1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/javaClasses.inline.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmClasses.hpp"
30 #include "code/codeCache.hpp"
31 #include "code/debugInfoRec.hpp"
32 #include "code/nmethod.hpp"
33 #include "code/pcDesc.hpp"
34 #include "code/scopeDesc.hpp"
35 #include "compiler/compilationPolicy.hpp"
36 #include "compiler/compilerDefinitions.inline.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "gc/shared/memAllocator.hpp"
39 #include "interpreter/bytecode.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/method.hpp"
55 #include "oops/objArrayKlass.hpp"
56 #include "oops/objArrayOop.inline.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "oops/typeArrayOop.inline.hpp"
59 #include "oops/verifyOopClosure.hpp"
60 #include "prims/jvmtiDeferredUpdates.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "prims/jvmtiThreadState.hpp"
63 #include "prims/methodHandles.hpp"
64 #include "prims/vectorSupport.hpp"
65 #include "runtime/atomic.hpp"
66 #include "runtime/basicLock.inline.hpp"
67 #include "runtime/continuation.hpp"
68 #include "runtime/continuationEntry.inline.hpp"
69 #include "runtime/deoptimization.hpp"
70 #include "runtime/escapeBarrier.hpp"
71 #include "runtime/fieldDescriptor.hpp"
72 #include "runtime/fieldDescriptor.inline.hpp"
73 #include "runtime/frame.inline.hpp"
74 #include "runtime/handles.inline.hpp"
75 #include "runtime/interfaceSupport.inline.hpp"
76 #include "runtime/javaThread.hpp"
77 #include "runtime/jniHandles.inline.hpp"
78 #include "runtime/keepStackGCProcessed.hpp"
79 #include "runtime/lightweightSynchronizer.hpp"
80 #include "runtime/lockStack.inline.hpp"
81 #include "runtime/objectMonitor.inline.hpp"
82 #include "runtime/osThread.hpp"
83 #include "runtime/perfData.inline.hpp"
84 #include "runtime/safepointVerifiers.hpp"
85 #include "runtime/sharedRuntime.hpp"
86 #include "runtime/signature.hpp"
87 #include "runtime/stackFrameStream.inline.hpp"
88 #include "runtime/stackValue.hpp"
89 #include "runtime/stackWatermarkSet.hpp"
90 #include "runtime/stubRoutines.hpp"
91 #include "runtime/synchronizer.inline.hpp"
92 #include "runtime/threadSMR.hpp"
93 #include "runtime/threadWXSetters.inline.hpp"
94 #include "runtime/vframe.hpp"
95 #include "runtime/vframeArray.hpp"
96 #include "runtime/vframe_hp.hpp"
97 #include "runtime/vmOperations.hpp"
98 #include "services/management.hpp"
99 #include "utilities/checkedCast.hpp"
100 #include "utilities/events.hpp"
101 #include "utilities/growableArray.hpp"
102 #include "utilities/macros.hpp"
103 #include "utilities/preserveException.hpp"
104 #include "utilities/xmlstream.hpp"
105 #if INCLUDE_JFR
106 #include "jfr/jfrEvents.hpp"
107 #include "jfr/metadata/jfrSerializer.hpp"
108 #endif
109
110 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
111 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
112 bool DeoptimizationScope::_committing_in_progress = false;
113
114 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
115 DEBUG_ONLY(_deopted = false;)
116
117 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
118 // If there is nothing to deopt _required_gen is the same as comitted.
119 _required_gen = DeoptimizationScope::_committed_deopt_gen;
120 }
121
122 DeoptimizationScope::~DeoptimizationScope() {
123 assert(_deopted, "Deopt not executed");
124 }
125
126 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
127 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
128
129 // If it's already marked but we still need it to be deopted.
130 if (nm->is_marked_for_deoptimization()) {
131 dependent(nm);
132 return;
133 }
134
135 nmethod::DeoptimizationStatus status =
136 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
137 Atomic::store(&nm->_deoptimization_status, status);
138
139 // Make sure active is not committed
140 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
141 assert(nm->_deoptimization_generation == 0, "Is already marked");
142
143 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
144 _required_gen = DeoptimizationScope::_active_deopt_gen;
145 }
146
147 void DeoptimizationScope::dependent(nmethod* nm) {
148 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
149
150 // A method marked by someone else may have a _required_gen lower than what we marked with.
151 // Therefore only store it if it's higher than _required_gen.
152 if (_required_gen < nm->_deoptimization_generation) {
153 _required_gen = nm->_deoptimization_generation;
154 }
155 }
156
157 void DeoptimizationScope::deoptimize_marked() {
158 assert(!_deopted, "Already deopted");
159
160 // We are not alive yet.
161 if (!Universe::is_fully_initialized()) {
162 DEBUG_ONLY(_deopted = true;)
163 return;
164 }
165
166 // Safepoints are a special case, handled here.
167 if (SafepointSynchronize::is_at_safepoint()) {
168 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
169 DeoptimizationScope::_active_deopt_gen++;
170 Deoptimization::deoptimize_all_marked();
171 DEBUG_ONLY(_deopted = true;)
172 return;
173 }
174
175 uint64_t comitting = 0;
176 bool wait = false;
177 while (true) {
178 {
179 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
180
181 // First we check if we or someone else already deopted the gen we want.
182 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
183 DEBUG_ONLY(_deopted = true;)
184 return;
185 }
186 if (!_committing_in_progress) {
187 // The version we are about to commit.
188 comitting = DeoptimizationScope::_active_deopt_gen;
189 // Make sure new marks use a higher gen.
190 DeoptimizationScope::_active_deopt_gen++;
191 _committing_in_progress = true;
192 wait = false;
193 } else {
194 // Another thread is handshaking and committing a gen.
195 wait = true;
196 }
197 }
198 if (wait) {
199 // Wait and let the concurrent handshake be performed.
200 ThreadBlockInVM tbivm(JavaThread::current());
201 os::naked_yield();
202 } else {
203 // Performs the handshake.
204 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
205 DEBUG_ONLY(_deopted = true;)
206 {
207 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
208
209 // Make sure that committed doesn't go backwards.
210 // Should only happen if we did a deopt during a safepoint above.
211 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
212 DeoptimizationScope::_committed_deopt_gen = comitting;
213 }
214 _committing_in_progress = false;
215
216 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
217
218 return;
219 }
220 }
221 }
222 }
223
224 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
225 int caller_adjustment,
226 int caller_actual_parameters,
227 int number_of_frames,
228 intptr_t* frame_sizes,
229 address* frame_pcs,
230 BasicType return_type,
231 int exec_mode) {
232 _size_of_deoptimized_frame = size_of_deoptimized_frame;
233 _caller_adjustment = caller_adjustment;
234 _caller_actual_parameters = caller_actual_parameters;
235 _number_of_frames = number_of_frames;
236 _frame_sizes = frame_sizes;
237 _frame_pcs = frame_pcs;
238 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
239 _return_type = return_type;
240 _initial_info = 0;
241 // PD (x86 only)
242 _counter_temp = 0;
243 _unpack_kind = exec_mode;
244 _sender_sp_temp = 0;
245
246 _total_frame_sizes = size_of_frames();
247 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
248 }
249
250 Deoptimization::UnrollBlock::~UnrollBlock() {
251 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
252 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
253 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
254 }
255
256 int Deoptimization::UnrollBlock::size_of_frames() const {
257 // Account first for the adjustment of the initial frame
258 intptr_t result = _caller_adjustment;
259 for (int index = 0; index < number_of_frames(); index++) {
260 result += frame_sizes()[index];
261 }
262 return checked_cast<int>(result);
263 }
264
265 void Deoptimization::UnrollBlock::print() {
266 ResourceMark rm;
267 stringStream st;
268 st.print_cr("UnrollBlock");
269 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
270 st.print( " frame_sizes: ");
271 for (int index = 0; index < number_of_frames(); index++) {
272 st.print(INTX_FORMAT " ", frame_sizes()[index]);
273 }
274 st.cr();
275 tty->print_raw(st.freeze());
276 }
277
278 // In order to make fetch_unroll_info work properly with escape
279 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
280 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
281 // which is called from the method fetch_unroll_info_helper below.
282 JRT_BLOCK_ENTRY_PROF(Deoptimization::UnrollBlock*, Deoptimization, fetch_unroll_info, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
283 // fetch_unroll_info() is called at the beginning of the deoptimization
284 // handler. Note this fact before we start generating temporary frames
285 // that can confuse an asynchronous stack walker. This counter is
286 // decremented at the end of unpack_frames().
287 current->inc_in_deopt_handler();
288
289 if (exec_mode == Unpack_exception) {
290 // When we get here, a callee has thrown an exception into a deoptimized
291 // frame. That throw might have deferred stack watermark checking until
292 // after unwinding. So we deal with such deferred requests here.
293 StackWatermarkSet::after_unwind(current);
294 }
295
296 return fetch_unroll_info_helper(current, exec_mode);
297 JRT_END
298
299 #if COMPILER2_OR_JVMCI
300 // print information about reallocated objects
301 static void print_objects(JavaThread* deoptee_thread,
302 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
303 ResourceMark rm;
304 stringStream st; // change to logStream with logging
305 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
306 fieldDescriptor fd;
307
308 for (int i = 0; i < objects->length(); i++) {
309 ObjectValue* sv = (ObjectValue*) objects->at(i);
310 Handle obj = sv->value();
311
312 if (obj.is_null()) {
313 st.print_cr(" nullptr");
314 continue;
315 }
316
317 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
318
319 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
320 k->print_value_on(&st);
321 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize);
322
323 if (Verbose && k != nullptr) {
324 k->oop_print_on(obj(), &st);
325 }
326 }
327 tty->print_raw(st.freeze());
328 }
329
330 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
331 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
332 bool& deoptimized_objects) {
333 bool realloc_failures = false;
334 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
335
336 JavaThread* deoptee_thread = chunk->at(0)->thread();
337 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
338 "a frame can only be deoptimized by the owner thread");
339
340 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
341
342 // The flag return_oop() indicates call sites which return oop
343 // in compiled code. Such sites include java method calls,
344 // runtime calls (for example, used to allocate new objects/arrays
345 // on slow code path) and any other calls generated in compiled code.
346 // It is not guaranteed that we can get such information here only
347 // by analyzing bytecode in deoptimized frames. This is why this flag
348 // is set during method compilation (see Compile::Process_OopMap_Node()).
349 // If the previous frame was popped or if we are dispatching an exception,
350 // we don't have an oop result.
351 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
352 Handle return_value;
353 if (save_oop_result) {
354 // Reallocation may trigger GC. If deoptimization happened on return from
355 // call which returns oop we need to save it since it is not in oopmap.
356 oop result = deoptee.saved_oop_result(&map);
357 assert(oopDesc::is_oop_or_null(result), "must be oop");
358 return_value = Handle(thread, result);
359 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
360 if (TraceDeoptimization) {
361 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
362 tty->cr();
363 }
364 }
365 if (objects != nullptr) {
366 if (exec_mode == Deoptimization::Unpack_none) {
367 assert(thread->thread_state() == _thread_in_vm, "assumption");
368 JavaThread* THREAD = thread; // For exception macros.
369 // Clear pending OOM if reallocation fails and return true indicating allocation failure
370 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
371 deoptimized_objects = true;
372 } else {
373 JavaThread* current = thread; // For JRT_BLOCK
374 JRT_BLOCK
375 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
376 JRT_END
377 }
378 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci();
379 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
380 if (TraceDeoptimization) {
381 print_objects(deoptee_thread, objects, realloc_failures);
382 }
383 }
384 if (save_oop_result) {
385 // Restore result.
