1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmClasses.hpp"
29 #include "code/codeCache.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/nmethod.hpp"
32 #include "code/pcDesc.hpp"
33 #include "code/scopeDesc.hpp"
34 #include "compiler/compilationPolicy.hpp"
35 #include "compiler/compilerDefinitions.inline.hpp"
36 #include "gc/shared/collectedHeap.hpp"
37 #include "gc/shared/memAllocator.hpp"
38 #include "interpreter/bytecode.hpp"
39 #include "interpreter/bytecode.inline.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/method.hpp"
55 #include "oops/objArrayKlass.hpp"
56 #include "oops/objArrayOop.inline.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "oops/typeArrayOop.inline.hpp"
59 #include "oops/verifyOopClosure.hpp"
60 #include "prims/jvmtiDeferredUpdates.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "prims/jvmtiThreadState.hpp"
63 #include "prims/methodHandles.hpp"
64 #include "prims/vectorSupport.hpp"
65 #include "runtime/atomic.hpp"
66 #include "runtime/basicLock.inline.hpp"
67 #include "runtime/continuation.hpp"
68 #include "runtime/continuationEntry.inline.hpp"
69 #include "runtime/deoptimization.hpp"
70 #include "runtime/escapeBarrier.hpp"
71 #include "runtime/fieldDescriptor.hpp"
72 #include "runtime/fieldDescriptor.inline.hpp"
73 #include "runtime/frame.inline.hpp"
74 #include "runtime/handles.inline.hpp"
75 #include "runtime/interfaceSupport.inline.hpp"
76 #include "runtime/javaThread.hpp"
77 #include "runtime/jniHandles.inline.hpp"
78 #include "runtime/keepStackGCProcessed.hpp"
79 #include "runtime/lightweightSynchronizer.hpp"
80 #include "runtime/lockStack.inline.hpp"
81 #include "runtime/objectMonitor.inline.hpp"
82 #include "runtime/osThread.hpp"
83 #include "runtime/perfData.inline.hpp"
84 #include "runtime/safepointVerifiers.hpp"
85 #include "runtime/sharedRuntime.hpp"
86 #include "runtime/signature.hpp"
87 #include "runtime/stackFrameStream.inline.hpp"
88 #include "runtime/stackValue.hpp"
89 #include "runtime/stackWatermarkSet.hpp"
90 #include "runtime/stubRoutines.hpp"
91 #include "runtime/synchronizer.inline.hpp"
92 #include "runtime/threadSMR.hpp"
93 #include "runtime/threadWXSetters.inline.hpp"
94 #include "runtime/vframe.hpp"
95 #include "runtime/vframeArray.hpp"
96 #include "runtime/vframe_hp.hpp"
97 #include "runtime/vmOperations.hpp"
98 #include "services/management.hpp"
99 #include "utilities/checkedCast.hpp"
100 #include "utilities/events.hpp"
101 #include "utilities/growableArray.hpp"
102 #include "utilities/macros.hpp"
103 #include "utilities/preserveException.hpp"
104 #include "utilities/xmlstream.hpp"
105 #if INCLUDE_JFR
106 #include "jfr/jfr.inline.hpp"
107 #include "jfr/jfrEvents.hpp"
108 #include "jfr/metadata/jfrSerializer.hpp"
109 #endif
110
111 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
112 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
113 bool DeoptimizationScope::_committing_in_progress = false;
114
115 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
116 DEBUG_ONLY(_deopted = false;)
117
118 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
119 // If there is nothing to deopt _required_gen is the same as comitted.
120 _required_gen = DeoptimizationScope::_committed_deopt_gen;
121 }
122
123 DeoptimizationScope::~DeoptimizationScope() {
124 assert(_deopted, "Deopt not executed");
125 }
126
127 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
128 if (!nm->can_be_deoptimized()) {
129 return;
130 }
131
132 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
133
134 // If it's already marked but we still need it to be deopted.
135 if (nm->is_marked_for_deoptimization()) {
136 dependent(nm);
137 return;
138 }
139
140 nmethod::DeoptimizationStatus status =
141 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
142 Atomic::store(&nm->_deoptimization_status, status);
143
144 // Make sure active is not committed
145 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
146 assert(nm->_deoptimization_generation == 0, "Is already marked");
147
148 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
149 _required_gen = DeoptimizationScope::_active_deopt_gen;
150 }
151
152 void DeoptimizationScope::dependent(nmethod* nm) {
153 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
154
155 // A method marked by someone else may have a _required_gen lower than what we marked with.
156 // Therefore only store it if it's higher than _required_gen.
157 if (_required_gen < nm->_deoptimization_generation) {
158 _required_gen = nm->_deoptimization_generation;
159 }
160 }
161
162 void DeoptimizationScope::deoptimize_marked() {
163 assert(!_deopted, "Already deopted");
164
165 // We are not alive yet.
166 if (!Universe::is_fully_initialized()) {
167 DEBUG_ONLY(_deopted = true;)
168 return;
169 }
170
171 // Safepoints are a special case, handled here.
172 if (SafepointSynchronize::is_at_safepoint()) {
173 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
174 DeoptimizationScope::_active_deopt_gen++;
175 Deoptimization::deoptimize_all_marked();
176 DEBUG_ONLY(_deopted = true;)
177 return;
178 }
179
180 uint64_t comitting = 0;
181 bool wait = false;
182 while (true) {
183 {
184 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
185
186 // First we check if we or someone else already deopted the gen we want.
187 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
188 DEBUG_ONLY(_deopted = true;)
189 return;
190 }
191 if (!_committing_in_progress) {
192 // The version we are about to commit.
193 comitting = DeoptimizationScope::_active_deopt_gen;
194 // Make sure new marks use a higher gen.
195 DeoptimizationScope::_active_deopt_gen++;
196 _committing_in_progress = true;
197 wait = false;
198 } else {
199 // Another thread is handshaking and committing a gen.
200 wait = true;
201 }
202 }
203 if (wait) {
204 // Wait and let the concurrent handshake be performed.
205 ThreadBlockInVM tbivm(JavaThread::current());
206 os::naked_yield();
207 } else {
208 // Performs the handshake.
209 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
210 DEBUG_ONLY(_deopted = true;)
211 {
212 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
213
214 // Make sure that committed doesn't go backwards.
215 // Should only happen if we did a deopt during a safepoint above.
216 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
217 DeoptimizationScope::_committed_deopt_gen = comitting;
218 }
219 _committing_in_progress = false;
220
221 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
222
223 return;
224 }
225 }
226 }
227 }
228
229 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
230 int caller_adjustment,
231 int caller_actual_parameters,
232 int number_of_frames,
233 intptr_t* frame_sizes,
234 address* frame_pcs,
235 BasicType return_type,
236 int exec_mode) {
237 _size_of_deoptimized_frame = size_of_deoptimized_frame;
238 _caller_adjustment = caller_adjustment;
239 _caller_actual_parameters = caller_actual_parameters;
240 _number_of_frames = number_of_frames;
241 _frame_sizes = frame_sizes;
242 _frame_pcs = frame_pcs;
243 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
244 _return_type = return_type;
245 _initial_info = 0;
246 // PD (x86 only)
247 _counter_temp = 0;
248 _unpack_kind = exec_mode;
249 _sender_sp_temp = 0;
250
251 _total_frame_sizes = size_of_frames();
252 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
253 }
254
255 Deoptimization::UnrollBlock::~UnrollBlock() {
256 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
257 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
258 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
259 }
260
261 int Deoptimization::UnrollBlock::size_of_frames() const {
262 // Account first for the adjustment of the initial frame
263 intptr_t result = _caller_adjustment;
264 for (int index = 0; index < number_of_frames(); index++) {
265 result += frame_sizes()[index];
266 }
267 return checked_cast<int>(result);
268 }
269
270 void Deoptimization::UnrollBlock::print() {
271 ResourceMark rm;
272 stringStream st;
273 st.print_cr("UnrollBlock");
274 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
275 st.print( " frame_sizes: ");
276 for (int index = 0; index < number_of_frames(); index++) {
277 st.print("%zd ", frame_sizes()[index]);
278 }
279 st.cr();
280 tty->print_raw(st.freeze());
281 }
282
283 // In order to make fetch_unroll_info work properly with escape
284 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
285 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
286 // which is called from the method fetch_unroll_info_helper below.
287 JRT_BLOCK_ENTRY_PROF(Deoptimization::UnrollBlock*, Deoptimization, fetch_unroll_info, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
288 // fetch_unroll_info() is called at the beginning of the deoptimization
289 // handler. Note this fact before we start generating temporary frames
290 // that can confuse an asynchronous stack walker. This counter is
291 // decremented at the end of unpack_frames().
292 current->inc_in_deopt_handler();
293
294 if (exec_mode == Unpack_exception) {
295 // When we get here, a callee has thrown an exception into a deoptimized
296 // frame. That throw might have deferred stack watermark checking until
297 // after unwinding. So we deal with such deferred requests here.
298 StackWatermarkSet::after_unwind(current);
299 }
300
301 return fetch_unroll_info_helper(current, exec_mode);
302 JRT_END
303
304 #if COMPILER2_OR_JVMCI
305 // print information about reallocated objects
306 static void print_objects(JavaThread* deoptee_thread,
307 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
308 ResourceMark rm;
309 stringStream st; // change to logStream with logging
310 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
311 fieldDescriptor fd;
312
313 for (int i = 0; i < objects->length(); i++) {
314 ObjectValue* sv = (ObjectValue*) objects->at(i);
315 Handle obj = sv->value();
316
317 if (obj.is_null()) {
318 st.print_cr(" nullptr");
319 continue;
320 }
321
322 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
323
324 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
325 k->print_value_on(&st);
326 st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize);
327
328 if (Verbose && k != nullptr) {
329 k->oop_print_on(obj(), &st);
330 }
331 }
332 tty->print_raw(st.freeze());
333 }
334
335 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
336 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
337 bool& deoptimized_objects) {
338 bool realloc_failures = false;
339 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
340
341 JavaThread* deoptee_thread = chunk->at(0)->thread();
342 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
343 "a frame can only be deoptimized by the owner thread");
344
345 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
346
347 // The flag return_oop() indicates call sites which return oop
348 // in compiled code. Such sites include java method calls,
349 // runtime calls (for example, used to allocate new objects/arrays
350 // on slow code path) and any other calls generated in compiled code.
351 // It is not guaranteed that we can get such information here only
352 // by analyzing bytecode in deoptimized frames. This is why this flag
353 // is set during method compilation (see Compile::Process_OopMap_Node()).
354 // If the previous frame was popped or if we are dispatching an exception,
355 // we don't have an oop result.
356 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
357 Handle return_value;
358 if (save_oop_result) {
359 // Reallocation may trigger GC. If deoptimization happened on return from
360 // call which returns oop we need to save it since it is not in oopmap.
361 oop result = deoptee.saved_oop_result(&map);
362 assert(oopDesc::is_oop_or_null(result), "must be oop");
363 return_value = Handle(thread, result);
364 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
365 if (TraceDeoptimization) {
366 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
367 tty->cr();
368 }
369 }
370 if (objects != nullptr) {
371 if (exec_mode == Deoptimization::Unpack_none) {
372 assert(thread->thread_state() == _thread_in_vm, "assumption");
373 JavaThread* THREAD = thread; // For exception macros.
374 // Clear pending OOM if reallocation fails and return true indicating allocation failure
375 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
376 deoptimized_objects = true;
377 } else {
378 JavaThread* current = thread; // For JRT_BLOCK
379 JRT_BLOCK
380 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
381 JRT_END
382 }
383 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
384 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
385 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci);
386 if (TraceDeoptimization) {
387 print_objects(deoptee_thread, objects, realloc_failures);
388 }
389 }
390 if (save_oop_result) {
391 // Restore result.
