1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmClasses.hpp"
29 #include "code/codeCache.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/nmethod.hpp"
32 #include "code/pcDesc.hpp"
33 #include "code/scopeDesc.hpp"
34 #include "compiler/compilationPolicy.hpp"
35 #include "compiler/compilerDefinitions.inline.hpp"
36 #include "gc/shared/collectedHeap.hpp"
37 #include "gc/shared/memAllocator.hpp"
38 #include "interpreter/bytecode.hpp"
39 #include "interpreter/bytecode.inline.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/flatArrayKlass.hpp"
54 #include "oops/flatArrayOop.hpp"
55 #include "oops/fieldStreams.inline.hpp"
56 #include "oops/method.hpp"
57 #include "oops/objArrayKlass.hpp"
58 #include "oops/objArrayOop.inline.hpp"
59 #include "oops/oop.inline.hpp"
60 #include "oops/inlineKlass.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/atomic.hpp"
69 #include "runtime/basicLock.inline.hpp"
70 #include "runtime/continuation.hpp"
71 #include "runtime/continuationEntry.inline.hpp"
72 #include "runtime/deoptimization.hpp"
73 #include "runtime/escapeBarrier.hpp"
74 #include "runtime/fieldDescriptor.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/lightweightSynchronizer.hpp"
83 #include "runtime/lockStack.inline.hpp"
84 #include "runtime/objectMonitor.inline.hpp"
85 #include "runtime/osThread.hpp"
86 #include "runtime/safepointVerifiers.hpp"
87 #include "runtime/sharedRuntime.hpp"
88 #include "runtime/signature.hpp"
89 #include "runtime/stackFrameStream.inline.hpp"
90 #include "runtime/stackValue.hpp"
91 #include "runtime/stackWatermarkSet.hpp"
92 #include "runtime/stubRoutines.hpp"
93 #include "runtime/synchronizer.inline.hpp"
94 #include "runtime/threadSMR.hpp"
95 #include "runtime/threadWXSetters.inline.hpp"
96 #include "runtime/vframe.hpp"
97 #include "runtime/vframeArray.hpp"
98 #include "runtime/vframe_hp.hpp"
99 #include "runtime/vmOperations.hpp"
100 #include "utilities/checkedCast.hpp"
101 #include "utilities/events.hpp"
102 #include "utilities/growableArray.hpp"
103 #include "utilities/macros.hpp"
104 #include "utilities/preserveException.hpp"
105 #include "utilities/xmlstream.hpp"
106 #if INCLUDE_JFR
107 #include "jfr/jfr.inline.hpp"
108 #include "jfr/jfrEvents.hpp"
109 #include "jfr/metadata/jfrSerializer.hpp"
110 #endif
111
112 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
113 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
114 bool DeoptimizationScope::_committing_in_progress = false;
115
116 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
117 DEBUG_ONLY(_deopted = false;)
118
119 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
120 // If there is nothing to deopt _required_gen is the same as comitted.
121 _required_gen = DeoptimizationScope::_committed_deopt_gen;
122 }
123
124 DeoptimizationScope::~DeoptimizationScope() {
125 assert(_deopted, "Deopt not executed");
126 }
127
128 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
129 if (!nm->can_be_deoptimized()) {
130 return;
131 }
132
133 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
134
135 // If it's already marked but we still need it to be deopted.
136 if (nm->is_marked_for_deoptimization()) {
137 dependent(nm);
138 return;
139 }
140
141 nmethod::DeoptimizationStatus status =
142 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
143 Atomic::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_refArray_klass() && !k->is_flatArray_klass(), "Unexpected refined 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 ArrayKlass::ArrayProperties props = static_cast<ArrayKlass::ArrayProperties>(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(ArrayKlass::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 && 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 // Set the deoptee nmethod
552 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
553 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
554 current->set_deopt_compiled_method(nm);
555
556 if (VerifyStack) {
557 current->validate_frame_layout();
558 }
559
560 // Create a growable array of VFrames where each VFrame represents an inlined
561 // Java frame. This storage is allocated with the usual system arena.
562 assert(deoptee.is_compiled_frame(), "Wrong frame type");
563 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
564 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
565 while (!vf->is_top()) {
566 assert(vf->is_compiled_frame(), "Wrong frame type");
567 chunk->push(compiledVFrame::cast(vf));
568 vf = vf->sender();
569 }
570 assert(vf->is_compiled_frame(), "Wrong frame type");
571 chunk->push(compiledVFrame::cast(vf));
572
573 bool realloc_failures = false;
574
575 #if COMPILER2_OR_JVMCI
576 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
577
578 // Reallocate the non-escaping objects and restore their fields. Then
579 // relock objects if synchronization on them was eliminated.
580 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
581 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
582 bool unused;
583 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
584 }
585 #endif // COMPILER2_OR_JVMCI
586
587 // Ensure that no safepoint is taken after pointers have been stored
588 // in fields of rematerialized objects. If a safepoint occurs from here on
589 // out the java state residing in the vframeArray will be missed.
590 // Locks may be rebaised in a safepoint.
591 NoSafepointVerifier no_safepoint;
592
593 #if COMPILER2_OR_JVMCI
594 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
595 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
596 bool unused = false;
597 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
598 }
599 #endif // COMPILER2_OR_JVMCI
600
601 ScopeDesc* trap_scope = chunk->at(0)->scope();
602 Handle exceptionObject;
603 if (trap_scope->rethrow_exception()) {
604 #ifndef PRODUCT
605 if (PrintDeoptimizationDetails) {
606 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());
607 }
608 #endif // !PRODUCT
609
610 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
611 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
612 ScopeValue* topOfStack = expressions->top();
613 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
614 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
615 }
616
617 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
618 #if COMPILER2_OR_JVMCI
619 if (realloc_failures) {
620 // This destroys all ScopedValue bindings.
621 current->clear_scopedValueBindings();
622 pop_frames_failed_reallocs(current, array);
623 }
624 #endif
625
626 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
627 current->set_vframe_array_head(array);
628
629 // Now that the vframeArray has been created if we have any deferred local writes
630 // added by jvmti then we can free up that structure as the data is now in the
631 // vframeArray
632
633 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
634
635 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
636 CodeBlob* cb = stub_frame.cb();
637 // Verify we have the right vframeArray
638 assert(cb->frame_size() >= 0, "Unexpected frame size");
639 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
640
641 // If the deopt call site is a MethodHandle invoke call site we have
642 // to adjust the unpack_sp.
643 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
644 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
645 unpack_sp = deoptee.unextended_sp();
646
647 #ifdef ASSERT
648 assert(cb->is_deoptimization_stub() ||
649 cb->is_uncommon_trap_stub() ||
650 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
651 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
652 "unexpected code blob: %s", cb->name());
653 #endif
654
655 // This is a guarantee instead of an assert because if vframe doesn't match
656 // we will unpack the wrong deoptimized frame and wind up in strange places
657 // where it will be very difficult to figure out what went wrong. Better
658 // to die an early death here than some very obscure death later when the
659 // trail is cold.
660 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
661
662 int number_of_frames = array->frames();
663
664 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
665 // virtual activation, which is the reverse of the elements in the vframes array.
666 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
667 // +1 because we always have an interpreter return address for the final slot.
668 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
669 int popframe_extra_args = 0;
670 // Create an interpreter return address for the stub to use as its return
671 // address so the skeletal frames are perfectly walkable
672 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
673
674 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
675 // activation be put back on the expression stack of the caller for reexecution
676 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
677 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
678 }
679
680 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
681 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
682 // than simply use array->sender.pc(). This requires us to walk the current set of frames
683 //
684 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
685 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
686
687 // It's possible that the number of parameters at the call site is
688 // different than number of arguments in the callee when method
689 // handles are used. If the caller is interpreted get the real
690 // value so that the proper amount of space can be added to it's
691 // frame.
692 bool caller_was_method_handle = false;
693 if (deopt_sender.is_interpreted_frame()) {
694 methodHandle method(current, deopt_sender.interpreter_frame_method());
695 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
696 if (cur.has_member_arg()) {
697 // This should cover all real-world cases. One exception is a pathological chain of
698 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we
699 // would need to get the size from the resolved method entry. Another exception would
700 // be an invokedynamic with an adapter that is really a MethodHandle linker.
701 caller_was_method_handle = true;
702 }
703 }
704
705 //
706 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
707 // frame_sizes/frame_pcs[1] next oldest frame (int)
708 // frame_sizes/frame_pcs[n] youngest frame (int)
709 //
710 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
711 // owns the space for the return address to it's caller). Confusing ain't it.
712 //
713 // The vframe array can address vframes with indices running from
714 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
715 // When we create the skeletal frames we need the oldest frame to be in the zero slot
716 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
717 // so things look a little strange in this loop.
718 //
719 int callee_parameters = 0;
720 int callee_locals = 0;
721 for (int index = 0; index < array->frames(); index++ ) {
722 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
723 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
724 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
725 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
726 callee_locals,
727 index == 0,
728 popframe_extra_args);
729 // This pc doesn't have to be perfect just good enough to identify the frame
730 // as interpreted so the skeleton frame will be walkable
731 // The correct pc will be set when the skeleton frame is completely filled out
732 // The final pc we store in the loop is wrong and will be overwritten below
733 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
734
735 callee_parameters = array->element(index)->method()->size_of_parameters();
736 callee_locals = array->element(index)->method()->max_locals();
737 popframe_extra_args = 0;
738 }
739
740 // Compute whether the root vframe returns a float or double value.
741 BasicType return_type;
742 {
743 methodHandle method(current, array->element(0)->method());
744 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
745 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
746 }
747
748 // Compute information for handling adapters and adjusting the frame size of the caller.
749 int caller_adjustment = 0;
750
751 // Compute the amount the oldest interpreter frame will have to adjust
752 // its caller's stack by. If the caller is a compiled frame then
753 // we pretend that the callee has no parameters so that the
754 // extension counts for the full amount of locals and not just
755 // locals-parms. This is because without a c2i adapter the parm
756 // area as created by the compiled frame will not be usable by
757 // the interpreter. (Depending on the calling convention there
758 // may not even be enough space).
759
760 // QQQ I'd rather see this pushed down into last_frame_adjust
761 // and have it take the sender (aka caller).
