1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "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.inline.hpp"
39 #include "interpreter/bytecodeStream.hpp"
40 #include "interpreter/interpreter.hpp"
41 #include "interpreter/oopMapCache.hpp"
42 #include "jvm.h"
43 #include "logging/log.hpp"
44 #include "logging/logLevel.hpp"
45 #include "logging/logMessage.hpp"
46 #include "logging/logStream.hpp"
47 #include "memory/allocation.inline.hpp"
48 #include "memory/oopFactory.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "memory/universe.hpp"
51 #include "oops/arrayOop.inline.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/flatArrayKlass.hpp"
55 #include "oops/flatArrayOop.hpp"
56 #include "oops/inlineKlass.inline.hpp"
57 #include "oops/method.hpp"
58 #include "oops/objArrayKlass.hpp"
59 #include "oops/objArrayOop.inline.hpp"
60 #include "oops/oop.inline.hpp"
61 #include "oops/typeArrayOop.inline.hpp"
62 #include "oops/verifyOopClosure.hpp"
63 #include "prims/jvmtiDeferredUpdates.hpp"
64 #include "prims/jvmtiExport.hpp"
65 #include "prims/jvmtiThreadState.hpp"
66 #include "prims/methodHandles.hpp"
67 #include "prims/vectorSupport.hpp"
68 #include "runtime/arguments.hpp"
69 #include "runtime/atomicAccess.hpp"
70 #include "runtime/basicLock.inline.hpp"
71 #include "runtime/continuation.hpp"
72 #include "runtime/continuationEntry.inline.hpp"
73 #include "runtime/deoptimization.hpp"
74 #include "runtime/escapeBarrier.hpp"
75 #include "runtime/fieldDescriptor.inline.hpp"
76 #include "runtime/frame.inline.hpp"
77 #include "runtime/handles.inline.hpp"
78 #include "runtime/interfaceSupport.inline.hpp"
79 #include "runtime/javaThread.hpp"
80 #include "runtime/jniHandles.inline.hpp"
81 #include "runtime/keepStackGCProcessed.hpp"
82 #include "runtime/lockStack.inline.hpp"
83 #include "runtime/objectMonitor.inline.hpp"
84 #include "runtime/osThread.hpp"
85 #include "runtime/safepointVerifiers.hpp"
86 #include "runtime/sharedRuntime.hpp"
87 #include "runtime/signature.hpp"
88 #include "runtime/stackFrameStream.inline.hpp"
89 #include "runtime/stackValue.hpp"
90 #include "runtime/stackWatermarkSet.hpp"
91 #include "runtime/stubRoutines.hpp"
92 #include "runtime/synchronizer.hpp"
93 #include "runtime/threadSMR.hpp"
94 #include "runtime/threadWXSetters.inline.hpp"
95 #include "runtime/vframe.hpp"
96 #include "runtime/vframe_hp.hpp"
97 #include "runtime/vframeArray.hpp"
98 #include "runtime/vmOperations.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 MACOS_AARCH64_ONLY(os::thread_wx_enable_write());
141 nmethod::DeoptimizationStatus status =
142 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
143 AtomicAccess::store(&nm->_deoptimization_status, status);
144
145 // Make sure active is not committed
146 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
147 assert(nm->_deoptimization_generation == 0, "Is already marked");
148
149 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
150 _required_gen = DeoptimizationScope::_active_deopt_gen;
151 }
152
153 void DeoptimizationScope::dependent(nmethod* nm) {
154 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
155
156 // A method marked by someone else may have a _required_gen lower than what we marked with.
157 // Therefore only store it if it's higher than _required_gen.
158 if (_required_gen < nm->_deoptimization_generation) {
159 _required_gen = nm->_deoptimization_generation;
160 }
161 }
162
163 void DeoptimizationScope::deoptimize_marked() {
164 assert(!_deopted, "Already deopted");
165
166 // We are not alive yet.
167 if (!Universe::is_fully_initialized()) {
168 DEBUG_ONLY(_deopted = true;)
169 return;
170 }
171
172 // Safepoints are a special case, handled here.
173 if (SafepointSynchronize::is_at_safepoint()) {
174 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
175 DeoptimizationScope::_active_deopt_gen++;
176 Deoptimization::deoptimize_all_marked();
177 DEBUG_ONLY(_deopted = true;)
178 return;
179 }
180
181 uint64_t comitting = 0;
182 bool wait = false;
183 while (true) {
184 {
185 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
186
187 // First we check if we or someone else already deopted the gen we want.
188 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
189 DEBUG_ONLY(_deopted = true;)
190 return;
191 }
192 if (!_committing_in_progress) {
193 // The version we are about to commit.
194 comitting = DeoptimizationScope::_active_deopt_gen;
195 // Make sure new marks use a higher gen.
196 DeoptimizationScope::_active_deopt_gen++;
197 _committing_in_progress = true;
198 wait = false;
199 } else {
200 // Another thread is handshaking and committing a gen.
201 wait = true;
202 }
203 }
204 if (wait) {
205 // Wait and let the concurrent handshake be performed.
206 ThreadBlockInVM tbivm(JavaThread::current());
207 os::naked_yield();
208 } else {
209 // Performs the handshake.
210 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
211 DEBUG_ONLY(_deopted = true;)
212 {
213 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
214
215 // Make sure that committed doesn't go backwards.
216 // Should only happen if we did a deopt during a safepoint above.
217 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
218 DeoptimizationScope::_committed_deopt_gen = comitting;
219 }
220 _committing_in_progress = false;
221
222 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
223
224 return;
225 }
226 }
227 }
228 }
229
230 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
231 int caller_adjustment,
232 int caller_actual_parameters,
233 int number_of_frames,
234 intptr_t* frame_sizes,
235 address* frame_pcs,
236 BasicType return_type,
237 int exec_mode) {
238 _size_of_deoptimized_frame = size_of_deoptimized_frame;
239 _caller_adjustment = caller_adjustment;
240 _caller_actual_parameters = caller_actual_parameters;
241 _number_of_frames = number_of_frames;
242 _frame_sizes = frame_sizes;
243 _frame_pcs = frame_pcs;
244 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
245 _return_type = return_type;
246 _initial_info = 0;
247 // PD (x86 only)
248 _counter_temp = 0;
249 _unpack_kind = exec_mode;
250 _sender_sp_temp = 0;
251
252 _total_frame_sizes = size_of_frames();
253 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
254 }
255
256 Deoptimization::UnrollBlock::~UnrollBlock() {
257 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
258 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
259 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
260 }
261
262 int Deoptimization::UnrollBlock::size_of_frames() const {
263 // Account first for the adjustment of the initial frame
264 intptr_t result = _caller_adjustment;
265 for (int index = 0; index < number_of_frames(); index++) {
266 result += frame_sizes()[index];
267 }
268 return checked_cast<int>(result);
269 }
270
271 void Deoptimization::UnrollBlock::print() {
272 ResourceMark rm;
273 stringStream st;
274 st.print_cr("UnrollBlock");
275 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
276 st.print( " frame_sizes: ");
277 for (int index = 0; index < number_of_frames(); index++) {
278 st.print("%zd ", frame_sizes()[index]);
279 }
280 st.cr();
281 tty->print_raw(st.freeze());
282 }
283
284 // In order to make fetch_unroll_info work properly with escape
285 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
286 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
287 // which is called from the method fetch_unroll_info_helper below.
288 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
289 // fetch_unroll_info() is called at the beginning of the deoptimization
290 // handler. Note this fact before we start generating temporary frames
291 // that can confuse an asynchronous stack walker. This counter is
292 // decremented at the end of unpack_frames().
293 current->inc_in_deopt_handler();
294
295 if (exec_mode == Unpack_exception) {
296 // When we get here, a callee has thrown an exception into a deoptimized
297 // frame. That throw might have deferred stack watermark checking until
298 // after unwinding. So we deal with such deferred requests here.
299 StackWatermarkSet::after_unwind(current);
300 }
301
302 return fetch_unroll_info_helper(current, exec_mode);
303 JRT_END
304
305 #if COMPILER2_OR_JVMCI
306
307 static Klass* get_refined_array_klass(Klass* k, frame* fr, RegisterMap* map, ObjectValue* sv, TRAPS) {
308 // If it's an array, get the properties
309 if (k->is_array_klass() && !k->is_typeArray_klass()) {
310 assert(k->is_unrefined_objArray_klass(), "Expected unrefined array klass");
311 nmethod* nm = fr->cb()->as_nmethod_or_null();
312 if (nm->is_compiled_by_c2()) {
313 assert(sv->has_properties(), "Property information is missing");
314 ArrayProperties props(checked_cast<ArrayProperties::Type>(StackValue::create_stack_value(fr, map, sv->properties())->get_jint()));
315 k = ObjArrayKlass::cast(k)->klass_with_properties(props, THREAD);
316 } else {
317 // TODO Graal needs to be fixed. Just go with the default properties for now
318 k = ObjArrayKlass::cast(k)->klass_with_properties(ArrayProperties::Default(), THREAD);
319 }
320 }
321 return k;
322 }
323
324 // print information about reallocated objects
325 static void print_objects(JavaThread* deoptee_thread, frame* deoptee, RegisterMap* map,
326 GrowableArray<ScopeValue*>* objects, bool realloc_failures, TRAPS) {
327 ResourceMark rm;
328 stringStream st; // change to logStream with logging
329 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
330 fieldDescriptor fd;
331
332 for (int i = 0; i < objects->length(); i++) {
333 ObjectValue* sv = (ObjectValue*) objects->at(i);
334 Handle obj = sv->value();
335
336 if (obj.is_null()) {
337 st.print_cr(" nullptr");
338 continue;
339 }
340
341 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
342 k = get_refined_array_klass(k, deoptee, map, sv, THREAD);
343
344 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
345 k->print_value_on(&st);
346 st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize);
347
348 if (Verbose && k != nullptr) {
349 k->oop_print_on(obj(), &st);
350 }
351 }
352 tty->print_raw(st.freeze());
353 }
354
355 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
356 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
357 bool& deoptimized_objects) {
358 bool realloc_failures = false;
359 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
360
361 JavaThread* deoptee_thread = chunk->at(0)->thread();
362 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
363 "a frame can only be deoptimized by the owner thread");
364
365 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
366
367 // The flag return_oop() indicates call sites which return oop
368 // in compiled code. Such sites include java method calls,
369 // runtime calls (for example, used to allocate new objects/arrays
370 // on slow code path) and any other calls generated in compiled code.
371 // It is not guaranteed that we can get such information here only
372 // by analyzing bytecode in deoptimized frames. This is why this flag
373 // is set during method compilation (see Compile::Process_OopMap_Node()).
374 // If the previous frame was popped or if we are dispatching an exception,
375 // we don't have an oop result.
376 ScopeDesc* scope = chunk->at(0)->scope();
377 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
378 // In case of the return of multiple values, we must take care
379 // of all oop return values.
380 GrowableArray<Handle> return_oops;
381 InlineKlass* vk = nullptr;
382 if (save_oop_result && scope->return_scalarized()) {
383 vk = InlineKlass::returned_inline_klass(map);
384 if (vk != nullptr) {
385 vk->save_oop_fields(map, return_oops);
386 save_oop_result = false;
387 }
388 }
389 if (save_oop_result) {
390 // Reallocation may trigger GC. If deoptimization happened on return from
391 // call which returns oop we need to save it since it is not in oopmap.
392 oop result = deoptee.saved_oop_result(&map);
393 assert(oopDesc::is_oop_or_null(result), "must be oop");
394 return_oops.push(Handle(thread, result));
395 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
396 if (TraceDeoptimization) {
397 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
398 tty->cr();
399 }
400 }
401 if (objects != nullptr || vk != nullptr) {
402 if (exec_mode == Deoptimization::Unpack_none) {
403 assert(thread->thread_state() == _thread_in_vm, "assumption");
404 JavaThread* THREAD = thread; // For exception macros.
405 // Clear pending OOM if reallocation fails and return true indicating allocation failure
406 if (vk != nullptr) {
407 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, CHECK_AND_CLEAR_(true));
408 }
409 if (objects != nullptr) {
410 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
411 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
412 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
413 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci, CHECK_AND_CLEAR_(true));
414 }
415 deoptimized_objects = true;
416 } else {
417 JavaThread* current = thread; // For JRT_BLOCK
418 JRT_BLOCK
419 if (vk != nullptr) {
420 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, THREAD);
421 }
422 if (objects != nullptr) {
423 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
424 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
425 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
426 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci, THREAD);
427 }
428 JRT_END
429 }
430 if (TraceDeoptimization && objects != nullptr) {
431 print_objects(deoptee_thread, &deoptee, &map, objects, realloc_failures, thread);
432 }
433 }
434 if (save_oop_result || vk != nullptr) {
435 // Restore result.
