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