762
763 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
764 caller_adjustment = last_frame_adjust(0, callee_locals);
765 } else if (callee_locals > callee_parameters) {
766 // The caller frame may need extending to accommodate
767 // non-parameter locals of the first unpacked interpreted frame.
768 // Compute that adjustment.
769 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
770 }
771
772 // If the sender is deoptimized we must retrieve the address of the handler
773 // since the frame will "magically" show the original pc before the deopt
774 // and we'd undo the deopt.
775
776 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
777 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
778 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
779 }
780
781 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
782
783 #if INCLUDE_JVMCI
784 if (exceptionObject() != nullptr) {
785 current->set_exception_oop(exceptionObject());
786 exec_mode = Unpack_exception;
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 // Return true if the execution after the provided bytecode continues at the
895 // next bytecode in the code. This is not the case for gotos, returns, and
896 // throws.
897 static bool falls_through(Bytecodes::Code bc) {
898 switch (bc) {
899 case Bytecodes::_goto:
900 case Bytecodes::_goto_w:
901 case Bytecodes::_athrow:
902 case Bytecodes::_areturn:
903 case Bytecodes::_dreturn:
904 case Bytecodes::_freturn:
905 case Bytecodes::_ireturn:
906 case Bytecodes::_lreturn:
907 case Bytecodes::_jsr:
908 case Bytecodes::_ret:
909 case Bytecodes::_return:
910 case Bytecodes::_lookupswitch:
911 case Bytecodes::_tableswitch:
912 return false;
913 default:
914 return true;
915 }
916 }
917 #endif
918
919 // Return BasicType of value being returned
920 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
921 assert(thread == JavaThread::current(), "pre-condition");
922
923 // We are already active in the special DeoptResourceMark any ResourceObj's we
924 // allocate will be freed at the end of the routine.
925
926 // JRT_LEAF methods don't normally allocate handles and there is a
927 // NoHandleMark to enforce that. It is actually safe to use Handles
928 // in a JRT_LEAF method, and sometimes desirable, but to do so we
929 // must use ResetNoHandleMark to bypass the NoHandleMark, and
930 // then use a HandleMark to ensure any Handles we do create are
931 // cleaned up in this scope.
932 ResetNoHandleMark rnhm;
933 HandleMark hm(thread);
934
935 frame stub_frame = thread->last_frame();
936
937 Continuation::notify_deopt(thread, stub_frame.sp());
938
939 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
940 // must point to the vframeArray for the unpack frame.
941 vframeArray* array = thread->vframe_array_head();
942 UnrollBlock* info = array->unroll_block();
943
944 // We set the last_Java frame. But the stack isn't really parsable here. So we
945 // clear it to make sure JFR understands not to try and walk stacks from events
946 // in here.
947 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
948 thread->frame_anchor()->set_last_Java_sp(nullptr);
949
950 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
951 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
952
953 thread->frame_anchor()->set_last_Java_sp(sp);
954
955 BasicType bt = info->return_type();
956
957 // If we have an exception pending, claim that the return type is an oop
958 // so the deopt_blob does not overwrite the exception_oop.
959
960 if (exec_mode == Unpack_exception)
961 bt = T_OBJECT;
962
963 // Cleanup thread deopt data
964 cleanup_deopt_info(thread, array);
965
966 #ifndef PRODUCT
967 if (VerifyStack) {
968 ResourceMark res_mark;
969 // Clear pending exception to not break verification code (restored afterwards)
970 PreserveExceptionMark pm(thread);
971
972 thread->validate_frame_layout();
973
974 // Verify that the just-unpacked frames match the interpreter's
975 // notions of expression stack and locals
976 vframeArray* cur_array = thread->vframe_array_last();
977 RegisterMap rm(thread,
978 RegisterMap::UpdateMap::skip,
979 RegisterMap::ProcessFrames::include,
980 RegisterMap::WalkContinuation::skip);
981 rm.set_include_argument_oops(false);
982 bool is_top_frame = true;
983 int callee_size_of_parameters = 0;
984 int callee_max_locals = 0;
985 for (int i = 0; i < cur_array->frames(); i++) {
986 vframeArrayElement* el = cur_array->element(i);
987 frame* iframe = el->iframe();
988 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
989
990 // Get the oop map for this bci
991 InterpreterOopMap mask;
992 int cur_invoke_parameter_size = 0;
993 bool try_next_mask = false;
994 int next_mask_expression_stack_size = -1;
995 int top_frame_expression_stack_adjustment = 0;
996 methodHandle mh(thread, iframe->interpreter_frame_method());
997 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
998 BytecodeStream str(mh, iframe->interpreter_frame_bci());
999 int max_bci = mh->code_size();
1000 // Get to the next bytecode if possible
1001 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
1002 // Check to see if we can grab the number of outgoing arguments
1003 // at an uncommon trap for an invoke (where the compiler
1004 // generates debug info before the invoke has executed)
1005 Bytecodes::Code cur_code = str.next();
1006 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
1007 if (Bytecodes::is_invoke(cur_code)) {
1008 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
1009 cur_invoke_parameter_size = invoke.size_of_parameters();
1010 if (i != 0 && invoke.has_member_arg()) {
1011 callee_size_of_parameters++;
1012 }
1013 }
1014 if (str.bci() < max_bci) {
1015 next_code = str.next();
1016 if (next_code >= 0) {
1017 // The interpreter oop map generator reports results before
1018 // the current bytecode has executed except in the case of
1019 // calls. It seems to be hard to tell whether the compiler
1020 // has emitted debug information matching the "state before"
1021 // a given bytecode or the state after, so we try both
1022 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
1023 // Get expression stack size for the next bytecode
1024 InterpreterOopMap next_mask;
1025 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
1026 next_mask_expression_stack_size = next_mask.expression_stack_size();
1027 if (Bytecodes::is_invoke(next_code)) {
1028 Bytecode_invoke invoke(mh, str.bci());
1029 next_mask_expression_stack_size += invoke.size_of_parameters();
1030 }
1031 // Need to subtract off the size of the result type of
1032 // the bytecode because this is not described in the
1033 // debug info but returned to the interpreter in the TOS
1034 // caching register
1035 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
1036 if (bytecode_result_type != T_ILLEGAL) {
1037 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
1038 }
1039 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
1040 try_next_mask = true;
1041 }
1042 }
1043 }
1044
1045 // Verify stack depth and oops in frame
1046 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
1047 if (!(
1048 /* SPARC */
1049 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
1050 /* x86 */
1051 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
1052 (try_next_mask &&
1053 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
1054 top_frame_expression_stack_adjustment))) ||
1055 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
1056 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
1057 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
1058 )) {
1059 {
1060 // Print out some information that will help us debug the problem
1061 tty->print_cr("Wrong number of expression stack elements during deoptimization");
1062 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
1063 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1064 if (try_next_mask) {
1065 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1066 }
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(" try_next_mask = %d", try_next_mask);
1071 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1072 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1073 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1074 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1075 tty->print_cr(" exec_mode = %d", exec_mode);
1076 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1077 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1078 tty->print_cr(" Interpreted frames:");
1079 for (int k = 0; k < cur_array->frames(); k++) {
1080 vframeArrayElement* el = cur_array->element(k);
1081 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1082 }
1083 cur_array->print_on_2(tty);
1084 }
1085 guarantee(false, "wrong number of expression stack elements during deopt");
1086 }
1087 VerifyOopClosure verify;
1088 iframe->oops_interpreted_do(&verify, &rm, false);
1089 callee_size_of_parameters = mh->size_of_parameters();
1090 callee_max_locals = mh->max_locals();
1091 is_top_frame = false;
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 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1150 if (ik->is_in_error_state()) {
1151 _low = 1;
1152 _high = 0;
1153 _cache = nullptr;
1154 } else {
1155 objArrayOop cache = CacheType::cache(ik);
1156 assert(cache->length() > 0, "Empty cache");
1157 _low = BoxType::value(cache->obj_at(0));
1158 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1159 _cache = JNIHandles::make_global(Handle(thread, cache));
1160 }
1161 }
1162 ~BoxCache() {
1163 JNIHandles::destroy_global(_cache);
1164 }
1165 public:
1166 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1167 if (_singleton == nullptr) {
1168 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1169 if (!Atomic::replace_if_null(&_singleton, s)) {
1170 delete s;
1171 }
1172 }
1173 return _singleton;
1174 }
1175 oop lookup(PrimitiveType value) {
1176 if (_low <= value && value <= _high) {
1177 int offset = checked_cast<int>(value - _low);
1178 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1179 }
1180 return nullptr;
1181 }
1182 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1183 if (_cache == nullptr) {
1184 cache_init_error = true;
1185 return nullptr;
1186 }
1187 // Have to cast to avoid little/big-endian problems.
1188 if (sizeof(PrimitiveType) > sizeof(jint)) {
1189 jlong value = (jlong)raw_value;
1190 return lookup(value);
1191 }
1192 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1193 return lookup(value);
1194 }
1195 };
1196
1197 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1198 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1199 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1200 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1201 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1202
1203 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1204 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1205 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1206 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1207 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1208
1209 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1210 jobject _true_cache;
1211 jobject _false_cache;
1212 protected:
1213 static BooleanBoxCache *_singleton;
1214 BooleanBoxCache(Thread *thread) {
1215 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1216 if (ik->is_in_error_state()) {
1217 _true_cache = nullptr;
1218 _false_cache = nullptr;
1219 } else {
1220 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1221 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1222 }
1223 }
1224 ~BooleanBoxCache() {
1225 JNIHandles::destroy_global(_true_cache);
1226 JNIHandles::destroy_global(_false_cache);
1227 }
1228 public:
1229 static BooleanBoxCache* singleton(Thread* thread) {
1230 if (_singleton == nullptr) {
1231 BooleanBoxCache* s = new BooleanBoxCache(thread);
1232 if (!Atomic::replace_if_null(&_singleton, s)) {
1233 delete s;
1234 }
1235 }
1236 return _singleton;
1237 }
1238 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1239 if (_true_cache == nullptr) {
1240 cache_in_error = true;
1241 return nullptr;
1242 }
1243 // Have to cast to avoid little/big-endian problems.
