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 "precompiled.hpp"
26 #include "classfile/javaClasses.inline.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmClasses.hpp"
30 #include "code/codeCache.hpp"
31 #include "code/debugInfoRec.hpp"
32 #include "code/nmethod.hpp"
33 #include "code/pcDesc.hpp"
34 #include "code/scopeDesc.hpp"
35 #include "compiler/compilationPolicy.hpp"
36 #include "compiler/compilerDefinitions.inline.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "gc/shared/memAllocator.hpp"
39 #include "interpreter/bytecode.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/method.hpp"
55 #include "oops/objArrayKlass.hpp"
56 #include "oops/objArrayOop.inline.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "oops/typeArrayOop.inline.hpp"
59 #include "oops/verifyOopClosure.hpp"
60 #include "prims/jvmtiDeferredUpdates.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "prims/jvmtiThreadState.hpp"
63 #include "prims/methodHandles.hpp"
64 #include "prims/vectorSupport.hpp"
65 #include "runtime/atomic.hpp"
66 #include "runtime/basicLock.inline.hpp"
67 #include "runtime/continuation.hpp"
68 #include "runtime/continuationEntry.inline.hpp"
69 #include "runtime/deoptimization.hpp"
70 #include "runtime/escapeBarrier.hpp"
71 #include "runtime/fieldDescriptor.hpp"
72 #include "runtime/fieldDescriptor.inline.hpp"
73 #include "runtime/frame.inline.hpp"
74 #include "runtime/handles.inline.hpp"
75 #include "runtime/interfaceSupport.inline.hpp"
76 #include "runtime/javaThread.hpp"
77 #include "runtime/jniHandles.inline.hpp"
78 #include "runtime/keepStackGCProcessed.hpp"
79 #include "runtime/lightweightSynchronizer.hpp"
80 #include "runtime/lockStack.inline.hpp"
81 #include "runtime/objectMonitor.inline.hpp"
82 #include "runtime/osThread.hpp"
83 #include "runtime/perfData.inline.hpp"
84 #include "runtime/safepointVerifiers.hpp"
85 #include "runtime/sharedRuntime.hpp"
86 #include "runtime/signature.hpp"
87 #include "runtime/stackFrameStream.inline.hpp"
88 #include "runtime/stackValue.hpp"
89 #include "runtime/stackWatermarkSet.hpp"
90 #include "runtime/stubRoutines.hpp"
91 #include "runtime/synchronizer.inline.hpp"
92 #include "runtime/threadSMR.hpp"
93 #include "runtime/threadWXSetters.inline.hpp"
94 #include "runtime/vframe.hpp"
95 #include "runtime/vframeArray.hpp"
96 #include "runtime/vframe_hp.hpp"
97 #include "runtime/vmOperations.hpp"
98 #include "services/management.hpp"
99 #include "utilities/checkedCast.hpp"
100 #include "utilities/events.hpp"
101 #include "utilities/growableArray.hpp"
102 #include "utilities/macros.hpp"
103 #include "utilities/preserveException.hpp"
104 #include "utilities/xmlstream.hpp"
105 #if INCLUDE_JFR
106 #include "jfr/jfrEvents.hpp"
107 #include "jfr/metadata/jfrSerializer.hpp"
108 #endif
109
110 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
111 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
112 bool DeoptimizationScope::_committing_in_progress = false;
113
114 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
115 DEBUG_ONLY(_deopted = false;)
116
117 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
118 // If there is nothing to deopt _required_gen is the same as comitted.
119 _required_gen = DeoptimizationScope::_committed_deopt_gen;
120 }
121
122 DeoptimizationScope::~DeoptimizationScope() {
123 assert(_deopted, "Deopt not executed");
124 }
125
126 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
127 if (!nm->can_be_deoptimized()) {
128 return;
129 }
130
131 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
132
133 // If it's already marked but we still need it to be deopted.
134 if (nm->is_marked_for_deoptimization()) {
135 dependent(nm);
136 return;
137 }
138
139 nmethod::DeoptimizationStatus status =
140 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
141 Atomic::store(&nm->_deoptimization_status, status);
142
143 // Make sure active is not committed
144 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
145 assert(nm->_deoptimization_generation == 0, "Is already marked");
146
147 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
148 _required_gen = DeoptimizationScope::_active_deopt_gen;
149 }
150
151 void DeoptimizationScope::dependent(nmethod* nm) {
152 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
153
154 // A method marked by someone else may have a _required_gen lower than what we marked with.
155 // Therefore only store it if it's higher than _required_gen.
156 if (_required_gen < nm->_deoptimization_generation) {
157 _required_gen = nm->_deoptimization_generation;
158 }
159 }
160
161 void DeoptimizationScope::deoptimize_marked() {
162 assert(!_deopted, "Already deopted");
163
164 // We are not alive yet.
165 if (!Universe::is_fully_initialized()) {
166 DEBUG_ONLY(_deopted = true;)
167 return;
168 }
169
170 // Safepoints are a special case, handled here.
171 if (SafepointSynchronize::is_at_safepoint()) {
172 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
173 DeoptimizationScope::_active_deopt_gen++;
174 Deoptimization::deoptimize_all_marked();
175 DEBUG_ONLY(_deopted = true;)
176 return;
177 }
178
179 uint64_t comitting = 0;
180 bool wait = false;
181 while (true) {
182 {
183 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
184
185 // First we check if we or someone else already deopted the gen we want.
186 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
187 DEBUG_ONLY(_deopted = true;)
188 return;
189 }
190 if (!_committing_in_progress) {
191 // The version we are about to commit.
192 comitting = DeoptimizationScope::_active_deopt_gen;
193 // Make sure new marks use a higher gen.
194 DeoptimizationScope::_active_deopt_gen++;
195 _committing_in_progress = true;
196 wait = false;
197 } else {
198 // Another thread is handshaking and committing a gen.
199 wait = true;
200 }
201 }
202 if (wait) {
203 // Wait and let the concurrent handshake be performed.
204 ThreadBlockInVM tbivm(JavaThread::current());
205 os::naked_yield();
206 } else {
207 // Performs the handshake.
208 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
209 DEBUG_ONLY(_deopted = true;)
210 {
211 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
212
213 // Make sure that committed doesn't go backwards.
214 // Should only happen if we did a deopt during a safepoint above.
215 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
216 DeoptimizationScope::_committed_deopt_gen = comitting;
217 }
218 _committing_in_progress = false;
219
220 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
221
222 return;
223 }
224 }
225 }
226 }
227
228 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
229 int caller_adjustment,
230 int caller_actual_parameters,
231 int number_of_frames,
232 intptr_t* frame_sizes,
233 address* frame_pcs,
234 BasicType return_type,
235 int exec_mode) {
236 _size_of_deoptimized_frame = size_of_deoptimized_frame;
237 _caller_adjustment = caller_adjustment;
238 _caller_actual_parameters = caller_actual_parameters;
239 _number_of_frames = number_of_frames;
240 _frame_sizes = frame_sizes;
241 _frame_pcs = frame_pcs;
242 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
243 _return_type = return_type;
244 _initial_info = 0;
245 // PD (x86 only)
246 _counter_temp = 0;
247 _unpack_kind = exec_mode;
248 _sender_sp_temp = 0;
249
250 _total_frame_sizes = size_of_frames();
251 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
252 }
253
254 Deoptimization::UnrollBlock::~UnrollBlock() {
255 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
256 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
257 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
258 }
259
260 int Deoptimization::UnrollBlock::size_of_frames() const {
261 // Account first for the adjustment of the initial frame
262 intptr_t result = _caller_adjustment;
263 for (int index = 0; index < number_of_frames(); index++) {
264 result += frame_sizes()[index];
265 }
266 return checked_cast<int>(result);
267 }
268
269 void Deoptimization::UnrollBlock::print() {
270 ResourceMark rm;
271 stringStream st;
272 st.print_cr("UnrollBlock");
273 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
274 st.print( " frame_sizes: ");
275 for (int index = 0; index < number_of_frames(); index++) {
276 st.print("%zd ", frame_sizes()[index]);
277 }
278 st.cr();
279 tty->print_raw(st.freeze());
280 }
281
282 // In order to make fetch_unroll_info work properly with escape
283 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
284 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
285 // which is called from the method fetch_unroll_info_helper below.
286 JRT_BLOCK_ENTRY_PROF(Deoptimization::UnrollBlock*, Deoptimization, fetch_unroll_info, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
287 // fetch_unroll_info() is called at the beginning of the deoptimization
288 // handler. Note this fact before we start generating temporary frames
289 // that can confuse an asynchronous stack walker. This counter is
290 // decremented at the end of unpack_frames().
291 current->inc_in_deopt_handler();
292
293 if (exec_mode == Unpack_exception) {
294 // When we get here, a callee has thrown an exception into a deoptimized
295 // frame. That throw might have deferred stack watermark checking until
296 // after unwinding. So we deal with such deferred requests here.
297 StackWatermarkSet::after_unwind(current);
298 }
299
300 return fetch_unroll_info_helper(current, exec_mode);
301 JRT_END
302
303 #if COMPILER2_OR_JVMCI
304 // print information about reallocated objects
305 static void print_objects(JavaThread* deoptee_thread,
306 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
307 ResourceMark rm;
308 stringStream st; // change to logStream with logging
309 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
310 fieldDescriptor fd;
311
312 for (int i = 0; i < objects->length(); i++) {
313 ObjectValue* sv = (ObjectValue*) objects->at(i);
314 Handle obj = sv->value();
315
316 if (obj.is_null()) {
317 st.print_cr(" nullptr");
318 continue;
319 }
320
321 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
322
323 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
324 k->print_value_on(&st);
325 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize);
326
327 if (Verbose && k != nullptr) {
328 k->oop_print_on(obj(), &st);
329 }
330 }
331 tty->print_raw(st.freeze());
332 }
333
334 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
335 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
336 bool& deoptimized_objects) {
337 bool realloc_failures = false;
338 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
339
340 JavaThread* deoptee_thread = chunk->at(0)->thread();
341 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
342 "a frame can only be deoptimized by the owner thread");
343
344 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
345
346 // The flag return_oop() indicates call sites which return oop
347 // in compiled code. Such sites include java method calls,
348 // runtime calls (for example, used to allocate new objects/arrays
349 // on slow code path) and any other calls generated in compiled code.
350 // It is not guaranteed that we can get such information here only
351 // by analyzing bytecode in deoptimized frames. This is why this flag
352 // is set during method compilation (see Compile::Process_OopMap_Node()).
353 // If the previous frame was popped or if we are dispatching an exception,
354 // we don't have an oop result.
355 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
356 Handle return_value;
357 if (save_oop_result) {
358 // Reallocation may trigger GC. If deoptimization happened on return from
359 // call which returns oop we need to save it since it is not in oopmap.
360 oop result = deoptee.saved_oop_result(&map);
361 assert(oopDesc::is_oop_or_null(result), "must be oop");
362 return_value = Handle(thread, result);
363 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
364 if (TraceDeoptimization) {
365 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
366 tty->cr();
367 }
368 }
369 if (objects != nullptr) {
370 if (exec_mode == Deoptimization::Unpack_none) {
371 assert(thread->thread_state() == _thread_in_vm, "assumption");
372 JavaThread* THREAD = thread; // For exception macros.
373 // Clear pending OOM if reallocation fails and return true indicating allocation failure
374 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
375 deoptimized_objects = true;
376 } else {
377 JavaThread* current = thread; // For JRT_BLOCK
378 JRT_BLOCK
379 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
380 JRT_END
381 }
382 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci();
383 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
384 if (TraceDeoptimization) {
385 print_objects(deoptee_thread, objects, realloc_failures);
386 }
387 }
388 if (save_oop_result) {
389 // Restore result.
