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