436 assert(return_oops.length() == 1, "no inline type");
437 deoptee.set_saved_oop_result(&map, return_oops.pop()());
438 }
439 return realloc_failures;
440 }
441
442 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
443 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
444 JavaThread* deoptee_thread = chunk->at(0)->thread();
445 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
446 assert(thread == Thread::current(), "should be");
447 HandleMark hm(thread);
448 #ifndef PRODUCT
449 bool first = true;
450 #endif // !PRODUCT
451 // Start locking from outermost/oldest frame
452 for (int i = (chunk->length() - 1); i >= 0; i--) {
453 compiledVFrame* cvf = chunk->at(i);
454 assert (cvf->scope() != nullptr,"expect only compiled java frames");
455 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
456 if (monitors->is_nonempty()) {
457 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
458 exec_mode, realloc_failures);
459 deoptimized_objects = deoptimized_objects || relocked;
460 #ifndef PRODUCT
461 if (PrintDeoptimizationDetails) {
462 ResourceMark rm;
463 stringStream st;
464 for (int j = 0; j < monitors->length(); j++) {
465 MonitorInfo* mi = monitors->at(j);
466 if (mi->eliminated()) {
467 if (first) {
468 first = false;
469 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
470 }
471 if (exec_mode == Deoptimization::Unpack_none) {
472 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
473 if (monitor != nullptr && monitor->object() == mi->owner()) {
474 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
475 continue;
476 }
477 }
478 if (mi->owner_is_scalar_replaced()) {
479 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
480 st.print_cr(" failed reallocation for klass %s", k->external_name());
481 } else {
482 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
483 }
484 }
485 }
486 tty->print_raw(st.freeze());
487 }
488 #endif // !PRODUCT
489 }
490 }
491 }
492
493 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
494 // The given vframes cover one physical frame.
495 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
496 bool& realloc_failures) {
497 frame deoptee = chunk->at(0)->fr();
498 JavaThread* deoptee_thread = chunk->at(0)->thread();
499 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
500 RegisterMap map(chunk->at(0)->register_map());
501 bool deoptimized_objects = false;
502
503 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
504
505 // Reallocate the non-escaping objects and restore their fields.
506 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || Arguments::is_valhalla_enabled()) && EliminateAllocations)
507 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
508 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
509 }
510
511 // MonitorInfo structures used in eliminate_locks are not GC safe.
512 NoSafepointVerifier no_safepoint;
513
514 // Now relock objects if synchronization on them was eliminated.
515 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
516 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
517 }
518 return deoptimized_objects;
519 }
520 #endif // COMPILER2_OR_JVMCI
521
522 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
523 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
524 JFR_ONLY(Jfr::check_and_process_sample_request(current);)
525 // When we get here we are about to unwind the deoptee frame. In order to
526 // catch not yet safe to use frames, the following stack watermark barrier
527 // poll will make such frames safe to use.
528 StackWatermarkSet::before_unwind(current);
529
530 // Note: there is a safepoint safety issue here. No matter whether we enter
531 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
532 // the vframeArray is created.
533 //
534
535 // Allocate our special deoptimization ResourceMark
536 DeoptResourceMark* dmark = new DeoptResourceMark(current);
537 assert(current->deopt_mark() == nullptr, "Pending deopt!");
538 current->set_deopt_mark(dmark);
539
540 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
541 RegisterMap map(current,
542 RegisterMap::UpdateMap::include,
543 RegisterMap::ProcessFrames::include,
544 RegisterMap::WalkContinuation::skip);
545 RegisterMap dummy_map(current,
546 RegisterMap::UpdateMap::skip,
547 RegisterMap::ProcessFrames::include,
548 RegisterMap::WalkContinuation::skip);
549 // Now get the deoptee with a valid map
550 frame deoptee = stub_frame.sender(&map);
551 if (exec_mode == Unpack_deopt) {
552 assert(deoptee.is_deoptimized_frame(), "frame is not marked for deoptimization");
553 }
554 // Set the deoptee nmethod
555 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
556 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
557 current->set_deopt_compiled_method(nm);
558
559 if (VerifyStack) {
560 current->validate_frame_layout();
561 }
562
563 // Create a growable array of VFrames where each VFrame represents an inlined
564 // Java frame. This storage is allocated with the usual system arena.
565 assert(deoptee.is_compiled_frame(), "Wrong frame type");
566 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
567 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
568 while (!vf->is_top()) {
569 assert(vf->is_compiled_frame(), "Wrong frame type");
570 chunk->push(compiledVFrame::cast(vf));
571 vf = vf->sender();
572 }
573 assert(vf->is_compiled_frame(), "Wrong frame type");
574 chunk->push(compiledVFrame::cast(vf));
575
576 bool realloc_failures = false;
577
578 #if COMPILER2_OR_JVMCI
579 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
580
581 // Reallocate the non-escaping objects and restore their fields. Then
582 // relock objects if synchronization on them was eliminated.
583 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || Arguments::is_valhalla_enabled()) && EliminateAllocations)
584 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
585 bool unused;
586 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
587 }
588 #endif // COMPILER2_OR_JVMCI
589
590 // Ensure that no safepoint is taken after pointers have been stored
591 // in fields of rematerialized objects. If a safepoint occurs from here on
592 // out the java state residing in the vframeArray will be missed.
593 // Locks may be rebaised in a safepoint.
594 NoSafepointVerifier no_safepoint;
595
596 #if COMPILER2_OR_JVMCI
597 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
598 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
599 bool unused = false;
600 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
601 }
602 #endif // COMPILER2_OR_JVMCI
603
604 ScopeDesc* trap_scope = chunk->at(0)->scope();
605 Handle exceptionObject;
606 if (trap_scope->rethrow_exception()) {
607 #ifndef PRODUCT
608 if (PrintDeoptimizationDetails) {
609 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());
610 }
611 #endif // !PRODUCT
612
613 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
614 guarantee(expressions != nullptr && expressions->length() == 1, "should have only exception on stack");
615 guarantee(exec_mode != Unpack_exception, "rethrow_exception set with Unpack_exception");
616 ScopeValue* topOfStack = expressions->top();
617 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
618 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
619 }
620
621 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
622 #if COMPILER2_OR_JVMCI
623 if (realloc_failures) {
624 // This destroys all ScopedValue bindings.
625 current->clear_scopedValueBindings();
626 pop_frames_failed_reallocs(current, array);
627 }
628 #endif
629
630 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
631 current->set_vframe_array_head(array);
632
633 // Now that the vframeArray has been created if we have any deferred local writes
634 // added by jvmti then we can free up that structure as the data is now in the
635 // vframeArray
636
637 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
638
639 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
640 CodeBlob* cb = stub_frame.cb();
641 // Verify we have the right vframeArray
642 assert(cb->frame_size() >= 0, "Unexpected frame size");
643 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
644 assert(unpack_sp == deoptee.unextended_sp(), "must be");
645
646 #ifdef ASSERT
647 assert(cb->is_deoptimization_stub() ||
648 cb->is_uncommon_trap_stub() ||
649 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
650 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
651 "unexpected code blob: %s", cb->name());
652 #endif
653
654 // This is a guarantee instead of an assert because if vframe doesn't match
655 // we will unpack the wrong deoptimized frame and wind up in strange places
656 // where it will be very difficult to figure out what went wrong. Better
657 // to die an early death here than some very obscure death later when the
658 // trail is cold.
659 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
660
661 int number_of_frames = array->frames();
662
663 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
664 // virtual activation, which is the reverse of the elements in the vframes array.
665 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
666 // +1 because we always have an interpreter return address for the final slot.
667 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
668 int popframe_extra_args = 0;
669 // Create an interpreter return address for the stub to use as its return
670 // address so the skeletal frames are perfectly walkable
671 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
672
673 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
674 // activation be put back on the expression stack of the caller for reexecution
675 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
676 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
677 }
678
679 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
680 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
681 // than simply use array->sender.pc(). This requires us to walk the current set of frames
682 //
683 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
684 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
685
686 // It's possible that the number of parameters at the call site is
687 // different than number of arguments in the callee when method
688 // handles are used. If the caller is interpreted get the real
689 // value so that the proper amount of space can be added to it's
690 // frame.
691 bool caller_was_method_handle = false;
692 if (deopt_sender.is_interpreted_frame()) {
693 methodHandle method(current, deopt_sender.interpreter_frame_method());
694 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
695 if (cur.has_member_arg()) {
696 // This should cover all real-world cases. One exception is a pathological chain of
697 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we
698 // would need to get the size from the resolved method entry. Another exception would
699 // be an invokedynamic with an adapter that is really a MethodHandle linker.
700 caller_was_method_handle = true;
701 }
702 }
703
704 //
705 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
706 // frame_sizes/frame_pcs[1] next oldest frame (int)
707 // frame_sizes/frame_pcs[n] youngest frame (int)
708 //
709 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
710 // owns the space for the return address to it's caller). Confusing ain't it.
711 //
712 // The vframe array can address vframes with indices running from
713 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
714 // When we create the skeletal frames we need the oldest frame to be in the zero slot
715 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
716 // so things look a little strange in this loop.
717 //
718 int callee_parameters = 0;
719 int callee_locals = 0;
720 for (int index = 0; index < array->frames(); index++ ) {
721 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
722 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
723 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
724 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
725 callee_locals,
726 index == 0,
727 popframe_extra_args);
728 // This pc doesn't have to be perfect just good enough to identify the frame
729 // as interpreted so the skeleton frame will be walkable
730 // The correct pc will be set when the skeleton frame is completely filled out
731 // The final pc we store in the loop is wrong and will be overwritten below
732 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
733
734 callee_parameters = array->element(index)->method()->size_of_parameters();
735 callee_locals = array->element(index)->method()->max_locals();
736 popframe_extra_args = 0;
737 }
738
739 // Compute whether the root vframe returns a float or double value.
740 BasicType return_type;
741 {
742 methodHandle method(current, array->element(0)->method());
743 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
744 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
745 }
746
747 // Compute information for handling adapters and adjusting the frame size of the caller.
748 int caller_adjustment = 0;
749
750 // Compute the amount the oldest interpreter frame will have to adjust
751 // its caller's stack by. If the caller is a compiled frame then
752 // we pretend that the callee has no parameters so that the
753 // extension counts for the full amount of locals and not just
754 // locals-parms. This is because without a c2i adapter the parm
755 // area as created by the compiled frame will not be usable by
756 // the interpreter. (Depending on the calling convention there
757 // may not even be enough space).
758
759 // QQQ I'd rather see this pushed down into last_frame_adjust
760 // and have it take the sender (aka caller).
761
762 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
763 caller_adjustment = last_frame_adjust(0, callee_locals);
764 } else if (callee_locals > callee_parameters) {
765 // The caller frame may need extending to accommodate
766 // non-parameter locals of the first unpacked interpreted frame.
767 // Compute that adjustment.
768 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
769 }
770
771 // If the sender is deoptimized we must retrieve the address of the handler
772 // since the frame will "magically" show the original pc before the deopt
773 // and we'd undo the deopt.
774
775 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
776 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
777 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
778 }
779
780 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
781
782 #if INCLUDE_JVMCI
783 if (exceptionObject() != nullptr) {
784 current->set_exception_oop(exceptionObject());
785 exec_mode = Unpack_exception;
786 assert(array->element(0)->rethrow_exception(), "must be");
787 }
788 #endif
789
790 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
791 assert(current->has_pending_exception(), "should have thrown OOME");
792 current->set_exception_oop(current->pending_exception());
793 current->clear_pending_exception();
794 exec_mode = Unpack_exception;
795 }
796
797 #if INCLUDE_JVMCI
798 if (current->frames_to_pop_failed_realloc() > 0) {
799 current->set_pending_monitorenter(false);
800 }
801 #endif
802
803 int caller_actual_parameters = -1; // value not used except for interpreted frames, see below
804 if (deopt_sender.is_interpreted_frame()) {
805 caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0);
806 }
807
808 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
809 caller_adjustment * BytesPerWord,
810 caller_actual_parameters,
811 number_of_frames,
812 frame_sizes,
813 frame_pcs,
814 return_type,
815 exec_mode);
816 // On some platforms, we need a way to pass some platform dependent
817 // information to the unpacking code so the skeletal frames come out
818 // correct (initial fp value, unextended sp, ...)
