----------------------------------
224 // Erase a previously saved exception from its map.
225 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
226 return ex_map->req() == ex_map->jvms()->endoff()+1;
227 }
228 #endif
229
230 //-------------------------make_exception_state--------------------------------
231 // Turn the current JVM state into an exception state, appending the ex_oop.
232 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
233 sync_jvms();
234 SafePointNode* ex_map = stop(); // do not manipulate this map any more
235 set_saved_ex_oop(ex_map, ex_oop);
236 return ex_map;
237 }
238
239
240 //--------------------------add_exception_state--------------------------------
241 // Add an exception to my list of exceptions.
242 void GraphKit::add_exception_state(SafePointNode* ex_map) {
243 if (ex_map == nullptr || ex_map->control() == top()) {
244 return;
245 }
246 #ifdef ASSERT
247 verify_exception_state(ex_map);
248 if (has_exceptions()) {
249 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
250 }
251 #endif
252
253 // If there is already an exception of exactly this type, merge with it.
254 // In particular, null-checks and other low-level exceptions common up here.
255 Node* ex_oop = saved_ex_oop(ex_map);
256 const Type* ex_type = _gvn.type(ex_oop);
257 if (ex_oop == top()) {
258 // No action needed.
259 return;
260 }
261 assert(ex_type->isa_instptr(), "exception must be an instance");
262 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
263 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
264 // We check sp also because call bytecodes can generate exceptions
265 // both before and after arguments are popped!
266 if (ex_type2 == ex_type
267 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
268 combine_exception_states(ex_map, e2);
269 return;
270 }
271 }
272
273 // No pre-existing exception of the same type. Chain it on the list.
274 push_exception_state(ex_map);
275 }
276
277 //-----------------------add_exception_states_from-----------------------------
278 void GraphKit::add_exception_states_from(JVMState* jvms) {
279 SafePointNode* ex_map = jvms->map()->next_exception();
280 if (ex_map != nullptr) {
281 jvms->map()->set_next_exception(nullptr);
282 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
283 next_map = ex_map->next_exception();
284 ex_map->set_next_exception(nullptr);
285 add_exception_state(ex_map);
286 }
287 }
288 }
289
290 //-----------------------transfer_exceptions_into_jvms-------------------------
291 JVMState* GraphKit::transfer_exceptions_into_jvms() {
292 if (map() == nullptr) {
293 // We need a JVMS to carry the exceptions, but the map has gone away.
294 // Create a scratch JVMS, cloned from any of the exception states...
295 if (has_exceptions()) {
296 _map = _exceptions;
297 _map = clone_map();
298 _map->set_next_exception(nullptr);
299 clear_saved_ex_oop(_map);
300 debug_only(verify_map());
301 } else {
302 // ...or created from scratch
303 JVMState* jvms = new (C) JVMState(_method, nullptr);
304 jvms->set_bci(_bci);
305 jvms->set_sp(_sp);
306 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
307 set_jvms(jvms);
308 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
309 set_all_memory(top());
310 while (map()->req() < jvms->endoff()) map()->add_req(top());
311 }
312 // (This is a kludge, in case you didn't notice.)
313 set_control(top());
314 }
315 JVMState* jvms = sync_jvms();
316 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
317 jvms->map()->set_next_exception(_exceptions);
318 _exceptions = nullptr; // done with this set of exceptions
319 return jvms;
320 }
321
322 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
323 assert(is_hidden_merge(dstphi), "must be a special merge node");
324 assert(is_hidden_merge(srcphi), "must be a special merge node");
325 uint limit = srcphi->req();
326 for (uint i = PhiNode::Input; i < limit; i++) {
327 dstphi->add_req(srcphi->in(i));
328 }
329 }
330 static inline void add_one_req(Node* dstphi, Node* src) {
331 assert(is_hidden_merge(dstphi), "must be a special merge node");
332 assert(!is_hidden_merge(src), "must not be a special merge node");
333 dstphi->add_req(src);
334 }
335
336 //-----------------------combine_exception_states------------------------------
337 // This helper function combines exception states by building phis on a
338 // specially marked state-merging region. These regions and phis are
339 // untransformed, and can build up gradually. The region is marked by
340 // having a control input of its exception map, rather than null. Such
341 // regions do not appear except in this function, and in use_exception_state.
342 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
343 if (failing_internal()) {
344 return; // dying anyway...
345 }
346 JVMState* ex_jvms = ex_map->_jvms;
347 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
348 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
349 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
350 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
351 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
352 assert(ex_map->req() == phi_map->req(), "matching maps");
353 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
354 Node* hidden_merge_mark = root();
355 Node* region = phi_map->control();
356 MergeMemNode* phi_mem = phi_map->merged_memory();
357 MergeMemNode* ex_mem = ex_map->merged_memory();
358 if (region->in(0) != hidden_merge_mark) {
359 // The control input is not (yet) a specially-marked region in phi_map.
360 // Make it so, and build some phis.
361 region = new RegionNode(2);
362 _gvn.set_type(region, Type::CONTROL);
363 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
364 region->init_req(1, phi_map->control());
365 phi_map->set_control(region);
366 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
367 record_for_igvn(io_phi);
368 _gvn.set_type(io_phi, Type::ABIO);
369 phi_map->set_i_o(io_phi);
370 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
371 Node* m = mms.memory();
372 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
373 record_for_igvn(m_phi);
374 _gvn.set_type(m_phi, Type::MEMORY);
375 mms.set_memory(m_phi);
376 }
377 }
378
379 // Either or both of phi_map and ex_map might already be converted into phis.
380 Node* ex_control = ex_map->control();
381 // if there is special marking on ex_map also, we add multiple edges from src
382 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
383 // how wide was the destination phi_map, originally?
384 uint orig_width = region->req();
385
386 if (add_multiple) {
387 add_n_reqs(region, ex_control);
388 add_n_reqs(phi_map->i_o(), ex_map->i_o());
389 } else {
390 // ex_map has no merges, so we just add single edges everywhere
391 add_one_req(region, ex_control);
392 add_one_req(phi_map->i_o(), ex_map->i_o());
393 }
394 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
395 if (mms.is_empty()) {
396 // get a copy of the base memory, and patch some inputs into it
397 const TypePtr* adr_type = mms.adr_type(C);
398 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
399 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
400 mms.set_memory(phi);
401 // Prepare to append interesting stuff onto the newly sliced phi:
402 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
403 }
404 // Append stuff from ex_map:
405 if (add_multiple) {
406 add_n_reqs(mms.memory(), mms.memory2());
407 } else {
408 add_one_req(mms.memory(), mms.memory2());
409 }
410 }
411 uint limit = ex_map->req();
412 for (uint i = TypeFunc::Parms; i < limit; i++) {
413 // Skip everything in the JVMS after tos. (The ex_oop follows.)
414 if (i == tos) i = ex_jvms->monoff();
415 Node* src = ex_map->in(i);
416 Node* dst = phi_map->in(i);
417 if (src != dst) {
418 PhiNode* phi;
419 if (dst->in(0) != region) {
420 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
421 record_for_igvn(phi);
422 _gvn.set_type(phi, phi->type());
423 phi_map->set_req(i, dst);
424 // Prepare to append interesting stuff onto the new phi:
425 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
426 } else {
427 assert(dst->is_Phi(), "nobody else uses a hidden region");
428 phi = dst->as_Phi();
429 }
430 if (add_multiple && src->in(0) == ex_control) {
431 // Both are phis.
432 add_n_reqs(dst, src);
433 } else {
434 while (dst->req() < region->req()) add_one_req(dst, src);
435 }
436 const Type* srctype = _gvn.type(src);
437 if (phi->type() != srctype) {
438 const Type* dsttype = phi->type()->meet_speculative(srctype);
439 if (phi->type() != dsttype) {
440 phi->set_type(dsttype);
441 _gvn.set_type(phi, dsttype);
442 }
443 }
444 }
445 }
446 phi_map->merge_replaced_nodes_with(ex_map);
447 }
448
449 //--------------------------use_exception_state--------------------------------
450 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
451 if (failing_internal()) { stop(); return top(); }
452 Node* region = phi_map->control();
453 Node* hidden_merge_mark = root();
454 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
455 Node* ex_oop = clear_saved_ex_oop(phi_map);
456 if (region->in(0) == hidden_merge_mark) {
457 // Special marking for internal ex-states. Process the phis now.
458 region->set_req(0, region); // now it's an ordinary region
459 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
460 // Note: Setting the jvms also sets the bci and sp.
461 set_control(_gvn.transform(region));
462 uint tos = jvms()->stkoff() + sp();
463 for (uint i = 1; i < tos; i++) {
464 Node* x = phi_map->in(i);
465 if (x->in(0) == region) {
466 assert(x->is_Phi(), "expected a special phi");
467 phi_map->set_req(i, _gvn.transform(x));
468 }
469 }
470 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
471 Node* x = mms.memory();
472 if (x->in(0) == region) {
473 assert(x->is_Phi(), "nobody else uses a hidden region");
474 mms.set_memory(_gvn.transform(x));
475 }
476 }
477 if (ex_oop->in(0) == region) {
478 assert(ex_oop->is_Phi(), "expected a special phi");
479 ex_oop = _gvn.transform(ex_oop);
480 }
481 } else {
482 set_jvms(phi_map->jvms());
483 }
484
485 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
486 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
487 return ex_oop;
488 }
489
490 //---------------------------------java_bc-------------------------------------
491 Bytecodes::Code GraphKit::java_bc() const {
492 ciMethod* method = this->method();
493 int bci = this->bci();
494 if (method != nullptr && bci != InvocationEntryBci)
495 return method->java_code_at_bci(bci);
496 else
497 return Bytecodes::_illegal;
498 }
499
500 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
501 bool must_throw) {
502 // if the exception capability is set, then we will generate code
503 // to check the JavaThread.should_post_on_exceptions flag to see
504 // if we actually need to report exception events (for this
505 // thread). If we don't need to report exception events, we will
506 // take the normal fast path provided by add_exception_events. If
507 // exception event reporting is enabled for this thread, we will
508 // take the uncommon_trap in the BuildCutout below.
509
510 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
511 Node* jthread = _gvn.transform(new ThreadLocalNode());
512 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
513 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
514
515 // Test the should_post_on_exceptions_flag vs. 0
516 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
517 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
518
519 // Branch to slow_path if should_post_on_exceptions_flag was true
520 { BuildCutout unless(this, tst, PROB_MAX);
521 // Do not try anything fancy if we're notifying the VM on every throw.
522 // Cf. case Bytecodes::_athrow in parse2.cpp.
523 uncommon_trap(reason, Deoptimization::Action_none,
524 (ciKlass*)nullptr, (char*)nullptr, must_throw);
525 }
526
527 }
528
529 //------------------------------builtin_throw----------------------------------
530 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
531 bool must_throw = true;
532
533 // If this particular condition has not yet happened at this
534 // bytecode, then use the uncommon trap mechanism, and allow for
535 // a future recompilation if several traps occur here.
536 // If the throw is hot, try to use a more complicated inline mechanism
537 // which keeps execution inside the compiled code.
538 bool treat_throw_as_hot = false;
539 ciMethodData* md = method()->method_data();
540
541 if (ProfileTraps) {
542 if (too_many_traps(reason)) {
543 treat_throw_as_hot = true;
544 }
545 // (If there is no MDO at all, assume it is early in
546 // execution, and that any deopts are part of the
547 // startup transient, and don't need to be remembered.)
548
549 // Also, if there is a local exception handler, treat all throws
550 // as hot if there has been at least one in this method.
551 if (C->trap_count(reason) != 0
552 && method()->method_data()->trap_count(reason) != 0
553 && has_exception_handler()) {
554 treat_throw_as_hot = true;
555 }
556 }
557
558 // If this throw happens frequently, an uncommon trap might cause
559 // a performance pothole. If there is a local exception handler,
560 // and if this particular bytecode appears to be deoptimizing often,
561 // let us handle the throw inline, with a preconstructed instance.
562 // Note: If the deopt count has blown up, the uncommon trap
563 // runtime is going to flush this nmethod, not matter what.
564 if (treat_throw_as_hot && method()->can_omit_stack_trace()) {
565 // If the throw is local, we use a pre-existing instance and
566 // punt on the backtrace. This would lead to a missing backtrace
567 // (a repeat of 4292742) if the backtrace object is ever asked
568 // for its backtrace.
569 // Fixing this remaining case of 4292742 requires some flavor of
570 // escape analysis. Leave that for the future.
571 ciInstance* ex_obj = nullptr;
572 switch (reason) {
573 case Deoptimization::Reason_null_check:
574 ex_obj = env()->NullPointerException_instance();
575 break;
576 case Deoptimization::Reason_div0_check:
577 ex_obj = env()->ArithmeticException_instance();
578 break;
579 case Deoptimization::Reason_range_check:
580 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
581 break;
582 case Deoptimization::Reason_class_check:
583 ex_obj = env()->ClassCastException_instance();
584 break;
585 case Deoptimization::Reason_array_check:
586 ex_obj = env()->ArrayStoreException_instance();
587 break;
588 default:
589 break;
590 }
591 if (failing()) { stop(); return; } // exception allocation might fail
592 if (ex_obj != nullptr) {
593 if (env()->jvmti_can_post_on_exceptions()) {
594 // check if we must post exception events, take uncommon trap if so
595 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
596 // here if should_post_on_exceptions is false
597 // continue on with the normal codegen
598 }
599
600 // Cheat with a preallocated exception object.
601 if (C->log() != nullptr)
602 C->log()->elem("hot_throw preallocated='1' reason='%s'",
603 Deoptimization::trap_reason_name(reason));
604 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
605 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
606
607 // Clear the detail message of the preallocated exception object.
608 // Weblogic sometimes mutates the detail message of exceptions
609 // using reflection.
610 int offset = java_lang_Throwable::get_detailMessage_offset();
611 const TypePtr* adr_typ = ex_con->add_offset(offset);
612
613 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
614 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
615 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
616
617 if (!method()->has_exception_handlers()) {
618 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
619 // This prevents issues with exception handling and late inlining.
620 set_sp(0);
621 clean_stack(0);
622 }
623
624 add_exception_state(make_exception_state(ex_node));
625 return;
626 }
627 }
628
629 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
630 // It won't be much cheaper than bailing to the interp., since we'll
631 // have to pass up all the debug-info, and the runtime will have to
632 // create the stack trace.
633
634 // Usual case: Bail to interpreter.
635 // Reserve the right to recompile if we haven't seen anything yet.
636
637 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
638 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
639 if (treat_throw_as_hot
640 && (method()->method_data()->trap_recompiled_at(bci(), m)
641 || C->too_many_traps(reason))) {
642 // We cannot afford to take more traps here. Suffer in the interpreter.
643 if (C->log() != nullptr)
644 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
645 Deoptimization::trap_reason_name(reason),
646 C->trap_count(reason));
647 action = Deoptimization::Action_none;
648 }
649
650 // "must_throw" prunes the JVM state to include only the stack, if there
651 // are no local exception handlers. This should cut down on register
652 // allocation time and code size, by drastically reducing the number
653 // of in-edges on the call to the uncommon trap.
654
655 uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw);
656 }
657
658
659 //----------------------------PreserveJVMState---------------------------------
660 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
661 debug_only(kit->verify_map());
662 _kit = kit;
663 _map = kit->map(); // preserve the map
664 _sp = kit->sp();
665 kit->set_map(clone_map ? kit->clone_map() : nullptr);
666 #ifdef ASSERT
667 _bci = kit->bci();
668 Parse* parser = kit->is_Parse();
669 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
670 _block = block;
671 #endif
672 }
673 PreserveJVMState::~PreserveJVMState() {
674 GraphKit* kit = _kit;
675 #ifdef ASSERT
676 assert(kit->bci() == _bci, "bci must not shift");
677 Parse* parser = kit->is_Parse();
678 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
679 assert(block == _block, "block must not shift");
680 #endif
681 kit->set_map(_map);
682 kit->set_sp(_sp);
683 }
684
685
686 //-----------------------------BuildCutout-------------------------------------
687 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
688 : PreserveJVMState(kit)
689 {
690 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
691 SafePointNode* outer_map = _map; // preserved map is caller's
692 SafePointNode* inner_map = kit->map();
693 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
694 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
695 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
696 }
697 BuildCutout::~BuildCutout() {
698 GraphKit* kit = _kit;
699 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
700 }
701
702 //---------------------------PreserveReexecuteState----------------------------
703 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
704 assert(!kit->stopped(), "must call stopped() before");
705 _kit = kit;
706 _sp = kit->sp();
707 _reexecute = kit->jvms()->_reexecute;
708 }
709 PreserveReexecuteState::~PreserveReexecuteState() {
710 if (_kit->stopped()) return;
711 _kit->jvms()->_reexecute = _reexecute;
712 _kit->set_sp(_sp);
713 }
714
715 //------------------------------clone_map--------------------------------------
716 // Implementation of PreserveJVMState
717 //
718 // Only clone_map(...) here. If this function is only used in the
719 // PreserveJVMState class we may want to get rid of this extra
720 // function eventually and do it all there.
