1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmClasses.hpp"
29 #include "code/codeCache.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/nmethod.hpp"
32 #include "code/pcDesc.hpp"
33 #include "code/scopeDesc.hpp"
34 #include "compiler/compilationPolicy.hpp"
35 #include "compiler/compilerDefinitions.inline.hpp"
36 #include "gc/shared/collectedHeap.hpp"
37 #include "gc/shared/memAllocator.hpp"
38 #include "interpreter/bytecode.hpp"
39 #include "interpreter/bytecode.inline.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/method.hpp"
55 #include "oops/objArrayKlass.hpp"
56 #include "oops/objArrayOop.inline.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "oops/typeArrayOop.inline.hpp"
59 #include "oops/verifyOopClosure.hpp"
60 #include "prims/jvmtiDeferredUpdates.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "prims/jvmtiThreadState.hpp"
63 #include "prims/methodHandles.hpp"
64 #include "prims/vectorSupport.hpp"
65 #include "runtime/atomic.hpp"
66 #include "runtime/basicLock.inline.hpp"
67 #include "runtime/continuation.hpp"
68 #include "runtime/continuationEntry.inline.hpp"
69 #include "runtime/deoptimization.hpp"
70 #include "runtime/escapeBarrier.hpp"
71 #include "runtime/fieldDescriptor.hpp"
72 #include "runtime/fieldDescriptor.inline.hpp"
73 #include "runtime/frame.inline.hpp"
74 #include "runtime/handles.inline.hpp"
75 #include "runtime/interfaceSupport.inline.hpp"
76 #include "runtime/javaThread.hpp"
77 #include "runtime/jniHandles.inline.hpp"
78 #include "runtime/keepStackGCProcessed.hpp"
79 #include "runtime/lightweightSynchronizer.hpp"
80 #include "runtime/lockStack.inline.hpp"
81 #include "runtime/objectMonitor.inline.hpp"
82 #include "runtime/osThread.hpp"
83 #include "runtime/safepointVerifiers.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stackFrameStream.inline.hpp"
87 #include "runtime/stackValue.hpp"
88 #include "runtime/stackWatermarkSet.hpp"
89 #include "runtime/stubRoutines.hpp"
90 #include "runtime/synchronizer.inline.hpp"
91 #include "runtime/threadSMR.hpp"
92 #include "runtime/threadWXSetters.inline.hpp"
93 #include "runtime/vframe.hpp"
94 #include "runtime/vframeArray.hpp"
95 #include "runtime/vframe_hp.hpp"
96 #include "runtime/vmOperations.hpp"
97 #include "utilities/checkedCast.hpp"
98 #include "utilities/events.hpp"
99 #include "utilities/growableArray.hpp"
100 #include "utilities/macros.hpp"
101 #include "utilities/preserveException.hpp"
102 #include "utilities/xmlstream.hpp"
103 #if INCLUDE_JFR
104 #include "jfr/jfr.inline.hpp"
105 #include "jfr/jfrEvents.hpp"
106 #include "jfr/metadata/jfrSerializer.hpp"
107 #endif
108
109 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
110 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
111 bool DeoptimizationScope::_committing_in_progress = false;
112
113 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
114 DEBUG_ONLY(_deopted = false;)
115
116 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
117 // If there is nothing to deopt _required_gen is the same as comitted.
118 _required_gen = DeoptimizationScope::_committed_deopt_gen;
119 }
120
121 DeoptimizationScope::~DeoptimizationScope() {
122 assert(_deopted, "Deopt not executed");
123 }
124
125 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
126 if (!nm->can_be_deoptimized()) {
127 return;
128 }
129
130 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
131
132 // If it's already marked but we still need it to be deopted.
133 if (nm->is_marked_for_deoptimization()) {
134 dependent(nm);
135 return;
136 }
137
138 nmethod::DeoptimizationStatus status =
139 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
140 Atomic::store(&nm->_deoptimization_status, status);
141
142 // Make sure active is not committed
143 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
144 assert(nm->_deoptimization_generation == 0, "Is already marked");
145
146 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
147 _required_gen = DeoptimizationScope::_active_deopt_gen;
148 }
149
150 void DeoptimizationScope::dependent(nmethod* nm) {
151 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
152
153 // A method marked by someone else may have a _required_gen lower than what we marked with.
154 // Therefore only store it if it's higher than _required_gen.
155 if (_required_gen < nm->_deoptimization_generation) {
156 _required_gen = nm->_deoptimization_generation;
157 }
158 }
159
160 void DeoptimizationScope::deoptimize_marked() {
161 assert(!_deopted, "Already deopted");
162
163 // We are not alive yet.
164 if (!Universe::is_fully_initialized()) {
165 DEBUG_ONLY(_deopted = true;)
166 return;
167 }
168
169 // Safepoints are a special case, handled here.
170 if (SafepointSynchronize::is_at_safepoint()) {
171 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
172 DeoptimizationScope::_active_deopt_gen++;
173 Deoptimization::deoptimize_all_marked();
174 DEBUG_ONLY(_deopted = true;)
175 return;
176 }
177
178 uint64_t comitting = 0;
179 bool wait = false;
180 while (true) {
181 {
182 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
183
184 // First we check if we or someone else already deopted the gen we want.
185 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
186 DEBUG_ONLY(_deopted = true;)
187 return;
188 }
189 if (!_committing_in_progress) {
190 // The version we are about to commit.
191 comitting = DeoptimizationScope::_active_deopt_gen;
192 // Make sure new marks use a higher gen.
193 DeoptimizationScope::_active_deopt_gen++;
194 _committing_in_progress = true;
195 wait = false;
196 } else {
197 // Another thread is handshaking and committing a gen.
198 wait = true;
199 }
200 }
201 if (wait) {
202 // Wait and let the concurrent handshake be performed.
203 ThreadBlockInVM tbivm(JavaThread::current());
204 os::naked_yield();
205 } else {
206 // Performs the handshake.
207 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
208 DEBUG_ONLY(_deopted = true;)
209 {
210 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
211
212 // Make sure that committed doesn't go backwards.
213 // Should only happen if we did a deopt during a safepoint above.
214 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
215 DeoptimizationScope::_committed_deopt_gen = comitting;
216 }
217 _committing_in_progress = false;
218
219 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
220
221 return;
222 }
223 }
224 }
225 }
226
227 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
228 int caller_adjustment,
229 int caller_actual_parameters,
230 int number_of_frames,
231 intptr_t* frame_sizes,
232 address* frame_pcs,
233 BasicType return_type,
234 int exec_mode) {
235 _size_of_deoptimized_frame = size_of_deoptimized_frame;
236 _caller_adjustment = caller_adjustment;
237 _caller_actual_parameters = caller_actual_parameters;
238 _number_of_frames = number_of_frames;
239 _frame_sizes = frame_sizes;
240 _frame_pcs = frame_pcs;
241 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
242 _return_type = return_type;
243 _initial_info = 0;
244 // PD (x86 only)
245 _counter_temp = 0;
246 _unpack_kind = exec_mode;
247 _sender_sp_temp = 0;
248
249 _total_frame_sizes = size_of_frames();
250 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
251 }
252
253 Deoptimization::UnrollBlock::~UnrollBlock() {
254 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
255 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
256 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
257 }
258
259 int Deoptimization::UnrollBlock::size_of_frames() const {
260 // Account first for the adjustment of the initial frame
261 intptr_t result = _caller_adjustment;
262 for (int index = 0; index < number_of_frames(); index++) {
263 result += frame_sizes()[index];
264 }
265 return checked_cast<int>(result);
266 }
267
268 void Deoptimization::UnrollBlock::print() {
269 ResourceMark rm;
270 stringStream st;
271 st.print_cr("UnrollBlock");
272 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
273 st.print( " frame_sizes: ");
274 for (int index = 0; index < number_of_frames(); index++) {
275 st.print("%zd ", frame_sizes()[index]);
276 }
277 st.cr();
278 tty->print_raw(st.freeze());
279 }
280
281 // In order to make fetch_unroll_info work properly with escape
282 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
283 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
284 // which is called from the method fetch_unroll_info_helper below.
285 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
286 // fetch_unroll_info() is called at the beginning of the deoptimization
287 // handler. Note this fact before we start generating temporary frames
288 // that can confuse an asynchronous stack walker. This counter is
289 // decremented at the end of unpack_frames().
290 current->inc_in_deopt_handler();
291
292 if (exec_mode == Unpack_exception) {
293 // When we get here, a callee has thrown an exception into a deoptimized
294 // frame. That throw might have deferred stack watermark checking until
295 // after unwinding. So we deal with such deferred requests here.
296 StackWatermarkSet::after_unwind(current);
297 }
298
299 return fetch_unroll_info_helper(current, exec_mode);
300 JRT_END
301
302 #if COMPILER2_OR_JVMCI
303 // print information about reallocated objects
304 static void print_objects(JavaThread* deoptee_thread,
305 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
306 ResourceMark rm;
307 stringStream st; // change to logStream with logging
308 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
309 fieldDescriptor fd;
310
311 for (int i = 0; i < objects->length(); i++) {
312 ObjectValue* sv = (ObjectValue*) objects->at(i);
313 Handle obj = sv->value();
314
315 if (obj.is_null()) {
316 st.print_cr(" nullptr");
317 continue;
318 }
319
320 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
321
322 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
323 k->print_value_on(&st);
324 st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize);
325
326 if (Verbose && k != nullptr) {
327 k->oop_print_on(obj(), &st);
328 }
329 }
330 tty->print_raw(st.freeze());
331 }
332
333 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
334 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
335 bool& deoptimized_objects) {
336 bool realloc_failures = false;
337 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
338
339 JavaThread* deoptee_thread = chunk->at(0)->thread();
340 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
341 "a frame can only be deoptimized by the owner thread");
342
343 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
344
345 // The flag return_oop() indicates call sites which return oop
346 // in compiled code. Such sites include java method calls,
347 // runtime calls (for example, used to allocate new objects/arrays
348 // on slow code path) and any other calls generated in compiled code.
349 // It is not guaranteed that we can get such information here only
350 // by analyzing bytecode in deoptimized frames. This is why this flag
351 // is set during method compilation (see Compile::Process_OopMap_Node()).
352 // If the previous frame was popped or if we are dispatching an exception,
353 // we don't have an oop result.
354 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
355 Handle return_value;
356 if (save_oop_result) {
357 // Reallocation may trigger GC. If deoptimization happened on return from
358 // call which returns oop we need to save it since it is not in oopmap.
359 oop result = deoptee.saved_oop_result(&map);
360 assert(oopDesc::is_oop_or_null(result), "must be oop");
361 return_value = Handle(thread, result);
362 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
363 if (TraceDeoptimization) {
364 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
365 tty->cr();
366 }
367 }
368 if (objects != nullptr) {
369 if (exec_mode == Deoptimization::Unpack_none) {
370 assert(thread->thread_state() == _thread_in_vm, "assumption");
371 JavaThread* THREAD = thread; // For exception macros.
372 // Clear pending OOM if reallocation fails and return true indicating allocation failure
373 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
374 deoptimized_objects = true;
375 } else {
376 JavaThread* current = thread; // For JRT_BLOCK
377 JRT_BLOCK
378 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
379 JRT_END
380 }
381 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
382 bool is_jvmci = compiled_method->is_compiled_by_jvmci();
383 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci);
384 if (TraceDeoptimization) {
385 print_objects(deoptee_thread, objects, realloc_failures);
386 }
387 }
388 if (save_oop_result) {
389 // Restore result.