390 deoptee.set_saved_oop_result(&map, return_value());
391 }
392 return realloc_failures;
393 }
394
395 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
396 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
397 JavaThread* deoptee_thread = chunk->at(0)->thread();
398 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
399 assert(thread == Thread::current(), "should be");
400 HandleMark hm(thread);
401 #ifndef PRODUCT
402 bool first = true;
403 #endif // !PRODUCT
404 // Start locking from outermost/oldest frame
405 for (int i = (chunk->length() - 1); i >= 0; i--) {
406 compiledVFrame* cvf = chunk->at(i);
407 assert (cvf->scope() != nullptr,"expect only compiled java frames");
408 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
409 if (monitors->is_nonempty()) {
410 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
411 exec_mode, realloc_failures);
412 deoptimized_objects = deoptimized_objects || relocked;
413 #ifndef PRODUCT
414 if (PrintDeoptimizationDetails) {
415 ResourceMark rm;
416 stringStream st;
417 for (int j = 0; j < monitors->length(); j++) {
418 MonitorInfo* mi = monitors->at(j);
419 if (mi->eliminated()) {
420 if (first) {
421 first = false;
422 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
423 }
424 if (exec_mode == Deoptimization::Unpack_none) {
425 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
426 if (monitor != nullptr && monitor->object() == mi->owner()) {
427 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
428 continue;
429 }
430 }
431 if (mi->owner_is_scalar_replaced()) {
432 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
433 st.print_cr(" failed reallocation for klass %s", k->external_name());
434 } else {
435 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
436 }
437 }
438 }
439 tty->print_raw(st.freeze());
440 }
441 #endif // !PRODUCT
442 }
443 }
444 }
445
446 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
447 // The given vframes cover one physical frame.
448 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
449 bool& realloc_failures) {
450 frame deoptee = chunk->at(0)->fr();
451 JavaThread* deoptee_thread = chunk->at(0)->thread();
452 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
453 RegisterMap map(chunk->at(0)->register_map());
454 bool deoptimized_objects = false;
455
456 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
457
458 // Reallocate the non-escaping objects and restore their fields.
459 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
460 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
461 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
462 }
463
464 // MonitorInfo structures used in eliminate_locks are not GC safe.
465 NoSafepointVerifier no_safepoint;
466
467 // Now relock objects if synchronization on them was eliminated.
468 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
469 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
470 }
471 return deoptimized_objects;
472 }
473 #endif // COMPILER2_OR_JVMCI
474
475 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
476 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
477 // When we get here we are about to unwind the deoptee frame. In order to
478 // catch not yet safe to use frames, the following stack watermark barrier
479 // poll will make such frames safe to use.
480 StackWatermarkSet::before_unwind(current);
481
482 // Note: there is a safepoint safety issue here. No matter whether we enter
483 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
484 // the vframeArray is created.
485 //
486
487 // Allocate our special deoptimization ResourceMark
488 DeoptResourceMark* dmark = new DeoptResourceMark(current);
489 assert(current->deopt_mark() == nullptr, "Pending deopt!");
490 current->set_deopt_mark(dmark);
491
492 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
493 RegisterMap map(current,
494 RegisterMap::UpdateMap::include,
495 RegisterMap::ProcessFrames::include,
496 RegisterMap::WalkContinuation::skip);
497 RegisterMap dummy_map(current,
498 RegisterMap::UpdateMap::skip,
499 RegisterMap::ProcessFrames::include,
500 RegisterMap::WalkContinuation::skip);
501 // Now get the deoptee with a valid map
502 frame deoptee = stub_frame.sender(&map);
503 // Set the deoptee nmethod
504 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
505 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
506 current->set_deopt_compiled_method(nm);
507
508 if (VerifyStack) {
509 current->validate_frame_layout();
510 }
511
512 // Create a growable array of VFrames where each VFrame represents an inlined
513 // Java frame. This storage is allocated with the usual system arena.
514 assert(deoptee.is_compiled_frame(), "Wrong frame type");
515 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
516 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
517 while (!vf->is_top()) {
518 assert(vf->is_compiled_frame(), "Wrong frame type");
519 chunk->push(compiledVFrame::cast(vf));
520 vf = vf->sender();
521 }
522 assert(vf->is_compiled_frame(), "Wrong frame type");
523 chunk->push(compiledVFrame::cast(vf));
524
525 bool realloc_failures = false;
526
527 #if COMPILER2_OR_JVMCI
528 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
529
530 // Reallocate the non-escaping objects and restore their fields. Then
531 // relock objects if synchronization on them was eliminated.
532 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
533 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
534 bool unused;
535 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
536 }
537 #endif // COMPILER2_OR_JVMCI
538
539 // Ensure that no safepoint is taken after pointers have been stored
540 // in fields of rematerialized objects. If a safepoint occurs from here on
541 // out the java state residing in the vframeArray will be missed.
542 // Locks may be rebaised in a safepoint.
543 NoSafepointVerifier no_safepoint;
544
545 #if COMPILER2_OR_JVMCI
546 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
547 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
548 bool unused = false;
549 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
550 }
551 #endif // COMPILER2_OR_JVMCI
552
553 ScopeDesc* trap_scope = chunk->at(0)->scope();
554 Handle exceptionObject;
555 if (trap_scope->rethrow_exception()) {
556 #ifndef PRODUCT
557 if (PrintDeoptimizationDetails) {
558 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());
559 }
560 #endif // !PRODUCT
561
562 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
563 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
564 ScopeValue* topOfStack = expressions->top();
565 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
566 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
567 }
568
569 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
570 #if COMPILER2_OR_JVMCI
571 if (realloc_failures) {
572 // This destroys all ScopedValue bindings.
573 current->clear_scopedValueBindings();
574 pop_frames_failed_reallocs(current, array);
575 }
576 #endif
577
578 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
579 current->set_vframe_array_head(array);
580
581 // Now that the vframeArray has been created if we have any deferred local writes
582 // added by jvmti then we can free up that structure as the data is now in the
583 // vframeArray
584
585 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
586
587 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
588 CodeBlob* cb = stub_frame.cb();
589 // Verify we have the right vframeArray
590 assert(cb->frame_size() >= 0, "Unexpected frame size");
591 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
592
593 // If the deopt call site is a MethodHandle invoke call site we have
594 // to adjust the unpack_sp.
595 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
596 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
597 unpack_sp = deoptee.unextended_sp();
598
599 #ifdef ASSERT
600 assert(cb->is_deoptimization_stub() ||
601 cb->is_uncommon_trap_stub() ||
602 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
603 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
604 "unexpected code blob: %s", cb->name());
605 #endif
606
607 // This is a guarantee instead of an assert because if vframe doesn't match
608 // we will unpack the wrong deoptimized frame and wind up in strange places
609 // where it will be very difficult to figure out what went wrong. Better
610 // to die an early death here than some very obscure death later when the
611 // trail is cold.
612 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
613
614 int number_of_frames = array->frames();
615
616 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
617 // virtual activation, which is the reverse of the elements in the vframes array.
618 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
619 // +1 because we always have an interpreter return address for the final slot.
620 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
621 int popframe_extra_args = 0;
622 // Create an interpreter return address for the stub to use as its return
623 // address so the skeletal frames are perfectly walkable
624 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
625
626 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
627 // activation be put back on the expression stack of the caller for reexecution
628 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
629 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
630 }
631
632 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
633 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
634 // than simply use array->sender.pc(). This requires us to walk the current set of frames
635 //
636 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
637 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
638
639 // It's possible that the number of parameters at the call site is
640 // different than number of arguments in the callee when method
641 // handles are used. If the caller is interpreted get the real
642 // value so that the proper amount of space can be added to it's
643 // frame.
644 bool caller_was_method_handle = false;
645 if (deopt_sender.is_interpreted_frame()) {
646 methodHandle method(current, deopt_sender.interpreter_frame_method());
647 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
648 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
649 // Method handle invokes may involve fairly arbitrary chains of
650 // calls so it's impossible to know how much actual space the
651 // caller has for locals.
652 caller_was_method_handle = true;
653 }
654 }
655
656 //
657 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
658 // frame_sizes/frame_pcs[1] next oldest frame (int)
659 // frame_sizes/frame_pcs[n] youngest frame (int)
660 //
661 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
662 // owns the space for the return address to it's caller). Confusing ain't it.
663 //
664 // The vframe array can address vframes with indices running from
665 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
666 // When we create the skeletal frames we need the oldest frame to be in the zero slot
667 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
668 // so things look a little strange in this loop.
669 //
670 int callee_parameters = 0;
671 int callee_locals = 0;
672 for (int index = 0; index < array->frames(); index++ ) {
673 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
674 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
675 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
676 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
677 callee_locals,
678 index == 0,
679 popframe_extra_args);
680 // This pc doesn't have to be perfect just good enough to identify the frame
681 // as interpreted so the skeleton frame will be walkable
682 // The correct pc will be set when the skeleton frame is completely filled out
683 // The final pc we store in the loop is wrong and will be overwritten below
684 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
685
686 callee_parameters = array->element(index)->method()->size_of_parameters();
687 callee_locals = array->element(index)->method()->max_locals();
688 popframe_extra_args = 0;
689 }
690
691 // Compute whether the root vframe returns a float or double value.
692 BasicType return_type;
693 {
694 methodHandle method(current, array->element(0)->method());
695 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
696 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
697 }
698
699 // Compute information for handling adapters and adjusting the frame size of the caller.
700 int caller_adjustment = 0;
701
702 // Compute the amount the oldest interpreter frame will have to adjust
703 // its caller's stack by. If the caller is a compiled frame then
704 // we pretend that the callee has no parameters so that the
705 // extension counts for the full amount of locals and not just
706 // locals-parms. This is because without a c2i adapter the parm
707 // area as created by the compiled frame will not be usable by
708 // the interpreter. (Depending on the calling convention there
709 // may not even be enough space).
710
711 // QQQ I'd rather see this pushed down into last_frame_adjust
712 // and have it take the sender (aka caller).
713
714 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
715 caller_adjustment = last_frame_adjust(0, callee_locals);
716 } else if (callee_locals > callee_parameters) {
717 // The caller frame may need extending to accommodate
718 // non-parameter locals of the first unpacked interpreted frame.
719 // Compute that adjustment.
720 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
721 }
722
723 // If the sender is deoptimized the we must retrieve the address of the handler
724 // since the frame will "magically" show the original pc before the deopt
725 // and we'd undo the deopt.
726
727 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
728 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
729 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
730 }
731
732 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
733
734 #if INCLUDE_JVMCI
735 if (exceptionObject() != nullptr) {
736 current->set_exception_oop(exceptionObject());
737 exec_mode = Unpack_exception;
738 }
739 #endif
740
741 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
742 assert(current->has_pending_exception(), "should have thrown OOME");
743 current->set_exception_oop(current->pending_exception());
744 current->clear_pending_exception();
745 exec_mode = Unpack_exception;
746 }
747
748 #if INCLUDE_JVMCI
749 if (current->frames_to_pop_failed_realloc() > 0) {
750 current->set_pending_monitorenter(false);
751 }
752 #endif
753
754 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
755 caller_adjustment * BytesPerWord,
756 caller_was_method_handle ? 0 : callee_parameters,
757 number_of_frames,
758 frame_sizes,
759 frame_pcs,
760 return_type,
761 exec_mode);
762 // On some platforms, we need a way to pass some platform dependent
763 // information to the unpacking code so the skeletal frames come out
764 // correct (initial fp value, unextended sp, ...)
765 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
766
767 if (array->frames() > 1) {
768 if (VerifyStack && TraceDeoptimization) {
769 tty->print_cr("Deoptimizing method containing inlining");
770 }
771 }
772
773 array->set_unroll_block(info);
774 return info;
775 }
776
777 // Called to cleanup deoptimization data structures in normal case
778 // after unpacking to stack and when stack overflow error occurs
779 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
780 vframeArray *array) {
781
782 // Get array if coming from exception
783 if (array == nullptr) {
784 array = thread->vframe_array_head();
785 }
786 thread->set_vframe_array_head(nullptr);
787
788 // Free the previous UnrollBlock
789 vframeArray* old_array = thread->vframe_array_last();
790 thread->set_vframe_array_last(array);
791
792 if (old_array != nullptr) {
793 UnrollBlock* old_info = old_array->unroll_block();
794 old_array->set_unroll_block(nullptr);
795 delete old_info;
796 delete old_array;
797 }
798
799 // Deallocate any resource creating in this routine and any ResourceObjs allocated
800 // inside the vframeArray (StackValueCollections)
801
802 delete thread->deopt_mark();
803 thread->set_deopt_mark(nullptr);
804 thread->set_deopt_compiled_method(nullptr);
805
806
807 if (JvmtiExport::can_pop_frame()) {
808 // Regardless of whether we entered this routine with the pending
809 // popframe condition bit set, we should always clear it now
810 thread->clear_popframe_condition();
811 }
812
813 // unpack_frames() is called at the end of the deoptimization handler
814 // and (in C2) at the end of the uncommon trap handler. Note this fact
815 // so that an asynchronous stack walker can work again. This counter is
816 // incremented at the beginning of fetch_unroll_info() and (in C2) at
817 // the beginning of uncommon_trap().
