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/jfrEvents.hpp"
108 #include "jfr/metadata/jfrSerializer.hpp"
109 #endif
110
111 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
112 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
113 bool DeoptimizationScope::_committing_in_progress = false;
114
115 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
116 DEBUG_ONLY(_deopted = false;)
117
118 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
119 // If there is nothing to deopt _required_gen is the same as comitted.
120 _required_gen = DeoptimizationScope::_committed_deopt_gen;
121 }
122
123 DeoptimizationScope::~DeoptimizationScope() {
124 assert(_deopted, "Deopt not executed");
125 }
126
127 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
128 if (!nm->can_be_deoptimized()) {
129 return;
130 }
131
132 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
133
134 // If it's already marked but we still need it to be deopted.
135 if (nm->is_marked_for_deoptimization()) {
136 dependent(nm);
137 return;
138 }
139
140 nmethod::DeoptimizationStatus status =
141 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
142 Atomic::store(&nm->_deoptimization_status, status);
143
144 // Make sure active is not committed
145 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
146 assert(nm->_deoptimization_generation == 0, "Is already marked");
147
148 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
149 _required_gen = DeoptimizationScope::_active_deopt_gen;
150 }
151
152 void DeoptimizationScope::dependent(nmethod* nm) {
153 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
154
155 // A method marked by someone else may have a _required_gen lower than what we marked with.
156 // Therefore only store it if it's higher than _required_gen.
157 if (_required_gen < nm->_deoptimization_generation) {
158 _required_gen = nm->_deoptimization_generation;
159 }
160 }
161
162 void DeoptimizationScope::deoptimize_marked() {
163 assert(!_deopted, "Already deopted");
164
165 // We are not alive yet.
166 if (!Universe::is_fully_initialized()) {
167 DEBUG_ONLY(_deopted = true;)
168 return;
169 }
170
171 // Safepoints are a special case, handled here.
172 if (SafepointSynchronize::is_at_safepoint()) {
173 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
174 DeoptimizationScope::_active_deopt_gen++;
175 Deoptimization::deoptimize_all_marked();
176 DEBUG_ONLY(_deopted = true;)
177 return;
178 }
179
180 uint64_t comitting = 0;
181 bool wait = false;
182 while (true) {
183 {
184 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
185
186 // First we check if we or someone else already deopted the gen we want.
187 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
188 DEBUG_ONLY(_deopted = true;)
189 return;
190 }
191 if (!_committing_in_progress) {
192 // The version we are about to commit.
193 comitting = DeoptimizationScope::_active_deopt_gen;
194 // Make sure new marks use a higher gen.
195 DeoptimizationScope::_active_deopt_gen++;
196 _committing_in_progress = true;
197 wait = false;
198 } else {
199 // Another thread is handshaking and committing a gen.
200 wait = true;
201 }
202 }
203 if (wait) {
204 // Wait and let the concurrent handshake be performed.
205 ThreadBlockInVM tbivm(JavaThread::current());
206 os::naked_yield();
207 } else {
208 // Performs the handshake.
209 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
210 DEBUG_ONLY(_deopted = true;)
211 {
212 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
213
214 // Make sure that committed doesn't go backwards.
215 // Should only happen if we did a deopt during a safepoint above.
216 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
217 DeoptimizationScope::_committed_deopt_gen = comitting;
218 }
219 _committing_in_progress = false;
220
221 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
222
223 return;
224 }
225 }
226 }
227 }
228
229 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
230 int caller_adjustment,
231 int caller_actual_parameters,
232 int number_of_frames,
233 intptr_t* frame_sizes,
234 address* frame_pcs,
235 BasicType return_type,
236 int exec_mode) {
237 _size_of_deoptimized_frame = size_of_deoptimized_frame;
238 _caller_adjustment = caller_adjustment;
239 _caller_actual_parameters = caller_actual_parameters;
240 _number_of_frames = number_of_frames;
241 _frame_sizes = frame_sizes;
242 _frame_pcs = frame_pcs;
243 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
244 _return_type = return_type;
245 _initial_info = 0;
246 // PD (x86 only)
247 _counter_temp = 0;
248 _unpack_kind = exec_mode;
249 _sender_sp_temp = 0;
250
251 _total_frame_sizes = size_of_frames();
252 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
253 }
254
255 Deoptimization::UnrollBlock::~UnrollBlock() {
256 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
257 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
258 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
259 }
260
261 int Deoptimization::UnrollBlock::size_of_frames() const {
262 // Account first for the adjustment of the initial frame
263 intptr_t result = _caller_adjustment;
264 for (int index = 0; index < number_of_frames(); index++) {
265 result += frame_sizes()[index];
266 }
267 return checked_cast<int>(result);
268 }
269
270 void Deoptimization::UnrollBlock::print() {
271 ResourceMark rm;
272 stringStream st;
273 st.print_cr("UnrollBlock");
274 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
275 st.print( " frame_sizes: ");
276 for (int index = 0; index < number_of_frames(); index++) {
277 st.print("%zd ", frame_sizes()[index]);
278 }
279 st.cr();
280 tty->print_raw(st.freeze());
281 }
282
283 // In order to make fetch_unroll_info work properly with escape
284 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
285 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
286 // which is called from the method fetch_unroll_info_helper below.
287 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
288 // fetch_unroll_info() is called at the beginning of the deoptimization
289 // handler. Note this fact before we start generating temporary frames
290 // that can confuse an asynchronous stack walker. This counter is
291 // decremented at the end of unpack_frames().
292 current->inc_in_deopt_handler();
293
294 if (exec_mode == Unpack_exception) {
295 // When we get here, a callee has thrown an exception into a deoptimized
296 // frame. That throw might have deferred stack watermark checking until
297 // after unwinding. So we deal with such deferred requests here.
298 StackWatermarkSet::after_unwind(current);
299 }
300
301 return fetch_unroll_info_helper(current, exec_mode);
302 JRT_END
303
304 #if COMPILER2_OR_JVMCI
305 // print information about reallocated objects
306 static void print_objects(JavaThread* deoptee_thread,
307 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
308 ResourceMark rm;
309 stringStream st; // change to logStream with logging
310 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
311 fieldDescriptor fd;
312
313 for (int i = 0; i < objects->length(); i++) {
314 ObjectValue* sv = (ObjectValue*) objects->at(i);
315 Handle obj = sv->value();
316
317 if (obj.is_null()) {
318 st.print_cr(" nullptr");
319 continue;
320 }
321
322 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
323
324 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
325 k->print_value_on(&st);
326 st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize);
327
328 if (Verbose && k != nullptr) {
329 k->oop_print_on(obj(), &st);
330 }
331 }
332 tty->print_raw(st.freeze());
333 }
334
335 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
336 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
337 bool& deoptimized_objects) {
338 bool realloc_failures = false;
339 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
340
341 JavaThread* deoptee_thread = chunk->at(0)->thread();
342 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
343 "a frame can only be deoptimized by the owner thread");
344
345 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
346
347 // The flag return_oop() indicates call sites which return oop
348 // in compiled code. Such sites include java method calls,
349 // runtime calls (for example, used to allocate new objects/arrays
350 // on slow code path) and any other calls generated in compiled code.
351 // It is not guaranteed that we can get such information here only
352 // by analyzing bytecode in deoptimized frames. This is why this flag
353 // is set during method compilation (see Compile::Process_OopMap_Node()).
354 // If the previous frame was popped or if we are dispatching an exception,
355 // we don't have an oop result.
356 ScopeDesc* scope = chunk->at(0)->scope();
357 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
358 // In case of the return of multiple values, we must take care
359 // of all oop return values.
360 GrowableArray<Handle> return_oops;
361 InlineKlass* vk = nullptr;
362 if (save_oop_result && scope->return_scalarized()) {
363 vk = InlineKlass::returned_inline_klass(map);
364 if (vk != nullptr) {
365 vk->save_oop_fields(map, return_oops);
366 save_oop_result = false;
367 }
368 }
369 if (save_oop_result) {
370 // Reallocation may trigger GC. If deoptimization happened on return from
371 // call which returns oop we need to save it since it is not in oopmap.
372 oop result = deoptee.saved_oop_result(&map);
373 assert(oopDesc::is_oop_or_null(result), "must be oop");
374 return_oops.push(Handle(thread, result));
375 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
376 if (TraceDeoptimization) {
377 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
378 tty->cr();
379 }
380 }
381 if (objects != nullptr || vk != nullptr) {
382 if (exec_mode == Deoptimization::Unpack_none) {
383 assert(thread->thread_state() == _thread_in_vm, "assumption");
384 JavaThread* THREAD = thread; // For exception macros.
385 // Clear pending OOM if reallocation fails and return true indicating allocation failure
386 if (vk != nullptr) {
387 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, CHECK_AND_CLEAR_(true));
388 }
389 if (objects != nullptr) {
390 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
391 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
392 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
393 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci, CHECK_AND_CLEAR_(true));
394 }
395 deoptimized_objects = true;
396 } else {
397 JavaThread* current = thread; // For JRT_BLOCK
398 JRT_BLOCK
399 if (vk != nullptr) {
400 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, THREAD);
401 }
402 if (objects != nullptr) {
403 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
404 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
405 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
406 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci, THREAD);
407 }
408 JRT_END
409 }
410 if (TraceDeoptimization && objects != nullptr) {
411 print_objects(deoptee_thread, objects, realloc_failures);
412 }
413 }
414 if (save_oop_result || vk != nullptr) {
415 // Restore result.
416 assert(return_oops.length() == 1, "no inline type");
417 deoptee.set_saved_oop_result(&map, return_oops.pop()());
418 }
419 return realloc_failures;
420 }
421
422 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
423 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
424 JavaThread* deoptee_thread = chunk->at(0)->thread();
425 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
426 assert(thread == Thread::current(), "should be");
427 HandleMark hm(thread);
428 #ifndef PRODUCT
429 bool first = true;
430 #endif // !PRODUCT
431 // Start locking from outermost/oldest frame
432 for (int i = (chunk->length() - 1); i >= 0; i--) {
433 compiledVFrame* cvf = chunk->at(i);
434 assert (cvf->scope() != nullptr,"expect only compiled java frames");
435 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
436 if (monitors->is_nonempty()) {
437 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
438 exec_mode, realloc_failures);
439 deoptimized_objects = deoptimized_objects || relocked;
440 #ifndef PRODUCT
441 if (PrintDeoptimizationDetails) {
442 ResourceMark rm;
443 stringStream st;
444 for (int j = 0; j < monitors->length(); j++) {
445 MonitorInfo* mi = monitors->at(j);
446 if (mi->eliminated()) {
447 if (first) {
448 first = false;
449 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
450 }
451 if (exec_mode == Deoptimization::Unpack_none) {
452 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
453 if (monitor != nullptr && monitor->object() == mi->owner()) {
454 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
455 continue;
456 }
457 }
458 if (mi->owner_is_scalar_replaced()) {
459 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
460 st.print_cr(" failed reallocation for klass %s", k->external_name());
461 } else {
462 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
463 }
464 }
465 }
466 tty->print_raw(st.freeze());
467 }
468 #endif // !PRODUCT
469 }
470 }
471 }
472
473 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
474 // The given vframes cover one physical frame.
475 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
476 bool& realloc_failures) {
477 frame deoptee = chunk->at(0)->fr();
478 JavaThread* deoptee_thread = chunk->at(0)->thread();
479 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
480 RegisterMap map(chunk->at(0)->register_map());
481 bool deoptimized_objects = false;
482
483 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
484
485 // Reallocate the non-escaping objects and restore their fields.
486 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
487 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
488 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
489 }
490
491 // MonitorInfo structures used in eliminate_locks are not GC safe.
492 NoSafepointVerifier no_safepoint;
493
494 // Now relock objects if synchronization on them was eliminated.
495 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
496 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
497 }
498 return deoptimized_objects;
499 }
500 #endif // COMPILER2_OR_JVMCI
501
502 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
503 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
504 // When we get here we are about to unwind the deoptee frame. In order to
505 // catch not yet safe to use frames, the following stack watermark barrier
506 // poll will make such frames safe to use.