390 deoptee.set_saved_oop_result(&map, return_value());
391 }
392 return realloc_failures;
393 }
394
395 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
396 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
397 JavaThread* deoptee_thread = chunk->at(0)->thread();
398 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
399 assert(thread == Thread::current(), "should be");
400 HandleMark hm(thread);
401 #ifndef PRODUCT
402 bool first = true;
403 #endif // !PRODUCT
404 // Start locking from outermost/oldest frame
405 for (int i = (chunk->length() - 1); i >= 0; i--) {
406 compiledVFrame* cvf = chunk->at(i);
407 assert (cvf->scope() != nullptr,"expect only compiled java frames");
408 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
409 if (monitors->is_nonempty()) {
410 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
411 exec_mode, realloc_failures);
412 deoptimized_objects = deoptimized_objects || relocked;
413 #ifndef PRODUCT
414 if (PrintDeoptimizationDetails) {
415 ResourceMark rm;
416 stringStream st;
417 for (int j = 0; j < monitors->length(); j++) {
418 MonitorInfo* mi = monitors->at(j);
419 if (mi->eliminated()) {
420 if (first) {
421 first = false;
422 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
423 }
424 if (exec_mode == Deoptimization::Unpack_none) {
425 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
426 if (monitor != nullptr && monitor->object() == mi->owner()) {
427 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
428 continue;
429 }
430 }
431 if (mi->owner_is_scalar_replaced()) {
432 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
433 st.print_cr(" failed reallocation for klass %s", k->external_name());
434 } else {
435 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
436 }
437 }
438 }
439 tty->print_raw(st.freeze());
440 }
441 #endif // !PRODUCT
442 }
443 }
444 }
445
446 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
447 // The given vframes cover one physical frame.
448 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
449 bool& realloc_failures) {
450 frame deoptee = chunk->at(0)->fr();
451 JavaThread* deoptee_thread = chunk->at(0)->thread();
452 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
453 RegisterMap map(chunk->at(0)->register_map());
454 bool deoptimized_objects = false;
455
456 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
457
458 // Reallocate the non-escaping objects and restore their fields.
459 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
460 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
461 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
462 }
463
464 // MonitorInfo structures used in eliminate_locks are not GC safe.
465 NoSafepointVerifier no_safepoint;
466
467 // Now relock objects if synchronization on them was eliminated.
468 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
469 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
470 }
471 return deoptimized_objects;
472 }
473 #endif // COMPILER2_OR_JVMCI
474
475 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
476 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
477 JFR_ONLY(Jfr::check_and_process_sample_request(current);)
478 // When we get here we are about to unwind the deoptee frame. In order to
479 // catch not yet safe to use frames, the following stack watermark barrier
480 // poll will make such frames safe to use.
481 StackWatermarkSet::before_unwind(current);
482
483 // Note: there is a safepoint safety issue here. No matter whether we enter
484 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
485 // the vframeArray is created.
486 //
487
488 // Allocate our special deoptimization ResourceMark
489 DeoptResourceMark* dmark = new DeoptResourceMark(current);
490 assert(current->deopt_mark() == nullptr, "Pending deopt!");
491 current->set_deopt_mark(dmark);
492
493 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
494 RegisterMap map(current,
495 RegisterMap::UpdateMap::include,
496 RegisterMap::ProcessFrames::include,
497 RegisterMap::WalkContinuation::skip);
498 RegisterMap dummy_map(current,
499 RegisterMap::UpdateMap::skip,
500 RegisterMap::ProcessFrames::include,
501 RegisterMap::WalkContinuation::skip);
502 // Now get the deoptee with a valid map
503 frame deoptee = stub_frame.sender(&map);
504 // Set the deoptee nmethod
505 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
506 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
507 current->set_deopt_compiled_method(nm);
508
509 if (VerifyStack) {
510 current->validate_frame_layout();
511 }
512
513 // Create a growable array of VFrames where each VFrame represents an inlined
514 // Java frame. This storage is allocated with the usual system arena.
515 assert(deoptee.is_compiled_frame(), "Wrong frame type");
516 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
517 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
518 while (!vf->is_top()) {
519 assert(vf->is_compiled_frame(), "Wrong frame type");
520 chunk->push(compiledVFrame::cast(vf));
521 vf = vf->sender();
522 }
523 assert(vf->is_compiled_frame(), "Wrong frame type");
524 chunk->push(compiledVFrame::cast(vf));
525
526 bool realloc_failures = false;
527
528 #if COMPILER2_OR_JVMCI
529 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
530
531 // Reallocate the non-escaping objects and restore their fields. Then
532 // relock objects if synchronization on them was eliminated.
533 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
534 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
535 bool unused;
536 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
537 }
538 #endif // COMPILER2_OR_JVMCI
539
540 // Ensure that no safepoint is taken after pointers have been stored
541 // in fields of rematerialized objects. If a safepoint occurs from here on
542 // out the java state residing in the vframeArray will be missed.
543 // Locks may be rebaised in a safepoint.
544 NoSafepointVerifier no_safepoint;
545
546 #if COMPILER2_OR_JVMCI
547 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
548 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
549 bool unused = false;
550 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
551 }
552 #endif // COMPILER2_OR_JVMCI
553
554 ScopeDesc* trap_scope = chunk->at(0)->scope();
555 Handle exceptionObject;
556 if (trap_scope->rethrow_exception()) {
557 #ifndef PRODUCT
558 if (PrintDeoptimizationDetails) {
559 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());
560 }
561 #endif // !PRODUCT
562
563 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
564 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
565 ScopeValue* topOfStack = expressions->top();
566 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
567 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
568 }
569
570 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
571 #if COMPILER2_OR_JVMCI
572 if (realloc_failures) {
573 // This destroys all ScopedValue bindings.
574 current->clear_scopedValueBindings();
575 pop_frames_failed_reallocs(current, array);
576 }
577 #endif
578
579 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
580 current->set_vframe_array_head(array);
581
582 // Now that the vframeArray has been created if we have any deferred local writes
583 // added by jvmti then we can free up that structure as the data is now in the
584 // vframeArray
585
586 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
587
588 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
589 CodeBlob* cb = stub_frame.cb();
590 // Verify we have the right vframeArray
591 assert(cb->frame_size() >= 0, "Unexpected frame size");
592 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
593
594 // If the deopt call site is a MethodHandle invoke call site we have
595 // to adjust the unpack_sp.
596 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
597 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
598 unpack_sp = deoptee.unextended_sp();
599
600 #ifdef ASSERT
601 assert(cb->is_deoptimization_stub() ||
602 cb->is_uncommon_trap_stub() ||
603 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
604 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
605 "unexpected code blob: %s", cb->name());
606 #endif
607
608 // This is a guarantee instead of an assert because if vframe doesn't match
609 // we will unpack the wrong deoptimized frame and wind up in strange places
610 // where it will be very difficult to figure out what went wrong. Better
611 // to die an early death here than some very obscure death later when the
612 // trail is cold.
613 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
614
615 int number_of_frames = array->frames();
616
617 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
618 // virtual activation, which is the reverse of the elements in the vframes array.
619 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
620 // +1 because we always have an interpreter return address for the final slot.
621 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
622 int popframe_extra_args = 0;
623 // Create an interpreter return address for the stub to use as its return
624 // address so the skeletal frames are perfectly walkable
625 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
626
627 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
628 // activation be put back on the expression stack of the caller for reexecution
629 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
630 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
631 }
632
633 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
634 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
635 // than simply use array->sender.pc(). This requires us to walk the current set of frames
636 //
637 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
638 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
639
640 // It's possible that the number of parameters at the call site is
641 // different than number of arguments in the callee when method
642 // handles are used. If the caller is interpreted get the real
643 // value so that the proper amount of space can be added to it's
644 // frame.
645 bool caller_was_method_handle = false;
646 if (deopt_sender.is_interpreted_frame()) {
647 methodHandle method(current, deopt_sender.interpreter_frame_method());
648 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
649 if (cur.has_member_arg()) {
650 // This should cover all real-world cases. One exception is a pathological chain of
651 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we
652 // would need to get the size from the resolved method entry. Another exception would
653 // be an invokedynamic with an adapter that is really a MethodHandle linker.
654 caller_was_method_handle = true;
655 }
656 }
657
658 //
659 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
660 // frame_sizes/frame_pcs[1] next oldest frame (int)
661 // frame_sizes/frame_pcs[n] youngest frame (int)
662 //
663 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
664 // owns the space for the return address to it's caller). Confusing ain't it.
665 //
666 // The vframe array can address vframes with indices running from
667 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
668 // When we create the skeletal frames we need the oldest frame to be in the zero slot
669 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
670 // so things look a little strange in this loop.
671 //
672 int callee_parameters = 0;
673 int callee_locals = 0;
674 for (int index = 0; index < array->frames(); index++ ) {
675 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
676 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
677 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
678 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
679 callee_locals,
680 index == 0,
681 popframe_extra_args);
682 // This pc doesn't have to be perfect just good enough to identify the frame
683 // as interpreted so the skeleton frame will be walkable
684 // The correct pc will be set when the skeleton frame is completely filled out
685 // The final pc we store in the loop is wrong and will be overwritten below
686 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
687
688 callee_parameters = array->element(index)->method()->size_of_parameters();
689 callee_locals = array->element(index)->method()->max_locals();
690 popframe_extra_args = 0;
691 }
692
693 // Compute whether the root vframe returns a float or double value.
694 BasicType return_type;
695 {
696 methodHandle method(current, array->element(0)->method());
697 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
698 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
699 }
700
701 // Compute information for handling adapters and adjusting the frame size of the caller.
702 int caller_adjustment = 0;
703
704 // Compute the amount the oldest interpreter frame will have to adjust
705 // its caller's stack by. If the caller is a compiled frame then
706 // we pretend that the callee has no parameters so that the
707 // extension counts for the full amount of locals and not just
708 // locals-parms. This is because without a c2i adapter the parm
709 // area as created by the compiled frame will not be usable by
710 // the interpreter. (Depending on the calling convention there
711 // may not even be enough space).
712
713 // QQQ I'd rather see this pushed down into last_frame_adjust
714 // and have it take the sender (aka caller).
715
716 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
717 caller_adjustment = last_frame_adjust(0, callee_locals);
718 } else if (callee_locals > callee_parameters) {
719 // The caller frame may need extending to accommodate
720 // non-parameter locals of the first unpacked interpreted frame.
721 // Compute that adjustment.
722 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
723 }
724
725 // If the sender is deoptimized the we must retrieve the address of the handler
726 // since the frame will "magically" show the original pc before the deopt
727 // and we'd undo the deopt.
728
729 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
730 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
731 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
732 }
733
734 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
735
736 #if INCLUDE_JVMCI
737 if (exceptionObject() != nullptr) {
738 current->set_exception_oop(exceptionObject());
739 exec_mode = Unpack_exception;
740 }
741 #endif
742
743 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
744 assert(current->has_pending_exception(), "should have thrown OOME");
745 current->set_exception_oop(current->pending_exception());
746 current->clear_pending_exception();
747 exec_mode = Unpack_exception;
748 }
749
750 #if INCLUDE_JVMCI
751 if (current->frames_to_pop_failed_realloc() > 0) {
752 current->set_pending_monitorenter(false);
753 }
754 #endif
755
756 int caller_actual_parameters = -1; // value not used except for interpreted frames, see below
757 if (deopt_sender.is_interpreted_frame()) {
758 caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0);
759 }
760
761 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
762 caller_adjustment * BytesPerWord,
763 caller_actual_parameters,
764 number_of_frames,
765 frame_sizes,
766 frame_pcs,
767 return_type,
768 exec_mode);
769 // On some platforms, we need a way to pass some platform dependent
770 // information to the unpacking code so the skeletal frames come out
771 // correct (initial fp value, unextended sp, ...)