818 thread->dec_in_deopt_handler();
819 }
820
821 // Moved from cpu directories because none of the cpus has callee save values.
822 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
823 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
824
825 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
826 // the days we had adapter frames. When we deoptimize a situation where a
827 // compiled caller calls a compiled caller will have registers it expects
828 // to survive the call to the callee. If we deoptimize the callee the only
829 // way we can restore these registers is to have the oldest interpreter
830 // frame that we create restore these values. That is what this routine
831 // will accomplish.
832
833 // At the moment we have modified c2 to not have any callee save registers
834 // so this problem does not exist and this routine is just a place holder.
835
836 assert(f->is_interpreted_frame(), "must be interpreted");
837 }
838
839 #ifndef PRODUCT
840 static bool falls_through(Bytecodes::Code bc) {
841 switch (bc) {
842 // List may be incomplete. Here we really only care about bytecodes where compiled code
843 // can deoptimize.
844 case Bytecodes::_goto:
845 case Bytecodes::_goto_w:
846 case Bytecodes::_athrow:
847 return false;
848 default:
849 return true;
850 }
851 }
852 #endif
853
854 // Return BasicType of value being returned
855 JRT_LEAF_PROF_NO_THREAD(BasicType, Deoptimization, unpack_frames, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
856 assert(thread == JavaThread::current(), "pre-condition");
857
858 // We are already active in the special DeoptResourceMark any ResourceObj's we
859 // allocate will be freed at the end of the routine.
860
861 // JRT_LEAF methods don't normally allocate handles and there is a
862 // NoHandleMark to enforce that. It is actually safe to use Handles
863 // in a JRT_LEAF method, and sometimes desirable, but to do so we
864 // must use ResetNoHandleMark to bypass the NoHandleMark, and
865 // then use a HandleMark to ensure any Handles we do create are
866 // cleaned up in this scope.
867 ResetNoHandleMark rnhm;
868 HandleMark hm(thread);
869
870 frame stub_frame = thread->last_frame();
871
872 Continuation::notify_deopt(thread, stub_frame.sp());
873
874 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
875 // must point to the vframeArray for the unpack frame.
876 vframeArray* array = thread->vframe_array_head();
877 UnrollBlock* info = array->unroll_block();
878
879 // We set the last_Java frame. But the stack isn't really parsable here. So we
880 // clear it to make sure JFR understands not to try and walk stacks from events
881 // in here.
882 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
883 thread->frame_anchor()->set_last_Java_sp(nullptr);
884
885 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
886 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
887
888 thread->frame_anchor()->set_last_Java_sp(sp);
889
890 BasicType bt = info->return_type();
891
892 // If we have an exception pending, claim that the return type is an oop
893 // so the deopt_blob does not overwrite the exception_oop.
894
895 if (exec_mode == Unpack_exception)
896 bt = T_OBJECT;
897
898 // Cleanup thread deopt data
899 cleanup_deopt_info(thread, array);
900
901 #ifndef PRODUCT
902 if (VerifyStack) {
903 ResourceMark res_mark;
904 // Clear pending exception to not break verification code (restored afterwards)
905 PreserveExceptionMark pm(thread);
906
907 thread->validate_frame_layout();
908
909 // Verify that the just-unpacked frames match the interpreter's
910 // notions of expression stack and locals
911 vframeArray* cur_array = thread->vframe_array_last();
912 RegisterMap rm(thread,
913 RegisterMap::UpdateMap::skip,
914 RegisterMap::ProcessFrames::include,
915 RegisterMap::WalkContinuation::skip);
916 rm.set_include_argument_oops(false);
917 bool is_top_frame = true;
918 int callee_size_of_parameters = 0;
919 int callee_max_locals = 0;
920 for (int i = 0; i < cur_array->frames(); i++) {
921 vframeArrayElement* el = cur_array->element(i);
922 frame* iframe = el->iframe();
923 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
924
925 // Get the oop map for this bci
926 InterpreterOopMap mask;
927 int cur_invoke_parameter_size = 0;
928 bool try_next_mask = false;
929 int next_mask_expression_stack_size = -1;
930 int top_frame_expression_stack_adjustment = 0;
931 methodHandle mh(thread, iframe->interpreter_frame_method());
932 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
933 BytecodeStream str(mh, iframe->interpreter_frame_bci());
934 int max_bci = mh->code_size();
935 // Get to the next bytecode if possible
936 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
937 // Check to see if we can grab the number of outgoing arguments
938 // at an uncommon trap for an invoke (where the compiler
939 // generates debug info before the invoke has executed)
940 Bytecodes::Code cur_code = str.next();
941 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
942 if (Bytecodes::is_invoke(cur_code)) {
943 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
944 cur_invoke_parameter_size = invoke.size_of_parameters();
945 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
946 callee_size_of_parameters++;
947 }
948 }
949 if (str.bci() < max_bci) {
950 next_code = str.next();
951 if (next_code >= 0) {
952 // The interpreter oop map generator reports results before
953 // the current bytecode has executed except in the case of
954 // calls. It seems to be hard to tell whether the compiler
955 // has emitted debug information matching the "state before"
956 // a given bytecode or the state after, so we try both
957 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
958 // Get expression stack size for the next bytecode
959 InterpreterOopMap next_mask;
960 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
961 next_mask_expression_stack_size = next_mask.expression_stack_size();
962 if (Bytecodes::is_invoke(next_code)) {
963 Bytecode_invoke invoke(mh, str.bci());
964 next_mask_expression_stack_size += invoke.size_of_parameters();
965 }
966 // Need to subtract off the size of the result type of
967 // the bytecode because this is not described in the
968 // debug info but returned to the interpreter in the TOS
969 // caching register
970 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
971 if (bytecode_result_type != T_ILLEGAL) {
972 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
973 }
974 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
975 try_next_mask = true;
976 }
977 }
978 }
979
980 // Verify stack depth and oops in frame
981 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
982 if (!(
983 /* SPARC */
984 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
985 /* x86 */
986 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
987 (try_next_mask &&
988 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
989 top_frame_expression_stack_adjustment))) ||
990 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
991 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
992 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
993 )) {
994 {
995 // Print out some information that will help us debug the problem
996 tty->print_cr("Wrong number of expression stack elements during deoptimization");
997 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
998 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
999 if (try_next_mask) {
1000 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1001 }
1002 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1003 iframe->interpreter_frame_expression_stack_size());
1004 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1005 tty->print_cr(" try_next_mask = %d", try_next_mask);
1006 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1007 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1008 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1009 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1010 tty->print_cr(" exec_mode = %d", exec_mode);
1011 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1012 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1013 tty->print_cr(" Interpreted frames:");
1014 for (int k = 0; k < cur_array->frames(); k++) {
1015 vframeArrayElement* el = cur_array->element(k);
1016 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1017 }
1018 cur_array->print_on_2(tty);
1019 }
1020 guarantee(false, "wrong number of expression stack elements during deopt");
1021 }
1022 VerifyOopClosure verify;
1023 iframe->oops_interpreted_do(&verify, &rm, false);
1024 callee_size_of_parameters = mh->size_of_parameters();
1025 callee_max_locals = mh->max_locals();
1026 is_top_frame = false;
1027 }
1028 }
1029 #endif // !PRODUCT
1030
1031 return bt;
1032 JRT_END
1033
1034 class DeoptimizeMarkedClosure : public HandshakeClosure {
1035 public:
1036 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1037 void do_thread(Thread* thread) {
1038 JavaThread* jt = JavaThread::cast(thread);
1039 jt->deoptimize_marked_methods();
1040 }
1041 };
1042
1043 void Deoptimization::deoptimize_all_marked() {
1044 ResourceMark rm;
1045
1046 // Make the dependent methods not entrant
1047 CodeCache::make_marked_nmethods_deoptimized();
1048
1049 DeoptimizeMarkedClosure deopt;
1050 if (SafepointSynchronize::is_at_safepoint()) {
1051 Threads::java_threads_do(&deopt);
1052 } else {
1053 Handshake::execute(&deopt);
1054 }
1055 }
1056
1057 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1058 = Deoptimization::Action_reinterpret;
1059
1060 #if INCLUDE_JVMCI
1061 template<typename CacheType>
1062 class BoxCacheBase : public CHeapObj<mtCompiler> {
1063 protected:
1064 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1065 ResourceMark rm(thread);
1066 char* klass_name_str = klass_name->as_C_string();
1067 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1068 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1069 if (!ik->is_in_error_state()) {
1070 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1071 CacheType::compute_offsets(ik);
1072 }
1073 return ik;
1074 }
1075 };
1076
1077 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1078 PrimitiveType _low;
1079 PrimitiveType _high;
1080 jobject _cache;
1081 protected:
1082 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1083 BoxCache(Thread* thread) {
1084 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1085 if (ik->is_in_error_state()) {
1086 _low = 1;
1087 _high = 0;
1088 _cache = nullptr;
1089 } else {
1090 objArrayOop cache = CacheType::cache(ik);
1091 assert(cache->length() > 0, "Empty cache");
1092 _low = BoxType::value(cache->obj_at(0));
1093 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1094 _cache = JNIHandles::make_global(Handle(thread, cache));
1095 }
1096 }
1097 ~BoxCache() {
1098 JNIHandles::destroy_global(_cache);
1099 }
1100 public:
1101 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1102 if (_singleton == nullptr) {
1103 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1104 if (!Atomic::replace_if_null(&_singleton, s)) {
1105 delete s;
1106 }
1107 }
1108 return _singleton;
1109 }
1110 oop lookup(PrimitiveType value) {
1111 if (_low <= value && value <= _high) {
1112 int offset = checked_cast<int>(value - _low);
1113 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1114 }
1115 return nullptr;
1116 }
1117 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1118 if (_cache == nullptr) {
1119 cache_init_error = true;
1120 return nullptr;
1121 }
1122 // Have to cast to avoid little/big-endian problems.
1123 if (sizeof(PrimitiveType) > sizeof(jint)) {
1124 jlong value = (jlong)raw_value;
1125 return lookup(value);
1126 }
1127 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1128 return lookup(value);
1129 }
1130 };
1131
1132 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1133 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1134 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1135 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1136 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1137
1138 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1139 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1140 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1141 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1142 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1143
1144 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1145 jobject _true_cache;
1146 jobject _false_cache;
1147 protected:
1148 static BooleanBoxCache *_singleton;
1149 BooleanBoxCache(Thread *thread) {
1150 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1151 if (ik->is_in_error_state()) {
1152 _true_cache = nullptr;
1153 _false_cache = nullptr;
1154 } else {
1155 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1156 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1157 }
1158 }
1159 ~BooleanBoxCache() {
1160 JNIHandles::destroy_global(_true_cache);
1161 JNIHandles::destroy_global(_false_cache);
1162 }
1163 public:
1164 static BooleanBoxCache* singleton(Thread* thread) {
1165 if (_singleton == nullptr) {
1166 BooleanBoxCache* s = new BooleanBoxCache(thread);
1167 if (!Atomic::replace_if_null(&_singleton, s)) {
1168 delete s;
1169 }
1170 }
1171 return _singleton;
1172 }
1173 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1174 if (_true_cache == nullptr) {
1175 cache_in_error = true;
1176 return nullptr;
1177 }
1178 // Have to cast to avoid little/big-endian problems.