1244 jboolean value = (jboolean)*((jint*)&raw_value);
1245 return lookup(value);
1246 }
1247 oop lookup(jboolean value) {
1248 if (value != 0) {
1249 return JNIHandles::resolve_non_null(_true_cache);
1250 }
1251 return JNIHandles::resolve_non_null(_false_cache);
1252 }
1253 };
1254
1255 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1256
1257 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1258 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1259 BasicType box_type = vmClasses::box_klass_type(k);
1260 if (box_type != T_OBJECT) {
1261 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1262 switch(box_type) {
1263 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1264 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1265 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1266 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1267 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1268 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1269 default:;
1270 }
1271 }
1272 return nullptr;
1273 }
1274 #endif // INCLUDE_JVMCI
1275
1276 #if COMPILER2_OR_JVMCI
1277 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1278 Handle pending_exception(THREAD, thread->pending_exception());
1279 const char* exception_file = thread->exception_file();
1280 int exception_line = thread->exception_line();
1281 thread->clear_pending_exception();
1282
1283 bool failures = false;
1284
1285 for (int i = 0; i < objects->length(); i++) {
1286 assert(objects->at(i)->is_object(), "invalid debug information");
1287 ObjectValue* sv = (ObjectValue*) objects->at(i);
1288 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1289 k = get_refined_array_klass(k, fr, reg_map, sv, THREAD);
1290
1291 // Check if the object may be null and has an additional null_marker input that needs
1292 // to be checked before using the field values. Skip re-allocation if it is null.
1293 if (k->is_inline_klass() && sv->has_properties()) {
1294 jint null_marker = StackValue::create_stack_value(fr, reg_map, sv->properties())->get_jint();
1295 if (null_marker == 0) {
1296 continue;
1297 }
1298 }
1299
1300 oop obj = nullptr;
1301 bool cache_init_error = false;
1302 if (k->is_instance_klass()) {
1303 #if INCLUDE_JVMCI
1304 nmethod* nm = fr->cb()->as_nmethod_or_null();
1305 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1306 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1307 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1308 if (obj != nullptr) {
1309 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1310 abv->set_cached(true);
1311 } else if (cache_init_error) {
1312 // Results in an OOME which is valid (as opposed to a class initialization error)
1313 // and is fine for the rare case a cache initialization failing.
1314 failures = true;
1315 }
1316 }
1317 #endif // INCLUDE_JVMCI
1318
1319 InstanceKlass* ik = InstanceKlass::cast(k);
1320 if (obj == nullptr && !cache_init_error) {
1321 InternalOOMEMark iom(THREAD);
1322 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1323 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1324 } else {
1325 obj = ik->allocate_instance(THREAD);
1326 }
1327 }
1328 } else if (k->is_flatArray_klass()) {
1329 FlatArrayKlass* ak = FlatArrayKlass::cast(k);
1330 // Inline type array must be zeroed because not all memory is reassigned
1331 obj = ak->allocate_instance(sv->field_size(), ak->properties(), THREAD);
1332 } else if (k->is_typeArray_klass()) {
1333 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1334 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1335 int len = sv->field_size() / type2size[ak->element_type()];
1336 InternalOOMEMark iom(THREAD);
1337 obj = ak->allocate_instance(len, THREAD);
1338 } else if (k->is_refArray_klass()) {
1339 RefArrayKlass* ak = RefArrayKlass::cast(k);
1340 InternalOOMEMark iom(THREAD);
1341 obj = ak->allocate_instance(sv->field_size(), ak->properties(), THREAD);
1342 }
1343
1344 if (obj == nullptr) {
1345 failures = true;
1346 }
1347
1348 assert(sv->value().is_null(), "redundant reallocation");
1349 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1350 CLEAR_PENDING_EXCEPTION;
1351 sv->set_value(obj);
1352 }
1353
1354 if (failures) {
1355 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1356 } else if (pending_exception.not_null()) {
1357 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1358 }
1359
1360 return failures;
1361 }
1362
1363 // We're deoptimizing at the return of a call, inline type fields are
1364 // in registers. When we go back to the interpreter, it will expect a
1365 // reference to an inline type instance. Allocate and initialize it from
1366 // the register values here.
1367 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
1368 oop new_vt = vk->realloc_result(map, return_oops, THREAD);
1369 if (new_vt == nullptr) {
1370 CLEAR_PENDING_EXCEPTION;
1371 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
1372 }
1373 return_oops.clear();
1374 return_oops.push(Handle(THREAD, new_vt));
1375 return false;
1376 }
1377
1378 #if INCLUDE_JVMCI
1379 /**
1380 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1381 * we need to somehow be able to recover the actual kind to be able to write the correct
1382 * amount of bytes.
1383 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1384 * the entries at index n + 1 to n + i are 'markers'.
1385 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1386 * expected form of the array would be:
1387 *
1388 * {b0, b1, b2, b3, INT, marker, b6, b7}
1389 *
1390 * Thus, in order to get back the size of the entry, we simply need to count the number
1391 * of marked entries
1392 *
1393 * @param virtualArray the virtualized byte array
1394 * @param i index of the virtual entry we are recovering
1395 * @return The number of bytes the entry spans
1396 */
1397 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1398 int index = i;
1399 while (++index < virtualArray->field_size() &&
1400 virtualArray->field_at(index)->is_marker()) {}
1401 return index - i;
1402 }
1403
1404 /**
1405 * If there was a guarantee for byte array to always start aligned to a long, we could
1406 * do a simple check on the parity of the index. Unfortunately, that is not always the
1407 * case. Thus, we check alignment of the actual address we are writing to.
1408 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1409 * write a long to index 3.
1410 */
1411 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1412 jbyte* res = obj->byte_at_addr(index);
1413 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1414 return res;
1415 }
1416
1417 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1418 switch (byte_count) {
1419 case 1:
1420 obj->byte_at_put(index, (jbyte) value->get_jint());
1421 break;
1422 case 2:
1423 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1424 break;
1425 case 4:
1426 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1427 break;
1428 case 8:
1429 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1430 break;
1431 default:
1432 ShouldNotReachHere();
1433 }
1434 }
1435 #endif // INCLUDE_JVMCI
1436
1437
1438 // restore elements of an eliminated type array
1439 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1440 int index = 0;
1441
1442 for (int i = 0; i < sv->field_size(); i++) {
1443 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1444 switch(type) {
1445 case T_LONG: case T_DOUBLE: {
1446 assert(value->type() == T_INT, "Agreement.");
1447 StackValue* low =
1448 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1449 #ifdef _LP64
1450 jlong res = (jlong)low->get_intptr();
1451 #else
1452 jlong res = jlong_from(value->get_jint(), low->get_jint());
1453 #endif
1454 obj->long_at_put(index, res);
1455 break;
1456 }
1457
1458 case T_INT: case T_FLOAT: { // 4 bytes.
1459 assert(value->type() == T_INT, "Agreement.");
1460 bool big_value = false;
1461 if (i + 1 < sv->field_size() && type == T_INT) {
1462 if (sv->field_at(i)->is_location()) {
1463 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1464 if (type == Location::dbl || type == Location::lng) {
1465 big_value = true;
1466 }
1467 } else if (sv->field_at(i)->is_constant_int()) {
1468 ScopeValue* next_scope_field = sv->field_at(i + 1);
1469 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1470 big_value = true;
1471 }
1472 }
1473 }
1474
1475 if (big_value) {
1476 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1477 #ifdef _LP64
1478 jlong res = (jlong)low->get_intptr();
1479 #else
1480 jlong res = jlong_from(value->get_jint(), low->get_jint());
1481 #endif
1482 obj->int_at_put(index, *(jint*)&res);
1483 obj->int_at_put(++index, *((jint*)&res + 1));
1484 } else {
1485 obj->int_at_put(index, value->get_jint());
1486 }
1487 break;
1488 }
1489
1490 case T_SHORT:
1491 assert(value->type() == T_INT, "Agreement.");
1492 obj->short_at_put(index, (jshort)value->get_jint());
1493 break;
1494
1495 case T_CHAR:
1496 assert(value->type() == T_INT, "Agreement.");
1497 obj->char_at_put(index, (jchar)value->get_jint());
1498 break;
1499
1500 case T_BYTE: {
1501 assert(value->type() == T_INT, "Agreement.");
1502 #if INCLUDE_JVMCI
1503 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1504 int byte_count = count_number_of_bytes_for_entry(sv, i);
1505 byte_array_put(obj, value, index, byte_count);
1506 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1507 i += byte_count - 1; // Balance the loop counter.
1508 index += byte_count;
1509 // index has been updated so continue at top of loop
1510 continue;
1511 #else
1512 obj->byte_at_put(index, (jbyte)value->get_jint());
1513 break;
1514 #endif // INCLUDE_JVMCI
1515 }
1516
1517 case T_BOOLEAN: {
1518 assert(value->type() == T_INT, "Agreement.");
1519 obj->bool_at_put(index, (jboolean)value->get_jint());
1520 break;
1521 }
1522
1523 default:
1524 ShouldNotReachHere();
1525 }
1526 index++;
1527 }
1528 }
1529
1530 // restore fields of an eliminated object array
1531 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1532 for (int i = 0; i < sv->field_size(); i++) {
1533 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1534 assert(value->type() == T_OBJECT, "object element expected");
1535 obj->obj_at_put(i, value->get_obj()());
1536 }
1537 }
1538
1539 class ReassignedField {
1540 public:
1541 int _offset;
1542 BasicType _type;
1543 InstanceKlass* _klass;
1544 bool _is_flat;
1545 bool _is_null_free;
1546 public:
1547 ReassignedField() : _offset(0), _type(T_ILLEGAL), _klass(nullptr), _is_flat(false), _is_null_free(false) { }
1548 };
1549
1550 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1551 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1552 InstanceKlass* super = klass->superklass();
1553 if (super != nullptr) {
1554 get_reassigned_fields(super, fields, is_jvmci);
1555 }
1556 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1557 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1558 ReassignedField field;
1559 field._offset = fs.offset();
1560 field._type = Signature::basic_type(fs.signature());
1561 if (fs.is_flat()) {
1562 field._is_flat = true;
1563 field._is_null_free = fs.is_null_free_inline_type();
1564 // Resolve klass of flat inline type field
1565 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index()));
1566 }
1567 fields->append(field);
1568 }
1569 }
1570 return fields;
1571 }
1572
1573 // Restore fields of an eliminated instance object employing the same field order used by the
1574 // compiler when it scalarizes an object at safepoints.