386 deoptee.set_saved_oop_result(&map, return_value());
387 }
388 return realloc_failures;
389 }
390
391 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
392 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
393 JavaThread* deoptee_thread = chunk->at(0)->thread();
394 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
395 assert(thread == Thread::current(), "should be");
396 HandleMark hm(thread);
397 #ifndef PRODUCT
398 bool first = true;
399 #endif // !PRODUCT
400 // Start locking from outermost/oldest frame
401 for (int i = (chunk->length() - 1); i >= 0; i--) {
402 compiledVFrame* cvf = chunk->at(i);
403 assert (cvf->scope() != nullptr,"expect only compiled java frames");
404 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
405 if (monitors->is_nonempty()) {
406 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
407 exec_mode, realloc_failures);
408 deoptimized_objects = deoptimized_objects || relocked;
409 #ifndef PRODUCT
410 if (PrintDeoptimizationDetails) {
411 ResourceMark rm;
412 stringStream st;
413 for (int j = 0; j < monitors->length(); j++) {
414 MonitorInfo* mi = monitors->at(j);
415 if (mi->eliminated()) {
416 if (first) {
417 first = false;
418 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
419 }
420 if (exec_mode == Deoptimization::Unpack_none) {
421 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
422 if (monitor != nullptr && monitor->object() == mi->owner()) {
423 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
424 continue;
425 }
426 }
427 if (mi->owner_is_scalar_replaced()) {
428 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
429 st.print_cr(" failed reallocation for klass %s", k->external_name());
430 } else {
431 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
432 }
433 }
434 }
435 tty->print_raw(st.freeze());
436 }
437 #endif // !PRODUCT
438 }
439 }
440 }
441
442 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
443 // The given vframes cover one physical frame.
444 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
445 bool& realloc_failures) {
446 frame deoptee = chunk->at(0)->fr();
447 JavaThread* deoptee_thread = chunk->at(0)->thread();
448 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
449 RegisterMap map(chunk->at(0)->register_map());
450 bool deoptimized_objects = false;
451
452 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
453
454 // Reallocate the non-escaping objects and restore their fields.
455 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
456 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
457 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
458 }
459
460 // MonitorInfo structures used in eliminate_locks are not GC safe.
461 NoSafepointVerifier no_safepoint;
462
463 // Now relock objects if synchronization on them was eliminated.
464 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
465 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
466 }
467 return deoptimized_objects;
468 }
469 #endif // COMPILER2_OR_JVMCI
470
471 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
472 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
473 // When we get here we are about to unwind the deoptee frame. In order to
474 // catch not yet safe to use frames, the following stack watermark barrier
475 // poll will make such frames safe to use.
476 StackWatermarkSet::before_unwind(current);
477
478 // Note: there is a safepoint safety issue here. No matter whether we enter
479 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
480 // the vframeArray is created.
481 //
482
483 // Allocate our special deoptimization ResourceMark
484 DeoptResourceMark* dmark = new DeoptResourceMark(current);
485 assert(current->deopt_mark() == nullptr, "Pending deopt!");
486 current->set_deopt_mark(dmark);
487
488 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
489 RegisterMap map(current,
490 RegisterMap::UpdateMap::include,
491 RegisterMap::ProcessFrames::include,
492 RegisterMap::WalkContinuation::skip);
493 RegisterMap dummy_map(current,
494 RegisterMap::UpdateMap::skip,
495 RegisterMap::ProcessFrames::include,
496 RegisterMap::WalkContinuation::skip);
497 // Now get the deoptee with a valid map
498 frame deoptee = stub_frame.sender(&map);
499 // Set the deoptee nmethod
500 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
501 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
502 current->set_deopt_compiled_method(nm);
503
504 if (VerifyStack) {
505 current->validate_frame_layout();
506 }
507
508 // Create a growable array of VFrames where each VFrame represents an inlined
509 // Java frame. This storage is allocated with the usual system arena.
510 assert(deoptee.is_compiled_frame(), "Wrong frame type");
511 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
512 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
513 while (!vf->is_top()) {
514 assert(vf->is_compiled_frame(), "Wrong frame type");
515 chunk->push(compiledVFrame::cast(vf));
516 vf = vf->sender();
517 }
518 assert(vf->is_compiled_frame(), "Wrong frame type");
519 chunk->push(compiledVFrame::cast(vf));
520
521 bool realloc_failures = false;
522
523 #if COMPILER2_OR_JVMCI
524 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
525
526 // Reallocate the non-escaping objects and restore their fields. Then
527 // relock objects if synchronization on them was eliminated.
528 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
529 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
530 bool unused;
531 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
532 }
533 #endif // COMPILER2_OR_JVMCI
534
535 // Ensure that no safepoint is taken after pointers have been stored
536 // in fields of rematerialized objects. If a safepoint occurs from here on
537 // out the java state residing in the vframeArray will be missed.
538 // Locks may be rebaised in a safepoint.
539 NoSafepointVerifier no_safepoint;
540
541 #if COMPILER2_OR_JVMCI
542 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
543 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
544 bool unused = false;
545 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
546 }
547 #endif // COMPILER2_OR_JVMCI
548
549 ScopeDesc* trap_scope = chunk->at(0)->scope();
550 Handle exceptionObject;
551 if (trap_scope->rethrow_exception()) {
552 #ifndef PRODUCT
553 if (PrintDeoptimizationDetails) {
554 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());
555 }
556 #endif // !PRODUCT
557
558 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
559 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
560 ScopeValue* topOfStack = expressions->top();
561 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
562 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
563 }
564
565 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
566 #if COMPILER2_OR_JVMCI
567 if (realloc_failures) {
568 // This destroys all ScopedValue bindings.
569 current->clear_scopedValueBindings();
570 pop_frames_failed_reallocs(current, array);
571 }
572 #endif
573
574 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
575 current->set_vframe_array_head(array);
576
577 // Now that the vframeArray has been created if we have any deferred local writes
578 // added by jvmti then we can free up that structure as the data is now in the
579 // vframeArray
580
581 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
582
583 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
584 CodeBlob* cb = stub_frame.cb();
585 // Verify we have the right vframeArray
586 assert(cb->frame_size() >= 0, "Unexpected frame size");
587 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
588
589 // If the deopt call site is a MethodHandle invoke call site we have
590 // to adjust the unpack_sp.
591 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
592 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
593 unpack_sp = deoptee.unextended_sp();
594
595 #ifdef ASSERT
596 assert(cb->is_deoptimization_stub() ||
597 cb->is_uncommon_trap_stub() ||
598 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
599 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
600 "unexpected code blob: %s", cb->name());
601 #endif
602
603 // This is a guarantee instead of an assert because if vframe doesn't match
604 // we will unpack the wrong deoptimized frame and wind up in strange places
605 // where it will be very difficult to figure out what went wrong. Better
606 // to die an early death here than some very obscure death later when the
607 // trail is cold.
608 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
609 // in that it will fail to detect a problem when there is one. This needs
610 // more work in tiger timeframe.
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 = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
647 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
648 // Method handle invokes may involve fairly arbitrary chains of
649 // calls so it's impossible to know how much actual space the
650 // caller has for locals.
651 caller_was_method_handle = true;
652 }
653 }
654
655 //
656 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
657 // frame_sizes/frame_pcs[1] next oldest frame (int)
658 // frame_sizes/frame_pcs[n] youngest frame (int)
659 //
660 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
661 // owns the space for the return address to it's caller). Confusing ain't it.
662 //
663 // The vframe array can address vframes with indices running from
664 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
665 // When we create the skeletal frames we need the oldest frame to be in the zero slot
666 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
667 // so things look a little strange in this loop.
668 //
669 int callee_parameters = 0;
670 int callee_locals = 0;
671 for (int index = 0; index < array->frames(); index++ ) {
672 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
673 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
674 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
675 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
676 callee_locals,
677 index == 0,
678 popframe_extra_args);
679 // This pc doesn't have to be perfect just good enough to identify the frame
680 // as interpreted so the skeleton frame will be walkable
681 // The correct pc will be set when the skeleton frame is completely filled out
682 // The final pc we store in the loop is wrong and will be overwritten below
683 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
684
685 callee_parameters = array->element(index)->method()->size_of_parameters();
686 callee_locals = array->element(index)->method()->max_locals();
687 popframe_extra_args = 0;
688 }
689
690 // Compute whether the root vframe returns a float or double value.
691 BasicType return_type;
692 {
693 methodHandle method(current, array->element(0)->method());
694 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
695 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
696 }
697
698 // Compute information for handling adapters and adjusting the frame size of the caller.
699 int caller_adjustment = 0;
700
701 // Compute the amount the oldest interpreter frame will have to adjust
702 // its caller's stack by. If the caller is a compiled frame then
703 // we pretend that the callee has no parameters so that the
704 // extension counts for the full amount of locals and not just
705 // locals-parms. This is because without a c2i adapter the parm
706 // area as created by the compiled frame will not be usable by
707 // the interpreter. (Depending on the calling convention there
708 // may not even be enough space).
709
710 // QQQ I'd rather see this pushed down into last_frame_adjust
711 // and have it take the sender (aka caller).
712
713 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
714 caller_adjustment = last_frame_adjust(0, callee_locals);
715 } else if (callee_locals > callee_parameters) {
716 // The caller frame may need extending to accommodate
717 // non-parameter locals of the first unpacked interpreted frame.
718 // Compute that adjustment.
719 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
720 }
721
722 // If the sender is deoptimized the we must retrieve the address of the handler
723 // since the frame will "magically" show the original pc before the deopt
724 // and we'd undo the deopt.
725
726 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
727 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
728 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
729 }
730
731 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
732
733 #if INCLUDE_JVMCI
734 if (exceptionObject() != nullptr) {
735 current->set_exception_oop(exceptionObject());
736 exec_mode = Unpack_exception;
737 }
738 #endif
739
740 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
741 assert(current->has_pending_exception(), "should have thrown OOME");
742 current->set_exception_oop(current->pending_exception());
743 current->clear_pending_exception();
744 exec_mode = Unpack_exception;
745 }
746
747 #if INCLUDE_JVMCI
748 if (current->frames_to_pop_failed_realloc() > 0) {
749 current->set_pending_monitorenter(false);
750 }
751 #endif
752
753 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
754 caller_adjustment * BytesPerWord,
755 caller_was_method_handle ? 0 : callee_parameters,
756 number_of_frames,
757 frame_sizes,
758 frame_pcs,
759 return_type,
760 exec_mode);
761 // On some platforms, we need a way to pass some platform dependent
762 // information to the unpacking code so the skeletal frames come out
763 // correct (initial fp value, unextended sp, ...)
764 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
765
766 if (array->frames() > 1) {
767 if (VerifyStack && TraceDeoptimization) {
768 tty->print_cr("Deoptimizing method containing inlining");
769 }
770 }
771
772 array->set_unroll_block(info);
773 return info;
774 }
775
776 // Called to cleanup deoptimization data structures in normal case
777 // after unpacking to stack and when stack overflow error occurs
778 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
779 vframeArray *array) {
780
781 // Get array if coming from exception
782 if (array == nullptr) {
783 array = thread->vframe_array_head();
784 }
785 thread->set_vframe_array_head(nullptr);
786
787 // Free the previous UnrollBlock
788 vframeArray* old_array = thread->vframe_array_last();
789 thread->set_vframe_array_last(array);
790
791 if (old_array != nullptr) {
792 UnrollBlock* old_info = old_array->unroll_block();
793 old_array->set_unroll_block(nullptr);
794 delete old_info;
795 delete old_array;
796 }
797
798 // Deallocate any resource creating in this routine and any ResourceObjs allocated
799 // inside the vframeArray (StackValueCollections)
800
801 delete thread->deopt_mark();
802 thread->set_deopt_mark(nullptr);
803 thread->set_deopt_compiled_method(nullptr);
804
805
806 if (JvmtiExport::can_pop_frame()) {
807 // Regardless of whether we entered this routine with the pending
808 // popframe condition bit set, we should always clear it now
809 thread->clear_popframe_condition();
810 }
811
812 // unpack_frames() is called at the end of the deoptimization handler
813 // and (in C2) at the end of the uncommon trap handler. Note this fact
814 // so that an asynchronous stack walker can work again. This counter is
815 // incremented at the beginning of fetch_unroll_info() and (in C2) at
816 // the beginning of uncommon_trap().