819 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
820
821 if (array->frames() > 1) {
822 if (VerifyStack && TraceDeoptimization) {
823 tty->print_cr("Deoptimizing method containing inlining");
824 }
825 }
826
827 array->set_unroll_block(info);
828 return info;
829 }
830
831 // Called to cleanup deoptimization data structures in normal case
832 // after unpacking to stack and when stack overflow error occurs
833 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
834 vframeArray *array) {
835
836 // Get array if coming from exception
837 if (array == nullptr) {
838 array = thread->vframe_array_head();
839 }
840 thread->set_vframe_array_head(nullptr);
841
842 // Free the previous UnrollBlock
843 vframeArray* old_array = thread->vframe_array_last();
844 thread->set_vframe_array_last(array);
845
846 if (old_array != nullptr) {
847 UnrollBlock* old_info = old_array->unroll_block();
848 old_array->set_unroll_block(nullptr);
849 delete old_info;
850 delete old_array;
851 }
852
853 // Deallocate any resource creating in this routine and any ResourceObjs allocated
854 // inside the vframeArray (StackValueCollections)
855
856 delete thread->deopt_mark();
857 thread->set_deopt_mark(nullptr);
858 thread->set_deopt_compiled_method(nullptr);
859
860
861 if (JvmtiExport::can_pop_frame()) {
862 // Regardless of whether we entered this routine with the pending
863 // popframe condition bit set, we should always clear it now
864 thread->clear_popframe_condition();
865 }
866
867 // unpack_frames() is called at the end of the deoptimization handler
868 // and (in C2) at the end of the uncommon trap handler. Note this fact
869 // so that an asynchronous stack walker can work again. This counter is
870 // incremented at the beginning of fetch_unroll_info() and (in C2) at
871 // the beginning of uncommon_trap().
872 thread->dec_in_deopt_handler();
873 }
874
875 // Moved from cpu directories because none of the cpus has callee save values.
876 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
877 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
878
879 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
880 // the days we had adapter frames. When we deoptimize a situation where a
881 // compiled caller calls a compiled caller will have registers it expects
882 // to survive the call to the callee. If we deoptimize the callee the only
883 // way we can restore these registers is to have the oldest interpreter
884 // frame that we create restore these values. That is what this routine
885 // will accomplish.
886
887 // At the moment we have modified c2 to not have any callee save registers
888 // so this problem does not exist and this routine is just a place holder.
889
890 assert(f->is_interpreted_frame(), "must be interpreted");
891 }
892
893 #ifndef PRODUCT
894 #ifdef ASSERT
895 // Return true if the execution after the provided bytecode continues at the
896 // next bytecode in the code. This is not the case for gotos, returns, and
897 // throws.
898 static bool falls_through(Bytecodes::Code bc) {
899 switch (bc) {
900 case Bytecodes::_goto:
901 case Bytecodes::_goto_w:
902 case Bytecodes::_athrow:
903 case Bytecodes::_areturn:
904 case Bytecodes::_dreturn:
905 case Bytecodes::_freturn:
906 case Bytecodes::_ireturn:
907 case Bytecodes::_lreturn:
908 case Bytecodes::_jsr:
909 case Bytecodes::_ret:
910 case Bytecodes::_return:
911 case Bytecodes::_lookupswitch:
912 case Bytecodes::_tableswitch:
913 return false;
914 default:
915 return true;
916 }
917 }
918 #endif
919 #endif
920
921 // Return BasicType of value being returned
922 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
923 assert(thread == JavaThread::current(), "pre-condition");
924
925 // We are already active in the special DeoptResourceMark any ResourceObj's we
926 // allocate will be freed at the end of the routine.
927
928 // JRT_LEAF methods don't normally allocate handles and there is a
929 // NoHandleMark to enforce that. It is actually safe to use Handles
930 // in a JRT_LEAF method, and sometimes desirable, but to do so we
931 // must use ResetNoHandleMark to bypass the NoHandleMark, and
932 // then use a HandleMark to ensure any Handles we do create are
933 // cleaned up in this scope.
934 ResetNoHandleMark rnhm;
935 HandleMark hm(thread);
936
937 frame stub_frame = thread->last_frame();
938
939 Continuation::notify_deopt(thread, stub_frame.sp());
940
941 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
942 // must point to the vframeArray for the unpack frame.
943 vframeArray* array = thread->vframe_array_head();
944 UnrollBlock* info = array->unroll_block();
945
946 // We set the last_Java frame. But the stack isn't really parsable here. So we
947 // clear it to make sure JFR understands not to try and walk stacks from events
948 // in here.
949 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
950 thread->frame_anchor()->set_last_Java_sp(nullptr);
951
952 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
953 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
954
955 thread->frame_anchor()->set_last_Java_sp(sp);
956
957 BasicType bt = info->return_type();
958
959 // If we have an exception pending, claim that the return type is an oop
960 // so the deopt_blob does not overwrite the exception_oop.
961
962 if (exec_mode == Unpack_exception)
963 bt = T_OBJECT;
964
965 // Cleanup thread deopt data
966 cleanup_deopt_info(thread, array);
967
968 #ifndef PRODUCT
969 if (VerifyStack) {
970 ResourceMark res_mark;
971 // Clear pending exception to not break verification code (restored afterwards)
972 PreserveExceptionMark pm(thread);
973
974 thread->validate_frame_layout();
975
976 // Verify that the just-unpacked frames match the interpreter's
977 // notions of expression stack and locals
978 vframeArray* cur_array = thread->vframe_array_last();
979 RegisterMap rm(thread,
980 RegisterMap::UpdateMap::skip,
981 RegisterMap::ProcessFrames::include,
982 RegisterMap::WalkContinuation::skip);
983 rm.set_include_argument_oops(false);
984 int callee_size_of_parameters = 0;
985 for (int frame_idx = 0; frame_idx < cur_array->frames(); frame_idx++) {
986 bool is_top_frame = (frame_idx == 0);
987 vframeArrayElement* el = cur_array->element(frame_idx);
988 frame* iframe = el->iframe();
989 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
990 methodHandle mh(thread, iframe->interpreter_frame_method());
991 bool reexecute = el->should_reexecute();
992
993 int cur_invoke_parameter_size = 0;
994 int top_frame_expression_stack_adjustment = 0;
995 int max_bci = mh->code_size();
996 BytecodeStream str(mh, iframe->interpreter_frame_bci());
997 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
998 Bytecodes::Code cur_code = str.next();
999
1000 if (!reexecute && !Bytecodes::is_invoke(cur_code)) {
1001 // We can only compute OopMaps for the before state, so we need to roll forward
1002 // to the next bytecode.
1003 assert(is_top_frame, "must be");
1004 assert(falls_through(cur_code), "must be");
1005 assert(cur_code != Bytecodes::_illegal, "illegal bytecode");
1006 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
1007
1008 // Need to subtract off the size of the result type of
1009 // the bytecode because this is not described in the
1010 // debug info but returned to the interpreter in the TOS
1011 // caching register
1012 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
1013 if (bytecode_result_type != T_ILLEGAL) {
1014 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
1015 }
1016 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
1017
1018 cur_code = str.next();
1019 // Reflect the fact that we have rolled forward and now need
1020 // top_frame_expression_stack_adjustment
1021 reexecute = true;
1022 }
1023
1024 assert(cur_code != Bytecodes::_illegal, "illegal bytecode");
1025 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
1026
1027 // Get the oop map for this bci
1028 InterpreterOopMap mask;
1029 OopMapCache::compute_one_oop_map(mh, str.bci(), &mask);
1030 // Check to see if we can grab the number of outgoing arguments
1031 // at an uncommon trap for an invoke (where the compiler
1032 // generates debug info before the invoke has executed)
1033 if (Bytecodes::is_invoke(cur_code)) {
1034 Bytecode_invoke invoke(mh, str.bci());
1035 cur_invoke_parameter_size = invoke.size_of_parameters();
1036 if (!is_top_frame && invoke.has_member_arg()) {
1037 callee_size_of_parameters++;
1038 }
1039 }
1040
1041 // Verify stack depth and oops in frame
1042 auto match = [&]() {
1043 int iframe_expr_ssize = iframe->interpreter_frame_expression_stack_size();
1044 #if INCLUDE_JVMCI
1045 if (is_top_frame && el->rethrow_exception()) {
1046 return iframe_expr_ssize == 1;
1047 }
1048 #endif
1049 // This should only be needed for C1
1050 if (is_top_frame && exec_mode == Unpack_exception && iframe_expr_ssize == 0) {
1051 return true;
1052 }
1053 if (reexecute) {
1054 int expr_ssize_before = iframe_expr_ssize + top_frame_expression_stack_adjustment;
1055 int oopmap_expr_invoke_ssize = mask.expression_stack_size() + cur_invoke_parameter_size;
1056 return expr_ssize_before == oopmap_expr_invoke_ssize;
1057 } else {
1058 int oopmap_expr_callee_ssize = mask.expression_stack_size() + callee_size_of_parameters;
1059 return iframe_expr_ssize == oopmap_expr_callee_ssize;
1060 }
1061 };
1062 if (!match()) {
1063 // Print out some information that will help us debug the problem
1064 tty->print_cr("Wrong number of expression stack elements during deoptimization");
1065 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", frame_idx, cur_array->frames() - 1);
1066 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1067 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1068 iframe->interpreter_frame_expression_stack_size());
1069 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1070 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1071 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1072 tty->print_cr(" exec_mode = %d", exec_mode);
1073 tty->print_cr(" original should_reexecute = %s", el->should_reexecute() ? "true" : "false");
1074 tty->print_cr(" reexecute = %s%s", reexecute ? "true" : "false",
1075 (reexecute != el->should_reexecute()) ? " (changed)" : "");
1076 #if INCLUDE_JVMCI
1077 tty->print_cr(" rethrow_exception = %s", el->rethrow_exception() ? "true" : "false");
1078 #endif
1079 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1080 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1081 tty->print_cr(" Interpreted frames:");
1082 for (int k = 0; k < cur_array->frames(); k++) {
1083 vframeArrayElement* el = cur_array->element(k);
1084 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1085 }
1086 cur_array->print_on_2(tty);
1087 guarantee(false, "wrong number of expression stack elements during deopt");
1088 }
1089 VerifyOopClosure verify;
1090 iframe->oops_interpreted_do(&verify, &rm, false);
1091 callee_size_of_parameters = mh->size_of_parameters();
1092 }
1093 }
1094 #endif // !PRODUCT
1095
1096 return bt;
1097 JRT_END
1098
1099 class DeoptimizeMarkedHandshakeClosure : public HandshakeClosure {
1100 public:
1101 DeoptimizeMarkedHandshakeClosure() : HandshakeClosure("Deoptimize") {}
1102 void do_thread(Thread* thread) {
1103 JavaThread* jt = JavaThread::cast(thread);
1104 jt->deoptimize_marked_methods();
1105 }
1106 };
1107
1108 void Deoptimization::deoptimize_all_marked() {
1109 ResourceMark rm;
1110
1111 // Make the dependent methods not entrant
1112 CodeCache::make_marked_nmethods_deoptimized();
1113
1114 DeoptimizeMarkedHandshakeClosure deopt;
1115 if (SafepointSynchronize::is_at_safepoint()) {
1116 Threads::java_threads_do(&deopt);
1117 } else {
1118 Handshake::execute(&deopt);
1119 }
1120 }
1121
1122 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1123 = Deoptimization::Action_reinterpret;
1124
1125 #if INCLUDE_JVMCI
1126 template<typename CacheType>
1127 class BoxCacheBase : public CHeapObj<mtCompiler> {
1128 protected:
1129 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1130 ResourceMark rm(thread);
1131 char* klass_name_str = klass_name->as_C_string();
1132 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1133 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1134 if (!ik->is_in_error_state()) {
1135 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1136 CacheType::compute_offsets(ik);
1137 }
1138 return ik;
1139 }
1140 };
1141
1142 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1143 PrimitiveType _low;
1144 PrimitiveType _high;
1145 jobject _cache;
1146 protected:
1147 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1148 BoxCache(Thread* thread) {
1149 assert(!Arguments::is_valhalla_enabled(), "Should not use box caches with enable preview");
1150 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1151 if (ik->is_in_error_state()) {
1152 _low = 1;
1153 _high = 0;
1154 _cache = nullptr;
1155 } else {
1156 refArrayOop cache = CacheType::cache(ik);
1157 assert(cache->length() > 0, "Empty cache");
1158 _low = BoxType::value(cache->obj_at(0));
1159 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1160 _cache = JNIHandles::make_global(Handle(thread, cache));
1161 }
1162 }
1163 ~BoxCache() {
1164 JNIHandles::destroy_global(_cache);
1165 }
1166 public:
1167 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1168 if (_singleton == nullptr) {
1169 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1170 if (!AtomicAccess::replace_if_null(&_singleton, s)) {
1171 delete s;
1172 }
1173 }
1174 return _singleton;
1175 }
1176 oop lookup(PrimitiveType value) {
1177 if (_low <= value && value <= _high) {
1178 int offset = checked_cast<int>(value - _low);
1179 return refArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1180 }
1181 return nullptr;
1182 }
1183 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1184 if (_cache == nullptr) {
1185 cache_init_error = true;
1186 return nullptr;
1187 }
1188 // Have to cast to avoid little/big-endian problems.