721
722 SafePointNode* GraphKit::clone_map() {
723 if (map() == nullptr) return nullptr;
724
725 // Clone the memory edge first
726 Node* mem = MergeMemNode::make(map()->memory());
727 gvn().set_type_bottom(mem);
728
729 SafePointNode *clonemap = (SafePointNode*)map()->clone();
730 JVMState* jvms = this->jvms();
731 JVMState* clonejvms = jvms->clone_shallow(C);
732 clonemap->set_memory(mem);
733 clonemap->set_jvms(clonejvms);
734 clonejvms->set_map(clonemap);
735 record_for_igvn(clonemap);
736 gvn().set_type_bottom(clonemap);
737 return clonemap;
738 }
739
740 //-----------------------------destruct_map_clone------------------------------
741 //
742 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
743 // to destruct/free/delete in the exact opposite order as clone_map().
744 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
745 if (sfp == nullptr) return;
746
747 Node* mem = sfp->memory();
748 JVMState* jvms = sfp->jvms();
749
750 if (jvms != nullptr) {
751 delete jvms;
752 }
753
754 remove_for_igvn(sfp);
755 gvn().clear_type(sfp);
756 sfp->destruct(&_gvn);
757
758 if (mem != nullptr) {
759 gvn().clear_type(mem);
760 mem->destruct(&_gvn);
761 }
762 }
763
764 //-----------------------------set_map_clone-----------------------------------
765 void GraphKit::set_map_clone(SafePointNode* m) {
766 _map = m;
767 _map = clone_map();
768 _map->set_next_exception(nullptr);
769 debug_only(verify_map());
770 }
771
772
773 //----------------------------kill_dead_locals---------------------------------
774 // Detect any locals which are known to be dead, and force them to top.
775 void GraphKit::kill_dead_locals() {
776 // Consult the liveness information for the locals. If any
777 // of them are unused, then they can be replaced by top(). This
778 // should help register allocation time and cut down on the size
779 // of the deoptimization information.
780
781 // This call is made from many of the bytecode handling
782 // subroutines called from the Big Switch in do_one_bytecode.
783 // Every bytecode which might include a slow path is responsible
784 // for killing its dead locals. The more consistent we
785 // are about killing deads, the fewer useless phis will be
786 // constructed for them at various merge points.
787
788 // bci can be -1 (InvocationEntryBci). We return the entry
789 // liveness for the method.
790
791 if (method() == nullptr || method()->code_size() == 0) {
792 // We are building a graph for a call to a native method.
793 // All locals are live.
794 return;
795 }
796
797 ResourceMark rm;
798
799 // Consult the liveness information for the locals. If any
800 // of them are unused, then they can be replaced by top(). This
801 // should help register allocation time and cut down on the size
802 // of the deoptimization information.
803 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
804
805 int len = (int)live_locals.size();
806 assert(len <= jvms()->loc_size(), "too many live locals");
807 for (int local = 0; local < len; local++) {
808 if (!live_locals.at(local)) {
809 set_local(local, top());
810 }
811 }
812 }
813
814 #ifdef ASSERT
815 //-------------------------dead_locals_are_killed------------------------------
816 // Return true if all dead locals are set to top in the map.
817 // Used to assert "clean" debug info at various points.
818 bool GraphKit::dead_locals_are_killed() {
819 if (method() == nullptr || method()->code_size() == 0) {
820 // No locals need to be dead, so all is as it should be.
821 return true;
822 }
823
824 // Make sure somebody called kill_dead_locals upstream.
825 ResourceMark rm;
826 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
827 if (jvms->loc_size() == 0) continue; // no locals to consult
828 SafePointNode* map = jvms->map();
829 ciMethod* method = jvms->method();
830 int bci = jvms->bci();
831 if (jvms == this->jvms()) {
832 bci = this->bci(); // it might not yet be synched
833 }
834 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
835 int len = (int)live_locals.size();
836 if (!live_locals.is_valid() || len == 0)
837 // This method is trivial, or is poisoned by a breakpoint.
838 return true;
839 assert(len == jvms->loc_size(), "live map consistent with locals map");
840 for (int local = 0; local < len; local++) {
841 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
842 if (PrintMiscellaneous && (Verbose || WizardMode)) {
843 tty->print_cr("Zombie local %d: ", local);
844 jvms->dump();
845 }
846 return false;
847 }
848 }
849 }
850 return true;
851 }
852
853 #endif //ASSERT
854
855 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
856 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
857 ciMethod* cur_method = jvms->method();
858 int cur_bci = jvms->bci();
859 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
860 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
861 return Interpreter::bytecode_should_reexecute(code) ||
862 (is_anewarray && code == Bytecodes::_multianewarray);
863 // Reexecute _multianewarray bytecode which was replaced with
864 // sequence of [a]newarray. See Parse::do_multianewarray().
865 //
866 // Note: interpreter should not have it set since this optimization
867 // is limited by dimensions and guarded by flag so in some cases
868 // multianewarray() runtime calls will be generated and
869 // the bytecode should not be reexecutes (stack will not be reset).
870 } else {
871 return false;
872 }
873 }
874
875 // Helper function for adding JVMState and debug information to node
876 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
877 // Add the safepoint edges to the call (or other safepoint).
878
879 // Make sure dead locals are set to top. This
880 // should help register allocation time and cut down on the size
881 // of the deoptimization information.
882 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
883
884 // Walk the inline list to fill in the correct set of JVMState's
885 // Also fill in the associated edges for each JVMState.
886
887 // If the bytecode needs to be reexecuted we need to put
888 // the arguments back on the stack.
889 const bool should_reexecute = jvms()->should_reexecute();
890 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
891
892 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
893 // undefined if the bci is different. This is normal for Parse but it
894 // should not happen for LibraryCallKit because only one bci is processed.
895 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
896 "in LibraryCallKit the reexecute bit should not change");
897
898 // If we are guaranteed to throw, we can prune everything but the
899 // input to the current bytecode.
900 bool can_prune_locals = false;
901 uint stack_slots_not_pruned = 0;
902 int inputs = 0, depth = 0;
903 if (must_throw) {
904 assert(method() == youngest_jvms->method(), "sanity");
905 if (compute_stack_effects(inputs, depth)) {
906 can_prune_locals = true;
907 stack_slots_not_pruned = inputs;
908 }
909 }
910
911 if (env()->should_retain_local_variables()) {
912 // At any safepoint, this method can get breakpointed, which would
913 // then require an immediate deoptimization.
914 can_prune_locals = false; // do not prune locals
915 stack_slots_not_pruned = 0;
916 }
917
918 // do not scribble on the input jvms
919 JVMState* out_jvms = youngest_jvms->clone_deep(C);
920 call->set_jvms(out_jvms); // Start jvms list for call node
921
922 // For a known set of bytecodes, the interpreter should reexecute them if
923 // deoptimization happens. We set the reexecute state for them here
924 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
925 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
926 #ifdef ASSERT
927 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
928 assert(method() == youngest_jvms->method(), "sanity");
929 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
930 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
931 #endif // ASSERT
932 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
933 }
934
935 // Presize the call:
936 DEBUG_ONLY(uint non_debug_edges = call->req());
937 call->add_req_batch(top(), youngest_jvms->debug_depth());
938 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
939
940 // Set up edges so that the call looks like this:
941 // Call [state:] ctl io mem fptr retadr
942 // [parms:] parm0 ... parmN
943 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
944 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
945 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
946 // Note that caller debug info precedes callee debug info.
947
948 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
949 uint debug_ptr = call->req();
950
951 // Loop over the map input edges associated with jvms, add them
952 // to the call node, & reset all offsets to match call node array.
953 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
954 uint debug_end = debug_ptr;
955 uint debug_start = debug_ptr - in_jvms->debug_size();
956 debug_ptr = debug_start; // back up the ptr
957
958 uint p = debug_start; // walks forward in [debug_start, debug_end)
959 uint j, k, l;
960 SafePointNode* in_map = in_jvms->map();
961 out_jvms->set_map(call);
962
963 if (can_prune_locals) {
964 assert(in_jvms->method() == out_jvms->method(), "sanity");
965 // If the current throw can reach an exception handler in this JVMS,
966 // then we must keep everything live that can reach that handler.
967 // As a quick and dirty approximation, we look for any handlers at all.
968 if (in_jvms->method()->has_exception_handlers()) {
969 can_prune_locals = false;
970 }
971 }
972
973 // Add the Locals
974 k = in_jvms->locoff();
975 l = in_jvms->loc_size();
976 out_jvms->set_locoff(p);
977 if (!can_prune_locals) {
978 for (j = 0; j < l; j++)
979 call->set_req(p++, in_map->in(k+j));
980 } else {
981 p += l; // already set to top above by add_req_batch
982 }
983
984 // Add the Expression Stack
985 k = in_jvms->stkoff();
986 l = in_jvms->sp();
987 out_jvms->set_stkoff(p);
988 if (!can_prune_locals) {
989 for (j = 0; j < l; j++)
990 call->set_req(p++, in_map->in(k+j));
991 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
992 // Divide stack into {S0,...,S1}, where S0 is set to top.
993 uint s1 = stack_slots_not_pruned;
994 stack_slots_not_pruned = 0; // for next iteration
995 if (s1 > l) s1 = l;
996 uint s0 = l - s1;
997 p += s0; // skip the tops preinstalled by add_req_batch
998 for (j = s0; j < l; j++)
999 call->set_req(p++, in_map->in(k+j));
1000 } else {
1001 p += l; // already set to top above by add_req_batch
1002 }
1003
1004 // Add the Monitors
1005 k = in_jvms->monoff();
1006 l = in_jvms->mon_size();
1007 out_jvms->set_monoff(p);
1008 for (j = 0; j < l; j++)
1009 call->set_req(p++, in_map->in(k+j));
1010
1011 // Copy any scalar object fields.
1012 k = in_jvms->scloff();
1013 l = in_jvms->scl_size();
1014 out_jvms->set_scloff(p);
1015 for (j = 0; j < l; j++)
1016 call->set_req(p++, in_map->in(k+j));
1017
1018 // Finish the new jvms.
1019 out_jvms->set_endoff(p);
1020
1021 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1022 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1023 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1024 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1025 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1026 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1027
1028 // Update the two tail pointers in parallel.
1029 out_jvms = out_jvms->caller();
1030 in_jvms = in_jvms->caller();
1031 }
1032
1033 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1034
1035 // Test the correctness of JVMState::debug_xxx accessors:
1036 assert(call->jvms()->debug_start() == non_debug_edges, "");
1037 assert(call->jvms()->debug_end() == call->req(), "");
1038 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1039 }
1040
1041 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1042 Bytecodes::Code code = java_bc();
1043 if (code == Bytecodes::_wide) {
1044 code = method()->java_code_at_bci(bci() + 1);
1045 }
1046
1047 if (code != Bytecodes::_illegal) {
1048 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1049 }
1050
1051 auto rsize = [&]() {
1052 assert(code != Bytecodes::_illegal, "code is illegal!");
1053 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1054 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1055 };
1056
1057 switch (code) {
1058 case Bytecodes::_illegal:
1059 return false;
1060
1061 case Bytecodes::_ldc:
1062 case Bytecodes::_ldc_w:
1063 case Bytecodes::_ldc2_w:
1064 inputs = 0;
1065 break;
1066
1067 case Bytecodes::_dup: inputs = 1; break;
1068 case Bytecodes::_dup_x1: inputs = 2; break;
1069 case Bytecodes::_dup_x2: inputs = 3; break;
1070 case Bytecodes::_dup2: inputs = 2; break;
1071 case Bytecodes::_dup2_x1: inputs = 3; break;
1072 case Bytecodes::_dup2_x2: inputs = 4; break;
1073 case Bytecodes::_swap: inputs = 2; break;
1074 case Bytecodes::_arraylength: inputs = 1; break;
1075
1076 case Bytecodes::_getstatic:
1077 case Bytecodes::_putstatic:
1078 case Bytecodes::_getfield:
1079 case Bytecodes::_putfield:
1080 {
1081 bool ignored_will_link;
1082 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1083 int size = field->type()->size();
1084 bool is_get = (depth >= 0), is_static = (depth & 1);
1085 inputs = (is_static ? 0 : 1);
1086 if (is_get) {
1087 depth = size - inputs;
1088 } else {
1089 inputs += size; // putxxx pops the value from the stack
1090 depth = - inputs;
1091 }
1092 }
1093 break;
1094
1095 case Bytecodes::_invokevirtual:
1096 case Bytecodes::_invokespecial:
1097 case Bytecodes::_invokestatic:
1098 case Bytecodes::_invokedynamic:
1099 case Bytecodes::_invokeinterface:
1100 {
1101 bool ignored_will_link;
1102 ciSignature* declared_signature = nullptr;
1103 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1104 assert(declared_signature != nullptr, "cannot be null");
1105 inputs = declared_signature->arg_size_for_bc(code);
1106 int size = declared_signature->return_type()->size();
1107 depth = size - inputs;
1108 }
1109 break;
1110
1111 case Bytecodes::_multianewarray:
1112 {
1113 ciBytecodeStream iter(method());
1114 iter.reset_to_bci(bci());
1115 iter.next();
1116 inputs = iter.get_dimensions();
1117 assert(rsize() == 1, "");
1118 depth = 1 - inputs;
1119 }
1120 break;
1121
1122 case Bytecodes::_ireturn:
1123 case Bytecodes::_lreturn:
1124 case Bytecodes::_freturn:
1125 case Bytecodes::_dreturn:
1126 case Bytecodes::_areturn:
1127 assert(rsize() == -depth, "");
1128 inputs = -depth;
1129 break;
1130
1131 case Bytecodes::_jsr:
1132 case Bytecodes::_jsr_w:
1133 inputs = 0;
1134 depth = 1; // S.B. depth=1, not zero
1135 break;
1136
1137 default:
1138 // bytecode produces a typed result
1139 inputs = rsize() - depth;
1140 assert(inputs >= 0, "");
1141 break;
1142 }
1143
1144 #ifdef ASSERT
1145 // spot check
1146 int outputs = depth + inputs;
1147 assert(outputs >= 0, "sanity");
1148 switch (code) {
1149 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1150 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1151 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1152 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1153 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1154 default: break;
1155 }
1156 #endif //ASSERT
1157
1158 return true;
1159 }
1160
1161
1162
1163 //------------------------------basic_plus_adr---------------------------------
1164 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1165 // short-circuit a common case
1166 if (offset == intcon(0)) return ptr;
1167 return _gvn.transform( new AddPNode(base, ptr, offset) );
1168 }
1169
1170 Node* GraphKit::ConvI2L(Node* offset) {
1171 // short-circuit a common case
1172 jint offset_con = find_int_con(offset, Type::OffsetBot);
1173 if (offset_con != Type::OffsetBot) {
1174 return longcon((jlong) offset_con);
1175 }
1176 return _gvn.transform( new ConvI2LNode(offset));
1177 }
1178
1179 Node* GraphKit::ConvI2UL(Node* offset) {
1180 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1181 if (offset_con != (juint) Type::OffsetBot) {
1182 return longcon((julong) offset_con);
1183 }
1184 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1185 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1186 return _gvn.transform( new AndLNode(conv, mask) );
1187 }
1188
1189 Node* GraphKit::ConvL2I(Node* offset) {
1190 // short-circuit a common case
1191 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1192 if (offset_con != (jlong)Type::OffsetBot) {
1193 return intcon((int) offset_con);
1194 }
1195 return _gvn.transform( new ConvL2INode(offset));
1196 }
1197
1198 //-------------------------load_object_klass-----------------------------------
1199 Node* GraphKit::load_object_klass(Node* obj) {
1200 // Special-case a fresh allocation to avoid building nodes:
1201 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1202 if (akls != nullptr) return akls;
1203 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1204 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1205 }
1206
1207 //-------------------------load_array_length-----------------------------------
1208 Node* GraphKit::load_array_length(Node* array) {
1209 // Special-case a fresh allocation to avoid building nodes:
1210 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1211 Node *alen;
1212 if (alloc == nullptr) {
1213 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1214 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1215 } else {
1216 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1217 }
1218 return alen;
1219 }
1220
1221 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1222 const TypeOopPtr* oop_type,
1223 bool replace_length_in_map) {
1224 Node* length = alloc->Ideal_length();
1225 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1226 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1227 if (ccast != length) {
1228 // do not transform ccast here, it might convert to top node for
1229 // negative array length and break assumptions in parsing stage.
