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