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/safepointVerifiers.hpp"
84 #include "runtime/sharedRuntime.hpp"
85 #include "runtime/signature.hpp"
86 #include "runtime/stackFrameStream.inline.hpp"
87 #include "runtime/stackValue.hpp"
88 #include "runtime/stackWatermarkSet.hpp"
89 #include "runtime/stubRoutines.hpp"
90 #include "runtime/synchronizer.inline.hpp"
91 #include "runtime/threadSMR.hpp"
92 #include "runtime/threadWXSetters.inline.hpp"
93 #include "runtime/vframe.hpp"
94 #include "runtime/vframeArray.hpp"
95 #include "runtime/vframe_hp.hpp"
96 #include "runtime/vmOperations.hpp"
97 #include "utilities/checkedCast.hpp"
98 #include "utilities/events.hpp"
99 #include "utilities/growableArray.hpp"
100 #include "utilities/macros.hpp"
101 #include "utilities/preserveException.hpp"
102 #include "utilities/xmlstream.hpp"
103 #if INCLUDE_JFR
104 #include "jfr/jfrEvents.hpp"
105 #include "jfr/metadata/jfrSerializer.hpp"
106 #endif
107
108 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
109 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
110 bool DeoptimizationScope::_committing_in_progress = false;
111
112 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
113 DEBUG_ONLY(_deopted = false;)
114
115 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
116 // If there is nothing to deopt _required_gen is the same as comitted.
117 _required_gen = DeoptimizationScope::_committed_deopt_gen;
118 }
119
120 DeoptimizationScope::~DeoptimizationScope() {
121 assert(_deopted, "Deopt not executed");
122 }
123
124 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
125 if (!nm->can_be_deoptimized()) {
126 return;
127 }
128
129 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
130
131 // If it's already marked but we still need it to be deopted.
132 if (nm->is_marked_for_deoptimization()) {
133 dependent(nm);
134 return;
135 }
136
137 nmethod::DeoptimizationStatus status =
138 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
139 Atomic::store(&nm->_deoptimization_status, status);
140
141 // Make sure active is not committed
142 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
143 assert(nm->_deoptimization_generation == 0, "Is already marked");
144
145 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
146 _required_gen = DeoptimizationScope::_active_deopt_gen;
147 }
148
149 void DeoptimizationScope::dependent(nmethod* nm) {
150 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
151
152 // A method marked by someone else may have a _required_gen lower than what we marked with.
153 // Therefore only store it if it's higher than _required_gen.
154 if (_required_gen < nm->_deoptimization_generation) {
155 _required_gen = nm->_deoptimization_generation;
156 }
157 }
158
159 void DeoptimizationScope::deoptimize_marked() {
160 assert(!_deopted, "Already deopted");
161
162 // We are not alive yet.
163 if (!Universe::is_fully_initialized()) {
164 DEBUG_ONLY(_deopted = true;)
165 return;
166 }
167
168 // Safepoints are a special case, handled here.
169 if (SafepointSynchronize::is_at_safepoint()) {
170 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
171 DeoptimizationScope::_active_deopt_gen++;
172 Deoptimization::deoptimize_all_marked();
173 DEBUG_ONLY(_deopted = true;)
174 return;
175 }
176
177 uint64_t comitting = 0;
178 bool wait = false;
179 while (true) {
180 {
181 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
182
183 // First we check if we or someone else already deopted the gen we want.
184 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
185 DEBUG_ONLY(_deopted = true;)
186 return;
187 }
188 if (!_committing_in_progress) {
189 // The version we are about to commit.
190 comitting = DeoptimizationScope::_active_deopt_gen;
191 // Make sure new marks use a higher gen.
192 DeoptimizationScope::_active_deopt_gen++;
193 _committing_in_progress = true;
194 wait = false;
195 } else {
196 // Another thread is handshaking and committing a gen.
197 wait = true;
198 }
199 }
200 if (wait) {
201 // Wait and let the concurrent handshake be performed.
202 ThreadBlockInVM tbivm(JavaThread::current());
203 os::naked_yield();
204 } else {
205 // Performs the handshake.