392 deoptee.set_saved_oop_result(&map, return_value());
393 }
394 return realloc_failures;
395 }
396
397 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
398 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
399 JavaThread* deoptee_thread = chunk->at(0)->thread();
400 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
401 assert(thread == Thread::current(), "should be");
402 HandleMark hm(thread);
403 #ifndef PRODUCT
404 bool first = true;
405 #endif // !PRODUCT
406 // Start locking from outermost/oldest frame
407 for (int i = (chunk->length() - 1); i >= 0; i--) {
408 compiledVFrame* cvf = chunk->at(i);
409 assert (cvf->scope() != nullptr,"expect only compiled java frames");
410 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
411 if (monitors->is_nonempty()) {
412 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
413 exec_mode, realloc_failures);
414 deoptimized_objects = deoptimized_objects || relocked;
415 #ifndef PRODUCT
416 if (PrintDeoptimizationDetails) {
417 ResourceMark rm;
418 stringStream st;
419 for (int j = 0; j < monitors->length(); j++) {
420 MonitorInfo* mi = monitors->at(j);
421 if (mi->eliminated()) {
422 if (first) {
423 first = false;
424 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
425 }
426 if (exec_mode == Deoptimization::Unpack_none) {
427 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
428 if (monitor != nullptr && monitor->object() == mi->owner()) {
429 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
430 continue;
431 }
432 }
433 if (mi->owner_is_scalar_replaced()) {
434 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
435 st.print_cr(" failed reallocation for klass %s", k->external_name());
436 } else {
437 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
438 }
439 }
440 }
441 tty->print_raw(st.freeze());
442 }
443 #endif // !PRODUCT
444 }
445 }
446 }
447
448 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
449 // The given vframes cover one physical frame.
450 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
451 bool& realloc_failures) {
452 frame deoptee = chunk->at(0)->fr();
453 JavaThread* deoptee_thread = chunk->at(0)->thread();
454 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
455 RegisterMap map(chunk->at(0)->register_map());
456 bool deoptimized_objects = false;
457
458 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
459
460 // Reallocate the non-escaping objects and restore their fields.
461 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
462 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
463 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
464 }
465
466 // MonitorInfo structures used in eliminate_locks are not GC safe.
467 NoSafepointVerifier no_safepoint;
468
469 // Now relock objects if synchronization on them was eliminated.
470 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
471 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
472 }
473 return deoptimized_objects;
474 }
475 #endif // COMPILER2_OR_JVMCI
476
477 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
478 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
479 JFR_ONLY(Jfr::check_and_process_sample_request(current);)
480 // When we get here we are about to unwind the deoptee frame. In order to
481 // catch not yet safe to use frames, the following stack watermark barrier
482 // poll will make such frames safe to use.
483 StackWatermarkSet::before_unwind(current);
484
485 // Note: there is a safepoint safety issue here. No matter whether we enter
486 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
487 // the vframeArray is created.
488 //
489
490 // Allocate our special deoptimization ResourceMark
491 DeoptResourceMark* dmark = new DeoptResourceMark(current);
492 assert(current->deopt_mark() == nullptr, "Pending deopt!");
493 current->set_deopt_mark(dmark);
494
495 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
496 RegisterMap map(current,
497 RegisterMap::UpdateMap::include,
498 RegisterMap::ProcessFrames::include,
499 RegisterMap::WalkContinuation::skip);
500 RegisterMap dummy_map(current,
501 RegisterMap::UpdateMap::skip,
502 RegisterMap::ProcessFrames::include,
503 RegisterMap::WalkContinuation::skip);
504 // Now get the deoptee with a valid map
505 frame deoptee = stub_frame.sender(&map);
506 // Set the deoptee nmethod
507 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
508 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
509 current->set_deopt_compiled_method(nm);
510
511 if (VerifyStack) {
512 current->validate_frame_layout();
513 }
514
515 // Create a growable array of VFrames where each VFrame represents an inlined
516 // Java frame. This storage is allocated with the usual system arena.
517 assert(deoptee.is_compiled_frame(), "Wrong frame type");
518 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
519 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
520 while (!vf->is_top()) {
521 assert(vf->is_compiled_frame(), "Wrong frame type");
522 chunk->push(compiledVFrame::cast(vf));
523 vf = vf->sender();
524 }
525 assert(vf->is_compiled_frame(), "Wrong frame type");
526 chunk->push(compiledVFrame::cast(vf));
527
528 bool realloc_failures = false;
529
530 #if COMPILER2_OR_JVMCI
531 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
532
533 // Reallocate the non-escaping objects and restore their fields. Then
534 // relock objects if synchronization on them was eliminated.
535 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
536 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
537 bool unused;
538 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
539 }
540 #endif // COMPILER2_OR_JVMCI
541
542 // Ensure that no safepoint is taken after pointers have been stored
543 // in fields of rematerialized objects. If a safepoint occurs from here on
544 // out the java state residing in the vframeArray will be missed.
545 // Locks may be rebaised in a safepoint.
546 NoSafepointVerifier no_safepoint;
547
548 #if COMPILER2_OR_JVMCI
549 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
550 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
551 bool unused = false;
552 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
553 }
554 #endif // COMPILER2_OR_JVMCI
555
556 ScopeDesc* trap_scope = chunk->at(0)->scope();
557 Handle exceptionObject;
558 if (trap_scope->rethrow_exception()) {
559 #ifndef PRODUCT
560 if (PrintDeoptimizationDetails) {
561 tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci());
562 }
563 #endif // !PRODUCT
564
565 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
566 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
567 ScopeValue* topOfStack = expressions->top();
568 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
569 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
570 }
571
572 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
573 #if COMPILER2_OR_JVMCI
574 if (realloc_failures) {
575 // This destroys all ScopedValue bindings.
576 current->clear_scopedValueBindings();
577 pop_frames_failed_reallocs(current, array);
578 }
579 #endif
580
581 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
582 current->set_vframe_array_head(array);
583
584 // Now that the vframeArray has been created if we have any deferred local writes
585 // added by jvmti then we can free up that structure as the data is now in the
586 // vframeArray
587
588 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
589
590 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
591 CodeBlob* cb = stub_frame.cb();
592 // Verify we have the right vframeArray
593 assert(cb->frame_size() >= 0, "Unexpected frame size");
594 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
595
596 // If the deopt call site is a MethodHandle invoke call site we have
597 // to adjust the unpack_sp.
598 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
599 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
600 unpack_sp = deoptee.unextended_sp();
601
602 #ifdef ASSERT
603 assert(cb->is_deoptimization_stub() ||
604 cb->is_uncommon_trap_stub() ||
605 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
606 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
607 "unexpected code blob: %s", cb->name());
608 #endif
609
610 // This is a guarantee instead of an assert because if vframe doesn't match
611 // we will unpack the wrong deoptimized frame and wind up in strange places
612 // where it will be very difficult to figure out what went wrong. Better
613 // to die an early death here than some very obscure death later when the
614 // trail is cold.
615 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
616
617 int number_of_frames = array->frames();
618
619 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
620 // virtual activation, which is the reverse of the elements in the vframes array.
621 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
622 // +1 because we always have an interpreter return address for the final slot.
623 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
624 int popframe_extra_args = 0;
625 // Create an interpreter return address for the stub to use as its return
626 // address so the skeletal frames are perfectly walkable
627 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
628
629 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
630 // activation be put back on the expression stack of the caller for reexecution
631 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
632 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
633 }
634
635 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
636 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
637 // than simply use array->sender.pc(). This requires us to walk the current set of frames
638 //
639 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
640 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
641
642 // It's possible that the number of parameters at the call site is
643 // different than number of arguments in the callee when method
644 // handles are used. If the caller is interpreted get the real
645 // value so that the proper amount of space can be added to it's
646 // frame.
647 bool caller_was_method_handle = false;
648 if (deopt_sender.is_interpreted_frame()) {
649 methodHandle method(current, deopt_sender.interpreter_frame_method());
650 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
651 if (cur.has_member_arg()) {
652 // This should cover all real-world cases. One exception is a pathological chain of
653 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we
654 // would need to get the size from the resolved method entry. Another exception would
655 // be an invokedynamic with an adapter that is really a MethodHandle linker.
656 caller_was_method_handle = true;
657 }
658 }
659
660 //
661 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
662 // frame_sizes/frame_pcs[1] next oldest frame (int)
663 // frame_sizes/frame_pcs[n] youngest frame (int)
664 //
665 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
666 // owns the space for the return address to it's caller). Confusing ain't it.
667 //
668 // The vframe array can address vframes with indices running from
669 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
670 // When we create the skeletal frames we need the oldest frame to be in the zero slot
671 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
672 // so things look a little strange in this loop.
673 //
674 int callee_parameters = 0;
675 int callee_locals = 0;
676 for (int index = 0; index < array->frames(); index++ ) {
677 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
678 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
679 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
680 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
681 callee_locals,
682 index == 0,
683 popframe_extra_args);
684 // This pc doesn't have to be perfect just good enough to identify the frame
685 // as interpreted so the skeleton frame will be walkable
686 // The correct pc will be set when the skeleton frame is completely filled out
687 // The final pc we store in the loop is wrong and will be overwritten below
688 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
689
690 callee_parameters = array->element(index)->method()->size_of_parameters();
691 callee_locals = array->element(index)->method()->max_locals();
692 popframe_extra_args = 0;
693 }
694
695 // Compute whether the root vframe returns a float or double value.
696 BasicType return_type;
697 {
698 methodHandle method(current, array->element(0)->method());
699 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
700 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
701 }
702
703 // Compute information for handling adapters and adjusting the frame size of the caller.
704 int caller_adjustment = 0;
705
706 // Compute the amount the oldest interpreter frame will have to adjust
707 // its caller's stack by. If the caller is a compiled frame then
708 // we pretend that the callee has no parameters so that the
709 // extension counts for the full amount of locals and not just
710 // locals-parms. This is because without a c2i adapter the parm
711 // area as created by the compiled frame will not be usable by
712 // the interpreter. (Depending on the calling convention there
713 // may not even be enough space).
714
715 // QQQ I'd rather see this pushed down into last_frame_adjust
716 // and have it take the sender (aka caller).
717
718 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
719 caller_adjustment = last_frame_adjust(0, callee_locals);
720 } else if (callee_locals > callee_parameters) {
721 // The caller frame may need extending to accommodate
722 // non-parameter locals of the first unpacked interpreted frame.
723 // Compute that adjustment.
724 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
725 }
726
727 // If the sender is deoptimized the we must retrieve the address of the handler
728 // since the frame will "magically" show the original pc before the deopt
729 // and we'd undo the deopt.
730
731 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
732 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
733 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
734 }
735
736 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
737
738 #if INCLUDE_JVMCI
739 if (exceptionObject() != nullptr) {
740 current->set_exception_oop(exceptionObject());
741 exec_mode = Unpack_exception;
742 }
743 #endif
744
745 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
746 assert(current->has_pending_exception(), "should have thrown OOME");
747 current->set_exception_oop(current->pending_exception());
748 current->clear_pending_exception();
749 exec_mode = Unpack_exception;
750 }
751
752 #if INCLUDE_JVMCI
753 if (current->frames_to_pop_failed_realloc() > 0) {
754 current->set_pending_monitorenter(false);
755 }
756 #endif
757
758 int caller_actual_parameters = -1; // value not used except for interpreted frames, see below
759 if (deopt_sender.is_interpreted_frame()) {
760 caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0);
761 }
762
763 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
764 caller_adjustment * BytesPerWord,
765 caller_actual_parameters,
766 number_of_frames,
767 frame_sizes,
768 frame_pcs,
769 return_type,
770 exec_mode);
771 // On some platforms, we need a way to pass some platform dependent
772 // information to the unpacking code so the skeletal frames come out
773 // correct (initial fp value, unextended sp, ...)
774 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
775
776 if (array->frames() > 1) {
777 if (VerifyStack && TraceDeoptimization) {
778 tty->print_cr("Deoptimizing method containing inlining");
779 }
780 }
781
782 array->set_unroll_block(info);
783 return info;
784 }
785
786 // Called to cleanup deoptimization data structures in normal case
787 // after unpacking to stack and when stack overflow error occurs
788 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
789 vframeArray *array) {
790
791 // Get array if coming from exception
792 if (array == nullptr) {
793 array = thread->vframe_array_head();
794 }
795 thread->set_vframe_array_head(nullptr);
796
797 // Free the previous UnrollBlock
798 vframeArray* old_array = thread->vframe_array_last();
799 thread->set_vframe_array_last(array);
800
801 if (old_array != nullptr) {
802 UnrollBlock* old_info = old_array->unroll_block();
803 old_array->set_unroll_block(nullptr);
804 delete old_info;
805 delete old_array;
806 }
807
808 // Deallocate any resource creating in this routine and any ResourceObjs allocated
809 // inside the vframeArray (StackValueCollections)
810
811 delete thread->deopt_mark();
812 thread->set_deopt_mark(nullptr);
813 thread->set_deopt_compiled_method(nullptr);
814
815
816 if (JvmtiExport::can_pop_frame()) {
817 // Regardless of whether we entered this routine with the pending
818 // popframe condition bit set, we should always clear it now
819 thread->clear_popframe_condition();
820 }
821
822 // unpack_frames() is called at the end of the deoptimization handler
823 // and (in C2) at the end of the uncommon trap handler. Note this fact
824 // so that an asynchronous stack walker can work again. This counter is
825 // incremented at the beginning of fetch_unroll_info() and (in C2) at
826 // the beginning of uncommon_trap().
