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