1179 jboolean value = (jboolean)*((jint*)&raw_value);
1180 return lookup(value);
1181 }
1182 oop lookup(jboolean value) {
1183 if (value != 0) {
1184 return JNIHandles::resolve_non_null(_true_cache);
1185 }
1186 return JNIHandles::resolve_non_null(_false_cache);
1187 }
1188 };
1189
1190 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1191
1192 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1193 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1194 BasicType box_type = vmClasses::box_klass_type(k);
1195 if (box_type != T_OBJECT) {
1196 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1197 switch(box_type) {
1198 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1199 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1200 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1201 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1202 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1203 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1204 default:;
1205 }
1206 }
1207 return nullptr;
1208 }
1209 #endif // INCLUDE_JVMCI
1210
1211 #if COMPILER2_OR_JVMCI
1212 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1213 Handle pending_exception(THREAD, thread->pending_exception());
1214 const char* exception_file = thread->exception_file();
1215 int exception_line = thread->exception_line();
1216 thread->clear_pending_exception();
1217
1218 bool failures = false;
1219
1220 for (int i = 0; i < objects->length(); i++) {
1221 assert(objects->at(i)->is_object(), "invalid debug information");
1222 ObjectValue* sv = (ObjectValue*) objects->at(i);
1223
1224 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1225 oop obj = nullptr;
1226
1227 bool cache_init_error = false;
1228 if (k->is_instance_klass()) {
1229 #if INCLUDE_JVMCI
1230 nmethod* nm = fr->cb()->as_nmethod_or_null();
1231 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1232 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1233 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1234 if (obj != nullptr) {
1235 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1236 abv->set_cached(true);
1237 } else if (cache_init_error) {
1238 // Results in an OOME which is valid (as opposed to a class initialization error)
1239 // and is fine for the rare case a cache initialization failing.
1240 failures = true;
1241 }
1242 }
1243 #endif // INCLUDE_JVMCI
1244
1245 InstanceKlass* ik = InstanceKlass::cast(k);
1246 if (obj == nullptr && !cache_init_error) {
1247 InternalOOMEMark iom(THREAD);
1248 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1249 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1250 } else {
1251 obj = ik->allocate_instance(THREAD);
1252 }
1253 }
1254 } else if (k->is_typeArray_klass()) {
1255 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1256 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1257 int len = sv->field_size() / type2size[ak->element_type()];
1258 InternalOOMEMark iom(THREAD);
1259 obj = ak->allocate(len, THREAD);
1260 } else if (k->is_objArray_klass()) {
1261 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1262 InternalOOMEMark iom(THREAD);
1263 obj = ak->allocate(sv->field_size(), THREAD);
1264 }
1265
1266 if (obj == nullptr) {
1267 failures = true;
1268 }
1269
1270 assert(sv->value().is_null(), "redundant reallocation");
1271 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1272 CLEAR_PENDING_EXCEPTION;
1273 sv->set_value(obj);
1274 }
1275
1276 if (failures) {
1277 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1278 } else if (pending_exception.not_null()) {
1279 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1280 }
1281
1282 return failures;
1283 }
1284
1285 #if INCLUDE_JVMCI
1286 /**
1287 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1288 * we need to somehow be able to recover the actual kind to be able to write the correct
1289 * amount of bytes.
1290 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1291 * the entries at index n + 1 to n + i are 'markers'.
1292 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1293 * expected form of the array would be:
1294 *
1295 * {b0, b1, b2, b3, INT, marker, b6, b7}
1296 *
1297 * Thus, in order to get back the size of the entry, we simply need to count the number
1298 * of marked entries
1299 *
1300 * @param virtualArray the virtualized byte array
1301 * @param i index of the virtual entry we are recovering
1302 * @return The number of bytes the entry spans
1303 */
1304 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1305 int index = i;
1306 while (++index < virtualArray->field_size() &&
1307 virtualArray->field_at(index)->is_marker()) {}
1308 return index - i;
1309 }
1310
1311 /**
1312 * If there was a guarantee for byte array to always start aligned to a long, we could
1313 * do a simple check on the parity of the index. Unfortunately, that is not always the
1314 * case. Thus, we check alignment of the actual address we are writing to.
1315 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1316 * write a long to index 3.
1317 */
1318 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1319 jbyte* res = obj->byte_at_addr(index);
1320 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1321 return res;
1322 }
1323
1324 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1325 switch (byte_count) {
1326 case 1:
1327 obj->byte_at_put(index, (jbyte) value->get_jint());
1328 break;
1329 case 2:
1330 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1331 break;
1332 case 4:
1333 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1334 break;
1335 case 8:
1336 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1337 break;
1338 default:
1339 ShouldNotReachHere();
1340 }
1341 }
1342 #endif // INCLUDE_JVMCI
1343
1344
1345 // restore elements of an eliminated type array
1346 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1347 int index = 0;
1348
1349 for (int i = 0; i < sv->field_size(); i++) {
1350 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1351 switch(type) {
1352 case T_LONG: case T_DOUBLE: {
1353 assert(value->type() == T_INT, "Agreement.");
1354 StackValue* low =
1355 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1356 #ifdef _LP64
1357 jlong res = (jlong)low->get_intptr();
1358 #else
1359 jlong res = jlong_from(value->get_jint(), low->get_jint());
1360 #endif
1361 obj->long_at_put(index, res);
1362 break;
1363 }
1364
1365 case T_INT: case T_FLOAT: { // 4 bytes.
1366 assert(value->type() == T_INT, "Agreement.");
1367 bool big_value = false;
1368 if (i + 1 < sv->field_size() && type == T_INT) {
1369 if (sv->field_at(i)->is_location()) {
1370 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1371 if (type == Location::dbl || type == Location::lng) {
1372 big_value = true;
1373 }
1374 } else if (sv->field_at(i)->is_constant_int()) {
1375 ScopeValue* next_scope_field = sv->field_at(i + 1);
1376 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1377 big_value = true;
1378 }
1379 }
1380 }
1381
1382 if (big_value) {
1383 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1384 #ifdef _LP64
1385 jlong res = (jlong)low->get_intptr();
1386 #else
1387 jlong res = jlong_from(value->get_jint(), low->get_jint());
1388 #endif
1389 obj->int_at_put(index, *(jint*)&res);
1390 obj->int_at_put(++index, *((jint*)&res + 1));
1391 } else {
1392 obj->int_at_put(index, value->get_jint());
1393 }
1394 break;
1395 }
1396
1397 case T_SHORT:
1398 assert(value->type() == T_INT, "Agreement.");
1399 obj->short_at_put(index, (jshort)value->get_jint());
1400 break;
1401
1402 case T_CHAR:
1403 assert(value->type() == T_INT, "Agreement.");
1404 obj->char_at_put(index, (jchar)value->get_jint());
1405 break;
1406
1407 case T_BYTE: {
1408 assert(value->type() == T_INT, "Agreement.");
1409 #if INCLUDE_JVMCI
1410 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1411 int byte_count = count_number_of_bytes_for_entry(sv, i);
1412 byte_array_put(obj, value, index, byte_count);
1413 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1414 i += byte_count - 1; // Balance the loop counter.
1415 index += byte_count;
1416 // index has been updated so continue at top of loop
1417 continue;
1418 #else
1419 obj->byte_at_put(index, (jbyte)value->get_jint());
1420 break;
1421 #endif // INCLUDE_JVMCI
1422 }
1423
1424 case T_BOOLEAN: {
1425 assert(value->type() == T_INT, "Agreement.");
1426 obj->bool_at_put(index, (jboolean)value->get_jint());
1427 break;
1428 }
1429
1430 default:
1431 ShouldNotReachHere();
1432 }
1433 index++;
1434 }
1435 }
1436
1437 // restore fields of an eliminated object array
1438 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1439 for (int i = 0; i < sv->field_size(); i++) {
1440 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1441 assert(value->type() == T_OBJECT, "object element expected");
1442 obj->obj_at_put(i, value->get_obj()());
1443 }
1444 }
1445
1446 class ReassignedField {
1447 public:
1448 int _offset;
1449 BasicType _type;
1450 public:
1451 ReassignedField() {
1452 _offset = 0;
1453 _type = T_ILLEGAL;
1454 }
1455 };
1456
1457 static int compare(ReassignedField* left, ReassignedField* right) {
1458 return left->_offset - right->_offset;
1459 }
1460
1461 // Restore fields of an eliminated instance object using the same field order
1462 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1463 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1464 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1465 InstanceKlass* ik = klass;
1466 while (ik != nullptr) {
1467 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1468 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1469 ReassignedField field;
1470 field._offset = fs.offset();
1471 field._type = Signature::basic_type(fs.signature());
1472 fields->append(field);
1473 }
1474 }
1475 ik = ik->superklass();
1476 }
1477 fields->sort(compare);
1478 for (int i = 0; i < fields->length(); i++) {
1479 ScopeValue* scope_field = sv->field_at(svIndex);
1480 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1481 int offset = fields->at(i)._offset;
1482 BasicType type = fields->at(i)._type;
1483 switch (type) {
1484 case T_OBJECT: case T_ARRAY:
1485 assert(value->type() == T_OBJECT, "Agreement.");
1486 obj->obj_field_put(offset, value->get_obj()());
1487 break;
1488
1489 case T_INT: case T_FLOAT: { // 4 bytes.
1490 assert(value->type() == T_INT, "Agreement.");
1491 bool big_value = false;
1492 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1493 if (scope_field->is_location()) {
1494 Location::Type type = ((LocationValue*) scope_field)->location().type();
1495 if (type == Location::dbl || type == Location::lng) {
1496 big_value = true;
1497 }
1498 }
1499 if (scope_field->is_constant_int()) {
1500 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1501 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1502 big_value = true;
1503 }
1504 }
1505 }
1506
1507 if (big_value) {
1508 i++;
1509 assert(i < fields->length(), "second T_INT field needed");
1510 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1511 } else {
1512 obj->int_field_put(offset, value->get_jint());
1513 break;
1514 }
1515 }
1516 /* no break */
1517
1518 case T_LONG: case T_DOUBLE: {
1519 assert(value->type() == T_INT, "Agreement.");
1520 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1521 #ifdef _LP64
1522 jlong res = (jlong)low->get_intptr();
1523 #else
1524 jlong res = jlong_from(value->get_jint(), low->get_jint());
1525 #endif
1526 obj->long_field_put(offset, res);
1527 break;
1528 }
1529
1530 case T_SHORT:
1531 assert(value->type() == T_INT, "Agreement.");
1532 obj->short_field_put(offset, (jshort)value->get_jint());
1533 break;
1534
1535 case T_CHAR:
1536 assert(value->type() == T_INT, "Agreement.");
1537 obj->char_field_put(offset, (jchar)value->get_jint());
1538 break;
1539
1540 case T_BYTE:
1541 assert(value->type() == T_INT, "Agreement.");
1542 obj->byte_field_put(offset, (jbyte)value->get_jint());
1543 break;
1544
1545 case T_BOOLEAN:
1546 assert(value->type() == T_INT, "Agreement.");
1547 obj->bool_field_put(offset, (jboolean)value->get_jint());
1548 break;
1549
1550 default:
1551 ShouldNotReachHere();
1552 }
1553 svIndex++;
1554 }
1555 return svIndex;
1556 }
1557
1558 // restore fields of all eliminated objects and arrays
1559 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1560 for (int i = 0; i < objects->length(); i++) {
1561 assert(objects->at(i)->is_object(), "invalid debug information");
1562 ObjectValue* sv = (ObjectValue*) objects->at(i);
1563 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1564 Handle obj = sv->value();
1565 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1566 #ifndef PRODUCT
1567 if (PrintDeoptimizationDetails) {
1568 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1569 }
1570 #endif // !PRODUCT
1571
1572 if (obj.is_null()) {
1573 continue;
1574 }
1575
1576 #if INCLUDE_JVMCI
1577 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1578 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1579 continue;
1580 }
1581 #endif // INCLUDE_JVMCI
1582 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1583 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1584 ScopeValue* payload = sv->field_at(0);
1585 if (payload->is_location() &&
1586 payload->as_LocationValue()->location().type() == Location::vector) {
1587 #ifndef PRODUCT
1588 if (PrintDeoptimizationDetails) {
1589 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1590 if (Verbose) {
1591 Handle obj = sv->value();
1592 k->oop_print_on(obj(), tty);
1593 }
1594 }
1595 #endif // !PRODUCT
1596 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1597 }
1598 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1599 // which could be restored after vector object allocation.