206 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
207 DEBUG_ONLY(_deopted = true;)
208 {
209 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
210
211 // Make sure that committed doesn't go backwards.
212 // Should only happen if we did a deopt during a safepoint above.
213 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
214 DeoptimizationScope::_committed_deopt_gen = comitting;
215 }
216 _committing_in_progress = false;
217
218 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
219
220 return;
221 }
222 }
223 }
224 }
225
226 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
227 int caller_adjustment,
228 int caller_actual_parameters,
229 int number_of_frames,
230 intptr_t* frame_sizes,
231 address* frame_pcs,
232 BasicType return_type,
233 int exec_mode) {
234 _size_of_deoptimized_frame = size_of_deoptimized_frame;
235 _caller_adjustment = caller_adjustment;
236 _caller_actual_parameters = caller_actual_parameters;
237 _number_of_frames = number_of_frames;
238 _frame_sizes = frame_sizes;
239 _frame_pcs = frame_pcs;
240 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
241 _return_type = return_type;
242 _initial_info = 0;
243 // PD (x86 only)
244 _counter_temp = 0;
245 _unpack_kind = exec_mode;
246 _sender_sp_temp = 0;
247
248 _total_frame_sizes = size_of_frames();
249 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
250 }
251
252 Deoptimization::UnrollBlock::~UnrollBlock() {
253 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
254 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
255 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
256 }
257
258 int Deoptimization::UnrollBlock::size_of_frames() const {
259 // Account first for the adjustment of the initial frame
260 intptr_t result = _caller_adjustment;
261 for (int index = 0; index < number_of_frames(); index++) {
262 result += frame_sizes()[index];
263 }
264 return checked_cast<int>(result);
265 }
266
267 void Deoptimization::UnrollBlock::print() {
268 ResourceMark rm;
269 stringStream st;
270 st.print_cr("UnrollBlock");
271 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
272 st.print( " frame_sizes: ");
273 for (int index = 0; index < number_of_frames(); index++) {
274 st.print(INTX_FORMAT " ", frame_sizes()[index]);
275 }
276 st.cr();
277 tty->print_raw(st.freeze());
278 }
279
280 // In order to make fetch_unroll_info work properly with escape
281 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
282 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
283 // which is called from the method fetch_unroll_info_helper below.
284 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
285 // fetch_unroll_info() is called at the beginning of the deoptimization
286 // handler. Note this fact before we start generating temporary frames
287 // that can confuse an asynchronous stack walker. This counter is
288 // decremented at the end of unpack_frames().
289 current->inc_in_deopt_handler();
290
291 if (exec_mode == Unpack_exception) {
292 // When we get here, a callee has thrown an exception into a deoptimized
293 // frame. That throw might have deferred stack watermark checking until
294 // after unwinding. So we deal with such deferred requests here.
295 StackWatermarkSet::after_unwind(current);
296 }
297
298 return fetch_unroll_info_helper(current, exec_mode);
299 JRT_END
300
301 #if COMPILER2_OR_JVMCI
302 // print information about reallocated objects
303 static void print_objects(JavaThread* deoptee_thread,
304 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
305 ResourceMark rm;
306 stringStream st; // change to logStream with logging
307 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
308 fieldDescriptor fd;
309
310 for (int i = 0; i < objects->length(); i++) {
311 ObjectValue* sv = (ObjectValue*) objects->at(i);
312 Handle obj = sv->value();
313
314 if (obj.is_null()) {
315 st.print_cr(" nullptr");
316 continue;
317 }
318
319 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
320
321 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
322 k->print_value_on(&st);
323 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize);
324
325 if (Verbose && k != nullptr) {
326 k->oop_print_on(obj(), &st);
327 }
328 }
329 tty->print_raw(st.freeze());
330 }
331
332 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
333 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
334 bool& deoptimized_objects) {
335 bool realloc_failures = false;
336 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
337
338 JavaThread* deoptee_thread = chunk->at(0)->thread();
339 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
340 "a frame can only be deoptimized by the owner thread");
341
342 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
343
344 // The flag return_oop() indicates call sites which return oop
345 // in compiled code. Such sites include java method calls,
346 // runtime calls (for example, used to allocate new objects/arrays
347 // on slow code path) and any other calls generated in compiled code.