1575 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) {
1576 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1577 for (int i = 0; i < fields->length(); i++) {
1578 BasicType type = fields->at(i)._type;
1579 int offset = base_offset + fields->at(i)._offset;
1580 // Check for flat inline type field before accessing the ScopeValue because it might not have any fields
1581 if (fields->at(i)._is_flat) {
1582 // Recursively re-assign flat inline type fields
1583 InstanceKlass* vk = fields->at(i)._klass;
1584 assert(vk != nullptr, "must be resolved");
1585 offset -= InlineKlass::cast(vk)->payload_offset(); // Adjust offset to omit oop header
1586 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, is_jvmci, offset, CHECK_0);
1587 if (!fields->at(i)._is_null_free) {
1588 ScopeValue* scope_field = sv->field_at(svIndex);
1589 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1590 int nm_offset = offset + InlineKlass::cast(vk)->null_marker_offset();
1591 obj->bool_field_put(nm_offset, value->get_jint() & 1);
1592 svIndex++;
1593 }
1594 continue; // Continue because we don't need to increment svIndex
1595 }
1596
1597 ScopeValue* scope_field = sv->field_at(svIndex);
1598 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1599 switch (type) {
1600 case T_OBJECT:
1601 case T_ARRAY:
1602 assert(value->type() == T_OBJECT, "Agreement.");
1603 obj->obj_field_put(offset, value->get_obj()());
1604 break;
1605
1606 case T_INT: case T_FLOAT: { // 4 bytes.
1607 assert(value->type() == T_INT, "Agreement.");
1608 bool big_value = false;
1609 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1610 if (scope_field->is_location()) {
1611 Location::Type type = ((LocationValue*) scope_field)->location().type();
1612 if (type == Location::dbl || type == Location::lng) {
1613 big_value = true;
1614 }
1615 }
1616 if (scope_field->is_constant_int()) {
1617 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1618 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1619 big_value = true;
1620 }
1621 }
1622 }
1623
1624 if (big_value) {
1625 i++;
1626 assert(i < fields->length(), "second T_INT field needed");
1627 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1628 } else {
1629 obj->int_field_put(offset, value->get_jint());
1630 break;
1631 }
1632 }
1633 /* no break */
1634
1635 case T_LONG: case T_DOUBLE: {
1636 assert(value->type() == T_INT, "Agreement.");
1637 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1638 #ifdef _LP64
1639 jlong res = (jlong)low->get_intptr();
1640 #else
1641 jlong res = jlong_from(value->get_jint(), low->get_jint());
1642 #endif
1643 obj->long_field_put(offset, res);
1644 break;
1645 }
1646
1647 case T_SHORT:
1648 assert(value->type() == T_INT, "Agreement.");
1649 obj->short_field_put(offset, (jshort)value->get_jint());
1650 break;
1651
1652 case T_CHAR:
1653 assert(value->type() == T_INT, "Agreement.");
1654 obj->char_field_put(offset, (jchar)value->get_jint());
1655 break;
1656
1657 case T_BYTE:
1658 assert(value->type() == T_INT, "Agreement.");
1659 obj->byte_field_put(offset, (jbyte)value->get_jint());
1660 break;
1661
1662 case T_BOOLEAN:
1663 assert(value->type() == T_INT, "Agreement.");
1664 obj->bool_field_put(offset, (jboolean)value->get_jint());
1665 break;
1666
1667 default:
1668 ShouldNotReachHere();
1669 }
1670 svIndex++;
1671 }
1672
1673 return svIndex;
1674 }
1675
1676 // restore fields of an eliminated inline type array
1677 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool is_jvmci, TRAPS) {
1678 InlineKlass* vk = vak->element_klass();
1679 assert(vk->maybe_flat_in_array(), "should only be used for flat inline type arrays");
1680 // Adjust offset to omit oop header
1681 int base_offset = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT) - InlineKlass::cast(vk)->payload_offset();
1682 // Initialize all elements of the flat inline type array
1683 for (int i = 0; i < sv->field_size(); i++) {
1684 ScopeValue* val = sv->field_at(i);
1685 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1686 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, is_jvmci, offset, CHECK);
1687 }
1688 }
1689
1690 // restore fields of all eliminated objects and arrays
1691 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci, TRAPS) {
1692 for (int i = 0; i < objects->length(); i++) {
1693 assert(objects->at(i)->is_object(), "invalid debug information");
1694 ObjectValue* sv = (ObjectValue*) objects->at(i);
1695 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1696 k = get_refined_array_klass(k, fr, reg_map, sv, THREAD);
1697
1698 Handle obj = sv->value();
1699 assert(obj.not_null() || realloc_failures || sv->has_properties(), "reallocation was missed");
1700 #ifndef PRODUCT
1701 if (PrintDeoptimizationDetails) {
1702 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1703 }
1704 #endif // !PRODUCT
1705
1706 if (obj.is_null()) {
1707 continue;
1708 }
1709
1710 #if INCLUDE_JVMCI
1711 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1712 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1713 continue;
1714 }
1715 #endif // INCLUDE_JVMCI
1716 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1717 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1718 ScopeValue* payload = sv->field_at(0);
1719 if (payload->is_location() &&
1720 payload->as_LocationValue()->location().type() == Location::vector) {
1721 #ifndef PRODUCT
1722 if (PrintDeoptimizationDetails) {
1723 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1724 if (Verbose) {
1725 Handle obj = sv->value();
1726 k->oop_print_on(obj(), tty);
1727 }
1728 }
1729 #endif // !PRODUCT
1730 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1731 }
1732 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1733 // which could be restored after vector object allocation.
1734 }
1735 if (k->is_instance_klass()) {
1736 InstanceKlass* ik = InstanceKlass::cast(k);
1737 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci, 0, CHECK);
1738 } else if (k->is_flatArray_klass()) {
1739 FlatArrayKlass* vak = FlatArrayKlass::cast(k);
1740 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, is_jvmci, CHECK);
1741 } else if (k->is_typeArray_klass()) {
1742 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1743 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1744 } else if (k->is_refArray_klass()) {
1745 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1746 }
1747 }
1748 // These objects may escape when we return to Interpreter after deoptimization.
1749 // We need barrier so that stores that initialize these objects can't be reordered
1750 // with subsequent stores that make these objects accessible by other threads.
1751 OrderAccess::storestore();
1752 }
1753
1754
1755 // relock objects for which synchronization was eliminated
1756 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1757 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1758 bool relocked_objects = false;
1759 for (int i = 0; i < monitors->length(); i++) {
1760 MonitorInfo* mon_info = monitors->at(i);
1761 if (mon_info->eliminated()) {
1762 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1763 relocked_objects = true;
1764 if (!mon_info->owner_is_scalar_replaced()) {
1765 Handle obj(thread, mon_info->owner());
1766 markWord mark = obj->mark();
1767 if (exec_mode == Unpack_none) {
1768 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1769 // With exec_mode == Unpack_none obj may be thread local and locked in
1770 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1771 markWord dmw = mark.displaced_mark_helper();
1772 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1773 obj->set_mark(dmw);
1774 }
1775 if (mark.has_monitor()) {
1776 // defer relocking if the deoptee thread is currently waiting for obj
1777 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1778 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1779 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1780 if (LockingMode == LM_LEGACY) {
1781 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1782 } else if (UseObjectMonitorTable) {
1783 mon_info->lock()->clear_object_monitor_cache();
1784 }
1785 #ifdef ASSERT
1786 else {
1787 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1788 mon_info->lock()->set_bad_metadata_deopt();
1789 }
1790 #endif
1791 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1792 continue;
1793 }
1794 }
1795 }
1796 BasicLock* lock = mon_info->lock();
1797 if (LockingMode == LM_LIGHTWEIGHT) {
1798 // We have lost information about the correct state of the lock stack.
1799 // Entering may create an invalid lock stack. Inflate the lock if it
1800 // was fast_locked to restore the valid lock stack.
1801 if (UseObjectMonitorTable) {
1802 // UseObjectMonitorTable expects the BasicLock cache to be either a
1803 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1804 // to nullptr.
1805 lock->clear_object_monitor_cache();
1806 }
1807 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1808 if (deoptee_thread->lock_stack().contains(obj())) {
1809 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1810 deoptee_thread, thread);
1811 }
1812 assert(mon_info->owner()->is_locked(), "object must be locked now");
1813 assert(obj->mark().has_monitor(), "must be");
1814 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1815 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1816 } else {
1817 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1818 assert(mon_info->owner()->is_locked(), "object must be locked now");
1819 }
1820 }
1821 }
1822 }
1823 return relocked_objects;
1824 }
1825 #endif // COMPILER2_OR_JVMCI
1826
1827 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1828 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1829
1830 // Register map for next frame (used for stack crawl). We capture
1831 // the state of the deopt'ing frame's caller. Thus if we need to
1832 // stuff a C2I adapter we can properly fill in the callee-save
1833 // register locations.
1834 frame caller = fr.sender(reg_map);
1835 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1836
1837 frame sender = caller;
1838
1839 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1840 // the vframeArray containing the unpacking information is allocated in the C heap.
1841 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1842 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1843
1844 // Compare the vframeArray to the collected vframes
1845 assert(array->structural_compare(thread, chunk), "just checking");
1846
1847 if (TraceDeoptimization) {
1848 ResourceMark rm;
1849 stringStream st;
1850 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1851 st.print(" ");
1852 fr.print_on(&st);
1853 st.print_cr(" Virtual frames (innermost/newest first):");
1854 for (int index = 0; index < chunk->length(); index++) {
1855 compiledVFrame* vf = chunk->at(index);
1856 int bci = vf->raw_bci();
1857 const char* code_name;
1858 if (bci == SynchronizationEntryBCI) {
1859 code_name = "sync entry";
1860 } else {
1861 Bytecodes::Code code = vf->method()->code_at(bci);
1862 code_name = Bytecodes::name(code);
1863 }
1864
1865 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1866 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1867 st.print(" - %s", code_name);
1868 st.print_cr(" @ bci=%d ", bci);
1869 }
1870 tty->print_raw(st.freeze());
1871 tty->cr();
1872 }
1873
1874 return array;
1875 }
1876
1877 #if COMPILER2_OR_JVMCI
1878 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1879 // Reallocation of some scalar replaced objects failed. Record
1880 // that we need to pop all the interpreter frames for the
1881 // deoptimized compiled frame.