1230 _gvn.set_type_bottom(ccast);
1231 record_for_igvn(ccast);
1232 if (replace_length_in_map) {
1233 replace_in_map(length, ccast);
1234 }
1235 return ccast;
1236 }
1237 }
1238 return length;
1239 }
1240
1241 //------------------------------do_null_check----------------------------------
1242 // Helper function to do a null pointer check. Returned value is
1243 // the incoming address with null casted away. You are allowed to use the
1244 // not-null value only if you are control dependent on the test.
1245 #ifndef PRODUCT
1246 extern uint explicit_null_checks_inserted,
1247 explicit_null_checks_elided;
1248 #endif
1249 Node* GraphKit::null_check_common(Node* value, BasicType type,
1250 // optional arguments for variations:
1251 bool assert_null,
1252 Node* *null_control,
1253 bool speculative) {
1254 assert(!assert_null || null_control == nullptr, "not both at once");
1255 if (stopped()) return top();
1256 NOT_PRODUCT(explicit_null_checks_inserted++);
1257
1258 // Construct null check
1259 Node *chk = nullptr;
1260 switch(type) {
1261 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1262 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1263 case T_ARRAY : // fall through
1264 type = T_OBJECT; // simplify further tests
1265 case T_OBJECT : {
1266 const Type *t = _gvn.type( value );
1267
1268 const TypeOopPtr* tp = t->isa_oopptr();
1269 if (tp != nullptr && !tp->is_loaded()
1270 // Only for do_null_check, not any of its siblings:
1271 && !assert_null && null_control == nullptr) {
1272 // Usually, any field access or invocation on an unloaded oop type
1273 // will simply fail to link, since the statically linked class is
1274 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1275 // the static class is loaded but the sharper oop type is not.
1276 // Rather than checking for this obscure case in lots of places,
1277 // we simply observe that a null check on an unloaded class
1278 // will always be followed by a nonsense operation, so we
1279 // can just issue the uncommon trap here.
1280 // Our access to the unloaded class will only be correct
1281 // after it has been loaded and initialized, which requires
1282 // a trip through the interpreter.
1283 ciKlass* klass = tp->unloaded_klass();
1284 #ifndef PRODUCT
1285 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1286 #endif
1287 uncommon_trap(Deoptimization::Reason_unloaded,
1288 Deoptimization::Action_reinterpret,
1289 klass, "!loaded");
1290 return top();
1291 }
1292
1293 if (assert_null) {
1294 // See if the type is contained in NULL_PTR.
1295 // If so, then the value is already null.
1296 if (t->higher_equal(TypePtr::NULL_PTR)) {
1297 NOT_PRODUCT(explicit_null_checks_elided++);
1298 return value; // Elided null assert quickly!
1299 }
1300 } else {
1301 // See if mixing in the null pointer changes type.
1302 // If so, then the null pointer was not allowed in the original
1303 // type. In other words, "value" was not-null.
1304 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1305 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1306 NOT_PRODUCT(explicit_null_checks_elided++);
1307 return value; // Elided null check quickly!
1308 }
1309 }
1310 chk = new CmpPNode( value, null() );
1311 break;
1312 }
1313
1314 default:
1315 fatal("unexpected type: %s", type2name(type));
1316 }
1317 assert(chk != nullptr, "sanity check");
1318 chk = _gvn.transform(chk);
1319
1320 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1321 BoolNode *btst = new BoolNode( chk, btest);
1322 Node *tst = _gvn.transform( btst );
1323
1324 //-----------
1325 // if peephole optimizations occurred, a prior test existed.
1326 // If a prior test existed, maybe it dominates as we can avoid this test.
1327 if (tst != btst && type == T_OBJECT) {
1328 // At this point we want to scan up the CFG to see if we can
1329 // find an identical test (and so avoid this test altogether).
1330 Node *cfg = control();
1331 int depth = 0;
1332 while( depth < 16 ) { // Limit search depth for speed
1333 if( cfg->Opcode() == Op_IfTrue &&
1334 cfg->in(0)->in(1) == tst ) {
1335 // Found prior test. Use "cast_not_null" to construct an identical
1336 // CastPP (and hence hash to) as already exists for the prior test.
1337 // Return that casted value.
1338 if (assert_null) {
1339 replace_in_map(value, null());
1340 return null(); // do not issue the redundant test
1341 }
1342 Node *oldcontrol = control();
1343 set_control(cfg);
1344 Node *res = cast_not_null(value);
1345 set_control(oldcontrol);
1346 NOT_PRODUCT(explicit_null_checks_elided++);
1347 return res;
1348 }
1349 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1350 if (cfg == nullptr) break; // Quit at region nodes
1351 depth++;
1352 }
1353 }
1354
1355 //-----------
1356 // Branch to failure if null
1357 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1358 Deoptimization::DeoptReason reason;
1359 if (assert_null) {
1360 reason = Deoptimization::reason_null_assert(speculative);
1361 } else if (type == T_OBJECT) {
1362 reason = Deoptimization::reason_null_check(speculative);
1363 } else {
1364 reason = Deoptimization::Reason_div0_check;
1365 }
1366 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1367 // ciMethodData::has_trap_at will return a conservative -1 if any
1368 // must-be-null assertion has failed. This could cause performance
1369 // problems for a method after its first do_null_assert failure.
1370 // Consider using 'Reason_class_check' instead?
1371
1372 // To cause an implicit null check, we set the not-null probability
1373 // to the maximum (PROB_MAX). For an explicit check the probability
1374 // is set to a smaller value.
1375 if (null_control != nullptr || too_many_traps(reason)) {
1376 // probability is less likely
1377 ok_prob = PROB_LIKELY_MAG(3);
1378 } else if (!assert_null &&
1379 (ImplicitNullCheckThreshold > 0) &&
1380 method() != nullptr &&
1381 (method()->method_data()->trap_count(reason)
1382 >= (uint)ImplicitNullCheckThreshold)) {
1383 ok_prob = PROB_LIKELY_MAG(3);
1384 }
1385
1386 if (null_control != nullptr) {
1387 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1388 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1389 set_control( _gvn.transform( new IfTrueNode(iff)));
1390 #ifndef PRODUCT
1391 if (null_true == top()) {
1392 explicit_null_checks_elided++;
1393 }
1394 #endif
1395 (*null_control) = null_true;
1396 } else {
1397 BuildCutout unless(this, tst, ok_prob);
1398 // Check for optimizer eliding test at parse time
1399 if (stopped()) {
1400 // Failure not possible; do not bother making uncommon trap.
1401 NOT_PRODUCT(explicit_null_checks_elided++);
1402 } else if (assert_null) {
1403 uncommon_trap(reason,
1404 Deoptimization::Action_make_not_entrant,
1405 nullptr, "assert_null");
1406 } else {
1407 replace_in_map(value, zerocon(type));
1408 builtin_throw(reason);
1409 }
1410 }
1411
1412 // Must throw exception, fall-thru not possible?
1413 if (stopped()) {
1414 return top(); // No result
1415 }
1416
1417 if (assert_null) {
1418 // Cast obj to null on this path.
1419 replace_in_map(value, zerocon(type));
1420 return zerocon(type);
1421 }
1422
1423 // Cast obj to not-null on this path, if there is no null_control.
1424 // (If there is a null_control, a non-null value may come back to haunt us.)
1425 if (type == T_OBJECT) {
1426 Node* cast = cast_not_null(value, false);
1427 if (null_control == nullptr || (*null_control) == top())
1428 replace_in_map(value, cast);
1429 value = cast;
1430 }
1431
1432 return value;
1433 }
1434
1435
1436 //------------------------------cast_not_null----------------------------------
1437 // Cast obj to not-null on this path
1438 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1439 const Type *t = _gvn.type(obj);
1440 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1441 // Object is already not-null?
1442 if( t == t_not_null ) return obj;
1443
1444 Node* cast = new CastPPNode(control(), obj,t_not_null);
1445 cast = _gvn.transform( cast );
1446
1447 // Scan for instances of 'obj' in the current JVM mapping.
1448 // These instances are known to be not-null after the test.
1449 if (do_replace_in_map)
1450 replace_in_map(obj, cast);
1451
1452 return cast; // Return casted value
1453 }
1454
1455 // Sometimes in intrinsics, we implicitly know an object is not null
1456 // (there's no actual null check) so we can cast it to not null. In
1457 // the course of optimizations, the input to the cast can become null.
1458 // In that case that data path will die and we need the control path
1459 // to become dead as well to keep the graph consistent. So we have to
1460 // add a check for null for which one branch can't be taken. It uses
1461 // an Opaque4 node that will cause the check to be removed after loop
1462 // opts so the test goes away and the compiled code doesn't execute a
1463 // useless check.
1464 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1465 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1466 return value;
1467 }
1468 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1469 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1470 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1471 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1472 _gvn.set_type(iff, iff->Value(&_gvn));
1473 if (!tst->is_Con()) {
1474 record_for_igvn(iff);
1475 }
1476 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1477 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1478 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1479 C->root()->add_req(halt);
1480 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1481 set_control(if_t);
1482 return cast_not_null(value, do_replace_in_map);
1483 }
1484
1485
1486 //--------------------------replace_in_map-------------------------------------
1487 void GraphKit::replace_in_map(Node* old, Node* neww) {
1488 if (old == neww) {
1489 return;
1490 }
1491
1492 map()->replace_edge(old, neww);
1493
1494 // Note: This operation potentially replaces any edge
1495 // on the map. This includes locals, stack, and monitors
1496 // of the current (innermost) JVM state.
1497
1498 // don't let inconsistent types from profiling escape this
1499 // method
1500
1501 const Type* told = _gvn.type(old);
1502 const Type* tnew = _gvn.type(neww);
1503
1504 if (!tnew->higher_equal(told)) {
1505 return;
1506 }
1507
1508 map()->record_replaced_node(old, neww);
1509 }
1510
1511
1512 //=============================================================================
1513 //--------------------------------memory---------------------------------------
1514 Node* GraphKit::memory(uint alias_idx) {
1515 MergeMemNode* mem = merged_memory();
1516 Node* p = mem->memory_at(alias_idx);
1517 assert(p != mem->empty_memory(), "empty");
1518 _gvn.set_type(p, Type::MEMORY); // must be mapped
1519 return p;
1520 }
1521
1522 //-----------------------------reset_memory------------------------------------
1523 Node* GraphKit::reset_memory() {
1524 Node* mem = map()->memory();
1525 // do not use this node for any more parsing!
1526 debug_only( map()->set_memory((Node*)nullptr) );
1527 return _gvn.transform( mem );
1528 }
1529
1530 //------------------------------set_all_memory---------------------------------
1531 void GraphKit::set_all_memory(Node* newmem) {
1532 Node* mergemem = MergeMemNode::make(newmem);
1533 gvn().set_type_bottom(mergemem);
1534 map()->set_memory(mergemem);
1535 }
1536
1537 //------------------------------set_all_memory_call----------------------------
1538 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1539 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1540 set_all_memory(newmem);
1541 }
1542
1543 //=============================================================================
1544 //
1545 // parser factory methods for MemNodes
1546 //
1547 // These are layered on top of the factory methods in LoadNode and StoreNode,
1548 // and integrate with the parser's memory state and _gvn engine.
1549 //
1550
1551 // factory methods in "int adr_idx"
1552 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1553 int adr_idx,
1554 MemNode::MemOrd mo,
1555 LoadNode::ControlDependency control_dependency,
1556 bool require_atomic_access,
1557 bool unaligned,
1558 bool mismatched,
1559 bool unsafe,
1560 uint8_t barrier_data) {
1561 assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1562 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1563 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1564 debug_only(adr_type = C->get_adr_type(adr_idx));
1565 Node* mem = memory(adr_idx);
1566 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1567 ld = _gvn.transform(ld);
1568 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1569 // Improve graph before escape analysis and boxing elimination.
1570 record_for_igvn(ld);
1571 if (ld->is_DecodeN()) {
1572 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1573 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1574 // a Phi). Recording such cases is still perfectly sound, but may be
1575 // unnecessary and result in some minor IGVN overhead.
1576 record_for_igvn(ld->in(1));
1577 }
1578 }
1579 return ld;
1580 }
1581
1582 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1583 int adr_idx,
1584 MemNode::MemOrd mo,
1585 bool require_atomic_access,
1586 bool unaligned,
1587 bool mismatched,
1588 bool unsafe,
1589 int barrier_data) {
1590 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1591 assert(adr_idx == C->get_alias_index(_gvn.type(adr)->isa_ptr()), "slice of address and input slice don't match");
1592 const TypePtr* adr_type = nullptr;
1593 debug_only(adr_type = C->get_adr_type(adr_idx));
1594 Node *mem = memory(adr_idx);
1595 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1596 if (unaligned) {
1597 st->as_Store()->set_unaligned_access();
1598 }
1599 if (mismatched) {
1600 st->as_Store()->set_mismatched_access();
1601 }
1602 if (unsafe) {
1603 st->as_Store()->set_unsafe_access();
1604 }
1605 st->as_Store()->set_barrier_data(barrier_data);
1606 st = _gvn.transform(st);
1607 set_memory(st, adr_idx);
1608 // Back-to-back stores can only remove intermediate store with DU info
1609 // so push on worklist for optimizer.
1610 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1611 record_for_igvn(st);
1612
1613 return st;
1614 }
1615
1616 Node* GraphKit::access_store_at(Node* obj,
1617 Node* adr,
1618 const TypePtr* adr_type,
1619 Node* val,
1620 const Type* val_type,
1621 BasicType bt,
1622 DecoratorSet decorators) {
1623 // Transformation of a value which could be null pointer (CastPP #null)
1624 // could be delayed during Parse (for example, in adjust_map_after_if()).
1625 // Execute transformation here to avoid barrier generation in such case.
1626 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1627 val = _gvn.makecon(TypePtr::NULL_PTR);
1628 }
1629
1630 if (stopped()) {
1631 return top(); // Dead path ?
1632 }
1633
1634 assert(val != nullptr, "not dead path");
1635
1636 C2AccessValuePtr addr(adr, adr_type);
1637 C2AccessValue value(val, val_type);
1638 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1639 if (access.is_raw()) {
1640 return _barrier_set->BarrierSetC2::store_at(access, value);
1641 } else {
1642 return _barrier_set->store_at(access, value);
1643 }
1644 }
1645
1646 Node* GraphKit::access_load_at(Node* obj, // containing obj
1647 Node* adr, // actual address to store val at
1648 const TypePtr* adr_type,
1649 const Type* val_type,
1650 BasicType bt,
1651 DecoratorSet decorators) {
1652 if (stopped()) {
1653 return top(); // Dead path ?
1654 }
1655
1656 C2AccessValuePtr addr(adr, adr_type);
1657 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1658 if (access.is_raw()) {
1659 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1660 } else {
1661 return _barrier_set->load_at(access, val_type);
1662 }
1663 }
1664
1665 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1666 const Type* val_type,
1667 BasicType bt,
1668 DecoratorSet decorators) {
1669 if (stopped()) {
1670 return top(); // Dead path ?