772 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
773
774 if (array->frames() > 1) {
775 if (VerifyStack && TraceDeoptimization) {
776 tty->print_cr("Deoptimizing method containing inlining");
777 }
778 }
779
780 array->set_unroll_block(info);
781 return info;
782 }
783
784 // Called to cleanup deoptimization data structures in normal case
785 // after unpacking to stack and when stack overflow error occurs
786 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
787 vframeArray *array) {
788
789 // Get array if coming from exception
790 if (array == nullptr) {
791 array = thread->vframe_array_head();
792 }
793 thread->set_vframe_array_head(nullptr);
794
795 // Free the previous UnrollBlock
796 vframeArray* old_array = thread->vframe_array_last();
797 thread->set_vframe_array_last(array);
798
799 if (old_array != nullptr) {
800 UnrollBlock* old_info = old_array->unroll_block();
801 old_array->set_unroll_block(nullptr);
802 delete old_info;
803 delete old_array;
804 }
805
806 // Deallocate any resource creating in this routine and any ResourceObjs allocated
807 // inside the vframeArray (StackValueCollections)
808
809 delete thread->deopt_mark();
810 thread->set_deopt_mark(nullptr);
811 thread->set_deopt_compiled_method(nullptr);
812
813
814 if (JvmtiExport::can_pop_frame()) {
815 // Regardless of whether we entered this routine with the pending
816 // popframe condition bit set, we should always clear it now
817 thread->clear_popframe_condition();
818 }
819
820 // unpack_frames() is called at the end of the deoptimization handler
821 // and (in C2) at the end of the uncommon trap handler. Note this fact
822 // so that an asynchronous stack walker can work again. This counter is
823 // incremented at the beginning of fetch_unroll_info() and (in C2) at
824 // the beginning of uncommon_trap().
825 thread->dec_in_deopt_handler();
826 }
827
828 // Moved from cpu directories because none of the cpus has callee save values.
829 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
830 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
831
832 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
833 // the days we had adapter frames. When we deoptimize a situation where a
834 // compiled caller calls a compiled caller will have registers it expects
835 // to survive the call to the callee. If we deoptimize the callee the only
836 // way we can restore these registers is to have the oldest interpreter
837 // frame that we create restore these values. That is what this routine
838 // will accomplish.
839
840 // At the moment we have modified c2 to not have any callee save registers
841 // so this problem does not exist and this routine is just a place holder.
842
843 assert(f->is_interpreted_frame(), "must be interpreted");
844 }
845
846 #ifndef PRODUCT
847 // Return true if the execution after the provided bytecode continues at the
848 // next bytecode in the code. This is not the case for gotos, returns, and
849 // throws.
850 static bool falls_through(Bytecodes::Code bc) {
851 switch (bc) {
852 case Bytecodes::_goto:
853 case Bytecodes::_goto_w:
854 case Bytecodes::_athrow:
855 case Bytecodes::_areturn:
856 case Bytecodes::_dreturn:
857 case Bytecodes::_freturn:
858 case Bytecodes::_ireturn:
859 case Bytecodes::_lreturn:
860 case Bytecodes::_jsr:
861 case Bytecodes::_ret:
862 case Bytecodes::_return:
863 case Bytecodes::_lookupswitch:
864 case Bytecodes::_tableswitch:
865 return false;
866 default:
867 return true;
868 }
869 }
870 #endif
871
872 // Return BasicType of value being returned
873 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
874 assert(thread == JavaThread::current(), "pre-condition");
875
876 // We are already active in the special DeoptResourceMark any ResourceObj's we
877 // allocate will be freed at the end of the routine.
878
879 // JRT_LEAF methods don't normally allocate handles and there is a
880 // NoHandleMark to enforce that. It is actually safe to use Handles
881 // in a JRT_LEAF method, and sometimes desirable, but to do so we
882 // must use ResetNoHandleMark to bypass the NoHandleMark, and
883 // then use a HandleMark to ensure any Handles we do create are
884 // cleaned up in this scope.
885 ResetNoHandleMark rnhm;
886 HandleMark hm(thread);
887
888 frame stub_frame = thread->last_frame();
889
890 Continuation::notify_deopt(thread, stub_frame.sp());
891
892 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
893 // must point to the vframeArray for the unpack frame.
894 vframeArray* array = thread->vframe_array_head();
895 UnrollBlock* info = array->unroll_block();
896
897 // We set the last_Java frame. But the stack isn't really parsable here. So we
898 // clear it to make sure JFR understands not to try and walk stacks from events
899 // in here.
900 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
901 thread->frame_anchor()->set_last_Java_sp(nullptr);
902
903 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
904 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
905
906 thread->frame_anchor()->set_last_Java_sp(sp);
907
908 BasicType bt = info->return_type();
909
910 // If we have an exception pending, claim that the return type is an oop
911 // so the deopt_blob does not overwrite the exception_oop.
912
913 if (exec_mode == Unpack_exception)
914 bt = T_OBJECT;
915
916 // Cleanup thread deopt data
917 cleanup_deopt_info(thread, array);
918
919 #ifndef PRODUCT
920 if (VerifyStack) {
921 ResourceMark res_mark;
922 // Clear pending exception to not break verification code (restored afterwards)
923 PreserveExceptionMark pm(thread);
924
925 thread->validate_frame_layout();
926
927 // Verify that the just-unpacked frames match the interpreter's
928 // notions of expression stack and locals
929 vframeArray* cur_array = thread->vframe_array_last();
930 RegisterMap rm(thread,
931 RegisterMap::UpdateMap::skip,
932 RegisterMap::ProcessFrames::include,
933 RegisterMap::WalkContinuation::skip);
934 rm.set_include_argument_oops(false);
935 bool is_top_frame = true;
936 int callee_size_of_parameters = 0;
937 int callee_max_locals = 0;
938 for (int i = 0; i < cur_array->frames(); i++) {
939 vframeArrayElement* el = cur_array->element(i);
940 frame* iframe = el->iframe();
941 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
942
943 // Get the oop map for this bci
944 InterpreterOopMap mask;
945 int cur_invoke_parameter_size = 0;
946 bool try_next_mask = false;
947 int next_mask_expression_stack_size = -1;
948 int top_frame_expression_stack_adjustment = 0;
949 methodHandle mh(thread, iframe->interpreter_frame_method());
950 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
951 BytecodeStream str(mh, iframe->interpreter_frame_bci());
952 int max_bci = mh->code_size();
953 // Get to the next bytecode if possible
954 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
955 // Check to see if we can grab the number of outgoing arguments
956 // at an uncommon trap for an invoke (where the compiler
957 // generates debug info before the invoke has executed)
958 Bytecodes::Code cur_code = str.next();
959 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
960 if (Bytecodes::is_invoke(cur_code)) {
961 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
962 cur_invoke_parameter_size = invoke.size_of_parameters();
963 if (i != 0 && invoke.has_member_arg()) {
964 callee_size_of_parameters++;
965 }
966 }
967 if (str.bci() < max_bci) {
968 next_code = str.next();
969 if (next_code >= 0) {
970 // The interpreter oop map generator reports results before
971 // the current bytecode has executed except in the case of
972 // calls. It seems to be hard to tell whether the compiler
973 // has emitted debug information matching the "state before"
974 // a given bytecode or the state after, so we try both
975 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
976 // Get expression stack size for the next bytecode
977 InterpreterOopMap next_mask;
978 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
979 next_mask_expression_stack_size = next_mask.expression_stack_size();
980 if (Bytecodes::is_invoke(next_code)) {
981 Bytecode_invoke invoke(mh, str.bci());
982 next_mask_expression_stack_size += invoke.size_of_parameters();
983 }
984 // Need to subtract off the size of the result type of
985 // the bytecode because this is not described in the
986 // debug info but returned to the interpreter in the TOS
987 // caching register
988 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
989 if (bytecode_result_type != T_ILLEGAL) {
990 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
991 }
992 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
993 try_next_mask = true;
994 }
995 }
996 }
997
998 // Verify stack depth and oops in frame
999 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
1000 if (!(
1001 /* SPARC */
1002 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
1003 /* x86 */
1004 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
1005 (try_next_mask &&
1006 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
1007 top_frame_expression_stack_adjustment))) ||
1008 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
1009 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
1010 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
1011 )) {
1012 {
1013 // Print out some information that will help us debug the problem
1014 tty->print_cr("Wrong number of expression stack elements during deoptimization");
1015 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
1016 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1017 if (try_next_mask) {
1018 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1019 }
1020 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1021 iframe->interpreter_frame_expression_stack_size());
1022 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1023 tty->print_cr(" try_next_mask = %d", try_next_mask);
1024 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1025 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1026 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1027 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1028 tty->print_cr(" exec_mode = %d", exec_mode);
1029 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1030 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1031 tty->print_cr(" Interpreted frames:");
1032 for (int k = 0; k < cur_array->frames(); k++) {
1033 vframeArrayElement* el = cur_array->element(k);
1034 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1035 }
1036 cur_array->print_on_2(tty);
1037 }
1038 guarantee(false, "wrong number of expression stack elements during deopt");
1039 }
1040 VerifyOopClosure verify;
1041 iframe->oops_interpreted_do(&verify, &rm, false);
1042 callee_size_of_parameters = mh->size_of_parameters();
1043 callee_max_locals = mh->max_locals();
1044 is_top_frame = false;
1045 }
1046 }
1047 #endif // !PRODUCT
1048
1049 return bt;
1050 JRT_END
1051
1052 class DeoptimizeMarkedClosure : public HandshakeClosure {
1053 public:
1054 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1055 void do_thread(Thread* thread) {
1056 JavaThread* jt = JavaThread::cast(thread);
1057 jt->deoptimize_marked_methods();
1058 }
1059 };
1060
1061 void Deoptimization::deoptimize_all_marked() {
1062 ResourceMark rm;
1063
1064 // Make the dependent methods not entrant
1065 CodeCache::make_marked_nmethods_deoptimized();
1066
1067 DeoptimizeMarkedClosure deopt;
1068 if (SafepointSynchronize::is_at_safepoint()) {
1069 Threads::java_threads_do(&deopt);
1070 } else {
1071 Handshake::execute(&deopt);
1072 }
1073 }
1074
1075 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1076 = Deoptimization::Action_reinterpret;
1077
1078 #if INCLUDE_JVMCI
1079 template<typename CacheType>
1080 class BoxCacheBase : public CHeapObj<mtCompiler> {
1081 protected:
1082 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1083 ResourceMark rm(thread);
1084 char* klass_name_str = klass_name->as_C_string();
1085 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1086 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1087 if (!ik->is_in_error_state()) {
1088 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1089 CacheType::compute_offsets(ik);
1090 }
1091 return ik;
1092 }
1093 };
1094
1095 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1096 PrimitiveType _low;
1097 PrimitiveType _high;
1098 jobject _cache;
1099 protected:
1100 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1101 BoxCache(Thread* thread) {
1102 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1103 if (ik->is_in_error_state()) {
1104 _low = 1;
1105 _high = 0;
1106 _cache = nullptr;
1107 } else {
1108 objArrayOop cache = CacheType::cache(ik);
1109 assert(cache->length() > 0, "Empty cache");
1110 _low = BoxType::value(cache->obj_at(0));
1111 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1112 _cache = JNIHandles::make_global(Handle(thread, cache));
1113 }
1114 }
1115 ~BoxCache() {
1116 JNIHandles::destroy_global(_cache);
1117 }
1118 public:
1119 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1120 if (_singleton == nullptr) {
1121 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1122 if (!Atomic::replace_if_null(&_singleton, s)) {
1123 delete s;
1124 }
1125 }
1126 return _singleton;
1127 }
1128 oop lookup(PrimitiveType value) {
1129 if (_low <= value && value <= _high) {
1130 int offset = checked_cast<int>(value - _low);
1131 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1132 }
1133 return nullptr;
1134 }
1135 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1136 if (_cache == nullptr) {
1137 cache_init_error = true;
1138 return nullptr;
1139 }
1140 // Have to cast to avoid little/big-endian problems.