1189 if (sizeof(PrimitiveType) > sizeof(jint)) {
1190 jlong value = (jlong)raw_value;
1191 return lookup(value);
1192 }
1193 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1194 return lookup(value);
1195 }
1196 };
1197
1198 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1199 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1200 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1201 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1202 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1203
1204 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1205 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1206 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1207 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1208 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1209
1210 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1211 jobject _true_cache;
1212 jobject _false_cache;
1213 protected:
1214 static BooleanBoxCache *_singleton;
1215 BooleanBoxCache(Thread *thread) {
1216 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1217 if (ik->is_in_error_state()) {
1218 _true_cache = nullptr;
1219 _false_cache = nullptr;
1220 } else {
1221 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1222 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1223 }
1224 }
1225 ~BooleanBoxCache() {
1226 JNIHandles::destroy_global(_true_cache);
1227 JNIHandles::destroy_global(_false_cache);
1228 }
1229 public:
1230 static BooleanBoxCache* singleton(Thread* thread) {
1231 if (_singleton == nullptr) {
1232 BooleanBoxCache* s = new BooleanBoxCache(thread);
1233 if (!AtomicAccess::replace_if_null(&_singleton, s)) {
1234 delete s;
1235 }
1236 }
1237 return _singleton;
1238 }
1239 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1240 if (_true_cache == nullptr) {
1241 cache_in_error = true;
1242 return nullptr;
1243 }
1244 // Have to cast to avoid little/big-endian problems.
1245 jboolean value = (jboolean)*((jint*)&raw_value);
1246 return lookup(value);
1247 }
1248 oop lookup(jboolean value) {
1249 if (value != 0) {
1250 return JNIHandles::resolve_non_null(_true_cache);
1251 }
1252 return JNIHandles::resolve_non_null(_false_cache);
1253 }
1254 };
1255
1256 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1257
1258 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1259 if (Arguments::enable_preview()) {
1260 // Box caches are not used with enable preview.
1261 return nullptr;
1262 }
1263 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1264 BasicType box_type = vmClasses::box_klass_type(k);
1265 if (box_type != T_OBJECT) {
1266 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1267 switch(box_type) {
1268 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1269 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1270 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1271 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1272 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1273 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1274 default:;
1275 }
1276 }
1277 return nullptr;
1278 }
1279 #endif // INCLUDE_JVMCI
1280
1281 #if COMPILER2_OR_JVMCI
1282 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1283 Handle pending_exception(THREAD, thread->pending_exception());
1284 const char* exception_file = thread->exception_file();
1285 int exception_line = thread->exception_line();
1286 thread->clear_pending_exception();
1287
1288 bool failures = false;
1289
1290 for (int i = 0; i < objects->length(); i++) {
1291 assert(objects->at(i)->is_object(), "invalid debug information");
1292 ObjectValue* sv = (ObjectValue*) objects->at(i);
1293
1294 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1295
1296 k = get_refined_array_klass(k, fr, reg_map, sv, THREAD);
1297
1298 // Check if the object may be null and has an additional null_marker input that needs
1299 // to be checked before using the field values. Skip re-allocation if it is null.
1300 if (k->is_inline_klass() && sv->has_properties()) {
1301 jint null_marker = StackValue::create_stack_value(fr, reg_map, sv->properties())->get_jint();
1302 if (null_marker == 0) {
1303 continue;
1304 }
1305 }
1306
1307 oop obj = nullptr;
1308 bool cache_init_error = false;
1309 if (k->is_instance_klass()) {
1310 #if INCLUDE_JVMCI
1311 nmethod* nm = fr->cb()->as_nmethod_or_null();
1312 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1313 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1314 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1315 if (obj != nullptr) {
1316 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1317 abv->set_cached(true);
1318 } else if (cache_init_error) {
1319 // Results in an OOME which is valid (as opposed to a class initialization error)
1320 // and is fine for the rare case a cache initialization failing.
1321 failures = true;
1322 }
1323 }
1324 #endif // INCLUDE_JVMCI
1325
1326 InstanceKlass* ik = InstanceKlass::cast(k);
1327 if (obj == nullptr && !cache_init_error) {
1328 InternalOOMEMark iom(THREAD);
1329 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1330 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1331 } else {
1332 obj = ik->allocate_instance(THREAD);
1333 }
1334 }
1335 } else if (k->is_flatArray_klass()) {
1336 FlatArrayKlass* ak = FlatArrayKlass::cast(k);
1337 // Inline type array must be zeroed because not all memory is reassigned
1338 InternalOOMEMark iom(THREAD);
1339 obj = ak->allocate_instance(sv->field_size(), THREAD);
1340 } else if (k->is_typeArray_klass()) {
1341 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1342 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1343 int len = sv->field_size() / type2size[ak->element_type()];
1344 InternalOOMEMark iom(THREAD);
1345 obj = ak->allocate_instance(len, THREAD);
1346 } else if (k->is_refArray_klass()) {
1347 RefArrayKlass* ak = RefArrayKlass::cast(k);
1348 InternalOOMEMark iom(THREAD);
1349 obj = ak->allocate_instance(sv->field_size(), THREAD);
1350 }
1351
1352 if (obj == nullptr) {
1353 failures = true;
1354 }
1355
1356 assert(sv->value().is_null(), "redundant reallocation");
1357 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1358 CLEAR_PENDING_EXCEPTION;
1359 sv->set_value(obj);
1360 }
1361
1362 if (failures) {
1363 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1364 } else if (pending_exception.not_null()) {
1365 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1366 }
1367
1368 return failures;
1369 }
1370
1371 // We're deoptimizing at the return of a call, inline type fields are
1372 // in registers. When we go back to the interpreter, it will expect a
1373 // reference to an inline type instance. Allocate and initialize it from
1374 // the register values here.
1375 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
1376 oop new_vt = vk->realloc_result(map, return_oops, THREAD);
1377 if (new_vt == nullptr) {
1378 CLEAR_PENDING_EXCEPTION;
1379 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
1380 }
1381 return_oops.clear();
1382 return_oops.push(Handle(THREAD, new_vt));
1383 return false;
1384 }
1385
1386 #if INCLUDE_JVMCI
1387 /**
1388 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1389 * we need to somehow be able to recover the actual kind to be able to write the correct
1390 * amount of bytes.
1391 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1392 * the entries at index n + 1 to n + i are 'markers'.
1393 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1394 * expected form of the array would be:
1395 *
1396 * {b0, b1, b2, b3, INT, marker, b6, b7}
1397 *
1398 * Thus, in order to get back the size of the entry, we simply need to count the number
1399 * of marked entries
1400 *
1401 * @param virtualArray the virtualized byte array
1402 * @param i index of the virtual entry we are recovering
1403 * @return The number of bytes the entry spans
1404 */
1405 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1406 int index = i;
1407 while (++index < virtualArray->field_size() &&
1408 virtualArray->field_at(index)->is_marker()) {}
1409 return index - i;
1410 }
1411
1412 /**
1413 * If there was a guarantee for byte array to always start aligned to a long, we could
1414 * do a simple check on the parity of the index. Unfortunately, that is not always the
1415 * case. Thus, we check alignment of the actual address we are writing to.
1416 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1417 * write a long to index 3.
1418 */
1419 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1420 jbyte* res = obj->byte_at_addr(index);
1421 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1422 return res;
1423 }
1424
1425 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1426 switch (byte_count) {
1427 case 1:
1428 obj->byte_at_put(index, (jbyte) value->get_jint());
1429 break;
1430 case 2:
1431 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1432 break;
1433 case 4:
1434 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1435 break;
1436 case 8:
1437 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1438 break;
1439 default:
1440 ShouldNotReachHere();
1441 }
1442 }
1443 #endif // INCLUDE_JVMCI
1444
1445
1446 // restore elements of an eliminated type array
1447 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1448 int index = 0;
1449
1450 for (int i = 0; i < sv->field_size(); i++) {
1451 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1452 switch(type) {
1453 case T_LONG: case T_DOUBLE: {
1454 assert(value->type() == T_INT, "Agreement.");
1455 StackValue* low =
1456 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1457 #ifdef _LP64
1458 jlong res = (jlong)low->get_intptr();
1459 #else
1460 jlong res = jlong_from(value->get_jint(), low->get_jint());
1461 #endif
1462 obj->long_at_put(index, res);
1463 break;
1464 }
1465
1466 case T_INT: case T_FLOAT: { // 4 bytes.
1467 assert(value->type() == T_INT, "Agreement.");
1468 #if INCLUDE_JVMCI
1469 // big_value allows encoding double/long value as e.g. [int = 0, long], and storing
1470 // the value in two array elements.
1471 bool big_value = false;
1472 if (i + 1 < sv->field_size() && type == T_INT) {
1473 if (sv->field_at(i)->is_location()) {
1474 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1475 if (type == Location::dbl || type == Location::lng) {
1476 big_value = true;
1477 }
1478 } else if (sv->field_at(i)->is_constant_int()) {
1479 ScopeValue* next_scope_field = sv->field_at(i + 1);
1480 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1481 big_value = true;
1482 }
1483 }
1484 }
1485
1486 if (big_value) {
1487 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1488 #ifdef _LP64
1489 jlong res = (jlong)low->get_intptr();
1490 #else
1491 jlong res = jlong_from(value->get_jint(), low->get_jint());
1492 #endif
1493 obj->int_at_put(index, *(jint*)&res);
1494 obj->int_at_put(++index, *((jint*)&res + 1));
1495 } else {
1496 obj->int_at_put(index, value->get_jint());
1497 }
1498 #else // not INCLUDE_JVMCI
1499 obj->int_at_put(index, value->get_jint());
1500 #endif // INCLUDE_JVMCI
1501 break;
1502 }
1503
1504 case T_SHORT:
1505 assert(value->type() == T_INT, "Agreement.");
1506 obj->short_at_put(index, (jshort)value->get_jint());
1507 break;
1508
1509 case T_CHAR:
1510 assert(value->type() == T_INT, "Agreement.");
1511 obj->char_at_put(index, (jchar)value->get_jint());
1512 break;
1513
1514 case T_BYTE: {
1515 assert(value->type() == T_INT, "Agreement.");
1516 #if INCLUDE_JVMCI
1517 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1518 int byte_count = count_number_of_bytes_for_entry(sv, i);
1519 byte_array_put(obj, value, index, byte_count);
1520 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1521 i += byte_count - 1; // Balance the loop counter.
1522 index += byte_count;
1523 // index has been updated so continue at top of loop
1524 continue;
1525 #else
1526 obj->byte_at_put(index, (jbyte)value->get_jint());
1527 break;
1528 #endif // INCLUDE_JVMCI
1529 }
1530
1531 case T_BOOLEAN: {
1532 assert(value->type() == T_INT, "Agreement.");
1533 obj->bool_at_put(index, (jboolean)value->get_jint());
1534 break;
1535 }
1536
1537 default:
1538 ShouldNotReachHere();
1539 }
1540 index++;
1541 }
1542 }
1543
1544 // restore fields of an eliminated object array
1545 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1546 for (int i = 0; i < sv->field_size(); i++) {
1547 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1548 assert(value->type() == T_OBJECT, "object element expected");
1549 obj->obj_at_put(i, value->get_obj()());
1550 }
1551 }
1552
1553 class ReassignedField {
1554 public:
1555 int _offset;
1556 BasicType _type;
1557 InstanceKlass* _klass;
1558 bool _is_flat;
1559 bool _is_null_free;
1560 public:
1561 ReassignedField() : _offset(0), _type(T_ILLEGAL), _klass(nullptr), _is_flat(false), _is_null_free(false) { }
1562 };
1563
1564 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1565 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1566 InstanceKlass* super = klass->super();
1567 if (super != nullptr) {
1568 get_reassigned_fields(super, fields, is_jvmci);
1569 }
1570 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1571 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1572 ReassignedField field;
1573 field._offset = fs.offset();
1574 field._type = Signature::basic_type(fs.signature());
1575 if (fs.is_flat()) {
1576 field._is_flat = true;
1577 field._is_null_free = fs.is_null_free_inline_type();
1578 // Resolve klass of flat inline type field
1579 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index()));
1580 }
1581 fields->append(field);
1582 }
1583 }
1584 return fields;
1585 }
1586
1587 // Restore fields of an eliminated instance object employing the same field order used by the
1588 // compiler when it scalarizes an object at safepoints.
1589 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool is_jvmci, int base_offset, TRAPS) {
1590 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1591 for (int i = 0; i < fields->length(); i++) {
1592 BasicType type = fields->at(i)._type;
1593 int offset = base_offset + fields->at(i)._offset;
1594 // Check for flat inline type field before accessing the ScopeValue because it might not have any fields
1595 if (fields->at(i)._is_flat) {
1596 // Recursively re-assign flat inline type fields
1597 InstanceKlass* vk = fields->at(i)._klass;
1598 assert(vk != nullptr, "must be resolved");
1599 offset -= InlineKlass::cast(vk)->payload_offset(); // Adjust offset to omit oop header
1600 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, is_jvmci, offset, CHECK_0);
1601 if (!fields->at(i)._is_null_free) {
1602 ScopeValue* scope_field = sv->field_at(svIndex);
1603 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1604 int nm_offset = offset + InlineKlass::cast(vk)->null_marker_offset();
1605 obj->bool_field_put(nm_offset, value->get_jint() & 1);
1606 svIndex++;
1607 }
1608 continue; // Continue because we don't need to increment svIndex
1609 }
1610
1611 ScopeValue* scope_field = sv->field_at(svIndex);
1612 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1613 switch (type) {
1614 case T_OBJECT: case T_ARRAY:
1615 assert(value->type() == T_OBJECT, "Agreement.");
1616 obj->obj_field_put(offset, value->get_obj()());
1617 break;
1618
1619 case T_INT: case T_FLOAT: { // 4 bytes.