507 StackWatermarkSet::before_unwind(current);
508
509 // Note: there is a safepoint safety issue here. No matter whether we enter
510 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
511 // the vframeArray is created.
512 //
513
514 // Allocate our special deoptimization ResourceMark
515 DeoptResourceMark* dmark = new DeoptResourceMark(current);
516 assert(current->deopt_mark() == nullptr, "Pending deopt!");
517 current->set_deopt_mark(dmark);
518
519 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
520 RegisterMap map(current,
521 RegisterMap::UpdateMap::include,
522 RegisterMap::ProcessFrames::include,
523 RegisterMap::WalkContinuation::skip);
524 RegisterMap dummy_map(current,
525 RegisterMap::UpdateMap::skip,
526 RegisterMap::ProcessFrames::include,
527 RegisterMap::WalkContinuation::skip);
528 // Now get the deoptee with a valid map
529 frame deoptee = stub_frame.sender(&map);
530 // Set the deoptee nmethod
531 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
532 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
533 current->set_deopt_compiled_method(nm);
534
535 if (VerifyStack) {
536 current->validate_frame_layout();
537 }
538
539 // Create a growable array of VFrames where each VFrame represents an inlined
540 // Java frame. This storage is allocated with the usual system arena.
541 assert(deoptee.is_compiled_frame(), "Wrong frame type");
542 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
543 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
544 while (!vf->is_top()) {
545 assert(vf->is_compiled_frame(), "Wrong frame type");
546 chunk->push(compiledVFrame::cast(vf));
547 vf = vf->sender();
548 }
549 assert(vf->is_compiled_frame(), "Wrong frame type");
550 chunk->push(compiledVFrame::cast(vf));
551
552 bool realloc_failures = false;
553
554 #if COMPILER2_OR_JVMCI
555 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
556
557 // Reallocate the non-escaping objects and restore their fields. Then
558 // relock objects if synchronization on them was eliminated.
559 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
560 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
561 bool unused;
562 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
563 }
564 #endif // COMPILER2_OR_JVMCI
565
566 // Ensure that no safepoint is taken after pointers have been stored
567 // in fields of rematerialized objects. If a safepoint occurs from here on
568 // out the java state residing in the vframeArray will be missed.
569 // Locks may be rebaised in a safepoint.
570 NoSafepointVerifier no_safepoint;
571
572 #if COMPILER2_OR_JVMCI
573 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
574 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
575 bool unused = false;
576 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
577 }
578 #endif // COMPILER2_OR_JVMCI
579
580 ScopeDesc* trap_scope = chunk->at(0)->scope();
581 Handle exceptionObject;
582 if (trap_scope->rethrow_exception()) {
583 #ifndef PRODUCT
584 if (PrintDeoptimizationDetails) {
585 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());
586 }
587 #endif // !PRODUCT
588
589 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
590 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
591 ScopeValue* topOfStack = expressions->top();
592 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
593 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
594 }
595
596 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
597 #if COMPILER2_OR_JVMCI
598 if (realloc_failures) {
599 // This destroys all ScopedValue bindings.
600 current->clear_scopedValueBindings();
601 pop_frames_failed_reallocs(current, array);
602 }
603 #endif
604
605 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
606 current->set_vframe_array_head(array);
607
608 // Now that the vframeArray has been created if we have any deferred local writes
609 // added by jvmti then we can free up that structure as the data is now in the
610 // vframeArray
611
612 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
613
614 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
615 CodeBlob* cb = stub_frame.cb();
616 // Verify we have the right vframeArray
617 assert(cb->frame_size() >= 0, "Unexpected frame size");
618 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
619
620 // If the deopt call site is a MethodHandle invoke call site we have
621 // to adjust the unpack_sp.
622 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
623 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
624 unpack_sp = deoptee.unextended_sp();
625
626 #ifdef ASSERT
627 assert(cb->is_deoptimization_stub() ||
628 cb->is_uncommon_trap_stub() ||
629 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
630 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
631 "unexpected code blob: %s", cb->name());
632 #endif
633
634 // This is a guarantee instead of an assert because if vframe doesn't match
635 // we will unpack the wrong deoptimized frame and wind up in strange places
636 // where it will be very difficult to figure out what went wrong. Better
637 // to die an early death here than some very obscure death later when the
638 // trail is cold.
639 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
640
641 int number_of_frames = array->frames();
642
643 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
644 // virtual activation, which is the reverse of the elements in the vframes array.
645 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
646 // +1 because we always have an interpreter return address for the final slot.
647 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
648 int popframe_extra_args = 0;
649 // Create an interpreter return address for the stub to use as its return
650 // address so the skeletal frames are perfectly walkable
651 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
652
653 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
654 // activation be put back on the expression stack of the caller for reexecution
655 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
656 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
657 }
658
659 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
660 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
661 // than simply use array->sender.pc(). This requires us to walk the current set of frames
662 //
663 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
664 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
665
666 // It's possible that the number of parameters at the call site is
667 // different than number of arguments in the callee when method
668 // handles are used. If the caller is interpreted get the real
669 // value so that the proper amount of space can be added to it's
670 // frame.
671 bool caller_was_method_handle = false;
672 if (deopt_sender.is_interpreted_frame()) {
673 methodHandle method(current, deopt_sender.interpreter_frame_method());
674 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
675 if (cur.has_member_arg()) {
676 // This should cover all real-world cases. One exception is a pathological chain of
677 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we
678 // would need to get the size from the resolved method entry. Another exception would
679 // be an invokedynamic with an adapter that is really a MethodHandle linker.
680 caller_was_method_handle = true;
681 }
682 }
683
684 //
685 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
686 // frame_sizes/frame_pcs[1] next oldest frame (int)
687 // frame_sizes/frame_pcs[n] youngest frame (int)
688 //
689 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
690 // owns the space for the return address to it's caller). Confusing ain't it.
691 //
692 // The vframe array can address vframes with indices running from
693 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
694 // When we create the skeletal frames we need the oldest frame to be in the zero slot
695 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
696 // so things look a little strange in this loop.
697 //
698 int callee_parameters = 0;
699 int callee_locals = 0;
700 for (int index = 0; index < array->frames(); index++ ) {
701 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
702 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
703 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
704 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
705 callee_locals,
706 index == 0,
707 popframe_extra_args);
708 // This pc doesn't have to be perfect just good enough to identify the frame
709 // as interpreted so the skeleton frame will be walkable
710 // The correct pc will be set when the skeleton frame is completely filled out
711 // The final pc we store in the loop is wrong and will be overwritten below
712 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
713
714 callee_parameters = array->element(index)->method()->size_of_parameters();
715 callee_locals = array->element(index)->method()->max_locals();
716 popframe_extra_args = 0;
717 }
718
719 // Compute whether the root vframe returns a float or double value.
720 BasicType return_type;
721 {
722 methodHandle method(current, array->element(0)->method());
723 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
724 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
725 }
726
727 // Compute information for handling adapters and adjusting the frame size of the caller.
728 int caller_adjustment = 0;
729
730 // Compute the amount the oldest interpreter frame will have to adjust
731 // its caller's stack by. If the caller is a compiled frame then
732 // we pretend that the callee has no parameters so that the
733 // extension counts for the full amount of locals and not just
734 // locals-parms. This is because without a c2i adapter the parm
735 // area as created by the compiled frame will not be usable by
736 // the interpreter. (Depending on the calling convention there
737 // may not even be enough space).
738
739 // QQQ I'd rather see this pushed down into last_frame_adjust
740 // and have it take the sender (aka caller).
741
742 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
743 caller_adjustment = last_frame_adjust(0, callee_locals);
744 } else if (callee_locals > callee_parameters) {
745 // The caller frame may need extending to accommodate
746 // non-parameter locals of the first unpacked interpreted frame.
747 // Compute that adjustment.
748 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
749 }
750
751 // If the sender is deoptimized we must retrieve the address of the handler
752 // since the frame will "magically" show the original pc before the deopt
753 // and we'd undo the deopt.
754
755 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
756 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
757 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
758 }
759
760 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
761
762 #if INCLUDE_JVMCI
763 if (exceptionObject() != nullptr) {
764 current->set_exception_oop(exceptionObject());
765 exec_mode = Unpack_exception;
766 }
767 #endif
768
769 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
770 assert(current->has_pending_exception(), "should have thrown OOME");
771 current->set_exception_oop(current->pending_exception());
772 current->clear_pending_exception();
773 exec_mode = Unpack_exception;
774 }
775
776 #if INCLUDE_JVMCI
777 if (current->frames_to_pop_failed_realloc() > 0) {
778 current->set_pending_monitorenter(false);
779 }
780 #endif
781
782 int caller_actual_parameters = -1; // value not used except for interpreted frames, see below
783 if (deopt_sender.is_interpreted_frame()) {
784 caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0);
785 }
786
787 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
788 caller_adjustment * BytesPerWord,
789 caller_actual_parameters,
790 number_of_frames,
791 frame_sizes,
792 frame_pcs,
793 return_type,
794 exec_mode);
795 // On some platforms, we need a way to pass some platform dependent
796 // information to the unpacking code so the skeletal frames come out
797 // correct (initial fp value, unextended sp, ...)
798 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
799
800 if (array->frames() > 1) {
801 if (VerifyStack && TraceDeoptimization) {
802 tty->print_cr("Deoptimizing method containing inlining");
803 }
804 }
805
806 array->set_unroll_block(info);
807 return info;
808 }
809
810 // Called to cleanup deoptimization data structures in normal case
811 // after unpacking to stack and when stack overflow error occurs
812 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
813 vframeArray *array) {
814
815 // Get array if coming from exception
816 if (array == nullptr) {
817 array = thread->vframe_array_head();
818 }
819 thread->set_vframe_array_head(nullptr);
820
821 // Free the previous UnrollBlock
822 vframeArray* old_array = thread->vframe_array_last();
823 thread->set_vframe_array_last(array);
824
825 if (old_array != nullptr) {
826 UnrollBlock* old_info = old_array->unroll_block();
827 old_array->set_unroll_block(nullptr);
828 delete old_info;
829 delete old_array;
830 }
831
832 // Deallocate any resource creating in this routine and any ResourceObjs allocated
833 // inside the vframeArray (StackValueCollections)
834
835 delete thread->deopt_mark();
836 thread->set_deopt_mark(nullptr);
837 thread->set_deopt_compiled_method(nullptr);
838
839
840 if (JvmtiExport::can_pop_frame()) {
841 // Regardless of whether we entered this routine with the pending
842 // popframe condition bit set, we should always clear it now
843 thread->clear_popframe_condition();
844 }
845
846 // unpack_frames() is called at the end of the deoptimization handler
847 // and (in C2) at the end of the uncommon trap handler. Note this fact
848 // so that an asynchronous stack walker can work again. This counter is
849 // incremented at the beginning of fetch_unroll_info() and (in C2) at
850 // the beginning of uncommon_trap().
851 thread->dec_in_deopt_handler();
852 }
853
854 // Moved from cpu directories because none of the cpus has callee save values.
855 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
856 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
857
858 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
859 // the days we had adapter frames. When we deoptimize a situation where a
860 // compiled caller calls a compiled caller will have registers it expects
861 // to survive the call to the callee. If we deoptimize the callee the only
862 // way we can restore these registers is to have the oldest interpreter
863 // frame that we create restore these values. That is what this routine
864 // will accomplish.
865
866 // At the moment we have modified c2 to not have any callee save registers
867 // so this problem does not exist and this routine is just a place holder.
868
869 assert(f->is_interpreted_frame(), "must be interpreted");
870 }
871
872 #ifndef PRODUCT
873 static bool falls_through(Bytecodes::Code bc) {
874 switch (bc) {
875 // List may be incomplete. Here we really only care about bytecodes where compiled code
876 // can deoptimize.
