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