1620 assert(value->type() == T_INT, "Agreement.");
1621 bool big_value = false;
1622 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1623 if (scope_field->is_location()) {
1624 Location::Type type = ((LocationValue*) scope_field)->location().type();
1625 if (type == Location::dbl || type == Location::lng) {
1626 big_value = true;
1627 }
1628 }
1629 if (scope_field->is_constant_int()) {
1630 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1631 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1632 big_value = true;
1633 }
1634 }
1635 }
1636
1637 if (big_value) {
1638 i++;
1639 assert(i < fields->length(), "second T_INT field needed");
1640 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1641 } else {
1642 obj->int_field_put(offset, value->get_jint());
1643 break;
1644 }
1645 }
1646 /* no break */
1647
1648 case T_LONG: case T_DOUBLE: {
1649 assert(value->type() == T_INT, "Agreement.");
1650 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1651 #ifdef _LP64
1652 jlong res = (jlong)low->get_intptr();
1653 #else
1654 jlong res = jlong_from(value->get_jint(), low->get_jint());
1655 #endif
1656 obj->long_field_put(offset, res);
1657 break;
1658 }
1659
1660 case T_SHORT:
1661 assert(value->type() == T_INT, "Agreement.");
1662 obj->short_field_put(offset, (jshort)value->get_jint());
1663 break;
1664
1665 case T_CHAR:
1666 assert(value->type() == T_INT, "Agreement.");
1667 obj->char_field_put(offset, (jchar)value->get_jint());
1668 break;
1669
1670 case T_BYTE:
1671 assert(value->type() == T_INT, "Agreement.");
1672 obj->byte_field_put(offset, (jbyte)value->get_jint());
1673 break;
1674
1675 case T_BOOLEAN:
1676 assert(value->type() == T_INT, "Agreement.");
1677 obj->bool_field_put(offset, (jboolean)value->get_jint());
1678 break;
1679
1680 default:
1681 ShouldNotReachHere();
1682 }
1683 svIndex++;
1684 }
1685 return svIndex;
1686 }
1687
1688 // restore fields of an eliminated inline type array
1689 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool is_jvmci, TRAPS) {
1690 InlineKlass* vk = vak->element_klass();
1691 assert(vk->maybe_flat_in_array(), "should only be used for flat inline type arrays");
1692 // Adjust offset to omit oop header
1693 int base_offset = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT) - vk->payload_offset();
1694 // Initialize all elements of the flat inline type array
1695 for (int i = 0; i < sv->field_size(); i++) {
1696 ObjectValue* val = sv->field_at(i)->as_ObjectValue();
1697 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1698 reassign_fields_by_klass(vk, fr, reg_map, val, 0, (oop)obj, is_jvmci, offset, CHECK);
1699 if (!obj->is_null_free_array()) {
1700 jboolean null_marker_value;
1701 if (val->has_properties()) {
1702 null_marker_value = StackValue::create_stack_value(fr, reg_map, val->properties())->get_jint() & 1;
1703 } else {
1704 null_marker_value = 1;
1705 }
1706 obj->bool_field_put(offset + vk->null_marker_offset(), null_marker_value);
1707 }
1708 }
1709 }
1710
1711 // restore fields of all eliminated objects and arrays
1712 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci, TRAPS) {
1713 for (int i = 0; i < objects->length(); i++) {
1714 assert(objects->at(i)->is_object(), "invalid debug information");
1715 ObjectValue* sv = (ObjectValue*) objects->at(i);
1716 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1717 k = get_refined_array_klass(k, fr, reg_map, sv, THREAD);
1718
1719 Handle obj = sv->value();
1720 assert(obj.not_null() || realloc_failures || sv->has_properties(), "reallocation was missed");
1721 #ifndef PRODUCT
1722 if (PrintDeoptimizationDetails) {
1723 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1724 }
1725 #endif // !PRODUCT
1726
1727 if (obj.is_null()) {
1728 continue;
1729 }
1730
1731 #if INCLUDE_JVMCI
1732 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1733 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1734 continue;
1735 }
1736 #endif // INCLUDE_JVMCI
1737 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1738 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1739 ScopeValue* payload = sv->field_at(0);
1740 if (payload->is_location() &&
1741 payload->as_LocationValue()->location().type() == Location::vector) {
1742 #ifndef PRODUCT
1743 if (PrintDeoptimizationDetails) {
1744 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1745 if (Verbose) {
1746 Handle obj = sv->value();
1747 k->oop_print_on(obj(), tty);
1748 }
1749 }
1750 #endif // !PRODUCT
1751 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1752 }
1753 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1754 // which could be restored after vector object allocation.
1755 }
1756 if (k->is_instance_klass()) {
1757 InstanceKlass* ik = InstanceKlass::cast(k);
1758 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci, 0, CHECK);
1759 } else if (k->is_flatArray_klass()) {
1760 FlatArrayKlass* vak = FlatArrayKlass::cast(k);
1761 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, is_jvmci, CHECK);
1762 } else if (k->is_typeArray_klass()) {
1763 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1764 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1765 } else if (k->is_refArray_klass()) {
1766 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1767 }
1768 }
1769 // These objects may escape when we return to Interpreter after deoptimization.
1770 // We need barrier so that stores that initialize these objects can't be reordered
1771 // with subsequent stores that make these objects accessible by other threads.
1772 OrderAccess::storestore();
1773 }
1774
1775
1776 // relock objects for which synchronization was eliminated
1777 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1778 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1779 bool relocked_objects = false;
1780 for (int i = 0; i < monitors->length(); i++) {
1781 MonitorInfo* mon_info = monitors->at(i);
1782 if (mon_info->eliminated()) {
1783 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1784 relocked_objects = true;
1785 if (!mon_info->owner_is_scalar_replaced()) {
1786 Handle obj(thread, mon_info->owner());
1787 markWord mark = obj->mark();
1788 if (exec_mode == Unpack_none) {
1789 if (mark.has_monitor()) {
1790 // defer relocking if the deoptee thread is currently waiting for obj
1791 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1792 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1793 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1794 if (UseObjectMonitorTable) {
1795 mon_info->lock()->clear_object_monitor_cache();
1796 }
1797 #ifdef ASSERT
1798 else {
1799 assert(!UseObjectMonitorTable, "must be");
1800 mon_info->lock()->set_bad_monitor_deopt();
1801 }
1802 #endif
1803 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1804 continue;
1805 }
1806 }
1807 }
1808 BasicLock* lock = mon_info->lock();
1809 // We have lost information about the correct state of the lock stack.
1810 // Entering may create an invalid lock stack. Inflate the lock if it
1811 // was fast_locked to restore the valid lock stack.
1812 if (UseObjectMonitorTable) {
1813 // UseObjectMonitorTable expects the BasicLock cache to be either a
1814 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1815 // to nullptr.
1816 lock->clear_object_monitor_cache();
1817 }
1818 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1819 if (deoptee_thread->lock_stack().contains(obj())) {
1820 ObjectSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1821 deoptee_thread, thread);
1822 }
1823 assert(mon_info->owner()->is_locked(), "object must be locked now");
1824 assert(obj->mark().has_monitor(), "must be");
1825 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1826 assert(ObjectSynchronizer::read_monitor(obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1827 }
1828 }
1829 }
1830 return relocked_objects;
1831 }
1832 #endif // COMPILER2_OR_JVMCI
1833
1834 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1835 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1836
1837 // Register map for next frame (used for stack crawl). We capture
1838 // the state of the deopt'ing frame's caller. Thus if we need to
1839 // stuff a C2I adapter we can properly fill in the callee-save
1840 // register locations.
1841 frame caller = fr.sender(reg_map);
1842 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1843
1844 frame sender = caller;
1845
1846 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1847 // the vframeArray containing the unpacking information is allocated in the C heap.
1848 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1849 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1850
1851 // Compare the vframeArray to the collected vframes
1852 assert(array->structural_compare(thread, chunk), "just checking");
1853
1854 if (TraceDeoptimization) {
1855 ResourceMark rm;
1856 stringStream st;
1857 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1858 st.print(" ");
1859 fr.print_on(&st);
1860 st.print_cr(" Virtual frames (innermost/newest first):");
1861 for (int index = 0; index < chunk->length(); index++) {
1862 compiledVFrame* vf = chunk->at(index);
1863 int bci = vf->raw_bci();
1864 const char* code_name;
1865 if (bci == SynchronizationEntryBCI) {
1866 code_name = "sync entry";
1867 } else {
1868 Bytecodes::Code code = vf->method()->code_at(bci);
1869 code_name = Bytecodes::name(code);
1870 }
1871
1872 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1873 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1874 st.print(" - %s", code_name);
1875 st.print_cr(" @ bci=%d ", bci);
1876 }
1877 tty->print_raw(st.freeze());
1878 tty->cr();
1879 }
1880
1881 return array;
1882 }
1883
1884 #if COMPILER2_OR_JVMCI
1885 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1886 // Reallocation of some scalar replaced objects failed. Record
1887 // that we need to pop all the interpreter frames for the
1888 // deoptimized compiled frame.
1889 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1890 thread->set_frames_to_pop_failed_realloc(array->frames());
1891 // Unlock all monitors here otherwise the interpreter will see a
1892 // mix of locked and unlocked monitors (because of failed
1893 // reallocations of synchronized objects) and be confused.
1894 for (int i = 0; i < array->frames(); i++) {
1895 MonitorChunk* monitors = array->element(i)->monitors();
1896 if (monitors != nullptr) {
1897 // Unlock in reverse order starting from most nested monitor.
1898 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1899 BasicObjectLock* src = monitors->at(j);
1900 if (src->obj() != nullptr) {
1901 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1902 }
1903 }
1904 array->element(i)->free_monitors();
1905 #ifdef ASSERT
1906 array->element(i)->set_removed_monitors();
1907 #endif
1908 }
1909 }
1910 }
1911 #endif
1912
1913 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1914 assert(fr.can_be_deoptimized(), "checking frame type");
1915
1916 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1917
1918 if (LogCompilation && xtty != nullptr) {
1919 nmethod* nm = fr.cb()->as_nmethod_or_null();
1920 assert(nm != nullptr, "only compiled methods can deopt");
1921
1922 ttyLocker ttyl;
1923 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1924 nm->log_identity(xtty);
1925 xtty->end_head();
1926 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1927 xtty->begin_elem("jvms bci='%d'", sd->bci());
1928 xtty->method(sd->method());
1929 xtty->end_elem();
1930 if (sd->is_top()) break;
1931 }
1932 xtty->tail("deoptimized");
1933 }
1934
1935 Continuation::notify_deopt(thread, fr.sp());
1936
1937 // Patch the compiled method so that when execution returns to it we will
1938 // deopt the execution state and return to the interpreter.
1939 fr.deoptimize(thread);
1940 }
1941
1942 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1943 // Deoptimize only if the frame comes from compiled code.
1944 // Do not deoptimize the frame which is already patched
1945 // during the execution of the loops below.
1946 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1947 return;
1948 }
1949 ResourceMark rm;
1950 deoptimize_single_frame(thread, fr, reason);
1951 }
1952
1953 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm, bool make_not_entrant) {
1954 // there is no exception handler for this pc => deoptimize
1955 if (make_not_entrant) {
1956 nm->make_not_entrant(nmethod::InvalidationReason::MISSING_EXCEPTION_HANDLER);
1957 }
1958
1959 // Use Deoptimization::deoptimize for all of its side-effects:
1960 // gathering traps statistics, logging...
1961 // it also patches the return pc but we do not care about that
1962 // since we return a continuation to the deopt_blob below.
1963 JavaThread* thread = JavaThread::current();
1964 RegisterMap reg_map(thread,
1965 RegisterMap::UpdateMap::skip,
1966 RegisterMap::ProcessFrames::include,
1967 RegisterMap::WalkContinuation::skip);
1968 frame runtime_frame = thread->last_frame();
1969 frame caller_frame = runtime_frame.sender(®_map);
1970 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1971
1972 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1973
1974 if (!nm->is_compiled_by_jvmci()) {
1975 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1976 }
1977
1978 #if INCLUDE_JVMCI
1979 // JVMCI support
1980 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1981 compiledVFrame* cvf = compiledVFrame::cast(vf);
1982 ScopeDesc* imm_scope = cvf->scope();
1983 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1984 if (imm_mdo != nullptr) {
1985 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1986 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1987
1988 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1989 if (pdata != nullptr && pdata->is_BitData()) {
1990 BitData* bit_data = (BitData*) pdata;
1991 bit_data->set_exception_seen();
1992 }
1993 }
1994
1995
1996 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1997 if (trap_mdo != nullptr) {
1998 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1999 }
2000 #endif
2001
2002 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
2003 }
2004
2005 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
2006 assert(thread == Thread::current() ||
2007 thread->is_handshake_safe_for(Thread::current()) ||
2008 SafepointSynchronize::is_at_safepoint(),
2009 "can only deoptimize other thread at a safepoint/handshake");
2010 // Compute frame and register map based on thread and sp.