877 case Bytecodes::_goto:
878 case Bytecodes::_goto_w:
879 case Bytecodes::_athrow:
880 return false;
881 default:
882 return true;
883 }
884 }
885 #endif
886
887 // Return BasicType of value being returned
888 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
889 assert(thread == JavaThread::current(), "pre-condition");
890
891 // We are already active in the special DeoptResourceMark any ResourceObj's we
892 // allocate will be freed at the end of the routine.
893
894 // JRT_LEAF methods don't normally allocate handles and there is a
895 // NoHandleMark to enforce that. It is actually safe to use Handles
896 // in a JRT_LEAF method, and sometimes desirable, but to do so we
897 // must use ResetNoHandleMark to bypass the NoHandleMark, and
898 // then use a HandleMark to ensure any Handles we do create are
899 // cleaned up in this scope.
900 ResetNoHandleMark rnhm;
901 HandleMark hm(thread);
902
903 frame stub_frame = thread->last_frame();
904
905 Continuation::notify_deopt(thread, stub_frame.sp());
906
907 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
908 // must point to the vframeArray for the unpack frame.
909 vframeArray* array = thread->vframe_array_head();
910 UnrollBlock* info = array->unroll_block();
911
912 // We set the last_Java frame. But the stack isn't really parsable here. So we
913 // clear it to make sure JFR understands not to try and walk stacks from events
914 // in here.
915 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
916 thread->frame_anchor()->set_last_Java_sp(nullptr);
917
918 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
919 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
920
921 thread->frame_anchor()->set_last_Java_sp(sp);
922
923 BasicType bt = info->return_type();
924
925 // If we have an exception pending, claim that the return type is an oop
926 // so the deopt_blob does not overwrite the exception_oop.
927
928 if (exec_mode == Unpack_exception)
929 bt = T_OBJECT;
930
931 // Cleanup thread deopt data
932 cleanup_deopt_info(thread, array);
933
934 #ifndef PRODUCT
935 if (VerifyStack) {
936 ResourceMark res_mark;
937 // Clear pending exception to not break verification code (restored afterwards)
938 PreserveExceptionMark pm(thread);
939
940 thread->validate_frame_layout();
941
942 // Verify that the just-unpacked frames match the interpreter's
943 // notions of expression stack and locals
944 vframeArray* cur_array = thread->vframe_array_last();
945 RegisterMap rm(thread,
946 RegisterMap::UpdateMap::skip,
947 RegisterMap::ProcessFrames::include,
948 RegisterMap::WalkContinuation::skip);
949 rm.set_include_argument_oops(false);
950 bool is_top_frame = true;
951 int callee_size_of_parameters = 0;
952 int callee_max_locals = 0;
953 for (int i = 0; i < cur_array->frames(); i++) {
954 vframeArrayElement* el = cur_array->element(i);
955 frame* iframe = el->iframe();
956 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
957
958 // Get the oop map for this bci
959 InterpreterOopMap mask;
960 int cur_invoke_parameter_size = 0;
961 bool try_next_mask = false;
962 int next_mask_expression_stack_size = -1;
963 int top_frame_expression_stack_adjustment = 0;
964 methodHandle mh(thread, iframe->interpreter_frame_method());
965 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
966 BytecodeStream str(mh, iframe->interpreter_frame_bci());
967 int max_bci = mh->code_size();
968 // Get to the next bytecode if possible
969 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
970 // Check to see if we can grab the number of outgoing arguments
971 // at an uncommon trap for an invoke (where the compiler
972 // generates debug info before the invoke has executed)
973 Bytecodes::Code cur_code = str.next();
974 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
975 if (Bytecodes::is_invoke(cur_code)) {
976 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
977 cur_invoke_parameter_size = invoke.size_of_parameters();
978 if (i != 0 && invoke.has_member_arg()) {
979 callee_size_of_parameters++;
980 }
981 }
982 if (str.bci() < max_bci) {
983 next_code = str.next();
984 if (next_code >= 0) {
985 // The interpreter oop map generator reports results before
986 // the current bytecode has executed except in the case of
987 // calls. It seems to be hard to tell whether the compiler
988 // has emitted debug information matching the "state before"
989 // a given bytecode or the state after, so we try both
990 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
991 // Get expression stack size for the next bytecode
992 InterpreterOopMap next_mask;
993 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
994 next_mask_expression_stack_size = next_mask.expression_stack_size();
995 if (Bytecodes::is_invoke(next_code)) {
996 Bytecode_invoke invoke(mh, str.bci());
997 next_mask_expression_stack_size += invoke.size_of_parameters();
998 }
999 // Need to subtract off the size of the result type of
1000 // the bytecode because this is not described in the
1001 // debug info but returned to the interpreter in the TOS
1002 // caching register
1003 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
1004 if (bytecode_result_type != T_ILLEGAL) {
1005 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
1006 }
1007 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
1008 try_next_mask = true;
1009 }
1010 }
1011 }
1012
1013 // Verify stack depth and oops in frame
1014 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
1015 if (!(
1016 /* SPARC */
1017 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
1018 /* x86 */
1019 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
1020 (try_next_mask &&
1021 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
1022 top_frame_expression_stack_adjustment))) ||
1023 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
1024 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
1025 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
1026 )) {
1027 {
1028 // Print out some information that will help us debug the problem
1029 tty->print_cr("Wrong number of expression stack elements during deoptimization");
1030 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
1031 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1032 if (try_next_mask) {
1033 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1034 }
1035 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1036 iframe->interpreter_frame_expression_stack_size());
1037 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1038 tty->print_cr(" try_next_mask = %d", try_next_mask);
1039 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1040 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1041 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1042 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1043 tty->print_cr(" exec_mode = %d", exec_mode);
1044 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1045 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1046 tty->print_cr(" Interpreted frames:");
1047 for (int k = 0; k < cur_array->frames(); k++) {
1048 vframeArrayElement* el = cur_array->element(k);
1049 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1050 }
1051 cur_array->print_on_2(tty);
1052 }
1053 guarantee(false, "wrong number of expression stack elements during deopt");
1054 }
1055 VerifyOopClosure verify;
1056 iframe->oops_interpreted_do(&verify, &rm, false);
1057 callee_size_of_parameters = mh->size_of_parameters();
1058 callee_max_locals = mh->max_locals();
1059 is_top_frame = false;
1060 }
1061 }
1062 #endif // !PRODUCT
1063
1064 return bt;
1065 JRT_END
1066
1067 class DeoptimizeMarkedClosure : public HandshakeClosure {
1068 public:
1069 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1070 void do_thread(Thread* thread) {
1071 JavaThread* jt = JavaThread::cast(thread);
1072 jt->deoptimize_marked_methods();
1073 }
1074 };
1075
1076 void Deoptimization::deoptimize_all_marked() {
1077 ResourceMark rm;
1078
1079 // Make the dependent methods not entrant
1080 CodeCache::make_marked_nmethods_deoptimized();
1081
1082 DeoptimizeMarkedClosure deopt;
1083 if (SafepointSynchronize::is_at_safepoint()) {
1084 Threads::java_threads_do(&deopt);
1085 } else {
1086 Handshake::execute(&deopt);
1087 }
1088 }
1089
1090 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1091 = Deoptimization::Action_reinterpret;
1092
1093 #if INCLUDE_JVMCI
1094 template<typename CacheType>
1095 class BoxCacheBase : public CHeapObj<mtCompiler> {
1096 protected:
1097 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1098 ResourceMark rm(thread);
1099 char* klass_name_str = klass_name->as_C_string();
1100 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1101 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1102 if (!ik->is_in_error_state()) {
1103 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1104 CacheType::compute_offsets(ik);
1105 }
1106 return ik;
1107 }
1108 };
1109
1110 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1111 PrimitiveType _low;
1112 PrimitiveType _high;
1113 jobject _cache;
1114 protected:
1115 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1116 BoxCache(Thread* thread) {
1117 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1118 if (ik->is_in_error_state()) {
1119 _low = 1;
1120 _high = 0;
1121 _cache = nullptr;
1122 } else {
1123 objArrayOop cache = CacheType::cache(ik);
1124 assert(cache->length() > 0, "Empty cache");
1125 _low = BoxType::value(cache->obj_at(0));
1126 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1127 _cache = JNIHandles::make_global(Handle(thread, cache));
1128 }
1129 }
1130 ~BoxCache() {
1131 JNIHandles::destroy_global(_cache);
1132 }
1133 public:
1134 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1135 if (_singleton == nullptr) {
1136 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1137 if (!Atomic::replace_if_null(&_singleton, s)) {
1138 delete s;
1139 }
1140 }
1141 return _singleton;
1142 }
1143 oop lookup(PrimitiveType value) {
1144 if (_low <= value && value <= _high) {
1145 int offset = checked_cast<int>(value - _low);
1146 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1147 }
1148 return nullptr;
1149 }
1150 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1151 if (_cache == nullptr) {
1152 cache_init_error = true;
1153 return nullptr;
1154 }
1155 // Have to cast to avoid little/big-endian problems.
1156 if (sizeof(PrimitiveType) > sizeof(jint)) {
1157 jlong value = (jlong)raw_value;
1158 return lookup(value);
1159 }
1160 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1161 return lookup(value);
1162 }
1163 };
1164
1165 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1166 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1167 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1168 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1169 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1170
1171 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1172 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1173 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1174 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1175 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1176
1177 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1178 jobject _true_cache;
1179 jobject _false_cache;
1180 protected:
1181 static BooleanBoxCache *_singleton;
1182 BooleanBoxCache(Thread *thread) {
1183 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1184 if (ik->is_in_error_state()) {
1185 _true_cache = nullptr;
1186 _false_cache = nullptr;
1187 } else {
1188 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1189 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1190 }
1191 }
1192 ~BooleanBoxCache() {
1193 JNIHandles::destroy_global(_true_cache);
1194 JNIHandles::destroy_global(_false_cache);
1195 }
1196 public:
1197 static BooleanBoxCache* singleton(Thread* thread) {
1198 if (_singleton == nullptr) {
1199 BooleanBoxCache* s = new BooleanBoxCache(thread);
1200 if (!Atomic::replace_if_null(&_singleton, s)) {
1201 delete s;
1202 }
1203 }
1204 return _singleton;
1205 }
1206 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1207 if (_true_cache == nullptr) {
1208 cache_in_error = true;
1209 return nullptr;
1210 }
1211 // Have to cast to avoid little/big-endian problems.
1212 jboolean value = (jboolean)*((jint*)&raw_value);
1213 return lookup(value);
1214 }
1215 oop lookup(jboolean value) {
1216 if (value != 0) {
1217 return JNIHandles::resolve_non_null(_true_cache);
1218 }
1219 return JNIHandles::resolve_non_null(_false_cache);
1220 }
1221 };
1222
1223 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1224
1225 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1226 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1227 BasicType box_type = vmClasses::box_klass_type(k);
1228 if (box_type != T_OBJECT) {
1229 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1230 switch(box_type) {
1231 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1232 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1233 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1234 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1235 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1236 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1237 default:;
1238 }
1239 }
1240 return nullptr;
1241 }
1242 #endif // INCLUDE_JVMCI
1243
1244 #if COMPILER2_OR_JVMCI
1245 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1246 Handle pending_exception(THREAD, thread->pending_exception());
1247 const char* exception_file = thread->exception_file();
1248 int exception_line = thread->exception_line();
1249 thread->clear_pending_exception();
1250
1251 bool failures = false;
1252
1253 for (int i = 0; i < objects->length(); i++) {
1254 assert(objects->at(i)->is_object(), "invalid debug information");
1255 ObjectValue* sv = (ObjectValue*) objects->at(i);
1256 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1257
1258 // Check if the object may be null and has an additional is_init input that needs
1259 // to be checked before using the field values. Skip re-allocation if it is null.