817 thread->dec_in_deopt_handler();
818 }
819
820 // Moved from cpu directories because none of the cpus has callee save values.
821 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
822 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
823
824 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
825 // the days we had adapter frames. When we deoptimize a situation where a
826 // compiled caller calls a compiled caller will have registers it expects
827 // to survive the call to the callee. If we deoptimize the callee the only
828 // way we can restore these registers is to have the oldest interpreter
829 // frame that we create restore these values. That is what this routine
830 // will accomplish.
831
832 // At the moment we have modified c2 to not have any callee save registers
833 // so this problem does not exist and this routine is just a place holder.
834
835 assert(f->is_interpreted_frame(), "must be interpreted");
836 }
837
838 #ifndef PRODUCT
839 static bool falls_through(Bytecodes::Code bc) {
840 switch (bc) {
841 // List may be incomplete. Here we really only care about bytecodes where compiled code
842 // can deoptimize.
843 case Bytecodes::_goto:
844 case Bytecodes::_goto_w:
845 case Bytecodes::_athrow:
846 return false;
847 default:
848 return true;
849 }
850 }
851 #endif
852
853 // Return BasicType of value being returned
854 JRT_LEAF_PROF_NO_THREAD(BasicType, Deoptimization, unpack_frames, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
855 assert(thread == JavaThread::current(), "pre-condition");
856
857 // We are already active in the special DeoptResourceMark any ResourceObj's we
858 // allocate will be freed at the end of the routine.
859
860 // JRT_LEAF methods don't normally allocate handles and there is a
861 // NoHandleMark to enforce that. It is actually safe to use Handles
862 // in a JRT_LEAF method, and sometimes desirable, but to do so we
863 // must use ResetNoHandleMark to bypass the NoHandleMark, and
864 // then use a HandleMark to ensure any Handles we do create are
865 // cleaned up in this scope.
866 ResetNoHandleMark rnhm;
867 HandleMark hm(thread);
868
869 frame stub_frame = thread->last_frame();
870
871 Continuation::notify_deopt(thread, stub_frame.sp());
872
873 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
874 // must point to the vframeArray for the unpack frame.
875 vframeArray* array = thread->vframe_array_head();
876 UnrollBlock* info = array->unroll_block();
877
878 // We set the last_Java frame. But the stack isn't really parsable here. So we
879 // clear it to make sure JFR understands not to try and walk stacks from events
880 // in here.
881 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
882 thread->frame_anchor()->set_last_Java_sp(nullptr);
883
884 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
885 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
886
887 thread->frame_anchor()->set_last_Java_sp(sp);
888
889 BasicType bt = info->return_type();
890
891 // If we have an exception pending, claim that the return type is an oop
892 // so the deopt_blob does not overwrite the exception_oop.
893
894 if (exec_mode == Unpack_exception)
895 bt = T_OBJECT;
896
897 // Cleanup thread deopt data
898 cleanup_deopt_info(thread, array);
899
900 #ifndef PRODUCT
901 if (VerifyStack) {
902 ResourceMark res_mark;
903 // Clear pending exception to not break verification code (restored afterwards)
904 PreserveExceptionMark pm(thread);
905
906 thread->validate_frame_layout();
907
908 // Verify that the just-unpacked frames match the interpreter's
909 // notions of expression stack and locals
910 vframeArray* cur_array = thread->vframe_array_last();
911 RegisterMap rm(thread,
912 RegisterMap::UpdateMap::skip,
913 RegisterMap::ProcessFrames::include,
914 RegisterMap::WalkContinuation::skip);
915 rm.set_include_argument_oops(false);
916 bool is_top_frame = true;
917 int callee_size_of_parameters = 0;
918 int callee_max_locals = 0;
919 for (int i = 0; i < cur_array->frames(); i++) {
920 vframeArrayElement* el = cur_array->element(i);
921 frame* iframe = el->iframe();
922 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
923
924 // Get the oop map for this bci
925 InterpreterOopMap mask;
926 int cur_invoke_parameter_size = 0;
927 bool try_next_mask = false;
928 int next_mask_expression_stack_size = -1;
929 int top_frame_expression_stack_adjustment = 0;
930 methodHandle mh(thread, iframe->interpreter_frame_method());
931 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
932 BytecodeStream str(mh, iframe->interpreter_frame_bci());
933 int max_bci = mh->code_size();
934 // Get to the next bytecode if possible
935 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
936 // Check to see if we can grab the number of outgoing arguments
937 // at an uncommon trap for an invoke (where the compiler
938 // generates debug info before the invoke has executed)
939 Bytecodes::Code cur_code = str.next();
940 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
941 if (Bytecodes::is_invoke(cur_code)) {
942 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
943 cur_invoke_parameter_size = invoke.size_of_parameters();
944 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
945 callee_size_of_parameters++;
946 }
947 }
948 if (str.bci() < max_bci) {
949 next_code = str.next();
950 if (next_code >= 0) {
951 // The interpreter oop map generator reports results before
952 // the current bytecode has executed except in the case of
953 // calls. It seems to be hard to tell whether the compiler
954 // has emitted debug information matching the "state before"
955 // a given bytecode or the state after, so we try both
956 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
957 // Get expression stack size for the next bytecode
958 InterpreterOopMap next_mask;
959 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
960 next_mask_expression_stack_size = next_mask.expression_stack_size();
961 if (Bytecodes::is_invoke(next_code)) {
962 Bytecode_invoke invoke(mh, str.bci());
963 next_mask_expression_stack_size += invoke.size_of_parameters();
964 }
965 // Need to subtract off the size of the result type of
966 // the bytecode because this is not described in the
967 // debug info but returned to the interpreter in the TOS
968 // caching register
969 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
970 if (bytecode_result_type != T_ILLEGAL) {
971 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
972 }
973 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
974 try_next_mask = true;
975 }
976 }
977 }
978
979 // Verify stack depth and oops in frame
980 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
981 if (!(
982 /* SPARC */
983 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
984 /* x86 */
985 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
986 (try_next_mask &&
987 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
988 top_frame_expression_stack_adjustment))) ||
989 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
990 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
991 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
992 )) {
993 {
994 // Print out some information that will help us debug the problem
995 tty->print_cr("Wrong number of expression stack elements during deoptimization");
996 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
997 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
998 if (try_next_mask) {
999 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1000 }
1001 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1002 iframe->interpreter_frame_expression_stack_size());
1003 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1004 tty->print_cr(" try_next_mask = %d", try_next_mask);
1005 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1006 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1007 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1008 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1009 tty->print_cr(" exec_mode = %d", exec_mode);
1010 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1011 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1012 tty->print_cr(" Interpreted frames:");
1013 for (int k = 0; k < cur_array->frames(); k++) {
1014 vframeArrayElement* el = cur_array->element(k);
1015 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1016 }
1017 cur_array->print_on_2(tty);
1018 }
1019 guarantee(false, "wrong number of expression stack elements during deopt");
1020 }
1021 VerifyOopClosure verify;
1022 iframe->oops_interpreted_do(&verify, &rm, false);
1023 callee_size_of_parameters = mh->size_of_parameters();
1024 callee_max_locals = mh->max_locals();
1025 is_top_frame = false;
1026 }
1027 }
1028 #endif // !PRODUCT
1029
1030 return bt;
1031 JRT_END
1032
1033 class DeoptimizeMarkedClosure : public HandshakeClosure {
1034 public:
1035 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1036 void do_thread(Thread* thread) {
1037 JavaThread* jt = JavaThread::cast(thread);
1038 jt->deoptimize_marked_methods();
1039 }
1040 };
1041
1042 void Deoptimization::deoptimize_all_marked() {
1043 ResourceMark rm;
1044
1045 // Make the dependent methods not entrant
1046 CodeCache::make_marked_nmethods_deoptimized();
1047
1048 DeoptimizeMarkedClosure deopt;
1049 if (SafepointSynchronize::is_at_safepoint()) {
1050 Threads::java_threads_do(&deopt);
1051 } else {
1052 Handshake::execute(&deopt);
1053 }
1054 }
1055
1056 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1057 = Deoptimization::Action_reinterpret;
1058
1059 #if INCLUDE_JVMCI
1060 template<typename CacheType>
1061 class BoxCacheBase : public CHeapObj<mtCompiler> {
1062 protected:
1063 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1064 ResourceMark rm(thread);
1065 char* klass_name_str = klass_name->as_C_string();
1066 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle());
1067 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1068 if (!ik->is_in_error_state()) {
1069 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1070 CacheType::compute_offsets(ik);
1071 }
1072 return ik;
1073 }
1074 };
1075
1076 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1077 PrimitiveType _low;
1078 PrimitiveType _high;
1079 jobject _cache;
1080 protected:
1081 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1082 BoxCache(Thread* thread) {
1083 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1084 if (ik->is_in_error_state()) {
1085 _low = 1;
1086 _high = 0;
1087 _cache = nullptr;
1088 } else {
1089 objArrayOop cache = CacheType::cache(ik);
1090 assert(cache->length() > 0, "Empty cache");
1091 _low = BoxType::value(cache->obj_at(0));
1092 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1093 _cache = JNIHandles::make_global(Handle(thread, cache));
1094 }
1095 }
1096 ~BoxCache() {
1097 JNIHandles::destroy_global(_cache);
1098 }
1099 public:
1100 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1101 if (_singleton == nullptr) {
1102 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1103 if (!Atomic::replace_if_null(&_singleton, s)) {
1104 delete s;
1105 }
1106 }
1107 return _singleton;
1108 }
1109 oop lookup(PrimitiveType value) {
1110 if (_low <= value && value <= _high) {
1111 int offset = checked_cast<int>(value - _low);
1112 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1113 }
1114 return nullptr;
1115 }
1116 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1117 if (_cache == nullptr) {
1118 cache_init_error = true;
1119 return nullptr;
1120 }
1121 // Have to cast to avoid little/big-endian problems.
1122 if (sizeof(PrimitiveType) > sizeof(jint)) {
1123 jlong value = (jlong)raw_value;
1124 return lookup(value);
1125 }
1126 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1127 return lookup(value);
1128 }
1129 };
1130
1131 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1132 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1133 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1134 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1135 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1136
1137 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1138 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1139 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1140 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1141 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1142
1143 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1144 jobject _true_cache;
1145 jobject _false_cache;
1146 protected:
1147 static BooleanBoxCache *_singleton;
1148 BooleanBoxCache(Thread *thread) {
1149 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1150 if (ik->is_in_error_state()) {
1151 _true_cache = nullptr;
1152 _false_cache = nullptr;
1153 } else {
1154 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1155 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1156 }
1157 }
1158 ~BooleanBoxCache() {
1159 JNIHandles::destroy_global(_true_cache);
1160 JNIHandles::destroy_global(_false_cache);
1161 }
1162 public:
1163 static BooleanBoxCache* singleton(Thread* thread) {
1164 if (_singleton == nullptr) {
1165 BooleanBoxCache* s = new BooleanBoxCache(thread);
1166 if (!Atomic::replace_if_null(&_singleton, s)) {
1167 delete s;
1168 }
1169 }
1170 return _singleton;
1171 }
1172 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1173 if (_true_cache == nullptr) {
1174 cache_in_error = true;
1175 return nullptr;
1176 }
1177 // Have to cast to avoid little/big-endian problems.