1671 }
1672
1673 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1674 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1675 if (access.is_raw()) {
1676 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1677 } else {
1678 return _barrier_set->load_at(access, val_type);
1679 }
1680 }
1681
1682 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1683 Node* adr,
1684 const TypePtr* adr_type,
1685 int alias_idx,
1686 Node* expected_val,
1687 Node* new_val,
1688 const Type* value_type,
1689 BasicType bt,
1690 DecoratorSet decorators) {
1691 C2AccessValuePtr addr(adr, adr_type);
1692 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1693 bt, obj, addr, alias_idx);
1694 if (access.is_raw()) {
1695 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1696 } else {
1697 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1698 }
1699 }
1700
1701 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1702 Node* adr,
1703 const TypePtr* adr_type,
1704 int alias_idx,
1705 Node* expected_val,
1706 Node* new_val,
1707 const Type* value_type,
1708 BasicType bt,
1709 DecoratorSet decorators) {
1710 C2AccessValuePtr addr(adr, adr_type);
1711 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1712 bt, obj, addr, alias_idx);
1713 if (access.is_raw()) {
1714 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1715 } else {
1716 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1717 }
1718 }
1719
1720 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1721 Node* adr,
1722 const TypePtr* adr_type,
1723 int alias_idx,
1724 Node* new_val,
1725 const Type* value_type,
1726 BasicType bt,
1727 DecoratorSet decorators) {
1728 C2AccessValuePtr addr(adr, adr_type);
1729 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1730 bt, obj, addr, alias_idx);
1731 if (access.is_raw()) {
1732 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1733 } else {
1734 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1735 }
1736 }
1737
1738 Node* GraphKit::access_atomic_add_at(Node* obj,
1739 Node* adr,
1740 const TypePtr* adr_type,
1741 int alias_idx,
1742 Node* new_val,
1743 const Type* value_type,
1744 BasicType bt,
1745 DecoratorSet decorators) {
1746 C2AccessValuePtr addr(adr, adr_type);
1747 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1748 if (access.is_raw()) {
1749 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1750 } else {
1751 return _barrier_set->atomic_add_at(access, new_val, value_type);
1752 }
1753 }
1754
1755 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1756 return _barrier_set->clone(this, src, dst, size, is_array);
1757 }
1758
1759 //-------------------------array_element_address-------------------------
1760 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1761 const TypeInt* sizetype, Node* ctrl) {
1762 uint shift = exact_log2(type2aelembytes(elembt));
1763 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1764
1765 // short-circuit a common case (saves lots of confusing waste motion)
1766 jint idx_con = find_int_con(idx, -1);
1767 if (idx_con >= 0) {
1768 intptr_t offset = header + ((intptr_t)idx_con << shift);
1769 return basic_plus_adr(ary, offset);
1770 }
1771
1772 // must be correct type for alignment purposes
1773 Node* base = basic_plus_adr(ary, header);
1774 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1775 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1776 return basic_plus_adr(ary, base, scale);
1777 }
1778
1779 //-------------------------load_array_element-------------------------
1780 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1781 const Type* elemtype = arytype->elem();
1782 BasicType elembt = elemtype->array_element_basic_type();
1783 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1784 if (elembt == T_NARROWOOP) {
1785 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1786 }
1787 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1788 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1789 return ld;
1790 }
1791
1792 //-------------------------set_arguments_for_java_call-------------------------
1793 // Arguments (pre-popped from the stack) are taken from the JVMS.
1794 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1795 // Add the call arguments:
1796 uint nargs = call->method()->arg_size();
1797 for (uint i = 0; i < nargs; i++) {
1798 Node* arg = argument(i);
1799 call->init_req(i + TypeFunc::Parms, arg);
1800 }
1801 }
1802
1803 //---------------------------set_edges_for_java_call---------------------------
1804 // Connect a newly created call into the current JVMS.
1805 // A return value node (if any) is returned from set_edges_for_java_call.
1806 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1807
1808 // Add the predefined inputs:
1809 call->init_req( TypeFunc::Control, control() );
1810 call->init_req( TypeFunc::I_O , i_o() );
1811 call->init_req( TypeFunc::Memory , reset_memory() );
1812 call->init_req( TypeFunc::FramePtr, frameptr() );
1813 call->init_req( TypeFunc::ReturnAdr, top() );
1814
1815 add_safepoint_edges(call, must_throw);
1816
1817 Node* xcall = _gvn.transform(call);
1818
1819 if (xcall == top()) {
1820 set_control(top());
1821 return;
1822 }
1823 assert(xcall == call, "call identity is stable");
1824
1825 // Re-use the current map to produce the result.
1826
1827 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1828 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1829 set_all_memory_call(xcall, separate_io_proj);
1830
1831 //return xcall; // no need, caller already has it
1832 }
1833
1834 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1835 if (stopped()) return top(); // maybe the call folded up?
1836
1837 // Capture the return value, if any.
1838 Node* ret;
1839 if (call->method() == nullptr ||
1840 call->method()->return_type()->basic_type() == T_VOID)
1841 ret = top();
1842 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1843
1844 // Note: Since any out-of-line call can produce an exception,
1845 // we always insert an I_O projection from the call into the result.
1846
1847 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1848
1849 if (separate_io_proj) {
1850 // The caller requested separate projections be used by the fall
1851 // through and exceptional paths, so replace the projections for
1852 // the fall through path.
1853 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1854 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1855 }
1856 return ret;
1857 }
1858
1859 //--------------------set_predefined_input_for_runtime_call--------------------
1860 // Reading and setting the memory state is way conservative here.
1861 // The real problem is that I am not doing real Type analysis on memory,
1862 // so I cannot distinguish card mark stores from other stores. Across a GC
1863 // point the Store Barrier and the card mark memory has to agree. I cannot
1864 // have a card mark store and its barrier split across the GC point from
1865 // either above or below. Here I get that to happen by reading ALL of memory.
1866 // A better answer would be to separate out card marks from other memory.
1867 // For now, return the input memory state, so that it can be reused
1868 // after the call, if this call has restricted memory effects.
1869 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1870 // Set fixed predefined input arguments
1871 Node* memory = reset_memory();
1872 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1873 call->init_req( TypeFunc::Control, control() );
1874 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1875 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1876 call->init_req( TypeFunc::FramePtr, frameptr() );
1877 call->init_req( TypeFunc::ReturnAdr, top() );
1878 return memory;
1879 }
1880
1881 //-------------------set_predefined_output_for_runtime_call--------------------
1882 // Set control and memory (not i_o) from the call.
1883 // If keep_mem is not null, use it for the output state,
1884 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1885 // If hook_mem is null, this call produces no memory effects at all.
1886 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1887 // then only that memory slice is taken from the call.
1888 // In the last case, we must put an appropriate memory barrier before
1889 // the call, so as to create the correct anti-dependencies on loads
1890 // preceding the call.
1891 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1892 Node* keep_mem,
1893 const TypePtr* hook_mem) {
1894 // no i/o
1895 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1896 if (keep_mem) {
1897 // First clone the existing memory state
1898 set_all_memory(keep_mem);
1899 if (hook_mem != nullptr) {
1900 // Make memory for the call
1901 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1902 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1903 // We also use hook_mem to extract specific effects from arraycopy stubs.
1904 set_memory(mem, hook_mem);
1905 }
1906 // ...else the call has NO memory effects.
1907
1908 // Make sure the call advertises its memory effects precisely.
1909 // This lets us build accurate anti-dependences in gcm.cpp.
1910 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1911 "call node must be constructed correctly");
1912 } else {
1913 assert(hook_mem == nullptr, "");
1914 // This is not a "slow path" call; all memory comes from the call.
1915 set_all_memory_call(call);
1916 }
1917 }
1918
1919 // Keep track of MergeMems feeding into other MergeMems
1920 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1921 if (!mem->is_MergeMem()) {
1922 return;
1923 }
1924 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1925 Node* use = i.get();
1926 if (use->is_MergeMem()) {
1927 wl.push(use);
1928 }
1929 }
1930 }
1931
1932 // Replace the call with the current state of the kit.
1933 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1934 JVMState* ejvms = nullptr;
1935 if (has_exceptions()) {
1936 ejvms = transfer_exceptions_into_jvms();
1937 }
1938
1939 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1940 ReplacedNodes replaced_nodes_exception;
1941 Node* ex_ctl = top();
1942
1943 SafePointNode* final_state = stop();
1944
1945 // Find all the needed outputs of this call
1946 CallProjections callprojs;
1947 call->extract_projections(&callprojs, true, do_asserts);
1948
1949 Unique_Node_List wl;
1950 Node* init_mem = call->in(TypeFunc::Memory);
1951 Node* final_mem = final_state->in(TypeFunc::Memory);
1952 Node* final_ctl = final_state->in(TypeFunc::Control);
1953 Node* final_io = final_state->in(TypeFunc::I_O);
1954
1955 // Replace all the old call edges with the edges from the inlining result
1956 if (callprojs.fallthrough_catchproj != nullptr) {
1957 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1958 }
1959 if (callprojs.fallthrough_memproj != nullptr) {
1960 if (final_mem->is_MergeMem()) {
1961 // Parser's exits MergeMem was not transformed but may be optimized
1962 final_mem = _gvn.transform(final_mem);
1963 }
1964 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1965 add_mergemem_users_to_worklist(wl, final_mem);
1966 }
1967 if (callprojs.fallthrough_ioproj != nullptr) {
1968 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1969 }
1970
1971 // Replace the result with the new result if it exists and is used
1972 if (callprojs.resproj != nullptr && result != nullptr) {
1973 C->gvn_replace_by(callprojs.resproj, result);
1974 }
1975
1976 if (ejvms == nullptr) {
1977 // No exception edges to simply kill off those paths
1978 if (callprojs.catchall_catchproj != nullptr) {
1979 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1980 }
1981 if (callprojs.catchall_memproj != nullptr) {
1982 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1983 }
1984 if (callprojs.catchall_ioproj != nullptr) {
1985 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1986 }
1987 // Replace the old exception object with top
1988 if (callprojs.exobj != nullptr) {
1989 C->gvn_replace_by(callprojs.exobj, C->top());
1990 }
1991 } else {
1992 GraphKit ekit(ejvms);
1993
1994 // Load my combined exception state into the kit, with all phis transformed:
1995 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1996 replaced_nodes_exception = ex_map->replaced_nodes();
1997
1998 Node* ex_oop = ekit.use_exception_state(ex_map);
1999
2000 if (callprojs.catchall_catchproj != nullptr) {
2001 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2002 ex_ctl = ekit.control();
2003 }
2004 if (callprojs.catchall_memproj != nullptr) {
2005 Node* ex_mem = ekit.reset_memory();
2006 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2007 add_mergemem_users_to_worklist(wl, ex_mem);
2008 }
2009 if (callprojs.catchall_ioproj != nullptr) {
2010 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2011 }
2012
2013 // Replace the old exception object with the newly created one
2014 if (callprojs.exobj != nullptr) {
2015 C->gvn_replace_by(callprojs.exobj, ex_oop);
2016 }
2017 }
2018
2019 // Disconnect the call from the graph
2020 call->disconnect_inputs(C);
2021 C->gvn_replace_by(call, C->top());
2022
2023 // Clean up any MergeMems that feed other MergeMems since the
2024 // optimizer doesn't like that.
2025 while (wl.size() > 0) {
2026 _gvn.transform(wl.pop());
2027 }
2028
2029 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2030 replaced_nodes.apply(C, final_ctl);
2031 }
2032 if (!ex_ctl->is_top() && do_replaced_nodes) {
2033 replaced_nodes_exception.apply(C, ex_ctl);
2034 }
2035 }
2036
2037
2038 //------------------------------increment_counter------------------------------
2039 // for statistics: increment a VM counter by 1
2040
2041 void GraphKit::increment_counter(address counter_addr) {
2042 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2043 increment_counter(adr1);
2044 }
2045
2046 void GraphKit::increment_counter(Node* counter_addr) {
2047 int adr_type = Compile::AliasIdxRaw;
2048 Node* ctrl = control();
2049 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2050 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2051 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2052 }
2053
2054
2055 //------------------------------uncommon_trap----------------------------------
2056 // Bail out to the interpreter in mid-method. Implemented by calling the
2057 // uncommon_trap blob. This helper function inserts a runtime call with the
2058 // right debug info.
2059 Node* GraphKit::uncommon_trap(int trap_request,
2060 ciKlass* klass, const char* comment,
2061 bool must_throw,
2062 bool keep_exact_action) {
2063 if (failing_internal()) {
2064 stop();
2065 }
2066 if (stopped()) return nullptr; // trap reachable?
2067
2068 // Note: If ProfileTraps is true, and if a deopt. actually
2069 // occurs here, the runtime will make sure an MDO exists. There is
2070 // no need to call method()->ensure_method_data() at this point.
2071
2072 // Set the stack pointer to the right value for reexecution:
2073 set_sp(reexecute_sp());
2074
2075 #ifdef ASSERT
2076 if (!must_throw) {
2077 // Make sure the stack has at least enough depth to execute
2078 // the current bytecode.
2079 int inputs, ignored_depth;
2080 if (compute_stack_effects(inputs, ignored_depth)) {
2081 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2082 Bytecodes::name(java_bc()), sp(), inputs);
2083 }
2084 }
2085 #endif
2086
2087 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2088 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2089
2090 switch (action) {
2091 case Deoptimization::Action_maybe_recompile:
2092 case Deoptimization::Action_reinterpret:
2093 // Temporary fix for 6529811 to allow virtual calls to be sure they
2094 // get the chance to go from mono->bi->mega
2095 if (!keep_exact_action &&
2096 Deoptimization::trap_request_index(trap_request) < 0 &&
2097 too_many_recompiles(reason)) {
2098 // This BCI is causing too many recompilations.
2099 if (C->log() != nullptr) {
2100 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2101 Deoptimization::trap_reason_name(reason),
2102 Deoptimization::trap_action_name(action));
2103 }
2104 action = Deoptimization::Action_none;
2105 trap_request = Deoptimization::make_trap_request(reason, action);
2106 } else {
2107 C->set_trap_can_recompile(true);
2108 }
2109 break;
2110 case Deoptimization::Action_make_not_entrant:
2111 C->set_trap_can_recompile(true);
2112 break;
2113 case Deoptimization::Action_none:
2114 case Deoptimization::Action_make_not_compilable:
2115 break;
2116 default:
2117 #ifdef ASSERT
2118 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2119 #endif
2120 break;
2121 }
2122
2123 if (TraceOptoParse) {
2124 char buf[100];
2125 tty->print_cr("Uncommon trap %s at bci:%d",
2126 Deoptimization::format_trap_request(buf, sizeof(buf),
2127 trap_request), bci());
2128 }
2129
2130 CompileLog* log = C->log();
2131 if (log != nullptr) {
2132 int kid = (klass == nullptr)? -1: log->identify(klass);
2133 log->begin_elem("uncommon_trap bci='%d'", bci());
2134 char buf[100];
2135 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2136 trap_request));
2137 if (kid >= 0) log->print(" klass='%d'", kid);
2138 if (comment != nullptr) log->print(" comment='%s'", comment);
2139 log->end_elem();
2140 }
2141
2142 // Make sure any guarding test views this path as very unlikely
2143 Node *i0 = control()->in(0);
2144 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2145 IfNode *iff = i0->as_If();
2146 float f = iff->_prob; // Get prob
2147 if (control()->Opcode() == Op_IfTrue) {
2148 if (f > PROB_UNLIKELY_MAG(4))
2149 iff->_prob = PROB_MIN;
2150 } else {
2151 if (f < PROB_LIKELY_MAG(4))
2152 iff->_prob = PROB_MAX;
2153 }
2154 }
2155
2156 // Clear out dead values from the debug info.
2157 kill_dead_locals();
2158
2159 // Now insert the uncommon trap subroutine call
2160 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2161 const TypePtr* no_memory_effects = nullptr;
2162 // Pass the index of the class to be loaded
2163 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2164 (must_throw ? RC_MUST_THROW : 0),
2165 OptoRuntime::uncommon_trap_Type(),
2166 call_addr, "uncommon_trap", no_memory_effects,
2167 intcon(trap_request));
2168 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2169 "must extract request correctly from the graph");
2170 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2171
2172 call->set_req(TypeFunc::ReturnAdr, returnadr());
2173 // The debug info is the only real input to this call.
2174
2175 // Halt-and-catch fire here. The above call should never return!
2176 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2177 PRODUCT_ONLY(COMMA /*reachable*/false));
2178 _gvn.set_type_bottom(halt);
2179 root()->add_req(halt);
2180
2181 stop_and_kill_map();
2182 return call;
2183 }
2184
2185
2186 //--------------------------just_allocated_object------------------------------
2187 // Report the object that was just allocated.
2188 // It must be the case that there are no intervening safepoints.
2189 // We use this to determine if an object is so "fresh" that
2190 // it does not require card marks.
2191 Node* GraphKit::just_allocated_object(Node* current_control) {
2192 Node* ctrl = current_control;
2193 // Object::<init> is invoked after allocation, most of invoke nodes
2194 // will be reduced, but a region node is kept in parse time, we check
2195 // the pattern and skip the region node if it degraded to a copy.
2196 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2197 ctrl->as_Region()->is_copy()) {
2198 ctrl = ctrl->as_Region()->is_copy();
2199 }
2200 if (C->recent_alloc_ctl() == ctrl) {
2201 return C->recent_alloc_obj();
2202 }
2203 return nullptr;
2204 }
2205
2206
2207 /**
2208 * Record profiling data exact_kls for Node n with the type system so
2209 * that it can propagate it (speculation)
2210 *
2211 * @param n node that the type applies to
2212 * @param exact_kls type from profiling
2213 * @param maybe_null did profiling see null?
2214 *
2215 * @return node with improved type
2216 */
2217 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2218 const Type* current_type = _gvn.type(n);
2219 assert(UseTypeSpeculation, "type speculation must be on");
2220
2221 const TypePtr* speculative = current_type->speculative();
2222
2223 // Should the klass from the profile be recorded in the speculative type?
2224 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2225 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2226 const TypeOopPtr* xtype = tklass->as_instance_type();
2227 assert(xtype->klass_is_exact(), "Should be exact");
2228 // Any reason to believe n is not null (from this profiling or a previous one)?