348 // It is not guaranteed that we can get such information here only
349 // by analyzing bytecode in deoptimized frames. This is why this flag
350 // is set during method compilation (see Compile::Process_OopMap_Node()).
351 // If the previous frame was popped or if we are dispatching an exception,
352 // we don't have an oop result.
353 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
354 Handle return_value;
355 if (save_oop_result) {
356 // Reallocation may trigger GC. If deoptimization happened on return from
357 // call which returns oop we need to save it since it is not in oopmap.
358 oop result = deoptee.saved_oop_result(&map);
359 assert(oopDesc::is_oop_or_null(result), "must be oop");
360 return_value = Handle(thread, result);
361 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
362 if (TraceDeoptimization) {
363 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
364 tty->cr();
365 }
366 }
367 if (objects != nullptr) {
368 if (exec_mode == Deoptimization::Unpack_none) {
369 assert(thread->thread_state() == _thread_in_vm, "assumption");
370 JavaThread* THREAD = thread; // For exception macros.
371 // Clear pending OOM if reallocation fails and return true indicating allocation failure
372 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
373 deoptimized_objects = true;
374 } else {
375 JavaThread* current = thread; // For JRT_BLOCK
376 JRT_BLOCK
377 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
378 JRT_END
379 }
380 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci();
381 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
382 if (TraceDeoptimization) {
383 print_objects(deoptee_thread, objects, realloc_failures);
384 }
385 }
386 if (save_oop_result) {
387 // Restore result.
388 deoptee.set_saved_oop_result(&map, return_value());
389 }
390 return realloc_failures;
391 }
392
393 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
394 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
395 JavaThread* deoptee_thread = chunk->at(0)->thread();
396 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
397 assert(thread == Thread::current(), "should be");
398 HandleMark hm(thread);
399 #ifndef PRODUCT
400 bool first = true;
401 #endif // !PRODUCT
402 // Start locking from outermost/oldest frame
403 for (int i = (chunk->length() - 1); i >= 0; i--) {
404 compiledVFrame* cvf = chunk->at(i);
405 assert (cvf->scope() != nullptr,"expect only compiled java frames");
406 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
407 if (monitors->is_nonempty()) {
408 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
409 exec_mode, realloc_failures);
410 deoptimized_objects = deoptimized_objects || relocked;
411 #ifndef PRODUCT
412 if (PrintDeoptimizationDetails) {
413 ResourceMark rm;
414 stringStream st;
415 for (int j = 0; j < monitors->length(); j++) {
416 MonitorInfo* mi = monitors->at(j);
417 if (mi->eliminated()) {
418 if (first) {
419 first = false;
420 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
421 }
422 if (exec_mode == Deoptimization::Unpack_none) {
423 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
424 if (monitor != nullptr && monitor->object() == mi->owner()) {
425 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
426 continue;
427 }
428 }
429 if (mi->owner_is_scalar_replaced()) {
430 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
431 st.print_cr(" failed reallocation for klass %s", k->external_name());
432 } else {
433 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
434 }
435 }
436 }
437 tty->print_raw(st.freeze());
438 }
439 #endif // !PRODUCT
440 }
441 }
442 }
443
444 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
445 // The given vframes cover one physical frame.
446 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
447 bool& realloc_failures) {
448 frame deoptee = chunk->at(0)->fr();
449 JavaThread* deoptee_thread = chunk->at(0)->thread();
450 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
451 RegisterMap map(chunk->at(0)->register_map());
452 bool deoptimized_objects = false;
453
454 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
455
456 // Reallocate the non-escaping objects and restore their fields.
457 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
458 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
459 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
460 }
461
462 // MonitorInfo structures used in eliminate_locks are not GC safe.
463 NoSafepointVerifier no_safepoint;
464
465 // Now relock objects if synchronization on them was eliminated.
466 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
467 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
468 }
469 return deoptimized_objects;
470 }
471 #endif // COMPILER2_OR_JVMCI
472
473 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
474 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
475 // When we get here we are about to unwind the deoptee frame. In order to
476 // catch not yet safe to use frames, the following stack watermark barrier
477 // poll will make such frames safe to use.