827 thread->dec_in_deopt_handler();
828 }
829
830 // Moved from cpu directories because none of the cpus has callee save values.
831 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
832 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
833
834 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
835 // the days we had adapter frames. When we deoptimize a situation where a
836 // compiled caller calls a compiled caller will have registers it expects
837 // to survive the call to the callee. If we deoptimize the callee the only
838 // way we can restore these registers is to have the oldest interpreter
839 // frame that we create restore these values. That is what this routine
840 // will accomplish.
841
842 // At the moment we have modified c2 to not have any callee save registers
843 // so this problem does not exist and this routine is just a place holder.
844
845 assert(f->is_interpreted_frame(), "must be interpreted");
846 }
847
848 #ifndef PRODUCT
849 // Return true if the execution after the provided bytecode continues at the
850 // next bytecode in the code. This is not the case for gotos, returns, and
851 // throws.
852 static bool falls_through(Bytecodes::Code bc) {
853 switch (bc) {
854 case Bytecodes::_goto:
855 case Bytecodes::_goto_w:
856 case Bytecodes::_athrow:
857 case Bytecodes::_areturn:
858 case Bytecodes::_dreturn:
859 case Bytecodes::_freturn:
860 case Bytecodes::_ireturn:
861 case Bytecodes::_lreturn:
862 case Bytecodes::_jsr:
863 case Bytecodes::_ret:
864 case Bytecodes::_return:
865 case Bytecodes::_lookupswitch:
866 case Bytecodes::_tableswitch:
867 return false;
868 default:
869 return true;
870 }
871 }
872 #endif
873
874 // Return BasicType of value being returned
875 JRT_LEAF_PROF_NO_THREAD(BasicType, Deoptimization, unpack_frames, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
876 assert(thread == JavaThread::current(), "pre-condition");
877
878 // We are already active in the special DeoptResourceMark any ResourceObj's we
879 // allocate will be freed at the end of the routine.
880
881 // JRT_LEAF methods don't normally allocate handles and there is a
882 // NoHandleMark to enforce that. It is actually safe to use Handles
883 // in a JRT_LEAF method, and sometimes desirable, but to do so we
884 // must use ResetNoHandleMark to bypass the NoHandleMark, and
885 // then use a HandleMark to ensure any Handles we do create are
886 // cleaned up in this scope.
887 ResetNoHandleMark rnhm;
888 HandleMark hm(thread);
889
890 frame stub_frame = thread->last_frame();
891
892 Continuation::notify_deopt(thread, stub_frame.sp());
893
894 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
895 // must point to the vframeArray for the unpack frame.
896 vframeArray* array = thread->vframe_array_head();
897 UnrollBlock* info = array->unroll_block();
898
899 // We set the last_Java frame. But the stack isn't really parsable here. So we
900 // clear it to make sure JFR understands not to try and walk stacks from events
901 // in here.
902 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
903 thread->frame_anchor()->set_last_Java_sp(nullptr);
904
905 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
906 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
907
908 thread->frame_anchor()->set_last_Java_sp(sp);
909
910 BasicType bt = info->return_type();
911
912 // If we have an exception pending, claim that the return type is an oop
913 // so the deopt_blob does not overwrite the exception_oop.
914
915 if (exec_mode == Unpack_exception)
916 bt = T_OBJECT;
917
918 // Cleanup thread deopt data
919 cleanup_deopt_info(thread, array);
920
921 #ifndef PRODUCT
922 if (VerifyStack) {
923 ResourceMark res_mark;
924 // Clear pending exception to not break verification code (restored afterwards)
925 PreserveExceptionMark pm(thread);
926
927 thread->validate_frame_layout();
928
929 // Verify that the just-unpacked frames match the interpreter's
930 // notions of expression stack and locals
931 vframeArray* cur_array = thread->vframe_array_last();
932 RegisterMap rm(thread,
933 RegisterMap::UpdateMap::skip,
934 RegisterMap::ProcessFrames::include,
935 RegisterMap::WalkContinuation::skip);
936 rm.set_include_argument_oops(false);
937 bool is_top_frame = true;
938 int callee_size_of_parameters = 0;
939 int callee_max_locals = 0;
940 for (int i = 0; i < cur_array->frames(); i++) {
941 vframeArrayElement* el = cur_array->element(i);
942 frame* iframe = el->iframe();
943 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
944
945 // Get the oop map for this bci
946 InterpreterOopMap mask;
947 int cur_invoke_parameter_size = 0;
948 bool try_next_mask = false;
949 int next_mask_expression_stack_size = -1;
950 int top_frame_expression_stack_adjustment = 0;
951 methodHandle mh(thread, iframe->interpreter_frame_method());
952 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
953 BytecodeStream str(mh, iframe->interpreter_frame_bci());
954 int max_bci = mh->code_size();
955 // Get to the next bytecode if possible
956 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
957 // Check to see if we can grab the number of outgoing arguments
958 // at an uncommon trap for an invoke (where the compiler
959 // generates debug info before the invoke has executed)
960 Bytecodes::Code cur_code = str.next();
961 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
962 if (Bytecodes::is_invoke(cur_code)) {
963 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
964 cur_invoke_parameter_size = invoke.size_of_parameters();
965 if (i != 0 && invoke.has_member_arg()) {
966 callee_size_of_parameters++;
967 }
968 }
969 if (str.bci() < max_bci) {
970 next_code = str.next();
971 if (next_code >= 0) {
972 // The interpreter oop map generator reports results before
973 // the current bytecode has executed except in the case of
974 // calls. It seems to be hard to tell whether the compiler
975 // has emitted debug information matching the "state before"
976 // a given bytecode or the state after, so we try both
977 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
978 // Get expression stack size for the next bytecode
979 InterpreterOopMap next_mask;
980 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
981 next_mask_expression_stack_size = next_mask.expression_stack_size();
982 if (Bytecodes::is_invoke(next_code)) {
983 Bytecode_invoke invoke(mh, str.bci());
984 next_mask_expression_stack_size += invoke.size_of_parameters();
985 }
986 // Need to subtract off the size of the result type of
987 // the bytecode because this is not described in the
988 // debug info but returned to the interpreter in the TOS
989 // caching register
990 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
991 if (bytecode_result_type != T_ILLEGAL) {
992 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
993 }
994 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
995 try_next_mask = true;
996 }
997 }
998 }
999
1000 // Verify stack depth and oops in frame
1001 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
1002 if (!(
1003 /* SPARC */
1004 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
1005 /* x86 */
1006 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
1007 (try_next_mask &&
1008 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
1009 top_frame_expression_stack_adjustment))) ||
1010 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
1011 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
1012 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
1013 )) {
1014 {
1015 // Print out some information that will help us debug the problem
1016 tty->print_cr("Wrong number of expression stack elements during deoptimization");
1017 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
1018 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1019 if (try_next_mask) {
1020 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1021 }
1022 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1023 iframe->interpreter_frame_expression_stack_size());
1024 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1025 tty->print_cr(" try_next_mask = %d", try_next_mask);
1026 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1027 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1028 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1029 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1030 tty->print_cr(" exec_mode = %d", exec_mode);
1031 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1032 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1033 tty->print_cr(" Interpreted frames:");
1034 for (int k = 0; k < cur_array->frames(); k++) {
1035 vframeArrayElement* el = cur_array->element(k);
1036 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1037 }
1038 cur_array->print_on_2(tty);
1039 }
1040 guarantee(false, "wrong number of expression stack elements during deopt");
1041 }
1042 VerifyOopClosure verify;
1043 iframe->oops_interpreted_do(&verify, &rm, false);
1044 callee_size_of_parameters = mh->size_of_parameters();
1045 callee_max_locals = mh->max_locals();
1046 is_top_frame = false;
1047 }
1048 }
1049 #endif // !PRODUCT
1050
1051 return bt;
1052 JRT_END
1053
1054 class DeoptimizeMarkedClosure : public HandshakeClosure {
1055 public:
1056 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1057 void do_thread(Thread* thread) {
1058 JavaThread* jt = JavaThread::cast(thread);
1059 jt->deoptimize_marked_methods();
1060 }
1061 };
1062
1063 void Deoptimization::deoptimize_all_marked() {
1064 ResourceMark rm;
1065
1066 // Make the dependent methods not entrant
1067 CodeCache::make_marked_nmethods_deoptimized();
1068
1069 DeoptimizeMarkedClosure deopt;
1070 if (SafepointSynchronize::is_at_safepoint()) {
1071 Threads::java_threads_do(&deopt);
1072 } else {
1073 Handshake::execute(&deopt);
1074 }
1075 }
1076
1077 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1078 = Deoptimization::Action_reinterpret;
1079
1080 #if INCLUDE_JVMCI
1081 template<typename CacheType>
1082 class BoxCacheBase : public CHeapObj<mtCompiler> {
1083 protected:
1084 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1085 ResourceMark rm(thread);
1086 char* klass_name_str = klass_name->as_C_string();
1087 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1088 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1089 if (!ik->is_in_error_state()) {
1090 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1091 CacheType::compute_offsets(ik);
1092 }
1093 return ik;
1094 }
1095 };
1096
1097 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1098 PrimitiveType _low;
1099 PrimitiveType _high;
1100 jobject _cache;
1101 protected:
1102 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1103 BoxCache(Thread* thread) {
1104 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1105 if (ik->is_in_error_state()) {
1106 _low = 1;
1107 _high = 0;
1108 _cache = nullptr;
1109 } else {
1110 objArrayOop cache = CacheType::cache(ik);
1111 assert(cache->length() > 0, "Empty cache");
1112 _low = BoxType::value(cache->obj_at(0));
1113 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1114 _cache = JNIHandles::make_global(Handle(thread, cache));
1115 }
1116 }
1117 ~BoxCache() {
1118 JNIHandles::destroy_global(_cache);
1119 }
1120 public:
1121 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1122 if (_singleton == nullptr) {
1123 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1124 if (!Atomic::replace_if_null(&_singleton, s)) {
1125 delete s;
1126 }
1127 }
1128 return _singleton;
1129 }
1130 oop lookup(PrimitiveType value) {
1131 if (_low <= value && value <= _high) {
1132 int offset = checked_cast<int>(value - _low);
1133 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1134 }
1135 return nullptr;
1136 }
1137 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1138 if (_cache == nullptr) {
1139 cache_init_error = true;
1140 return nullptr;
1141 }
1142 // Have to cast to avoid little/big-endian problems.
1143 if (sizeof(PrimitiveType) > sizeof(jint)) {
1144 jlong value = (jlong)raw_value;
1145 return lookup(value);
1146 }
1147 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1148 return lookup(value);
1149 }
1150 };
1151
1152 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1153 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1154 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1155 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1156 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1157
1158 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1159 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1160 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1161 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1162 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1163
1164 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1165 jobject _true_cache;
1166 jobject _false_cache;
1167 protected:
1168 static BooleanBoxCache *_singleton;
1169 BooleanBoxCache(Thread *thread) {
1170 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1171 if (ik->is_in_error_state()) {
1172 _true_cache = nullptr;
1173 _false_cache = nullptr;
1174 } else {
1175 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1176 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1177 }
1178 }
1179 ~BooleanBoxCache() {
1180 JNIHandles::destroy_global(_true_cache);
1181 JNIHandles::destroy_global(_false_cache);
1182 }
1183 public:
1184 static BooleanBoxCache* singleton(Thread* thread) {
1185 if (_singleton == nullptr) {
1186 BooleanBoxCache* s = new BooleanBoxCache(thread);
1187 if (!Atomic::replace_if_null(&_singleton, s)) {
1188 delete s;
1189 }
1190 }
1191 return _singleton;
1192 }
1193 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1194 if (_true_cache == nullptr) {
1195 cache_in_error = true;
1196 return nullptr;
1197 }
1198 // Have to cast to avoid little/big-endian problems.