2229 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2230 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2231 // record the new speculative type's depth
2232 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2233 speculative = speculative->with_inline_depth(jvms()->depth());
2234 } else if (current_type->would_improve_ptr(ptr_kind)) {
2235 // Profiling report that null was never seen so we can change the
2236 // speculative type to non null ptr.
2237 if (ptr_kind == ProfileAlwaysNull) {
2238 speculative = TypePtr::NULL_PTR;
2239 } else {
2240 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2241 const TypePtr* ptr = TypePtr::NOTNULL;
2242 if (speculative != nullptr) {
2243 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2244 } else {
2245 speculative = ptr;
2246 }
2247 }
2248 }
2249
2250 if (speculative != current_type->speculative()) {
2251 // Build a type with a speculative type (what we think we know
2252 // about the type but will need a guard when we use it)
2253 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2254 // We're changing the type, we need a new CheckCast node to carry
2255 // the new type. The new type depends on the control: what
2256 // profiling tells us is only valid from here as far as we can
2257 // tell.
2258 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2259 cast = _gvn.transform(cast);
2260 replace_in_map(n, cast);
2261 n = cast;
2262 }
2263
2264 return n;
2265 }
2266
2267 /**
2268 * Record profiling data from receiver profiling at an invoke with the
2269 * type system so that it can propagate it (speculation)
2270 *
2271 * @param n receiver node
2272 *
2273 * @return node with improved type
2274 */
2275 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2276 if (!UseTypeSpeculation) {
2277 return n;
2278 }
2279 ciKlass* exact_kls = profile_has_unique_klass();
2280 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2281 if ((java_bc() == Bytecodes::_checkcast ||
2282 java_bc() == Bytecodes::_instanceof ||
2283 java_bc() == Bytecodes::_aastore) &&
2284 method()->method_data()->is_mature()) {
2285 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2286 if (data != nullptr) {
2287 if (!data->as_BitData()->null_seen()) {
2288 ptr_kind = ProfileNeverNull;
2289 } else {
2290 assert(data->is_ReceiverTypeData(), "bad profile data type");
2291 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2292 uint i = 0;
2293 for (; i < call->row_limit(); i++) {
2294 ciKlass* receiver = call->receiver(i);
2295 if (receiver != nullptr) {
2296 break;
2297 }
2298 }
2299 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2300 }
2301 }
2302 }
2303 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2304 }
2305
2306 /**
2307 * Record profiling data from argument profiling at an invoke with the
2308 * type system so that it can propagate it (speculation)
2309 *
2310 * @param dest_method target method for the call
2311 * @param bc what invoke bytecode is this?
2312 */
2313 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2314 if (!UseTypeSpeculation) {
2315 return;
2316 }
2317 const TypeFunc* tf = TypeFunc::make(dest_method);
2318 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2319 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2320 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2321 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2322 if (is_reference_type(targ->basic_type())) {
2323 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2324 ciKlass* better_type = nullptr;
2325 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2326 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2327 }
2328 i++;
2329 }
2330 }
2331 }
2332
2333 /**
2334 * Record profiling data from parameter profiling at an invoke with
2335 * the type system so that it can propagate it (speculation)
2336 */
2337 void GraphKit::record_profiled_parameters_for_speculation() {
2338 if (!UseTypeSpeculation) {
2339 return;
2340 }
2341 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2342 if (_gvn.type(local(i))->isa_oopptr()) {
2343 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2344 ciKlass* better_type = nullptr;
2345 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2346 record_profile_for_speculation(local(i), better_type, ptr_kind);
2347 }
2348 j++;
2349 }
2350 }
2351 }
2352
2353 /**
2354 * Record profiling data from return value profiling at an invoke with
2355 * the type system so that it can propagate it (speculation)
2356 */
2357 void GraphKit::record_profiled_return_for_speculation() {
2358 if (!UseTypeSpeculation) {
2359 return;
2360 }
2361 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2362 ciKlass* better_type = nullptr;
2363 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2364 // If profiling reports a single type for the return value,
2365 // feed it to the type system so it can propagate it as a
2366 // speculative type
2367 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2368 }
2369 }
2370
2371 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2372 if (Matcher::strict_fp_requires_explicit_rounding) {
2373 // (Note: TypeFunc::make has a cache that makes this fast.)
2374 const TypeFunc* tf = TypeFunc::make(dest_method);
2375 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2376 for (int j = 0; j < nargs; j++) {
2377 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2378 if (targ->basic_type() == T_DOUBLE) {
2379 // If any parameters are doubles, they must be rounded before
2380 // the call, dprecision_rounding does gvn.transform
2381 Node *arg = argument(j);
2382 arg = dprecision_rounding(arg);
2383 set_argument(j, arg);
2384 }
2385 }
2386 }
2387 }
2388
2389 // rounding for strict float precision conformance
2390 Node* GraphKit::precision_rounding(Node* n) {
2391 if (Matcher::strict_fp_requires_explicit_rounding) {
2392 #ifdef IA32
2393 if (UseSSE == 0) {
2394 return _gvn.transform(new RoundFloatNode(nullptr, n));
2395 }
2396 #else
2397 Unimplemented();
2398 #endif // IA32
2399 }
2400 return n;
2401 }
2402
2403 // rounding for strict double precision conformance
2404 Node* GraphKit::dprecision_rounding(Node *n) {
2405 if (Matcher::strict_fp_requires_explicit_rounding) {
2406 #ifdef IA32
2407 if (UseSSE < 2) {
2408 return _gvn.transform(new RoundDoubleNode(nullptr, n));
2409 }
2410 #else
2411 Unimplemented();
2412 #endif // IA32
2413 }
2414 return n;
2415 }
2416
2417 //=============================================================================
2418 // Generate a fast path/slow path idiom. Graph looks like:
2419 // [foo] indicates that 'foo' is a parameter
2420 //
2421 // [in] null
2422 // \ /
2423 // CmpP
2424 // Bool ne
2425 // If
2426 // / \
2427 // True False-<2>
2428 // / |
2429 // / cast_not_null
2430 // Load | | ^
2431 // [fast_test] | |
2432 // gvn to opt_test | |
2433 // / \ | <1>
2434 // True False |
2435 // | \\ |
2436 // [slow_call] \[fast_result]
2437 // Ctl Val \ \
2438 // | \ \
2439 // Catch <1> \ \
2440 // / \ ^ \ \
2441 // Ex No_Ex | \ \
2442 // | \ \ | \ <2> \
2443 // ... \ [slow_res] | | \ [null_result]
2444 // \ \--+--+--- | |
2445 // \ | / \ | /
2446 // --------Region Phi
2447 //
2448 //=============================================================================
2449 // Code is structured as a series of driver functions all called 'do_XXX' that
2450 // call a set of helper functions. Helper functions first, then drivers.
2451
2452 //------------------------------null_check_oop---------------------------------
2453 // Null check oop. Set null-path control into Region in slot 3.
2454 // Make a cast-not-nullness use the other not-null control. Return cast.
2455 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2456 bool never_see_null,
2457 bool safe_for_replace,
2458 bool speculative) {
2459 // Initial null check taken path
2460 (*null_control) = top();
2461 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2462
2463 // Generate uncommon_trap:
2464 if (never_see_null && (*null_control) != top()) {
2465 // If we see an unexpected null at a check-cast we record it and force a
2466 // recompile; the offending check-cast will be compiled to handle nulls.
2467 // If we see more than one offending BCI, then all checkcasts in the
2468 // method will be compiled to handle nulls.
2469 PreserveJVMState pjvms(this);
2470 set_control(*null_control);
2471 replace_in_map(value, null());
2472 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2473 uncommon_trap(reason,
2474 Deoptimization::Action_make_not_entrant);
2475 (*null_control) = top(); // null path is dead
2476 }
2477 if ((*null_control) == top() && safe_for_replace) {
2478 replace_in_map(value, cast);
2479 }
2480
2481 // Cast away null-ness on the result
2482 return cast;
2483 }
2484
2485 //------------------------------opt_iff----------------------------------------
2486 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2487 // Return slow-path control.
2488 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2489 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2490
2491 // Fast path taken; set region slot 2
2492 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2493 region->init_req(2,fast_taken); // Capture fast-control
2494
2495 // Fast path not-taken, i.e. slow path
2496 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2497 return slow_taken;
2498 }
2499
2500 //-----------------------------make_runtime_call-------------------------------
2501 Node* GraphKit::make_runtime_call(int flags,
2502 const TypeFunc* call_type, address call_addr,
2503 const char* call_name,
2504 const TypePtr* adr_type,
2505 // The following parms are all optional.
2506 // The first null ends the list.
2507 Node* parm0, Node* parm1,
2508 Node* parm2, Node* parm3,
2509 Node* parm4, Node* parm5,
2510 Node* parm6, Node* parm7) {
2511 assert(call_addr != nullptr, "must not call null targets");
2512
2513 // Slow-path call
2514 bool is_leaf = !(flags & RC_NO_LEAF);
2515 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2516 if (call_name == nullptr) {
2517 assert(!is_leaf, "must supply name for leaf");
2518 call_name = OptoRuntime::stub_name(call_addr);
2519 }
2520 CallNode* call;
2521 if (!is_leaf) {
2522 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2523 } else if (flags & RC_NO_FP) {
2524 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2525 } else if (flags & RC_VECTOR){
2526 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2527 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2528 } else {
2529 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2530 }
2531
2532 // The following is similar to set_edges_for_java_call,
2533 // except that the memory effects of the call are restricted to AliasIdxRaw.
2534
2535 // Slow path call has no side-effects, uses few values
2536 bool wide_in = !(flags & RC_NARROW_MEM);
2537 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2538
2539 Node* prev_mem = nullptr;
2540 if (wide_in) {
2541 prev_mem = set_predefined_input_for_runtime_call(call);
2542 } else {
2543 assert(!wide_out, "narrow in => narrow out");
2544 Node* narrow_mem = memory(adr_type);
2545 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2546 }
2547
2548 // Hook each parm in order. Stop looking at the first null.
2549 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2550 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2551 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2552 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2553 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2554 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2555 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2556 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2557 /* close each nested if ===> */ } } } } } } } }
2558 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2559
2560 if (!is_leaf) {
2561 // Non-leaves can block and take safepoints:
2562 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2563 }
2564 // Non-leaves can throw exceptions:
2565 if (has_io) {
2566 call->set_req(TypeFunc::I_O, i_o());
2567 }
2568
2569 if (flags & RC_UNCOMMON) {
2570 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2571 // (An "if" probability corresponds roughly to an unconditional count.
2572 // Sort of.)
2573 call->set_cnt(PROB_UNLIKELY_MAG(4));
2574 }
2575
2576 Node* c = _gvn.transform(call);
2577 assert(c == call, "cannot disappear");
2578
2579 if (wide_out) {
2580 // Slow path call has full side-effects.
2581 set_predefined_output_for_runtime_call(call);
2582 } else {
2583 // Slow path call has few side-effects, and/or sets few values.
2584 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2585 }
2586
2587 if (has_io) {
2588 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2589 }
2590 return call;
2591
2592 }
2593
2594 // i2b
2595 Node* GraphKit::sign_extend_byte(Node* in) {
2596 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2597 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2598 }
2599
2600 // i2s
2601 Node* GraphKit::sign_extend_short(Node* in) {
2602 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2603 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2604 }
2605
2606 //------------------------------merge_memory-----------------------------------
2607 // Merge memory from one path into the current memory state.
2608 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2609 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2610 Node* old_slice = mms.force_memory();
2611 Node* new_slice = mms.memory2();
2612 if (old_slice != new_slice) {
2613 PhiNode* phi;
2614 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2615 if (mms.is_empty()) {
2616 // clone base memory Phi's inputs for this memory slice
2617 assert(old_slice == mms.base_memory(), "sanity");
2618 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2619 _gvn.set_type(phi, Type::MEMORY);
2620 for (uint i = 1; i < phi->req(); i++) {
2621 phi->init_req(i, old_slice->in(i));
2622 }
2623 } else {
2624 phi = old_slice->as_Phi(); // Phi was generated already
2625 }
2626 } else {
2627 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2628 _gvn.set_type(phi, Type::MEMORY);
2629 }
2630 phi->set_req(new_path, new_slice);
2631 mms.set_memory(phi);
2632 }
2633 }
2634 }
2635
2636 //------------------------------make_slow_call_ex------------------------------
2637 // Make the exception handler hookups for the slow call
2638 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2639 if (stopped()) return;
2640
2641 // Make a catch node with just two handlers: fall-through and catch-all
2642 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2643 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2644 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2645 _gvn.set_type_bottom(norm);
2646 C->record_for_igvn(norm);
2647 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2648
2649 { PreserveJVMState pjvms(this);
2650 set_control(excp);
2651 set_i_o(i_o);
2652
2653 if (excp != top()) {
2654 if (deoptimize) {
2655 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2656 uncommon_trap(Deoptimization::Reason_unhandled,
2657 Deoptimization::Action_none);
2658 } else {
2659 // Create an exception state also.
2660 // Use an exact type if the caller has a specific exception.
2661 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2662 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2663 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2664 }
2665 }
2666 }
2667
2668 // Get the no-exception control from the CatchNode.
2669 set_control(norm);
2670 }
2671
2672 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2673 Node* cmp = nullptr;
2674 switch(bt) {
2675 case T_INT: cmp = new CmpINode(in1, in2); break;
2676 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2677 default: fatal("unexpected comparison type %s", type2name(bt));
2678 }
2679 cmp = gvn.transform(cmp);
2680 Node* bol = gvn.transform(new BoolNode(cmp, test));
2681 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2682 gvn.transform(iff);
2683 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2684 return iff;
2685 }
2686
2687 //-------------------------------gen_subtype_check-----------------------------
2688 // Generate a subtyping check. Takes as input the subtype and supertype.
2689 // Returns 2 values: sets the default control() to the true path and returns
2690 // the false path. Only reads invariant memory; sets no (visible) memory.
2691 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2692 // but that's not exposed to the optimizer. This call also doesn't take in an
2693 // Object; if you wish to check an Object you need to load the Object's class
2694 // prior to coming here.
2695 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2696 ciMethod* method, int bci) {
2697 Compile* C = gvn.C;
2698 if ((*ctrl)->is_top()) {
2699 return C->top();
2700 }
2701
2702 // Fast check for identical types, perhaps identical constants.
2703 // The types can even be identical non-constants, in cases
2704 // involving Array.newInstance, Object.clone, etc.
2705 if (subklass == superklass)
2706 return C->top(); // false path is dead; no test needed.
2707
2708 if (gvn.type(superklass)->singleton()) {
2709 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2710 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2711
2712 // In the common case of an exact superklass, try to fold up the
2713 // test before generating code. You may ask, why not just generate
2714 // the code and then let it fold up? The answer is that the generated
2715 // code will necessarily include null checks, which do not always
2716 // completely fold away. If they are also needless, then they turn
2717 // into a performance loss. Example:
2718 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2719 // Here, the type of 'fa' is often exact, so the store check
2720 // of fa[1]=x will fold up, without testing the nullness of x.
2721 //
2722 // At macro expansion, we would have already folded the SubTypeCheckNode
2723 // being expanded here because we always perform the static sub type
2724 // check in SubTypeCheckNode::sub() regardless of whether
2725 // StressReflectiveCode is set or not. We can therefore skip this
2726 // static check when StressReflectiveCode is on.
2727 switch (C->static_subtype_check(superk, subk)) {
2728 case Compile::SSC_always_false:
2729 {
2730 Node* always_fail = *ctrl;
2731 *ctrl = gvn.C->top();
2732 return always_fail;
2733 }
2734 case Compile::SSC_always_true:
2735 return C->top();
2736 case Compile::SSC_easy_test:
2737 {
2738 // Just do a direct pointer compare and be done.
2739 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2740 *ctrl = gvn.transform(new IfTrueNode(iff));
2741 return gvn.transform(new IfFalseNode(iff));
2742 }
2743 case Compile::SSC_full_test:
2744 break;
2745 default:
2746 ShouldNotReachHere();
2747 }
2748 }
2749
2750 // %%% Possible further optimization: Even if the superklass is not exact,
2751 // if the subklass is the unique subtype of the superklass, the check
2752 // will always succeed. We could leave a dependency behind to ensure this.
2753
2754 // First load the super-klass's check-offset
2755 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2756 Node* m = C->immutable_memory();
2757 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2758 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2759 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2760 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2761 bool might_be_cache = (chk_off_con == cacheoff_con);
2762
2763 // Load from the sub-klass's super-class display list, or a 1-word cache of
2764 // the secondary superclass list, or a failing value with a sentinel offset
2765 // if the super-klass is an interface or exceptionally deep in the Java
2766 // hierarchy and we have to scan the secondary superclass list the hard way.