478 StackWatermarkSet::before_unwind(current);
479
480 // Note: there is a safepoint safety issue here. No matter whether we enter
481 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
482 // the vframeArray is created.
483 //
484
485 // Allocate our special deoptimization ResourceMark
486 DeoptResourceMark* dmark = new DeoptResourceMark(current);
487 assert(current->deopt_mark() == nullptr, "Pending deopt!");
488 current->set_deopt_mark(dmark);
489
490 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
491 RegisterMap map(current,
492 RegisterMap::UpdateMap::include,
493 RegisterMap::ProcessFrames::include,
494 RegisterMap::WalkContinuation::skip);
495 RegisterMap dummy_map(current,
496 RegisterMap::UpdateMap::skip,
497 RegisterMap::ProcessFrames::include,
498 RegisterMap::WalkContinuation::skip);
499 // Now get the deoptee with a valid map
500 frame deoptee = stub_frame.sender(&map);
501 // Set the deoptee nmethod
502 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
503 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
504 current->set_deopt_compiled_method(nm);
505
506 if (VerifyStack) {
507 current->validate_frame_layout();
508 }
509
510 // Create a growable array of VFrames where each VFrame represents an inlined
511 // Java frame. This storage is allocated with the usual system arena.
512 assert(deoptee.is_compiled_frame(), "Wrong frame type");
513 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
514 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
515 while (!vf->is_top()) {
516 assert(vf->is_compiled_frame(), "Wrong frame type");
517 chunk->push(compiledVFrame::cast(vf));
518 vf = vf->sender();
519 }
520 assert(vf->is_compiled_frame(), "Wrong frame type");
521 chunk->push(compiledVFrame::cast(vf));
522
523 bool realloc_failures = false;
524
525 #if COMPILER2_OR_JVMCI
526 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
527
528 // Reallocate the non-escaping objects and restore their fields. Then
529 // relock objects if synchronization on them was eliminated.
530 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
531 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
532 bool unused;
533 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
534 }
535 #endif // COMPILER2_OR_JVMCI
536
537 // Ensure that no safepoint is taken after pointers have been stored
538 // in fields of rematerialized objects. If a safepoint occurs from here on
539 // out the java state residing in the vframeArray will be missed.
540 // Locks may be rebaised in a safepoint.
541 NoSafepointVerifier no_safepoint;
542
543 #if COMPILER2_OR_JVMCI
544 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
545 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
546 bool unused = false;
547 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
548 }
549 #endif // COMPILER2_OR_JVMCI
550
551 ScopeDesc* trap_scope = chunk->at(0)->scope();
552 Handle exceptionObject;
553 if (trap_scope->rethrow_exception()) {
554 #ifndef PRODUCT
555 if (PrintDeoptimizationDetails) {
556 tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci());
557 }
558 #endif // !PRODUCT
559
560 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
561 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
562 ScopeValue* topOfStack = expressions->top();
563 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
564 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
565 }
566
567 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
568 #if COMPILER2_OR_JVMCI
569 if (realloc_failures) {
570 // This destroys all ScopedValue bindings.
571 current->clear_scopedValueBindings();
572 pop_frames_failed_reallocs(current, array);
573 }
574 #endif
575
576 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
577 current->set_vframe_array_head(array);
578
579 // Now that the vframeArray has been created if we have any deferred local writes
580 // added by jvmti then we can free up that structure as the data is now in the
581 // vframeArray
582
583 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
584
585 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
586 CodeBlob* cb = stub_frame.cb();
587 // Verify we have the right vframeArray
588 assert(cb->frame_size() >= 0, "Unexpected frame size");
589 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
590
591 // If the deopt call site is a MethodHandle invoke call site we have
592 // to adjust the unpack_sp.
593 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
594 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
595 unpack_sp = deoptee.unextended_sp();
596
597 #ifdef ASSERT
598 assert(cb->is_deoptimization_stub() ||
599 cb->is_uncommon_trap_stub() ||
600 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
601 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
602 "unexpected code blob: %s", cb->name());
603 #endif
604
605 // This is a guarantee instead of an assert because if vframe doesn't match
606 // we will unpack the wrong deoptimized frame and wind up in strange places
607 // where it will be very difficult to figure out what went wrong. Better
608 // to die an early death here than some very obscure death later when the
609 // trail is cold.