1882 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1883 thread->set_frames_to_pop_failed_realloc(array->frames());
1884 // Unlock all monitors here otherwise the interpreter will see a
1885 // mix of locked and unlocked monitors (because of failed
1886 // reallocations of synchronized objects) and be confused.
1887 for (int i = 0; i < array->frames(); i++) {
1888 MonitorChunk* monitors = array->element(i)->monitors();
1889 if (monitors != nullptr) {
1890 // Unlock in reverse order starting from most nested monitor.
1891 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1892 BasicObjectLock* src = monitors->at(j);
1893 if (src->obj() != nullptr) {
1894 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1895 }
1896 }
1897 array->element(i)->free_monitors();
1898 #ifdef ASSERT
1899 array->element(i)->set_removed_monitors();
1900 #endif
1901 }
1902 }
1903 }
1904 #endif
1905
1906 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1907 assert(fr.can_be_deoptimized(), "checking frame type");
1908
1909 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1910
1911 if (LogCompilation && xtty != nullptr) {
1912 nmethod* nm = fr.cb()->as_nmethod_or_null();
1913 assert(nm != nullptr, "only compiled methods can deopt");
1914
1915 ttyLocker ttyl;
1916 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1917 nm->log_identity(xtty);
1918 xtty->end_head();
1919 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1920 xtty->begin_elem("jvms bci='%d'", sd->bci());
1921 xtty->method(sd->method());
1922 xtty->end_elem();
1923 if (sd->is_top()) break;
1924 }
1925 xtty->tail("deoptimized");
1926 }
1927
1928 Continuation::notify_deopt(thread, fr.sp());
1929
1930 // Patch the compiled method so that when execution returns to it we will
1931 // deopt the execution state and return to the interpreter.
1932 fr.deoptimize(thread);
1933 }
1934
1935 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1936 // Deoptimize only if the frame comes from compiled code.
1937 // Do not deoptimize the frame which is already patched
1938 // during the execution of the loops below.
1939 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1940 return;
1941 }
1942 ResourceMark rm;
1943 deoptimize_single_frame(thread, fr, reason);
1944 }
1945
1946 #if INCLUDE_JVMCI
1947 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1948 // there is no exception handler for this pc => deoptimize
1949 nm->make_not_entrant(nmethod::InvalidationReason::MISSING_EXCEPTION_HANDLER);
1950
1951 // Use Deoptimization::deoptimize for all of its side-effects:
1952 // gathering traps statistics, logging...
1953 // it also patches the return pc but we do not care about that
1954 // since we return a continuation to the deopt_blob below.
1955 JavaThread* thread = JavaThread::current();
1956 RegisterMap reg_map(thread,
1957 RegisterMap::UpdateMap::skip,
1958 RegisterMap::ProcessFrames::include,
1959 RegisterMap::WalkContinuation::skip);
1960 frame runtime_frame = thread->last_frame();
1961 frame caller_frame = runtime_frame.sender(®_map);
1962 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1963 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1964 compiledVFrame* cvf = compiledVFrame::cast(vf);
1965 ScopeDesc* imm_scope = cvf->scope();
1966 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1967 if (imm_mdo != nullptr) {
1968 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1969 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1970
1971 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1972 if (pdata != nullptr && pdata->is_BitData()) {
1973 BitData* bit_data = (BitData*) pdata;
1974 bit_data->set_exception_seen();
1975 }
1976 }
1977
1978 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1979
1980 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1981 if (trap_mdo != nullptr) {
1982 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1983 }
1984
1985 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1986 }
1987 #endif
1988
1989 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1990 assert(thread == Thread::current() ||
1991 thread->is_handshake_safe_for(Thread::current()) ||
1992 SafepointSynchronize::is_at_safepoint(),
1993 "can only deoptimize other thread at a safepoint/handshake");
1994 // Compute frame and register map based on thread and sp.
1995 RegisterMap reg_map(thread,
1996 RegisterMap::UpdateMap::skip,
1997 RegisterMap::ProcessFrames::include,
1998 RegisterMap::WalkContinuation::skip);
1999 frame fr = thread->last_frame();
2000 while (fr.id() != id) {
2001 fr = fr.sender(®_map);
2002 }
2003 deoptimize(thread, fr, reason);
2004 }
2005
2006
2007 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
2008 Thread* current = Thread::current();
2009 if (thread == current || thread->is_handshake_safe_for(current)) {
2010 Deoptimization::deoptimize_frame_internal(thread, id, reason);
2011 } else {
2012 VM_DeoptimizeFrame deopt(thread, id, reason);
2013 VMThread::execute(&deopt);
2014 }
2015 }
2016
2017 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
2018 deoptimize_frame(thread, id, Reason_constraint);
2019 }
2020
2021 // JVMTI PopFrame support
2022 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
2023 {
2024 assert(thread == JavaThread::current(), "pre-condition");
2025 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
2026 }
2027 JRT_END
2028
2029 MethodData*
2030 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
2031 bool create_if_missing) {
2032 JavaThread* THREAD = thread; // For exception macros.
2033 MethodData* mdo = m()->method_data();
2034 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
2035 // Build an MDO. Ignore errors like OutOfMemory;
2036 // that simply means we won't have an MDO to update.
2037 Method::build_profiling_method_data(m, THREAD);
2038 if (HAS_PENDING_EXCEPTION) {
2039 // Only metaspace OOM is expected. No Java code executed.
2040 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
2041 CLEAR_PENDING_EXCEPTION;
2042 }
2043 mdo = m()->method_data();
2044 }
2045 return mdo;
2046 }
2047
2048 #if COMPILER2_OR_JVMCI
2049 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
2050 // In case of an unresolved klass entry, load the class.
2051 // This path is exercised from case _ldc in Parse::do_one_bytecode,
2052 // and probably nowhere else.
2053 // Even that case would benefit from simply re-interpreting the
2054 // bytecode, without paying special attention to the class index.
2055 // So this whole "class index" feature should probably be removed.
2056
2057 if (constant_pool->tag_at(index).is_unresolved_klass()) {
2058 Klass* tk = constant_pool->klass_at(index, THREAD);
2059 if (HAS_PENDING_EXCEPTION) {
2060 // Exception happened during classloading. We ignore the exception here, since it
2061 // is going to be rethrown since the current activation is going to be deoptimized and
2062 // the interpreter will re-execute the bytecode.
2063 // Do not clear probable Async Exceptions.
2064 CLEAR_PENDING_NONASYNC_EXCEPTION;
2065 // Class loading called java code which may have caused a stack
2066 // overflow. If the exception was thrown right before the return
2067 // to the runtime the stack is no longer guarded. Reguard the
2068 // stack otherwise if we return to the uncommon trap blob and the
2069 // stack bang causes a stack overflow we crash.
2070 JavaThread* jt = THREAD;
2071 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
2072 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
2073 }
2074 return;
2075 }
2076
2077 assert(!constant_pool->tag_at(index).is_symbol(),
2078 "no symbolic names here, please");
2079 }
2080
2081 #if INCLUDE_JFR
2082
2083 class DeoptReasonSerializer : public JfrSerializer {
2084 public:
2085 void serialize(JfrCheckpointWriter& writer) {
2086 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
2087 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
2088 writer.write_key((u8)i);
2089 writer.write(Deoptimization::trap_reason_name(i));
2090 }
2091 }
2092 };
2093
2094 class DeoptActionSerializer : public JfrSerializer {
2095 public:
2096 void serialize(JfrCheckpointWriter& writer) {
2097 static const u4 nof_actions = Deoptimization::Action_LIMIT;
2098 writer.write_count(nof_actions);
2099 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
2100 writer.write_key(i);
2101 writer.write(Deoptimization::trap_action_name((int)i));
2102 }
2103 }
2104 };
2105
2106 static void register_serializers() {
2107 static int critical_section = 0;
2108 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
2109 return;
2110 }
2111 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
2112 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
2113 }
2114
2115 static void post_deoptimization_event(nmethod* nm,
2116 const Method* method,
2117 int trap_bci,
2118 int instruction,
2119 Deoptimization::DeoptReason reason,
2120 Deoptimization::DeoptAction action) {
2121 assert(nm != nullptr, "invariant");
2122 assert(method != nullptr, "invariant");
2123 if (EventDeoptimization::is_enabled()) {
2124 static bool serializers_registered = false;
2125 if (!serializers_registered) {
2126 register_serializers();
2127 serializers_registered = true;
2128 }
2129 EventDeoptimization event;
2130 event.set_compileId(nm->compile_id());
2131 event.set_compiler(nm->compiler_type());
2132 event.set_method(method);
2133 event.set_lineNumber(method->line_number_from_bci(trap_bci));
2134 event.set_bci(trap_bci);
2135 event.set_instruction(instruction);
2136 event.set_reason(reason);
2137 event.set_action(action);
2138 event.commit();
2139 }
2140 }
2141
2142 #endif // INCLUDE_JFR
2143
2144 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2145 const char* reason_name, const char* reason_action) {
2146 LogTarget(Debug, deoptimization) lt;
2147 if (lt.is_enabled()) {
2148 LogStream ls(lt);
2149 bool is_osr = nm->is_osr_method();
2150 ls.print("cid=%4d %s level=%d",
2151 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2152 ls.print(" %s", tm->name_and_sig_as_C_string());
2153 ls.print(" trap_bci=%d ", trap_bci);
2154 if (is_osr) {
2155 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2156 }
2157 ls.print("%s ", reason_name);
2158 ls.print("%s ", reason_action);
2159 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2160 pc, fr.pc() - nm->code_begin());
2161 }
2162 }
2163
2164 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2165 HandleMark hm(current);
2166
2167 // uncommon_trap() is called at the beginning of the uncommon trap
2168 // handler. Note this fact before we start generating temporary frames
2169 // that can confuse an asynchronous stack walker. This counter is
2170 // decremented at the end of unpack_frames().