1600 }
1601 if (k->is_instance_klass()) {
1602 InstanceKlass* ik = InstanceKlass::cast(k);
1603 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1604 } else if (k->is_typeArray_klass()) {
1605 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1606 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1607 } else if (k->is_objArray_klass()) {
1608 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1609 }
1610 }
1611 // These objects may escape when we return to Interpreter after deoptimization.
1612 // We need barrier so that stores that initialize these objects can't be reordered
1613 // with subsequent stores that make these objects accessible by other threads.
1614 OrderAccess::storestore();
1615 }
1616
1617
1618 // relock objects for which synchronization was eliminated
1619 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1620 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1621 bool relocked_objects = false;
1622 for (int i = 0; i < monitors->length(); i++) {
1623 MonitorInfo* mon_info = monitors->at(i);
1624 if (mon_info->eliminated()) {
1625 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1626 relocked_objects = true;
1627 if (!mon_info->owner_is_scalar_replaced()) {
1628 Handle obj(thread, mon_info->owner());
1629 markWord mark = obj->mark();
1630 if (exec_mode == Unpack_none) {
1631 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1632 // With exec_mode == Unpack_none obj may be thread local and locked in
1633 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1634 markWord dmw = mark.displaced_mark_helper();
1635 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1636 obj->set_mark(dmw);
1637 }
1638 if (mark.has_monitor()) {
1639 // defer relocking if the deoptee thread is currently waiting for obj
1640 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1641 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1642 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1643 if (LockingMode == LM_LEGACY) {
1644 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1645 } else if (UseObjectMonitorTable) {
1646 mon_info->lock()->clear_object_monitor_cache();
1647 }
1648 #ifdef ASSERT
1649 else {
1650 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1651 mon_info->lock()->set_bad_metadata_deopt();
1652 }
1653 #endif
1654 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1655 continue;
1656 }
1657 }
1658 }
1659 BasicLock* lock = mon_info->lock();
1660 if (LockingMode == LM_LIGHTWEIGHT) {
1661 // We have lost information about the correct state of the lock stack.
1662 // Entering may create an invalid lock stack. Inflate the lock if it
1663 // was fast_locked to restore the valid lock stack.
1664 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1665 if (deoptee_thread->lock_stack().contains(obj())) {
1666 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1667 deoptee_thread, thread);
1668 }
1669 assert(mon_info->owner()->is_locked(), "object must be locked now");
1670 assert(obj->mark().has_monitor(), "must be");
1671 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1672 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1673 } else {
1674 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1675 assert(mon_info->owner()->is_locked(), "object must be locked now");
1676 }
1677 }
1678 }
1679 }
1680 return relocked_objects;
1681 }
1682 #endif // COMPILER2_OR_JVMCI
1683
1684 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1685 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1686
1687 // Register map for next frame (used for stack crawl). We capture
1688 // the state of the deopt'ing frame's caller. Thus if we need to
1689 // stuff a C2I adapter we can properly fill in the callee-save
1690 // register locations.
1691 frame caller = fr.sender(reg_map);
1692 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1693
1694 frame sender = caller;
1695
1696 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1697 // the vframeArray containing the unpacking information is allocated in the C heap.
1698 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1699 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1700
1701 // Compare the vframeArray to the collected vframes
1702 assert(array->structural_compare(thread, chunk), "just checking");
1703
1704 if (TraceDeoptimization) {
1705 ResourceMark rm;
1706 stringStream st;
1707 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1708 st.print(" ");
1709 fr.print_on(&st);
1710 st.print_cr(" Virtual frames (innermost/newest first):");
1711 for (int index = 0; index < chunk->length(); index++) {
1712 compiledVFrame* vf = chunk->at(index);
1713 int bci = vf->raw_bci();
1714 const char* code_name;
1715 if (bci == SynchronizationEntryBCI) {
1716 code_name = "sync entry";
1717 } else {
1718 Bytecodes::Code code = vf->method()->code_at(bci);
1719 code_name = Bytecodes::name(code);
1720 }
1721
1722 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1723 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1724 st.print(" - %s", code_name);
1725 st.print_cr(" @ bci=%d ", bci);
1726 }
1727 tty->print_raw(st.freeze());
1728 tty->cr();
1729 }
1730
1731 return array;
1732 }
1733
1734 #if COMPILER2_OR_JVMCI
1735 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1736 // Reallocation of some scalar replaced objects failed. Record
1737 // that we need to pop all the interpreter frames for the
1738 // deoptimized compiled frame.
1739 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1740 thread->set_frames_to_pop_failed_realloc(array->frames());
1741 // Unlock all monitors here otherwise the interpreter will see a
1742 // mix of locked and unlocked monitors (because of failed
1743 // reallocations of synchronized objects) and be confused.
1744 for (int i = 0; i < array->frames(); i++) {
1745 MonitorChunk* monitors = array->element(i)->monitors();
1746 if (monitors != nullptr) {
1747 // Unlock in reverse order starting from most nested monitor.
1748 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1749 BasicObjectLock* src = monitors->at(j);
1750 if (src->obj() != nullptr) {
1751 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1752 }
1753 }
1754 array->element(i)->free_monitors();
1755 #ifdef ASSERT
1756 array->element(i)->set_removed_monitors();
1757 #endif
1758 }
1759 }
1760 }
1761 #endif
1762
1763 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1764 assert(fr.can_be_deoptimized(), "checking frame type");
1765
1766 nmethod* nm = fr.cb()->as_nmethod_or_null();
1767 assert(nm != nullptr, "only compiled methods can deopt");
1768 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none);
1769 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc());
1770 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry
1771 : cur_sd->method()->java_code_at(cur_sd->bci()));
1772 gather_statistics(nm, reason, action, bc);
1773
1774 if (LogCompilation && xtty != nullptr) {
1775 ttyLocker ttyl;
1776 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1777 nm->log_identity(xtty);
1778 xtty->end_head();
1779 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1780 xtty->begin_elem("jvms bci='%d'", sd->bci());
1781 xtty->method(sd->method());
1782 xtty->end_elem();
1783 if (sd->is_top()) break;
1784 }
1785 xtty->tail("deoptimized");
1786 }
1787
1788 Continuation::notify_deopt(thread, fr.sp());
1789
1790 // Patch the compiled method so that when execution returns to it we will
1791 // deopt the execution state and return to the interpreter.
1792 fr.deoptimize(thread);
1793 }
1794
1795 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1796 // Deoptimize only if the frame comes from compile code.
1797 // Do not deoptimize the frame which is already patched
1798 // during the execution of the loops below.
1799 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1800 return;
1801 }
1802 ResourceMark rm;
1803 deoptimize_single_frame(thread, fr, reason);
1804 }
1805
1806 #if INCLUDE_JVMCI
1807 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1808 // there is no exception handler for this pc => deoptimize
1809 nm->make_not_entrant();
1810
1811 // Use Deoptimization::deoptimize for all of its side-effects:
1812 // gathering traps statistics, logging...
1813 // it also patches the return pc but we do not care about that
1814 // since we return a continuation to the deopt_blob below.
1815 JavaThread* thread = JavaThread::current();
1816 RegisterMap reg_map(thread,
1817 RegisterMap::UpdateMap::skip,
1818 RegisterMap::ProcessFrames::include,
1819 RegisterMap::WalkContinuation::skip);
1820 frame runtime_frame = thread->last_frame();
1821 frame caller_frame = runtime_frame.sender(®_map);
1822 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1823 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1824 compiledVFrame* cvf = compiledVFrame::cast(vf);
1825 ScopeDesc* imm_scope = cvf->scope();
1826 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1827 if (imm_mdo != nullptr) {
1828 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1829 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1830
1831 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1832 if (pdata != nullptr && pdata->is_BitData()) {
1833 BitData* bit_data = (BitData*) pdata;
1834 bit_data->set_exception_seen();
1835 }
1836 }
1837
1838 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1839
1840 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1841 if (trap_mdo != nullptr) {
1842 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1843 }
1844
1845 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1846 }
1847 #endif
1848
1849 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1850 assert(thread == Thread::current() ||
1851 thread->is_handshake_safe_for(Thread::current()) ||
1852 SafepointSynchronize::is_at_safepoint(),
1853 "can only deoptimize other thread at a safepoint/handshake");
1854 // Compute frame and register map based on thread and sp.
1855 RegisterMap reg_map(thread,
1856 RegisterMap::UpdateMap::skip,
1857 RegisterMap::ProcessFrames::include,
1858 RegisterMap::WalkContinuation::skip);
1859 frame fr = thread->last_frame();
1860 while (fr.id() != id) {
1861 fr = fr.sender(®_map);
1862 }
1863 deoptimize(thread, fr, reason);
1864 }
1865
1866
1867 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1868 Thread* current = Thread::current();
1869 if (thread == current || thread->is_handshake_safe_for(current)) {
1870 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1871 } else {
1872 VM_DeoptimizeFrame deopt(thread, id, reason);
1873 VMThread::execute(&deopt);
1874 }
1875 }
1876
1877 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1878 deoptimize_frame(thread, id, Reason_constraint);
1879 }
1880
1881 // JVMTI PopFrame support
1882 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1883 {
1884 assert(thread == JavaThread::current(), "pre-condition");
1885 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1886 }
1887 JRT_END
1888
1889 MethodData*
1890 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1891 bool create_if_missing) {
1892 JavaThread* THREAD = thread; // For exception macros.
1893 MethodData* mdo = m()->method_data();
1894 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1895 // Build an MDO. Ignore errors like OutOfMemory;
1896 // that simply means we won't have an MDO to update.
1897 Method::build_profiling_method_data(m, THREAD);
1898 if (HAS_PENDING_EXCEPTION) {
1899 // Only metaspace OOM is expected. No Java code executed.
1900 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1901 CLEAR_PENDING_EXCEPTION;
1902 }
1903 mdo = m()->method_data();
1904 }
1905 return mdo;
1906 }
1907
1908 #if COMPILER2_OR_JVMCI
1909 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1910 // In case of an unresolved klass entry, load the class.
1911 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1912 // and probably nowhere else.
1913 // Even that case would benefit from simply re-interpreting the
1914 // bytecode, without paying special attention to the class index.
1915 // So this whole "class index" feature should probably be removed.
1916
1917 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1918 Klass* tk = constant_pool->klass_at(index, THREAD);
1919 if (HAS_PENDING_EXCEPTION) {
1920 // Exception happened during classloading. We ignore the exception here, since it
1921 // is going to be rethrown since the current activation is going to be deoptimized and
1922 // the interpreter will re-execute the bytecode.
1923 // Do not clear probable Async Exceptions.
1924 CLEAR_PENDING_NONASYNC_EXCEPTION;
1925 // Class loading called java code which may have caused a stack
1926 // overflow. If the exception was thrown right before the return
1927 // to the runtime the stack is no longer guarded. Reguard the
1928 // stack otherwise if we return to the uncommon trap blob and the
1929 // stack bang causes a stack overflow we crash.