1260 if (sv->maybe_null()) {
1261 assert(k->is_inline_klass(), "must be an inline klass");
1262 jint is_init = StackValue::create_stack_value(fr, reg_map, sv->is_init())->get_jint();
1263 if (is_init == 0) {
1264 continue;
1265 }
1266 }
1267
1268 oop obj = nullptr;
1269 bool cache_init_error = false;
1270 if (k->is_instance_klass()) {
1271 #if INCLUDE_JVMCI
1272 nmethod* nm = fr->cb()->as_nmethod_or_null();
1273 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1274 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1275 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1276 if (obj != nullptr) {
1277 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1278 abv->set_cached(true);
1279 } else if (cache_init_error) {
1280 // Results in an OOME which is valid (as opposed to a class initialization error)
1281 // and is fine for the rare case a cache initialization failing.
1282 failures = true;
1283 }
1284 }
1285 #endif // INCLUDE_JVMCI
1286
1287 InstanceKlass* ik = InstanceKlass::cast(k);
1288 if (obj == nullptr && !cache_init_error) {
1289 InternalOOMEMark iom(THREAD);
1290 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1291 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1292 } else {
1293 obj = ik->allocate_instance(THREAD);
1294 }
1295 }
1296 } else if (k->is_flatArray_klass()) {
1297 FlatArrayKlass* ak = FlatArrayKlass::cast(k);
1298 // Inline type array must be zeroed because not all memory is reassigned
1299 obj = ak->allocate(sv->field_size(), ak->layout_kind(), THREAD);
1300 } else if (k->is_typeArray_klass()) {
1301 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1302 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1303 int len = sv->field_size() / type2size[ak->element_type()];
1304 InternalOOMEMark iom(THREAD);
1305 obj = ak->allocate(len, THREAD);
1306 } else if (k->is_objArray_klass()) {
1307 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1308 InternalOOMEMark iom(THREAD);
1309 obj = ak->allocate(sv->field_size(), THREAD);
1310 }
1311
1312 if (obj == nullptr) {
1313 failures = true;
1314 }
1315
1316 assert(sv->value().is_null(), "redundant reallocation");
1317 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1318 CLEAR_PENDING_EXCEPTION;
1319 sv->set_value(obj);
1320 }
1321
1322 if (failures) {
1323 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1324 } else if (pending_exception.not_null()) {
1325 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1326 }
1327
1328 return failures;
1329 }
1330
1331 // We're deoptimizing at the return of a call, inline type fields are
1332 // in registers. When we go back to the interpreter, it will expect a
1333 // reference to an inline type instance. Allocate and initialize it from
1334 // the register values here.
1335 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
1336 oop new_vt = vk->realloc_result(map, return_oops, THREAD);
1337 if (new_vt == nullptr) {
1338 CLEAR_PENDING_EXCEPTION;
1339 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
1340 }
1341 return_oops.clear();
1342 return_oops.push(Handle(THREAD, new_vt));
1343 return false;
1344 }
1345
1346 #if INCLUDE_JVMCI
1347 /**
1348 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1349 * we need to somehow be able to recover the actual kind to be able to write the correct
1350 * amount of bytes.
1351 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1352 * the entries at index n + 1 to n + i are 'markers'.
1353 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1354 * expected form of the array would be:
1355 *
1356 * {b0, b1, b2, b3, INT, marker, b6, b7}
1357 *
1358 * Thus, in order to get back the size of the entry, we simply need to count the number
1359 * of marked entries
1360 *
1361 * @param virtualArray the virtualized byte array
1362 * @param i index of the virtual entry we are recovering
1363 * @return The number of bytes the entry spans
1364 */
1365 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1366 int index = i;
1367 while (++index < virtualArray->field_size() &&
1368 virtualArray->field_at(index)->is_marker()) {}
1369 return index - i;
1370 }
1371
1372 /**
1373 * If there was a guarantee for byte array to always start aligned to a long, we could
1374 * do a simple check on the parity of the index. Unfortunately, that is not always the
1375 * case. Thus, we check alignment of the actual address we are writing to.
1376 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1377 * write a long to index 3.
1378 */
1379 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1380 jbyte* res = obj->byte_at_addr(index);
1381 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1382 return res;
1383 }
1384
1385 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1386 switch (byte_count) {
1387 case 1:
1388 obj->byte_at_put(index, (jbyte) value->get_jint());
1389 break;
1390 case 2:
1391 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1392 break;
1393 case 4:
1394 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1395 break;
1396 case 8:
1397 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1398 break;
1399 default:
1400 ShouldNotReachHere();
1401 }
1402 }
1403 #endif // INCLUDE_JVMCI
1404
1405
1406 // restore elements of an eliminated type array
1407 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1408 int index = 0;
1409
1410 for (int i = 0; i < sv->field_size(); i++) {
1411 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1412 switch(type) {
1413 case T_LONG: case T_DOUBLE: {
1414 assert(value->type() == T_INT, "Agreement.");
1415 StackValue* low =
1416 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1417 #ifdef _LP64
1418 jlong res = (jlong)low->get_intptr();
1419 #else
1420 jlong res = jlong_from(value->get_jint(), low->get_jint());
1421 #endif
1422 obj->long_at_put(index, res);
1423 break;
1424 }
1425
1426 case T_INT: case T_FLOAT: { // 4 bytes.
1427 assert(value->type() == T_INT, "Agreement.");
1428 bool big_value = false;
1429 if (i + 1 < sv->field_size() && type == T_INT) {
1430 if (sv->field_at(i)->is_location()) {
1431 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1432 if (type == Location::dbl || type == Location::lng) {
1433 big_value = true;
1434 }
1435 } else if (sv->field_at(i)->is_constant_int()) {
1436 ScopeValue* next_scope_field = sv->field_at(i + 1);
1437 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1438 big_value = true;
1439 }
1440 }
1441 }
1442
1443 if (big_value) {
1444 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1445 #ifdef _LP64
1446 jlong res = (jlong)low->get_intptr();
1447 #else
1448 jlong res = jlong_from(value->get_jint(), low->get_jint());
1449 #endif
1450 obj->int_at_put(index, *(jint*)&res);
1451 obj->int_at_put(++index, *((jint*)&res + 1));
1452 } else {
1453 obj->int_at_put(index, value->get_jint());
1454 }
1455 break;
1456 }
1457
1458 case T_SHORT:
1459 assert(value->type() == T_INT, "Agreement.");
1460 obj->short_at_put(index, (jshort)value->get_jint());
1461 break;
1462
1463 case T_CHAR:
1464 assert(value->type() == T_INT, "Agreement.");
1465 obj->char_at_put(index, (jchar)value->get_jint());
1466 break;
1467
1468 case T_BYTE: {
1469 assert(value->type() == T_INT, "Agreement.");
1470 #if INCLUDE_JVMCI
1471 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1472 int byte_count = count_number_of_bytes_for_entry(sv, i);
1473 byte_array_put(obj, value, index, byte_count);
1474 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1475 i += byte_count - 1; // Balance the loop counter.
1476 index += byte_count;
1477 // index has been updated so continue at top of loop
1478 continue;
1479 #else
1480 obj->byte_at_put(index, (jbyte)value->get_jint());
1481 break;
1482 #endif // INCLUDE_JVMCI
1483 }
1484
1485 case T_BOOLEAN: {
1486 assert(value->type() == T_INT, "Agreement.");
1487 obj->bool_at_put(index, (jboolean)value->get_jint());
1488 break;
1489 }
1490
1491 default:
1492 ShouldNotReachHere();
1493 }
1494 index++;
1495 }
1496 }
1497
1498 // restore fields of an eliminated object array
1499 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1500 for (int i = 0; i < sv->field_size(); i++) {
1501 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1502 assert(value->type() == T_OBJECT, "object element expected");
1503 obj->obj_at_put(i, value->get_obj()());
1504 }
1505 }
1506
1507 class ReassignedField {
1508 public:
1509 int _offset;
1510 BasicType _type;
1511 InstanceKlass* _klass;
1512 bool _is_flat;
1513 bool _is_null_free;
1514 public:
1515 ReassignedField() : _offset(0), _type(T_ILLEGAL), _klass(nullptr), _is_flat(false), _is_null_free(false) { }
1516 };
1517
1518 static int compare(ReassignedField* left, ReassignedField* right) {
1519 return left->_offset - right->_offset;
1520 }
1521
1522
1523 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1524 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1525 InstanceKlass* super = klass->superklass();
1526 if (super != nullptr) {
1527 get_reassigned_fields(super, fields, is_jvmci);
1528 }
1529 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1530 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1531 ReassignedField field;
1532 field._offset = fs.offset();
1533 field._type = Signature::basic_type(fs.signature());
1534 if (fs.is_flat()) {
1535 field._is_flat = true;
1536 field._is_null_free = fs.is_null_free_inline_type();
1537 // Resolve klass of flat inline type field
1538 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index()));
1539 }
1540 fields->append(field);
1541 }
1542 }
1543 return fields;
1544 }
1545
1546 // Restore fields of an eliminated instance object employing the same field order used by the
1547 // compiler when it scalarizes an object at safepoints.
1548 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) {
1549 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1550 fields->sort(compare);
1551
1552 for (int i = 0; i < fields->length(); i++) {
1553 BasicType type = fields->at(i)._type;
1554 int offset = base_offset + fields->at(i)._offset;
1555 // Check for flat inline type field before accessing the ScopeValue because it might not have any fields
1556 if (fields->at(i)._is_flat) {
1557 // Recursively re-assign flat inline type fields
1558 InstanceKlass* vk = fields->at(i)._klass;
1559 assert(vk != nullptr, "must be resolved");
1560 offset -= InlineKlass::cast(vk)->payload_offset(); // Adjust offset to omit oop header
1561 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, is_jvmci, offset, CHECK_0);
1562 if (!fields->at(i)._is_null_free) {
1563 ScopeValue* scope_field = sv->field_at(svIndex);
1564 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1565 int nm_offset = offset + InlineKlass::cast(vk)->null_marker_offset();
1566 obj->bool_field_put(nm_offset, value->get_jint() & 1);
1567 svIndex++;
1568 }
1569 continue; // Continue because we don't need to increment svIndex
1570 }
1571
1572 ScopeValue* scope_field = sv->field_at(svIndex);
1573 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1574 switch (type) {
1575 case T_OBJECT:
1576 case T_ARRAY:
1577 assert(value->type() == T_OBJECT, "Agreement.");
1578 obj->obj_field_put(offset, value->get_obj()());
1579 break;
1580
1581 case T_INT: case T_FLOAT: { // 4 bytes.