1178 jboolean value = (jboolean)*((jint*)&raw_value);
1179 return lookup(value);
1180 }
1181 oop lookup(jboolean value) {
1182 if (value != 0) {
1183 return JNIHandles::resolve_non_null(_true_cache);
1184 }
1185 return JNIHandles::resolve_non_null(_false_cache);
1186 }
1187 };
1188
1189 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1190
1191 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1192 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1193 BasicType box_type = vmClasses::box_klass_type(k);
1194 if (box_type != T_OBJECT) {
1195 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1196 switch(box_type) {
1197 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1198 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1199 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1200 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1201 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1202 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1203 default:;
1204 }
1205 }
1206 return nullptr;
1207 }
1208 #endif // INCLUDE_JVMCI
1209
1210 #if COMPILER2_OR_JVMCI
1211 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1212 Handle pending_exception(THREAD, thread->pending_exception());
1213 const char* exception_file = thread->exception_file();
1214 int exception_line = thread->exception_line();
1215 thread->clear_pending_exception();
1216
1217 bool failures = false;
1218
1219 for (int i = 0; i < objects->length(); i++) {
1220 assert(objects->at(i)->is_object(), "invalid debug information");
1221 ObjectValue* sv = (ObjectValue*) objects->at(i);
1222
1223 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1224 oop obj = nullptr;
1225
1226 bool cache_init_error = false;
1227 if (k->is_instance_klass()) {
1228 #if INCLUDE_JVMCI
1229 nmethod* nm = fr->cb()->as_nmethod_or_null();
1230 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1231 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1232 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1233 if (obj != nullptr) {
1234 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1235 abv->set_cached(true);
1236 } else if (cache_init_error) {
1237 // Results in an OOME which is valid (as opposed to a class initialization error)
1238 // and is fine for the rare case a cache initialization failing.
1239 failures = true;
1240 }
1241 }
1242 #endif // INCLUDE_JVMCI
1243
1244 InstanceKlass* ik = InstanceKlass::cast(k);
1245 if (obj == nullptr && !cache_init_error) {
1246 InternalOOMEMark iom(THREAD);
1247 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1248 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1249 } else {
1250 obj = ik->allocate_instance(THREAD);
1251 }
1252 }
1253 } else if (k->is_typeArray_klass()) {
1254 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1255 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1256 int len = sv->field_size() / type2size[ak->element_type()];
1257 InternalOOMEMark iom(THREAD);
1258 obj = ak->allocate(len, THREAD);
1259 } else if (k->is_objArray_klass()) {
1260 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1261 InternalOOMEMark iom(THREAD);
1262 obj = ak->allocate(sv->field_size(), THREAD);
1263 }
1264
1265 if (obj == nullptr) {
1266 failures = true;
1267 }
1268
1269 assert(sv->value().is_null(), "redundant reallocation");
1270 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1271 CLEAR_PENDING_EXCEPTION;
1272 sv->set_value(obj);
1273 }
1274
1275 if (failures) {
1276 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1277 } else if (pending_exception.not_null()) {
1278 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1279 }
1280
1281 return failures;
1282 }
1283
1284 #if INCLUDE_JVMCI
1285 /**
1286 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1287 * we need to somehow be able to recover the actual kind to be able to write the correct
1288 * amount of bytes.
1289 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1290 * the entries at index n + 1 to n + i are 'markers'.
1291 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1292 * expected form of the array would be:
1293 *
1294 * {b0, b1, b2, b3, INT, marker, b6, b7}
1295 *
1296 * Thus, in order to get back the size of the entry, we simply need to count the number
1297 * of marked entries
1298 *
1299 * @param virtualArray the virtualized byte array
1300 * @param i index of the virtual entry we are recovering
1301 * @return The number of bytes the entry spans
1302 */
1303 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1304 int index = i;
1305 while (++index < virtualArray->field_size() &&
1306 virtualArray->field_at(index)->is_marker()) {}
1307 return index - i;
1308 }
1309
1310 /**
1311 * If there was a guarantee for byte array to always start aligned to a long, we could
1312 * do a simple check on the parity of the index. Unfortunately, that is not always the
1313 * case. Thus, we check alignment of the actual address we are writing to.
1314 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1315 * write a long to index 3.
1316 */
1317 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1318 jbyte* res = obj->byte_at_addr(index);
1319 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1320 return res;
1321 }
1322
1323 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1324 switch (byte_count) {
1325 case 1:
1326 obj->byte_at_put(index, (jbyte) value->get_jint());
1327 break;
1328 case 2:
1329 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1330 break;
1331 case 4:
1332 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1333 break;
1334 case 8:
1335 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1336 break;
1337 default:
1338 ShouldNotReachHere();
1339 }
1340 }
1341 #endif // INCLUDE_JVMCI
1342
1343
1344 // restore elements of an eliminated type array
1345 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1346 int index = 0;
1347
1348 for (int i = 0; i < sv->field_size(); i++) {
1349 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1350 switch(type) {
1351 case T_LONG: case T_DOUBLE: {
1352 assert(value->type() == T_INT, "Agreement.");
1353 StackValue* low =
1354 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1355 #ifdef _LP64
1356 jlong res = (jlong)low->get_intptr();
1357 #else
1358 jlong res = jlong_from(value->get_jint(), low->get_jint());
1359 #endif
1360 obj->long_at_put(index, res);
1361 break;
1362 }
1363
1364 case T_INT: case T_FLOAT: { // 4 bytes.
1365 assert(value->type() == T_INT, "Agreement.");
1366 bool big_value = false;
1367 if (i + 1 < sv->field_size() && type == T_INT) {
1368 if (sv->field_at(i)->is_location()) {
1369 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1370 if (type == Location::dbl || type == Location::lng) {
1371 big_value = true;
1372 }
1373 } else if (sv->field_at(i)->is_constant_int()) {
1374 ScopeValue* next_scope_field = sv->field_at(i + 1);
1375 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1376 big_value = true;
1377 }
1378 }
1379 }
1380
1381 if (big_value) {
1382 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1383 #ifdef _LP64
1384 jlong res = (jlong)low->get_intptr();
1385 #else
1386 jlong res = jlong_from(value->get_jint(), low->get_jint());
1387 #endif
1388 obj->int_at_put(index, *(jint*)&res);
1389 obj->int_at_put(++index, *((jint*)&res + 1));
1390 } else {
1391 obj->int_at_put(index, value->get_jint());
1392 }
1393 break;
1394 }
1395
1396 case T_SHORT:
1397 assert(value->type() == T_INT, "Agreement.");
1398 obj->short_at_put(index, (jshort)value->get_jint());
1399 break;
1400
1401 case T_CHAR:
1402 assert(value->type() == T_INT, "Agreement.");
1403 obj->char_at_put(index, (jchar)value->get_jint());
1404 break;
1405
1406 case T_BYTE: {
1407 assert(value->type() == T_INT, "Agreement.");
1408 #if INCLUDE_JVMCI
1409 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1410 int byte_count = count_number_of_bytes_for_entry(sv, i);
1411 byte_array_put(obj, value, index, byte_count);
1412 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1413 i += byte_count - 1; // Balance the loop counter.
1414 index += byte_count;
1415 // index has been updated so continue at top of loop
1416 continue;
1417 #else
1418 obj->byte_at_put(index, (jbyte)value->get_jint());
1419 break;
1420 #endif // INCLUDE_JVMCI
1421 }
1422
1423 case T_BOOLEAN: {
1424 assert(value->type() == T_INT, "Agreement.");
1425 obj->bool_at_put(index, (jboolean)value->get_jint());
1426 break;
1427 }
1428
1429 default:
1430 ShouldNotReachHere();
1431 }
1432 index++;
1433 }
1434 }
1435
1436 // restore fields of an eliminated object array
1437 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1438 for (int i = 0; i < sv->field_size(); i++) {
1439 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1440 assert(value->type() == T_OBJECT, "object element expected");
1441 obj->obj_at_put(i, value->get_obj()());
1442 }
1443 }
1444
1445 class ReassignedField {
1446 public:
1447 int _offset;
1448 BasicType _type;
1449 public:
1450 ReassignedField() {
1451 _offset = 0;
1452 _type = T_ILLEGAL;
1453 }
1454 };
1455
1456 static int compare(ReassignedField* left, ReassignedField* right) {
1457 return left->_offset - right->_offset;
1458 }
1459
1460 // Restore fields of an eliminated instance object using the same field order
1461 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1462 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1463 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1464 InstanceKlass* ik = klass;
1465 while (ik != nullptr) {
1466 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1467 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1468 ReassignedField field;
1469 field._offset = fs.offset();
1470 field._type = Signature::basic_type(fs.signature());
1471 fields->append(field);
1472 }
1473 }
1474 ik = ik->superklass();
1475 }
1476 fields->sort(compare);
1477 for (int i = 0; i < fields->length(); i++) {
1478 ScopeValue* scope_field = sv->field_at(svIndex);
1479 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1480 int offset = fields->at(i)._offset;
1481 BasicType type = fields->at(i)._type;
1482 switch (type) {
1483 case T_OBJECT: case T_ARRAY:
1484 assert(value->type() == T_OBJECT, "Agreement.");
1485 obj->obj_field_put(offset, value->get_obj()());
1486 break;
1487
1488 case T_INT: case T_FLOAT: { // 4 bytes.
1489 assert(value->type() == T_INT, "Agreement.");
1490 bool big_value = false;
1491 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1492 if (scope_field->is_location()) {
1493 Location::Type type = ((LocationValue*) scope_field)->location().type();
1494 if (type == Location::dbl || type == Location::lng) {
1495 big_value = true;
1496 }
1497 }
1498 if (scope_field->is_constant_int()) {
1499 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1500 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1501 big_value = true;
1502 }
1503 }
1504 }
1505
1506 if (big_value) {
1507 i++;
1508 assert(i < fields->length(), "second T_INT field needed");
1509 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1510 } else {
1511 obj->int_field_put(offset, value->get_jint());
1512 break;
1513 }
1514 }
1515 /* no break */
1516
1517 case T_LONG: case T_DOUBLE: {
1518 assert(value->type() == T_INT, "Agreement.");
1519 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1520 #ifdef _LP64
1521 jlong res = (jlong)low->get_intptr();
1522 #else
1523 jlong res = jlong_from(value->get_jint(), low->get_jint());
1524 #endif
1525 obj->long_field_put(offset, res);
1526 break;
1527 }
1528
1529 case T_SHORT:
1530 assert(value->type() == T_INT, "Agreement.");
1531 obj->short_field_put(offset, (jshort)value->get_jint());
1532 break;
1533
1534 case T_CHAR:
1535 assert(value->type() == T_INT, "Agreement.");
1536 obj->char_field_put(offset, (jchar)value->get_jint());
1537 break;
1538
1539 case T_BYTE:
1540 assert(value->type() == T_INT, "Agreement.");
1541 obj->byte_field_put(offset, (jbyte)value->get_jint());
1542 break;
1543
1544 case T_BOOLEAN:
1545 assert(value->type() == T_INT, "Agreement.");
1546 obj->bool_field_put(offset, (jboolean)value->get_jint());
1547 break;
1548
1549 default:
1550 ShouldNotReachHere();
1551 }
1552 svIndex++;
1553 }
1554 return svIndex;
1555 }
1556
1557 // restore fields of all eliminated objects and arrays
1558 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1559 for (int i = 0; i < objects->length(); i++) {
1560 assert(objects->at(i)->is_object(), "invalid debug information");
1561 ObjectValue* sv = (ObjectValue*) objects->at(i);
1562 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1563 Handle obj = sv->value();
1564 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1565 #ifndef PRODUCT
1566 if (PrintDeoptimizationDetails) {
1567 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1568 }
1569 #endif // !PRODUCT
1570
1571 if (obj.is_null()) {
1572 continue;
1573 }
1574
1575 #if INCLUDE_JVMCI
1576 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1577 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1578 continue;
1579 }
1580 #endif // INCLUDE_JVMCI
1581 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1582 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1583 ScopeValue* payload = sv->field_at(0);
1584 if (payload->is_location() &&
1585 payload->as_LocationValue()->location().type() == Location::vector) {
1586 #ifndef PRODUCT
1587 if (PrintDeoptimizationDetails) {
1588 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1589 if (Verbose) {
1590 Handle obj = sv->value();
1591 k->oop_print_on(obj(), tty);
1592 }
1593 }
1594 #endif // !PRODUCT
1595 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1596 }
1597 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1598 // which could be restored after vector object allocation.