610 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
611 // in that it will fail to detect a problem when there is one. This needs
612 // more work in tiger timeframe.
613 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
614
615 int number_of_frames = array->frames();
616
617 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
618 // virtual activation, which is the reverse of the elements in the vframes array.
619 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
620 // +1 because we always have an interpreter return address for the final slot.
621 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
622 int popframe_extra_args = 0;
623 // Create an interpreter return address for the stub to use as its return
624 // address so the skeletal frames are perfectly walkable
625 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
626
627 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
628 // activation be put back on the expression stack of the caller for reexecution
629 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
630 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
631 }
632
633 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
634 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
635 // than simply use array->sender.pc(). This requires us to walk the current set of frames
636 //
637 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
638 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
639
640 // It's possible that the number of parameters at the call site is
641 // different than number of arguments in the callee when method
642 // handles are used. If the caller is interpreted get the real
643 // value so that the proper amount of space can be added to it's
644 // frame.
645 bool caller_was_method_handle = false;
646 if (deopt_sender.is_interpreted_frame()) {
647 methodHandle method(current, deopt_sender.interpreter_frame_method());
648 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
649 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
650 // Method handle invokes may involve fairly arbitrary chains of
651 // calls so it's impossible to know how much actual space the
652 // caller has for locals.
653 caller_was_method_handle = true;
654 }
655 }
656
657 //
658 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
659 // frame_sizes/frame_pcs[1] next oldest frame (int)
660 // frame_sizes/frame_pcs[n] youngest frame (int)
661 //
662 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
663 // owns the space for the return address to it's caller). Confusing ain't it.
664 //
665 // The vframe array can address vframes with indices running from
666 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
667 // When we create the skeletal frames we need the oldest frame to be in the zero slot
668 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
669 // so things look a little strange in this loop.
670 //
671 int callee_parameters = 0;
672 int callee_locals = 0;
673 for (int index = 0; index < array->frames(); index++ ) {
674 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
675 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
676 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
677 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
678 callee_locals,
679 index == 0,
680 popframe_extra_args);
681 // This pc doesn't have to be perfect just good enough to identify the frame
682 // as interpreted so the skeleton frame will be walkable
683 // The correct pc will be set when the skeleton frame is completely filled out
684 // The final pc we store in the loop is wrong and will be overwritten below
685 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
686
687 callee_parameters = array->element(index)->method()->size_of_parameters();
688 callee_locals = array->element(index)->method()->max_locals();
689 popframe_extra_args = 0;
690 }
691
692 // Compute whether the root vframe returns a float or double value.
693 BasicType return_type;
694 {
695 methodHandle method(current, array->element(0)->method());
696 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
697 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
698 }
699
700 // Compute information for handling adapters and adjusting the frame size of the caller.
701 int caller_adjustment = 0;
702
703 // Compute the amount the oldest interpreter frame will have to adjust
704 // its caller's stack by. If the caller is a compiled frame then
705 // we pretend that the callee has no parameters so that the
706 // extension counts for the full amount of locals and not just
707 // locals-parms. This is because without a c2i adapter the parm
708 // area as created by the compiled frame will not be usable by
709 // the interpreter. (Depending on the calling convention there
710 // may not even be enough space).
711
712 // QQQ I'd rather see this pushed down into last_frame_adjust
713 // and have it take the sender (aka caller).
714
715 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
716 caller_adjustment = last_frame_adjust(0, callee_locals);
717 } else if (callee_locals > callee_parameters) {
718 // The caller frame may need extending to accommodate
719 // non-parameter locals of the first unpacked interpreted frame.
720 // Compute that adjustment.
721 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
722 }
723
724 // If the sender is deoptimized the we must retrieve the address of the handler
725 // since the frame will "magically" show the original pc before the deopt
726 // and we'd undo the deopt.