1141 if (sizeof(PrimitiveType) > sizeof(jint)) {
1142 jlong value = (jlong)raw_value;
1143 return lookup(value);
1144 }
1145 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1146 return lookup(value);
1147 }
1148 };
1149
1150 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1151 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1152 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1153 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1154 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1155
1156 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1157 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1158 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1159 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1160 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1161
1162 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1163 jobject _true_cache;
1164 jobject _false_cache;
1165 protected:
1166 static BooleanBoxCache *_singleton;
1167 BooleanBoxCache(Thread *thread) {
1168 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1169 if (ik->is_in_error_state()) {
1170 _true_cache = nullptr;
1171 _false_cache = nullptr;
1172 } else {
1173 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1174 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1175 }
1176 }
1177 ~BooleanBoxCache() {
1178 JNIHandles::destroy_global(_true_cache);
1179 JNIHandles::destroy_global(_false_cache);
1180 }
1181 public:
1182 static BooleanBoxCache* singleton(Thread* thread) {
1183 if (_singleton == nullptr) {
1184 BooleanBoxCache* s = new BooleanBoxCache(thread);
1185 if (!Atomic::replace_if_null(&_singleton, s)) {
1186 delete s;
1187 }
1188 }
1189 return _singleton;
1190 }
1191 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1192 if (_true_cache == nullptr) {
1193 cache_in_error = true;
1194 return nullptr;
1195 }
1196 // Have to cast to avoid little/big-endian problems.
1197 jboolean value = (jboolean)*((jint*)&raw_value);
1198 return lookup(value);
1199 }
1200 oop lookup(jboolean value) {
1201 if (value != 0) {
1202 return JNIHandles::resolve_non_null(_true_cache);
1203 }
1204 return JNIHandles::resolve_non_null(_false_cache);
1205 }
1206 };
1207
1208 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1209
1210 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1211 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1212 BasicType box_type = vmClasses::box_klass_type(k);
1213 if (box_type != T_OBJECT) {
1214 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1215 switch(box_type) {
1216 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1217 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1218 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1219 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1220 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1221 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1222 default:;
1223 }
1224 }
1225 return nullptr;
1226 }
1227 #endif // INCLUDE_JVMCI
1228
1229 #if COMPILER2_OR_JVMCI
1230 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1231 Handle pending_exception(THREAD, thread->pending_exception());
1232 const char* exception_file = thread->exception_file();
1233 int exception_line = thread->exception_line();
1234 thread->clear_pending_exception();
1235
1236 bool failures = false;
1237
1238 for (int i = 0; i < objects->length(); i++) {
1239 assert(objects->at(i)->is_object(), "invalid debug information");
1240 ObjectValue* sv = (ObjectValue*) objects->at(i);
1241
1242 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1243 oop obj = nullptr;
1244
1245 bool cache_init_error = false;
1246 if (k->is_instance_klass()) {
1247 #if INCLUDE_JVMCI
1248 nmethod* nm = fr->cb()->as_nmethod_or_null();
1249 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1250 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1251 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1252 if (obj != nullptr) {
1253 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1254 abv->set_cached(true);
1255 } else if (cache_init_error) {
1256 // Results in an OOME which is valid (as opposed to a class initialization error)
1257 // and is fine for the rare case a cache initialization failing.
1258 failures = true;
1259 }
1260 }
1261 #endif // INCLUDE_JVMCI
1262
1263 InstanceKlass* ik = InstanceKlass::cast(k);
1264 if (obj == nullptr && !cache_init_error) {
1265 InternalOOMEMark iom(THREAD);
1266 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1267 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1268 } else {
1269 obj = ik->allocate_instance(THREAD);
1270 }
1271 }
1272 } else if (k->is_typeArray_klass()) {
1273 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1274 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1275 int len = sv->field_size() / type2size[ak->element_type()];
1276 InternalOOMEMark iom(THREAD);
1277 obj = ak->allocate(len, THREAD);
1278 } else if (k->is_objArray_klass()) {
1279 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1280 InternalOOMEMark iom(THREAD);
1281 obj = ak->allocate(sv->field_size(), THREAD);
1282 }
1283
1284 if (obj == nullptr) {
1285 failures = true;
1286 }
1287
1288 assert(sv->value().is_null(), "redundant reallocation");
1289 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1290 CLEAR_PENDING_EXCEPTION;
1291 sv->set_value(obj);
1292 }
1293
1294 if (failures) {
1295 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1296 } else if (pending_exception.not_null()) {
1297 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1298 }
1299
1300 return failures;
1301 }
1302
1303 #if INCLUDE_JVMCI
1304 /**
1305 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1306 * we need to somehow be able to recover the actual kind to be able to write the correct
1307 * amount of bytes.
1308 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1309 * the entries at index n + 1 to n + i are 'markers'.
1310 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1311 * expected form of the array would be:
1312 *
1313 * {b0, b1, b2, b3, INT, marker, b6, b7}
1314 *
1315 * Thus, in order to get back the size of the entry, we simply need to count the number
1316 * of marked entries
1317 *
1318 * @param virtualArray the virtualized byte array
1319 * @param i index of the virtual entry we are recovering
1320 * @return The number of bytes the entry spans
1321 */
1322 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1323 int index = i;
1324 while (++index < virtualArray->field_size() &&
1325 virtualArray->field_at(index)->is_marker()) {}
1326 return index - i;
1327 }
1328
1329 /**
1330 * If there was a guarantee for byte array to always start aligned to a long, we could
1331 * do a simple check on the parity of the index. Unfortunately, that is not always the
1332 * case. Thus, we check alignment of the actual address we are writing to.
1333 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1334 * write a long to index 3.
1335 */
1336 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1337 jbyte* res = obj->byte_at_addr(index);
1338 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1339 return res;
1340 }
1341
1342 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1343 switch (byte_count) {
1344 case 1:
1345 obj->byte_at_put(index, (jbyte) value->get_jint());
1346 break;
1347 case 2:
1348 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1349 break;
1350 case 4:
1351 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1352 break;
1353 case 8:
1354 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1355 break;
1356 default:
1357 ShouldNotReachHere();
1358 }
1359 }
1360 #endif // INCLUDE_JVMCI
1361
1362
1363 // restore elements of an eliminated type array
1364 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1365 int index = 0;
1366
1367 for (int i = 0; i < sv->field_size(); i++) {
1368 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1369 switch(type) {
1370 case T_LONG: case T_DOUBLE: {
1371 assert(value->type() == T_INT, "Agreement.");
1372 StackValue* low =
1373 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1374 #ifdef _LP64
1375 jlong res = (jlong)low->get_intptr();
1376 #else
1377 jlong res = jlong_from(value->get_jint(), low->get_jint());
1378 #endif
1379 obj->long_at_put(index, res);
1380 break;
1381 }
1382
1383 case T_INT: case T_FLOAT: { // 4 bytes.
1384 assert(value->type() == T_INT, "Agreement.");
1385 bool big_value = false;
1386 if (i + 1 < sv->field_size() && type == T_INT) {
1387 if (sv->field_at(i)->is_location()) {
1388 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1389 if (type == Location::dbl || type == Location::lng) {
1390 big_value = true;
1391 }
1392 } else if (sv->field_at(i)->is_constant_int()) {
1393 ScopeValue* next_scope_field = sv->field_at(i + 1);
1394 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1395 big_value = true;
1396 }
1397 }
1398 }
1399
1400 if (big_value) {
1401 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1402 #ifdef _LP64
1403 jlong res = (jlong)low->get_intptr();
1404 #else
1405 jlong res = jlong_from(value->get_jint(), low->get_jint());
1406 #endif
1407 obj->int_at_put(index, *(jint*)&res);
1408 obj->int_at_put(++index, *((jint*)&res + 1));
1409 } else {
1410 obj->int_at_put(index, value->get_jint());
1411 }
1412 break;
1413 }
1414
1415 case T_SHORT:
1416 assert(value->type() == T_INT, "Agreement.");
1417 obj->short_at_put(index, (jshort)value->get_jint());
1418 break;
1419
1420 case T_CHAR:
1421 assert(value->type() == T_INT, "Agreement.");
1422 obj->char_at_put(index, (jchar)value->get_jint());
1423 break;
1424
1425 case T_BYTE: {
1426 assert(value->type() == T_INT, "Agreement.");
1427 #if INCLUDE_JVMCI
1428 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1429 int byte_count = count_number_of_bytes_for_entry(sv, i);
1430 byte_array_put(obj, value, index, byte_count);
1431 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1432 i += byte_count - 1; // Balance the loop counter.
1433 index += byte_count;
1434 // index has been updated so continue at top of loop
1435 continue;
1436 #else
1437 obj->byte_at_put(index, (jbyte)value->get_jint());
1438 break;
1439 #endif // INCLUDE_JVMCI
1440 }
1441
1442 case T_BOOLEAN: {
1443 assert(value->type() == T_INT, "Agreement.");
1444 obj->bool_at_put(index, (jboolean)value->get_jint());
1445 break;
1446 }
1447
1448 default:
1449 ShouldNotReachHere();
1450 }
1451 index++;
1452 }
1453 }
1454
1455 // restore fields of an eliminated object array
1456 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1457 for (int i = 0; i < sv->field_size(); i++) {
1458 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1459 assert(value->type() == T_OBJECT, "object element expected");
1460 obj->obj_at_put(i, value->get_obj()());
1461 }
1462 }
1463
1464 class ReassignedField {
1465 public:
1466 int _offset;
1467 BasicType _type;
1468 public:
1469 ReassignedField() {
1470 _offset = 0;
1471 _type = T_ILLEGAL;
1472 }
1473 };
1474
1475 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1476 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1477 InstanceKlass* super = klass->superklass();
1478 if (super != nullptr) {
1479 get_reassigned_fields(super, fields, is_jvmci);
1480 }
1481 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1482 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1483 ReassignedField field;
1484 field._offset = fs.offset();
1485 field._type = Signature::basic_type(fs.signature());
1486 fields->append(field);
1487 }
1488 }
1489 return fields;
1490 }
1491
1492 // Restore fields of an eliminated instance object employing the same field order used by the compiler.
1493 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool is_jvmci) {
1494 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1495 for (int i = 0; i < fields->length(); i++) {
1496 ScopeValue* scope_field = sv->field_at(svIndex);
1497 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1498 int offset = fields->at(i)._offset;
1499 BasicType type = fields->at(i)._type;
1500 switch (type) {
1501 case T_OBJECT: case T_ARRAY:
1502 assert(value->type() == T_OBJECT, "Agreement.");
1503 obj->obj_field_put(offset, value->get_obj()());
1504 break;
1505
1506 case T_INT: case T_FLOAT: { // 4 bytes.