1599 }
1600 if (k->is_instance_klass()) {
1601 InstanceKlass* ik = InstanceKlass::cast(k);
1602 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1603 } else if (k->is_typeArray_klass()) {
1604 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1605 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1606 } else if (k->is_objArray_klass()) {
1607 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1608 }
1609 }
1610 // These objects may escape when we return to Interpreter after deoptimization.
1611 // We need barrier so that stores that initialize these objects can't be reordered
1612 // with subsequent stores that make these objects accessible by other threads.
1613 OrderAccess::storestore();
1614 }
1615
1616
1617 // relock objects for which synchronization was eliminated
1618 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1619 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1620 bool relocked_objects = false;
1621 for (int i = 0; i < monitors->length(); i++) {
1622 MonitorInfo* mon_info = monitors->at(i);
1623 if (mon_info->eliminated()) {
1624 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1625 relocked_objects = true;
1626 if (!mon_info->owner_is_scalar_replaced()) {
1627 Handle obj(thread, mon_info->owner());
1628 markWord mark = obj->mark();
1629 if (exec_mode == Unpack_none) {
1630 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1631 // With exec_mode == Unpack_none obj may be thread local and locked in
1632 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1633 markWord dmw = mark.displaced_mark_helper();
1634 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1635 obj->set_mark(dmw);
1636 }
1637 if (mark.has_monitor()) {
1638 // defer relocking if the deoptee thread is currently waiting for obj
1639 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1640 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1641 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1642 if (LockingMode == LM_LEGACY) {
1643 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1644 } else if (UseObjectMonitorTable) {
1645 mon_info->lock()->clear_object_monitor_cache();
1646 }
1647 #ifdef ASSERT
1648 else {
1649 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1650 mon_info->lock()->set_bad_metadata_deopt();
1651 }
1652 #endif
1653 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1654 continue;
1655 }
1656 }
1657 }
1658 BasicLock* lock = mon_info->lock();
1659 if (LockingMode == LM_LIGHTWEIGHT) {
1660 // We have lost information about the correct state of the lock stack.
1661 // Entering may create an invalid lock stack. Inflate the lock if it
1662 // was fast_locked to restore the valid lock stack.
1663 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1664 if (deoptee_thread->lock_stack().contains(obj())) {
1665 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1666 deoptee_thread, thread);
1667 }
1668 assert(mon_info->owner()->is_locked(), "object must be locked now");
1669 assert(obj->mark().has_monitor(), "must be");
1670 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1671 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->owner() == deoptee_thread, "must be");
1672 } else {
1673 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1674 assert(mon_info->owner()->is_locked(), "object must be locked now");
1675 }
1676 }
1677 }
1678 }
1679 return relocked_objects;
1680 }
1681 #endif // COMPILER2_OR_JVMCI
1682
1683 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1684 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1685
1686 // Register map for next frame (used for stack crawl). We capture
1687 // the state of the deopt'ing frame's caller. Thus if we need to
1688 // stuff a C2I adapter we can properly fill in the callee-save
1689 // register locations.
1690 frame caller = fr.sender(reg_map);
1691 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1692
1693 frame sender = caller;
1694
1695 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1696 // the vframeArray containing the unpacking information is allocated in the C heap.
1697 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1698 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1699
1700 // Compare the vframeArray to the collected vframes
1701 assert(array->structural_compare(thread, chunk), "just checking");
1702
1703 if (TraceDeoptimization) {
1704 ResourceMark rm;
1705 stringStream st;
1706 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1707 st.print(" ");
1708 fr.print_on(&st);
1709 st.print_cr(" Virtual frames (innermost/newest first):");
1710 for (int index = 0; index < chunk->length(); index++) {
1711 compiledVFrame* vf = chunk->at(index);
1712 int bci = vf->raw_bci();
1713 const char* code_name;
1714 if (bci == SynchronizationEntryBCI) {
1715 code_name = "sync entry";
1716 } else {
1717 Bytecodes::Code code = vf->method()->code_at(bci);
1718 code_name = Bytecodes::name(code);
1719 }
1720
1721 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1722 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1723 st.print(" - %s", code_name);
1724 st.print_cr(" @ bci=%d ", bci);
1725 }
1726 tty->print_raw(st.freeze());
1727 tty->cr();
1728 }
1729
1730 return array;
1731 }
1732
1733 #if COMPILER2_OR_JVMCI
1734 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1735 // Reallocation of some scalar replaced objects failed. Record
1736 // that we need to pop all the interpreter frames for the
1737 // deoptimized compiled frame.
1738 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1739 thread->set_frames_to_pop_failed_realloc(array->frames());
1740 // Unlock all monitors here otherwise the interpreter will see a
1741 // mix of locked and unlocked monitors (because of failed
1742 // reallocations of synchronized objects) and be confused.
1743 for (int i = 0; i < array->frames(); i++) {
1744 MonitorChunk* monitors = array->element(i)->monitors();
1745 if (monitors != nullptr) {
1746 // Unlock in reverse order starting from most nested monitor.
1747 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1748 BasicObjectLock* src = monitors->at(j);
1749 if (src->obj() != nullptr) {
1750 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1751 }
1752 }
1753 array->element(i)->free_monitors();
1754 #ifdef ASSERT
1755 array->element(i)->set_removed_monitors();
1756 #endif
1757 }
1758 }
1759 }
1760 #endif
1761
1762 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1763 assert(fr.can_be_deoptimized(), "checking frame type");
1764
1765 nmethod* nm = fr.cb()->as_nmethod_or_null();
1766 assert(nm != nullptr, "only compiled methods can deopt");
1767 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none);
1768 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc());
1769 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry
1770 : cur_sd->method()->java_code_at(cur_sd->bci()));
1771 gather_statistics(nm, reason, action, bc);
1772
1773 if (LogCompilation && xtty != nullptr) {
1774 ttyLocker ttyl;
1775 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1776 nm->log_identity(xtty);
1777 xtty->end_head();
1778 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1779 xtty->begin_elem("jvms bci='%d'", sd->bci());
1780 xtty->method(sd->method());
1781 xtty->end_elem();
1782 if (sd->is_top()) break;
1783 }
1784 xtty->tail("deoptimized");
1785 }
1786
1787 Continuation::notify_deopt(thread, fr.sp());
1788
1789 // Patch the compiled method so that when execution returns to it we will
1790 // deopt the execution state and return to the interpreter.
1791 fr.deoptimize(thread);
1792 }
1793
1794 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1795 // Deoptimize only if the frame comes from compile code.
1796 // Do not deoptimize the frame which is already patched
1797 // during the execution of the loops below.
1798 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1799 return;
1800 }
1801 ResourceMark rm;
1802 deoptimize_single_frame(thread, fr, reason);
1803 }
1804
1805 #if INCLUDE_JVMCI
1806 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1807 // there is no exception handler for this pc => deoptimize
1808 nm->make_not_entrant();
1809
1810 // Use Deoptimization::deoptimize for all of its side-effects:
1811 // gathering traps statistics, logging...
1812 // it also patches the return pc but we do not care about that
1813 // since we return a continuation to the deopt_blob below.
1814 JavaThread* thread = JavaThread::current();
1815 RegisterMap reg_map(thread,
1816 RegisterMap::UpdateMap::skip,
1817 RegisterMap::ProcessFrames::include,
1818 RegisterMap::WalkContinuation::skip);
1819 frame runtime_frame = thread->last_frame();
1820 frame caller_frame = runtime_frame.sender(®_map);
1821 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1822 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1823 compiledVFrame* cvf = compiledVFrame::cast(vf);
1824 ScopeDesc* imm_scope = cvf->scope();
1825 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1826 if (imm_mdo != nullptr) {
1827 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1828 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1829
1830 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1831 if (pdata != nullptr && pdata->is_BitData()) {
1832 BitData* bit_data = (BitData*) pdata;
1833 bit_data->set_exception_seen();
1834 }
1835 }
1836
1837 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1838
1839 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1840 if (trap_mdo != nullptr) {
1841 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1842 }
1843
1844 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1845 }
1846 #endif
1847
1848 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1849 assert(thread == Thread::current() ||
1850 thread->is_handshake_safe_for(Thread::current()) ||
1851 SafepointSynchronize::is_at_safepoint(),
1852 "can only deoptimize other thread at a safepoint/handshake");
1853 // Compute frame and register map based on thread and sp.
1854 RegisterMap reg_map(thread,
1855 RegisterMap::UpdateMap::skip,
1856 RegisterMap::ProcessFrames::include,
1857 RegisterMap::WalkContinuation::skip);
1858 frame fr = thread->last_frame();
1859 while (fr.id() != id) {
1860 fr = fr.sender(®_map);
1861 }
1862 deoptimize(thread, fr, reason);
1863 }
1864
1865
1866 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1867 Thread* current = Thread::current();
1868 if (thread == current || thread->is_handshake_safe_for(current)) {
1869 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1870 } else {
1871 VM_DeoptimizeFrame deopt(thread, id, reason);
1872 VMThread::execute(&deopt);
1873 }
1874 }
1875
1876 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1877 deoptimize_frame(thread, id, Reason_constraint);
1878 }
1879
1880 // JVMTI PopFrame support
1881 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1882 {
1883 assert(thread == JavaThread::current(), "pre-condition");
1884 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1885 }
1886 JRT_END
1887
1888 MethodData*
1889 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1890 bool create_if_missing) {
1891 JavaThread* THREAD = thread; // For exception macros.
1892 MethodData* mdo = m()->method_data();
1893 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1894 // Build an MDO. Ignore errors like OutOfMemory;
1895 // that simply means we won't have an MDO to update.
1896 Method::build_profiling_method_data(m, THREAD);
1897 if (HAS_PENDING_EXCEPTION) {
1898 // Only metaspace OOM is expected. No Java code executed.
1899 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1900 CLEAR_PENDING_EXCEPTION;
1901 }
1902 mdo = m()->method_data();
1903 }
1904 return mdo;
1905 }
1906
1907 #if COMPILER2_OR_JVMCI
1908 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1909 // In case of an unresolved klass entry, load the class.
1910 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1911 // and probably nowhere else.
1912 // Even that case would benefit from simply re-interpreting the
1913 // bytecode, without paying special attention to the class index.
1914 // So this whole "class index" feature should probably be removed.
1915
1916 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1917 Klass* tk = constant_pool->klass_at(index, THREAD);
1918 if (HAS_PENDING_EXCEPTION) {
1919 // Exception happened during classloading. We ignore the exception here, since it
1920 // is going to be rethrown since the current activation is going to be deoptimized and
1921 // the interpreter will re-execute the bytecode.
1922 // Do not clear probable Async Exceptions.
1923 CLEAR_PENDING_NONASYNC_EXCEPTION;
1924 // Class loading called java code which may have caused a stack
1925 // overflow. If the exception was thrown right before the return
1926 // to the runtime the stack is no longer guarded. Reguard the
1927 // stack otherwise if we return to the uncommon trap blob and the
1928 // stack bang causes a stack overflow we crash.