727
728 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
729 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
730 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
731 }
732
733 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
734
735 #if INCLUDE_JVMCI
736 if (exceptionObject() != nullptr) {
737 current->set_exception_oop(exceptionObject());
738 exec_mode = Unpack_exception;
739 }
740 #endif
741
742 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
743 assert(current->has_pending_exception(), "should have thrown OOME");
744 current->set_exception_oop(current->pending_exception());
745 current->clear_pending_exception();
746 exec_mode = Unpack_exception;
747 }
748
749 #if INCLUDE_JVMCI
750 if (current->frames_to_pop_failed_realloc() > 0) {
751 current->set_pending_monitorenter(false);
752 }
753 #endif
754
755 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
756 caller_adjustment * BytesPerWord,
757 caller_was_method_handle ? 0 : callee_parameters,
758 number_of_frames,
759 frame_sizes,
760 frame_pcs,
761 return_type,
762 exec_mode);
763 // On some platforms, we need a way to pass some platform dependent
764 // information to the unpacking code so the skeletal frames come out
765 // correct (initial fp value, unextended sp, ...)
766 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
767
768 if (array->frames() > 1) {
769 if (VerifyStack && TraceDeoptimization) {
770 tty->print_cr("Deoptimizing method containing inlining");
771 }
772 }
773
774 array->set_unroll_block(info);
775 return info;
776 }
777
778 // Called to cleanup deoptimization data structures in normal case
779 // after unpacking to stack and when stack overflow error occurs
780 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
781 vframeArray *array) {
782
783 // Get array if coming from exception
784 if (array == nullptr) {
785 array = thread->vframe_array_head();
786 }
787 thread->set_vframe_array_head(nullptr);
788
789 // Free the previous UnrollBlock
790 vframeArray* old_array = thread->vframe_array_last();
791 thread->set_vframe_array_last(array);
792
793 if (old_array != nullptr) {
794 UnrollBlock* old_info = old_array->unroll_block();
795 old_array->set_unroll_block(nullptr);
796 delete old_info;
797 delete old_array;
798 }
799
800 // Deallocate any resource creating in this routine and any ResourceObjs allocated
801 // inside the vframeArray (StackValueCollections)
802
803 delete thread->deopt_mark();
804 thread->set_deopt_mark(nullptr);
805 thread->set_deopt_compiled_method(nullptr);
806
807
808 if (JvmtiExport::can_pop_frame()) {
809 // Regardless of whether we entered this routine with the pending
810 // popframe condition bit set, we should always clear it now
811 thread->clear_popframe_condition();
812 }
813
814 // unpack_frames() is called at the end of the deoptimization handler
815 // and (in C2) at the end of the uncommon trap handler. Note this fact
816 // so that an asynchronous stack walker can work again. This counter is
817 // incremented at the beginning of fetch_unroll_info() and (in C2) at
818 // the beginning of uncommon_trap().
819 thread->dec_in_deopt_handler();
820 }
821
822 // Moved from cpu directories because none of the cpus has callee save values.
823 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
824 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
825
826 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
827 // the days we had adapter frames. When we deoptimize a situation where a
828 // compiled caller calls a compiled caller will have registers it expects
829 // to survive the call to the callee. If we deoptimize the callee the only
830 // way we can restore these registers is to have the oldest interpreter
831 // frame that we create restore these values. That is what this routine
832 // will accomplish.
833
834 // At the moment we have modified c2 to not have any callee save registers
835 // so this problem does not exist and this routine is just a place holder.
836
837 assert(f->is_interpreted_frame(), "must be interpreted");
838 }
839
840 #ifndef PRODUCT
841 static bool falls_through(Bytecodes::Code bc) {
842 switch (bc) {
843 // List may be incomplete. Here we really only care about bytecodes where compiled code
844 // can deoptimize.
845 case Bytecodes::_goto:
846 case Bytecodes::_goto_w:
847 case Bytecodes::_athrow:
848 return false;
849 default:
850 return true;
851 }
852 }
853 #endif
854
855 // Return BasicType of value being returned
856 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
857 assert(thread == JavaThread::current(), "pre-condition");
858
859 // We are already active in the special DeoptResourceMark any ResourceObj's we
860 // allocate will be freed at the end of the routine.