1582 assert(value->type() == T_INT, "Agreement.");
1583 bool big_value = false;
1584 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1585 if (scope_field->is_location()) {
1586 Location::Type type = ((LocationValue*) scope_field)->location().type();
1587 if (type == Location::dbl || type == Location::lng) {
1588 big_value = true;
1589 }
1590 }
1591 if (scope_field->is_constant_int()) {
1592 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1593 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1594 big_value = true;
1595 }
1596 }
1597 }
1598
1599 if (big_value) {
1600 i++;
1601 assert(i < fields->length(), "second T_INT field needed");
1602 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1603 } else {
1604 obj->int_field_put(offset, value->get_jint());
1605 break;
1606 }
1607 }
1608 /* no break */
1609
1610 case T_LONG: case T_DOUBLE: {
1611 assert(value->type() == T_INT, "Agreement.");
1612 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1613 #ifdef _LP64
1614 jlong res = (jlong)low->get_intptr();
1615 #else
1616 jlong res = jlong_from(value->get_jint(), low->get_jint());
1617 #endif
1618 obj->long_field_put(offset, res);
1619 break;
1620 }
1621
1622 case T_SHORT:
1623 assert(value->type() == T_INT, "Agreement.");
1624 obj->short_field_put(offset, (jshort)value->get_jint());
1625 break;
1626
1627 case T_CHAR:
1628 assert(value->type() == T_INT, "Agreement.");
1629 obj->char_field_put(offset, (jchar)value->get_jint());
1630 break;
1631
1632 case T_BYTE:
1633 assert(value->type() == T_INT, "Agreement.");
1634 obj->byte_field_put(offset, (jbyte)value->get_jint());
1635 break;
1636
1637 case T_BOOLEAN:
1638 assert(value->type() == T_INT, "Agreement.");
1639 obj->bool_field_put(offset, (jboolean)value->get_jint());
1640 break;
1641
1642 default:
1643 ShouldNotReachHere();
1644 }
1645 svIndex++;
1646 }
1647
1648 return svIndex;
1649 }
1650
1651 // restore fields of an eliminated inline type array
1652 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool is_jvmci, TRAPS) {
1653 InlineKlass* vk = vak->element_klass();
1654 assert(vk->maybe_flat_in_array(), "should only be used for flat inline type arrays");
1655 // Adjust offset to omit oop header
1656 int base_offset = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT) - InlineKlass::cast(vk)->payload_offset();
1657 // Initialize all elements of the flat inline type array
1658 for (int i = 0; i < sv->field_size(); i++) {
1659 ScopeValue* val = sv->field_at(i);
1660 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1661 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, is_jvmci, offset, CHECK);
1662 }
1663 }
1664
1665 // restore fields of all eliminated objects and arrays
1666 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci, TRAPS) {
1667 for (int i = 0; i < objects->length(); i++) {
1668 assert(objects->at(i)->is_object(), "invalid debug information");
1669 ObjectValue* sv = (ObjectValue*) objects->at(i);
1670 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1671 Handle obj = sv->value();
1672 assert(obj.not_null() || realloc_failures || sv->maybe_null(), "reallocation was missed");
1673 #ifndef PRODUCT
1674 if (PrintDeoptimizationDetails) {
1675 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1676 }
1677 #endif // !PRODUCT
1678
1679 if (obj.is_null()) {
1680 continue;
1681 }
1682
1683 #if INCLUDE_JVMCI
1684 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1685 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1686 continue;
1687 }
1688 #endif // INCLUDE_JVMCI
1689 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1690 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1691 ScopeValue* payload = sv->field_at(0);
1692 if (payload->is_location() &&
1693 payload->as_LocationValue()->location().type() == Location::vector) {
1694 #ifndef PRODUCT
1695 if (PrintDeoptimizationDetails) {
1696 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1697 if (Verbose) {
1698 Handle obj = sv->value();
1699 k->oop_print_on(obj(), tty);
1700 }
1701 }
1702 #endif // !PRODUCT
1703 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1704 }
1705 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1706 // which could be restored after vector object allocation.
1707 }
1708 if (k->is_instance_klass()) {
1709 InstanceKlass* ik = InstanceKlass::cast(k);
1710 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci, 0, CHECK);
1711 } else if (k->is_flatArray_klass()) {
1712 FlatArrayKlass* vak = FlatArrayKlass::cast(k);
1713 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, is_jvmci, CHECK);
1714 } else if (k->is_typeArray_klass()) {
1715 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1716 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1717 } else if (k->is_objArray_klass()) {
1718 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1719 }
1720 }
1721 // These objects may escape when we return to Interpreter after deoptimization.
1722 // We need barrier so that stores that initialize these objects can't be reordered
1723 // with subsequent stores that make these objects accessible by other threads.
1724 OrderAccess::storestore();
1725 }
1726
1727
1728 // relock objects for which synchronization was eliminated
1729 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1730 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1731 bool relocked_objects = false;
1732 for (int i = 0; i < monitors->length(); i++) {
1733 MonitorInfo* mon_info = monitors->at(i);
1734 if (mon_info->eliminated()) {
1735 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1736 relocked_objects = true;
1737 if (!mon_info->owner_is_scalar_replaced()) {
1738 Handle obj(thread, mon_info->owner());
1739 markWord mark = obj->mark();
1740 if (exec_mode == Unpack_none) {
1741 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1742 // With exec_mode == Unpack_none obj may be thread local and locked in
1743 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1744 markWord dmw = mark.displaced_mark_helper();
1745 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1746 obj->set_mark(dmw);
1747 }
1748 if (mark.has_monitor()) {
1749 // defer relocking if the deoptee thread is currently waiting for obj
1750 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1751 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1752 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1753 if (LockingMode == LM_LEGACY) {
1754 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1755 } else if (UseObjectMonitorTable) {
1756 mon_info->lock()->clear_object_monitor_cache();
1757 }
1758 #ifdef ASSERT
1759 else {
1760 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1761 mon_info->lock()->set_bad_metadata_deopt();
1762 }
1763 #endif
1764 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1765 continue;
1766 }
1767 }
1768 }
1769 BasicLock* lock = mon_info->lock();
1770 if (LockingMode == LM_LIGHTWEIGHT) {
1771 // We have lost information about the correct state of the lock stack.
1772 // Entering may create an invalid lock stack. Inflate the lock if it
1773 // was fast_locked to restore the valid lock stack.
1774 if (UseObjectMonitorTable) {
1775 // UseObjectMonitorTable expects the BasicLock cache to be either a
1776 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1777 // to nullptr.
1778 lock->clear_object_monitor_cache();
1779 }
1780 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1781 if (deoptee_thread->lock_stack().contains(obj())) {
1782 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1783 deoptee_thread, thread);
1784 }
1785 assert(mon_info->owner()->is_locked(), "object must be locked now");
1786 assert(obj->mark().has_monitor(), "must be");
1787 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1788 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1789 } else {
1790 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1791 assert(mon_info->owner()->is_locked(), "object must be locked now");
1792 }
1793 }
1794 }
1795 }
1796 return relocked_objects;
1797 }
1798 #endif // COMPILER2_OR_JVMCI
1799
1800 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1801 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1802
1803 // Register map for next frame (used for stack crawl). We capture
1804 // the state of the deopt'ing frame's caller. Thus if we need to
1805 // stuff a C2I adapter we can properly fill in the callee-save
1806 // register locations.
1807 frame caller = fr.sender(reg_map);
1808 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1809
1810 frame sender = caller;
1811
1812 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1813 // the vframeArray containing the unpacking information is allocated in the C heap.
1814 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1815 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1816
1817 // Compare the vframeArray to the collected vframes
1818 assert(array->structural_compare(thread, chunk), "just checking");
1819
1820 if (TraceDeoptimization) {
1821 ResourceMark rm;
1822 stringStream st;
1823 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1824 st.print(" ");
1825 fr.print_on(&st);
1826 st.print_cr(" Virtual frames (innermost/newest first):");
1827 for (int index = 0; index < chunk->length(); index++) {
1828 compiledVFrame* vf = chunk->at(index);
1829 int bci = vf->raw_bci();
1830 const char* code_name;
1831 if (bci == SynchronizationEntryBCI) {
1832 code_name = "sync entry";
1833 } else {
1834 Bytecodes::Code code = vf->method()->code_at(bci);
1835 code_name = Bytecodes::name(code);
1836 }
1837
1838 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1839 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1840 st.print(" - %s", code_name);
1841 st.print_cr(" @ bci=%d ", bci);
1842 }
1843 tty->print_raw(st.freeze());
1844 tty->cr();
1845 }
1846
1847 return array;
1848 }
1849
1850 #if COMPILER2_OR_JVMCI
1851 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1852 // Reallocation of some scalar replaced objects failed. Record
1853 // that we need to pop all the interpreter frames for the
1854 // deoptimized compiled frame.
1855 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1856 thread->set_frames_to_pop_failed_realloc(array->frames());
1857 // Unlock all monitors here otherwise the interpreter will see a
1858 // mix of locked and unlocked monitors (because of failed
1859 // reallocations of synchronized objects) and be confused.
1860 for (int i = 0; i < array->frames(); i++) {
1861 MonitorChunk* monitors = array->element(i)->monitors();
1862 if (monitors != nullptr) {
1863 // Unlock in reverse order starting from most nested monitor.
1864 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1865 BasicObjectLock* src = monitors->at(j);
1866 if (src->obj() != nullptr) {
1867 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1868 }
1869 }
1870 array->element(i)->free_monitors();
1871 #ifdef ASSERT
1872 array->element(i)->set_removed_monitors();
1873 #endif
1874 }
1875 }
1876 }
1877 #endif
1878
1879 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1880 assert(fr.can_be_deoptimized(), "checking frame type");
1881
1882 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1883
1884 if (LogCompilation && xtty != nullptr) {
1885 nmethod* nm = fr.cb()->as_nmethod_or_null();
1886 assert(nm != nullptr, "only compiled methods can deopt");
1887
1888 ttyLocker ttyl;
1889 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1890 nm->log_identity(xtty);
1891 xtty->end_head();
1892 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1893 xtty->begin_elem("jvms bci='%d'", sd->bci());
1894 xtty->method(sd->method());
1895 xtty->end_elem();
1896 if (sd->is_top()) break;
1897 }
1898 xtty->tail("deoptimized");
1899 }
1900
1901 Continuation::notify_deopt(thread, fr.sp());
1902
1903 // Patch the compiled method so that when execution returns to it we will
1904 // deopt the execution state and return to the interpreter.
1905 fr.deoptimize(thread);
1906 }
1907
1908 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1909 // Deoptimize only if the frame comes from compiled code.
1910 // Do not deoptimize the frame which is already patched
1911 // during the execution of the loops below.
1912 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1913 return;
1914 }
1915 ResourceMark rm;
1916 deoptimize_single_frame(thread, fr, reason);
1917 }
1918
1919 #if INCLUDE_JVMCI
1920 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1921 // there is no exception handler for this pc => deoptimize
1922 nm->make_not_entrant("missing exception handler");
1923
1924 // Use Deoptimization::deoptimize for all of its side-effects:
1925 // gathering traps statistics, logging...
1926 // it also patches the return pc but we do not care about that
1927 // since we return a continuation to the deopt_blob below.
1928 JavaThread* thread = JavaThread::current();
1929 RegisterMap reg_map(thread,
1930 RegisterMap::UpdateMap::skip,
1931 RegisterMap::ProcessFrames::include,
1932 RegisterMap::WalkContinuation::skip);
1933 frame runtime_frame = thread->last_frame();
1934 frame caller_frame = runtime_frame.sender(®_map);
1935 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1936 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1937 compiledVFrame* cvf = compiledVFrame::cast(vf);
1938 ScopeDesc* imm_scope = cvf->scope();
1939 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1940 if (imm_mdo != nullptr) {
1941 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1942 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1943
1944 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1945 if (pdata != nullptr && pdata->is_BitData()) {
1946 BitData* bit_data = (BitData*) pdata;
1947 bit_data->set_exception_seen();
1948 }
1949 }
1950
1951 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1952
1953 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1954 if (trap_mdo != nullptr) {
1955 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1956 }
1957
1958 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1959 }
1960 #endif
1961
1962 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1963 assert(thread == Thread::current() ||
1964 thread->is_handshake_safe_for(Thread::current()) ||
1965 SafepointSynchronize::is_at_safepoint(),
1966 "can only deoptimize other thread at a safepoint/handshake");
1967 // Compute frame and register map based on thread and sp.
1968 RegisterMap reg_map(thread,
1969 RegisterMap::UpdateMap::skip,
1970 RegisterMap::ProcessFrames::include,
1971 RegisterMap::WalkContinuation::skip);
1972 frame fr = thread->last_frame();
1973 while (fr.id() != id) {
1974 fr = fr.sender(®_map);
1975 }
1976 deoptimize(thread, fr, reason);
1977 }
1978
1979
1980 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1981 Thread* current = Thread::current();
1982 if (thread == current || thread->is_handshake_safe_for(current)) {
1983 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1984 } else {
1985 VM_DeoptimizeFrame deopt(thread, id, reason);
1986 VMThread::execute(&deopt);
1987 }
1988 }
1989
1990 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1991 deoptimize_frame(thread, id, Reason_constraint);
1992 }
1993
1994 // JVMTI PopFrame support
1995 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1996 {
1997 assert(thread == JavaThread::current(), "pre-condition");
1998 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1999 }
2000 JRT_END
2001
2002 MethodData*
2003 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
2004 bool create_if_missing) {
2005 JavaThread* THREAD = thread; // For exception macros.
2006 MethodData* mdo = m()->method_data();
2007 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
2008 // Build an MDO. Ignore errors like OutOfMemory;
2009 // that simply means we won't have an MDO to update.