1507 assert(value->type() == T_INT, "Agreement.");
1508 bool big_value = false;
1509 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1510 if (scope_field->is_location()) {
1511 Location::Type type = ((LocationValue*) scope_field)->location().type();
1512 if (type == Location::dbl || type == Location::lng) {
1513 big_value = true;
1514 }
1515 }
1516 if (scope_field->is_constant_int()) {
1517 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1518 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1519 big_value = true;
1520 }
1521 }
1522 }
1523
1524 if (big_value) {
1525 i++;
1526 assert(i < fields->length(), "second T_INT field needed");
1527 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1528 } else {
1529 obj->int_field_put(offset, value->get_jint());
1530 break;
1531 }
1532 }
1533 /* no break */
1534
1535 case T_LONG: case T_DOUBLE: {
1536 assert(value->type() == T_INT, "Agreement.");
1537 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1538 #ifdef _LP64
1539 jlong res = (jlong)low->get_intptr();
1540 #else
1541 jlong res = jlong_from(value->get_jint(), low->get_jint());
1542 #endif
1543 obj->long_field_put(offset, res);
1544 break;
1545 }
1546
1547 case T_SHORT:
1548 assert(value->type() == T_INT, "Agreement.");
1549 obj->short_field_put(offset, (jshort)value->get_jint());
1550 break;
1551
1552 case T_CHAR:
1553 assert(value->type() == T_INT, "Agreement.");
1554 obj->char_field_put(offset, (jchar)value->get_jint());
1555 break;
1556
1557 case T_BYTE:
1558 assert(value->type() == T_INT, "Agreement.");
1559 obj->byte_field_put(offset, (jbyte)value->get_jint());
1560 break;
1561
1562 case T_BOOLEAN:
1563 assert(value->type() == T_INT, "Agreement.");
1564 obj->bool_field_put(offset, (jboolean)value->get_jint());
1565 break;
1566
1567 default:
1568 ShouldNotReachHere();
1569 }
1570 svIndex++;
1571 }
1572 return svIndex;
1573 }
1574
1575 // restore fields of all eliminated objects and arrays
1576 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci) {
1577 for (int i = 0; i < objects->length(); i++) {
1578 assert(objects->at(i)->is_object(), "invalid debug information");
1579 ObjectValue* sv = (ObjectValue*) objects->at(i);
1580 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1581 Handle obj = sv->value();
1582 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1583 #ifndef PRODUCT
1584 if (PrintDeoptimizationDetails) {
1585 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1586 }
1587 #endif // !PRODUCT
1588
1589 if (obj.is_null()) {
1590 continue;
1591 }
1592
1593 #if INCLUDE_JVMCI
1594 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1595 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1596 continue;
1597 }
1598 #endif // INCLUDE_JVMCI
1599 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1600 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1601 ScopeValue* payload = sv->field_at(0);
1602 if (payload->is_location() &&
1603 payload->as_LocationValue()->location().type() == Location::vector) {
1604 #ifndef PRODUCT
1605 if (PrintDeoptimizationDetails) {
1606 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1607 if (Verbose) {
1608 Handle obj = sv->value();
1609 k->oop_print_on(obj(), tty);
1610 }
1611 }
1612 #endif // !PRODUCT
1613 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1614 }
1615 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1616 // which could be restored after vector object allocation.
1617 }
1618 if (k->is_instance_klass()) {
1619 InstanceKlass* ik = InstanceKlass::cast(k);
1620 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci);
1621 } else if (k->is_typeArray_klass()) {
1622 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1623 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1624 } else if (k->is_objArray_klass()) {
1625 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1626 }
1627 }
1628 // These objects may escape when we return to Interpreter after deoptimization.
1629 // We need barrier so that stores that initialize these objects can't be reordered
1630 // with subsequent stores that make these objects accessible by other threads.
1631 OrderAccess::storestore();
1632 }
1633
1634
1635 // relock objects for which synchronization was eliminated
1636 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1637 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1638 bool relocked_objects = false;
1639 for (int i = 0; i < monitors->length(); i++) {
1640 MonitorInfo* mon_info = monitors->at(i);
1641 if (mon_info->eliminated()) {
1642 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1643 relocked_objects = true;
1644 if (!mon_info->owner_is_scalar_replaced()) {
1645 Handle obj(thread, mon_info->owner());
1646 markWord mark = obj->mark();
1647 if (exec_mode == Unpack_none) {
1648 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1649 // With exec_mode == Unpack_none obj may be thread local and locked in
1650 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1651 markWord dmw = mark.displaced_mark_helper();
1652 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1653 obj->set_mark(dmw);
1654 }
1655 if (mark.has_monitor()) {
1656 // defer relocking if the deoptee thread is currently waiting for obj
1657 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1658 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1659 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1660 if (LockingMode == LM_LEGACY) {
1661 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1662 } else if (UseObjectMonitorTable) {
1663 mon_info->lock()->clear_object_monitor_cache();
1664 }
1665 #ifdef ASSERT
1666 else {
1667 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1668 mon_info->lock()->set_bad_metadata_deopt();
1669 }
1670 #endif
1671 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1672 continue;
1673 }
1674 }
1675 }
1676 BasicLock* lock = mon_info->lock();
1677 if (LockingMode == LM_LIGHTWEIGHT) {
1678 // We have lost information about the correct state of the lock stack.
1679 // Entering may create an invalid lock stack. Inflate the lock if it
1680 // was fast_locked to restore the valid lock stack.
1681 if (UseObjectMonitorTable) {
1682 // UseObjectMonitorTable expects the BasicLock cache to be either a
1683 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1684 // to nullptr.
1685 lock->clear_object_monitor_cache();
1686 }
1687 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1688 if (deoptee_thread->lock_stack().contains(obj())) {
1689 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1690 deoptee_thread, thread);
1691 }
1692 assert(mon_info->owner()->is_locked(), "object must be locked now");
1693 assert(obj->mark().has_monitor(), "must be");
1694 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1695 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1696 } else {
1697 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1698 assert(mon_info->owner()->is_locked(), "object must be locked now");
1699 }
1700 }
1701 }
1702 }
1703 return relocked_objects;
1704 }
1705 #endif // COMPILER2_OR_JVMCI
1706
1707 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1708 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1709
1710 // Register map for next frame (used for stack crawl). We capture
1711 // the state of the deopt'ing frame's caller. Thus if we need to
1712 // stuff a C2I adapter we can properly fill in the callee-save
1713 // register locations.
1714 frame caller = fr.sender(reg_map);
1715 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1716
1717 frame sender = caller;
1718
1719 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1720 // the vframeArray containing the unpacking information is allocated in the C heap.
1721 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1722 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1723
1724 // Compare the vframeArray to the collected vframes
1725 assert(array->structural_compare(thread, chunk), "just checking");
1726
1727 if (TraceDeoptimization) {
1728 ResourceMark rm;
1729 stringStream st;
1730 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1731 st.print(" ");
1732 fr.print_on(&st);
1733 st.print_cr(" Virtual frames (innermost/newest first):");
1734 for (int index = 0; index < chunk->length(); index++) {
1735 compiledVFrame* vf = chunk->at(index);
1736 int bci = vf->raw_bci();
1737 const char* code_name;
1738 if (bci == SynchronizationEntryBCI) {
1739 code_name = "sync entry";
1740 } else {
1741 Bytecodes::Code code = vf->method()->code_at(bci);
1742 code_name = Bytecodes::name(code);
1743 }
1744
1745 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1746 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1747 st.print(" - %s", code_name);
1748 st.print_cr(" @ bci=%d ", bci);
1749 }
1750 tty->print_raw(st.freeze());
1751 tty->cr();
1752 }
1753
1754 return array;
1755 }
1756
1757 #if COMPILER2_OR_JVMCI
1758 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1759 // Reallocation of some scalar replaced objects failed. Record
1760 // that we need to pop all the interpreter frames for the
1761 // deoptimized compiled frame.
1762 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1763 thread->set_frames_to_pop_failed_realloc(array->frames());
1764 // Unlock all monitors here otherwise the interpreter will see a
1765 // mix of locked and unlocked monitors (because of failed
1766 // reallocations of synchronized objects) and be confused.
1767 for (int i = 0; i < array->frames(); i++) {
1768 MonitorChunk* monitors = array->element(i)->monitors();
1769 if (monitors != nullptr) {
1770 // Unlock in reverse order starting from most nested monitor.
1771 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1772 BasicObjectLock* src = monitors->at(j);
1773 if (src->obj() != nullptr) {
1774 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1775 }
1776 }
1777 array->element(i)->free_monitors();
1778 #ifdef ASSERT
1779 array->element(i)->set_removed_monitors();
1780 #endif
1781 }
1782 }
1783 }
1784 #endif
1785
1786 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1787 assert(fr.can_be_deoptimized(), "checking frame type");
1788
1789 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1790
1791 if (LogCompilation && xtty != nullptr) {
1792 nmethod* nm = fr.cb()->as_nmethod_or_null();
1793 assert(nm != nullptr, "only compiled methods can deopt");
1794
1795 ttyLocker ttyl;
1796 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1797 nm->log_identity(xtty);
1798 xtty->end_head();
1799 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1800 xtty->begin_elem("jvms bci='%d'", sd->bci());
1801 xtty->method(sd->method());
1802 xtty->end_elem();
1803 if (sd->is_top()) break;
1804 }
1805 xtty->tail("deoptimized");
1806 }
1807
1808 Continuation::notify_deopt(thread, fr.sp());
1809
1810 // Patch the compiled method so that when execution returns to it we will
1811 // deopt the execution state and return to the interpreter.
1812 fr.deoptimize(thread);
1813 }
1814
1815 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1816 // Deoptimize only if the frame comes from compile code.
1817 // Do not deoptimize the frame which is already patched
1818 // during the execution of the loops below.
1819 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1820 return;
1821 }
1822 ResourceMark rm;
1823 deoptimize_single_frame(thread, fr, reason);
1824 }
1825
1826 #if INCLUDE_JVMCI
1827 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1828 // there is no exception handler for this pc => deoptimize
1829 nm->make_not_entrant(nmethod::ChangeReason::missing_exception_handler);
1830
1831 // Use Deoptimization::deoptimize for all of its side-effects:
1832 // gathering traps statistics, logging...
1833 // it also patches the return pc but we do not care about that
1834 // since we return a continuation to the deopt_blob below.
1835 JavaThread* thread = JavaThread::current();
1836 RegisterMap reg_map(thread,
1837 RegisterMap::UpdateMap::skip,
1838 RegisterMap::ProcessFrames::include,
1839 RegisterMap::WalkContinuation::skip);
1840 frame runtime_frame = thread->last_frame();
1841 frame caller_frame = runtime_frame.sender(®_map);
1842 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1843 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1844 compiledVFrame* cvf = compiledVFrame::cast(vf);
1845 ScopeDesc* imm_scope = cvf->scope();
1846 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1847 if (imm_mdo != nullptr) {
1848 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1849 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1850
1851 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1852 if (pdata != nullptr && pdata->is_BitData()) {
1853 BitData* bit_data = (BitData*) pdata;
1854 bit_data->set_exception_seen();
1855 }
1856 }
1857
1858 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1859
1860 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1861 if (trap_mdo != nullptr) {
1862 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1863 }
1864
1865 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1866 }
1867 #endif
1868
1869 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1870 assert(thread == Thread::current() ||
1871 thread->is_handshake_safe_for(Thread::current()) ||
1872 SafepointSynchronize::is_at_safepoint(),
1873 "can only deoptimize other thread at a safepoint/handshake");
1874 // Compute frame and register map based on thread and sp.
1875 RegisterMap reg_map(thread,
1876 RegisterMap::UpdateMap::skip,
1877 RegisterMap::ProcessFrames::include,
1878 RegisterMap::WalkContinuation::skip);
1879 frame fr = thread->last_frame();
1880 while (fr.id() != id) {
1881 fr = fr.sender(®_map);
1882 }
1883 deoptimize(thread, fr, reason);
1884 }
1885
1886
1887 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1888 Thread* current = Thread::current();
1889 if (thread == current || thread->is_handshake_safe_for(current)) {
1890 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1891 } else {
1892 VM_DeoptimizeFrame deopt(thread, id, reason);
1893 VMThread::execute(&deopt);
1894 }
1895 }
1896
1897 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1898 deoptimize_frame(thread, id, Reason_constraint);
1899 }
1900
1901 // JVMTI PopFrame support
1902 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1903 {
1904 assert(thread == JavaThread::current(), "pre-condition");
1905 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1906 }
1907 JRT_END
1908
1909 MethodData*
1910 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1911 bool create_if_missing) {
1912 JavaThread* THREAD = thread; // For exception macros.
1913 MethodData* mdo = m()->method_data();
1914 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1915 // Build an MDO. Ignore errors like OutOfMemory;
1916 // that simply means we won't have an MDO to update.