861
862 // JRT_LEAF methods don't normally allocate handles and there is a
863 // NoHandleMark to enforce that. It is actually safe to use Handles
864 // in a JRT_LEAF method, and sometimes desirable, but to do so we
865 // must use ResetNoHandleMark to bypass the NoHandleMark, and
866 // then use a HandleMark to ensure any Handles we do create are
867 // cleaned up in this scope.
868 ResetNoHandleMark rnhm;
869 HandleMark hm(thread);
870
871 frame stub_frame = thread->last_frame();
872
873 Continuation::notify_deopt(thread, stub_frame.sp());
874
875 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
876 // must point to the vframeArray for the unpack frame.
877 vframeArray* array = thread->vframe_array_head();
878 UnrollBlock* info = array->unroll_block();
879
880 // We set the last_Java frame. But the stack isn't really parsable here. So we
881 // clear it to make sure JFR understands not to try and walk stacks from events
882 // in here.
883 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
884 thread->frame_anchor()->set_last_Java_sp(nullptr);
885
886 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
887 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
888
889 thread->frame_anchor()->set_last_Java_sp(sp);
890
891 BasicType bt = info->return_type();
892
893 // If we have an exception pending, claim that the return type is an oop
894 // so the deopt_blob does not overwrite the exception_oop.
895
896 if (exec_mode == Unpack_exception)
897 bt = T_OBJECT;
898
899 // Cleanup thread deopt data
900 cleanup_deopt_info(thread, array);
901
902 #ifndef PRODUCT
903 if (VerifyStack) {
904 ResourceMark res_mark;
905 // Clear pending exception to not break verification code (restored afterwards)
906 PreserveExceptionMark pm(thread);
907
908 thread->validate_frame_layout();
909
910 // Verify that the just-unpacked frames match the interpreter's
911 // notions of expression stack and locals
912 vframeArray* cur_array = thread->vframe_array_last();
913 RegisterMap rm(thread,
914 RegisterMap::UpdateMap::skip,
915 RegisterMap::ProcessFrames::include,
916 RegisterMap::WalkContinuation::skip);
917 rm.set_include_argument_oops(false);
918 bool is_top_frame = true;
919 int callee_size_of_parameters = 0;
920 int callee_max_locals = 0;
921 for (int i = 0; i < cur_array->frames(); i++) {
922 vframeArrayElement* el = cur_array->element(i);
923 frame* iframe = el->iframe();
924 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
925
926 // Get the oop map for this bci
927 InterpreterOopMap mask;
928 int cur_invoke_parameter_size = 0;
929 bool try_next_mask = false;
930 int next_mask_expression_stack_size = -1;
931 int top_frame_expression_stack_adjustment = 0;
932 methodHandle mh(thread, iframe->interpreter_frame_method());
933 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
934 BytecodeStream str(mh, iframe->interpreter_frame_bci());
935 int max_bci = mh->code_size();
936 // Get to the next bytecode if possible
937 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
938 // Check to see if we can grab the number of outgoing arguments
939 // at an uncommon trap for an invoke (where the compiler
940 // generates debug info before the invoke has executed)
941 Bytecodes::Code cur_code = str.next();
942 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
943 if (Bytecodes::is_invoke(cur_code)) {
944 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
945 cur_invoke_parameter_size = invoke.size_of_parameters();
946 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
947 callee_size_of_parameters++;
948 }
949 }
950 if (str.bci() < max_bci) {
951 next_code = str.next();
952 if (next_code >= 0) {
953 // The interpreter oop map generator reports results before
954 // the current bytecode has executed except in the case of
955 // calls. It seems to be hard to tell whether the compiler
956 // has emitted debug information matching the "state before"
957 // a given bytecode or the state after, so we try both
958 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
959 // Get expression stack size for the next bytecode
960 InterpreterOopMap next_mask;
961 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
962 next_mask_expression_stack_size = next_mask.expression_stack_size();
963 if (Bytecodes::is_invoke(next_code)) {
964 Bytecode_invoke invoke(mh, str.bci());
965 next_mask_expression_stack_size += invoke.size_of_parameters();
966 }
967 // Need to subtract off the size of the result type of
968 // the bytecode because this is not described in the
969 // debug info but returned to the interpreter in the TOS
970 // caching register
971 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
972 if (bytecode_result_type != T_ILLEGAL) {
973 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
974 }
975 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
976 try_next_mask = true;
977 }
978 }