2010 Method::build_profiling_method_data(m, THREAD);
2011 if (HAS_PENDING_EXCEPTION) {
2012 // Only metaspace OOM is expected. No Java code executed.
2013 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
2014 CLEAR_PENDING_EXCEPTION;
2015 }
2016 mdo = m()->method_data();
2017 }
2018 return mdo;
2019 }
2020
2021 #if COMPILER2_OR_JVMCI
2022 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
2023 // In case of an unresolved klass entry, load the class.
2024 // This path is exercised from case _ldc in Parse::do_one_bytecode,
2025 // and probably nowhere else.
2026 // Even that case would benefit from simply re-interpreting the
2027 // bytecode, without paying special attention to the class index.
2028 // So this whole "class index" feature should probably be removed.
2029
2030 if (constant_pool->tag_at(index).is_unresolved_klass()) {
2031 Klass* tk = constant_pool->klass_at(index, THREAD);
2032 if (HAS_PENDING_EXCEPTION) {
2033 // Exception happened during classloading. We ignore the exception here, since it
2034 // is going to be rethrown since the current activation is going to be deoptimized and
2035 // the interpreter will re-execute the bytecode.
2036 // Do not clear probable Async Exceptions.
2037 CLEAR_PENDING_NONASYNC_EXCEPTION;
2038 // Class loading called java code which may have caused a stack
2039 // overflow. If the exception was thrown right before the return
2040 // to the runtime the stack is no longer guarded. Reguard the
2041 // stack otherwise if we return to the uncommon trap blob and the
2042 // stack bang causes a stack overflow we crash.
2043 JavaThread* jt = THREAD;
2044 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
2045 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
2046 }
2047 return;
2048 }
2049
2050 assert(!constant_pool->tag_at(index).is_symbol(),
2051 "no symbolic names here, please");
2052 }
2053
2054 #if INCLUDE_JFR
2055
2056 class DeoptReasonSerializer : public JfrSerializer {
2057 public:
2058 void serialize(JfrCheckpointWriter& writer) {
2059 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
2060 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
2061 writer.write_key((u8)i);
2062 writer.write(Deoptimization::trap_reason_name(i));
2063 }
2064 }
2065 };
2066
2067 class DeoptActionSerializer : public JfrSerializer {
2068 public:
2069 void serialize(JfrCheckpointWriter& writer) {
2070 static const u4 nof_actions = Deoptimization::Action_LIMIT;
2071 writer.write_count(nof_actions);
2072 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
2073 writer.write_key(i);
2074 writer.write(Deoptimization::trap_action_name((int)i));
2075 }
2076 }
2077 };
2078
2079 static void register_serializers() {
2080 static int critical_section = 0;
2081 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
2082 return;
2083 }
2084 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
2085 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
2086 }
2087
2088 static void post_deoptimization_event(nmethod* nm,
2089 const Method* method,
2090 int trap_bci,
2091 int instruction,
2092 Deoptimization::DeoptReason reason,
2093 Deoptimization::DeoptAction action) {
2094 assert(nm != nullptr, "invariant");
2095 assert(method != nullptr, "invariant");
2096 if (EventDeoptimization::is_enabled()) {
2097 static bool serializers_registered = false;
2098 if (!serializers_registered) {
2099 register_serializers();
2100 serializers_registered = true;
2101 }
2102 EventDeoptimization event;
2103 event.set_compileId(nm->compile_id());
2104 event.set_compiler(nm->compiler_type());
2105 event.set_method(method);
2106 event.set_lineNumber(method->line_number_from_bci(trap_bci));
2107 event.set_bci(trap_bci);
2108 event.set_instruction(instruction);
2109 event.set_reason(reason);
2110 event.set_action(action);
2111 event.commit();
2112 }
2113 }
2114
2115 #endif // INCLUDE_JFR
2116
2117 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2118 const char* reason_name, const char* reason_action) {
2119 LogTarget(Debug, deoptimization) lt;
2120 if (lt.is_enabled()) {
2121 LogStream ls(lt);
2122 bool is_osr = nm->is_osr_method();
2123 ls.print("cid=%4d %s level=%d",
2124 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2125 ls.print(" %s", tm->name_and_sig_as_C_string());
2126 ls.print(" trap_bci=%d ", trap_bci);
2127 if (is_osr) {
2128 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2129 }
2130 ls.print("%s ", reason_name);
2131 ls.print("%s ", reason_action);
2132 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2133 pc, fr.pc() - nm->code_begin());
2134 }
2135 }
2136
2137 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2138 HandleMark hm(current);
2139
2140 // uncommon_trap() is called at the beginning of the uncommon trap
2141 // handler. Note this fact before we start generating temporary frames
2142 // that can confuse an asynchronous stack walker. This counter is
2143 // decremented at the end of unpack_frames().
2144
2145 current->inc_in_deopt_handler();
2146
2147 #if INCLUDE_JVMCI
2148 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2149 RegisterMap reg_map(current,
2150 RegisterMap::UpdateMap::include,
2151 RegisterMap::ProcessFrames::include,
2152 RegisterMap::WalkContinuation::skip);
2153 #else
2154 RegisterMap reg_map(current,
2155 RegisterMap::UpdateMap::skip,
2156 RegisterMap::ProcessFrames::include,
2157 RegisterMap::WalkContinuation::skip);
2158 #endif
2159 frame stub_frame = current->last_frame();
2160 frame fr = stub_frame.sender(®_map);
2161
2162 // Log a message
2163 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2164 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2165
2166 {
2167 ResourceMark rm;
2168
2169 DeoptReason reason = trap_request_reason(trap_request);
2170 DeoptAction action = trap_request_action(trap_request);
2171 #if INCLUDE_JVMCI
2172 int debug_id = trap_request_debug_id(trap_request);
2173 #endif
2174 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2175
2176 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2177 compiledVFrame* cvf = compiledVFrame::cast(vf);
2178
2179 nmethod* nm = cvf->code();
2180
2181 ScopeDesc* trap_scope = cvf->scope();
2182
2183 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2184
2185 if (is_receiver_constraint_failure) {
2186 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2187 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2188 JVMCI_ONLY(COMMA debug_id));
2189 }
2190
2191 methodHandle trap_method(current, trap_scope->method());
2192 int trap_bci = trap_scope->bci();
2193 #if INCLUDE_JVMCI
2194 jlong speculation = current->pending_failed_speculation();
2195 if (nm->is_compiled_by_jvmci()) {
2196 nm->update_speculation(current);
2197 } else {
2198 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2199 }
2200
2201 if (trap_bci == SynchronizationEntryBCI) {
2202 trap_bci = 0;
2203 current->set_pending_monitorenter(true);
2204 }
2205
2206 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2207 current->set_pending_transfer_to_interpreter(true);
2208 }
2209 #endif
2210
2211 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2212 // Record this event in the histogram.
2213 gather_statistics(reason, action, trap_bc);
2214
2215 // Ensure that we can record deopt. history:
2216 bool create_if_missing = ProfileTraps;
2217
2218 methodHandle profiled_method;
2219 #if INCLUDE_JVMCI
2220 if (nm->is_compiled_by_jvmci()) {
2221 profiled_method = methodHandle(current, nm->method());
2222 } else {
2223 profiled_method = trap_method;
2224 }
2225 #else
2226 profiled_method = trap_method;
2227 #endif
2228
2229 MethodData* trap_mdo =
2230 get_method_data(current, profiled_method, create_if_missing);
2231
2232 { // Log Deoptimization event for JFR, UL and event system
2233 Method* tm = trap_method();
2234 const char* reason_name = trap_reason_name(reason);
2235 const char* reason_action = trap_action_name(action);
2236 intptr_t pc = p2i(fr.pc());
2237
2238 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2239 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2240 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2241 reason_name, reason_action, pc,
2242 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2243 }
2244
2245 // Print a bunch of diagnostics, if requested.
2246 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2247 ResourceMark rm;
2248
2249 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2250 // We must do this already now, since we cannot acquire this lock while
2251 // holding the tty lock (lock ordering by rank).
2252 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2253
2254 ttyLocker ttyl;
2255
2256 char buf[100];
2257 if (xtty != nullptr) {
2258 xtty->begin_head("uncommon_trap thread='%zu' %s",
2259 os::current_thread_id(),
2260 format_trap_request(buf, sizeof(buf), trap_request));
2261 #if INCLUDE_JVMCI
2262 if (speculation != 0) {
2263 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2264 }
2265 #endif
2266 nm->log_identity(xtty);
2267 }
2268 Symbol* class_name = nullptr;
2269 bool unresolved = false;
2270 if (unloaded_class_index >= 0) {
2271 constantPoolHandle constants (current, trap_method->constants());
2272 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2273 class_name = constants->klass_name_at(unloaded_class_index);
2274 unresolved = true;
2275 if (xtty != nullptr)
2276 xtty->print(" unresolved='1'");
2277 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2278 class_name = constants->symbol_at(unloaded_class_index);
2279 }
2280 if (xtty != nullptr)
2281 xtty->name(class_name);
2282 }
2283 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2284 // Dump the relevant MDO state.
2285 // This is the deopt count for the current reason, any previous
2286 // reasons or recompiles seen at this point.
2287 int dcnt = trap_mdo->trap_count(reason);
2288 if (dcnt != 0)
2289 xtty->print(" count='%d'", dcnt);
2290
2291 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2292 // lock extra_data_lock before the tty lock.
2293 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2294 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2295 if (dos != 0) {
2296 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2297 if (trap_state_is_recompiled(dos)) {
2298 int recnt2 = trap_mdo->overflow_recompile_count();
2299 if (recnt2 != 0)
2300 xtty->print(" recompiles2='%d'", recnt2);
2301 }
2302 }
2303 }
2304 if (xtty != nullptr) {
2305 xtty->stamp();
2306 xtty->end_head();
2307 }
2308 if (TraceDeoptimization) { // make noise on the tty
2309 stringStream st;
2310 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2311 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2312 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2313 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2314 #if INCLUDE_JVMCI
2315 if (nm->is_compiled_by_jvmci()) {
2316 const char* installed_code_name = nm->jvmci_name();
2317 if (installed_code_name != nullptr) {
2318 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2319 }
2320 }
2321 #endif
2322 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2323 p2i(fr.pc()),
2324 os::current_thread_id(),
2325 trap_reason_name(reason),
2326 trap_action_name(action),
2327 unloaded_class_index
2328 #if INCLUDE_JVMCI
2329 , debug_id
2330 #endif
2331 );
2332 if (class_name != nullptr) {
2333 st.print(unresolved ? " unresolved class: " : " symbol: ");
2334 class_name->print_symbol_on(&st);
2335 }
2336 st.cr();
2337 tty->print_raw(st.freeze());
2338 }
2339 if (xtty != nullptr) {
2340 // Log the precise location of the trap.
2341 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2342 xtty->begin_elem("jvms bci='%d'", sd->bci());
2343 xtty->method(sd->method());
2344 xtty->end_elem();
2345 if (sd->is_top()) break;
2346 }
2347 xtty->tail("uncommon_trap");
2348 }
2349 }
2350 // (End diagnostic printout.)
2351
2352 if (is_receiver_constraint_failure) {
2353 fatal("missing receiver type check");
2354 }
2355
2356 // Load class if necessary
2357 if (unloaded_class_index >= 0) {
2358 constantPoolHandle constants(current, trap_method->constants());
2359 load_class_by_index(constants, unloaded_class_index, THREAD);
2360 }
2361
2362 // Flush the nmethod if necessary and desirable.