1199 jboolean value = (jboolean)*((jint*)&raw_value);
1200 return lookup(value);
1201 }
1202 oop lookup(jboolean value) {
1203 if (value != 0) {
1204 return JNIHandles::resolve_non_null(_true_cache);
1205 }
1206 return JNIHandles::resolve_non_null(_false_cache);
1207 }
1208 };
1209
1210 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1211
1212 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1213 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1214 BasicType box_type = vmClasses::box_klass_type(k);
1215 if (box_type != T_OBJECT) {
1216 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1217 switch(box_type) {
1218 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1219 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1220 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1221 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1222 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1223 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1224 default:;
1225 }
1226 }
1227 return nullptr;
1228 }
1229 #endif // INCLUDE_JVMCI
1230
1231 #if COMPILER2_OR_JVMCI
1232 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1233 Handle pending_exception(THREAD, thread->pending_exception());
1234 const char* exception_file = thread->exception_file();
1235 int exception_line = thread->exception_line();
1236 thread->clear_pending_exception();
1237
1238 bool failures = false;
1239
1240 for (int i = 0; i < objects->length(); i++) {
1241 assert(objects->at(i)->is_object(), "invalid debug information");
1242 ObjectValue* sv = (ObjectValue*) objects->at(i);
1243
1244 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1245 oop obj = nullptr;
1246
1247 bool cache_init_error = false;
1248 if (k->is_instance_klass()) {
1249 #if INCLUDE_JVMCI
1250 nmethod* nm = fr->cb()->as_nmethod_or_null();
1251 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1252 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1253 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1254 if (obj != nullptr) {
1255 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1256 abv->set_cached(true);
1257 } else if (cache_init_error) {
1258 // Results in an OOME which is valid (as opposed to a class initialization error)
1259 // and is fine for the rare case a cache initialization failing.
1260 failures = true;
1261 }
1262 }
1263 #endif // INCLUDE_JVMCI
1264
1265 InstanceKlass* ik = InstanceKlass::cast(k);
1266 if (obj == nullptr && !cache_init_error) {
1267 InternalOOMEMark iom(THREAD);
1268 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1269 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1270 } else {
1271 obj = ik->allocate_instance(THREAD);
1272 }
1273 }
1274 } else if (k->is_typeArray_klass()) {
1275 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1276 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1277 int len = sv->field_size() / type2size[ak->element_type()];
1278 InternalOOMEMark iom(THREAD);
1279 obj = ak->allocate(len, THREAD);
1280 } else if (k->is_objArray_klass()) {
1281 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1282 InternalOOMEMark iom(THREAD);
1283 obj = ak->allocate(sv->field_size(), THREAD);
1284 }
1285
1286 if (obj == nullptr) {
1287 failures = true;
1288 }
1289
1290 assert(sv->value().is_null(), "redundant reallocation");
1291 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1292 CLEAR_PENDING_EXCEPTION;
1293 sv->set_value(obj);
1294 }
1295
1296 if (failures) {
1297 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1298 } else if (pending_exception.not_null()) {
1299 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1300 }
1301
1302 return failures;
1303 }
1304
1305 #if INCLUDE_JVMCI
1306 /**
1307 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1308 * we need to somehow be able to recover the actual kind to be able to write the correct
1309 * amount of bytes.
1310 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1311 * the entries at index n + 1 to n + i are 'markers'.
1312 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1313 * expected form of the array would be:
1314 *
1315 * {b0, b1, b2, b3, INT, marker, b6, b7}
1316 *
1317 * Thus, in order to get back the size of the entry, we simply need to count the number
1318 * of marked entries
1319 *
1320 * @param virtualArray the virtualized byte array
1321 * @param i index of the virtual entry we are recovering
1322 * @return The number of bytes the entry spans
1323 */
1324 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1325 int index = i;
1326 while (++index < virtualArray->field_size() &&
1327 virtualArray->field_at(index)->is_marker()) {}
1328 return index - i;
1329 }
1330
1331 /**
1332 * If there was a guarantee for byte array to always start aligned to a long, we could
1333 * do a simple check on the parity of the index. Unfortunately, that is not always the
1334 * case. Thus, we check alignment of the actual address we are writing to.
1335 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1336 * write a long to index 3.
1337 */
1338 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1339 jbyte* res = obj->byte_at_addr(index);
1340 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1341 return res;
1342 }
1343
1344 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1345 switch (byte_count) {
1346 case 1:
1347 obj->byte_at_put(index, (jbyte) value->get_jint());
1348 break;
1349 case 2:
1350 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1351 break;
1352 case 4:
1353 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1354 break;
1355 case 8:
1356 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1357 break;
1358 default:
1359 ShouldNotReachHere();
1360 }
1361 }
1362 #endif // INCLUDE_JVMCI
1363
1364
1365 // restore elements of an eliminated type array
1366 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1367 int index = 0;
1368
1369 for (int i = 0; i < sv->field_size(); i++) {
1370 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1371 switch(type) {
1372 case T_LONG: case T_DOUBLE: {
1373 assert(value->type() == T_INT, "Agreement.");
1374 StackValue* low =
1375 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1376 #ifdef _LP64
1377 jlong res = (jlong)low->get_intptr();
1378 #else
1379 jlong res = jlong_from(value->get_jint(), low->get_jint());
1380 #endif
1381 obj->long_at_put(index, res);
1382 break;
1383 }
1384
1385 case T_INT: case T_FLOAT: { // 4 bytes.
1386 assert(value->type() == T_INT, "Agreement.");
1387 bool big_value = false;
1388 if (i + 1 < sv->field_size() && type == T_INT) {
1389 if (sv->field_at(i)->is_location()) {
1390 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1391 if (type == Location::dbl || type == Location::lng) {
1392 big_value = true;
1393 }
1394 } else if (sv->field_at(i)->is_constant_int()) {
1395 ScopeValue* next_scope_field = sv->field_at(i + 1);
1396 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1397 big_value = true;
1398 }
1399 }
1400 }
1401
1402 if (big_value) {
1403 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1404 #ifdef _LP64
1405 jlong res = (jlong)low->get_intptr();
1406 #else
1407 jlong res = jlong_from(value->get_jint(), low->get_jint());
1408 #endif
1409 obj->int_at_put(index, *(jint*)&res);
1410 obj->int_at_put(++index, *((jint*)&res + 1));
1411 } else {
1412 obj->int_at_put(index, value->get_jint());
1413 }
1414 break;
1415 }
1416
1417 case T_SHORT:
1418 assert(value->type() == T_INT, "Agreement.");
1419 obj->short_at_put(index, (jshort)value->get_jint());
1420 break;
1421
1422 case T_CHAR:
1423 assert(value->type() == T_INT, "Agreement.");
1424 obj->char_at_put(index, (jchar)value->get_jint());
1425 break;
1426
1427 case T_BYTE: {
1428 assert(value->type() == T_INT, "Agreement.");
1429 #if INCLUDE_JVMCI
1430 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1431 int byte_count = count_number_of_bytes_for_entry(sv, i);
1432 byte_array_put(obj, value, index, byte_count);
1433 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1434 i += byte_count - 1; // Balance the loop counter.
1435 index += byte_count;
1436 // index has been updated so continue at top of loop
1437 continue;
1438 #else
1439 obj->byte_at_put(index, (jbyte)value->get_jint());
1440 break;
1441 #endif // INCLUDE_JVMCI
1442 }
1443
1444 case T_BOOLEAN: {
1445 assert(value->type() == T_INT, "Agreement.");
1446 obj->bool_at_put(index, (jboolean)value->get_jint());
1447 break;
1448 }
1449
1450 default:
1451 ShouldNotReachHere();
1452 }
1453 index++;
1454 }
1455 }
1456
1457 // restore fields of an eliminated object array
1458 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1459 for (int i = 0; i < sv->field_size(); i++) {
1460 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1461 assert(value->type() == T_OBJECT, "object element expected");
1462 obj->obj_at_put(i, value->get_obj()());
1463 }
1464 }
1465
1466 class ReassignedField {
1467 public:
1468 int _offset;
1469 BasicType _type;
1470 public:
1471 ReassignedField() {
1472 _offset = 0;
1473 _type = T_ILLEGAL;
1474 }
1475 };
1476
1477 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1478 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1479 InstanceKlass* super = klass->superklass();
1480 if (super != nullptr) {
1481 get_reassigned_fields(super, fields, is_jvmci);
1482 }
1483 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1484 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1485 ReassignedField field;
1486 field._offset = fs.offset();
1487 field._type = Signature::basic_type(fs.signature());
1488 fields->append(field);
1489 }
1490 }
1491 return fields;
1492 }
1493
1494 // Restore fields of an eliminated instance object employing the same field order used by the compiler.
1495 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool is_jvmci) {
1496 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1497 for (int i = 0; i < fields->length(); i++) {
1498 ScopeValue* scope_field = sv->field_at(svIndex);
1499 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1500 int offset = fields->at(i)._offset;
1501 BasicType type = fields->at(i)._type;
1502 switch (type) {
1503 case T_OBJECT: case T_ARRAY:
1504 assert(value->type() == T_OBJECT, "Agreement.");
1505 obj->obj_field_put(offset, value->get_obj()());
1506 break;
1507
1508 case T_INT: case T_FLOAT: { // 4 bytes.
1509 assert(value->type() == T_INT, "Agreement.");
1510 bool big_value = false;
1511 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1512 if (scope_field->is_location()) {
1513 Location::Type type = ((LocationValue*) scope_field)->location().type();
1514 if (type == Location::dbl || type == Location::lng) {
1515 big_value = true;
1516 }
1517 }
1518 if (scope_field->is_constant_int()) {
1519 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1520 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1521 big_value = true;
1522 }
1523 }
1524 }
1525
1526 if (big_value) {
1527 i++;
1528 assert(i < fields->length(), "second T_INT field needed");
1529 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1530 } else {
1531 obj->int_field_put(offset, value->get_jint());
1532 break;
1533 }
1534 }
1535 /* no break */
1536
1537 case T_LONG: case T_DOUBLE: {
1538 assert(value->type() == T_INT, "Agreement.");
1539 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1540 #ifdef _LP64
1541 jlong res = (jlong)low->get_intptr();
1542 #else
1543 jlong res = jlong_from(value->get_jint(), low->get_jint());
1544 #endif
1545 obj->long_field_put(offset, res);
1546 break;
1547 }
1548
1549 case T_SHORT:
1550 assert(value->type() == T_INT, "Agreement.");
1551 obj->short_field_put(offset, (jshort)value->get_jint());
1552 break;
1553
1554 case T_CHAR:
1555 assert(value->type() == T_INT, "Agreement.");
1556 obj->char_field_put(offset, (jchar)value->get_jint());
1557 break;
1558
1559 case T_BYTE:
1560 assert(value->type() == T_INT, "Agreement.");
1561 obj->byte_field_put(offset, (jbyte)value->get_jint());
1562 break;
1563
1564 case T_BOOLEAN:
1565 assert(value->type() == T_INT, "Agreement.");
1566 obj->bool_field_put(offset, (jboolean)value->get_jint());
1567 break;
1568
1569 default:
1570 ShouldNotReachHere();
1571 }
1572 svIndex++;
1573 }
1574 return svIndex;
1575 }
1576
1577 // restore fields of all eliminated objects and arrays
1578 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci) {
1579 for (int i = 0; i < objects->length(); i++) {
1580 assert(objects->at(i)->is_object(), "invalid debug information");
1581 ObjectValue* sv = (ObjectValue*) objects->at(i);
1582 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1583 Handle obj = sv->value();
1584 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1585 #ifndef PRODUCT
1586 if (PrintDeoptimizationDetails) {
1587 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1588 }
1589 #endif // !PRODUCT
1590
1591 if (obj.is_null()) {
1592 continue;
1593 }
1594
1595 #if INCLUDE_JVMCI
1596 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1597 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1598 continue;
1599 }
1600 #endif // INCLUDE_JVMCI
1601 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1602 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1603 ScopeValue* payload = sv->field_at(0);
1604 if (payload->is_location() &&
1605 payload->as_LocationValue()->location().type() == Location::vector) {
1606 #ifndef PRODUCT
1607 if (PrintDeoptimizationDetails) {
1608 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1609 if (Verbose) {
1610 Handle obj = sv->value();
1611 k->oop_print_on(obj(), tty);
1612 }
1613 }
1614 #endif // !PRODUCT
1615 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1616 }
1617 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1618 // which could be restored after vector object allocation.