2171
2172 current->inc_in_deopt_handler();
2173
2174 #if INCLUDE_JVMCI
2175 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2176 RegisterMap reg_map(current,
2177 RegisterMap::UpdateMap::include,
2178 RegisterMap::ProcessFrames::include,
2179 RegisterMap::WalkContinuation::skip);
2180 #else
2181 RegisterMap reg_map(current,
2182 RegisterMap::UpdateMap::skip,
2183 RegisterMap::ProcessFrames::include,
2184 RegisterMap::WalkContinuation::skip);
2185 #endif
2186 frame stub_frame = current->last_frame();
2187 frame fr = stub_frame.sender(®_map);
2188
2189 // Log a message
2190 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2191 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2192
2193 {
2194 ResourceMark rm;
2195
2196 DeoptReason reason = trap_request_reason(trap_request);
2197 DeoptAction action = trap_request_action(trap_request);
2198 #if INCLUDE_JVMCI
2199 int debug_id = trap_request_debug_id(trap_request);
2200 #endif
2201 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2202
2203 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2204 compiledVFrame* cvf = compiledVFrame::cast(vf);
2205
2206 nmethod* nm = cvf->code();
2207
2208 ScopeDesc* trap_scope = cvf->scope();
2209
2210 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2211
2212 if (is_receiver_constraint_failure) {
2213 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2214 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2215 JVMCI_ONLY(COMMA debug_id));
2216 }
2217
2218 methodHandle trap_method(current, trap_scope->method());
2219 int trap_bci = trap_scope->bci();
2220 #if INCLUDE_JVMCI
2221 jlong speculation = current->pending_failed_speculation();
2222 if (nm->is_compiled_by_jvmci()) {
2223 nm->update_speculation(current);
2224 } else {
2225 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2226 }
2227
2228 if (trap_bci == SynchronizationEntryBCI) {
2229 trap_bci = 0;
2230 current->set_pending_monitorenter(true);
2231 }
2232
2233 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2234 current->set_pending_transfer_to_interpreter(true);
2235 }
2236 #endif
2237
2238 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2239 // Record this event in the histogram.
2240 gather_statistics(reason, action, trap_bc);
2241
2242 // Ensure that we can record deopt. history:
2243 bool create_if_missing = ProfileTraps;
2244
2245 methodHandle profiled_method;
2246 #if INCLUDE_JVMCI
2247 if (nm->is_compiled_by_jvmci()) {
2248 profiled_method = methodHandle(current, nm->method());
2249 } else {
2250 profiled_method = trap_method;
2251 }
2252 #else
2253 profiled_method = trap_method;
2254 #endif
2255
2256 MethodData* trap_mdo =
2257 get_method_data(current, profiled_method, create_if_missing);
2258
2259 { // Log Deoptimization event for JFR, UL and event system
2260 Method* tm = trap_method();
2261 const char* reason_name = trap_reason_name(reason);
2262 const char* reason_action = trap_action_name(action);
2263 intptr_t pc = p2i(fr.pc());
2264
2265 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2266 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2267 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2268 reason_name, reason_action, pc,
2269 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2270 }
2271
2272 // Print a bunch of diagnostics, if requested.
2273 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2274 ResourceMark rm;
2275
2276 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2277 // We must do this already now, since we cannot acquire this lock while
2278 // holding the tty lock (lock ordering by rank).
2279 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2280
2281 ttyLocker ttyl;
2282
2283 char buf[100];
2284 if (xtty != nullptr) {
2285 xtty->begin_head("uncommon_trap thread='%zu' %s",
2286 os::current_thread_id(),
2287 format_trap_request(buf, sizeof(buf), trap_request));
2288 #if INCLUDE_JVMCI
2289 if (speculation != 0) {
2290 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2291 }
2292 #endif
2293 nm->log_identity(xtty);
2294 }
2295 Symbol* class_name = nullptr;
2296 bool unresolved = false;
2297 if (unloaded_class_index >= 0) {
2298 constantPoolHandle constants (current, trap_method->constants());
2299 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2300 class_name = constants->klass_name_at(unloaded_class_index);
2301 unresolved = true;
2302 if (xtty != nullptr)
2303 xtty->print(" unresolved='1'");
2304 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2305 class_name = constants->symbol_at(unloaded_class_index);
2306 }
2307 if (xtty != nullptr)
2308 xtty->name(class_name);
2309 }
2310 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2311 // Dump the relevant MDO state.
2312 // This is the deopt count for the current reason, any previous
2313 // reasons or recompiles seen at this point.
2314 int dcnt = trap_mdo->trap_count(reason);
2315 if (dcnt != 0)
2316 xtty->print(" count='%d'", dcnt);
2317
2318 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2319 // lock extra_data_lock before the tty lock.
2320 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2321 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2322 if (dos != 0) {
2323 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2324 if (trap_state_is_recompiled(dos)) {
2325 int recnt2 = trap_mdo->overflow_recompile_count();
2326 if (recnt2 != 0)
2327 xtty->print(" recompiles2='%d'", recnt2);
2328 }
2329 }
2330 }
2331 if (xtty != nullptr) {
2332 xtty->stamp();
2333 xtty->end_head();
2334 }
2335 if (TraceDeoptimization) { // make noise on the tty
2336 stringStream st;
2337 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2338 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2339 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2340 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2341 #if INCLUDE_JVMCI
2342 if (nm->is_compiled_by_jvmci()) {
2343 const char* installed_code_name = nm->jvmci_name();
2344 if (installed_code_name != nullptr) {
2345 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2346 }
2347 }
2348 #endif
2349 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2350 p2i(fr.pc()),
2351 os::current_thread_id(),
2352 trap_reason_name(reason),
2353 trap_action_name(action),
2354 unloaded_class_index
2355 #if INCLUDE_JVMCI
2356 , debug_id
2357 #endif
2358 );
2359 if (class_name != nullptr) {
2360 st.print(unresolved ? " unresolved class: " : " symbol: ");
2361 class_name->print_symbol_on(&st);
2362 }
2363 st.cr();
2364 tty->print_raw(st.freeze());
2365 }
2366 if (xtty != nullptr) {
2367 // Log the precise location of the trap.
2368 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2369 xtty->begin_elem("jvms bci='%d'", sd->bci());
2370 xtty->method(sd->method());
2371 xtty->end_elem();
2372 if (sd->is_top()) break;
2373 }
2374 xtty->tail("uncommon_trap");
2375 }
2376 }
2377 // (End diagnostic printout.)
2378
2379 if (is_receiver_constraint_failure) {
2380 fatal("missing receiver type check");
2381 }
2382
2383 // Load class if necessary
2384 if (unloaded_class_index >= 0) {
2385 constantPoolHandle constants(current, trap_method->constants());
2386 load_class_by_index(constants, unloaded_class_index, THREAD);
2387 }
2388
2389 // Flush the nmethod if necessary and desirable.
2390 //
2391 // We need to avoid situations where we are re-flushing the nmethod
2392 // because of a hot deoptimization site. Repeated flushes at the same
2393 // point need to be detected by the compiler and avoided. If the compiler
2394 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2395 // module must take measures to avoid an infinite cycle of recompilation
2396 // and deoptimization. There are several such measures:
2397 //
2398 // 1. If a recompilation is ordered a second time at some site X
2399 // and for the same reason R, the action is adjusted to 'reinterpret',
2400 // to give the interpreter time to exercise the method more thoroughly.
2401 // If this happens, the method's overflow_recompile_count is incremented.
2402 //
2403 // 2. If the compiler fails to reduce the deoptimization rate, then
2404 // the method's overflow_recompile_count will begin to exceed the set
2405 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2406 // is adjusted to 'make_not_compilable', and the method is abandoned
2407 // to the interpreter. This is a performance hit for hot methods,
2408 // but is better than a disastrous infinite cycle of recompilations.
2409 // (Actually, only the method containing the site X is abandoned.)
2410 //
2411 // 3. In parallel with the previous measures, if the total number of
2412 // recompilations of a method exceeds the much larger set limit
2413 // PerMethodRecompilationCutoff, the method is abandoned.
2414 // This should only happen if the method is very large and has
2415 // many "lukewarm" deoptimizations. The code which enforces this
2416 // limit is elsewhere (class nmethod, class Method).
2417 //
2418 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2419 // to recompile at each bytecode independently of the per-BCI cutoff.
2420 //
2421 // The decision to update code is up to the compiler, and is encoded
2422 // in the Action_xxx code. If the compiler requests Action_none
2423 // no trap state is changed, no compiled code is changed, and the
2424 // computation suffers along in the interpreter.
2425 //
2426 // The other action codes specify various tactics for decompilation
2427 // and recompilation. Action_maybe_recompile is the loosest, and
2428 // allows the compiled code to stay around until enough traps are seen,
2429 // and until the compiler gets around to recompiling the trapping method.
2430 //
2431 // The other actions cause immediate removal of the present code.
2432
2433 // Traps caused by injected profile shouldn't pollute trap counts.
2434 bool injected_profile_trap = trap_method->has_injected_profile() &&
2435 (reason == Reason_intrinsic || reason == Reason_unreached);
2436
2437 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2438 bool make_not_entrant = false;
2439 bool make_not_compilable = false;
2440 bool reprofile = false;
2441 switch (action) {
2442 case Action_none:
2443 // Keep the old code.
2444 update_trap_state = false;
2445 break;
2446 case Action_maybe_recompile:
2447 // Do not need to invalidate the present code, but we can
2448 // initiate another
2449 // Start compiler without (necessarily) invalidating the nmethod.
2450 // The system will tolerate the old code, but new code should be
2451 // generated when possible.
2452 break;
2453 case Action_reinterpret:
2454 // Go back into the interpreter for a while, and then consider
2455 // recompiling form scratch.
2456 make_not_entrant = true;
2457 // Reset invocation counter for outer most method.
2458 // This will allow the interpreter to exercise the bytecodes
2459 // for a while before recompiling.
2460 // By contrast, Action_make_not_entrant is immediate.
2461 //
2462 // Note that the compiler will track null_check, null_assert,
2463 // range_check, and class_check events and log them as if they
2464 // had been traps taken from compiled code. This will update
2465 // the MDO trap history so that the next compilation will
2466 // properly detect hot trap sites.
2467 reprofile = true;
2468 break;
2469 case Action_make_not_entrant:
2470 // Request immediate recompilation, and get rid of the old code.
2471 // Make them not entrant, so next time they are called they get
2472 // recompiled. Unloaded classes are loaded now so recompile before next
2473 // time they are called. Same for uninitialized. The interpreter will
2474 // link the missing class, if any.
2475 make_not_entrant = true;
2476 break;
2477 case Action_make_not_compilable:
2478 // Give up on compiling this method at all.