1930 JavaThread* jt = THREAD;
1931 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1932 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1933 }
1934 return;
1935 }
1936
1937 assert(!constant_pool->tag_at(index).is_symbol(),
1938 "no symbolic names here, please");
1939 }
1940
1941 #if INCLUDE_JFR
1942
1943 class DeoptReasonSerializer : public JfrSerializer {
1944 public:
1945 void serialize(JfrCheckpointWriter& writer) {
1946 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1947 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1948 writer.write_key((u8)i);
1949 writer.write(Deoptimization::trap_reason_name(i));
1950 }
1951 }
1952 };
1953
1954 class DeoptActionSerializer : public JfrSerializer {
1955 public:
1956 void serialize(JfrCheckpointWriter& writer) {
1957 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1958 writer.write_count(nof_actions);
1959 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1960 writer.write_key(i);
1961 writer.write(Deoptimization::trap_action_name((int)i));
1962 }
1963 }
1964 };
1965
1966 static void register_serializers() {
1967 static int critical_section = 0;
1968 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1969 return;
1970 }
1971 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1972 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1973 }
1974
1975 static void post_deoptimization_event(nmethod* nm,
1976 const Method* method,
1977 int trap_bci,
1978 int instruction,
1979 Deoptimization::DeoptReason reason,
1980 Deoptimization::DeoptAction action) {
1981 assert(nm != nullptr, "invariant");
1982 assert(method != nullptr, "invariant");
1983 if (EventDeoptimization::is_enabled()) {
1984 static bool serializers_registered = false;
1985 if (!serializers_registered) {
1986 register_serializers();
1987 serializers_registered = true;
1988 }
1989 EventDeoptimization event;
1990 event.set_compileId(nm->compile_id());
1991 event.set_compiler(nm->compiler_type());
1992 event.set_method(method);
1993 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1994 event.set_bci(trap_bci);
1995 event.set_instruction(instruction);
1996 event.set_reason(reason);
1997 event.set_action(action);
1998 event.commit();
1999 }
2000 }
2001
2002 #endif // INCLUDE_JFR
2003
2004 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2005 const char* reason_name, const char* reason_action) {
2006 LogTarget(Debug, deoptimization) lt;
2007 if (lt.is_enabled()) {
2008 LogStream ls(lt);
2009 bool is_osr = nm->is_osr_method();
2010 ls.print("cid=%4d %s level=%d",
2011 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2012 ls.print(" %s", tm->name_and_sig_as_C_string());
2013 ls.print(" trap_bci=%d ", trap_bci);
2014 if (is_osr) {
2015 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2016 }
2017 ls.print("%s ", reason_name);
2018 ls.print("%s ", reason_action);
2019 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2020 pc, fr.pc() - nm->code_begin());
2021 }
2022 }
2023
2024 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2025 HandleMark hm(current);
2026
2027 // uncommon_trap() is called at the beginning of the uncommon trap
2028 // handler. Note this fact before we start generating temporary frames
2029 // that can confuse an asynchronous stack walker. This counter is
2030 // decremented at the end of unpack_frames().
2031
2032 current->inc_in_deopt_handler();
2033
2034 #if INCLUDE_JVMCI
2035 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2036 RegisterMap reg_map(current,
2037 RegisterMap::UpdateMap::include,
2038 RegisterMap::ProcessFrames::include,
2039 RegisterMap::WalkContinuation::skip);
2040 #else
2041 RegisterMap reg_map(current,
2042 RegisterMap::UpdateMap::skip,
2043 RegisterMap::ProcessFrames::include,
2044 RegisterMap::WalkContinuation::skip);
2045 #endif
2046 frame stub_frame = current->last_frame();
2047 frame fr = stub_frame.sender(®_map);
2048
2049 // Log a message
2050 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2051 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2052
2053 {
2054 ResourceMark rm;
2055
2056 DeoptReason reason = trap_request_reason(trap_request);
2057 DeoptAction action = trap_request_action(trap_request);
2058 #if INCLUDE_JVMCI
2059 int debug_id = trap_request_debug_id(trap_request);
2060 #endif
2061 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2062
2063 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2064 compiledVFrame* cvf = compiledVFrame::cast(vf);
2065
2066 nmethod* nm = cvf->code();
2067
2068 ScopeDesc* trap_scope = cvf->scope();
2069
2070 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2071
2072 if (is_receiver_constraint_failure) {
2073 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2074 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2075 JVMCI_ONLY(COMMA debug_id));
2076 }
2077
2078 methodHandle trap_method(current, trap_scope->method());
2079 int trap_bci = trap_scope->bci();
2080 #if INCLUDE_JVMCI
2081 jlong speculation = current->pending_failed_speculation();
2082 if (nm->is_compiled_by_jvmci()) {
2083 nm->update_speculation(current);
2084 } else {
2085 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2086 }
2087
2088 if (trap_bci == SynchronizationEntryBCI) {
2089 trap_bci = 0;
2090 current->set_pending_monitorenter(true);
2091 }
2092
2093 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2094 current->set_pending_transfer_to_interpreter(true);
2095 }
2096 #endif
2097
2098 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2099 // Record this event in the histogram.
2100 gather_statistics(nm, reason, action, trap_bc);
2101
2102 // Ensure that we can record deopt. history:
2103 bool create_if_missing = ProfileTraps;
2104
2105 methodHandle profiled_method;
2106 #if INCLUDE_JVMCI
2107 if (nm->is_compiled_by_jvmci()) {
2108 profiled_method = methodHandle(current, nm->method());
2109 } else {
2110 profiled_method = trap_method;
2111 }
2112 #else
2113 profiled_method = trap_method;
2114 #endif
2115
2116 MethodData* trap_mdo =
2117 get_method_data(current, profiled_method, create_if_missing);
2118
2119 { // Log Deoptimization event for JFR, UL and event system
2120 Method* tm = trap_method();
2121 const char* reason_name = trap_reason_name(reason);
2122 const char* reason_action = trap_action_name(action);
2123 intptr_t pc = p2i(fr.pc());
2124
2125 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2126 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2127 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2128 reason_name, reason_action, pc,
2129 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2130 }
2131
2132 // Print a bunch of diagnostics, if requested.
2133 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2134 ResourceMark rm;
2135
2136 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2137 // We must do this already now, since we cannot acquire this lock while
2138 // holding the tty lock (lock ordering by rank).
2139 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2140
2141 ttyLocker ttyl;
2142
2143 char buf[100];
2144 if (xtty != nullptr) {
2145 xtty->begin_head("uncommon_trap thread='%zu' %s",
2146 os::current_thread_id(),
2147 format_trap_request(buf, sizeof(buf), trap_request));
2148 #if INCLUDE_JVMCI
2149 if (speculation != 0) {
2150 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2151 }
2152 #endif
2153 nm->log_identity(xtty);
2154 }
2155 Symbol* class_name = nullptr;
2156 bool unresolved = false;
2157 if (unloaded_class_index >= 0) {
2158 constantPoolHandle constants (current, trap_method->constants());
2159 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2160 class_name = constants->klass_name_at(unloaded_class_index);
2161 unresolved = true;
2162 if (xtty != nullptr)
2163 xtty->print(" unresolved='1'");
2164 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2165 class_name = constants->symbol_at(unloaded_class_index);
2166 }
2167 if (xtty != nullptr)
2168 xtty->name(class_name);
2169 }
2170 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2171 // Dump the relevant MDO state.
2172 // This is the deopt count for the current reason, any previous
2173 // reasons or recompiles seen at this point.
2174 int dcnt = trap_mdo->trap_count(reason);
2175 if (dcnt != 0)
2176 xtty->print(" count='%d'", dcnt);
2177
2178 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2179 // lock extra_data_lock before the tty lock.
2180 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2181 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2182 if (dos != 0) {
2183 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2184 if (trap_state_is_recompiled(dos)) {
2185 int recnt2 = trap_mdo->overflow_recompile_count();
2186 if (recnt2 != 0)
2187 xtty->print(" recompiles2='%d'", recnt2);
2188 }
2189 }
2190 }
2191 if (xtty != nullptr) {
2192 xtty->stamp();
2193 xtty->end_head();
2194 }
2195 if (TraceDeoptimization) { // make noise on the tty
2196 stringStream st;
2197 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2198 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2199 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2200 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2201 #if INCLUDE_JVMCI
2202 if (nm->is_compiled_by_jvmci()) {
2203 const char* installed_code_name = nm->jvmci_name();
2204 if (installed_code_name != nullptr) {
2205 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2206 }
2207 }
2208 #endif
2209 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2210 p2i(fr.pc()),
2211 os::current_thread_id(),
2212 trap_reason_name(reason),
2213 trap_action_name(action),
2214 unloaded_class_index
2215 #if INCLUDE_JVMCI
2216 , debug_id
2217 #endif
2218 );
2219 if (class_name != nullptr) {
2220 st.print(unresolved ? " unresolved class: " : " symbol: ");
2221 class_name->print_symbol_on(&st);
2222 }
2223 st.cr();
2224 tty->print_raw(st.freeze());
2225 }
2226 if (xtty != nullptr) {
2227 // Log the precise location of the trap.
2228 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2229 xtty->begin_elem("jvms bci='%d'", sd->bci());
2230 xtty->method(sd->method());
2231 xtty->end_elem();
2232 if (sd->is_top()) break;
2233 }
2234 xtty->tail("uncommon_trap");
2235 }
2236 }
2237 // (End diagnostic printout.)
2238
2239 if (is_receiver_constraint_failure) {
2240 fatal("missing receiver type check");
2241 }
2242
2243 // Load class if necessary
2244 if (unloaded_class_index >= 0) {
2245 constantPoolHandle constants(current, trap_method->constants());
2246 load_class_by_index(constants, unloaded_class_index, THREAD);
2247 }
2248
2249 // Flush the nmethod if necessary and desirable.
2250 //
2251 // We need to avoid situations where we are re-flushing the nmethod
2252 // because of a hot deoptimization site. Repeated flushes at the same
2253 // point need to be detected by the compiler and avoided. If the compiler
2254 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2255 // module must take measures to avoid an infinite cycle of recompilation
2256 // and deoptimization. There are several such measures:
2257 //
2258 // 1. If a recompilation is ordered a second time at some site X
2259 // and for the same reason R, the action is adjusted to 'reinterpret',
2260 // to give the interpreter time to exercise the method more thoroughly.
2261 // If this happens, the method's overflow_recompile_count is incremented.
2262 //
2263 // 2. If the compiler fails to reduce the deoptimization rate, then
2264 // the method's overflow_recompile_count will begin to exceed the set
2265 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2266 // is adjusted to 'make_not_compilable', and the method is abandoned
2267 // to the interpreter. This is a performance hit for hot methods,
2268 // but is better than a disastrous infinite cycle of recompilations.
2269 // (Actually, only the method containing the site X is abandoned.)
2270 //
2271 // 3. In parallel with the previous measures, if the total number of
2272 // recompilations of a method exceeds the much larger set limit
2273 // PerMethodRecompilationCutoff, the method is abandoned.
2274 // This should only happen if the method is very large and has
2275 // many "lukewarm" deoptimizations. The code which enforces this
2276 // limit is elsewhere (class nmethod, class Method).
2277 //
2278 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2279 // to recompile at each bytecode independently of the per-BCI cutoff.
2280 //
2281 // The decision to update code is up to the compiler, and is encoded
2282 // in the Action_xxx code. If the compiler requests Action_none
2283 // no trap state is changed, no compiled code is changed, and the
2284 // computation suffers along in the interpreter.
2285 //
2286 // The other action codes specify various tactics for decompilation
2287 // and recompilation. Action_maybe_recompile is the loosest, and
2288 // allows the compiled code to stay around until enough traps are seen,
2289 // and until the compiler gets around to recompiling the trapping method.
2290 //
2291 // The other actions cause immediate removal of the present code.
2292
2293 // Traps caused by injected profile shouldn't pollute trap counts.
2294 bool injected_profile_trap = trap_method->has_injected_profile() &&
2295 (reason == Reason_intrinsic || reason == Reason_unreached);
2296
2297 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2298 bool make_not_entrant = false;
2299 bool make_not_compilable = false;
2300 bool reprofile = false;
2301 switch (action) {
2302 case Action_none:
2303 // Keep the old code.
2304 update_trap_state = false;
2305 break;
2306 case Action_maybe_recompile:
2307 // Do not need to invalidate the present code, but we can
2308 // initiate another
2309 // Start compiler without (necessarily) invalidating the nmethod.
2310 // The system will tolerate the old code, but new code should be
2311 // generated when possible.
2312 break;
2313 case Action_reinterpret:
2314 // Go back into the interpreter for a while, and then consider
2315 // recompiling form scratch.
2316 make_not_entrant = true;
2317 // Reset invocation counter for outer most method.
2318 // This will allow the interpreter to exercise the bytecodes
2319 // for a while before recompiling.
2320 // By contrast, Action_make_not_entrant is immediate.