1929 JavaThread* jt = THREAD;
1930 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1931 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1932 }
1933 return;
1934 }
1935
1936 assert(!constant_pool->tag_at(index).is_symbol(),
1937 "no symbolic names here, please");
1938 }
1939
1940 #if INCLUDE_JFR
1941
1942 class DeoptReasonSerializer : public JfrSerializer {
1943 public:
1944 void serialize(JfrCheckpointWriter& writer) {
1945 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1946 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1947 writer.write_key((u8)i);
1948 writer.write(Deoptimization::trap_reason_name(i));
1949 }
1950 }
1951 };
1952
1953 class DeoptActionSerializer : public JfrSerializer {
1954 public:
1955 void serialize(JfrCheckpointWriter& writer) {
1956 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1957 writer.write_count(nof_actions);
1958 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1959 writer.write_key(i);
1960 writer.write(Deoptimization::trap_action_name((int)i));
1961 }
1962 }
1963 };
1964
1965 static void register_serializers() {
1966 static int critical_section = 0;
1967 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1968 return;
1969 }
1970 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1971 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1972 }
1973
1974 static void post_deoptimization_event(nmethod* nm,
1975 const Method* method,
1976 int trap_bci,
1977 int instruction,
1978 Deoptimization::DeoptReason reason,
1979 Deoptimization::DeoptAction action) {
1980 assert(nm != nullptr, "invariant");
1981 assert(method != nullptr, "invariant");
1982 if (EventDeoptimization::is_enabled()) {
1983 static bool serializers_registered = false;
1984 if (!serializers_registered) {
1985 register_serializers();
1986 serializers_registered = true;
1987 }
1988 EventDeoptimization event;
1989 event.set_compileId(nm->compile_id());
1990 event.set_compiler(nm->compiler_type());
1991 event.set_method(method);
1992 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1993 event.set_bci(trap_bci);
1994 event.set_instruction(instruction);
1995 event.set_reason(reason);
1996 event.set_action(action);
1997 event.commit();
1998 }
1999 }
2000
2001 #endif // INCLUDE_JFR
2002
2003 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2004 const char* reason_name, const char* reason_action) {
2005 LogTarget(Debug, deoptimization) lt;
2006 if (lt.is_enabled()) {
2007 LogStream ls(lt);
2008 bool is_osr = nm->is_osr_method();
2009 ls.print("cid=%4d %s level=%d",
2010 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2011 ls.print(" %s", tm->name_and_sig_as_C_string());
2012 ls.print(" trap_bci=%d ", trap_bci);
2013 if (is_osr) {
2014 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2015 }
2016 ls.print("%s ", reason_name);
2017 ls.print("%s ", reason_action);
2018 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2019 pc, fr.pc() - nm->code_begin());
2020 }
2021 }
2022
2023 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2024 HandleMark hm(current);
2025
2026 // uncommon_trap() is called at the beginning of the uncommon trap
2027 // handler. Note this fact before we start generating temporary frames
2028 // that can confuse an asynchronous stack walker. This counter is
2029 // decremented at the end of unpack_frames().
2030
2031 current->inc_in_deopt_handler();
2032
2033 #if INCLUDE_JVMCI
2034 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2035 RegisterMap reg_map(current,
2036 RegisterMap::UpdateMap::include,
2037 RegisterMap::ProcessFrames::include,
2038 RegisterMap::WalkContinuation::skip);
2039 #else
2040 RegisterMap reg_map(current,
2041 RegisterMap::UpdateMap::skip,
2042 RegisterMap::ProcessFrames::include,
2043 RegisterMap::WalkContinuation::skip);
2044 #endif
2045 frame stub_frame = current->last_frame();
2046 frame fr = stub_frame.sender(®_map);
2047
2048 // Log a message
2049 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2050 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2051
2052 {
2053 ResourceMark rm;
2054
2055 DeoptReason reason = trap_request_reason(trap_request);
2056 DeoptAction action = trap_request_action(trap_request);
2057 #if INCLUDE_JVMCI
2058 int debug_id = trap_request_debug_id(trap_request);
2059 #endif
2060 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2061
2062 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2063 compiledVFrame* cvf = compiledVFrame::cast(vf);
2064
2065 nmethod* nm = cvf->code();
2066
2067 ScopeDesc* trap_scope = cvf->scope();
2068
2069 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2070
2071 if (is_receiver_constraint_failure) {
2072 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2073 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2074 JVMCI_ONLY(COMMA debug_id));
2075 }
2076
2077 methodHandle trap_method(current, trap_scope->method());
2078 int trap_bci = trap_scope->bci();
2079 #if INCLUDE_JVMCI
2080 jlong speculation = current->pending_failed_speculation();
2081 if (nm->is_compiled_by_jvmci()) {
2082 nm->update_speculation(current);
2083 } else {
2084 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2085 }
2086
2087 if (trap_bci == SynchronizationEntryBCI) {
2088 trap_bci = 0;
2089 current->set_pending_monitorenter(true);
2090 }
2091
2092 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2093 current->set_pending_transfer_to_interpreter(true);
2094 }
2095 #endif
2096
2097 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2098 // Record this event in the histogram.
2099 gather_statistics(nm, reason, action, trap_bc);
2100
2101 // Ensure that we can record deopt. history:
2102 bool create_if_missing = ProfileTraps;
2103
2104 methodHandle profiled_method;
2105 #if INCLUDE_JVMCI
2106 if (nm->is_compiled_by_jvmci()) {
2107 profiled_method = methodHandle(current, nm->method());
2108 } else {
2109 profiled_method = trap_method;
2110 }
2111 #else
2112 profiled_method = trap_method;
2113 #endif
2114
2115 MethodData* trap_mdo =
2116 get_method_data(current, profiled_method, create_if_missing);
2117
2118 { // Log Deoptimization event for JFR, UL and event system
2119 Method* tm = trap_method();
2120 const char* reason_name = trap_reason_name(reason);
2121 const char* reason_action = trap_action_name(action);
2122 intptr_t pc = p2i(fr.pc());
2123
2124 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2125 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2126 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2127 reason_name, reason_action, pc,
2128 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2129 }
2130
2131 // Print a bunch of diagnostics, if requested.
2132 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2133 ResourceMark rm;
2134
2135 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2136 // We must do this already now, since we cannot acquire this lock while
2137 // holding the tty lock (lock ordering by rank).
2138 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2139
2140 ttyLocker ttyl;
2141
2142 char buf[100];
2143 if (xtty != nullptr) {
2144 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2145 os::current_thread_id(),
2146 format_trap_request(buf, sizeof(buf), trap_request));
2147 #if INCLUDE_JVMCI
2148 if (speculation != 0) {
2149 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2150 }
2151 #endif
2152 nm->log_identity(xtty);
2153 }
2154 Symbol* class_name = nullptr;
2155 bool unresolved = false;
2156 if (unloaded_class_index >= 0) {
2157 constantPoolHandle constants (current, trap_method->constants());
2158 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2159 class_name = constants->klass_name_at(unloaded_class_index);
2160 unresolved = true;
2161 if (xtty != nullptr)
2162 xtty->print(" unresolved='1'");
2163 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2164 class_name = constants->symbol_at(unloaded_class_index);
2165 }
2166 if (xtty != nullptr)
2167 xtty->name(class_name);
2168 }
2169 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2170 // Dump the relevant MDO state.
2171 // This is the deopt count for the current reason, any previous
2172 // reasons or recompiles seen at this point.
2173 int dcnt = trap_mdo->trap_count(reason);
2174 if (dcnt != 0)
2175 xtty->print(" count='%d'", dcnt);
2176
2177 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2178 // lock extra_data_lock before the tty lock.
2179 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2180 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2181 if (dos != 0) {
2182 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2183 if (trap_state_is_recompiled(dos)) {
2184 int recnt2 = trap_mdo->overflow_recompile_count();
2185 if (recnt2 != 0)
2186 xtty->print(" recompiles2='%d'", recnt2);
2187 }
2188 }
2189 }
2190 if (xtty != nullptr) {
2191 xtty->stamp();
2192 xtty->end_head();
2193 }
2194 if (TraceDeoptimization) { // make noise on the tty
2195 stringStream st;
2196 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2197 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2198 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2199 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2200 #if INCLUDE_JVMCI
2201 if (nm->is_compiled_by_jvmci()) {
2202 const char* installed_code_name = nm->jvmci_name();
2203 if (installed_code_name != nullptr) {
2204 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2205 }
2206 }
2207 #endif
2208 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2209 p2i(fr.pc()),
2210 os::current_thread_id(),
2211 trap_reason_name(reason),
2212 trap_action_name(action),
2213 unloaded_class_index
2214 #if INCLUDE_JVMCI
2215 , debug_id
2216 #endif
2217 );
2218 if (class_name != nullptr) {
2219 st.print(unresolved ? " unresolved class: " : " symbol: ");
2220 class_name->print_symbol_on(&st);
2221 }
2222 st.cr();
2223 tty->print_raw(st.freeze());
2224 }
2225 if (xtty != nullptr) {
2226 // Log the precise location of the trap.
2227 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2228 xtty->begin_elem("jvms bci='%d'", sd->bci());
2229 xtty->method(sd->method());
2230 xtty->end_elem();
2231 if (sd->is_top()) break;
2232 }
2233 xtty->tail("uncommon_trap");
2234 }
2235 }
2236 // (End diagnostic printout.)
2237
2238 if (is_receiver_constraint_failure) {
2239 fatal("missing receiver type check");
2240 }
2241
2242 // Load class if necessary
2243 if (unloaded_class_index >= 0) {
2244 constantPoolHandle constants(current, trap_method->constants());
2245 load_class_by_index(constants, unloaded_class_index, THREAD);
2246 }
2247
2248 // Flush the nmethod if necessary and desirable.
2249 //
2250 // We need to avoid situations where we are re-flushing the nmethod
2251 // because of a hot deoptimization site. Repeated flushes at the same
2252 // point need to be detected by the compiler and avoided. If the compiler
2253 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2254 // module must take measures to avoid an infinite cycle of recompilation
2255 // and deoptimization. There are several such measures:
2256 //
2257 // 1. If a recompilation is ordered a second time at some site X
2258 // and for the same reason R, the action is adjusted to 'reinterpret',
2259 // to give the interpreter time to exercise the method more thoroughly.
2260 // If this happens, the method's overflow_recompile_count is incremented.
2261 //
2262 // 2. If the compiler fails to reduce the deoptimization rate, then
2263 // the method's overflow_recompile_count will begin to exceed the set
2264 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2265 // is adjusted to 'make_not_compilable', and the method is abandoned
2266 // to the interpreter. This is a performance hit for hot methods,
2267 // but is better than a disastrous infinite cycle of recompilations.
2268 // (Actually, only the method containing the site X is abandoned.)
2269 //
2270 // 3. In parallel with the previous measures, if the total number of
2271 // recompilations of a method exceeds the much larger set limit
2272 // PerMethodRecompilationCutoff, the method is abandoned.
2273 // This should only happen if the method is very large and has
2274 // many "lukewarm" deoptimizations. The code which enforces this
2275 // limit is elsewhere (class nmethod, class Method).
2276 //
2277 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2278 // to recompile at each bytecode independently of the per-BCI cutoff.
2279 //
2280 // The decision to update code is up to the compiler, and is encoded
2281 // in the Action_xxx code. If the compiler requests Action_none
2282 // no trap state is changed, no compiled code is changed, and the
2283 // computation suffers along in the interpreter.
2284 //
2285 // The other action codes specify various tactics for decompilation
2286 // and recompilation. Action_maybe_recompile is the loosest, and
2287 // allows the compiled code to stay around until enough traps are seen,
2288 // and until the compiler gets around to recompiling the trapping method.
2289 //
2290 // The other actions cause immediate removal of the present code.
2291
2292 // Traps caused by injected profile shouldn't pollute trap counts.
2293 bool injected_profile_trap = trap_method->has_injected_profile() &&
2294 (reason == Reason_intrinsic || reason == Reason_unreached);
2295
2296 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2297 bool make_not_entrant = false;
2298 bool make_not_compilable = false;
2299 bool reprofile = false;
2300 switch (action) {
2301 case Action_none:
2302 // Keep the old code.
2303 update_trap_state = false;
2304 break;
2305 case Action_maybe_recompile:
2306 // Do not need to invalidate the present code, but we can
2307 // initiate another
2308 // Start compiler without (necessarily) invalidating the nmethod.
2309 // The system will tolerate the old code, but new code should be
2310 // generated when possible.
2311 break;
2312 case Action_reinterpret:
2313 // Go back into the interpreter for a while, and then consider
2314 // recompiling form scratch.
2315 make_not_entrant = true;
2316 // Reset invocation counter for outer most method.
2317 // This will allow the interpreter to exercise the bytecodes
2318 // for a while before recompiling.
2319 // By contrast, Action_make_not_entrant is immediate.
2320 //
2321 // Note that the compiler will track null_check, null_assert,
2322 // range_check, and class_check events and log them as if they
2323 // had been traps taken from compiled code. This will update
2324 // the MDO trap history so that the next compilation will
2325 // properly detect hot trap sites.