2479 make_not_entrant = true;
2480 make_not_compilable = true;
2481 break;
2482 default:
2483 ShouldNotReachHere();
2484 }
2485
2486 #if INCLUDE_JVMCI
2487 // Deoptimization count is used by the CompileBroker to reason about compilations
2488 // it requests so do not pollute the count for deoptimizations in non-default (i.e.
2489 // non-CompilerBroker) compilations.
2490 if (nm->is_jvmci_hosted()) {
2491 update_trap_state = false;
2492 }
2493 #endif
2494 // Setting +ProfileTraps fixes the following, on all platforms:
2495 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2496 // recompile relies on a MethodData* to record heroic opt failures.
2497
2498 // Whether the interpreter is producing MDO data or not, we also need
2499 // to use the MDO to detect hot deoptimization points and control
2500 // aggressive optimization.
2501 bool inc_recompile_count = false;
2502
2503 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2504 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2505 (trap_mdo != nullptr),
2506 Mutex::_no_safepoint_check_flag);
2507 ProfileData* pdata = nullptr;
2508 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2509 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2510 uint this_trap_count = 0;
2511 bool maybe_prior_trap = false;
2512 bool maybe_prior_recompile = false;
2513
2514 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2515 #if INCLUDE_JVMCI
2516 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2517 #endif
2518 nm->method(),
2519 //outputs:
2520 this_trap_count,
2521 maybe_prior_trap,
2522 maybe_prior_recompile);
2523 // Because the interpreter also counts null, div0, range, and class
2524 // checks, these traps from compiled code are double-counted.
2525 // This is harmless; it just means that the PerXTrapLimit values
2526 // are in effect a little smaller than they look.
2527
2528 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2529 if (per_bc_reason != Reason_none) {
2530 // Now take action based on the partially known per-BCI history.
2531 if (maybe_prior_trap
2532 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2533 // If there are too many traps at this BCI, force a recompile.
2534 // This will allow the compiler to see the limit overflow, and
2535 // take corrective action, if possible. The compiler generally
2536 // does not use the exact PerBytecodeTrapLimit value, but instead
2537 // changes its tactics if it sees any traps at all. This provides
2538 // a little hysteresis, delaying a recompile until a trap happens
2539 // several times.
2540 //
2541 // Actually, since there is only one bit of counter per BCI,
2542 // the possible per-BCI counts are {0,1,(per-method count)}.
2543 // This produces accurate results if in fact there is only
2544 // one hot trap site, but begins to get fuzzy if there are
2545 // many sites. For example, if there are ten sites each
2546 // trapping two or more times, they each get the blame for
2547 // all of their traps.
2548 make_not_entrant = true;
2549 }
2550
2551 // Detect repeated recompilation at the same BCI, and enforce a limit.
2552 if (make_not_entrant && maybe_prior_recompile) {
2553 // More than one recompile at this point.
2554 inc_recompile_count = maybe_prior_trap;
2555 }
2556 } else {
2557 // For reasons which are not recorded per-bytecode, we simply
2558 // force recompiles unconditionally.
2559 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2560 make_not_entrant = true;
2561 }
2562
2563 // Go back to the compiler if there are too many traps in this method.
2564 if (this_trap_count >= per_method_trap_limit(reason)) {
2565 // If there are too many traps in this method, force a recompile.
2566 // This will allow the compiler to see the limit overflow, and
2567 // take corrective action, if possible.
2568 // (This condition is an unlikely backstop only, because the
2569 // PerBytecodeTrapLimit is more likely to take effect first,
2570 // if it is applicable.)
2571 make_not_entrant = true;
2572 }
2573
2574 // Here's more hysteresis: If there has been a recompile at
2575 // this trap point already, run the method in the interpreter
2576 // for a while to exercise it more thoroughly.
2577 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2578 reprofile = true;
2579 }
2580 }
2581
2582 // Take requested actions on the method:
2583
2584 // Recompile
2585 if (make_not_entrant) {
2586 if (!nm->make_not_entrant(nmethod::InvalidationReason::UNCOMMON_TRAP)) {
2587 return; // the call did not change nmethod's state
2588 }
2589
2590 if (pdata != nullptr) {
2591 // Record the recompilation event, if any.
2592 int tstate0 = pdata->trap_state();
2593 int tstate1 = trap_state_set_recompiled(tstate0, true);
2594 if (tstate1 != tstate0)
2595 pdata->set_trap_state(tstate1);
2596 }
2597
2598 // For code aging we count traps separately here, using make_not_entrant()
2599 // as a guard against simultaneous deopts in multiple threads.
2600 if (reason == Reason_tenured && trap_mdo != nullptr) {
2601 trap_mdo->inc_tenure_traps();
2602 }
2603 }
2604 if (inc_recompile_count) {
2605 trap_mdo->inc_overflow_recompile_count();
2606 if ((uint)trap_mdo->overflow_recompile_count() >
2607 (uint)PerBytecodeRecompilationCutoff) {
2608 // Give up on the method containing the bad BCI.
2609 if (trap_method() == nm->method()) {
2610 make_not_compilable = true;
2611 } else {
2612 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2613 // But give grace to the enclosing nm->method().
2614 }
2615 }
2616 }
2617
2618 // Reprofile
2619 if (reprofile) {
2620 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2621 }
2622
2623 // Give up compiling
2624 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2625 assert(make_not_entrant, "consistent");
2626 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2627 }
2628
2629 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2630 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2631 if (exception_handler_data != nullptr) {
2632 // uncommon trap at the start of an exception handler.
2633 // C2 generates these for un-entered exception handlers.
2634 // mark the handler as entered to avoid generating
2635 // another uncommon trap the next time the handler is compiled
2636 exception_handler_data->set_exception_handler_entered();
2637 }
2638 }
2639
2640 } // Free marked resources
2641
2642 }
2643 JRT_END
2644
2645 ProfileData*
2646 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2647 int trap_bci,
2648 Deoptimization::DeoptReason reason,
2649 bool update_total_trap_count,
2650 #if INCLUDE_JVMCI
2651 bool is_osr,
2652 #endif
2653 Method* compiled_method,
2654 //outputs:
2655 uint& ret_this_trap_count,
2656 bool& ret_maybe_prior_trap,
2657 bool& ret_maybe_prior_recompile) {
2658 trap_mdo->check_extra_data_locked();
2659
2660 bool maybe_prior_trap = false;
2661 bool maybe_prior_recompile = false;
2662 uint this_trap_count = 0;
2663 if (update_total_trap_count) {
2664 uint idx = reason;
2665 #if INCLUDE_JVMCI
2666 if (is_osr) {
2667 // Upper half of history array used for traps in OSR compilations
2668 idx += Reason_TRAP_HISTORY_LENGTH;
2669 }
2670 #endif
2671 uint prior_trap_count = trap_mdo->trap_count(idx);
2672 this_trap_count = trap_mdo->inc_trap_count(idx);
2673
2674 // If the runtime cannot find a place to store trap history,
2675 // it is estimated based on the general condition of the method.
2676 // If the method has ever been recompiled, or has ever incurred
2677 // a trap with the present reason , then this BCI is assumed
2678 // (pessimistically) to be the culprit.
2679 maybe_prior_trap = (prior_trap_count != 0);
2680 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2681 }
2682 ProfileData* pdata = nullptr;
2683
2684
2685 // For reasons which are recorded per bytecode, we check per-BCI data.
2686 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2687 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2688 if (per_bc_reason != Reason_none) {
2689 // Find the profile data for this BCI. If there isn't one,
2690 // try to allocate one from the MDO's set of spares.
2691 // This will let us detect a repeated trap at this point.
2692 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2693
2694 if (pdata != nullptr) {
2695 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2696 if (LogCompilation && xtty != nullptr) {
2697 ttyLocker ttyl;
2698 // no more room for speculative traps in this MDO
2699 xtty->elem("speculative_traps_oom");
2700 }
2701 }
2702 // Query the trap state of this profile datum.
2703 int tstate0 = pdata->trap_state();
2704 if (!trap_state_has_reason(tstate0, per_bc_reason))
2705 maybe_prior_trap = false;
2706 if (!trap_state_is_recompiled(tstate0))
2707 maybe_prior_recompile = false;
2708
2709 // Update the trap state of this profile datum.
2710 int tstate1 = tstate0;
2711 // Record the reason.
2712 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2713 // Store the updated state on the MDO, for next time.
2714 if (tstate1 != tstate0)
2715 pdata->set_trap_state(tstate1);
2716 } else {
2717 if (LogCompilation && xtty != nullptr) {
2718 ttyLocker ttyl;
2719 // Missing MDP? Leave a small complaint in the log.
2720 xtty->elem("missing_mdp bci='%d'", trap_bci);
2721 }
2722 }
2723 }
2724
2725 // Return results:
2726 ret_this_trap_count = this_trap_count;
2727 ret_maybe_prior_trap = maybe_prior_trap;
2728 ret_maybe_prior_recompile = maybe_prior_recompile;
2729 return pdata;
2730 }
2731
2732 void
2733 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2734 ResourceMark rm;
2735 // Ignored outputs:
2736 uint ignore_this_trap_count;
2737 bool ignore_maybe_prior_trap;
2738 bool ignore_maybe_prior_recompile;
2739 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2740 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2741 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2742
2743 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2744 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2745
2746 query_update_method_data(trap_mdo, trap_bci,
2747 (DeoptReason)reason,
2748 update_total_counts,
2749 #if INCLUDE_JVMCI
2750 false,
2751 #endif
2752 nullptr,
2753 ignore_this_trap_count,
2754 ignore_maybe_prior_trap,
2755 ignore_maybe_prior_recompile);
2756 }
2757
2758 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2759 // Enable WXWrite: current function is called from methods compiled by C2 directly
2760 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2761
2762 // Still in Java no safepoints
2763 {
2764 // This enters VM and may safepoint
2765 uncommon_trap_inner(current, trap_request);
2766 }
2767 HandleMark hm(current);
2768 return fetch_unroll_info_helper(current, exec_mode);
2769 }
2770
2771 // Local derived constants.
2772 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2773 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2774 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2775
2776 //---------------------------trap_state_reason---------------------------------
2777 Deoptimization::DeoptReason
2778 Deoptimization::trap_state_reason(int trap_state) {
2779 // This assert provides the link between the width of DataLayout::trap_bits
2780 // and the encoding of "recorded" reasons. It ensures there are enough
2781 // bits to store all needed reasons in the per-BCI MDO profile.