1619 }
1620 if (k->is_instance_klass()) {
1621 InstanceKlass* ik = InstanceKlass::cast(k);
1622 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci);
1623 } else if (k->is_typeArray_klass()) {
1624 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1625 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1626 } else if (k->is_objArray_klass()) {
1627 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1628 }
1629 }
1630 // These objects may escape when we return to Interpreter after deoptimization.
1631 // We need barrier so that stores that initialize these objects can't be reordered
1632 // with subsequent stores that make these objects accessible by other threads.
1633 OrderAccess::storestore();
1634 }
1635
1636
1637 // relock objects for which synchronization was eliminated
1638 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1639 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1640 bool relocked_objects = false;
1641 for (int i = 0; i < monitors->length(); i++) {
1642 MonitorInfo* mon_info = monitors->at(i);
1643 if (mon_info->eliminated()) {
1644 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1645 relocked_objects = true;
1646 if (!mon_info->owner_is_scalar_replaced()) {
1647 Handle obj(thread, mon_info->owner());
1648 markWord mark = obj->mark();
1649 if (exec_mode == Unpack_none) {
1650 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1651 // With exec_mode == Unpack_none obj may be thread local and locked in
1652 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1653 markWord dmw = mark.displaced_mark_helper();
1654 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1655 obj->set_mark(dmw);
1656 }
1657 if (mark.has_monitor()) {
1658 // defer relocking if the deoptee thread is currently waiting for obj
1659 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1660 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1661 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1662 if (LockingMode == LM_LEGACY) {
1663 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1664 } else if (UseObjectMonitorTable) {
1665 mon_info->lock()->clear_object_monitor_cache();
1666 }
1667 #ifdef ASSERT
1668 else {
1669 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1670 mon_info->lock()->set_bad_metadata_deopt();
1671 }
1672 #endif
1673 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1674 continue;
1675 }
1676 }
1677 }
1678 BasicLock* lock = mon_info->lock();
1679 if (LockingMode == LM_LIGHTWEIGHT) {
1680 // We have lost information about the correct state of the lock stack.
1681 // Entering may create an invalid lock stack. Inflate the lock if it
1682 // was fast_locked to restore the valid lock stack.
1683 if (UseObjectMonitorTable) {
1684 // UseObjectMonitorTable expects the BasicLock cache to be either a
1685 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1686 // to nullptr.
1687 lock->clear_object_monitor_cache();
1688 }
1689 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1690 if (deoptee_thread->lock_stack().contains(obj())) {
1691 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1692 deoptee_thread, thread);
1693 }
1694 assert(mon_info->owner()->is_locked(), "object must be locked now");
1695 assert(obj->mark().has_monitor(), "must be");
1696 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1697 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1698 } else {
1699 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1700 assert(mon_info->owner()->is_locked(), "object must be locked now");
1701 }
1702 }
1703 }
1704 }
1705 return relocked_objects;
1706 }
1707 #endif // COMPILER2_OR_JVMCI
1708
1709 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1710 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1711
1712 // Register map for next frame (used for stack crawl). We capture
1713 // the state of the deopt'ing frame's caller. Thus if we need to
1714 // stuff a C2I adapter we can properly fill in the callee-save
1715 // register locations.
1716 frame caller = fr.sender(reg_map);
1717 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1718
1719 frame sender = caller;
1720
1721 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1722 // the vframeArray containing the unpacking information is allocated in the C heap.
1723 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1724 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1725
1726 // Compare the vframeArray to the collected vframes
1727 assert(array->structural_compare(thread, chunk), "just checking");
1728
1729 if (TraceDeoptimization) {
1730 ResourceMark rm;
1731 stringStream st;
1732 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1733 st.print(" ");
1734 fr.print_on(&st);
1735 st.print_cr(" Virtual frames (innermost/newest first):");
1736 for (int index = 0; index < chunk->length(); index++) {
1737 compiledVFrame* vf = chunk->at(index);
1738 int bci = vf->raw_bci();
1739 const char* code_name;
1740 if (bci == SynchronizationEntryBCI) {
1741 code_name = "sync entry";
1742 } else {
1743 Bytecodes::Code code = vf->method()->code_at(bci);
1744 code_name = Bytecodes::name(code);
1745 }
1746
1747 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1748 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1749 st.print(" - %s", code_name);
1750 st.print_cr(" @ bci=%d ", bci);
1751 }
1752 tty->print_raw(st.freeze());
1753 tty->cr();
1754 }
1755
1756 return array;
1757 }
1758
1759 #if COMPILER2_OR_JVMCI
1760 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1761 // Reallocation of some scalar replaced objects failed. Record
1762 // that we need to pop all the interpreter frames for the
1763 // deoptimized compiled frame.
1764 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1765 thread->set_frames_to_pop_failed_realloc(array->frames());
1766 // Unlock all monitors here otherwise the interpreter will see a
1767 // mix of locked and unlocked monitors (because of failed
1768 // reallocations of synchronized objects) and be confused.
1769 for (int i = 0; i < array->frames(); i++) {
1770 MonitorChunk* monitors = array->element(i)->monitors();
1771 if (monitors != nullptr) {
1772 // Unlock in reverse order starting from most nested monitor.
1773 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1774 BasicObjectLock* src = monitors->at(j);
1775 if (src->obj() != nullptr) {
1776 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1777 }
1778 }
1779 array->element(i)->free_monitors();
1780 #ifdef ASSERT
1781 array->element(i)->set_removed_monitors();
1782 #endif
1783 }
1784 }
1785 }
1786 #endif
1787
1788 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1789 assert(fr.can_be_deoptimized(), "checking frame type");
1790
1791 nmethod* nm = fr.cb()->as_nmethod_or_null();
1792 assert(nm != nullptr, "only compiled methods can deopt");
1793 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none);
1794 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc());
1795 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry
1796 : cur_sd->method()->java_code_at(cur_sd->bci()));
1797 gather_statistics(nm, reason, action, bc);
1798
1799 if (LogCompilation && xtty != nullptr) {
1800 ttyLocker ttyl;
1801 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1802 nm->log_identity(xtty);
1803 xtty->end_head();
1804 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1805 xtty->begin_elem("jvms bci='%d'", sd->bci());
1806 xtty->method(sd->method());
1807 xtty->end_elem();
1808 if (sd->is_top()) break;
1809 }
1810 xtty->tail("deoptimized");
1811 }
1812
1813 Continuation::notify_deopt(thread, fr.sp());
1814
1815 // Patch the compiled method so that when execution returns to it we will
1816 // deopt the execution state and return to the interpreter.
1817 fr.deoptimize(thread);
1818 }
1819
1820 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1821 // Deoptimize only if the frame comes from compile code.
1822 // Do not deoptimize the frame which is already patched
1823 // during the execution of the loops below.
1824 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1825 return;
1826 }
1827 ResourceMark rm;
1828 deoptimize_single_frame(thread, fr, reason);
1829 }
1830
1831 #if INCLUDE_JVMCI
1832 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1833 // there is no exception handler for this pc => deoptimize
1834 nm->make_not_entrant("missing exception handler");
1835
1836 // Use Deoptimization::deoptimize for all of its side-effects:
1837 // gathering traps statistics, logging...
1838 // it also patches the return pc but we do not care about that
1839 // since we return a continuation to the deopt_blob below.
1840 JavaThread* thread = JavaThread::current();
1841 RegisterMap reg_map(thread,
1842 RegisterMap::UpdateMap::skip,
1843 RegisterMap::ProcessFrames::include,
1844 RegisterMap::WalkContinuation::skip);
1845 frame runtime_frame = thread->last_frame();
1846 frame caller_frame = runtime_frame.sender(®_map);
1847 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1848 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1849 compiledVFrame* cvf = compiledVFrame::cast(vf);
1850 ScopeDesc* imm_scope = cvf->scope();
1851 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1852 if (imm_mdo != nullptr) {
1853 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1854 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1855
1856 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1857 if (pdata != nullptr && pdata->is_BitData()) {
1858 BitData* bit_data = (BitData*) pdata;
1859 bit_data->set_exception_seen();
1860 }
1861 }
1862
1863 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1864
1865 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1866 if (trap_mdo != nullptr) {
1867 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1868 }
1869
1870 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1871 }
1872 #endif
1873
1874 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1875 assert(thread == Thread::current() ||
1876 thread->is_handshake_safe_for(Thread::current()) ||
1877 SafepointSynchronize::is_at_safepoint(),
1878 "can only deoptimize other thread at a safepoint/handshake");
1879 // Compute frame and register map based on thread and sp.
1880 RegisterMap reg_map(thread,
1881 RegisterMap::UpdateMap::skip,
1882 RegisterMap::ProcessFrames::include,
1883 RegisterMap::WalkContinuation::skip);
1884 frame fr = thread->last_frame();
1885 while (fr.id() != id) {
1886 fr = fr.sender(®_map);
1887 }
1888 deoptimize(thread, fr, reason);
1889 }
1890
1891
1892 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1893 Thread* current = Thread::current();
1894 if (thread == current || thread->is_handshake_safe_for(current)) {
1895 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1896 } else {
1897 VM_DeoptimizeFrame deopt(thread, id, reason);
1898 VMThread::execute(&deopt);
1899 }
1900 }
1901
1902 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1903 deoptimize_frame(thread, id, Reason_constraint);
1904 }
1905
1906 // JVMTI PopFrame support
1907 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1908 {
1909 assert(thread == JavaThread::current(), "pre-condition");
1910 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1911 }
1912 JRT_END
1913
1914 MethodData*
1915 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1916 bool create_if_missing) {
1917 JavaThread* THREAD = thread; // For exception macros.
1918 MethodData* mdo = m()->method_data();
1919 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1920 // Build an MDO. Ignore errors like OutOfMemory;
1921 // that simply means we won't have an MDO to update.
1922 Method::build_profiling_method_data(m, THREAD);
1923 if (HAS_PENDING_EXCEPTION) {
1924 // Only metaspace OOM is expected. No Java code executed.
1925 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1926 CLEAR_PENDING_EXCEPTION;
1927 }
1928 mdo = m()->method_data();
1929 }
1930 return mdo;
1931 }
1932
1933 #if COMPILER2_OR_JVMCI
1934 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1935 // In case of an unresolved klass entry, load the class.
1936 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1937 // and probably nowhere else.
1938 // Even that case would benefit from simply re-interpreting the
1939 // bytecode, without paying special attention to the class index.
1940 // So this whole "class index" feature should probably be removed.
1941
1942 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1943 Klass* tk = constant_pool->klass_at(index, THREAD);
1944 if (HAS_PENDING_EXCEPTION) {
1945 // Exception happened during classloading. We ignore the exception here, since it
1946 // is going to be rethrown since the current activation is going to be deoptimized and
1947 // the interpreter will re-execute the bytecode.
1948 // Do not clear probable Async Exceptions.
1949 CLEAR_PENDING_NONASYNC_EXCEPTION;
1950 // Class loading called java code which may have caused a stack
1951 // overflow. If the exception was thrown right before the return
1952 // to the runtime the stack is no longer guarded. Reguard the
1953 // stack otherwise if we return to the uncommon trap blob and the
1954 // stack bang causes a stack overflow we crash.
