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