2363 //
2364 // We need to avoid situations where we are re-flushing the nmethod
2365 // because of a hot deoptimization site. Repeated flushes at the same
2366 // point need to be detected by the compiler and avoided. If the compiler
2367 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2368 // module must take measures to avoid an infinite cycle of recompilation
2369 // and deoptimization. There are several such measures:
2370 //
2371 // 1. If a recompilation is ordered a second time at some site X
2372 // and for the same reason R, the action is adjusted to 'reinterpret',
2373 // to give the interpreter time to exercise the method more thoroughly.
2374 // If this happens, the method's overflow_recompile_count is incremented.
2375 //
2376 // 2. If the compiler fails to reduce the deoptimization rate, then
2377 // the method's overflow_recompile_count will begin to exceed the set
2378 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2379 // is adjusted to 'make_not_compilable', and the method is abandoned
2380 // to the interpreter. This is a performance hit for hot methods,
2381 // but is better than a disastrous infinite cycle of recompilations.
2382 // (Actually, only the method containing the site X is abandoned.)
2383 //
2384 // 3. In parallel with the previous measures, if the total number of
2385 // recompilations of a method exceeds the much larger set limit
2386 // PerMethodRecompilationCutoff, the method is abandoned.
2387 // This should only happen if the method is very large and has
2388 // many "lukewarm" deoptimizations. The code which enforces this
2389 // limit is elsewhere (class nmethod, class Method).
2390 //
2391 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2392 // to recompile at each bytecode independently of the per-BCI cutoff.
2393 //
2394 // The decision to update code is up to the compiler, and is encoded
2395 // in the Action_xxx code. If the compiler requests Action_none
2396 // no trap state is changed, no compiled code is changed, and the
2397 // computation suffers along in the interpreter.
2398 //
2399 // The other action codes specify various tactics for decompilation
2400 // and recompilation. Action_maybe_recompile is the loosest, and
2401 // allows the compiled code to stay around until enough traps are seen,
2402 // and until the compiler gets around to recompiling the trapping method.
2403 //
2404 // The other actions cause immediate removal of the present code.
2405
2406 // Traps caused by injected profile shouldn't pollute trap counts.
2407 bool injected_profile_trap = trap_method->has_injected_profile() &&
2408 (reason == Reason_intrinsic || reason == Reason_unreached);
2409
2410 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2411 bool make_not_entrant = false;
2412 bool make_not_compilable = false;
2413 bool reprofile = false;
2414 switch (action) {
2415 case Action_none:
2416 // Keep the old code.
2417 update_trap_state = false;
2418 break;
2419 case Action_maybe_recompile:
2420 // Do not need to invalidate the present code, but we can
2421 // initiate another
2422 // Start compiler without (necessarily) invalidating the nmethod.
2423 // The system will tolerate the old code, but new code should be
2424 // generated when possible.
2425 break;
2426 case Action_reinterpret:
2427 // Go back into the interpreter for a while, and then consider
2428 // recompiling form scratch.
2429 make_not_entrant = true;
2430 // Reset invocation counter for outer most method.
2431 // This will allow the interpreter to exercise the bytecodes
2432 // for a while before recompiling.
2433 // By contrast, Action_make_not_entrant is immediate.
2434 //
2435 // Note that the compiler will track null_check, null_assert,
2436 // range_check, and class_check events and log them as if they
2437 // had been traps taken from compiled code. This will update
2438 // the MDO trap history so that the next compilation will
2439 // properly detect hot trap sites.
2440 reprofile = true;
2441 break;
2442 case Action_make_not_entrant:
2443 // Request immediate recompilation, and get rid of the old code.
2444 // Make them not entrant, so next time they are called they get
2445 // recompiled. Unloaded classes are loaded now so recompile before next
2446 // time they are called. Same for uninitialized. The interpreter will
2447 // link the missing class, if any.
2448 make_not_entrant = true;
2449 break;
2450 case Action_make_not_compilable:
2451 // Give up on compiling this method at all.
2452 make_not_entrant = true;
2453 make_not_compilable = true;
2454 break;
2455 default:
2456 ShouldNotReachHere();
2457 }
2458
2459 // Setting +ProfileTraps fixes the following, on all platforms:
2460 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2461 // recompile relies on a MethodData* to record heroic opt failures.
2462
2463 // Whether the interpreter is producing MDO data or not, we also need
2464 // to use the MDO to detect hot deoptimization points and control
2465 // aggressive optimization.
2466 bool inc_recompile_count = false;
2467
2468 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2469 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2470 (trap_mdo != nullptr),
2471 Mutex::_no_safepoint_check_flag);
2472 ProfileData* pdata = nullptr;
2473 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2474 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2475 uint this_trap_count = 0;
2476 bool maybe_prior_trap = false;
2477 bool maybe_prior_recompile = false;
2478
2479 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2480 #if INCLUDE_JVMCI
2481 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2482 #endif
2483 nm->method(),
2484 //outputs:
2485 this_trap_count,
2486 maybe_prior_trap,
2487 maybe_prior_recompile);
2488 // Because the interpreter also counts null, div0, range, and class
2489 // checks, these traps from compiled code are double-counted.
2490 // This is harmless; it just means that the PerXTrapLimit values
2491 // are in effect a little smaller than they look.
2492
2493 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2494 if (per_bc_reason != Reason_none) {
2495 // Now take action based on the partially known per-BCI history.
2496 if (maybe_prior_trap
2497 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2498 // If there are too many traps at this BCI, force a recompile.
2499 // This will allow the compiler to see the limit overflow, and
2500 // take corrective action, if possible. The compiler generally
2501 // does not use the exact PerBytecodeTrapLimit value, but instead
2502 // changes its tactics if it sees any traps at all. This provides
2503 // a little hysteresis, delaying a recompile until a trap happens
2504 // several times.
2505 //
2506 // Actually, since there is only one bit of counter per BCI,
2507 // the possible per-BCI counts are {0,1,(per-method count)}.
2508 // This produces accurate results if in fact there is only
2509 // one hot trap site, but begins to get fuzzy if there are
2510 // many sites. For example, if there are ten sites each
2511 // trapping two or more times, they each get the blame for
2512 // all of their traps.
2513 make_not_entrant = true;
2514 }
2515
2516 // Detect repeated recompilation at the same BCI, and enforce a limit.
2517 if (make_not_entrant && maybe_prior_recompile) {
2518 // More than one recompile at this point.
2519 inc_recompile_count = maybe_prior_trap;
2520 }
2521 } else {
2522 // For reasons which are not recorded per-bytecode, we simply
2523 // force recompiles unconditionally.
2524 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2525 make_not_entrant = true;
2526 }
2527
2528 // Go back to the compiler if there are too many traps in this method.
2529 if (this_trap_count >= per_method_trap_limit(reason)) {
2530 // If there are too many traps in this method, force a recompile.
2531 // This will allow the compiler to see the limit overflow, and
2532 // take corrective action, if possible.
2533 // (This condition is an unlikely backstop only, because the
2534 // PerBytecodeTrapLimit is more likely to take effect first,
2535 // if it is applicable.)
2536 make_not_entrant = true;
2537 }
2538
2539 // Here's more hysteresis: If there has been a recompile at
2540 // this trap point already, run the method in the interpreter
2541 // for a while to exercise it more thoroughly.
2542 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2543 reprofile = true;
2544 }
2545 }
2546
2547 // Take requested actions on the method:
2548
2549 // Recompile
2550 if (make_not_entrant) {
2551 if (!nm->make_not_entrant("uncommon trap")) {
2552 return; // the call did not change nmethod's state
2553 }
2554
2555 if (pdata != nullptr) {
2556 // Record the recompilation event, if any.
2557 int tstate0 = pdata->trap_state();
2558 int tstate1 = trap_state_set_recompiled(tstate0, true);
2559 if (tstate1 != tstate0)
2560 pdata->set_trap_state(tstate1);
2561 }
2562
2563 // For code aging we count traps separately here, using make_not_entrant()
2564 // as a guard against simultaneous deopts in multiple threads.
2565 if (reason == Reason_tenured && trap_mdo != nullptr) {
2566 trap_mdo->inc_tenure_traps();
2567 }
2568 }
2569
2570 if (inc_recompile_count) {
2571 trap_mdo->inc_overflow_recompile_count();
2572 if ((uint)trap_mdo->overflow_recompile_count() >
2573 (uint)PerBytecodeRecompilationCutoff) {
2574 // Give up on the method containing the bad BCI.
2575 if (trap_method() == nm->method()) {
2576 make_not_compilable = true;
2577 } else {
2578 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2579 // But give grace to the enclosing nm->method().
2580 }
2581 }
2582 }
2583
2584 // Reprofile
2585 if (reprofile) {
2586 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2587 }
2588
2589 // Give up compiling
2590 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2591 assert(make_not_entrant, "consistent");
2592 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2593 }
2594
2595 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2596 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2597 if (exception_handler_data != nullptr) {
2598 // uncommon trap at the start of an exception handler.
2599 // C2 generates these for un-entered exception handlers.
2600 // mark the handler as entered to avoid generating
2601 // another uncommon trap the next time the handler is compiled
2602 exception_handler_data->set_exception_handler_entered();
2603 }
2604 }
2605
2606 } // Free marked resources
2607
2608 }
2609 JRT_END
2610
2611 ProfileData*
2612 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2613 int trap_bci,
2614 Deoptimization::DeoptReason reason,
2615 bool update_total_trap_count,
2616 #if INCLUDE_JVMCI
2617 bool is_osr,
2618 #endif
2619 Method* compiled_method,
2620 //outputs:
2621 uint& ret_this_trap_count,
2622 bool& ret_maybe_prior_trap,
2623 bool& ret_maybe_prior_recompile) {
2624 trap_mdo->check_extra_data_locked();
2625
2626 bool maybe_prior_trap = false;
2627 bool maybe_prior_recompile = false;
2628 uint this_trap_count = 0;
2629 if (update_total_trap_count) {
2630 uint idx = reason;
2631 #if INCLUDE_JVMCI
2632 if (is_osr) {
2633 // Upper half of history array used for traps in OSR compilations
2634 idx += Reason_TRAP_HISTORY_LENGTH;
2635 }
2636 #endif
2637 uint prior_trap_count = trap_mdo->trap_count(idx);
2638 this_trap_count = trap_mdo->inc_trap_count(idx);
2639
2640 // If the runtime cannot find a place to store trap history,
2641 // it is estimated based on the general condition of the method.
2642 // If the method has ever been recompiled, or has ever incurred
2643 // a trap with the present reason , then this BCI is assumed
2644 // (pessimistically) to be the culprit.
2645 maybe_prior_trap = (prior_trap_count != 0);
2646 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2647 }
2648 ProfileData* pdata = nullptr;
2649
2650
2651 // For reasons which are recorded per bytecode, we check per-BCI data.
2652 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2653 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2654 if (per_bc_reason != Reason_none) {
2655 // Find the profile data for this BCI. If there isn't one,
2656 // try to allocate one from the MDO's set of spares.
2657 // This will let us detect a repeated trap at this point.
2658 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2659
2660 if (pdata != nullptr) {
2661 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2662 if (LogCompilation && xtty != nullptr) {
2663 ttyLocker ttyl;
2664 // no more room for speculative traps in this MDO
2665 xtty->elem("speculative_traps_oom");
2666 }
2667 }
2668 // Query the trap state of this profile datum.
2669 int tstate0 = pdata->trap_state();
2670 if (!trap_state_has_reason(tstate0, per_bc_reason))
2671 maybe_prior_trap = false;
2672 if (!trap_state_is_recompiled(tstate0))
2673 maybe_prior_recompile = false;
2674
2675 // Update the trap state of this profile datum.
2676 int tstate1 = tstate0;
2677 // Record the reason.
2678 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2679 // Store the updated state on the MDO, for next time.
2680 if (tstate1 != tstate0)
2681 pdata->set_trap_state(tstate1);
2682 } else {
2683 if (LogCompilation && xtty != nullptr) {
2684 ttyLocker ttyl;
2685 // Missing MDP? Leave a small complaint in the log.