1955 JavaThread* jt = THREAD;
1956 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1957 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1958 }
1959 return;
1960 }
1961
1962 assert(!constant_pool->tag_at(index).is_symbol(),
1963 "no symbolic names here, please");
1964 }
1965
1966 #if INCLUDE_JFR
1967
1968 class DeoptReasonSerializer : public JfrSerializer {
1969 public:
1970 void serialize(JfrCheckpointWriter& writer) {
1971 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1972 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1973 writer.write_key((u8)i);
1974 writer.write(Deoptimization::trap_reason_name(i));
1975 }
1976 }
1977 };
1978
1979 class DeoptActionSerializer : public JfrSerializer {
1980 public:
1981 void serialize(JfrCheckpointWriter& writer) {
1982 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1983 writer.write_count(nof_actions);
1984 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1985 writer.write_key(i);
1986 writer.write(Deoptimization::trap_action_name((int)i));
1987 }
1988 }
1989 };
1990
1991 static void register_serializers() {
1992 static int critical_section = 0;
1993 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1994 return;
1995 }
1996 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1997 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1998 }
1999
2000 static void post_deoptimization_event(nmethod* nm,
2001 const Method* method,
2002 int trap_bci,
2003 int instruction,
2004 Deoptimization::DeoptReason reason,
2005 Deoptimization::DeoptAction action) {
2006 assert(nm != nullptr, "invariant");
2007 assert(method != nullptr, "invariant");
2008 if (EventDeoptimization::is_enabled()) {
2009 static bool serializers_registered = false;
2010 if (!serializers_registered) {
2011 register_serializers();
2012 serializers_registered = true;
2013 }
2014 EventDeoptimization event;
2015 event.set_compileId(nm->compile_id());
2016 event.set_compiler(nm->compiler_type());
2017 event.set_method(method);
2018 event.set_lineNumber(method->line_number_from_bci(trap_bci));
2019 event.set_bci(trap_bci);
2020 event.set_instruction(instruction);
2021 event.set_reason(reason);
2022 event.set_action(action);
2023 event.commit();
2024 }
2025 }
2026
2027 #endif // INCLUDE_JFR
2028
2029 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2030 const char* reason_name, const char* reason_action) {
2031 LogTarget(Debug, deoptimization) lt;
2032 if (lt.is_enabled()) {
2033 LogStream ls(lt);
2034 bool is_osr = nm->is_osr_method();
2035 ls.print("cid=%4d %s%s level=%d",
2036 nm->compile_id(), (is_osr ? "osr" : " "), (nm->preloaded() ? "preload" : ""), nm->comp_level());
2037 ls.print(" %s", tm->name_and_sig_as_C_string());
2038 ls.print(" trap_bci=%d ", trap_bci);
2039 if (is_osr) {
2040 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2041 }
2042 ls.print("%s ", reason_name);
2043 ls.print("%s ", reason_action);
2044 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2045 pc, fr.pc() - nm->code_begin());
2046 }
2047 }
2048
2049 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2050 HandleMark hm(current);
2051
2052 // uncommon_trap() is called at the beginning of the uncommon trap
2053 // handler. Note this fact before we start generating temporary frames
2054 // that can confuse an asynchronous stack walker. This counter is
2055 // decremented at the end of unpack_frames().
2056
2057 current->inc_in_deopt_handler();
2058
2059 #if INCLUDE_JVMCI
2060 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2061 RegisterMap reg_map(current,
2062 RegisterMap::UpdateMap::include,
2063 RegisterMap::ProcessFrames::include,
2064 RegisterMap::WalkContinuation::skip);
2065 #else
2066 RegisterMap reg_map(current,
2067 RegisterMap::UpdateMap::skip,
2068 RegisterMap::ProcessFrames::include,
2069 RegisterMap::WalkContinuation::skip);
2070 #endif
2071 frame stub_frame = current->last_frame();
2072 frame fr = stub_frame.sender(®_map);
2073
2074 // Log a message
2075 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2076 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2077
2078 {
2079 ResourceMark rm;
2080
2081 DeoptReason reason = trap_request_reason(trap_request);
2082 DeoptAction action = trap_request_action(trap_request);
2083 #if INCLUDE_JVMCI
2084 int debug_id = trap_request_debug_id(trap_request);
2085 #endif
2086 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2087
2088 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2089 compiledVFrame* cvf = compiledVFrame::cast(vf);
2090
2091 nmethod* nm = cvf->code();
2092
2093 ScopeDesc* trap_scope = cvf->scope();
2094
2095 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2096
2097 if (is_receiver_constraint_failure) {
2098 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2099 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2100 JVMCI_ONLY(COMMA debug_id));
2101 }
2102
2103 methodHandle trap_method(current, trap_scope->method());
2104 int trap_bci = trap_scope->bci();
2105 #if INCLUDE_JVMCI
2106 jlong speculation = current->pending_failed_speculation();
2107 if (nm->is_compiled_by_jvmci()) {
2108 nm->update_speculation(current);
2109 } else {
2110 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2111 }
2112
2113 if (trap_bci == SynchronizationEntryBCI) {
2114 trap_bci = 0;
2115 current->set_pending_monitorenter(true);
2116 }
2117
2118 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2119 current->set_pending_transfer_to_interpreter(true);
2120 }
2121 #endif
2122
2123 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2124 // Record this event in the histogram.
2125 gather_statistics(nm, reason, action, trap_bc);
2126
2127 // Ensure that we can record deopt. history:
2128 bool create_if_missing = ProfileTraps;
2129
2130 methodHandle profiled_method;
2131 #if INCLUDE_JVMCI
2132 if (nm->is_compiled_by_jvmci()) {
2133 profiled_method = methodHandle(current, nm->method());
2134 } else {
2135 profiled_method = trap_method;
2136 }
2137 #else
2138 profiled_method = trap_method;
2139 #endif
2140
2141 MethodData* trap_mdo =
2142 get_method_data(current, profiled_method, create_if_missing);
2143
2144 { // Log Deoptimization event for JFR, UL and event system
2145 Method* tm = trap_method();
2146 const char* reason_name = trap_reason_name(reason);
2147 const char* reason_action = trap_action_name(action);
2148 intptr_t pc = p2i(fr.pc());
2149
2150 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2151 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2152 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2153 reason_name, reason_action, pc,
2154 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2155 }
2156
2157 // Print a bunch of diagnostics, if requested.
2158 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2159 ResourceMark rm;
2160
2161 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2162 // We must do this already now, since we cannot acquire this lock while
2163 // holding the tty lock (lock ordering by rank).
2164 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2165
2166 ttyLocker ttyl;
2167
2168 char buf[100];
2169 if (xtty != nullptr) {
2170 xtty->begin_head("uncommon_trap thread='%zu' %s",
2171 os::current_thread_id(),
2172 format_trap_request(buf, sizeof(buf), trap_request));
2173 #if INCLUDE_JVMCI
2174 if (speculation != 0) {
2175 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2176 }
2177 #endif
2178 nm->log_identity(xtty);
2179 }
2180 Symbol* class_name = nullptr;
2181 bool unresolved = false;
2182 if (unloaded_class_index >= 0) {
2183 constantPoolHandle constants (current, trap_method->constants());
2184 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2185 class_name = constants->klass_name_at(unloaded_class_index);
2186 unresolved = true;
2187 if (xtty != nullptr)
2188 xtty->print(" unresolved='1'");
2189 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2190 class_name = constants->symbol_at(unloaded_class_index);
2191 }
2192 if (xtty != nullptr)
2193 xtty->name(class_name);
2194 }
2195 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2196 // Dump the relevant MDO state.
2197 // This is the deopt count for the current reason, any previous
2198 // reasons or recompiles seen at this point.
2199 int dcnt = trap_mdo->trap_count(reason);
2200 if (dcnt != 0)
2201 xtty->print(" count='%d'", dcnt);
2202
2203 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2204 // lock extra_data_lock before the tty lock.
2205 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2206 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2207 if (dos != 0) {
2208 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2209 if (trap_state_is_recompiled(dos)) {
2210 int recnt2 = trap_mdo->overflow_recompile_count();
2211 if (recnt2 != 0)
2212 xtty->print(" recompiles2='%d'", recnt2);
2213 }
2214 }
2215 }
2216 if (xtty != nullptr) {
2217 xtty->stamp();
2218 xtty->end_head();
2219 }
2220 if (TraceDeoptimization) { // make noise on the tty
2221 stringStream st;
2222 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2223 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2224 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2225 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2226 #if INCLUDE_JVMCI
2227 if (nm->is_compiled_by_jvmci()) {
2228 const char* installed_code_name = nm->jvmci_name();
2229 if (installed_code_name != nullptr) {
2230 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2231 }
2232 }
2233 #endif
2234 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2235 p2i(fr.pc()),
2236 os::current_thread_id(),
2237 trap_reason_name(reason),
2238 trap_action_name(action),
2239 unloaded_class_index
2240 #if INCLUDE_JVMCI
2241 , debug_id
2242 #endif
2243 );
2244 if (class_name != nullptr) {
2245 st.print(unresolved ? " unresolved class: " : " symbol: ");
2246 class_name->print_symbol_on(&st);
2247 }
2248 st.cr();
2249 tty->print_raw(st.freeze());
2250 }
2251 if (xtty != nullptr) {
2252 // Log the precise location of the trap.
2253 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2254 xtty->begin_elem("jvms bci='%d'", sd->bci());
2255 xtty->method(sd->method());
2256 xtty->end_elem();
2257 if (sd->is_top()) break;
2258 }
2259 xtty->tail("uncommon_trap");
2260 }
2261 }
2262 // (End diagnostic printout.)
2263
2264 if (is_receiver_constraint_failure) {
2265 fatal("missing receiver type check");
2266 }
2267
2268 // Load class if necessary
2269 if (unloaded_class_index >= 0) {
2270 constantPoolHandle constants(current, trap_method->constants());
2271 load_class_by_index(constants, unloaded_class_index, THREAD);
2272 }
2273
2274 // Flush the nmethod if necessary and desirable.
2275 //
2276 // We need to avoid situations where we are re-flushing the nmethod
2277 // because of a hot deoptimization site. Repeated flushes at the same
2278 // point need to be detected by the compiler and avoided. If the compiler
2279 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2280 // module must take measures to avoid an infinite cycle of recompilation
2281 // and deoptimization. There are several such measures:
2282 //
2283 // 1. If a recompilation is ordered a second time at some site X
2284 // and for the same reason R, the action is adjusted to 'reinterpret',
2285 // to give the interpreter time to exercise the method more thoroughly.
2286 // If this happens, the method's overflow_recompile_count is incremented.
2287 //
2288 // 2. If the compiler fails to reduce the deoptimization rate, then
2289 // the method's overflow_recompile_count will begin to exceed the set
2290 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2291 // is adjusted to 'make_not_compilable', and the method is abandoned
2292 // to the interpreter. This is a performance hit for hot methods,
2293 // but is better than a disastrous infinite cycle of recompilations.
2294 // (Actually, only the method containing the site X is abandoned.)
2295 //
2296 // 3. In parallel with the previous measures, if the total number of
2297 // recompilations of a method exceeds the much larger set limit
2298 // PerMethodRecompilationCutoff, the method is abandoned.
2299 // This should only happen if the method is very large and has
2300 // many "lukewarm" deoptimizations. The code which enforces this
2301 // limit is elsewhere (class nmethod, class Method).
2302 //
2303 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2304 // to recompile at each bytecode independently of the per-BCI cutoff.
2305 //
2306 // The decision to update code is up to the compiler, and is encoded
2307 // in the Action_xxx code. If the compiler requests Action_none
2308 // no trap state is changed, no compiled code is changed, and the
2309 // computation suffers along in the interpreter.
2310 //
2311 // The other action codes specify various tactics for decompilation
2312 // and recompilation. Action_maybe_recompile is the loosest, and
2313 // allows the compiled code to stay around until enough traps are seen,
2314 // and until the compiler gets around to recompiling the trapping method.
2315 //
2316 // The other actions cause immediate removal of the present code.
2317
2318 // Traps caused by injected profile shouldn't pollute trap counts.
2319 bool injected_profile_trap = trap_method->has_injected_profile() &&
2320 (reason == Reason_intrinsic || reason == Reason_unreached);
2321
2322 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2323 bool make_not_entrant = false;
2324 bool make_not_compilable = false;
2325 bool reprofile = false;
2326 switch (action) {
2327 case Action_none:
2328 // Keep the old code.
2329 update_trap_state = false;
2330 break;
2331 case Action_maybe_recompile:
2332 // Do not need to invalidate the present code, but we can
2333 // initiate another
2334 // Start compiler without (necessarily) invalidating the nmethod.
2335 // The system will tolerate the old code, but new code should be
2336 // generated when possible.
2337 break;
2338 case Action_reinterpret:
2339 // Go back into the interpreter for a while, and then consider
2340 // recompiling form scratch.