2321 //
2322 // Note that the compiler will track null_check, null_assert,
2323 // range_check, and class_check events and log them as if they
2324 // had been traps taken from compiled code. This will update
2325 // the MDO trap history so that the next compilation will
2326 // properly detect hot trap sites.
2327 reprofile = true;
2328 break;
2329 case Action_make_not_entrant:
2330 // Request immediate recompilation, and get rid of the old code.
2331 // Make them not entrant, so next time they are called they get
2332 // recompiled. Unloaded classes are loaded now so recompile before next
2333 // time they are called. Same for uninitialized. The interpreter will
2334 // link the missing class, if any.
2335 make_not_entrant = true;
2336 break;
2337 case Action_make_not_compilable:
2338 // Give up on compiling this method at all.
2339 make_not_entrant = true;
2340 make_not_compilable = true;
2341 break;
2342 default:
2343 ShouldNotReachHere();
2344 }
2345
2346 // Setting +ProfileTraps fixes the following, on all platforms:
2347 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2348 // recompile relies on a MethodData* to record heroic opt failures.
2349
2350 // Whether the interpreter is producing MDO data or not, we also need
2351 // to use the MDO to detect hot deoptimization points and control
2352 // aggressive optimization.
2353 bool inc_recompile_count = false;
2354
2355 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2356 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2357 (trap_mdo != nullptr),
2358 Mutex::_no_safepoint_check_flag);
2359 ProfileData* pdata = nullptr;
2360 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2361 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2362 uint this_trap_count = 0;
2363 bool maybe_prior_trap = false;
2364 bool maybe_prior_recompile = false;
2365
2366 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2367 #if INCLUDE_JVMCI
2368 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2369 #endif
2370 nm->method(),
2371 //outputs:
2372 this_trap_count,
2373 maybe_prior_trap,
2374 maybe_prior_recompile);
2375 // Because the interpreter also counts null, div0, range, and class
2376 // checks, these traps from compiled code are double-counted.
2377 // This is harmless; it just means that the PerXTrapLimit values
2378 // are in effect a little smaller than they look.
2379
2380 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2381 if (per_bc_reason != Reason_none) {
2382 // Now take action based on the partially known per-BCI history.
2383 if (maybe_prior_trap
2384 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2385 // If there are too many traps at this BCI, force a recompile.
2386 // This will allow the compiler to see the limit overflow, and
2387 // take corrective action, if possible. The compiler generally
2388 // does not use the exact PerBytecodeTrapLimit value, but instead
2389 // changes its tactics if it sees any traps at all. This provides
2390 // a little hysteresis, delaying a recompile until a trap happens
2391 // several times.
2392 //
2393 // Actually, since there is only one bit of counter per BCI,
2394 // the possible per-BCI counts are {0,1,(per-method count)}.
2395 // This produces accurate results if in fact there is only
2396 // one hot trap site, but begins to get fuzzy if there are
2397 // many sites. For example, if there are ten sites each
2398 // trapping two or more times, they each get the blame for
2399 // all of their traps.
2400 make_not_entrant = true;
2401 }
2402
2403 // Detect repeated recompilation at the same BCI, and enforce a limit.
2404 if (make_not_entrant && maybe_prior_recompile) {
2405 // More than one recompile at this point.
2406 inc_recompile_count = maybe_prior_trap;
2407 }
2408 } else {
2409 // For reasons which are not recorded per-bytecode, we simply
2410 // force recompiles unconditionally.
2411 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2412 make_not_entrant = true;
2413 }
2414
2415 // Go back to the compiler if there are too many traps in this method.
2416 if (this_trap_count >= per_method_trap_limit(reason)) {
2417 // If there are too many traps in this method, force a recompile.
2418 // This will allow the compiler to see the limit overflow, and
2419 // take corrective action, if possible.
2420 // (This condition is an unlikely backstop only, because the
2421 // PerBytecodeTrapLimit is more likely to take effect first,
2422 // if it is applicable.)
2423 make_not_entrant = true;
2424 }
2425
2426 // Here's more hysteresis: If there has been a recompile at
2427 // this trap point already, run the method in the interpreter
2428 // for a while to exercise it more thoroughly.
2429 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2430 reprofile = true;
2431 }
2432 }
2433
2434 // Take requested actions on the method:
2435
2436 // Recompile
2437 if (make_not_entrant) {
2438 if (!nm->make_not_entrant()) {
2439 return; // the call did not change nmethod's state
2440 }
2441
2442 if (pdata != nullptr) {
2443 // Record the recompilation event, if any.
2444 int tstate0 = pdata->trap_state();
2445 int tstate1 = trap_state_set_recompiled(tstate0, true);
2446 if (tstate1 != tstate0)
2447 pdata->set_trap_state(tstate1);
2448 }
2449
2450 // For code aging we count traps separately here, using make_not_entrant()
2451 // as a guard against simultaneous deopts in multiple threads.
2452 if (reason == Reason_tenured && trap_mdo != nullptr) {
2453 trap_mdo->inc_tenure_traps();
2454 }
2455 }
2456
2457 if (inc_recompile_count) {
2458 trap_mdo->inc_overflow_recompile_count();
2459 if ((uint)trap_mdo->overflow_recompile_count() >
2460 (uint)PerBytecodeRecompilationCutoff) {
2461 // Give up on the method containing the bad BCI.
2462 if (trap_method() == nm->method()) {
2463 make_not_compilable = true;
2464 } else {
2465 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2466 // But give grace to the enclosing nm->method().
2467 }
2468 }
2469 }
2470
2471 // Reprofile
2472 if (reprofile) {
2473 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2474 }
2475
2476 // Give up compiling
2477 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2478 assert(make_not_entrant, "consistent");
2479 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2480 }
2481
2482 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2483 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2484 if (exception_handler_data != nullptr) {
2485 // uncommon trap at the start of an exception handler.
2486 // C2 generates these for un-entered exception handlers.
2487 // mark the handler as entered to avoid generating
2488 // another uncommon trap the next time the handler is compiled
2489 exception_handler_data->set_exception_handler_entered();
2490 }
2491 }
2492
2493 } // Free marked resources
2494
2495 }
2496 JRT_END
2497
2498 ProfileData*
2499 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2500 int trap_bci,
2501 Deoptimization::DeoptReason reason,
2502 bool update_total_trap_count,
2503 #if INCLUDE_JVMCI
2504 bool is_osr,
2505 #endif
2506 Method* compiled_method,
2507 //outputs:
2508 uint& ret_this_trap_count,
2509 bool& ret_maybe_prior_trap,
2510 bool& ret_maybe_prior_recompile) {
2511 trap_mdo->check_extra_data_locked();
2512
2513 bool maybe_prior_trap = false;
2514 bool maybe_prior_recompile = false;
2515 uint this_trap_count = 0;
2516 if (update_total_trap_count) {
2517 uint idx = reason;
2518 #if INCLUDE_JVMCI
2519 if (is_osr) {
2520 // Upper half of history array used for traps in OSR compilations
2521 idx += Reason_TRAP_HISTORY_LENGTH;
2522 }
2523 #endif
2524 uint prior_trap_count = trap_mdo->trap_count(idx);
2525 this_trap_count = trap_mdo->inc_trap_count(idx);
2526
2527 // If the runtime cannot find a place to store trap history,
2528 // it is estimated based on the general condition of the method.
2529 // If the method has ever been recompiled, or has ever incurred
2530 // a trap with the present reason , then this BCI is assumed
2531 // (pessimistically) to be the culprit.
2532 maybe_prior_trap = (prior_trap_count != 0);
2533 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2534 }
2535 ProfileData* pdata = nullptr;
2536
2537
2538 // For reasons which are recorded per bytecode, we check per-BCI data.
2539 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2540 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2541 if (per_bc_reason != Reason_none) {
2542 // Find the profile data for this BCI. If there isn't one,
2543 // try to allocate one from the MDO's set of spares.
2544 // This will let us detect a repeated trap at this point.
2545 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2546
2547 if (pdata != nullptr) {
2548 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2549 if (LogCompilation && xtty != nullptr) {
2550 ttyLocker ttyl;
2551 // no more room for speculative traps in this MDO
2552 xtty->elem("speculative_traps_oom");
2553 }
2554 }
2555 // Query the trap state of this profile datum.
2556 int tstate0 = pdata->trap_state();
2557 if (!trap_state_has_reason(tstate0, per_bc_reason))
2558 maybe_prior_trap = false;
2559 if (!trap_state_is_recompiled(tstate0))
2560 maybe_prior_recompile = false;
2561
2562 // Update the trap state of this profile datum.
2563 int tstate1 = tstate0;
2564 // Record the reason.
2565 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2566 // Store the updated state on the MDO, for next time.
2567 if (tstate1 != tstate0)
2568 pdata->set_trap_state(tstate1);
2569 } else {
2570 if (LogCompilation && xtty != nullptr) {
2571 ttyLocker ttyl;
2572 // Missing MDP? Leave a small complaint in the log.
2573 xtty->elem("missing_mdp bci='%d'", trap_bci);
2574 }
2575 }
2576 }
2577
2578 // Return results:
2579 ret_this_trap_count = this_trap_count;
2580 ret_maybe_prior_trap = maybe_prior_trap;
2581 ret_maybe_prior_recompile = maybe_prior_recompile;
2582 return pdata;
2583 }
2584
2585 void
2586 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2587 ResourceMark rm;
2588 // Ignored outputs:
2589 uint ignore_this_trap_count;
2590 bool ignore_maybe_prior_trap;
2591 bool ignore_maybe_prior_recompile;
2592 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2593 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2594 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2595
2596 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2597 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2598
2599 query_update_method_data(trap_mdo, trap_bci,
2600 (DeoptReason)reason,
2601 update_total_counts,
2602 #if INCLUDE_JVMCI
2603 false,
2604 #endif
2605 nullptr,
2606 ignore_this_trap_count,
2607 ignore_maybe_prior_trap,
2608 ignore_maybe_prior_recompile);
2609 }
2610
2611 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode))
2612 // Enable WXWrite: current function is called from methods compiled by C2 directly
2613 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2614
2615 // Still in Java no safepoints
2616 {
2617 // This enters VM and may safepoint
2618 uncommon_trap_inner(current, trap_request);
2619 }
2620 HandleMark hm(current);
2621 return fetch_unroll_info_helper(current, exec_mode);
2622 PROF_END
2623
2624 // Local derived constants.
2625 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2626 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2627 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2628
2629 //---------------------------trap_state_reason---------------------------------
2630 Deoptimization::DeoptReason
2631 Deoptimization::trap_state_reason(int trap_state) {
2632 // This assert provides the link between the width of DataLayout::trap_bits
2633 // and the encoding of "recorded" reasons. It ensures there are enough
2634 // bits to store all needed reasons in the per-BCI MDO profile.
2635 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2636 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2637 trap_state -= recompile_bit;
2638 if (trap_state == DS_REASON_MASK) {
2639 return Reason_many;
2640 } else {
2641 assert((int)Reason_none == 0, "state=0 => Reason_none");
2642 return (DeoptReason)trap_state;
2643 }
2644 }
2645 //-------------------------trap_state_has_reason-------------------------------
2646 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2647 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2648 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2649 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2650 trap_state -= recompile_bit;
2651 if (trap_state == DS_REASON_MASK) {
2652 return -1; // true, unspecifically (bottom of state lattice)
2653 } else if (trap_state == reason) {
2654 return 1; // true, definitely
2655 } else if (trap_state == 0) {
2656 return 0; // false, definitely (top of state lattice)
2657 } else {
2658 return 0; // false, definitely
2659 }
2660 }
2661 //-------------------------trap_state_add_reason-------------------------------
2662 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2663 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2664 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2665 trap_state -= recompile_bit;
2666 if (trap_state == DS_REASON_MASK) {
2667 return trap_state + recompile_bit; // already at state lattice bottom
2668 } else if (trap_state == reason) {
2669 return trap_state + recompile_bit; // the condition is already true
2670 } else if (trap_state == 0) {
2671 return reason + recompile_bit; // no condition has yet been true
2672 } else {
2673 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2674 }
2675 }
2676 //-----------------------trap_state_is_recompiled------------------------------
2677 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2678 return (trap_state & DS_RECOMPILE_BIT) != 0;
2679 }
2680 //-----------------------trap_state_set_recompiled-----------------------------
2681 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2682 if (z) return trap_state | DS_RECOMPILE_BIT;
2683 else return trap_state & ~DS_RECOMPILE_BIT;
2684 }
2685 //---------------------------format_trap_state---------------------------------
2686 // This is used for debugging and diagnostics, including LogFile output.