2011 RegisterMap reg_map(thread,
2012 RegisterMap::UpdateMap::skip,
2013 RegisterMap::ProcessFrames::include,
2014 RegisterMap::WalkContinuation::skip);
2015 frame fr = thread->last_frame();
2016 while (fr.id() != id) {
2017 fr = fr.sender(®_map);
2018 }
2019 deoptimize(thread, fr, reason);
2020 }
2021
2022
2023 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
2024 Thread* current = Thread::current();
2025 if (thread == current || thread->is_handshake_safe_for(current)) {
2026 Deoptimization::deoptimize_frame_internal(thread, id, reason);
2027 } else {
2028 VM_DeoptimizeFrame deopt(thread, id, reason);
2029 VMThread::execute(&deopt);
2030 }
2031 }
2032
2033 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
2034 deoptimize_frame(thread, id, Reason_constraint);
2035 }
2036
2037 // JVMTI PopFrame support
2038 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
2039 {
2040 assert(thread == JavaThread::current(), "pre-condition");
2041 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
2042 }
2043 JRT_END
2044
2045 MethodData*
2046 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
2047 bool create_if_missing) {
2048 JavaThread* THREAD = thread; // For exception macros.
2049 MethodData* mdo = m()->method_data();
2050 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
2051 // Build an MDO. Ignore errors like OutOfMemory;
2052 // that simply means we won't have an MDO to update.
2053 Method::build_profiling_method_data(m, THREAD);
2054 if (HAS_PENDING_EXCEPTION) {
2055 // Only metaspace OOM is expected. No Java code executed.
2056 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
2057 CLEAR_PENDING_EXCEPTION;
2058 }
2059 mdo = m()->method_data();
2060 }
2061 return mdo;
2062 }
2063
2064 #if COMPILER2_OR_JVMCI
2065 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
2066 // In case of an unresolved klass entry, load the class.
2067 // This path is exercised from case _ldc in Parse::do_one_bytecode,
2068 // and probably nowhere else.
2069 // Even that case would benefit from simply re-interpreting the
2070 // bytecode, without paying special attention to the class index.
2071 // So this whole "class index" feature should probably be removed.
2072
2073 if (constant_pool->tag_at(index).is_unresolved_klass()) {
2074 Klass* tk = constant_pool->klass_at(index, THREAD);
2075 if (HAS_PENDING_EXCEPTION) {
2076 // Exception happened during classloading. We ignore the exception here, since it
2077 // is going to be rethrown since the current activation is going to be deoptimized and
2078 // the interpreter will re-execute the bytecode.
2079 // Do not clear probable Async Exceptions.
2080 CLEAR_PENDING_NONASYNC_EXCEPTION;
2081 // Class loading called java code which may have caused a stack
2082 // overflow. If the exception was thrown right before the return
2083 // to the runtime the stack is no longer guarded. Reguard the
2084 // stack otherwise if we return to the uncommon trap blob and the
2085 // stack bang causes a stack overflow we crash.
2086 JavaThread* jt = THREAD;
2087 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
2088 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
2089 }
2090 return;
2091 }
2092
2093 assert(!constant_pool->tag_at(index).is_symbol(),
2094 "no symbolic names here, please");
2095 }
2096
2097 #if INCLUDE_JFR
2098
2099 class DeoptReasonSerializer : public JfrSerializer {
2100 public:
2101 void serialize(JfrCheckpointWriter& writer) {
2102 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
2103 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
2104 writer.write_key((u8)i);
2105 writer.write(Deoptimization::trap_reason_name(i));
2106 }
2107 }
2108 };
2109
2110 class DeoptActionSerializer : public JfrSerializer {
2111 public:
2112 void serialize(JfrCheckpointWriter& writer) {
2113 static const u4 nof_actions = Deoptimization::Action_LIMIT;
2114 writer.write_count(nof_actions);
2115 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
2116 writer.write_key(i);
2117 writer.write(Deoptimization::trap_action_name((int)i));
2118 }
2119 }
2120 };
2121
2122 static void register_serializers() {
2123 static int critical_section = 0;
2124 if (1 == critical_section || AtomicAccess::cmpxchg(&critical_section, 0, 1) == 1) {
2125 return;
2126 }
2127 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
2128 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
2129 }
2130
2131 static void post_deoptimization_event(nmethod* nm,
2132 const Method* method,
2133 int trap_bci,
2134 int instruction,
2135 Deoptimization::DeoptReason reason,
2136 Deoptimization::DeoptAction action) {
2137 assert(nm != nullptr, "invariant");
2138 assert(method != nullptr, "invariant");
2139 if (EventDeoptimization::is_enabled()) {
2140 static bool serializers_registered = false;
2141 if (!serializers_registered) {
2142 register_serializers();
2143 serializers_registered = true;
2144 }
2145 EventDeoptimization event;
2146 event.set_compileId(nm->compile_id());
2147 event.set_compiler(nm->compiler_type());
2148 event.set_method(method);
2149 event.set_lineNumber(method->line_number_from_bci(trap_bci));
2150 event.set_bci(trap_bci);
2151 event.set_instruction(instruction);
2152 event.set_reason(reason);
2153 event.set_action(action);
2154 event.commit();
2155 }
2156 }
2157
2158 #endif // INCLUDE_JFR
2159
2160 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2161 const char* reason_name, const char* reason_action, const char* class_name) {
2162 LogTarget(Debug, deoptimization) lt;
2163 if (lt.is_enabled()) {
2164 LogStream ls(lt);
2165 bool is_osr = nm->is_osr_method();
2166 ls.print("cid=%4d %s level=%d",
2167 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2168 ls.print(" %s", tm->name_and_sig_as_C_string());
2169 ls.print(" trap_bci=%d ", trap_bci);
2170 if (is_osr) {
2171 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2172 }
2173 ls.print("%s ", reason_name);
2174 ls.print("%s ", reason_action);
2175 if (class_name != nullptr) {
2176 ls.print("%s ", class_name);
2177 }
2178 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2179 pc, fr.pc() - nm->code_begin());
2180 }
2181 }
2182
2183 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2184 HandleMark hm(current);
2185
2186 // uncommon_trap() is called at the beginning of the uncommon trap
2187 // handler. Note this fact before we start generating temporary frames
2188 // that can confuse an asynchronous stack walker. This counter is
2189 // decremented at the end of unpack_frames().
2190
2191 current->inc_in_deopt_handler();
2192
2193 #if INCLUDE_JVMCI
2194 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2195 RegisterMap reg_map(current,
2196 RegisterMap::UpdateMap::include,
2197 RegisterMap::ProcessFrames::include,
2198 RegisterMap::WalkContinuation::skip);
2199 #else
2200 RegisterMap reg_map(current,
2201 RegisterMap::UpdateMap::skip,
2202 RegisterMap::ProcessFrames::include,
2203 RegisterMap::WalkContinuation::skip);
2204 #endif
2205 frame stub_frame = current->last_frame();
2206 frame fr = stub_frame.sender(®_map);
2207
2208 // Log a message
2209 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2210 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2211
2212 {
2213 ResourceMark rm;
2214
2215 DeoptReason reason = trap_request_reason(trap_request);
2216 DeoptAction action = trap_request_action(trap_request);
2217 #if INCLUDE_JVMCI
2218 int debug_id = trap_request_debug_id(trap_request);
2219 #endif
2220 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2221
2222 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2223 compiledVFrame* cvf = compiledVFrame::cast(vf);
2224
2225 nmethod* nm = cvf->code();
2226
2227 ScopeDesc* trap_scope = cvf->scope();
2228
2229 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2230
2231 if (is_receiver_constraint_failure) {
2232 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2233 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2234 JVMCI_ONLY(COMMA debug_id));
2235 }
2236
2237 methodHandle trap_method(current, trap_scope->method());
2238 int trap_bci = trap_scope->bci();
2239 #if INCLUDE_JVMCI
2240 jlong speculation = current->pending_failed_speculation();
2241 if (nm->is_compiled_by_jvmci()) {
2242 nm->update_speculation(current);
2243 } else {
2244 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2245 }
2246
2247 if (trap_bci == SynchronizationEntryBCI) {
2248 trap_bci = 0;
2249 current->set_pending_monitorenter(true);
2250 }
2251
2252 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2253 current->set_pending_transfer_to_interpreter(true);
2254 }
2255 #endif
2256
2257 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2258 // Record this event in the histogram.
2259 gather_statistics(reason, action, trap_bc);
2260
2261 // Ensure that we can record deopt. history:
2262 bool create_if_missing = ProfileTraps;
2263
2264 methodHandle profiled_method;
2265 #if INCLUDE_JVMCI
2266 if (nm->is_compiled_by_jvmci()) {
2267 profiled_method = methodHandle(current, nm->method());
2268 } else {
2269 profiled_method = trap_method;
2270 }
2271 #else
2272 profiled_method = trap_method;
2273 #endif
2274
2275 MethodData* trap_mdo =
2276 get_method_data(current, profiled_method, create_if_missing);
2277
2278 Symbol* class_name = nullptr;
2279 bool unresolved = false;
2280 if (unloaded_class_index >= 0) {
2281 constantPoolHandle constants (current, trap_method->constants());
2282 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2283 class_name = constants->klass_name_at(unloaded_class_index);
2284 unresolved = true;
2285 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2286 class_name = constants->symbol_at(unloaded_class_index);
2287 }
2288 }
2289 { // Log Deoptimization event for JFR, UL and event system
2290 Method* tm = trap_method();
2291 const char* reason_name = trap_reason_name(reason);
2292 const char* reason_action = trap_action_name(action);
2293 intptr_t pc = p2i(fr.pc());
2294
2295 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2296
2297 ResourceMark rm;
2298
2299 const char* class_name_str = nullptr;
2300 const char* class_name_msg = nullptr;
2301 stringStream st, stm;
2302 if (class_name != nullptr) {
2303 class_name->print_symbol_on(&st);
2304 class_name_str = st.freeze();
2305 stm.print("class=%s ", class_name_str);
2306 class_name_msg = stm.freeze();
2307 } else {
2308 class_name_msg = "";
2309 }
2310 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action, class_name_str);
2311 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s%s",
2312 reason_name, reason_action, pc,
2313 tm->name_and_sig_as_C_string(), trap_bci, class_name_msg, nm->compiler_name());
2314 }
2315
2316 // Print a bunch of diagnostics, if requested.
2317 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2318 ResourceMark rm;
2319
2320 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2321 // We must do this already now, since we cannot acquire this lock while
2322 // holding the tty lock (lock ordering by rank).
2323 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2324 (trap_mdo != nullptr),
2325 Mutex::_no_safepoint_check_flag);
2326
2327 ttyLocker ttyl;
2328
2329 char buf[100];
2330 if (xtty != nullptr) {
2331 xtty->begin_head("uncommon_trap thread='%zu' %s",
2332 os::current_thread_id(),
2333 format_trap_request(buf, sizeof(buf), trap_request));
2334 #if INCLUDE_JVMCI
2335 if (speculation != 0) {
2336 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2337 }
2338 #endif
2339 nm->log_identity(xtty);
2340 }
2341 if (class_name != nullptr) {
2342 if (xtty != nullptr) {
2343 if (unresolved) {
2344 xtty->print(" unresolved='1'");
2345 }
2346 xtty->name(class_name);
2347 }
2348 }
2349 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2350 // Dump the relevant MDO state.
2351 // This is the deopt count for the current reason, any previous
2352 // reasons or recompiles seen at this point.
2353 int dcnt = trap_mdo->trap_count(reason);
2354 if (dcnt != 0)
2355 xtty->print(" count='%d'", dcnt);
2356
2357 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2358 // lock extra_data_lock before the tty lock.
2359 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2360 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2361 if (dos != 0) {
2362 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2363 if (trap_state_is_recompiled(dos)) {
2364 int recnt2 = trap_mdo->overflow_recompile_count();
2365 if (recnt2 != 0)
2366 xtty->print(" recompiles2='%d'", recnt2);
2367 }
2368 }
2369 }
2370 if (xtty != nullptr) {
2371 xtty->stamp();
2372 xtty->end_head();
2373 }
2374 if (TraceDeoptimization) { // make noise on the tty
2375 stringStream st;
2376 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2377 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2378 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2379 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2380 #if INCLUDE_JVMCI
2381 if (nm->is_compiled_by_jvmci()) {
2382 const char* installed_code_name = nm->jvmci_name();
2383 if (installed_code_name != nullptr) {
2384 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2385 }
2386 }
2387 #endif
2388 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2389 p2i(fr.pc()),
2390 os::current_thread_id(),
2391 trap_reason_name(reason),
2392 trap_action_name(action),
2393 unloaded_class_index
2394 #if INCLUDE_JVMCI
2395 , debug_id
2396 #endif
2397 );
2398 if (class_name != nullptr) {
2399 st.print(unresolved ? " unresolved class: " : " symbol: ");
2400 class_name->print_symbol_on(&st);
2401 }
2402 st.cr();
2403 tty->print_raw(st.freeze());
2404 }
2405 if (xtty != nullptr) {
2406 // Log the precise location of the trap.