2767 // Worst-case type is a little odd: null is allowed as a result (usually
2768 // klass loads can never produce a null).
2769 Node *chk_off_X = chk_off;
2770 #ifdef _LP64
2771 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2772 #endif
2773 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2774 // For some types like interfaces the following loadKlass is from a 1-word
2775 // cache which is mutable so can't use immutable memory. Other
2776 // types load from the super-class display table which is immutable.
2777 Node *kmem = C->immutable_memory();
2778 // secondary_super_cache is not immutable but can be treated as such because:
2779 // - no ideal node writes to it in a way that could cause an
2780 // incorrect/missed optimization of the following Load.
2781 // - it's a cache so, worse case, not reading the latest value
2782 // wouldn't cause incorrect execution
2783 if (might_be_cache && mem != nullptr) {
2784 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2785 }
2786 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2787
2788 // Compile speed common case: ARE a subtype and we canNOT fail
2789 if (superklass == nkls) {
2790 return C->top(); // false path is dead; no test needed.
2791 }
2792
2793 // Gather the various success & failures here
2794 RegionNode* r_not_subtype = new RegionNode(3);
2795 gvn.record_for_igvn(r_not_subtype);
2796 RegionNode* r_ok_subtype = new RegionNode(4);
2797 gvn.record_for_igvn(r_ok_subtype);
2798
2799 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2800 // SubTypeCheck node
2801 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2802 ciCallProfile profile = method->call_profile_at_bci(bci);
2803 float total_prob = 0;
2804 for (int i = 0; profile.has_receiver(i); ++i) {
2805 float prob = profile.receiver_prob(i);
2806 total_prob += prob;
2807 }
2808 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2809 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2810 for (int i = 0; profile.has_receiver(i); ++i) {
2811 ciKlass* klass = profile.receiver(i);
2812 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2813 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2814 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2815 continue;
2816 }
2817 float prob = profile.receiver_prob(i);
2818 ConNode* klass_node = gvn.makecon(klass_t);
2819 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2820 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2821
2822 if (result == Compile::SSC_always_true) {
2823 r_ok_subtype->add_req(iftrue);
2824 } else {
2825 assert(result == Compile::SSC_always_false, "");
2826 r_not_subtype->add_req(iftrue);
2827 }
2828 *ctrl = gvn.transform(new IfFalseNode(iff));
2829 }
2830 }
2831 }
2832
2833 // See if we get an immediate positive hit. Happens roughly 83% of the
2834 // time. Test to see if the value loaded just previously from the subklass
2835 // is exactly the superklass.
2836 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2837 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2838 *ctrl = gvn.transform(new IfFalseNode(iff1));
2839
2840 // Compile speed common case: Check for being deterministic right now. If
2841 // chk_off is a constant and not equal to cacheoff then we are NOT a
2842 // subklass. In this case we need exactly the 1 test above and we can
2843 // return those results immediately.
2844 if (!might_be_cache) {
2845 Node* not_subtype_ctrl = *ctrl;
2846 *ctrl = iftrue1; // We need exactly the 1 test above
2847 PhaseIterGVN* igvn = gvn.is_IterGVN();
2848 if (igvn != nullptr) {
2849 igvn->remove_globally_dead_node(r_ok_subtype);
2850 igvn->remove_globally_dead_node(r_not_subtype);
2851 }
2852 return not_subtype_ctrl;
2853 }
2854
2855 r_ok_subtype->init_req(1, iftrue1);
2856
2857 // Check for immediate negative hit. Happens roughly 11% of the time (which
2858 // is roughly 63% of the remaining cases). Test to see if the loaded
2859 // check-offset points into the subklass display list or the 1-element
2860 // cache. If it points to the display (and NOT the cache) and the display
2861 // missed then it's not a subtype.
2862 Node *cacheoff = gvn.intcon(cacheoff_con);
2863 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2864 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2865 *ctrl = gvn.transform(new IfFalseNode(iff2));
2866
2867 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2868 // No performance impact (too rare) but allows sharing of secondary arrays
2869 // which has some footprint reduction.
2870 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2871 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2872 *ctrl = gvn.transform(new IfFalseNode(iff3));
2873
2874 // -- Roads not taken here: --
2875 // We could also have chosen to perform the self-check at the beginning
2876 // of this code sequence, as the assembler does. This would not pay off
2877 // the same way, since the optimizer, unlike the assembler, can perform
2878 // static type analysis to fold away many successful self-checks.
2879 // Non-foldable self checks work better here in second position, because
2880 // the initial primary superclass check subsumes a self-check for most
2881 // types. An exception would be a secondary type like array-of-interface,
2882 // which does not appear in its own primary supertype display.
2883 // Finally, we could have chosen to move the self-check into the
2884 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2885 // dependent manner. But it is worthwhile to have the check here,
2886 // where it can be perhaps be optimized. The cost in code space is
2887 // small (register compare, branch).
2888
2889 // Now do a linear scan of the secondary super-klass array. Again, no real
2890 // performance impact (too rare) but it's gotta be done.
2891 // Since the code is rarely used, there is no penalty for moving it
2892 // out of line, and it can only improve I-cache density.
2893 // The decision to inline or out-of-line this final check is platform
2894 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2895 Node* psc = gvn.transform(
2896 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2897
2898 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2899 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2900 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2901
2902 // Return false path; set default control to true path.
2903 *ctrl = gvn.transform(r_ok_subtype);
2904 return gvn.transform(r_not_subtype);
2905 }
2906
2907 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2908 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2909 if (expand_subtype_check) {
2910 MergeMemNode* mem = merged_memory();
2911 Node* ctrl = control();
2912 Node* subklass = obj_or_subklass;
2913 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2914 subklass = load_object_klass(obj_or_subklass);
2915 }
2916
2917 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2918 set_control(ctrl);
2919 return n;
2920 }
2921
2922 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2923 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2924 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2925 set_control(_gvn.transform(new IfTrueNode(iff)));
2926 return _gvn.transform(new IfFalseNode(iff));
2927 }
2928
2929 // Profile-driven exact type check:
2930 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2931 float prob,
2932 Node* *casted_receiver) {
2933 assert(!klass->is_interface(), "no exact type check on interfaces");
2934
2935 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2936 Node* recv_klass = load_object_klass(receiver);
2937 Node* want_klass = makecon(tklass);
2938 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2939 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2940 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2941 set_control( _gvn.transform(new IfTrueNode (iff)));
2942 Node* fail = _gvn.transform(new IfFalseNode(iff));
2943
2944 if (!stopped()) {
2945 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2946 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2947 assert(recvx_type->klass_is_exact(), "");
2948
2949 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2950 // Subsume downstream occurrences of receiver with a cast to
2951 // recv_xtype, since now we know what the type will be.
2952 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2953 (*casted_receiver) = _gvn.transform(cast);
2954 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2955 // (User must make the replace_in_map call.)
2956 }
2957 }
2958
2959 return fail;
2960 }
2961
2962 //------------------------------subtype_check_receiver-------------------------
2963 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2964 Node** casted_receiver) {
2965 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2966 Node* want_klass = makecon(tklass);
2967
2968 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2969
2970 // Ignore interface type information until interface types are properly tracked.
2971 if (!stopped() && !klass->is_interface()) {
2972 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2973 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2974 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2975 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2976 (*casted_receiver) = _gvn.transform(cast);
2977 }
2978 }
2979
2980 return slow_ctl;
2981 }
2982
2983 //------------------------------seems_never_null-------------------------------
2984 // Use null_seen information if it is available from the profile.
2985 // If we see an unexpected null at a type check we record it and force a
2986 // recompile; the offending check will be recompiled to handle nulls.
2987 // If we see several offending BCIs, then all checks in the
2988 // method will be recompiled.
2989 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2990 speculating = !_gvn.type(obj)->speculative_maybe_null();
2991 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2992 if (UncommonNullCast // Cutout for this technique
2993 && obj != null() // And not the -Xcomp stupid case?
2994 && !too_many_traps(reason)
2995 ) {
2996 if (speculating) {
2997 return true;
2998 }
2999 if (data == nullptr)
3000 // Edge case: no mature data. Be optimistic here.
3001 return true;
3002 // If the profile has not seen a null, assume it won't happen.
3003 assert(java_bc() == Bytecodes::_checkcast ||
3004 java_bc() == Bytecodes::_instanceof ||
3005 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3006 return !data->as_BitData()->null_seen();
3007 }
3008 speculating = false;
3009 return false;
3010 }
3011
3012 void GraphKit::guard_klass_being_initialized(Node* klass) {
3013 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3014 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3015 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3016 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3017 T_BYTE, MemNode::acquire);
3018 init_state = _gvn.transform(init_state);
3019
3020 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3021
3022 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3023 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3024
3025 { BuildCutout unless(this, tst, PROB_MAX);
3026 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3027 }
3028 }
3029
3030 void GraphKit::guard_init_thread(Node* klass) {
3031 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3032 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3033
3034 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3035 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3036 T_ADDRESS, MemNode::unordered);
3037 init_thread = _gvn.transform(init_thread);
3038
3039 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3040
3041 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3042 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3043
3044 { BuildCutout unless(this, tst, PROB_MAX);
3045 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3046 }
3047 }
3048
3049 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3050 if (ik->is_being_initialized()) {
3051 if (C->needs_clinit_barrier(ik, context)) {
3052 Node* klass = makecon(TypeKlassPtr::make(ik));
3053 guard_klass_being_initialized(klass);
3054 guard_init_thread(klass);
3055 insert_mem_bar(Op_MemBarCPUOrder);
3056 }
3057 } else if (ik->is_initialized()) {
3058 return; // no barrier needed
3059 } else {
3060 uncommon_trap(Deoptimization::Reason_uninitialized,
3061 Deoptimization::Action_reinterpret,
3062 nullptr);
3063 }
3064 }
3065
3066 //------------------------maybe_cast_profiled_receiver-------------------------
3067 // If the profile has seen exactly one type, narrow to exactly that type.
3068 // Subsequent type checks will always fold up.
3069 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3070 const TypeKlassPtr* require_klass,
3071 ciKlass* spec_klass,
3072 bool safe_for_replace) {
3073 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3074
3075 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3076
3077 // Make sure we haven't already deoptimized from this tactic.
3078 if (too_many_traps_or_recompiles(reason))
3079 return nullptr;
3080
3081 // (No, this isn't a call, but it's enough like a virtual call
3082 // to use the same ciMethod accessor to get the profile info...)
3083 // If we have a speculative type use it instead of profiling (which
3084 // may not help us)
3085 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3086 if (exact_kls != nullptr) {// no cast failures here
3087 if (require_klass == nullptr ||
3088 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3089 // If we narrow the type to match what the type profile sees or
3090 // the speculative type, we can then remove the rest of the
3091 // cast.
3092 // This is a win, even if the exact_kls is very specific,
3093 // because downstream operations, such as method calls,
3094 // will often benefit from the sharper type.
3095 Node* exact_obj = not_null_obj; // will get updated in place...
3096 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3097 &exact_obj);
3098 { PreserveJVMState pjvms(this);
3099 set_control(slow_ctl);
3100 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3101 }
3102 if (safe_for_replace) {
3103 replace_in_map(not_null_obj, exact_obj);
3104 }
3105 return exact_obj;
3106 }
3107 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3108 }
3109
3110 return nullptr;
3111 }
3112
3113 /**
3114 * Cast obj to type and emit guard unless we had too many traps here
3115 * already
3116 *
3117 * @param obj node being casted
3118 * @param type type to cast the node to
3119 * @param not_null true if we know node cannot be null
3120 */
3121 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3122 ciKlass* type,
3123 bool not_null) {
3124 if (stopped()) {
3125 return obj;
3126 }
3127
3128 // type is null if profiling tells us this object is always null
3129 if (type != nullptr) {
3130 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3131 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3132
3133 if (!too_many_traps_or_recompiles(null_reason) &&
3134 !too_many_traps_or_recompiles(class_reason)) {
3135 Node* not_null_obj = nullptr;
3136 // not_null is true if we know the object is not null and
3137 // there's no need for a null check
3138 if (!not_null) {
3139 Node* null_ctl = top();
3140 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3141 assert(null_ctl->is_top(), "no null control here");
3142 } else {
3143 not_null_obj = obj;
3144 }
3145
3146 Node* exact_obj = not_null_obj;
3147 ciKlass* exact_kls = type;
3148 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3149 &exact_obj);
3150 {
3151 PreserveJVMState pjvms(this);
3152 set_control(slow_ctl);
3153 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3154 }
3155 replace_in_map(not_null_obj, exact_obj);
3156 obj = exact_obj;
3157 }
3158 } else {
3159 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3160 Node* exact_obj = null_assert(obj);
3161 replace_in_map(obj, exact_obj);
3162 obj = exact_obj;
3163 }
3164 }
3165 return obj;
3166 }
3167
3168 //-------------------------------gen_instanceof--------------------------------
3169 // Generate an instance-of idiom. Used by both the instance-of bytecode
3170 // and the reflective instance-of call.
3171 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3172 kill_dead_locals(); // Benefit all the uncommon traps
3173 assert( !stopped(), "dead parse path should be checked in callers" );
3174 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3175 "must check for not-null not-dead klass in callers");
3176
3177 // Make the merge point
3178 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3179 RegionNode* region = new RegionNode(PATH_LIMIT);
3180 Node* phi = new PhiNode(region, TypeInt::BOOL);
3181 C->set_has_split_ifs(true); // Has chance for split-if optimization
3182
3183 ciProfileData* data = nullptr;
3184 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3185 data = method()->method_data()->bci_to_data(bci());
3186 }
3187 bool speculative_not_null = false;
3188 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3189 && seems_never_null(obj, data, speculative_not_null));
3190
3191 // Null check; get casted pointer; set region slot 3
3192 Node* null_ctl = top();
3193 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3194
3195 // If not_null_obj is dead, only null-path is taken
3196 if (stopped()) { // Doing instance-of on a null?
3197 set_control(null_ctl);
3198 return intcon(0);
3199 }
3200 region->init_req(_null_path, null_ctl);
3201 phi ->init_req(_null_path, intcon(0)); // Set null path value
3202 if (null_ctl == top()) {
3203 // Do this eagerly, so that pattern matches like is_diamond_phi
3204 // will work even during parsing.
3205 assert(_null_path == PATH_LIMIT-1, "delete last");
3206 region->del_req(_null_path);
3207 phi ->del_req(_null_path);
3208 }
3209
3210 // Do we know the type check always succeed?
3211 bool known_statically = false;
3212 if (_gvn.type(superklass)->singleton()) {
3213 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3214 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3215 if (subk->is_loaded()) {
3216 int static_res = C->static_subtype_check(superk, subk);
3217 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3218 }
3219 }
3220
3221 if (!known_statically) {
3222 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3223 // We may not have profiling here or it may not help us. If we
3224 // have a speculative type use it to perform an exact cast.
3225 ciKlass* spec_obj_type = obj_type->speculative_type();
3226 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3227 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3228 if (stopped()) { // Profile disagrees with this path.
3229 set_control(null_ctl); // Null is the only remaining possibility.
3230 return intcon(0);
3231 }
3232 if (cast_obj != nullptr) {
3233 not_null_obj = cast_obj;
3234 }
3235 }
3236 }
3237
3238 // Generate the subtype check
3239 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3240
3241 // Plug in the success path to the general merge in slot 1.
3242 region->init_req(_obj_path, control());
3243 phi ->init_req(_obj_path, intcon(1));
3244
3245 // Plug in the failing path to the general merge in slot 2.
3246 region->init_req(_fail_path, not_subtype_ctrl);
3247 phi ->init_req(_fail_path, intcon(0));
3248
3249 // Return final merged results
3250 set_control( _gvn.transform(region) );
3251 record_for_igvn(region);
3252
3253 // If we know the type check always succeeds then we don't use the
3254 // profiling data at this bytecode. Don't lose it, feed it to the
3255 // type system as a speculative type.
3256 if (safe_for_replace) {
3257 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3258 replace_in_map(obj, casted_obj);
3259 }
3260
3261 return _gvn.transform(phi);
3262 }
3263
3264 //-------------------------------gen_checkcast---------------------------------
3265 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3266 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3267 // uncommon-trap paths work. Adjust stack after this call.
3268 // If failure_control is supplied and not null, it is filled in with
3269 // the control edge for the cast failure. Otherwise, an appropriate
3270 // uncommon trap or exception is thrown.
3271 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3272 Node* *failure_control) {
3273 kill_dead_locals(); // Benefit all the uncommon traps
3274 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3275 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3276 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3277
3278 // Fast cutout: Check the case that the cast is vacuously true.
3279 // This detects the common cases where the test will short-circuit
3280 // away completely. We do this before we perform the null check,
3281 // because if the test is going to turn into zero code, we don't
3282 // want a residual null check left around. (Causes a slowdown,
3283 // for example, in some objArray manipulations, such as a[i]=a[j].)