2341 make_not_entrant = true;
2342 // Reset invocation counter for outer most method.
2343 // This will allow the interpreter to exercise the bytecodes
2344 // for a while before recompiling.
2345 // By contrast, Action_make_not_entrant is immediate.
2346 //
2347 // Note that the compiler will track null_check, null_assert,
2348 // range_check, and class_check events and log them as if they
2349 // had been traps taken from compiled code. This will update
2350 // the MDO trap history so that the next compilation will
2351 // properly detect hot trap sites.
2352 reprofile = true;
2353 break;
2354 case Action_make_not_entrant:
2355 // Request immediate recompilation, and get rid of the old code.
2356 // Make them not entrant, so next time they are called they get
2357 // recompiled. Unloaded classes are loaded now so recompile before next
2358 // time they are called. Same for uninitialized. The interpreter will
2359 // link the missing class, if any.
2360 make_not_entrant = true;
2361 break;
2362 case Action_make_not_compilable:
2363 // Give up on compiling this method at all.
2364 make_not_entrant = true;
2365 make_not_compilable = true;
2366 break;
2367 default:
2368 ShouldNotReachHere();
2369 }
2370
2371 // Setting +ProfileTraps fixes the following, on all platforms:
2372 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2373 // recompile relies on a MethodData* to record heroic opt failures.
2374
2375 // Whether the interpreter is producing MDO data or not, we also need
2376 // to use the MDO to detect hot deoptimization points and control
2377 // aggressive optimization.
2378 bool inc_recompile_count = false;
2379
2380 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2381 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2382 (trap_mdo != nullptr),
2383 Mutex::_no_safepoint_check_flag);
2384 ProfileData* pdata = nullptr;
2385 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2386 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2387 uint this_trap_count = 0;
2388 bool maybe_prior_trap = false;
2389 bool maybe_prior_recompile = false;
2390
2391 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2392 #if INCLUDE_JVMCI
2393 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2394 #endif
2395 nm->method(),
2396 //outputs:
2397 this_trap_count,
2398 maybe_prior_trap,
2399 maybe_prior_recompile);
2400 // Because the interpreter also counts null, div0, range, and class
2401 // checks, these traps from compiled code are double-counted.
2402 // This is harmless; it just means that the PerXTrapLimit values
2403 // are in effect a little smaller than they look.
2404
2405 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2406 if (per_bc_reason != Reason_none) {
2407 // Now take action based on the partially known per-BCI history.
2408 if (maybe_prior_trap
2409 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2410 // If there are too many traps at this BCI, force a recompile.
2411 // This will allow the compiler to see the limit overflow, and
2412 // take corrective action, if possible. The compiler generally
2413 // does not use the exact PerBytecodeTrapLimit value, but instead
2414 // changes its tactics if it sees any traps at all. This provides
2415 // a little hysteresis, delaying a recompile until a trap happens
2416 // several times.
2417 //
2418 // Actually, since there is only one bit of counter per BCI,
2419 // the possible per-BCI counts are {0,1,(per-method count)}.
2420 // This produces accurate results if in fact there is only
2421 // one hot trap site, but begins to get fuzzy if there are
2422 // many sites. For example, if there are ten sites each
2423 // trapping two or more times, they each get the blame for
2424 // all of their traps.
2425 make_not_entrant = true;
2426 }
2427
2428 // Detect repeated recompilation at the same BCI, and enforce a limit.
2429 if (make_not_entrant && maybe_prior_recompile) {
2430 // More than one recompile at this point.
2431 inc_recompile_count = maybe_prior_trap;
2432 }
2433 } else {
2434 // For reasons which are not recorded per-bytecode, we simply
2435 // force recompiles unconditionally.
2436 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2437 make_not_entrant = true;
2438 }
2439
2440 // Go back to the compiler if there are too many traps in this method.
2441 if (this_trap_count >= per_method_trap_limit(reason)) {
2442 // If there are too many traps in this method, force a recompile.
2443 // This will allow the compiler to see the limit overflow, and
2444 // take corrective action, if possible.
2445 // (This condition is an unlikely backstop only, because the
2446 // PerBytecodeTrapLimit is more likely to take effect first,
2447 // if it is applicable.)
2448 make_not_entrant = true;
2449 }
2450
2451 // Here's more hysteresis: If there has been a recompile at
2452 // this trap point already, run the method in the interpreter
2453 // for a while to exercise it more thoroughly.
2454 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2455 reprofile = true;
2456 }
2457 }
2458
2459 // Take requested actions on the method:
2460
2461 // Recompile
2462 if (make_not_entrant) {
2463 if (!nm->make_not_entrant("uncommon trap")) {
2464 return; // the call did not change nmethod's state
2465 }
2466
2467 if (pdata != nullptr) {
2468 // Record the recompilation event, if any.
2469 int tstate0 = pdata->trap_state();
2470 int tstate1 = trap_state_set_recompiled(tstate0, true);
2471 if (tstate1 != tstate0)
2472 pdata->set_trap_state(tstate1);
2473 }
2474
2475 // For code aging we count traps separately here, using make_not_entrant()
2476 // as a guard against simultaneous deopts in multiple threads.
2477 if (reason == Reason_tenured && trap_mdo != nullptr) {
2478 trap_mdo->inc_tenure_traps();
2479 }
2480 }
2481
2482 if (inc_recompile_count) {
2483 trap_mdo->inc_overflow_recompile_count();
2484 if ((uint)trap_mdo->overflow_recompile_count() >
2485 (uint)PerBytecodeRecompilationCutoff) {
2486 // Give up on the method containing the bad BCI.
2487 if (trap_method() == nm->method()) {
2488 make_not_compilable = true;
2489 } else {
2490 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2491 // But give grace to the enclosing nm->method().
2492 }
2493 }
2494 }
2495
2496 // Reprofile
2497 if (reprofile) {
2498 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2499 }
2500
2501 // Give up compiling
2502 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2503 assert(make_not_entrant, "consistent");
2504 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2505 }
2506
2507 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2508 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2509 if (exception_handler_data != nullptr) {
2510 // uncommon trap at the start of an exception handler.
2511 // C2 generates these for un-entered exception handlers.
2512 // mark the handler as entered to avoid generating
2513 // another uncommon trap the next time the handler is compiled
2514 exception_handler_data->set_exception_handler_entered();
2515 }
2516 }
2517
2518 } // Free marked resources
2519
2520 }
2521 JRT_END
2522
2523 ProfileData*
2524 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2525 int trap_bci,
2526 Deoptimization::DeoptReason reason,
2527 bool update_total_trap_count,
2528 #if INCLUDE_JVMCI
2529 bool is_osr,
2530 #endif
2531 Method* compiled_method,
2532 //outputs:
2533 uint& ret_this_trap_count,
2534 bool& ret_maybe_prior_trap,
2535 bool& ret_maybe_prior_recompile) {
2536 trap_mdo->check_extra_data_locked();
2537
2538 bool maybe_prior_trap = false;
2539 bool maybe_prior_recompile = false;
2540 uint this_trap_count = 0;
2541 if (update_total_trap_count) {
2542 uint idx = reason;
2543 #if INCLUDE_JVMCI
2544 if (is_osr) {
2545 // Upper half of history array used for traps in OSR compilations
2546 idx += Reason_TRAP_HISTORY_LENGTH;
2547 }
2548 #endif
2549 uint prior_trap_count = trap_mdo->trap_count(idx);
2550 this_trap_count = trap_mdo->inc_trap_count(idx);
2551
2552 // If the runtime cannot find a place to store trap history,
2553 // it is estimated based on the general condition of the method.
2554 // If the method has ever been recompiled, or has ever incurred
2555 // a trap with the present reason , then this BCI is assumed
2556 // (pessimistically) to be the culprit.
2557 maybe_prior_trap = (prior_trap_count != 0);
2558 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2559 }
2560 ProfileData* pdata = nullptr;
2561
2562
2563 // For reasons which are recorded per bytecode, we check per-BCI data.
2564 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2565 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2566 if (per_bc_reason != Reason_none) {
2567 // Find the profile data for this BCI. If there isn't one,
2568 // try to allocate one from the MDO's set of spares.
2569 // This will let us detect a repeated trap at this point.
2570 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2571
2572 if (pdata != nullptr) {
2573 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2574 if (LogCompilation && xtty != nullptr) {
2575 ttyLocker ttyl;
2576 // no more room for speculative traps in this MDO
2577 xtty->elem("speculative_traps_oom");
2578 }
2579 }
2580 // Query the trap state of this profile datum.
2581 int tstate0 = pdata->trap_state();
2582 if (!trap_state_has_reason(tstate0, per_bc_reason))
2583 maybe_prior_trap = false;
2584 if (!trap_state_is_recompiled(tstate0))
2585 maybe_prior_recompile = false;
2586
2587 // Update the trap state of this profile datum.
2588 int tstate1 = tstate0;
2589 // Record the reason.
2590 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2591 // Store the updated state on the MDO, for next time.
2592 if (tstate1 != tstate0)
2593 pdata->set_trap_state(tstate1);
2594 } else {
2595 if (LogCompilation && xtty != nullptr) {
2596 ttyLocker ttyl;
2597 // Missing MDP? Leave a small complaint in the log.
2598 xtty->elem("missing_mdp bci='%d'", trap_bci);
2599 }
2600 }
2601 }
2602
2603 // Return results:
2604 ret_this_trap_count = this_trap_count;
2605 ret_maybe_prior_trap = maybe_prior_trap;
2606 ret_maybe_prior_recompile = maybe_prior_recompile;
2607 return pdata;
2608 }
2609
2610 void
2611 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2612 ResourceMark rm;
2613 // Ignored outputs:
2614 uint ignore_this_trap_count;
2615 bool ignore_maybe_prior_trap;
2616 bool ignore_maybe_prior_recompile;
2617 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2618 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2619 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2620
2621 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2622 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2623
2624 query_update_method_data(trap_mdo, trap_bci,
2625 (DeoptReason)reason,
2626 update_total_counts,
2627 #if INCLUDE_JVMCI
2628 false,
2629 #endif
2630 nullptr,
2631 ignore_this_trap_count,
2632 ignore_maybe_prior_trap,
2633 ignore_maybe_prior_recompile);
2634 }
2635
2636 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode))
2637 // Enable WXWrite: current function is called from methods compiled by C2 directly
2638 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2639
2640 // Still in Java no safepoints
2641 {
2642 // This enters VM and may safepoint
2643 uncommon_trap_inner(current, trap_request);
2644 }
2645 HandleMark hm(current);
2646 return fetch_unroll_info_helper(current, exec_mode);
2647 PROF_END
2648
2649 // Local derived constants.
2650 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2651 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2652 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2653
2654 //---------------------------trap_state_reason---------------------------------
2655 Deoptimization::DeoptReason
2656 Deoptimization::trap_state_reason(int trap_state) {
2657 // This assert provides the link between the width of DataLayout::trap_bits
2658 // and the encoding of "recorded" reasons. It ensures there are enough
2659 // bits to store all needed reasons in the per-BCI MDO profile.
2660 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2661 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2662 trap_state -= recompile_bit;
2663 if (trap_state == DS_REASON_MASK) {
2664 return Reason_many;
2665 } else {
2666 assert((int)Reason_none == 0, "state=0 => Reason_none");
2667 return (DeoptReason)trap_state;
2668 }
2669 }
2670 //-------------------------trap_state_has_reason-------------------------------
2671 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2672 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2673 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2674 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2675 trap_state -= recompile_bit;
2676 if (trap_state == DS_REASON_MASK) {
2677 return -1; // true, unspecifically (bottom of state lattice)
2678 } else if (trap_state == reason) {
2679 return 1; // true, definitely
2680 } else if (trap_state == 0) {
2681 return 0; // false, definitely (top of state lattice)
2682 } else {
2683 return 0; // false, definitely
2684 }
2685 }
2686 //-------------------------trap_state_add_reason-------------------------------
2687 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2688 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2689 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2690 trap_state -= recompile_bit;
2691 if (trap_state == DS_REASON_MASK) {
2692 return trap_state + recompile_bit; // already at state lattice bottom
2693 } else if (trap_state == reason) {
2694 return trap_state + recompile_bit; // the condition is already true
2695 } else if (trap_state == 0) {
2696 return reason + recompile_bit; // no condition has yet been true
2697 } else {
2698 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2699 }
2700 }
2701 //-----------------------trap_state_is_recompiled------------------------------
2702 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2703 return (trap_state & DS_RECOMPILE_BIT) != 0;
2704 }
2705 //-----------------------trap_state_set_recompiled-----------------------------
2706 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2707 if (z) return trap_state | DS_RECOMPILE_BIT;
2708 else return trap_state & ~DS_RECOMPILE_BIT;
2709 }
2710 //---------------------------format_trap_state---------------------------------
2711 // This is used for debugging and diagnostics, including LogFile output.