2407 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2408 xtty->begin_elem("jvms bci='%d'", sd->bci());
2409 xtty->method(sd->method());
2410 xtty->end_elem();
2411 if (sd->is_top()) break;
2412 }
2413 xtty->tail("uncommon_trap");
2414 }
2415 }
2416 // (End diagnostic printout.)
2417
2418 if (is_receiver_constraint_failure) {
2419 fatal("missing receiver type check");
2420 }
2421
2422 // Load class if necessary
2423 if (unloaded_class_index >= 0) {
2424 constantPoolHandle constants(current, trap_method->constants());
2425 load_class_by_index(constants, unloaded_class_index, THREAD);
2426 }
2427
2428 // Flush the nmethod if necessary and desirable.
2429 //
2430 // We need to avoid situations where we are re-flushing the nmethod
2431 // because of a hot deoptimization site. Repeated flushes at the same
2432 // point need to be detected by the compiler and avoided. If the compiler
2433 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2434 // module must take measures to avoid an infinite cycle of recompilation
2435 // and deoptimization. There are several such measures:
2436 //
2437 // 1. If a recompilation is ordered a second time at some site X
2438 // and for the same reason R, the action is adjusted to 'reinterpret',
2439 // to give the interpreter time to exercise the method more thoroughly.
2440 // If this happens, the method's overflow_recompile_count is incremented.
2441 //
2442 // 2. If the compiler fails to reduce the deoptimization rate, then
2443 // the method's overflow_recompile_count will begin to exceed the set
2444 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2445 // is adjusted to 'make_not_compilable', and the method is abandoned
2446 // to the interpreter. This is a performance hit for hot methods,
2447 // but is better than a disastrous infinite cycle of recompilations.
2448 // (Actually, only the method containing the site X is abandoned.)
2449 //
2450 // 3. In parallel with the previous measures, if the total number of
2451 // recompilations of a method exceeds the much larger set limit
2452 // PerMethodRecompilationCutoff, the method is abandoned.
2453 // This should only happen if the method is very large and has
2454 // many "lukewarm" deoptimizations. The code which enforces this
2455 // limit is elsewhere (class nmethod, class Method).
2456 //
2457 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2458 // to recompile at each bytecode independently of the per-BCI cutoff.
2459 //
2460 // The decision to update code is up to the compiler, and is encoded
2461 // in the Action_xxx code. If the compiler requests Action_none
2462 // no trap state is changed, no compiled code is changed, and the
2463 // computation suffers along in the interpreter.
2464 //
2465 // The other action codes specify various tactics for decompilation
2466 // and recompilation. Action_maybe_recompile is the loosest, and
2467 // allows the compiled code to stay around until enough traps are seen,
2468 // and until the compiler gets around to recompiling the trapping method.
2469 //
2470 // The other actions cause immediate removal of the present code.
2471
2472 // Traps caused by injected profile shouldn't pollute trap counts.
2473 bool injected_profile_trap = trap_method->has_injected_profile() &&
2474 (reason == Reason_intrinsic || reason == Reason_unreached);
2475
2476 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2477 bool make_not_entrant = false;
2478 bool make_not_compilable = false;
2479 bool reprofile = false;
2480 switch (action) {
2481 case Action_none:
2482 // Keep the old code.
2483 update_trap_state = false;
2484 break;
2485 case Action_maybe_recompile:
2486 // Do not need to invalidate the present code, but we can
2487 // initiate another
2488 // Start compiler without (necessarily) invalidating the nmethod.
2489 // The system will tolerate the old code, but new code should be
2490 // generated when possible.
2491 break;
2492 case Action_reinterpret:
2493 // Go back into the interpreter for a while, and then consider
2494 // recompiling form scratch.
2495 make_not_entrant = true;
2496 // Reset invocation counter for outer most method.
2497 // This will allow the interpreter to exercise the bytecodes
2498 // for a while before recompiling.
2499 // By contrast, Action_make_not_entrant is immediate.
2500 //
2501 // Note that the compiler will track null_check, null_assert,
2502 // range_check, and class_check events and log them as if they
2503 // had been traps taken from compiled code. This will update
2504 // the MDO trap history so that the next compilation will
2505 // properly detect hot trap sites.
2506 reprofile = true;
2507 break;
2508 case Action_make_not_entrant:
2509 // Request immediate recompilation, and get rid of the old code.
2510 // Make them not entrant, so next time they are called they get
2511 // recompiled. Unloaded classes are loaded now so recompile before next
2512 // time they are called. Same for uninitialized. The interpreter will
2513 // link the missing class, if any.
2514 make_not_entrant = true;
2515 break;
2516 case Action_make_not_compilable:
2517 // Give up on compiling this method at all.
2518 make_not_entrant = true;
2519 make_not_compilable = true;
2520 break;
2521 default:
2522 ShouldNotReachHere();
2523 }
2524
2525 #if INCLUDE_JVMCI
2526 // Deoptimization count is used by the CompileBroker to reason about compilations
2527 // it requests so do not pollute the count for deoptimizations in non-default (i.e.
2528 // non-CompilerBroker) compilations.
2529 if (nm->jvmci_skip_profile_deopt()) {
2530 update_trap_state = false;
2531 }
2532 #endif
2533 // Setting +ProfileTraps fixes the following, on all platforms:
2534 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2535 // recompile relies on a MethodData* to record heroic opt failures.
2536
2537 // Whether the interpreter is producing MDO data or not, we also need
2538 // to use the MDO to detect hot deoptimization points and control
2539 // aggressive optimization.
2540 bool inc_recompile_count = false;
2541
2542 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2543 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2544 (trap_mdo != nullptr),
2545 Mutex::_no_safepoint_check_flag);
2546 ProfileData* pdata = nullptr;
2547 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2548 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2549 uint this_trap_count = 0;
2550 bool maybe_prior_trap = false;
2551 bool maybe_prior_recompile = false;
2552
2553 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2554 #if INCLUDE_JVMCI
2555 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2556 #endif
2557 nm->method(),
2558 //outputs:
2559 this_trap_count,
2560 maybe_prior_trap,
2561 maybe_prior_recompile);
2562 // Because the interpreter also counts null, div0, range, and class
2563 // checks, these traps from compiled code are double-counted.
2564 // This is harmless; it just means that the PerXTrapLimit values
2565 // are in effect a little smaller than they look.
2566
2567 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2568 if (per_bc_reason != Reason_none) {
2569 // Now take action based on the partially known per-BCI history.
2570 if (maybe_prior_trap
2571 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2572 // If there are too many traps at this BCI, force a recompile.
2573 // This will allow the compiler to see the limit overflow, and
2574 // take corrective action, if possible. The compiler generally
2575 // does not use the exact PerBytecodeTrapLimit value, but instead
2576 // changes its tactics if it sees any traps at all. This provides
2577 // a little hysteresis, delaying a recompile until a trap happens
2578 // several times.
2579 //
2580 // Actually, since there is only one bit of counter per BCI,
2581 // the possible per-BCI counts are {0,1,(per-method count)}.
2582 // This produces accurate results if in fact there is only
2583 // one hot trap site, but begins to get fuzzy if there are
2584 // many sites. For example, if there are ten sites each
2585 // trapping two or more times, they each get the blame for
2586 // all of their traps.
2587 make_not_entrant = true;
2588 }
2589
2590 // Detect repeated recompilation at the same BCI, and enforce a limit.
2591 if (make_not_entrant && maybe_prior_recompile) {
2592 // More than one recompile at this point.
2593 inc_recompile_count = maybe_prior_trap;
2594 }
2595 } else {
2596 // For reasons which are not recorded per-bytecode, we simply
2597 // force recompiles unconditionally.
2598 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2599 make_not_entrant = true;
2600 }
2601
2602 // Go back to the compiler if there are too many traps in this method.
2603 if (this_trap_count >= per_method_trap_limit(reason)) {
2604 // If there are too many traps in this method, force a recompile.
2605 // This will allow the compiler to see the limit overflow, and
2606 // take corrective action, if possible.
2607 // (This condition is an unlikely backstop only, because the
2608 // PerBytecodeTrapLimit is more likely to take effect first,
2609 // if it is applicable.)
2610 make_not_entrant = true;
2611 }
2612
2613 // Here's more hysteresis: If there has been a recompile at
2614 // this trap point already, run the method in the interpreter
2615 // for a while to exercise it more thoroughly.
2616 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2617 reprofile = true;
2618 }
2619 }
2620
2621 // Take requested actions on the method:
2622
2623 // Recompile
2624 if (make_not_entrant) {
2625 if (!nm->make_not_entrant(nmethod::InvalidationReason::UNCOMMON_TRAP)) {
2626 return; // the call did not change nmethod's state
2627 }
2628
2629 if (pdata != nullptr) {
2630 // Record the recompilation event, if any.
2631 int tstate0 = pdata->trap_state();
2632 int tstate1 = trap_state_set_recompiled(tstate0, true);
2633 if (tstate1 != tstate0)
2634 pdata->set_trap_state(tstate1);
2635 }
2636
2637 // For code aging we count traps separately here, using make_not_entrant()
2638 // as a guard against simultaneous deopts in multiple threads.
2639 if (reason == Reason_tenured && trap_mdo != nullptr) {
2640 trap_mdo->inc_tenure_traps();
2641 }
2642 }
2643 if (inc_recompile_count) {
2644 trap_mdo->inc_overflow_recompile_count();
2645 if ((uint)trap_mdo->overflow_recompile_count() >
2646 (uint)PerBytecodeRecompilationCutoff) {
2647 // Give up on the method containing the bad BCI.
2648 if (trap_method() == nm->method()) {
2649 make_not_compilable = true;
2650 } else {
2651 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2652 // But give grace to the enclosing nm->method().
2653 }
2654 }
2655 }
2656
2657 // Reprofile
2658 if (reprofile) {
2659 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2660 }
2661
2662 // Give up compiling
2663 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2664 assert(make_not_entrant, "consistent");
2665 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2666 }
2667
2668 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2669 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2670 if (exception_handler_data != nullptr) {
2671 // uncommon trap at the start of an exception handler.
2672 // C2 generates these for un-entered exception handlers.
2673 // mark the handler as entered to avoid generating
2674 // another uncommon trap the next time the handler is compiled
2675 exception_handler_data->set_exception_handler_entered();
2676 }
2677 }
2678
2679 } // Free marked resources
2680
2681 }
2682 JRT_END
2683
2684 ProfileData*
2685 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2686 int trap_bci,
2687 Deoptimization::DeoptReason reason,
2688 bool update_total_trap_count,
2689 #if INCLUDE_JVMCI
2690 bool is_osr,
2691 #endif
2692 Method* compiled_method,
2693 //outputs:
2694 uint& ret_this_trap_count,
2695 bool& ret_maybe_prior_trap,
2696 bool& ret_maybe_prior_recompile) {
2697 trap_mdo->check_extra_data_locked();
2698
2699 bool maybe_prior_trap = false;
2700 bool maybe_prior_recompile = false;
2701 uint this_trap_count = 0;
2702 if (update_total_trap_count) {
2703 uint idx = reason;
2704 #if INCLUDE_JVMCI
2705 if (is_osr) {
2706 // Upper half of history array used for traps in OSR compilations
2707 idx += Reason_TRAP_HISTORY_LENGTH;
2708 }
2709 #endif
2710 uint prior_trap_count = trap_mdo->trap_count(idx);
2711 this_trap_count = trap_mdo->inc_trap_count(idx);
2712
2713 // If the runtime cannot find a place to store trap history,
2714 // it is estimated based on the general condition of the method.
2715 // If the method has ever been recompiled, or has ever incurred
2716 // a trap with the present reason , then this BCI is assumed
2717 // (pessimistically) to be the culprit.
2718 maybe_prior_trap = (prior_trap_count != 0);
2719 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2720 }
2721 ProfileData* pdata = nullptr;
2722
2723
2724 // For reasons which are recorded per bytecode, we check per-BCI data.
2725 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2726 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2727 if (per_bc_reason != Reason_none) {
2728 // Find the profile data for this BCI. If there isn't one,
2729 // try to allocate one from the MDO's set of spares.
2730 // This will let us detect a repeated trap at this point.
2731 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2732
2733 if (pdata != nullptr) {
2734 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2735 if (LogCompilation && xtty != nullptr) {
2736 ttyLocker ttyl;
2737 // no more room for speculative traps in this MDO
2738 xtty->elem("speculative_traps_oom");
2739 }
2740 }
2741 // Query the trap state of this profile datum.