3284 if (improved_klass_ptr_type->singleton()) {
3285 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3286 if (objtp != nullptr) {
3287 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3288 case Compile::SSC_always_true:
3289 // If we know the type check always succeed then we don't use
3290 // the profiling data at this bytecode. Don't lose it, feed it
3291 // to the type system as a speculative type.
3292 return record_profiled_receiver_for_speculation(obj);
3293 case Compile::SSC_always_false:
3294 // It needs a null check because a null will *pass* the cast check.
3295 // A non-null value will always produce an exception.
3296 if (!objtp->maybe_null()) {
3297 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3298 Deoptimization::DeoptReason reason = is_aastore ?
3299 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3300 builtin_throw(reason);
3301 return top();
3302 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3303 return null_assert(obj);
3304 }
3305 break; // Fall through to full check
3306 default:
3307 break;
3308 }
3309 }
3310 }
3311
3312 ciProfileData* data = nullptr;
3313 bool safe_for_replace = false;
3314 if (failure_control == nullptr) { // use MDO in regular case only
3315 assert(java_bc() == Bytecodes::_aastore ||
3316 java_bc() == Bytecodes::_checkcast,
3317 "interpreter profiles type checks only for these BCs");
3318 data = method()->method_data()->bci_to_data(bci());
3319 safe_for_replace = true;
3320 }
3321
3322 // Make the merge point
3323 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3324 RegionNode* region = new RegionNode(PATH_LIMIT);
3325 Node* phi = new PhiNode(region, toop);
3326 C->set_has_split_ifs(true); // Has chance for split-if optimization
3327
3328 // Use null-cast information if it is available
3329 bool speculative_not_null = false;
3330 bool never_see_null = ((failure_control == nullptr) // regular case only
3331 && seems_never_null(obj, data, speculative_not_null));
3332
3333 // Null check; get casted pointer; set region slot 3
3334 Node* null_ctl = top();
3335 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3336
3337 // If not_null_obj is dead, only null-path is taken
3338 if (stopped()) { // Doing instance-of on a null?
3339 set_control(null_ctl);
3340 return null();
3341 }
3342 region->init_req(_null_path, null_ctl);
3343 phi ->init_req(_null_path, null()); // Set null path value
3344 if (null_ctl == top()) {
3345 // Do this eagerly, so that pattern matches like is_diamond_phi
3346 // will work even during parsing.
3347 assert(_null_path == PATH_LIMIT-1, "delete last");
3348 region->del_req(_null_path);
3349 phi ->del_req(_null_path);
3350 }
3351
3352 Node* cast_obj = nullptr;
3353 if (improved_klass_ptr_type->klass_is_exact()) {
3354 // The following optimization tries to statically cast the speculative type of the object
3355 // (for example obtained during profiling) to the type of the superklass and then do a
3356 // dynamic check that the type of the object is what we expect. To work correctly
3357 // for checkcast and aastore the type of superklass should be exact.
3358 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3359 // We may not have profiling here or it may not help us. If we have
3360 // a speculative type use it to perform an exact cast.
3361 ciKlass* spec_obj_type = obj_type->speculative_type();
3362 if (spec_obj_type != nullptr || data != nullptr) {
3363 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3364 if (cast_obj != nullptr) {
3365 if (failure_control != nullptr) // failure is now impossible
3366 (*failure_control) = top();
3367 // adjust the type of the phi to the exact klass:
3368 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3369 }
3370 }
3371 }
3372
3373 if (cast_obj == nullptr) {
3374 // Generate the subtype check
3375 Node* improved_superklass = superklass;
3376 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3377 improved_superklass = makecon(improved_klass_ptr_type);
3378 }
3379 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3380
3381 // Plug in success path into the merge
3382 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3383 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3384 if (failure_control == nullptr) {
3385 if (not_subtype_ctrl != top()) { // If failure is possible
3386 PreserveJVMState pjvms(this);
3387 set_control(not_subtype_ctrl);
3388 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3389 Deoptimization::DeoptReason reason = is_aastore ?
3390 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3391 builtin_throw(reason);
3392 }
3393 } else {
3394 (*failure_control) = not_subtype_ctrl;
3395 }
3396 }
3397
3398 region->init_req(_obj_path, control());
3399 phi ->init_req(_obj_path, cast_obj);
3400
3401 // A merge of null or Casted-NotNull obj
3402 Node* res = _gvn.transform(phi);
3403
3404 // Note I do NOT always 'replace_in_map(obj,result)' here.
3405 // if( tk->klass()->can_be_primary_super() )
3406 // This means that if I successfully store an Object into an array-of-String
3407 // I 'forget' that the Object is really now known to be a String. I have to
3408 // do this because we don't have true union types for interfaces - if I store
3409 // a Baz into an array-of-Interface and then tell the optimizer it's an
3410 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3411 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3412 // replace_in_map( obj, res );
3413
3414 // Return final merged results
3415 set_control( _gvn.transform(region) );
3416 record_for_igvn(region);
3417
3418 return record_profiled_receiver_for_speculation(res);
3419 }
3420
3421 //------------------------------next_monitor-----------------------------------
3422 // What number should be given to the next monitor?
3423 int GraphKit::next_monitor() {
3424 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3425 int next = current + C->sync_stack_slots();
3426 // Keep the toplevel high water mark current:
3427 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3428 return current;
3429 }
3430
3431 //------------------------------insert_mem_bar---------------------------------
3432 // Memory barrier to avoid floating things around
3433 // The membar serves as a pinch point between both control and all memory slices.
3434 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3435 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3436 mb->init_req(TypeFunc::Control, control());
3437 mb->init_req(TypeFunc::Memory, reset_memory());
3438 Node* membar = _gvn.transform(mb);
3439 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3440 set_all_memory_call(membar);
3441 return membar;
3442 }
3443
3444 //-------------------------insert_mem_bar_volatile----------------------------
3445 // Memory barrier to avoid floating things around
3446 // The membar serves as a pinch point between both control and memory(alias_idx).
3447 // If you want to make a pinch point on all memory slices, do not use this
3448 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3449 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3450 // When Parse::do_put_xxx updates a volatile field, it appends a series
3451 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3452 // The first membar is on the same memory slice as the field store opcode.
3453 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3454 // All the other membars (for other volatile slices, including AliasIdxBot,
3455 // which stands for all unknown volatile slices) are control-dependent
3456 // on the first membar. This prevents later volatile loads or stores
3457 // from sliding up past the just-emitted store.
3458
3459 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3460 mb->set_req(TypeFunc::Control,control());
3461 if (alias_idx == Compile::AliasIdxBot) {
3462 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3463 } else {
3464 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3465 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3466 }
3467 Node* membar = _gvn.transform(mb);
3468 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3469 if (alias_idx == Compile::AliasIdxBot) {
3470 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3471 } else {
3472 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3473 }
3474 return membar;
3475 }
3476
3477 //------------------------------shared_lock------------------------------------
3478 // Emit locking code.
3479 FastLockNode* GraphKit::shared_lock(Node* obj) {
3480 // bci is either a monitorenter bc or InvocationEntryBci
3481 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3482 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3483
3484 if( !GenerateSynchronizationCode )
3485 return nullptr; // Not locking things?
3486 if (stopped()) // Dead monitor?
3487 return nullptr;
3488
3489 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3490
3491 // Box the stack location
3492 Node* box = new BoxLockNode(next_monitor());
3493 // Check for bailout after new BoxLockNode
3494 if (failing()) { return nullptr; }
3495 box = _gvn.transform(box);
3496 Node* mem = reset_memory();
3497
3498 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3499
3500 // Add monitor to debug info for the slow path. If we block inside the
3501 // slow path and de-opt, we need the monitor hanging around
3502 map()->push_monitor( flock );
3503
3504 const TypeFunc *tf = LockNode::lock_type();
3505 LockNode *lock = new LockNode(C, tf);
3506
3507 lock->init_req( TypeFunc::Control, control() );
3508 lock->init_req( TypeFunc::Memory , mem );
3509 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3510 lock->init_req( TypeFunc::FramePtr, frameptr() );
3511 lock->init_req( TypeFunc::ReturnAdr, top() );
3512
3513 lock->init_req(TypeFunc::Parms + 0, obj);
3514 lock->init_req(TypeFunc::Parms + 1, box);
3515 lock->init_req(TypeFunc::Parms + 2, flock);
3516 add_safepoint_edges(lock);
3517
3518 lock = _gvn.transform( lock )->as_Lock();
3519
3520 // lock has no side-effects, sets few values
3521 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3522
3523 insert_mem_bar(Op_MemBarAcquireLock);
3524
3525 // Add this to the worklist so that the lock can be eliminated
3526 record_for_igvn(lock);
3527
3528 #ifndef PRODUCT
3529 if (PrintLockStatistics) {
3530 // Update the counter for this lock. Don't bother using an atomic
3531 // operation since we don't require absolute accuracy.
3532 lock->create_lock_counter(map()->jvms());
3533 increment_counter(lock->counter()->addr());
3534 }
3535 #endif
3536
3537 return flock;
3538 }
3539
3540
3541 //------------------------------shared_unlock----------------------------------
3542 // Emit unlocking code.
3543 void GraphKit::shared_unlock(Node* box, Node* obj) {
3544 // bci is either a monitorenter bc or InvocationEntryBci
3545 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3546 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3547
3548 if( !GenerateSynchronizationCode )
3549 return;
3550 if (stopped()) { // Dead monitor?
3551 map()->pop_monitor(); // Kill monitor from debug info
3552 return;
3553 }
3554
3555 // Memory barrier to avoid floating things down past the locked region
3556 insert_mem_bar(Op_MemBarReleaseLock);
3557
3558 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3559 UnlockNode *unlock = new UnlockNode(C, tf);
3560 #ifdef ASSERT
3561 unlock->set_dbg_jvms(sync_jvms());
3562 #endif
3563 uint raw_idx = Compile::AliasIdxRaw;
3564 unlock->init_req( TypeFunc::Control, control() );
3565 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3566 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3567 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3568 unlock->init_req( TypeFunc::ReturnAdr, top() );
3569
3570 unlock->init_req(TypeFunc::Parms + 0, obj);
3571 unlock->init_req(TypeFunc::Parms + 1, box);
3572 unlock = _gvn.transform(unlock)->as_Unlock();
3573
3574 Node* mem = reset_memory();
3575
3576 // unlock has no side-effects, sets few values
3577 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3578
3579 // Kill monitor from debug info
3580 map()->pop_monitor( );
3581 }
3582
3583 //-------------------------------get_layout_helper-----------------------------
3584 // If the given klass is a constant or known to be an array,
3585 // fetch the constant layout helper value into constant_value
3586 // and return null. Otherwise, load the non-constant
3587 // layout helper value, and return the node which represents it.
3588 // This two-faced routine is useful because allocation sites
3589 // almost always feature constant types.
3590 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3591 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3592 if (!StressReflectiveCode && klass_t != nullptr) {
3593 bool xklass = klass_t->klass_is_exact();
3594 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3595 jint lhelper;
3596 if (klass_t->isa_aryklassptr()) {
3597 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3598 if (is_reference_type(elem, true)) {
3599 elem = T_OBJECT;
3600 }
3601 lhelper = Klass::array_layout_helper(elem);
3602 } else {
3603 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3604 }
3605 if (lhelper != Klass::_lh_neutral_value) {
3606 constant_value = lhelper;
3607 return (Node*) nullptr;
3608 }
3609 }
3610 }
3611 constant_value = Klass::_lh_neutral_value; // put in a known value
3612 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3613 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3614 }
3615
3616 // We just put in an allocate/initialize with a big raw-memory effect.
3617 // Hook selected additional alias categories on the initialization.
3618 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3619 MergeMemNode* init_in_merge,
3620 Node* init_out_raw) {
3621 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3622 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3623
3624 Node* prevmem = kit.memory(alias_idx);
3625 init_in_merge->set_memory_at(alias_idx, prevmem);
3626 kit.set_memory(init_out_raw, alias_idx);
3627 }
3628
3629 //---------------------------set_output_for_allocation-------------------------
3630 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3631 const TypeOopPtr* oop_type,
3632 bool deoptimize_on_exception) {
3633 int rawidx = Compile::AliasIdxRaw;
3634 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3635 add_safepoint_edges(alloc);
3636 Node* allocx = _gvn.transform(alloc);
3637 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3638 // create memory projection for i_o
3639 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3640 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3641
3642 // create a memory projection as for the normal control path
3643 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3644 set_memory(malloc, rawidx);
3645
3646 // a normal slow-call doesn't change i_o, but an allocation does
3647 // we create a separate i_o projection for the normal control path
3648 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3649 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3650
3651 // put in an initialization barrier
3652 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3653 rawoop)->as_Initialize();
3654 assert(alloc->initialization() == init, "2-way macro link must work");
3655 assert(init ->allocation() == alloc, "2-way macro link must work");
3656 {
3657 // Extract memory strands which may participate in the new object's
3658 // initialization, and source them from the new InitializeNode.
3659 // This will allow us to observe initializations when they occur,
3660 // and link them properly (as a group) to the InitializeNode.
3661 assert(init->in(InitializeNode::Memory) == malloc, "");
3662 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3663 init->set_req(InitializeNode::Memory, minit_in);
3664 record_for_igvn(minit_in); // fold it up later, if possible
3665 Node* minit_out = memory(rawidx);
3666 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3667 // Add an edge in the MergeMem for the header fields so an access
3668 // to one of those has correct memory state
3669 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3670 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3671 if (oop_type->isa_aryptr()) {
3672 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3673 int elemidx = C->get_alias_index(telemref);
3674 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3675 } else if (oop_type->isa_instptr()) {
3676 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3677 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3678 ciField* field = ik->nonstatic_field_at(i);
3679 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3680 continue; // do not bother to track really large numbers of fields
3681 // Find (or create) the alias category for this field:
3682 int fieldidx = C->alias_type(field)->index();
3683 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3684 }
3685 }
3686 }
3687
3688 // Cast raw oop to the real thing...
3689 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3690 javaoop = _gvn.transform(javaoop);
3691 C->set_recent_alloc(control(), javaoop);
3692 assert(just_allocated_object(control()) == javaoop, "just allocated");
3693
3694 #ifdef ASSERT
3695 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3696 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
3697 "Ideal_allocation works");
3698 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
3699 "Ideal_allocation works");
3700 if (alloc->is_AllocateArray()) {
3701 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
3702 "Ideal_allocation works");
3703 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
3704 "Ideal_allocation works");
3705 } else {
3706 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3707 }
3708 }
3709 #endif //ASSERT
3710
3711 return javaoop;
3712 }
3713
3714 //---------------------------new_instance--------------------------------------
3715 // This routine takes a klass_node which may be constant (for a static type)
3716 // or may be non-constant (for reflective code). It will work equally well
3717 // for either, and the graph will fold nicely if the optimizer later reduces
3718 // the type to a constant.
3719 // The optional arguments are for specialized use by intrinsics:
3720 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3721 // - If 'return_size_val', report the total object size to the caller.
3722 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3723 Node* GraphKit::new_instance(Node* klass_node,
3724 Node* extra_slow_test,
3725 Node* *return_size_val,
3726 bool deoptimize_on_exception) {
3727 // Compute size in doublewords
3728 // The size is always an integral number of doublewords, represented
3729 // as a positive bytewise size stored in the klass's layout_helper.
3730 // The layout_helper also encodes (in a low bit) the need for a slow path.
3731 jint layout_con = Klass::_lh_neutral_value;
3732 Node* layout_val = get_layout_helper(klass_node, layout_con);
3733 int layout_is_con = (layout_val == nullptr);
3734
3735 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3736 // Generate the initial go-slow test. It's either ALWAYS (return a
3737 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3738 // case) a computed value derived from the layout_helper.
3739 Node* initial_slow_test = nullptr;
3740 if (layout_is_con) {
3741 assert(!StressReflectiveCode, "stress mode does not use these paths");
3742 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3743 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3744 } else { // reflective case
3745 // This reflective path is used by Unsafe.allocateInstance.
3746 // (It may be stress-tested by specifying StressReflectiveCode.)
3747 // Basically, we want to get into the VM is there's an illegal argument.
3748 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3749 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3750 if (extra_slow_test != intcon(0)) {
3751 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3752 }
3753 // (Macro-expander will further convert this to a Bool, if necessary.)
3754 }
3755
3756 // Find the size in bytes. This is easy; it's the layout_helper.
3757 // The size value must be valid even if the slow path is taken.
3758 Node* size = nullptr;
3759 if (layout_is_con) {
3760 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3761 } else { // reflective case
3762 // This reflective path is used by clone and Unsafe.allocateInstance.