2326 reprofile = true;
2327 break;
2328 case Action_make_not_entrant:
2329 // Request immediate recompilation, and get rid of the old code.
2330 // Make them not entrant, so next time they are called they get
2331 // recompiled. Unloaded classes are loaded now so recompile before next
2332 // time they are called. Same for uninitialized. The interpreter will
2333 // link the missing class, if any.
2334 make_not_entrant = true;
2335 break;
2336 case Action_make_not_compilable:
2337 // Give up on compiling this method at all.
2338 make_not_entrant = true;
2339 make_not_compilable = true;
2340 break;
2341 default:
2342 ShouldNotReachHere();
2343 }
2344
2345 // Setting +ProfileTraps fixes the following, on all platforms:
2346 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2347 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2348 // recompile relies on a MethodData* to record heroic opt failures.
2349
2350 // Whether the interpreter is producing MDO data or not, we also need
2351 // to use the MDO to detect hot deoptimization points and control
2352 // aggressive optimization.
2353 bool inc_recompile_count = false;
2354
2355 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2356 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2357 (trap_mdo != nullptr),
2358 Mutex::_no_safepoint_check_flag);
2359 ProfileData* pdata = nullptr;
2360 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2361 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2362 uint this_trap_count = 0;
2363 bool maybe_prior_trap = false;
2364 bool maybe_prior_recompile = false;
2365
2366 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2367 #if INCLUDE_JVMCI
2368 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2369 #endif
2370 nm->method(),
2371 //outputs:
2372 this_trap_count,
2373 maybe_prior_trap,
2374 maybe_prior_recompile);
2375 // Because the interpreter also counts null, div0, range, and class
2376 // checks, these traps from compiled code are double-counted.
2377 // This is harmless; it just means that the PerXTrapLimit values
2378 // are in effect a little smaller than they look.
2379
2380 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2381 if (per_bc_reason != Reason_none) {
2382 // Now take action based on the partially known per-BCI history.
2383 if (maybe_prior_trap
2384 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2385 // If there are too many traps at this BCI, force a recompile.
2386 // This will allow the compiler to see the limit overflow, and
2387 // take corrective action, if possible. The compiler generally
2388 // does not use the exact PerBytecodeTrapLimit value, but instead
2389 // changes its tactics if it sees any traps at all. This provides
2390 // a little hysteresis, delaying a recompile until a trap happens
2391 // several times.
2392 //
2393 // Actually, since there is only one bit of counter per BCI,
2394 // the possible per-BCI counts are {0,1,(per-method count)}.
2395 // This produces accurate results if in fact there is only
2396 // one hot trap site, but begins to get fuzzy if there are
2397 // many sites. For example, if there are ten sites each
2398 // trapping two or more times, they each get the blame for
2399 // all of their traps.
2400 make_not_entrant = true;
2401 }
2402
2403 // Detect repeated recompilation at the same BCI, and enforce a limit.
2404 if (make_not_entrant && maybe_prior_recompile) {
2405 // More than one recompile at this point.
2406 inc_recompile_count = maybe_prior_trap;
2407 }
2408 } else {
2409 // For reasons which are not recorded per-bytecode, we simply
2410 // force recompiles unconditionally.
2411 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2412 make_not_entrant = true;
2413 }
2414
2415 // Go back to the compiler if there are too many traps in this method.
2416 if (this_trap_count >= per_method_trap_limit(reason)) {
2417 // If there are too many traps in this method, force a recompile.
2418 // This will allow the compiler to see the limit overflow, and
2419 // take corrective action, if possible.
2420 // (This condition is an unlikely backstop only, because the
2421 // PerBytecodeTrapLimit is more likely to take effect first,
2422 // if it is applicable.)
2423 make_not_entrant = true;
2424 }
2425
2426 // Here's more hysteresis: If there has been a recompile at
2427 // this trap point already, run the method in the interpreter
2428 // for a while to exercise it more thoroughly.
2429 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2430 reprofile = true;
2431 }
2432 }
2433
2434 // Take requested actions on the method:
2435
2436 // Recompile
2437 if (make_not_entrant) {
2438 if (!nm->make_not_entrant()) {
2439 return; // the call did not change nmethod's state
2440 }
2441
2442 if (pdata != nullptr) {
2443 // Record the recompilation event, if any.
2444 int tstate0 = pdata->trap_state();
2445 int tstate1 = trap_state_set_recompiled(tstate0, true);
2446 if (tstate1 != tstate0)
2447 pdata->set_trap_state(tstate1);
2448 }
2449
2450 // For code aging we count traps separately here, using make_not_entrant()
2451 // as a guard against simultaneous deopts in multiple threads.
2452 if (reason == Reason_tenured && trap_mdo != nullptr) {
2453 trap_mdo->inc_tenure_traps();
2454 }
2455 }
2456
2457 if (inc_recompile_count) {
2458 trap_mdo->inc_overflow_recompile_count();
2459 if ((uint)trap_mdo->overflow_recompile_count() >
2460 (uint)PerBytecodeRecompilationCutoff) {
2461 // Give up on the method containing the bad BCI.
2462 if (trap_method() == nm->method()) {
2463 make_not_compilable = true;
2464 } else {
2465 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2466 // But give grace to the enclosing nm->method().
2467 }
2468 }
2469 }
2470
2471 // Reprofile
2472 if (reprofile) {
2473 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2474 }
2475
2476 // Give up compiling
2477 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2478 assert(make_not_entrant, "consistent");
2479 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2480 }
2481
2482 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2483 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2484 if (exception_handler_data != nullptr) {
2485 // uncommon trap at the start of an exception handler.
2486 // C2 generates these for un-entered exception handlers.
2487 // mark the handler as entered to avoid generating
2488 // another uncommon trap the next time the handler is compiled
2489 exception_handler_data->set_exception_handler_entered();
2490 }
2491 }
2492
2493 } // Free marked resources
2494
2495 }
2496 JRT_END
2497
2498 ProfileData*
2499 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2500 int trap_bci,
2501 Deoptimization::DeoptReason reason,
2502 bool update_total_trap_count,
2503 #if INCLUDE_JVMCI
2504 bool is_osr,
2505 #endif
2506 Method* compiled_method,
2507 //outputs:
2508 uint& ret_this_trap_count,
2509 bool& ret_maybe_prior_trap,
2510 bool& ret_maybe_prior_recompile) {
2511 trap_mdo->check_extra_data_locked();
2512
2513 bool maybe_prior_trap = false;
2514 bool maybe_prior_recompile = false;
2515 uint this_trap_count = 0;
2516 if (update_total_trap_count) {
2517 uint idx = reason;
2518 #if INCLUDE_JVMCI
2519 if (is_osr) {
2520 // Upper half of history array used for traps in OSR compilations
2521 idx += Reason_TRAP_HISTORY_LENGTH;
2522 }
2523 #endif
2524 uint prior_trap_count = trap_mdo->trap_count(idx);
2525 this_trap_count = trap_mdo->inc_trap_count(idx);
2526
2527 // If the runtime cannot find a place to store trap history,
2528 // it is estimated based on the general condition of the method.
2529 // If the method has ever been recompiled, or has ever incurred
2530 // a trap with the present reason , then this BCI is assumed
2531 // (pessimistically) to be the culprit.
2532 maybe_prior_trap = (prior_trap_count != 0);
2533 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2534 }
2535 ProfileData* pdata = nullptr;
2536
2537
2538 // For reasons which are recorded per bytecode, we check per-BCI data.
2539 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2540 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2541 if (per_bc_reason != Reason_none) {
2542 // Find the profile data for this BCI. If there isn't one,
2543 // try to allocate one from the MDO's set of spares.
2544 // This will let us detect a repeated trap at this point.
2545 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2546
2547 if (pdata != nullptr) {
2548 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2549 if (LogCompilation && xtty != nullptr) {
2550 ttyLocker ttyl;
2551 // no more room for speculative traps in this MDO
2552 xtty->elem("speculative_traps_oom");
2553 }
2554 }
2555 // Query the trap state of this profile datum.
2556 int tstate0 = pdata->trap_state();
2557 if (!trap_state_has_reason(tstate0, per_bc_reason))
2558 maybe_prior_trap = false;
2559 if (!trap_state_is_recompiled(tstate0))
2560 maybe_prior_recompile = false;
2561
2562 // Update the trap state of this profile datum.
2563 int tstate1 = tstate0;
2564 // Record the reason.
2565 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2566 // Store the updated state on the MDO, for next time.
2567 if (tstate1 != tstate0)
2568 pdata->set_trap_state(tstate1);
2569 } else {
2570 if (LogCompilation && xtty != nullptr) {
2571 ttyLocker ttyl;
2572 // Missing MDP? Leave a small complaint in the log.
2573 xtty->elem("missing_mdp bci='%d'", trap_bci);
2574 }
2575 }
2576 }
2577
2578 // Return results:
2579 ret_this_trap_count = this_trap_count;
2580 ret_maybe_prior_trap = maybe_prior_trap;
2581 ret_maybe_prior_recompile = maybe_prior_recompile;
2582 return pdata;
2583 }
2584
2585 void
2586 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2587 ResourceMark rm;
2588 // Ignored outputs:
2589 uint ignore_this_trap_count;
2590 bool ignore_maybe_prior_trap;
2591 bool ignore_maybe_prior_recompile;
2592 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2593 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2594 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2595
2596 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2597 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2598
2599 query_update_method_data(trap_mdo, trap_bci,
2600 (DeoptReason)reason,
2601 update_total_counts,
2602 #if INCLUDE_JVMCI
2603 false,
2604 #endif
2605 nullptr,
2606 ignore_this_trap_count,
2607 ignore_maybe_prior_trap,
2608 ignore_maybe_prior_recompile);
2609 }
2610
2611 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode))
2612 // Enable WXWrite: current function is called from methods compiled by C2 directly
2613 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2614
2615 // Still in Java no safepoints
2616 {
2617 // This enters VM and may safepoint
2618 uncommon_trap_inner(current, trap_request);
2619 }
2620 HandleMark hm(current);
2621 return fetch_unroll_info_helper(current, exec_mode);
2622 PROF_END
2623
2624 // Local derived constants.
2625 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2626 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2627 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2628
2629 //---------------------------trap_state_reason---------------------------------
2630 Deoptimization::DeoptReason
2631 Deoptimization::trap_state_reason(int trap_state) {
2632 // This assert provides the link between the width of DataLayout::trap_bits
2633 // and the encoding of "recorded" reasons. It ensures there are enough
2634 // bits to store all needed reasons in the per-BCI MDO profile.
2635 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2636 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2637 trap_state -= recompile_bit;
2638 if (trap_state == DS_REASON_MASK) {
2639 return Reason_many;
2640 } else {
2641 assert((int)Reason_none == 0, "state=0 => Reason_none");
2642 return (DeoptReason)trap_state;
2643 }
2644 }
2645 //-------------------------trap_state_has_reason-------------------------------
2646 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2647 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2648 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2649 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2650 trap_state -= recompile_bit;
2651 if (trap_state == DS_REASON_MASK) {
2652 return -1; // true, unspecifically (bottom of state lattice)
2653 } else if (trap_state == reason) {
2654 return 1; // true, definitely
2655 } else if (trap_state == 0) {
2656 return 0; // false, definitely (top of state lattice)
2657 } else {
2658 return 0; // false, definitely
2659 }
2660 }
2661 //-------------------------trap_state_add_reason-------------------------------
2662 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2663 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2664 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2665 trap_state -= recompile_bit;
2666 if (trap_state == DS_REASON_MASK) {
2667 return trap_state + recompile_bit; // already at state lattice bottom
2668 } else if (trap_state == reason) {
2669 return trap_state + recompile_bit; // the condition is already true
2670 } else if (trap_state == 0) {
2671 return reason + recompile_bit; // no condition has yet been true
2672 } else {
2673 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2674 }
2675 }
2676 //-----------------------trap_state_is_recompiled------------------------------
2677 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2678 return (trap_state & DS_RECOMPILE_BIT) != 0;
2679 }
2680 //-----------------------trap_state_set_recompiled-----------------------------
2681 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2682 if (z) return trap_state | DS_RECOMPILE_BIT;
2683 else return trap_state & ~DS_RECOMPILE_BIT;
2684 }
2685 //---------------------------format_trap_state---------------------------------
2686 // This is used for debugging and diagnostics, including LogFile output.