2742 int tstate0 = pdata->trap_state();
2743 if (!trap_state_has_reason(tstate0, per_bc_reason))
2744 maybe_prior_trap = false;
2745 if (!trap_state_is_recompiled(tstate0))
2746 maybe_prior_recompile = false;
2747
2748 // Update the trap state of this profile datum.
2749 int tstate1 = tstate0;
2750 // Record the reason.
2751 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2752 // Store the updated state on the MDO, for next time.
2753 if (tstate1 != tstate0)
2754 pdata->set_trap_state(tstate1);
2755 } else {
2756 if (LogCompilation && xtty != nullptr) {
2757 ttyLocker ttyl;
2758 // Missing MDP? Leave a small complaint in the log.
2759 xtty->elem("missing_mdp bci='%d'", trap_bci);
2760 }
2761 }
2762 }
2763
2764 // Return results:
2765 ret_this_trap_count = this_trap_count;
2766 ret_maybe_prior_trap = maybe_prior_trap;
2767 ret_maybe_prior_recompile = maybe_prior_recompile;
2768 return pdata;
2769 }
2770
2771 void
2772 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2773 ResourceMark rm;
2774 // Ignored outputs:
2775 uint ignore_this_trap_count;
2776 bool ignore_maybe_prior_trap;
2777 bool ignore_maybe_prior_recompile;
2778 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2779 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2780 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2781
2782 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2783 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2784
2785 query_update_method_data(trap_mdo, trap_bci,
2786 (DeoptReason)reason,
2787 update_total_counts,
2788 #if INCLUDE_JVMCI
2789 false,
2790 #endif
2791 nullptr,
2792 ignore_this_trap_count,
2793 ignore_maybe_prior_trap,
2794 ignore_maybe_prior_recompile);
2795 }
2796
2797 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2798 // Enable WXWrite: current function is called from methods compiled by C2 directly
2799 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2800
2801 // Still in Java no safepoints
2802 {
2803 // This enters VM and may safepoint
2804 uncommon_trap_inner(current, trap_request);
2805 }
2806 HandleMark hm(current);
2807 return fetch_unroll_info_helper(current, exec_mode);
2808 }
2809
2810 // Local derived constants.
2811 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2812 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2813 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2814
2815 //---------------------------trap_state_reason---------------------------------
2816 Deoptimization::DeoptReason
2817 Deoptimization::trap_state_reason(int trap_state) {
2818 // This assert provides the link between the width of DataLayout::trap_bits
2819 // and the encoding of "recorded" reasons. It ensures there are enough
2820 // bits to store all needed reasons in the per-BCI MDO profile.
2821 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2822 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2823 trap_state -= recompile_bit;
2824 if (trap_state == DS_REASON_MASK) {
2825 return Reason_many;
2826 } else {
2827 assert((int)Reason_none == 0, "state=0 => Reason_none");
2828 return (DeoptReason)trap_state;
2829 }
2830 }
2831 //-------------------------trap_state_has_reason-------------------------------
2832 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2833 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2834 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2835 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2836 trap_state -= recompile_bit;
2837 if (trap_state == DS_REASON_MASK) {
2838 return -1; // true, unspecifically (bottom of state lattice)
2839 } else if (trap_state == reason) {
2840 return 1; // true, definitely
2841 } else if (trap_state == 0) {
2842 return 0; // false, definitely (top of state lattice)
2843 } else {
2844 return 0; // false, definitely
2845 }
2846 }
2847 //-------------------------trap_state_add_reason-------------------------------
2848 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2849 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2850 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2851 trap_state -= recompile_bit;
2852 if (trap_state == DS_REASON_MASK) {
2853 return trap_state + recompile_bit; // already at state lattice bottom
2854 } else if (trap_state == reason) {
2855 return trap_state + recompile_bit; // the condition is already true
2856 } else if (trap_state == 0) {
2857 return reason + recompile_bit; // no condition has yet been true
2858 } else {
2859 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2860 }
2861 }
2862 //-----------------------trap_state_is_recompiled------------------------------
2863 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2864 return (trap_state & DS_RECOMPILE_BIT) != 0;
2865 }
2866 //-----------------------trap_state_set_recompiled-----------------------------
2867 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2868 if (z) return trap_state | DS_RECOMPILE_BIT;
2869 else return trap_state & ~DS_RECOMPILE_BIT;
2870 }
2871 //---------------------------format_trap_state---------------------------------
2872 // This is used for debugging and diagnostics, including LogFile output.
2873 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2874 int trap_state) {
2875 assert(buflen > 0, "sanity");
2876 DeoptReason reason = trap_state_reason(trap_state);
2877 bool recomp_flag = trap_state_is_recompiled(trap_state);
2878 // Re-encode the state from its decoded components.
2879 int decoded_state = 0;
2880 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2881 decoded_state = trap_state_add_reason(decoded_state, reason);
2882 if (recomp_flag)
2883 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2884 // If the state re-encodes properly, format it symbolically.
2885 // Because this routine is used for debugging and diagnostics,
2886 // be robust even if the state is a strange value.
2887 size_t len;
2888 if (decoded_state != trap_state) {
2889 // Random buggy state that doesn't decode??
2890 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2891 } else {
2892 len = jio_snprintf(buf, buflen, "%s%s",
2893 trap_reason_name(reason),
2894 recomp_flag ? " recompiled" : "");
2895 }
2896 return buf;
2897 }
2898
2899
2900 //--------------------------------statics--------------------------------------
2901 const char* Deoptimization::_trap_reason_name[] = {
2902 // Note: Keep this in sync. with enum DeoptReason.
2903 "none",
2904 "null_check",
2905 "null_assert" JVMCI_ONLY("_or_unreached0"),
2906 "range_check",
2907 "class_check",
2908 "array_check",
2909 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2910 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2911 "profile_predicate",
2912 "auto_vectorization_check",
2913 "unloaded",
2914 "uninitialized",
2915 "initialized",
2916 "unreached",
2917 "unhandled",
2918 "constraint",
2919 "div0_check",
2920 "age",
2921 "predicate",
2922 "loop_limit_check",
2923 "speculate_class_check",
2924 "speculate_null_check",
2925 "speculate_null_assert",
2926 "unstable_if",
2927 "unstable_fused_if",
2928 "receiver_constraint",
2929 "not_compiled_exception_handler",
2930 "short_running_loop" JVMCI_ONLY("_or_aliasing"),
2931 #if INCLUDE_JVMCI
2932 "transfer_to_interpreter",
2933 "unresolved",
2934 "jsr_mismatch",
2935 #endif
2936 "tenured"
2937 };
2938 const char* Deoptimization::_trap_action_name[] = {
2939 // Note: Keep this in sync. with enum DeoptAction.
2940 "none",
2941 "maybe_recompile",
2942 "reinterpret",
2943 "make_not_entrant",
2944 "make_not_compilable"
2945 };
2946
2947 const char* Deoptimization::trap_reason_name(int reason) {
2948 // Check that every reason has a name
2949 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2950
2951 if (reason == Reason_many) return "many";
2952 if ((uint)reason < Reason_LIMIT)
2953 return _trap_reason_name[reason];
2954 static char buf[20];
2955 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2956 return buf;
2957 }
2958 const char* Deoptimization::trap_action_name(int action) {
2959 // Check that every action has a name
2960 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2961
2962 if ((uint)action < Action_LIMIT)
2963 return _trap_action_name[action];
2964 static char buf[20];
2965 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2966 return buf;
2967 }
2968
2969 // This is used for debugging and diagnostics, including LogFile output.
2970 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2971 int trap_request) {
2972 jint unloaded_class_index = trap_request_index(trap_request);
2973 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2974 const char* action = trap_action_name(trap_request_action(trap_request));
2975 #if INCLUDE_JVMCI
2976 int debug_id = trap_request_debug_id(trap_request);
2977 #endif
2978 size_t len;
2979 if (unloaded_class_index < 0) {
2980 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2981 reason, action
2982 #if INCLUDE_JVMCI
2983 ,debug_id
2984 #endif
2985 );
2986 } else {
2987 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2988 reason, action, unloaded_class_index
2989 #if INCLUDE_JVMCI
2990 ,debug_id
2991 #endif
2992 );
2993 }
2994 return buf;
2995 }
2996
2997 juint Deoptimization::_deoptimization_hist
2998 [Deoptimization::Reason_LIMIT]
2999 [1 + Deoptimization::Action_LIMIT]
3000 [Deoptimization::BC_CASE_LIMIT]
3001 = {0};
3002
3003 enum {
3004 LSB_BITS = 8,
3005 LSB_MASK = right_n_bits(LSB_BITS)
3006 };
3007
3008 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3009 Bytecodes::Code bc) {
3010 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
3011 assert(action >= 0 && action < Action_LIMIT, "oob");
3012 _deoptimization_hist[Reason_none][0][0] += 1; // total
3013 _deoptimization_hist[reason][0][0] += 1; // per-reason total
3014 juint* cases = _deoptimization_hist[reason][1+action];
3015 juint* bc_counter_addr = nullptr;
3016 juint bc_counter = 0;
3017 // Look for an unused counter, or an exact match to this BC.
3018 if (bc != Bytecodes::_illegal) {
3019 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3020 juint* counter_addr = &cases[bc_case];
3021 juint counter = *counter_addr;
3022 if ((counter == 0 && bc_counter_addr == nullptr)
3023 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
3024 // this counter is either free or is already devoted to this BC
3025 bc_counter_addr = counter_addr;
3026 bc_counter = counter | bc;
3027 }
3028 }
3029 }
3030 if (bc_counter_addr == nullptr) {
3031 // Overflow, or no given bytecode.
3032 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
3033 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
3034 }
3035 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
3036 }
3037
3038 jint Deoptimization::total_deoptimization_count() {
3039 return _deoptimization_hist[Reason_none][0][0];
3040 }
3041
3042 // Get the deopt count for a specific reason and a specific action. If either
3043 // one of 'reason' or 'action' is null, the method returns the sum of all
3044 // deoptimizations with the specific 'action' or 'reason' respectively.
3045 // If both arguments are null, the method returns the total deopt count.
3046 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3047 if (reason_str == nullptr && action_str == nullptr) {
3048 return total_deoptimization_count();
3049 }
3050 juint counter = 0;
3051 for (int reason = 0; reason < Reason_LIMIT; reason++) {
3052 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
3053 for (int action = 0; action < Action_LIMIT; action++) {
3054 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
3055 juint* cases = _deoptimization_hist[reason][1+action];
3056 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3057 counter += cases[bc_case] >> LSB_BITS;
3058 }
3059 }
3060 }
3061 }
3062 }
3063 return counter;
3064 }
3065
3066 void Deoptimization::print_statistics() {
3067 juint total = total_deoptimization_count();
3068 juint account = total;
3069 if (total != 0) {
3070 ttyLocker ttyl;
3071 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
3072 tty->print_cr("Deoptimization traps recorded:");
3073 #define PRINT_STAT_LINE(name, r) \
3074 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
3075 PRINT_STAT_LINE("total", total);
3076 // For each non-zero entry in the histogram, print the reason,
3077 // the action, and (if specifically known) the type of bytecode.
3078 for (int reason = 0; reason < Reason_LIMIT; reason++) {
3079 for (int action = 0; action < Action_LIMIT; action++) {
3080 juint* cases = _deoptimization_hist[reason][1+action];
3081 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3082 juint counter = cases[bc_case];
3083 if (counter != 0) {
3084 char name[1*K];
3085 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
3086 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
3087 trap_reason_name(reason),
3088 trap_action_name(action),
3089 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
3090 juint r = counter >> LSB_BITS;
3091 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
3092 account -= r;
3093 }
3094 }
3095 }
3096 }
3097 if (account != 0) {
3098 PRINT_STAT_LINE("unaccounted", account);
3099 }
3100 #undef PRINT_STAT_LINE
3101 if (xtty != nullptr) xtty->tail("statistics");
3102 }
3103 }
3104
3105 #else // COMPILER2_OR_JVMCI
3106
3107
3108 // Stubs for C1 only system.
3109 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3110 return false;
3111 }
3112
3113 const char* Deoptimization::trap_reason_name(int reason) {
3114 return "unknown";
3115 }
3116
3117 jint Deoptimization::total_deoptimization_count() {
3118 return 0;
3119 }
3120
3121 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3122 return 0;
3123 }
3124
3125 void Deoptimization::print_statistics() {
3126 // no output
3127 }
3128
3129 void
3130 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3131 // no update
3132 }
3133
3134 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3135 return 0;
3136 }
3137
3138 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3139 Bytecodes::Code bc) {
3140 // no update
3141 }
3142
3143 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3144 int trap_state) {
3145 jio_snprintf(buf, buflen, "#%d", trap_state);
3146 return buf;
3147 }
3148
3149 #endif // COMPILER2_OR_JVMCI