1917 Method::build_profiling_method_data(m, THREAD);
1918 if (HAS_PENDING_EXCEPTION) {
1919 // Only metaspace OOM is expected. No Java code executed.
1920 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1921 CLEAR_PENDING_EXCEPTION;
1922 }
1923 mdo = m()->method_data();
1924 }
1925 return mdo;
1926 }
1927
1928 #if COMPILER2_OR_JVMCI
1929 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1930 // In case of an unresolved klass entry, load the class.
1931 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1932 // and probably nowhere else.
1933 // Even that case would benefit from simply re-interpreting the
1934 // bytecode, without paying special attention to the class index.
1935 // So this whole "class index" feature should probably be removed.
1936
1937 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1938 Klass* tk = constant_pool->klass_at(index, THREAD);
1939 if (HAS_PENDING_EXCEPTION) {
1940 // Exception happened during classloading. We ignore the exception here, since it
1941 // is going to be rethrown since the current activation is going to be deoptimized and
1942 // the interpreter will re-execute the bytecode.
1943 // Do not clear probable Async Exceptions.
1944 CLEAR_PENDING_NONASYNC_EXCEPTION;
1945 // Class loading called java code which may have caused a stack
1946 // overflow. If the exception was thrown right before the return
1947 // to the runtime the stack is no longer guarded. Reguard the
1948 // stack otherwise if we return to the uncommon trap blob and the
1949 // stack bang causes a stack overflow we crash.
1950 JavaThread* jt = THREAD;
1951 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1952 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1953 }
1954 return;
1955 }
1956
1957 assert(!constant_pool->tag_at(index).is_symbol(),
1958 "no symbolic names here, please");
1959 }
1960
1961 #if INCLUDE_JFR
1962
1963 class DeoptReasonSerializer : public JfrSerializer {
1964 public:
1965 void serialize(JfrCheckpointWriter& writer) {
1966 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1967 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1968 writer.write_key((u8)i);
1969 writer.write(Deoptimization::trap_reason_name(i));
1970 }
1971 }
1972 };
1973
1974 class DeoptActionSerializer : public JfrSerializer {
1975 public:
1976 void serialize(JfrCheckpointWriter& writer) {
1977 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1978 writer.write_count(nof_actions);
1979 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1980 writer.write_key(i);
1981 writer.write(Deoptimization::trap_action_name((int)i));
1982 }
1983 }
1984 };
1985
1986 static void register_serializers() {
1987 static int critical_section = 0;
1988 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1989 return;
1990 }
1991 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1992 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1993 }
1994
1995 static void post_deoptimization_event(nmethod* nm,
1996 const Method* method,
1997 int trap_bci,
1998 int instruction,
1999 Deoptimization::DeoptReason reason,
2000 Deoptimization::DeoptAction action) {
2001 assert(nm != nullptr, "invariant");
2002 assert(method != nullptr, "invariant");
2003 if (EventDeoptimization::is_enabled()) {
2004 static bool serializers_registered = false;
2005 if (!serializers_registered) {
2006 register_serializers();
2007 serializers_registered = true;
2008 }
2009 EventDeoptimization event;
2010 event.set_compileId(nm->compile_id());
2011 event.set_compiler(nm->compiler_type());
2012 event.set_method(method);
2013 event.set_lineNumber(method->line_number_from_bci(trap_bci));
2014 event.set_bci(trap_bci);
2015 event.set_instruction(instruction);
2016 event.set_reason(reason);
2017 event.set_action(action);
2018 event.commit();
2019 }
2020 }
2021
2022 #endif // INCLUDE_JFR
2023
2024 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2025 const char* reason_name, const char* reason_action) {
2026 LogTarget(Debug, deoptimization) lt;
2027 if (lt.is_enabled()) {
2028 LogStream ls(lt);
2029 bool is_osr = nm->is_osr_method();
2030 ls.print("cid=%4d %s level=%d",
2031 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2032 ls.print(" %s", tm->name_and_sig_as_C_string());
2033 ls.print(" trap_bci=%d ", trap_bci);
2034 if (is_osr) {
2035 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2036 }
2037 ls.print("%s ", reason_name);
2038 ls.print("%s ", reason_action);
2039 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2040 pc, fr.pc() - nm->code_begin());
2041 }
2042 }
2043
2044 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2045 HandleMark hm(current);
2046
2047 // uncommon_trap() is called at the beginning of the uncommon trap
2048 // handler. Note this fact before we start generating temporary frames
2049 // that can confuse an asynchronous stack walker. This counter is
2050 // decremented at the end of unpack_frames().
2051
2052 current->inc_in_deopt_handler();
2053
2054 #if INCLUDE_JVMCI
2055 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2056 RegisterMap reg_map(current,
2057 RegisterMap::UpdateMap::include,
2058 RegisterMap::ProcessFrames::include,
2059 RegisterMap::WalkContinuation::skip);
2060 #else
2061 RegisterMap reg_map(current,
2062 RegisterMap::UpdateMap::skip,
2063 RegisterMap::ProcessFrames::include,
2064 RegisterMap::WalkContinuation::skip);
2065 #endif
2066 frame stub_frame = current->last_frame();
2067 frame fr = stub_frame.sender(®_map);
2068
2069 // Log a message
2070 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2071 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2072
2073 {
2074 ResourceMark rm;
2075
2076 DeoptReason reason = trap_request_reason(trap_request);
2077 DeoptAction action = trap_request_action(trap_request);
2078 #if INCLUDE_JVMCI
2079 int debug_id = trap_request_debug_id(trap_request);
2080 #endif
2081 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2082
2083 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2084 compiledVFrame* cvf = compiledVFrame::cast(vf);
2085
2086 nmethod* nm = cvf->code();
2087
2088 ScopeDesc* trap_scope = cvf->scope();
2089
2090 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2091
2092 if (is_receiver_constraint_failure) {
2093 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2094 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2095 JVMCI_ONLY(COMMA debug_id));
2096 }
2097
2098 methodHandle trap_method(current, trap_scope->method());
2099 int trap_bci = trap_scope->bci();
2100 #if INCLUDE_JVMCI
2101 jlong speculation = current->pending_failed_speculation();
2102 if (nm->is_compiled_by_jvmci()) {
2103 nm->update_speculation(current);
2104 } else {
2105 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2106 }
2107
2108 if (trap_bci == SynchronizationEntryBCI) {
2109 trap_bci = 0;
2110 current->set_pending_monitorenter(true);
2111 }
2112
2113 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2114 current->set_pending_transfer_to_interpreter(true);
2115 }
2116 #endif
2117
2118 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2119 // Record this event in the histogram.
2120 gather_statistics(reason, action, trap_bc);
2121
2122 // Ensure that we can record deopt. history:
2123 bool create_if_missing = ProfileTraps;
2124
2125 methodHandle profiled_method;
2126 #if INCLUDE_JVMCI
2127 if (nm->is_compiled_by_jvmci()) {
2128 profiled_method = methodHandle(current, nm->method());
2129 } else {
2130 profiled_method = trap_method;
2131 }
2132 #else
2133 profiled_method = trap_method;
2134 #endif
2135
2136 MethodData* trap_mdo =
2137 get_method_data(current, profiled_method, create_if_missing);
2138
2139 { // Log Deoptimization event for JFR, UL and event system
2140 Method* tm = trap_method();
2141 const char* reason_name = trap_reason_name(reason);
2142 const char* reason_action = trap_action_name(action);
2143 intptr_t pc = p2i(fr.pc());
2144
2145 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2146 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2147 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2148 reason_name, reason_action, pc,
2149 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2150 }
2151
2152 // Print a bunch of diagnostics, if requested.
2153 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2154 ResourceMark rm;
2155
2156 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2157 // We must do this already now, since we cannot acquire this lock while
2158 // holding the tty lock (lock ordering by rank).
2159 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2160
2161 ttyLocker ttyl;
2162
2163 char buf[100];
2164 if (xtty != nullptr) {
2165 xtty->begin_head("uncommon_trap thread='%zu' %s",
2166 os::current_thread_id(),
2167 format_trap_request(buf, sizeof(buf), trap_request));
2168 #if INCLUDE_JVMCI
2169 if (speculation != 0) {
2170 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2171 }
2172 #endif
2173 nm->log_identity(xtty);
2174 }
2175 Symbol* class_name = nullptr;
2176 bool unresolved = false;
2177 if (unloaded_class_index >= 0) {
2178 constantPoolHandle constants (current, trap_method->constants());
2179 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2180 class_name = constants->klass_name_at(unloaded_class_index);
2181 unresolved = true;
2182 if (xtty != nullptr)
2183 xtty->print(" unresolved='1'");
2184 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2185 class_name = constants->symbol_at(unloaded_class_index);
2186 }
2187 if (xtty != nullptr)
2188 xtty->name(class_name);
2189 }
2190 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2191 // Dump the relevant MDO state.
2192 // This is the deopt count for the current reason, any previous
2193 // reasons or recompiles seen at this point.
2194 int dcnt = trap_mdo->trap_count(reason);
2195 if (dcnt != 0)
2196 xtty->print(" count='%d'", dcnt);
2197
2198 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2199 // lock extra_data_lock before the tty lock.
2200 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2201 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2202 if (dos != 0) {
2203 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2204 if (trap_state_is_recompiled(dos)) {
2205 int recnt2 = trap_mdo->overflow_recompile_count();
2206 if (recnt2 != 0)
2207 xtty->print(" recompiles2='%d'", recnt2);
2208 }
2209 }
2210 }
2211 if (xtty != nullptr) {
2212 xtty->stamp();
2213 xtty->end_head();
2214 }
2215 if (TraceDeoptimization) { // make noise on the tty
2216 stringStream st;
2217 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2218 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2219 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2220 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2221 #if INCLUDE_JVMCI
2222 if (nm->is_compiled_by_jvmci()) {
2223 const char* installed_code_name = nm->jvmci_name();
2224 if (installed_code_name != nullptr) {
2225 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2226 }
2227 }
2228 #endif
2229 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2230 p2i(fr.pc()),
2231 os::current_thread_id(),
2232 trap_reason_name(reason),
2233 trap_action_name(action),
2234 unloaded_class_index
2235 #if INCLUDE_JVMCI
2236 , debug_id
2237 #endif
2238 );
2239 if (class_name != nullptr) {
2240 st.print(unresolved ? " unresolved class: " : " symbol: ");
2241 class_name->print_symbol_on(&st);
2242 }
2243 st.cr();
2244 tty->print_raw(st.freeze());
2245 }
2246 if (xtty != nullptr) {
2247 // Log the precise location of the trap.
2248 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2249 xtty->begin_elem("jvms bci='%d'", sd->bci());
2250 xtty->method(sd->method());
2251 xtty->end_elem();
2252 if (sd->is_top()) break;
2253 }
2254 xtty->tail("uncommon_trap");
2255 }
2256 }
2257 // (End diagnostic printout.)
2258
2259 if (is_receiver_constraint_failure) {
2260 fatal("missing receiver type check");
2261 }
2262
2263 // Load class if necessary
2264 if (unloaded_class_index >= 0) {
2265 constantPoolHandle constants(current, trap_method->constants());
2266 load_class_by_index(constants, unloaded_class_index, THREAD);
2267 }
2268
2269 // Flush the nmethod if necessary and desirable.
2270 //
2271 // We need to avoid situations where we are re-flushing the nmethod
2272 // because of a hot deoptimization site. Repeated flushes at the same
2273 // point need to be detected by the compiler and avoided. If the compiler
2274 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2275 // module must take measures to avoid an infinite cycle of recompilation
2276 // and deoptimization. There are several such measures:
2277 //
2278 // 1. If a recompilation is ordered a second time at some site X
2279 // and for the same reason R, the action is adjusted to 'reinterpret',
2280 // to give the interpreter time to exercise the method more thoroughly.