2712 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2713 int trap_state) {
2714 assert(buflen > 0, "sanity");
2715 DeoptReason reason = trap_state_reason(trap_state);
2716 bool recomp_flag = trap_state_is_recompiled(trap_state);
2717 // Re-encode the state from its decoded components.
2718 int decoded_state = 0;
2719 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2720 decoded_state = trap_state_add_reason(decoded_state, reason);
2721 if (recomp_flag)
2722 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2723 // If the state re-encodes properly, format it symbolically.
2724 // Because this routine is used for debugging and diagnostics,
2725 // be robust even if the state is a strange value.
2726 size_t len;
2727 if (decoded_state != trap_state) {
2728 // Random buggy state that doesn't decode??
2729 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2730 } else {
2731 len = jio_snprintf(buf, buflen, "%s%s",
2732 trap_reason_name(reason),
2733 recomp_flag ? " recompiled" : "");
2734 }
2735 return buf;
2736 }
2737
2738
2739 //--------------------------------statics--------------------------------------
2740 const char* Deoptimization::_trap_reason_name[] = {
2741 // Note: Keep this in sync. with enum DeoptReason.
2742 "none",
2743 "null_check",
2744 "null_assert" JVMCI_ONLY("_or_unreached0"),
2745 "range_check",
2746 "class_check",
2747 "array_check",
2748 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2749 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2750 "profile_predicate",
2751 "auto_vectorization_check",
2752 "unloaded",
2753 "uninitialized",
2754 "initialized",
2755 "unreached",
2756 "unhandled",
2757 "constraint",
2758 "div0_check",
2759 "age",
2760 "predicate",
2761 "loop_limit_check",
2762 "speculate_class_check",
2763 "speculate_null_check",
2764 "speculate_null_assert",
2765 "unstable_if",
2766 "unstable_fused_if",
2767 "receiver_constraint",
2768 #if INCLUDE_JVMCI
2769 "aliasing",
2770 "transfer_to_interpreter",
2771 "not_compiled_exception_handler",
2772 "unresolved",
2773 "jsr_mismatch",
2774 #endif
2775 "tenured"
2776 };
2777 const char* Deoptimization::_trap_action_name[] = {
2778 // Note: Keep this in sync. with enum DeoptAction.
2779 "none",
2780 "maybe_recompile",
2781 "reinterpret",
2782 "make_not_entrant",
2783 "make_not_compilable"
2784 };
2785
2786 const char* Deoptimization::trap_reason_name(int reason) {
2787 // Check that every reason has a name
2788 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2789
2790 if (reason == Reason_many) return "many";
2791 if ((uint)reason < Reason_LIMIT)
2792 return _trap_reason_name[reason];
2793 static char buf[20];
2794 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2795 return buf;
2796 }
2797 const char* Deoptimization::trap_action_name(int action) {
2798 // Check that every action has a name
2799 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2800
2801 if ((uint)action < Action_LIMIT)
2802 return _trap_action_name[action];
2803 static char buf[20];
2804 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2805 return buf;
2806 }
2807
2808 // This is used for debugging and diagnostics, including LogFile output.
2809 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2810 int trap_request) {
2811 jint unloaded_class_index = trap_request_index(trap_request);
2812 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2813 const char* action = trap_action_name(trap_request_action(trap_request));
2814 #if INCLUDE_JVMCI
2815 int debug_id = trap_request_debug_id(trap_request);
2816 #endif
2817 size_t len;
2818 if (unloaded_class_index < 0) {
2819 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2820 reason, action
2821 #if INCLUDE_JVMCI
2822 ,debug_id
2823 #endif
2824 );
2825 } else {
2826 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2827 reason, action, unloaded_class_index
2828 #if INCLUDE_JVMCI
2829 ,debug_id
2830 #endif
2831 );
2832 }
2833 return buf;
2834 }
2835
2836 juint Deoptimization::_deoptimization_hist
2837 [1 + 4 + 5] // total + online + archived
2838 [Deoptimization::Reason_LIMIT]
2839 [1 + Deoptimization::Action_LIMIT]
2840 [Deoptimization::BC_CASE_LIMIT]
2841 = {0};
2842
2843 enum {
2844 LSB_BITS = 8,
2845 LSB_MASK = right_n_bits(LSB_BITS)
2846 };
2847
2848 static void update(juint* cases, Bytecodes::Code bc) {
2849 juint* bc_counter_addr = nullptr;
2850 juint bc_counter = 0;
2851 // Look for an unused counter, or an exact match to this BC.
2852 if (bc != Bytecodes::_illegal) {
2853 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) {
2854 juint* counter_addr = &cases[bc_case];
2855 juint counter = *counter_addr;
2856 if ((counter == 0 && bc_counter_addr == nullptr)
2857 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2858 // this counter is either free or is already devoted to this BC
2859 bc_counter_addr = counter_addr;
2860 bc_counter = counter | bc;
2861 }
2862 }
2863 }
2864 if (bc_counter_addr == nullptr) {
2865 // Overflow, or no given bytecode.
2866 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1];
2867 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2868 }
2869 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2870 }
2871
2872
2873 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
2874 Bytecodes::Code bc) {
2875 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2876 assert(action >= 0 && action < Action_LIMIT, "oob");
2877 _deoptimization_hist[0][Reason_none][0][0] += 1; // total
2878 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total
2879
2880 update(_deoptimization_hist[0][reason][1+action], bc);
2881
2882 uint lvl = nm->comp_level() + (nm->is_aot() ? 4 : 0) + (nm->preloaded() ? 1 : 0);
2883 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total
2884 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total
2885 update(_deoptimization_hist[lvl][reason][1+action], bc);
2886 }
2887
2888 jint Deoptimization::total_deoptimization_count() {
2889 return _deoptimization_hist[0][Reason_none][0][0];
2890 }
2891
2892 // Get the deopt count for a specific reason and a specific action. If either
2893 // one of 'reason' or 'action' is null, the method returns the sum of all
2894 // deoptimizations with the specific 'action' or 'reason' respectively.
2895 // If both arguments are null, the method returns the total deopt count.
2896 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2897 if (reason_str == nullptr && action_str == nullptr) {
2898 return total_deoptimization_count();
2899 }
2900 juint counter = 0;
2901 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2902 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2903 for (int action = 0; action < Action_LIMIT; action++) {
2904 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2905 juint* cases = _deoptimization_hist[0][reason][1+action];
2906 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2907 counter += cases[bc_case] >> LSB_BITS;
2908 }
2909 }
2910 }
2911 }
2912 }
2913 return counter;
2914 }
2915
2916 void Deoptimization::print_statistics() {
2917 ttyLocker ttyl;
2918 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2919 tty->print_cr("Deoptimization traps recorded:");
2920 print_statistics_on(tty);
2921 if (xtty != nullptr) xtty->tail("statistics");
2922 }
2923
2924 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) {
2925 juint total = _deoptimization_hist[lvl][Reason_none][0][0];
2926 juint account = total;
2927 #define PRINT_STAT_LINE(name, r) \
2928 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name);
2929 if (total > 0) {
2930 st->print(" %s: ", title);
2931 PRINT_STAT_LINE("total", total);
2932 // For each non-zero entry in the histogram, print the reason,
2933 // the action, and (if specifically known) the type of bytecode.
2934 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2935 for (int action = 0; action < Action_LIMIT; action++) {
2936 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action];
2937 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2938 juint counter = cases[bc_case];
2939 if (counter != 0) {
2940 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2941 const char* bc_name = "other";
2942 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) {
2943 // overwritten
2944 } else if (Bytecodes::is_defined(bc)) {
2945 bc_name = Bytecodes::name(bc);
2946 }
2947 juint r = counter >> LSB_BITS;
2948 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)",
2949 trap_reason_name(reason), trap_action_name(action), bc_name,
2950 r, (r * 100.0) / total);
2951 account -= r;
2952 }
2953 }
2954 }
2955 }
2956 if (account != 0) {
2957 PRINT_STAT_LINE("unaccounted", account);
2958 }
2959 #undef PRINT_STAT_LINE
2960 }
2961 }
2962
2963 void Deoptimization::print_statistics_on(outputStream* st) {
2964 // print_statistics_on("Total", 0, st);
2965 print_statistics_on("Tier1", 1, st);
2966 print_statistics_on("Tier2", 2, st);
2967 print_statistics_on("Tier3", 3, st);
2968 print_statistics_on("Tier4", 4, st);
2969
2970 print_statistics_on("AOT Code Tier1", 5, st);
2971 print_statistics_on("AOT Code Tier2", 6, st);
2972 print_statistics_on("AOT Code Tier4", 8, st);
2973 print_statistics_on("AOT Code Tier5 (preloaded)", 9, st);
2974 }
2975
2976 #define DO_COUNTERS(macro) \
2977 macro(Deoptimization, fetch_unroll_info) \
2978 macro(Deoptimization, unpack_frames) \
2979 macro(Deoptimization, uncommon_trap_inner) \
2980 macro(Deoptimization, uncommon_trap)
2981
2982 #define INIT_COUNTER(sub, name) \
2983 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \
2984 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count");
2985
2986 void Deoptimization::init_counters() {
2987 if (ProfileRuntimeCalls && UsePerfData) {
2988 EXCEPTION_MARK;
2989
2990 DO_COUNTERS(INIT_COUNTER)
2991
2992 if (HAS_PENDING_EXCEPTION) {
2993 vm_exit_during_initialization("jvm_perf_init failed unexpectedly");
2994 }
2995 }
2996 }
2997 #undef INIT_COUNTER
2998
2999 #define PRINT_COUNTER(sub, name) { \
3000 jlong count = _perf_##sub##_##name##_count->get_value(); \
3001 if (count > 0) { \
3002 st->print_cr(" %-50s = " JLONG_FORMAT_W(6) "us (elapsed) " JLONG_FORMAT_W(6) "us (thread) (" JLONG_FORMAT_W(5) " events)", #sub "::" #name, \
3003 _perf_##sub##_##name##_timer->elapsed_counter_value_us(), \
3004 _perf_##sub##_##name##_timer->thread_counter_value_us(), \
3005 count); \
3006 }}
3007
3008 void Deoptimization::print_counters_on(outputStream* st) {
3009 if (ProfileRuntimeCalls && UsePerfData) {
3010 DO_COUNTERS(PRINT_COUNTER)
3011 } else {
3012 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData"));
3013 }
3014 }
3015
3016 #undef PRINT_COUNTER
3017 #undef DO_COUNTERS
3018
3019 #else // COMPILER2_OR_JVMCI
3020
3021
3022 // Stubs for C1 only system.
3023 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3024 return false;
3025 }
3026
3027 const char* Deoptimization::trap_reason_name(int reason) {
3028 return "unknown";
3029 }
3030
3031 jint Deoptimization::total_deoptimization_count() {
3032 return 0;
3033 }
3034
3035 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3036 return 0;
3037 }
3038
3039 void Deoptimization::print_statistics() {
3040 // no output
3041 }
3042
3043 void
3044 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3045 // no update
3046 }
3047
3048 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3049 return 0;
3050 }
3051
3052 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
3053 Bytecodes::Code bc) {
3054 // no update
3055 }
3056
3057 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3058 int trap_state) {
3059 jio_snprintf(buf, buflen, "#%d", trap_state);
3060 return buf;
3061 }
3062
3063 void Deoptimization::init_counters() {
3064 // nothing to do
3065 }
3066
3067 void Deoptimization::print_counters_on(outputStream* st) {
3068 // no output
3069 }
3070
3071 void Deoptimization::print_statistics_on(outputStream* st) {
3072 // no output
3073 }
3074
3075 #endif // COMPILER2_OR_JVMCI