2687 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2688 int trap_state) {
2689 assert(buflen > 0, "sanity");
2690 DeoptReason reason = trap_state_reason(trap_state);
2691 bool recomp_flag = trap_state_is_recompiled(trap_state);
2692 // Re-encode the state from its decoded components.
2693 int decoded_state = 0;
2694 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2695 decoded_state = trap_state_add_reason(decoded_state, reason);
2696 if (recomp_flag)
2697 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2698 // If the state re-encodes properly, format it symbolically.
2699 // Because this routine is used for debugging and diagnostics,
2700 // be robust even if the state is a strange value.
2701 size_t len;
2702 if (decoded_state != trap_state) {
2703 // Random buggy state that doesn't decode??
2704 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2705 } else {
2706 len = jio_snprintf(buf, buflen, "%s%s",
2707 trap_reason_name(reason),
2708 recomp_flag ? " recompiled" : "");
2709 }
2710 return buf;
2711 }
2712
2713
2714 //--------------------------------statics--------------------------------------
2715 const char* Deoptimization::_trap_reason_name[] = {
2716 // Note: Keep this in sync. with enum DeoptReason.
2717 "none",
2718 "null_check",
2719 "null_assert" JVMCI_ONLY("_or_unreached0"),
2720 "range_check",
2721 "class_check",
2722 "array_check",
2723 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2724 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2725 "profile_predicate",
2726 "unloaded",
2727 "uninitialized",
2728 "initialized",
2729 "unreached",
2730 "unhandled",
2731 "constraint",
2732 "div0_check",
2733 "age",
2734 "predicate",
2735 "loop_limit_check",
2736 "speculate_class_check",
2737 "speculate_null_check",
2738 "speculate_null_assert",
2739 "unstable_if",
2740 "unstable_fused_if",
2741 "receiver_constraint",
2742 #if INCLUDE_JVMCI
2743 "aliasing",
2744 "transfer_to_interpreter",
2745 "not_compiled_exception_handler",
2746 "unresolved",
2747 "jsr_mismatch",
2748 #endif
2749 "tenured"
2750 };
2751 const char* Deoptimization::_trap_action_name[] = {
2752 // Note: Keep this in sync. with enum DeoptAction.
2753 "none",
2754 "maybe_recompile",
2755 "reinterpret",
2756 "make_not_entrant",
2757 "make_not_compilable"
2758 };
2759
2760 const char* Deoptimization::trap_reason_name(int reason) {
2761 // Check that every reason has a name
2762 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2763
2764 if (reason == Reason_many) return "many";
2765 if ((uint)reason < Reason_LIMIT)
2766 return _trap_reason_name[reason];
2767 static char buf[20];
2768 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2769 return buf;
2770 }
2771 const char* Deoptimization::trap_action_name(int action) {
2772 // Check that every action has a name
2773 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2774
2775 if ((uint)action < Action_LIMIT)
2776 return _trap_action_name[action];
2777 static char buf[20];
2778 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2779 return buf;
2780 }
2781
2782 // This is used for debugging and diagnostics, including LogFile output.
2783 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2784 int trap_request) {
2785 jint unloaded_class_index = trap_request_index(trap_request);
2786 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2787 const char* action = trap_action_name(trap_request_action(trap_request));
2788 #if INCLUDE_JVMCI
2789 int debug_id = trap_request_debug_id(trap_request);
2790 #endif
2791 size_t len;
2792 if (unloaded_class_index < 0) {
2793 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2794 reason, action
2795 #if INCLUDE_JVMCI
2796 ,debug_id
2797 #endif
2798 );
2799 } else {
2800 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2801 reason, action, unloaded_class_index
2802 #if INCLUDE_JVMCI
2803 ,debug_id
2804 #endif
2805 );
2806 }
2807 return buf;
2808 }
2809
2810 juint Deoptimization::_deoptimization_hist
2811 [1 + 4 + 5] // total + online + archived
2812 [Deoptimization::Reason_LIMIT]
2813 [1 + Deoptimization::Action_LIMIT]
2814 [Deoptimization::BC_CASE_LIMIT]
2815 = {0};
2816
2817 enum {
2818 LSB_BITS = 8,
2819 LSB_MASK = right_n_bits(LSB_BITS)
2820 };
2821
2822 static void update(juint* cases, Bytecodes::Code bc) {
2823 juint* bc_counter_addr = nullptr;
2824 juint bc_counter = 0;
2825 // Look for an unused counter, or an exact match to this BC.
2826 if (bc != Bytecodes::_illegal) {
2827 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) {
2828 juint* counter_addr = &cases[bc_case];
2829 juint counter = *counter_addr;
2830 if ((counter == 0 && bc_counter_addr == nullptr)
2831 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2832 // this counter is either free or is already devoted to this BC
2833 bc_counter_addr = counter_addr;
2834 bc_counter = counter | bc;
2835 }
2836 }
2837 }
2838 if (bc_counter_addr == nullptr) {
2839 // Overflow, or no given bytecode.
2840 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1];
2841 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2842 }
2843 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2844 }
2845
2846
2847 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
2848 Bytecodes::Code bc) {
2849 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2850 assert(action >= 0 && action < Action_LIMIT, "oob");
2851 _deoptimization_hist[0][Reason_none][0][0] += 1; // total
2852 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total
2853
2854 update(_deoptimization_hist[0][reason][1+action], bc);
2855
2856 uint lvl = nm->comp_level() + (nm->is_scc() ? 4 : 0) + (nm->preloaded() ? 1 : 0);
2857 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total
2858 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total
2859 update(_deoptimization_hist[lvl][reason][1+action], bc);
2860 }
2861
2862 jint Deoptimization::total_deoptimization_count() {
2863 return _deoptimization_hist[0][Reason_none][0][0];
2864 }
2865
2866 // Get the deopt count for a specific reason and a specific action. If either
2867 // one of 'reason' or 'action' is null, the method returns the sum of all
2868 // deoptimizations with the specific 'action' or 'reason' respectively.
2869 // If both arguments are null, the method returns the total deopt count.
2870 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2871 if (reason_str == nullptr && action_str == nullptr) {
2872 return total_deoptimization_count();
2873 }
2874 juint counter = 0;
2875 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2876 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2877 for (int action = 0; action < Action_LIMIT; action++) {
2878 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2879 juint* cases = _deoptimization_hist[0][reason][1+action];
2880 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2881 counter += cases[bc_case] >> LSB_BITS;
2882 }
2883 }
2884 }
2885 }
2886 }
2887 return counter;
2888 }
2889
2890 void Deoptimization::print_statistics() {
2891 ttyLocker ttyl;
2892 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2893 tty->print_cr("Deoptimization traps recorded:");
2894 print_statistics_on(tty);
2895 if (xtty != nullptr) xtty->tail("statistics");
2896 }
2897
2898 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) {
2899 juint total = _deoptimization_hist[lvl][Reason_none][0][0];
2900 juint account = total;
2901 #define PRINT_STAT_LINE(name, r) \
2902 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name);
2903 if (total > 0) {
2904 st->print(" %s: ", title);
2905 PRINT_STAT_LINE("total", total);
2906 // For each non-zero entry in the histogram, print the reason,
2907 // the action, and (if specifically known) the type of bytecode.
2908 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2909 for (int action = 0; action < Action_LIMIT; action++) {
2910 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action];
2911 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2912 juint counter = cases[bc_case];
2913 if (counter != 0) {
2914 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2915 const char* bc_name = "other";
2916 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) {
2917 // overwritten
2918 } else if (Bytecodes::is_defined(bc)) {
2919 bc_name = Bytecodes::name(bc);
2920 }
2921 juint r = counter >> LSB_BITS;
2922 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)",
2923 trap_reason_name(reason), trap_action_name(action), bc_name,
2924 r, (r * 100.0) / total);
2925 account -= r;
2926 }
2927 }
2928 }
2929 }
2930 if (account != 0) {
2931 PRINT_STAT_LINE("unaccounted", account);
2932 }
2933 #undef PRINT_STAT_LINE
2934 }
2935 }
2936
2937 void Deoptimization::print_statistics_on(outputStream* st) {
2938 // print_statistics_on("Total", 0, st);
2939 print_statistics_on("Tier1", 1, st);
2940 print_statistics_on("Tier2", 2, st);
2941 print_statistics_on("Tier3", 3, st);
2942 print_statistics_on("Tier4", 4, st);
2943
2944 print_statistics_on("SC Tier1", 5, st);
2945 print_statistics_on("SC Tier2", 6, st);
2946 print_statistics_on("SC Tier4", 8, st);
2947 print_statistics_on("SC Tier5 (preloaded)", 9, st);
2948 }
2949
2950 #define DO_COUNTERS(macro) \
2951 macro(Deoptimization, fetch_unroll_info) \
2952 macro(Deoptimization, unpack_frames) \
2953 macro(Deoptimization, uncommon_trap_inner) \
2954 macro(Deoptimization, uncommon_trap)
2955
2956 #define INIT_COUNTER(sub, name) \
2957 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \
2958 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count");
2959
2960 void Deoptimization::init_counters() {
2961 if (ProfileRuntimeCalls && UsePerfData) {
2962 EXCEPTION_MARK;
2963
2964 DO_COUNTERS(INIT_COUNTER)
2965
2966 if (HAS_PENDING_EXCEPTION) {
2967 vm_exit_during_initialization("jvm_perf_init failed unexpectedly");
2968 }
2969 }
2970 }
2971 #undef INIT_COUNTER
2972
2973 #define PRINT_COUNTER(sub, name) { \
2974 jlong count = _perf_##sub##_##name##_count->get_value(); \
2975 if (count > 0) { \
2976 st->print_cr(" %-50s = " JLONG_FORMAT_W(6) "us (elapsed) " JLONG_FORMAT_W(6) "us (thread) (" JLONG_FORMAT_W(5) " events)", #sub "::" #name, \
2977 _perf_##sub##_##name##_timer->elapsed_counter_value_us(), \
2978 _perf_##sub##_##name##_timer->thread_counter_value_us(), \
2979 count); \
2980 }}
2981
2982 void Deoptimization::print_counters_on(outputStream* st) {
2983 if (ProfileRuntimeCalls && UsePerfData) {
2984 DO_COUNTERS(PRINT_COUNTER)
2985 } else {
2986 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData"));
2987 }
2988 }
2989
2990 #undef PRINT_COUNTER
2991 #undef DO_COUNTERS
2992
2993 #else // COMPILER2_OR_JVMCI
2994
2995
2996 // Stubs for C1 only system.
2997 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2998 return false;
2999 }
3000
3001 const char* Deoptimization::trap_reason_name(int reason) {
3002 return "unknown";
3003 }
3004
3005 jint Deoptimization::total_deoptimization_count() {
3006 return 0;
3007 }
3008
3009 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
3010 return 0;
3011 }
3012
3013 void Deoptimization::print_statistics() {
3014 // no output
3015 }
3016
3017 void
3018 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3019 // no update
3020 }
3021
3022 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3023 return 0;
3024 }
3025
3026 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
3027 Bytecodes::Code bc) {
3028 // no update
3029 }
3030
3031 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3032 int trap_state) {
3033 jio_snprintf(buf, buflen, "#%d", trap_state);
3034 return buf;
3035 }
3036
3037 void Deoptimization::init_counters() {
3038 // nothing to do
3039 }
3040
3041 void Deoptimization::print_counters_on(outputStream* st) {
3042 // no output
3043 }
3044
3045 void Deoptimization::print_statistics_on(outputStream* st) {
3046 // no output
3047 }
3048
3049 #endif // COMPILER2_OR_JVMCI