2687 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2688 int trap_state) {
2689 assert(buflen > 0, "sanity");
2690 DeoptReason reason = trap_state_reason(trap_state);
2691 bool recomp_flag = trap_state_is_recompiled(trap_state);
2692 // Re-encode the state from its decoded components.
2693 int decoded_state = 0;
2694 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2695 decoded_state = trap_state_add_reason(decoded_state, reason);
2696 if (recomp_flag)
2697 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2698 // If the state re-encodes properly, format it symbolically.
2699 // Because this routine is used for debugging and diagnostics,
2700 // be robust even if the state is a strange value.
2701 size_t len;
2702 if (decoded_state != trap_state) {
2703 // Random buggy state that doesn't decode??
2704 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2705 } else {
2706 len = jio_snprintf(buf, buflen, "%s%s",
2707 trap_reason_name(reason),
2708 recomp_flag ? " recompiled" : "");
2709 }
2710 return buf;
2711 }
2712
2713
2714 //--------------------------------statics--------------------------------------
2715 const char* Deoptimization::_trap_reason_name[] = {
2716 // Note: Keep this in sync. with enum DeoptReason.
2717 "none",
2718 "null_check",
2719 "null_assert" JVMCI_ONLY("_or_unreached0"),
2720 "range_check",
2721 "class_check",
2722 "array_check",
2723 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2724 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2725 "profile_predicate",
2726 "unloaded",
2727 "uninitialized",
2728 "initialized",
2729 "unreached",
2730 "unhandled",
2731 "constraint",
2732 "div0_check",
2733 "age",
2734 "predicate",
2735 "loop_limit_check",
2736 "speculate_class_check",
2737 "speculate_null_check",
2738 "speculate_null_assert",
2739 "unstable_if",
2740 "unstable_fused_if",
2741 "receiver_constraint",
2742 #if INCLUDE_JVMCI
2743 "aliasing",
2744 "transfer_to_interpreter",
2745 "not_compiled_exception_handler",
2746 "unresolved",
2747 "jsr_mismatch",
2748 #endif
2749 "tenured"
2750 };
2751 const char* Deoptimization::_trap_action_name[] = {
2752 // Note: Keep this in sync. with enum DeoptAction.
2753 "none",
2754 "maybe_recompile",
2755 "reinterpret",
2756 "make_not_entrant",
2757 "make_not_compilable"
2758 };
2759
2760 const char* Deoptimization::trap_reason_name(int reason) {
2761 // Check that every reason has a name
2762 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2763
2764 if (reason == Reason_many) return "many";
2765 if ((uint)reason < Reason_LIMIT)
2766 return _trap_reason_name[reason];
2767 static char buf[20];
2768 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2769 return buf;
2770 }
2771 const char* Deoptimization::trap_action_name(int action) {
2772 // Check that every action has a name
2773 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2774
2775 if ((uint)action < Action_LIMIT)
2776 return _trap_action_name[action];
2777 static char buf[20];
2778 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2779 return buf;
2780 }
2781
2782 // This is used for debugging and diagnostics, including LogFile output.
2783 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2784 int trap_request) {
2785 jint unloaded_class_index = trap_request_index(trap_request);
2786 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2787 const char* action = trap_action_name(trap_request_action(trap_request));
2788 #if INCLUDE_JVMCI
2789 int debug_id = trap_request_debug_id(trap_request);
2790 #endif
2791 size_t len;
2792 if (unloaded_class_index < 0) {
2793 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2794 reason, action
2795 #if INCLUDE_JVMCI
2796 ,debug_id
2797 #endif
2798 );
2799 } else {
2800 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2801 reason, action, unloaded_class_index
2802 #if INCLUDE_JVMCI
2803 ,debug_id
2804 #endif
2805 );
2806 }
2807 return buf;
2808 }
2809
2810 juint Deoptimization::_deoptimization_hist
2811 [1 + 4 + 5] // total + online + archived
2812 [Deoptimization::Reason_LIMIT]
2813 [1 + Deoptimization::Action_LIMIT]
2814 [Deoptimization::BC_CASE_LIMIT]
2815 = {0};
2816
2817 enum {
2818 LSB_BITS = 8,
2819 LSB_MASK = right_n_bits(LSB_BITS)
2820 };
2821
2822 static void update(juint* cases, Bytecodes::Code bc) {
2823 juint* bc_counter_addr = nullptr;
2824 juint bc_counter = 0;
2825 // Look for an unused counter, or an exact match to this BC.
2826 if (bc != Bytecodes::_illegal) {
2827 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) {
2828 juint* counter_addr = &cases[bc_case];
2829 juint counter = *counter_addr;
2830 if ((counter == 0 && bc_counter_addr == nullptr)
2831 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2832 // this counter is either free or is already devoted to this BC
2833 bc_counter_addr = counter_addr;
2834 bc_counter = counter | bc;
2835 }
2836 }
2837 }
2838 if (bc_counter_addr == nullptr) {
2839 // Overflow, or no given bytecode.
2840 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1];
2841 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2842 }
2843 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2844 }
2845
2846
2847 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
2848 Bytecodes::Code bc) {
2849 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2850 assert(action >= 0 && action < Action_LIMIT, "oob");
2851 _deoptimization_hist[0][Reason_none][0][0] += 1; // total
2852 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total
2853
2854 update(_deoptimization_hist[0][reason][1+action], bc);
2855
2856 uint lvl = nm->comp_level() + (nm->is_scc() ? 4 : 0) + (nm->preloaded() ? 1 : 0);
2857 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total
2858 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total
2859 update(_deoptimization_hist[lvl][reason][1+action], bc);
2860 }
2861
2862 jint Deoptimization::total_deoptimization_count() {
2863 return _deoptimization_hist[0][Reason_none][0][0];
2864 }
2865
2866 // Get the deopt count for a specific reason and a specific action. If either
2867 // one of 'reason' or 'action' is null, the method returns the sum of all
2868 // deoptimizations with the specific 'action' or 'reason' respectively.
2869 // If both arguments are null, the method returns the total deopt count.
2870 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2871 if (reason_str == nullptr && action_str == nullptr) {
2872 return total_deoptimization_count();
2873 }
2874 juint counter = 0;
2875 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2876 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2877 for (int action = 0; action < Action_LIMIT; action++) {
2878 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2879 juint* cases = _deoptimization_hist[0][reason][1+action];
2880 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2881 counter += cases[bc_case] >> LSB_BITS;
2882 }
2883 }
2884 }
2885 }
2886 }
2887 return counter;
2888 }
2889
2890 void Deoptimization::print_statistics() {
2891 ttyLocker ttyl;
2892 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2893 tty->print_cr("Deoptimization traps recorded:");
2894 print_statistics_on(tty);
2895 if (xtty != nullptr) xtty->tail("statistics");
2896 }
2897
2898 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) {
2899 juint total = _deoptimization_hist[lvl][Reason_none][0][0];
2900 juint account = total;
2901 #define PRINT_STAT_LINE(name, r) \
2902 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name);
2903 if (total > 0) {
2904 st->print(" %s: ", title);
2905 PRINT_STAT_LINE("total", total);
2906 // For each non-zero entry in the histogram, print the reason,
2907 // the action, and (if specifically known) the type of bytecode.
2908 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2909 for (int action = 0; action < Action_LIMIT; action++) {
2910 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action];
2911 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2912 juint counter = cases[bc_case];
2913 if (counter != 0) {
2914 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2915 const char* bc_name = "other";
2916 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) {
2917 // overwritten
2918 } else if (Bytecodes::is_defined(bc)) {
2919 bc_name = Bytecodes::name(bc);
2920 }
2921 juint r = counter >> LSB_BITS;
2922 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)",
2923 trap_reason_name(reason), trap_action_name(action), bc_name,
2924 r, (r * 100.0) / total);
2925 account -= r;
2926 }
2927 }
2928 }
2929 }
2930 if (account != 0) {
2931 PRINT_STAT_LINE("unaccounted", account);
2932 }
2933 #undef PRINT_STAT_LINE
2934 }
2935 }
2936
2937 void Deoptimization::print_statistics_on(outputStream* st) {
2938 // print_statistics_on("Total", 0, st);
2939 print_statistics_on("Tier1", 1, st);
2940 print_statistics_on("Tier2", 2, st);
2941 print_statistics_on("Tier3", 3, st);
2942 print_statistics_on("Tier4", 4, st);
2943
2944 print_statistics_on("SC Tier1", 5, st);
2945 print_statistics_on("SC Tier2", 6, st);
2946 print_statistics_on("SC Tier4", 8, st);
2947 print_statistics_on("SC Tier5 (preloaded)", 9, st);
2948 }
2949
2950 #define DO_COUNTERS(macro) \
2951 macro(Deoptimization, fetch_unroll_info) \
2952 macro(Deoptimization, unpack_frames) \
2953 macro(Deoptimization, uncommon_trap_inner) \
2954 macro(Deoptimization, uncommon_trap)
2955
2956 #define INIT_COUNTER(sub, name) \
2957 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \
2958 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count");
2959
2960 void Deoptimization::init_counters() {
2961 if (ProfileRuntimeCalls && UsePerfData) {
2962 EXCEPTION_MARK;
2963
2964 DO_COUNTERS(INIT_COUNTER)
2965
2966 if (HAS_PENDING_EXCEPTION) {
2967 vm_exit_during_initialization("jvm_perf_init failed unexpectedly");
2968 }
2969 }
2970 }
2971 #undef INIT_COUNTER
2972
2973 #define PRINT_COUNTER(sub, name) { \
2974 jlong count = _perf_##sub##_##name##_count->get_value(); \
2975 if (count > 0) { \
2976 st->print_cr(" %-50s = " JLONG_FORMAT_W(6) "us (elapsed) " JLONG_FORMAT_W(6) "us (thread) (" JLONG_FORMAT_W(5) " events)", #sub "::" #name, \
2977 _perf_##sub##_##name##_timer->elapsed_counter_value_us(), \
2978 _perf_##sub##_##name##_timer->thread_counter_value_us(), \
2979 count); \
2980 }}
2981
2982 void Deoptimization::print_counters_on(outputStream* st) {
2983 if (ProfileRuntimeCalls && UsePerfData) {
2984 DO_COUNTERS(PRINT_COUNTER)
2985 } else {
2986 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData"));
2987 }
2988 }
2989
2990 #undef PRINT_COUNTER
2991 #undef DO_COUNTERS
2992
2993 #else // COMPILER2_OR_JVMCI
2994
2995
2996 // Stubs for C1 only system.
2997 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2998 return false;
2999 }
3000
3001 const char* Deoptimization::trap_reason_name(int reason) {
3002 return "unknown";
3003 }
3004
3005 jint Deoptimization::total_deoptimization_count() {
3006 return 0;
3007 }
3008
3009 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3010 return 0;
3011 }
3012
3013 void Deoptimization::print_statistics() {
3014 // no output
3015 }
3016
3017 void
3018 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3019 // no update
3020 }
3021
3022 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3023 return 0;
3024 }
3025
3026 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
3027 Bytecodes::Code bc) {
3028 // no update
3029 }
3030
3031 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3032 int trap_state) {
3033 jio_snprintf(buf, buflen, "#%d", trap_state);
3034 return buf;
3035 }
3036
3037 void Deoptimization::init_counters() {
3038 // nothing to do
3039 }
3040
3041 void Deoptimization::print_counters_on(outputStream* st) {
3042 // no output
3043 }
3044
3045 void Deoptimization::print_statistics_on(outputStream* st) {
3046 // no output
3047 }
3048
3049 #endif // COMPILER2_OR_JVMCI