3763 size = ConvI2X(layout_val);
3764
3765 // Clear the low bits to extract layout_helper_size_in_bytes:
3766 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3767 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3768 size = _gvn.transform( new AndXNode(size, mask) );
3769 }
3770 if (return_size_val != nullptr) {
3771 (*return_size_val) = size;
3772 }
3773
3774 // This is a precise notnull oop of the klass.
3775 // (Actually, it need not be precise if this is a reflective allocation.)
3776 // It's what we cast the result to.
3777 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3778 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3779 const TypeOopPtr* oop_type = tklass->as_instance_type();
3780
3781 // Now generate allocation code
3782
3783 // The entire memory state is needed for slow path of the allocation
3784 // since GC and deoptimization can happened.
3785 Node *mem = reset_memory();
3786 set_all_memory(mem); // Create new memory state
3787
3788 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3789 control(), mem, i_o(),
3790 size, klass_node,
3791 initial_slow_test);
3792
3793 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3794 }
3795
3796 //-------------------------------new_array-------------------------------------
3797 // helper for both newarray and anewarray
3798 // The 'length' parameter is (obviously) the length of the array.
3799 // The optional arguments are for specialized use by intrinsics:
3800 // - If 'return_size_val', report the non-padded array size (sum of header size
3801 // and array body) to the caller.
3802 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3803 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3804 Node* length, // number of array elements
3805 int nargs, // number of arguments to push back for uncommon trap
3806 Node* *return_size_val,
3807 bool deoptimize_on_exception) {
3808 jint layout_con = Klass::_lh_neutral_value;
3809 Node* layout_val = get_layout_helper(klass_node, layout_con);
3810 int layout_is_con = (layout_val == nullptr);
3811
3812 if (!layout_is_con && !StressReflectiveCode &&
3813 !too_many_traps(Deoptimization::Reason_class_check)) {
3814 // This is a reflective array creation site.
3815 // Optimistically assume that it is a subtype of Object[],
3816 // so that we can fold up all the address arithmetic.
3817 layout_con = Klass::array_layout_helper(T_OBJECT);
3818 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3819 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3820 { BuildCutout unless(this, bol_lh, PROB_MAX);
3821 inc_sp(nargs);
3822 uncommon_trap(Deoptimization::Reason_class_check,
3823 Deoptimization::Action_maybe_recompile);
3824 }
3825 layout_val = nullptr;
3826 layout_is_con = true;
3827 }
3828
3829 // Generate the initial go-slow test. Make sure we do not overflow
3830 // if length is huge (near 2Gig) or negative! We do not need
3831 // exact double-words here, just a close approximation of needed
3832 // double-words. We can't add any offset or rounding bits, lest we
3833 // take a size -1 of bytes and make it positive. Use an unsigned
3834 // compare, so negative sizes look hugely positive.
3835 int fast_size_limit = FastAllocateSizeLimit;
3836 if (layout_is_con) {
3837 assert(!StressReflectiveCode, "stress mode does not use these paths");
3838 // Increase the size limit if we have exact knowledge of array type.
3839 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3840 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3841 }
3842
3843 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3844 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3845
3846 // --- Size Computation ---
3847 // array_size = round_to_heap(array_header + (length << elem_shift));
3848 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3849 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3850 // The rounding mask is strength-reduced, if possible.
3851 int round_mask = MinObjAlignmentInBytes - 1;
3852 Node* header_size = nullptr;
3853 // (T_BYTE has the weakest alignment and size restrictions...)
3854 if (layout_is_con) {
3855 int hsize = Klass::layout_helper_header_size(layout_con);
3856 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3857 if ((round_mask & ~right_n_bits(eshift)) == 0)
3858 round_mask = 0; // strength-reduce it if it goes away completely
3859 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3860 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3861 assert(header_size_min <= hsize, "generic minimum is smallest");
3862 header_size = intcon(hsize);
3863 } else {
3864 Node* hss = intcon(Klass::_lh_header_size_shift);
3865 Node* hsm = intcon(Klass::_lh_header_size_mask);
3866 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3867 header_size = _gvn.transform(new AndINode(header_size, hsm));
3868 }
3869
3870 Node* elem_shift = nullptr;
3871 if (layout_is_con) {
3872 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3873 if (eshift != 0)
3874 elem_shift = intcon(eshift);
3875 } else {
3876 // There is no need to mask or shift this value.
3877 // The semantics of LShiftINode include an implicit mask to 0x1F.
3878 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3879 elem_shift = layout_val;
3880 }
3881
3882 // Transition to native address size for all offset calculations:
3883 Node* lengthx = ConvI2X(length);
3884 Node* headerx = ConvI2X(header_size);
3885 #ifdef _LP64
3886 { const TypeInt* tilen = _gvn.find_int_type(length);
3887 if (tilen != nullptr && tilen->_lo < 0) {
3888 // Add a manual constraint to a positive range. Cf. array_element_address.
3889 jint size_max = fast_size_limit;
3890 if (size_max > tilen->_hi) size_max = tilen->_hi;
3891 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3892
3893 // Only do a narrow I2L conversion if the range check passed.
3894 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3895 _gvn.transform(iff);
3896 RegionNode* region = new RegionNode(3);
3897 _gvn.set_type(region, Type::CONTROL);
3898 lengthx = new PhiNode(region, TypeLong::LONG);
3899 _gvn.set_type(lengthx, TypeLong::LONG);
3900
3901 // Range check passed. Use ConvI2L node with narrow type.
3902 Node* passed = IfFalse(iff);
3903 region->init_req(1, passed);
3904 // Make I2L conversion control dependent to prevent it from
3905 // floating above the range check during loop optimizations.
3906 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3907
3908 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3909 region->init_req(2, IfTrue(iff));
3910 lengthx->init_req(2, ConvI2X(length));
3911
3912 set_control(region);
3913 record_for_igvn(region);
3914 record_for_igvn(lengthx);
3915 }
3916 }
3917 #endif
3918
3919 // Combine header size and body size for the array copy part, then align (if
3920 // necessary) for the allocation part. This computation cannot overflow,
3921 // because it is used only in two places, one where the length is sharply
3922 // limited, and the other after a successful allocation.
3923 Node* abody = lengthx;
3924 if (elem_shift != nullptr) {
3925 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
3926 }
3927 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3928
3929 if (return_size_val != nullptr) {
3930 // This is the size
3931 (*return_size_val) = non_rounded_size;
3932 }
3933
3934 Node* size = non_rounded_size;
3935 if (round_mask != 0) {
3936 Node* mask1 = MakeConX(round_mask);
3937 size = _gvn.transform(new AddXNode(size, mask1));
3938 Node* mask2 = MakeConX(~round_mask);
3939 size = _gvn.transform(new AndXNode(size, mask2));
3940 }
3941 // else if round_mask == 0, the size computation is self-rounding
3942
3943 // Now generate allocation code
3944
3945 // The entire memory state is needed for slow path of the allocation
3946 // since GC and deoptimization can happened.
3947 Node *mem = reset_memory();
3948 set_all_memory(mem); // Create new memory state
3949
3950 if (initial_slow_test->is_Bool()) {
3951 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3952 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3953 }
3954
3955 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3956 Node* valid_length_test = _gvn.intcon(1);
3957 if (ary_type->isa_aryptr()) {
3958 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3959 jint max = TypeAryPtr::max_array_length(bt);
3960 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3961 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3962 }
3963
3964 // Create the AllocateArrayNode and its result projections
3965 AllocateArrayNode* alloc
3966 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3967 control(), mem, i_o(),
3968 size, klass_node,
3969 initial_slow_test,
3970 length, valid_length_test);
3971
3972 // Cast to correct type. Note that the klass_node may be constant or not,
3973 // and in the latter case the actual array type will be inexact also.
3974 // (This happens via a non-constant argument to inline_native_newArray.)
3975 // In any case, the value of klass_node provides the desired array type.
3976 const TypeInt* length_type = _gvn.find_int_type(length);
3977 if (ary_type->isa_aryptr() && length_type != nullptr) {
3978 // Try to get a better type than POS for the size
3979 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3980 }
3981
3982 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3983
3984 array_ideal_length(alloc, ary_type, true);
3985 return javaoop;
3986 }
3987
3988 // The following "Ideal_foo" functions are placed here because they recognize
3989 // the graph shapes created by the functions immediately above.
3990
3991 //---------------------------Ideal_allocation----------------------------------
3992 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3993 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
3994 if (ptr == nullptr) { // reduce dumb test in callers
3995 return nullptr;
3996 }
3997
3998 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3999 ptr = bs->step_over_gc_barrier(ptr);
4000
4001 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4002 ptr = ptr->in(1);
4003 if (ptr == nullptr) return nullptr;
4004 }
4005 // Return null for allocations with several casts:
4006 // j.l.reflect.Array.newInstance(jobject, jint)
4007 // Object.clone()
4008 // to keep more precise type from last cast.
4009 if (ptr->is_Proj()) {
4010 Node* allo = ptr->in(0);
4011 if (allo != nullptr && allo->is_Allocate()) {
4012 return allo->as_Allocate();
4013 }
4014 }
4015 // Report failure to match.
4016 return nullptr;
4017 }
4018
4019 // Fancy version which also strips off an offset (and reports it to caller).
4020 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4021 intptr_t& offset) {
4022 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4023 if (base == nullptr) return nullptr;
4024 return Ideal_allocation(base);
4025 }
4026
4027 // Trace Initialize <- Proj[Parm] <- Allocate
4028 AllocateNode* InitializeNode::allocation() {
4029 Node* rawoop = in(InitializeNode::RawAddress);
4030 if (rawoop->is_Proj()) {
4031 Node* alloc = rawoop->in(0);
4032 if (alloc->is_Allocate()) {
4033 return alloc->as_Allocate();
4034 }
4035 }
4036 return nullptr;
4037 }
4038
4039 // Trace Allocate -> Proj[Parm] -> Initialize
4040 InitializeNode* AllocateNode::initialization() {
4041 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4042 if (rawoop == nullptr) return nullptr;
4043 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4044 Node* init = rawoop->fast_out(i);
4045 if (init->is_Initialize()) {
4046 assert(init->as_Initialize()->allocation() == this, "2-way link");
4047 return init->as_Initialize();
4048 }
4049 }
4050 return nullptr;
4051 }
4052
4053 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4054 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4055 // Too many traps seen?
4056 if (too_many_traps(reason)) {
4057 #ifdef ASSERT
4058 if (TraceLoopPredicate) {
4059 int tc = C->trap_count(reason);
4060 tty->print("too many traps=%s tcount=%d in ",
4061 Deoptimization::trap_reason_name(reason), tc);
4062 method()->print(); // which method has too many predicate traps
4063 tty->cr();
4064 }
4065 #endif
4066 // We cannot afford to take more traps here,
4067 // do not generate Parse Predicate.
4068 return;
4069 }
4070
4071 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4072 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4073 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4074 {
4075 PreserveJVMState pjvms(this);
4076 set_control(if_false);
4077 inc_sp(nargs);
4078 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4079 }
4080 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4081 set_control(if_true);
4082 }
4083
4084 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4085 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4086 void GraphKit::add_parse_predicates(int nargs) {
4087 if (UseLoopPredicate) {
4088 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4089 }
4090 if (UseProfiledLoopPredicate) {
4091 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4092 }
4093 // Loop Limit Check Predicate should be near the loop.
4094 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4095 }
4096
4097 void GraphKit::sync_kit(IdealKit& ideal) {
4098 set_all_memory(ideal.merged_memory());
4099 set_i_o(ideal.i_o());
4100 set_control(ideal.ctrl());
4101 }
4102
4103 void GraphKit::final_sync(IdealKit& ideal) {
4104 // Final sync IdealKit and graphKit.
4105 sync_kit(ideal);
4106 }
4107
4108 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4109 Node* len = load_array_length(load_String_value(str, set_ctrl));
4110 Node* coder = load_String_coder(str, set_ctrl);
4111 // Divide length by 2 if coder is UTF16
4112 return _gvn.transform(new RShiftINode(len, coder));
4113 }
4114
4115 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4116 int value_offset = java_lang_String::value_offset();
4117 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4118 false, nullptr, 0);
4119 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4120 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4121 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4122 ciTypeArrayKlass::make(T_BYTE), true, 0);
4123 Node* p = basic_plus_adr(str, str, value_offset);
4124 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4125 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4126 return load;
4127 }
4128
4129 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4130 if (!CompactStrings) {
4131 return intcon(java_lang_String::CODER_UTF16);
4132 }
4133 int coder_offset = java_lang_String::coder_offset();
4134 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4135 false, nullptr, 0);
4136 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4137
4138 Node* p = basic_plus_adr(str, str, coder_offset);
4139 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4140 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4141 return load;
4142 }
4143
4144 void GraphKit::store_String_value(Node* str, Node* value) {
4145 int value_offset = java_lang_String::value_offset();
4146 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4147 false, nullptr, 0);
4148 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4149
4150 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4151 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4152 }
4153
4154 void GraphKit::store_String_coder(Node* str, Node* value) {
4155 int coder_offset = java_lang_String::coder_offset();
4156 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4157 false, nullptr, 0);
4158 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4159
4160 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4161 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4162 }
4163
4164 // Capture src and dst memory state with a MergeMemNode
4165 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4166 if (src_type == dst_type) {
4167 // Types are equal, we don't need a MergeMemNode
4168 return memory(src_type);
4169 }
4170 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4171 record_for_igvn(merge); // fold it up later, if possible
4172 int src_idx = C->get_alias_index(src_type);
4173 int dst_idx = C->get_alias_index(dst_type);
4174 merge->set_memory_at(src_idx, memory(src_idx));
4175 merge->set_memory_at(dst_idx, memory(dst_idx));
4176 return merge;
4177 }
4178
4179 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4180 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4181 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4182 // If input and output memory types differ, capture both states to preserve
4183 // the dependency between preceding and subsequent loads/stores.
4184 // For example, the following program:
4185 // StoreB
4186 // compress_string
4187 // LoadB
4188 // has this memory graph (use->def):
4189 // LoadB -> compress_string -> CharMem
4190 // ... -> StoreB -> ByteMem
4191 // The intrinsic hides the dependency between LoadB and StoreB, causing
4192 // the load to read from memory not containing the result of the StoreB.
4193 // The correct memory graph should look like this:
4194 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4195 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4196 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4197 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4198 set_memory(res_mem, TypeAryPtr::BYTES);
4199 return str;
4200 }
4201
4202 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4203 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4204 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4205 // Capture src and dst memory (see comment in 'compress_string').
4206 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4207 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4208 set_memory(_gvn.transform(str), dst_type);
4209 }
4210
4211 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4212 /**
4213 * int i_char = start;
4214 * for (int i_byte = 0; i_byte < count; i_byte++) {
4215 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4216 * }
4217 */
4218 add_parse_predicates();
4219 C->set_has_loops(true);
4220
4221 RegionNode* head = new RegionNode(3);
4222 head->init_req(1, control());
4223 gvn().set_type(head, Type::CONTROL);
4224 record_for_igvn(head);
4225
4226 Node* i_byte = new PhiNode(head, TypeInt::INT);
4227 i_byte->init_req(1, intcon(0));
4228 gvn().set_type(i_byte, TypeInt::INT);
4229 record_for_igvn(i_byte);
4230
4231 Node* i_char = new PhiNode(head, TypeInt::INT);
4232 i_char->init_req(1, start);
4233 gvn().set_type(i_char, TypeInt::INT);
4234 record_for_igvn(i_char);
4235
4236 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4237 gvn().set_type(mem, Type::MEMORY);
4238 record_for_igvn(mem);
4239 set_control(head);
4240 set_memory(mem, TypeAryPtr::BYTES);
4241 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4242 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4243 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4244 false, false, true /* mismatched */);
4245
4246 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4247 head->init_req(2, IfTrue(iff));
4248 mem->init_req(2, st);
4249 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4250 i_char->init_req(2, AddI(i_char, intcon(2)));
4251
4252 set_control(IfFalse(iff));
4253 set_memory(st, TypeAryPtr::BYTES);
4254 }
4255
4256 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4257 if (!field->is_constant()) {
4258 return nullptr; // Field not marked as constant.
4259 }
4260 ciInstance* holder = nullptr;
4261 if (!field->is_static()) {
4262 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4263 if (const_oop != nullptr && const_oop->is_instance()) {
4264 holder = const_oop->as_instance();
4265 }
4266 }
4267 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4268 /*is_unsigned_load=*/false);
4269 if (con_type != nullptr) {
4270 return makecon(con_type);
4271 }
4272 return nullptr;
4273 }
4274
4275 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4276 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4277 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4278 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4279 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4280 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4281 return casted_obj;
4282 }
4283 return obj;
4284 }