2782 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2783 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2784 trap_state -= recompile_bit;
2785 if (trap_state == DS_REASON_MASK) {
2786 return Reason_many;
2787 } else {
2788 assert((int)Reason_none == 0, "state=0 => Reason_none");
2789 return (DeoptReason)trap_state;
2790 }
2791 }
2792 //-------------------------trap_state_has_reason-------------------------------
2793 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2794 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2795 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2796 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2797 trap_state -= recompile_bit;
2798 if (trap_state == DS_REASON_MASK) {
2799 return -1; // true, unspecifically (bottom of state lattice)
2800 } else if (trap_state == reason) {
2801 return 1; // true, definitely
2802 } else if (trap_state == 0) {
2803 return 0; // false, definitely (top of state lattice)
2804 } else {
2805 return 0; // false, definitely
2806 }
2807 }
2808 //-------------------------trap_state_add_reason-------------------------------
2809 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2810 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2811 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2812 trap_state -= recompile_bit;
2813 if (trap_state == DS_REASON_MASK) {
2814 return trap_state + recompile_bit; // already at state lattice bottom
2815 } else if (trap_state == reason) {
2816 return trap_state + recompile_bit; // the condition is already true
2817 } else if (trap_state == 0) {
2818 return reason + recompile_bit; // no condition has yet been true
2819 } else {
2820 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2821 }
2822 }
2823 //-----------------------trap_state_is_recompiled------------------------------
2824 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2825 return (trap_state & DS_RECOMPILE_BIT) != 0;
2826 }
2827 //-----------------------trap_state_set_recompiled-----------------------------
2828 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2829 if (z) return trap_state | DS_RECOMPILE_BIT;
2830 else return trap_state & ~DS_RECOMPILE_BIT;
2831 }
2832 //---------------------------format_trap_state---------------------------------
2833 // This is used for debugging and diagnostics, including LogFile output.
2834 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2835 int trap_state) {
2836 assert(buflen > 0, "sanity");
2837 DeoptReason reason = trap_state_reason(trap_state);
2838 bool recomp_flag = trap_state_is_recompiled(trap_state);
2839 // Re-encode the state from its decoded components.
2840 int decoded_state = 0;
2841 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2842 decoded_state = trap_state_add_reason(decoded_state, reason);
2843 if (recomp_flag)
2844 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2845 // If the state re-encodes properly, format it symbolically.
2846 // Because this routine is used for debugging and diagnostics,
2847 // be robust even if the state is a strange value.
2848 size_t len;
2849 if (decoded_state != trap_state) {
2850 // Random buggy state that doesn't decode??
2851 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2852 } else {
2853 len = jio_snprintf(buf, buflen, "%s%s",
2854 trap_reason_name(reason),
2855 recomp_flag ? " recompiled" : "");
2856 }
2857 return buf;
2858 }
2859
2860
2861 //--------------------------------statics--------------------------------------
2862 const char* Deoptimization::_trap_reason_name[] = {
2863 // Note: Keep this in sync. with enum DeoptReason.
2864 "none",
2865 "null_check",
2866 "null_assert" JVMCI_ONLY("_or_unreached0"),
2867 "range_check",
2868 "class_check",
2869 "array_check",
2870 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2871 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2872 "profile_predicate",
2873 "auto_vectorization_check",
2874 "unloaded",
2875 "uninitialized",
2876 "initialized",
2877 "unreached",
2878 "unhandled",
2879 "constraint",
2880 "div0_check",
2881 "age",
2882 "predicate",
2883 "loop_limit_check",
2884 "speculate_class_check",
2885 "speculate_null_check",
2886 "speculate_null_assert",
2887 "unstable_if",
2888 "unstable_fused_if",
2889 "receiver_constraint",
2890 #if INCLUDE_JVMCI
2891 "aliasing",
2892 "transfer_to_interpreter",
2893 "not_compiled_exception_handler",
2894 "unresolved",
2895 "jsr_mismatch",
2896 #endif
2897 "tenured"
2898 };
2899 const char* Deoptimization::_trap_action_name[] = {
2900 // Note: Keep this in sync. with enum DeoptAction.
2901 "none",
2902 "maybe_recompile",
2903 "reinterpret",
2904 "make_not_entrant",
2905 "make_not_compilable"
2906 };
2907
2908 const char* Deoptimization::trap_reason_name(int reason) {
2909 // Check that every reason has a name
2910 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2911
2912 if (reason == Reason_many) return "many";
2913 if ((uint)reason < Reason_LIMIT)
2914 return _trap_reason_name[reason];
2915 static char buf[20];
2916 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2917 return buf;
2918 }
2919 const char* Deoptimization::trap_action_name(int action) {
2920 // Check that every action has a name
2921 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2922
2923 if ((uint)action < Action_LIMIT)
2924 return _trap_action_name[action];
2925 static char buf[20];
2926 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2927 return buf;
2928 }
2929
2930 // This is used for debugging and diagnostics, including LogFile output.
2931 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2932 int trap_request) {
2933 jint unloaded_class_index = trap_request_index(trap_request);
2934 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2935 const char* action = trap_action_name(trap_request_action(trap_request));
2936 #if INCLUDE_JVMCI
2937 int debug_id = trap_request_debug_id(trap_request);
2938 #endif
2939 size_t len;
2940 if (unloaded_class_index < 0) {
2941 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2942 reason, action
2943 #if INCLUDE_JVMCI
2944 ,debug_id
2945 #endif
2946 );
2947 } else {
2948 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2949 reason, action, unloaded_class_index
2950 #if INCLUDE_JVMCI
2951 ,debug_id
2952 #endif
2953 );
2954 }
2955 return buf;
2956 }
2957
2958 juint Deoptimization::_deoptimization_hist
2959 [Deoptimization::Reason_LIMIT]
2960 [1 + Deoptimization::Action_LIMIT]
2961 [Deoptimization::BC_CASE_LIMIT]
2962 = {0};
2963
2964 enum {
2965 LSB_BITS = 8,
2966 LSB_MASK = right_n_bits(LSB_BITS)
2967 };
2968
2969 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2970 Bytecodes::Code bc) {
2971 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2972 assert(action >= 0 && action < Action_LIMIT, "oob");
2973 _deoptimization_hist[Reason_none][0][0] += 1; // total
2974 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2975 juint* cases = _deoptimization_hist[reason][1+action];
2976 juint* bc_counter_addr = nullptr;
2977 juint bc_counter = 0;
2978 // Look for an unused counter, or an exact match to this BC.
2979 if (bc != Bytecodes::_illegal) {
2980 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2981 juint* counter_addr = &cases[bc_case];
2982 juint counter = *counter_addr;
2983 if ((counter == 0 && bc_counter_addr == nullptr)
2984 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2985 // this counter is either free or is already devoted to this BC
2986 bc_counter_addr = counter_addr;
2987 bc_counter = counter | bc;
2988 }
2989 }
2990 }
2991 if (bc_counter_addr == nullptr) {
2992 // Overflow, or no given bytecode.
2993 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2994 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2995 }
2996 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2997 }
2998
2999 jint Deoptimization::total_deoptimization_count() {
3000 return _deoptimization_hist[Reason_none][0][0];
3001 }
3002
3003 // Get the deopt count for a specific reason and a specific action. If either
3004 // one of 'reason' or 'action' is null, the method returns the sum of all
3005 // deoptimizations with the specific 'action' or 'reason' respectively.
3006 // If both arguments are null, the method returns the total deopt count.
3007 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3008 if (reason_str == nullptr && action_str == nullptr) {
3009 return total_deoptimization_count();
3010 }
3011 juint counter = 0;
3012 for (int reason = 0; reason < Reason_LIMIT; reason++) {
3013 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
3014 for (int action = 0; action < Action_LIMIT; action++) {
3015 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
3016 juint* cases = _deoptimization_hist[reason][1+action];
3017 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3018 counter += cases[bc_case] >> LSB_BITS;
3019 }
3020 }
3021 }
3022 }
3023 }
3024 return counter;
3025 }
3026
3027 void Deoptimization::print_statistics() {
3028 juint total = total_deoptimization_count();
3029 juint account = total;
3030 if (total != 0) {
3031 ttyLocker ttyl;
3032 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
3033 tty->print_cr("Deoptimization traps recorded:");
3034 #define PRINT_STAT_LINE(name, r) \
3035 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
3036 PRINT_STAT_LINE("total", total);
3037 // For each non-zero entry in the histogram, print the reason,
3038 // the action, and (if specifically known) the type of bytecode.
3039 for (int reason = 0; reason < Reason_LIMIT; reason++) {
3040 for (int action = 0; action < Action_LIMIT; action++) {
3041 juint* cases = _deoptimization_hist[reason][1+action];
3042 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3043 juint counter = cases[bc_case];
3044 if (counter != 0) {
3045 char name[1*K];
3046 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
3047 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
3048 bc = Bytecodes::_illegal;
3049 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
3050 trap_reason_name(reason),
3051 trap_action_name(action),
3052 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
3053 juint r = counter >> LSB_BITS;
3054 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
3055 account -= r;
3056 }
3057 }
3058 }
3059 }
3060 if (account != 0) {
3061 PRINT_STAT_LINE("unaccounted", account);
3062 }
3063 #undef PRINT_STAT_LINE
3064 if (xtty != nullptr) xtty->tail("statistics");
3065 }
3066 }
3067
3068 #else // COMPILER2_OR_JVMCI
3069
3070
3071 // Stubs for C1 only system.
3072 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3073 return false;
3074 }
3075
3076 const char* Deoptimization::trap_reason_name(int reason) {
3077 return "unknown";
3078 }
3079
3080 jint Deoptimization::total_deoptimization_count() {
3081 return 0;
3082 }
3083
3084 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3085 return 0;
3086 }
3087
3088 void Deoptimization::print_statistics() {
3089 // no output
3090 }
3091
3092 void
3093 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3094 // no update
3095 }
3096
3097 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3098 return 0;
3099 }
3100
3101 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3102 Bytecodes::Code bc) {
3103 // no update
3104 }
3105
3106 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3107 int trap_state) {
3108 jio_snprintf(buf, buflen, "#%d", trap_state);
3109 return buf;
3110 }
3111
3112 #endif // COMPILER2_OR_JVMCI