2281 // If this happens, the method's overflow_recompile_count is incremented.
2282 //
2283 // 2. If the compiler fails to reduce the deoptimization rate, then
2284 // the method's overflow_recompile_count will begin to exceed the set
2285 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2286 // is adjusted to 'make_not_compilable', and the method is abandoned
2287 // to the interpreter. This is a performance hit for hot methods,
2288 // but is better than a disastrous infinite cycle of recompilations.
2289 // (Actually, only the method containing the site X is abandoned.)
2290 //
2291 // 3. In parallel with the previous measures, if the total number of
2292 // recompilations of a method exceeds the much larger set limit
2293 // PerMethodRecompilationCutoff, the method is abandoned.
2294 // This should only happen if the method is very large and has
2295 // many "lukewarm" deoptimizations. The code which enforces this
2296 // limit is elsewhere (class nmethod, class Method).
2297 //
2298 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2299 // to recompile at each bytecode independently of the per-BCI cutoff.
2300 //
2301 // The decision to update code is up to the compiler, and is encoded
2302 // in the Action_xxx code. If the compiler requests Action_none
2303 // no trap state is changed, no compiled code is changed, and the
2304 // computation suffers along in the interpreter.
2305 //
2306 // The other action codes specify various tactics for decompilation
2307 // and recompilation. Action_maybe_recompile is the loosest, and
2308 // allows the compiled code to stay around until enough traps are seen,
2309 // and until the compiler gets around to recompiling the trapping method.
2310 //
2311 // The other actions cause immediate removal of the present code.
2312
2313 // Traps caused by injected profile shouldn't pollute trap counts.
2314 bool injected_profile_trap = trap_method->has_injected_profile() &&
2315 (reason == Reason_intrinsic || reason == Reason_unreached);
2316
2317 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2318 bool make_not_entrant = false;
2319 bool make_not_compilable = false;
2320 bool reprofile = false;
2321 switch (action) {
2322 case Action_none:
2323 // Keep the old code.
2324 update_trap_state = false;
2325 break;
2326 case Action_maybe_recompile:
2327 // Do not need to invalidate the present code, but we can
2328 // initiate another
2329 // Start compiler without (necessarily) invalidating the nmethod.
2330 // The system will tolerate the old code, but new code should be
2331 // generated when possible.
2332 break;
2333 case Action_reinterpret:
2334 // Go back into the interpreter for a while, and then consider
2335 // recompiling form scratch.
2336 make_not_entrant = true;
2337 // Reset invocation counter for outer most method.
2338 // This will allow the interpreter to exercise the bytecodes
2339 // for a while before recompiling.
2340 // By contrast, Action_make_not_entrant is immediate.
2341 //
2342 // Note that the compiler will track null_check, null_assert,
2343 // range_check, and class_check events and log them as if they
2344 // had been traps taken from compiled code. This will update
2345 // the MDO trap history so that the next compilation will
2346 // properly detect hot trap sites.
2347 reprofile = true;
2348 break;
2349 case Action_make_not_entrant:
2350 // Request immediate recompilation, and get rid of the old code.
2351 // Make them not entrant, so next time they are called they get
2352 // recompiled. Unloaded classes are loaded now so recompile before next
2353 // time they are called. Same for uninitialized. The interpreter will
2354 // link the missing class, if any.
2355 make_not_entrant = true;
2356 break;
2357 case Action_make_not_compilable:
2358 // Give up on compiling this method at all.
2359 make_not_entrant = true;
2360 make_not_compilable = true;
2361 break;
2362 default:
2363 ShouldNotReachHere();
2364 }
2365
2366 // Setting +ProfileTraps fixes the following, on all platforms:
2367 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2368 // recompile relies on a MethodData* to record heroic opt failures.
2369
2370 // Whether the interpreter is producing MDO data or not, we also need
2371 // to use the MDO to detect hot deoptimization points and control
2372 // aggressive optimization.
2373 bool inc_recompile_count = false;
2374
2375 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2376 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2377 (trap_mdo != nullptr),
2378 Mutex::_no_safepoint_check_flag);
2379 ProfileData* pdata = nullptr;
2380 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2381 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2382 uint this_trap_count = 0;
2383 bool maybe_prior_trap = false;
2384 bool maybe_prior_recompile = false;
2385
2386 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2387 #if INCLUDE_JVMCI
2388 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2389 #endif
2390 nm->method(),
2391 //outputs:
2392 this_trap_count,
2393 maybe_prior_trap,
2394 maybe_prior_recompile);
2395 // Because the interpreter also counts null, div0, range, and class
2396 // checks, these traps from compiled code are double-counted.
2397 // This is harmless; it just means that the PerXTrapLimit values
2398 // are in effect a little smaller than they look.
2399
2400 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2401 if (per_bc_reason != Reason_none) {
2402 // Now take action based on the partially known per-BCI history.
2403 if (maybe_prior_trap
2404 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2405 // If there are too many traps at this BCI, force a recompile.
2406 // This will allow the compiler to see the limit overflow, and
2407 // take corrective action, if possible. The compiler generally
2408 // does not use the exact PerBytecodeTrapLimit value, but instead
2409 // changes its tactics if it sees any traps at all. This provides
2410 // a little hysteresis, delaying a recompile until a trap happens
2411 // several times.
2412 //
2413 // Actually, since there is only one bit of counter per BCI,
2414 // the possible per-BCI counts are {0,1,(per-method count)}.
2415 // This produces accurate results if in fact there is only
2416 // one hot trap site, but begins to get fuzzy if there are
2417 // many sites. For example, if there are ten sites each
2418 // trapping two or more times, they each get the blame for
2419 // all of their traps.
2420 make_not_entrant = true;
2421 }
2422
2423 // Detect repeated recompilation at the same BCI, and enforce a limit.
2424 if (make_not_entrant && maybe_prior_recompile) {
2425 // More than one recompile at this point.
2426 inc_recompile_count = maybe_prior_trap;
2427 }
2428 } else {
2429 // For reasons which are not recorded per-bytecode, we simply
2430 // force recompiles unconditionally.
2431 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2432 make_not_entrant = true;
2433 }
2434
2435 // Go back to the compiler if there are too many traps in this method.
2436 if (this_trap_count >= per_method_trap_limit(reason)) {
2437 // If there are too many traps in this method, force a recompile.
2438 // This will allow the compiler to see the limit overflow, and
2439 // take corrective action, if possible.
2440 // (This condition is an unlikely backstop only, because the
2441 // PerBytecodeTrapLimit is more likely to take effect first,
2442 // if it is applicable.)
2443 make_not_entrant = true;
2444 }
2445
2446 // Here's more hysteresis: If there has been a recompile at
2447 // this trap point already, run the method in the interpreter
2448 // for a while to exercise it more thoroughly.
2449 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2450 reprofile = true;
2451 }
2452 }
2453
2454 // Take requested actions on the method:
2455
2456 // Recompile
2457 if (make_not_entrant) {
2458 if (!nm->make_not_entrant(nmethod::ChangeReason::uncommon_trap)) {
2459 return; // the call did not change nmethod's state
2460 }
2461
2462 if (pdata != nullptr) {
2463 // Record the recompilation event, if any.
2464 int tstate0 = pdata->trap_state();
2465 int tstate1 = trap_state_set_recompiled(tstate0, true);
2466 if (tstate1 != tstate0)
2467 pdata->set_trap_state(tstate1);
2468 }
2469
2470 // For code aging we count traps separately here, using make_not_entrant()
2471 // as a guard against simultaneous deopts in multiple threads.
2472 if (reason == Reason_tenured && trap_mdo != nullptr) {
2473 trap_mdo->inc_tenure_traps();
2474 }
2475 }
2476
2477 if (inc_recompile_count) {
2478 trap_mdo->inc_overflow_recompile_count();
2479 if ((uint)trap_mdo->overflow_recompile_count() >
2480 (uint)PerBytecodeRecompilationCutoff) {
2481 // Give up on the method containing the bad BCI.
2482 if (trap_method() == nm->method()) {
2483 make_not_compilable = true;
2484 } else {
2485 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2486 // But give grace to the enclosing nm->method().
2487 }
2488 }
2489 }
2490
2491 // Reprofile
2492 if (reprofile) {
2493 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2494 }
2495
2496 // Give up compiling
2497 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2498 assert(make_not_entrant, "consistent");
2499 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2500 }
2501
2502 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2503 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2504 if (exception_handler_data != nullptr) {
2505 // uncommon trap at the start of an exception handler.
2506 // C2 generates these for un-entered exception handlers.
2507 // mark the handler as entered to avoid generating
2508 // another uncommon trap the next time the handler is compiled
2509 exception_handler_data->set_exception_handler_entered();
2510 }
2511 }
2512
2513 } // Free marked resources
2514
2515 }
2516 JRT_END
2517
2518 ProfileData*
2519 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2520 int trap_bci,
2521 Deoptimization::DeoptReason reason,
2522 bool update_total_trap_count,
2523 #if INCLUDE_JVMCI
2524 bool is_osr,
2525 #endif
2526 Method* compiled_method,
2527 //outputs:
2528 uint& ret_this_trap_count,
2529 bool& ret_maybe_prior_trap,
2530 bool& ret_maybe_prior_recompile) {
2531 trap_mdo->check_extra_data_locked();
2532
2533 bool maybe_prior_trap = false;
2534 bool maybe_prior_recompile = false;
2535 uint this_trap_count = 0;
2536 if (update_total_trap_count) {
2537 uint idx = reason;
2538 #if INCLUDE_JVMCI
2539 if (is_osr) {
2540 // Upper half of history array used for traps in OSR compilations
2541 idx += Reason_TRAP_HISTORY_LENGTH;
2542 }
2543 #endif
2544 uint prior_trap_count = trap_mdo->trap_count(idx);
2545 this_trap_count = trap_mdo->inc_trap_count(idx);
2546
2547 // If the runtime cannot find a place to store trap history,
2548 // it is estimated based on the general condition of the method.
2549 // If the method has ever been recompiled, or has ever incurred
2550 // a trap with the present reason , then this BCI is assumed
2551 // (pessimistically) to be the culprit.
2552 maybe_prior_trap = (prior_trap_count != 0);
2553 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2554 }
2555 ProfileData* pdata = nullptr;
2556
2557
2558 // For reasons which are recorded per bytecode, we check per-BCI data.
2559 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2560 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2561 if (per_bc_reason != Reason_none) {
2562 // Find the profile data for this BCI. If there isn't one,
2563 // try to allocate one from the MDO's set of spares.
2564 // This will let us detect a repeated trap at this point.
2565 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2566
2567 if (pdata != nullptr) {
2568 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2569 if (LogCompilation && xtty != nullptr) {
2570 ttyLocker ttyl;
2571 // no more room for speculative traps in this MDO
2572 xtty->elem("speculative_traps_oom");
2573 }
2574 }
2575 // Query the trap state of this profile datum.
2576 int tstate0 = pdata->trap_state();
2577 if (!trap_state_has_reason(tstate0, per_bc_reason))
2578 maybe_prior_trap = false;
2579 if (!trap_state_is_recompiled(tstate0))
2580 maybe_prior_recompile = false;
2581
2582 // Update the trap state of this profile datum.
2583 int tstate1 = tstate0;
2584 // Record the reason.
2585 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2586 // Store the updated state on the MDO, for next time.
2587 if (tstate1 != tstate0)
2588 pdata->set_trap_state(tstate1);
2589 } else {
2590 if (LogCompilation && xtty != nullptr) {
2591 ttyLocker ttyl;
2592 // Missing MDP? Leave a small complaint in the log.