2686 xtty->elem("missing_mdp bci='%d'", trap_bci);
2687 }
2688 }
2689 }
2690
2691 // Return results:
2692 ret_this_trap_count = this_trap_count;
2693 ret_maybe_prior_trap = maybe_prior_trap;
2694 ret_maybe_prior_recompile = maybe_prior_recompile;
2695 return pdata;
2696 }
2697
2698 void
2699 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2700 ResourceMark rm;
2701 // Ignored outputs:
2702 uint ignore_this_trap_count;
2703 bool ignore_maybe_prior_trap;
2704 bool ignore_maybe_prior_recompile;
2705 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2706 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2707 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2708
2709 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2710 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2711
2712 query_update_method_data(trap_mdo, trap_bci,
2713 (DeoptReason)reason,
2714 update_total_counts,
2715 #if INCLUDE_JVMCI
2716 false,
2717 #endif
2718 nullptr,
2719 ignore_this_trap_count,
2720 ignore_maybe_prior_trap,
2721 ignore_maybe_prior_recompile);
2722 }
2723
2724 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2725 // Enable WXWrite: current function is called from methods compiled by C2 directly
2726 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2727
2728 // Still in Java no safepoints
2729 {
2730 // This enters VM and may safepoint
2731 uncommon_trap_inner(current, trap_request);
2732 }
2733 HandleMark hm(current);
2734 return fetch_unroll_info_helper(current, exec_mode);
2735 }
2736
2737 // Local derived constants.
2738 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2739 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2740 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2741
2742 //---------------------------trap_state_reason---------------------------------
2743 Deoptimization::DeoptReason
2744 Deoptimization::trap_state_reason(int trap_state) {
2745 // This assert provides the link between the width of DataLayout::trap_bits
2746 // and the encoding of "recorded" reasons. It ensures there are enough
2747 // bits to store all needed reasons in the per-BCI MDO profile.
2748 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2749 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2750 trap_state -= recompile_bit;
2751 if (trap_state == DS_REASON_MASK) {
2752 return Reason_many;
2753 } else {
2754 assert((int)Reason_none == 0, "state=0 => Reason_none");
2755 return (DeoptReason)trap_state;
2756 }
2757 }
2758 //-------------------------trap_state_has_reason-------------------------------
2759 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2760 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2761 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2762 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2763 trap_state -= recompile_bit;
2764 if (trap_state == DS_REASON_MASK) {
2765 return -1; // true, unspecifically (bottom of state lattice)
2766 } else if (trap_state == reason) {
2767 return 1; // true, definitely
2768 } else if (trap_state == 0) {
2769 return 0; // false, definitely (top of state lattice)
2770 } else {
2771 return 0; // false, definitely
2772 }
2773 }
2774 //-------------------------trap_state_add_reason-------------------------------
2775 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2776 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2777 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2778 trap_state -= recompile_bit;
2779 if (trap_state == DS_REASON_MASK) {
2780 return trap_state + recompile_bit; // already at state lattice bottom
2781 } else if (trap_state == reason) {
2782 return trap_state + recompile_bit; // the condition is already true
2783 } else if (trap_state == 0) {
2784 return reason + recompile_bit; // no condition has yet been true
2785 } else {
2786 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2787 }
2788 }
2789 //-----------------------trap_state_is_recompiled------------------------------
2790 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2791 return (trap_state & DS_RECOMPILE_BIT) != 0;
2792 }
2793 //-----------------------trap_state_set_recompiled-----------------------------
2794 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2795 if (z) return trap_state | DS_RECOMPILE_BIT;
2796 else return trap_state & ~DS_RECOMPILE_BIT;
2797 }
2798 //---------------------------format_trap_state---------------------------------
2799 // This is used for debugging and diagnostics, including LogFile output.
2800 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2801 int trap_state) {
2802 assert(buflen > 0, "sanity");
2803 DeoptReason reason = trap_state_reason(trap_state);
2804 bool recomp_flag = trap_state_is_recompiled(trap_state);
2805 // Re-encode the state from its decoded components.
2806 int decoded_state = 0;
2807 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2808 decoded_state = trap_state_add_reason(decoded_state, reason);
2809 if (recomp_flag)
2810 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2811 // If the state re-encodes properly, format it symbolically.
2812 // Because this routine is used for debugging and diagnostics,
2813 // be robust even if the state is a strange value.
2814 size_t len;
2815 if (decoded_state != trap_state) {
2816 // Random buggy state that doesn't decode??
2817 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2818 } else {
2819 len = jio_snprintf(buf, buflen, "%s%s",
2820 trap_reason_name(reason),
2821 recomp_flag ? " recompiled" : "");
2822 }
2823 return buf;
2824 }
2825
2826
2827 //--------------------------------statics--------------------------------------
2828 const char* Deoptimization::_trap_reason_name[] = {
2829 // Note: Keep this in sync. with enum DeoptReason.
2830 "none",
2831 "null_check",
2832 "null_assert" JVMCI_ONLY("_or_unreached0"),
2833 "range_check",
2834 "class_check",
2835 "array_check",
2836 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2837 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2838 "profile_predicate",
2839 "auto_vectorization_check",
2840 "unloaded",
2841 "uninitialized",
2842 "initialized",
2843 "unreached",
2844 "unhandled",
2845 "constraint",
2846 "div0_check",
2847 "age",
2848 "predicate",
2849 "loop_limit_check",
2850 "speculate_class_check",
2851 "speculate_null_check",
2852 "speculate_null_assert",
2853 "unstable_if",
2854 "unstable_fused_if",
2855 "receiver_constraint",
2856 #if INCLUDE_JVMCI
2857 "aliasing",
2858 "transfer_to_interpreter",
2859 "not_compiled_exception_handler",
2860 "unresolved",
2861 "jsr_mismatch",
2862 #endif
2863 "tenured"
2864 };
2865 const char* Deoptimization::_trap_action_name[] = {
2866 // Note: Keep this in sync. with enum DeoptAction.
2867 "none",
2868 "maybe_recompile",
2869 "reinterpret",
2870 "make_not_entrant",
2871 "make_not_compilable"
2872 };
2873
2874 const char* Deoptimization::trap_reason_name(int reason) {
2875 // Check that every reason has a name
2876 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2877
2878 if (reason == Reason_many) return "many";
2879 if ((uint)reason < Reason_LIMIT)
2880 return _trap_reason_name[reason];
2881 static char buf[20];
2882 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2883 return buf;
2884 }
2885 const char* Deoptimization::trap_action_name(int action) {
2886 // Check that every action has a name
2887 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2888
2889 if ((uint)action < Action_LIMIT)
2890 return _trap_action_name[action];
2891 static char buf[20];
2892 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2893 return buf;
2894 }
2895
2896 // This is used for debugging and diagnostics, including LogFile output.
2897 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2898 int trap_request) {
2899 jint unloaded_class_index = trap_request_index(trap_request);
2900 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2901 const char* action = trap_action_name(trap_request_action(trap_request));
2902 #if INCLUDE_JVMCI
2903 int debug_id = trap_request_debug_id(trap_request);
2904 #endif
2905 size_t len;
2906 if (unloaded_class_index < 0) {
2907 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2908 reason, action
2909 #if INCLUDE_JVMCI
2910 ,debug_id
2911 #endif
2912 );
2913 } else {
2914 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2915 reason, action, unloaded_class_index
2916 #if INCLUDE_JVMCI
2917 ,debug_id
2918 #endif
2919 );
2920 }
2921 return buf;
2922 }
2923
2924 juint Deoptimization::_deoptimization_hist
2925 [Deoptimization::Reason_LIMIT]
2926 [1 + Deoptimization::Action_LIMIT]
2927 [Deoptimization::BC_CASE_LIMIT]
2928 = {0};
2929
2930 enum {
2931 LSB_BITS = 8,
2932 LSB_MASK = right_n_bits(LSB_BITS)
2933 };
2934
2935 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2936 Bytecodes::Code bc) {
2937 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2938 assert(action >= 0 && action < Action_LIMIT, "oob");
2939 _deoptimization_hist[Reason_none][0][0] += 1; // total
2940 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2941 juint* cases = _deoptimization_hist[reason][1+action];
2942 juint* bc_counter_addr = nullptr;
2943 juint bc_counter = 0;
2944 // Look for an unused counter, or an exact match to this BC.
2945 if (bc != Bytecodes::_illegal) {
2946 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2947 juint* counter_addr = &cases[bc_case];
2948 juint counter = *counter_addr;
2949 if ((counter == 0 && bc_counter_addr == nullptr)
2950 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2951 // this counter is either free or is already devoted to this BC
2952 bc_counter_addr = counter_addr;
2953 bc_counter = counter | bc;
2954 }
2955 }
2956 }
2957 if (bc_counter_addr == nullptr) {
2958 // Overflow, or no given bytecode.
2959 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2960 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2961 }
2962 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2963 }
2964
2965 jint Deoptimization::total_deoptimization_count() {
2966 return _deoptimization_hist[Reason_none][0][0];
2967 }
2968
2969 // Get the deopt count for a specific reason and a specific action. If either
2970 // one of 'reason' or 'action' is null, the method returns the sum of all
2971 // deoptimizations with the specific 'action' or 'reason' respectively.
2972 // If both arguments are null, the method returns the total deopt count.
2973 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2974 if (reason_str == nullptr && action_str == nullptr) {
2975 return total_deoptimization_count();
2976 }
2977 juint counter = 0;
2978 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2979 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2980 for (int action = 0; action < Action_LIMIT; action++) {
2981 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2982 juint* cases = _deoptimization_hist[reason][1+action];
2983 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2984 counter += cases[bc_case] >> LSB_BITS;
2985 }
2986 }
2987 }
2988 }
2989 }
2990 return counter;
2991 }
2992
2993 void Deoptimization::print_statistics() {
2994 juint total = total_deoptimization_count();
2995 juint account = total;
2996 if (total != 0) {
2997 ttyLocker ttyl;
2998 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2999 tty->print_cr("Deoptimization traps recorded:");
3000 #define PRINT_STAT_LINE(name, r) \
3001 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
3002 PRINT_STAT_LINE("total", total);
3003 // For each non-zero entry in the histogram, print the reason,
3004 // the action, and (if specifically known) the type of bytecode.
3005 for (int reason = 0; reason < Reason_LIMIT; reason++) {
3006 for (int action = 0; action < Action_LIMIT; action++) {
3007 juint* cases = _deoptimization_hist[reason][1+action];
3008 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
3009 juint counter = cases[bc_case];
3010 if (counter != 0) {
3011 char name[1*K];
3012 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
3013 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
3014 bc = Bytecodes::_illegal;
3015 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
3016 trap_reason_name(reason),
3017 trap_action_name(action),
3018 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
3019 juint r = counter >> LSB_BITS;
3020 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
3021 account -= r;
3022 }
3023 }
3024 }
3025 }
3026 if (account != 0) {
3027 PRINT_STAT_LINE("unaccounted", account);
3028 }
3029 #undef PRINT_STAT_LINE
3030 if (xtty != nullptr) xtty->tail("statistics");
3031 }
3032 }
3033
3034 #else // COMPILER2_OR_JVMCI
3035
3036
3037 // Stubs for C1 only system.
3038 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3039 return false;
3040 }
3041
3042 const char* Deoptimization::trap_reason_name(int reason) {
3043 return "unknown";
3044 }
3045
3046 jint Deoptimization::total_deoptimization_count() {
3047 return 0;
3048 }
3049
3050 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3051 return 0;
3052 }
3053
3054 void Deoptimization::print_statistics() {
3055 // no output
3056 }
3057
3058 void
3059 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3060 // no update
3061 }
3062
3063 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3064 return 0;
3065 }
3066
3067 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3068 Bytecodes::Code bc) {
3069 // no update
3070 }
3071
3072 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3073 int trap_state) {
3074 jio_snprintf(buf, buflen, "#%d", trap_state);
3075 return buf;
3076 }
3077
3078 #endif // COMPILER2_OR_JVMCI