2593 xtty->elem("missing_mdp bci='%d'", trap_bci);
2594 }
2595 }
2596 }
2597
2598 // Return results:
2599 ret_this_trap_count = this_trap_count;
2600 ret_maybe_prior_trap = maybe_prior_trap;
2601 ret_maybe_prior_recompile = maybe_prior_recompile;
2602 return pdata;
2603 }
2604
2605 void
2606 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2607 ResourceMark rm;
2608 // Ignored outputs:
2609 uint ignore_this_trap_count;
2610 bool ignore_maybe_prior_trap;
2611 bool ignore_maybe_prior_recompile;
2612 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2613 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2614 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2615
2616 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2617 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2618
2619 query_update_method_data(trap_mdo, trap_bci,
2620 (DeoptReason)reason,
2621 update_total_counts,
2622 #if INCLUDE_JVMCI
2623 false,
2624 #endif
2625 nullptr,
2626 ignore_this_trap_count,
2627 ignore_maybe_prior_trap,
2628 ignore_maybe_prior_recompile);
2629 }
2630
2631 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2632 // Enable WXWrite: current function is called from methods compiled by C2 directly
2633 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2634
2635 // Still in Java no safepoints
2636 {
2637 // This enters VM and may safepoint
2638 uncommon_trap_inner(current, trap_request);
2639 }
2640 HandleMark hm(current);
2641 return fetch_unroll_info_helper(current, exec_mode);
2642 }
2643
2644 // Local derived constants.
2645 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2646 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2647 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2648
2649 //---------------------------trap_state_reason---------------------------------
2650 Deoptimization::DeoptReason
2651 Deoptimization::trap_state_reason(int trap_state) {
2652 // This assert provides the link between the width of DataLayout::trap_bits
2653 // and the encoding of "recorded" reasons. It ensures there are enough
2654 // bits to store all needed reasons in the per-BCI MDO profile.
2655 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2656 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2657 trap_state -= recompile_bit;
2658 if (trap_state == DS_REASON_MASK) {
2659 return Reason_many;
2660 } else {
2661 assert((int)Reason_none == 0, "state=0 => Reason_none");
2662 return (DeoptReason)trap_state;
2663 }
2664 }
2665 //-------------------------trap_state_has_reason-------------------------------
2666 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2667 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2668 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2669 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2670 trap_state -= recompile_bit;
2671 if (trap_state == DS_REASON_MASK) {
2672 return -1; // true, unspecifically (bottom of state lattice)
2673 } else if (trap_state == reason) {
2674 return 1; // true, definitely
2675 } else if (trap_state == 0) {
2676 return 0; // false, definitely (top of state lattice)
2677 } else {
2678 return 0; // false, definitely
2679 }
2680 }
2681 //-------------------------trap_state_add_reason-------------------------------
2682 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2683 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2684 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2685 trap_state -= recompile_bit;
2686 if (trap_state == DS_REASON_MASK) {
2687 return trap_state + recompile_bit; // already at state lattice bottom
2688 } else if (trap_state == reason) {
2689 return trap_state + recompile_bit; // the condition is already true
2690 } else if (trap_state == 0) {
2691 return reason + recompile_bit; // no condition has yet been true
2692 } else {
2693 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2694 }
2695 }
2696 //-----------------------trap_state_is_recompiled------------------------------
2697 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2698 return (trap_state & DS_RECOMPILE_BIT) != 0;
2699 }
2700 //-----------------------trap_state_set_recompiled-----------------------------
2701 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2702 if (z) return trap_state | DS_RECOMPILE_BIT;
2703 else return trap_state & ~DS_RECOMPILE_BIT;
2704 }
2705 //---------------------------format_trap_state---------------------------------
2706 // This is used for debugging and diagnostics, including LogFile output.
2707 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2708 int trap_state) {
2709 assert(buflen > 0, "sanity");
2710 DeoptReason reason = trap_state_reason(trap_state);
2711 bool recomp_flag = trap_state_is_recompiled(trap_state);
2712 // Re-encode the state from its decoded components.
2713 int decoded_state = 0;
2714 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2715 decoded_state = trap_state_add_reason(decoded_state, reason);
2716 if (recomp_flag)
2717 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2718 // If the state re-encodes properly, format it symbolically.
2719 // Because this routine is used for debugging and diagnostics,
2720 // be robust even if the state is a strange value.
2721 size_t len;
2722 if (decoded_state != trap_state) {
2723 // Random buggy state that doesn't decode??
2724 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2725 } else {
2726 len = jio_snprintf(buf, buflen, "%s%s",
2727 trap_reason_name(reason),
2728 recomp_flag ? " recompiled" : "");
2729 }
2730 return buf;
2731 }
2732
2733
2734 //--------------------------------statics--------------------------------------
2735 const char* Deoptimization::_trap_reason_name[] = {
2736 // Note: Keep this in sync. with enum DeoptReason.
2737 "none",
2738 "null_check",
2739 "null_assert" JVMCI_ONLY("_or_unreached0"),
2740 "range_check",
2741 "class_check",
2742 "array_check",
2743 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2744 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2745 "profile_predicate",
2746 "auto_vectorization_check",
2747 "unloaded",
2748 "uninitialized",
2749 "initialized",
2750 "unreached",
2751 "unhandled",
2752 "constraint",
2753 "div0_check",
2754 "age",
2755 "predicate",
2756 "loop_limit_check",
2757 "speculate_class_check",
2758 "speculate_null_check",
2759 "speculate_null_assert",
2760 "unstable_if",
2761 "unstable_fused_if",
2762 "receiver_constraint",
2763 #if INCLUDE_JVMCI
2764 "aliasing",
2765 "transfer_to_interpreter",
2766 "not_compiled_exception_handler",
2767 "unresolved",
2768 "jsr_mismatch",
2769 #endif
2770 "tenured"
2771 };
2772 const char* Deoptimization::_trap_action_name[] = {
2773 // Note: Keep this in sync. with enum DeoptAction.
2774 "none",
2775 "maybe_recompile",
2776 "reinterpret",
2777 "make_not_entrant",
2778 "make_not_compilable"
2779 };
2780
2781 const char* Deoptimization::trap_reason_name(int reason) {
2782 // Check that every reason has a name
2783 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2784
2785 if (reason == Reason_many) return "many";
2786 if ((uint)reason < Reason_LIMIT)
2787 return _trap_reason_name[reason];
2788 static char buf[20];
2789 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2790 return buf;
2791 }
2792 const char* Deoptimization::trap_action_name(int action) {
2793 // Check that every action has a name
2794 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2795
2796 if ((uint)action < Action_LIMIT)
2797 return _trap_action_name[action];
2798 static char buf[20];
2799 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2800 return buf;
2801 }
2802
2803 // This is used for debugging and diagnostics, including LogFile output.
2804 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2805 int trap_request) {
2806 jint unloaded_class_index = trap_request_index(trap_request);
2807 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2808 const char* action = trap_action_name(trap_request_action(trap_request));
2809 #if INCLUDE_JVMCI
2810 int debug_id = trap_request_debug_id(trap_request);
2811 #endif
2812 size_t len;
2813 if (unloaded_class_index < 0) {
2814 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2815 reason, action
2816 #if INCLUDE_JVMCI
2817 ,debug_id
2818 #endif
2819 );
2820 } else {
2821 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2822 reason, action, unloaded_class_index
2823 #if INCLUDE_JVMCI
2824 ,debug_id
2825 #endif
2826 );
2827 }
2828 return buf;
2829 }
2830
2831 juint Deoptimization::_deoptimization_hist
2832 [Deoptimization::Reason_LIMIT]
2833 [1 + Deoptimization::Action_LIMIT]
2834 [Deoptimization::BC_CASE_LIMIT]
2835 = {0};
2836
2837 enum {
2838 LSB_BITS = 8,
2839 LSB_MASK = right_n_bits(LSB_BITS)
2840 };
2841
2842 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2843 Bytecodes::Code bc) {
2844 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2845 assert(action >= 0 && action < Action_LIMIT, "oob");
2846 _deoptimization_hist[Reason_none][0][0] += 1; // total
2847 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2848 juint* cases = _deoptimization_hist[reason][1+action];
2849 juint* bc_counter_addr = nullptr;
2850 juint bc_counter = 0;
2851 // Look for an unused counter, or an exact match to this BC.
2852 if (bc != Bytecodes::_illegal) {
2853 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2854 juint* counter_addr = &cases[bc_case];
2855 juint counter = *counter_addr;
2856 if ((counter == 0 && bc_counter_addr == nullptr)
2857 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2858 // this counter is either free or is already devoted to this BC
2859 bc_counter_addr = counter_addr;
2860 bc_counter = counter | bc;
2861 }
2862 }
2863 }
2864 if (bc_counter_addr == nullptr) {
2865 // Overflow, or no given bytecode.
2866 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2867 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2868 }
2869 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2870 }
2871
2872 jint Deoptimization::total_deoptimization_count() {
2873 return _deoptimization_hist[Reason_none][0][0];
2874 }
2875
2876 // Get the deopt count for a specific reason and a specific action. If either
2877 // one of 'reason' or 'action' is null, the method returns the sum of all
2878 // deoptimizations with the specific 'action' or 'reason' respectively.
2879 // If both arguments are null, the method returns the total deopt count.
2880 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2881 if (reason_str == nullptr && action_str == nullptr) {
2882 return total_deoptimization_count();
2883 }
2884 juint counter = 0;
2885 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2886 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2887 for (int action = 0; action < Action_LIMIT; action++) {
2888 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2889 juint* cases = _deoptimization_hist[reason][1+action];
2890 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2891 counter += cases[bc_case] >> LSB_BITS;
2892 }
2893 }
2894 }
2895 }
2896 }
2897 return counter;
2898 }
2899
2900 void Deoptimization::print_statistics() {
2901 juint total = total_deoptimization_count();
2902 juint account = total;
2903 if (total != 0) {
2904 ttyLocker ttyl;
2905 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2906 tty->print_cr("Deoptimization traps recorded:");
2907 #define PRINT_STAT_LINE(name, r) \
2908 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2909 PRINT_STAT_LINE("total", total);
2910 // For each non-zero entry in the histogram, print the reason,
2911 // the action, and (if specifically known) the type of bytecode.
2912 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2913 for (int action = 0; action < Action_LIMIT; action++) {
2914 juint* cases = _deoptimization_hist[reason][1+action];
2915 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2916 juint counter = cases[bc_case];
2917 if (counter != 0) {
2918 char name[1*K];
2919 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2920 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2921 bc = Bytecodes::_illegal;
2922 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
2923 trap_reason_name(reason),
2924 trap_action_name(action),
2925 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2926 juint r = counter >> LSB_BITS;
2927 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2928 account -= r;
2929 }
2930 }
2931 }
2932 }
2933 if (account != 0) {
2934 PRINT_STAT_LINE("unaccounted", account);
2935 }
2936 #undef PRINT_STAT_LINE
2937 if (xtty != nullptr) xtty->tail("statistics");
2938 }
2939 }
2940
2941 #else // COMPILER2_OR_JVMCI
2942
2943
2944 // Stubs for C1 only system.
2945 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2946 return false;
2947 }
2948
2949 const char* Deoptimization::trap_reason_name(int reason) {
2950 return "unknown";
2951 }
2952
2953 jint Deoptimization::total_deoptimization_count() {
2954 return 0;
2955 }
2956
2957 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2958 return 0;
2959 }
2960
2961 void Deoptimization::print_statistics() {
2962 // no output
2963 }
2964
2965 void
2966 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2967 // no update
2968 }
2969
2970 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2971 return 0;
2972 }
2973
2974 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2975 Bytecodes::Code bc) {
2976 // no update
2977 }
2978
2979 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2980 int trap_state) {
2981 jio_snprintf(buf, buflen, "#%d", trap_state);
2982 return buf;
2983 }
2984
2985 #endif // COMPILER2_OR_JVMCI