979 }
980
981 // Verify stack depth and oops in frame
982 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
983 if (!(
984 /* SPARC */
985 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
986 /* x86 */
987 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
988 (try_next_mask &&
989 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
990 top_frame_expression_stack_adjustment))) ||
991 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
992 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
993 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
994 )) {
995 {
996 // Print out some information that will help us debug the problem
997 tty->print_cr("Wrong number of expression stack elements during deoptimization");
998 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
999 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
1000 if (try_next_mask) {
1001 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
1002 }
1003 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
1004 iframe->interpreter_frame_expression_stack_size());
1005 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1006 tty->print_cr(" try_next_mask = %d", try_next_mask);
1007 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1008 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1009 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1010 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1011 tty->print_cr(" exec_mode = %d", exec_mode);
1012 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1013 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1014 tty->print_cr(" Interpreted frames:");
1015 for (int k = 0; k < cur_array->frames(); k++) {
1016 vframeArrayElement* el = cur_array->element(k);
1017 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1018 }
1019 cur_array->print_on_2(tty);
1020 }
1021 guarantee(false, "wrong number of expression stack elements during deopt");
1022 }
1023 VerifyOopClosure verify;
1024 iframe->oops_interpreted_do(&verify, &rm, false);
1025 callee_size_of_parameters = mh->size_of_parameters();
1026 callee_max_locals = mh->max_locals();
1027 is_top_frame = false;
1028 }
1029 }
1030 #endif // !PRODUCT
1031
1032 return bt;
1033 JRT_END
1034
1035 class DeoptimizeMarkedClosure : public HandshakeClosure {
1036 public:
1037 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1038 void do_thread(Thread* thread) {
1039 JavaThread* jt = JavaThread::cast(thread);
1040 jt->deoptimize_marked_methods();
1041 }
1042 };
1043
1044 void Deoptimization::deoptimize_all_marked() {
1045 ResourceMark rm;
1046
1047 // Make the dependent methods not entrant
1048 CodeCache::make_marked_nmethods_deoptimized();
1049
1050 DeoptimizeMarkedClosure deopt;
1051 if (SafepointSynchronize::is_at_safepoint()) {
1052 Threads::java_threads_do(&deopt);
1053 } else {
1054 Handshake::execute(&deopt);
1055 }
1056 }
1057
1058 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1059 = Deoptimization::Action_reinterpret;
1060
1061 #if INCLUDE_JVMCI
1062 template<typename CacheType>
1063 class BoxCacheBase : public CHeapObj<mtCompiler> {
1064 protected:
1065 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1066 ResourceMark rm(thread);
1067 char* klass_name_str = klass_name->as_C_string();
1068 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle());
1069 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1070 if (!ik->is_in_error_state()) {
1071 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1072 CacheType::compute_offsets(ik);
1073 }
1074 return ik;
1075 }
1076 };
1077
1078 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1079 PrimitiveType _low;
1080 PrimitiveType _high;
1081 jobject _cache;
1082 protected:
1083 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1084 BoxCache(Thread* thread) {
1085 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1086 if (ik->is_in_error_state()) {
1087 _low = 1;
1088 _high = 0;
1089 _cache = nullptr;
1090 } else {
1091 objArrayOop cache = CacheType::cache(ik);
1092 assert(cache->length() > 0, "Empty cache");
1093 _low = BoxType::value(cache->obj_at(0));
1094 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1095 _cache = JNIHandles::make_global(Handle(thread, cache));
1096 }
1097 }
1098 ~BoxCache() {
1099 JNIHandles::destroy_global(_cache);
1100 }
1101 public:
1102 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1103 if (_singleton == nullptr) {
1104 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1105 if (!Atomic::replace_if_null(&_singleton, s)) {
1106 delete s;
1107 }
1108 }
1109 return _singleton;
1110 }
1111 oop lookup(PrimitiveType value) {
1112 if (_low <= value && value <= _high) {
1113 int offset = checked_cast<int>(value - _low);
1114 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1115 }
1116 return nullptr;
1117 }
1118 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1119 if (_cache == nullptr) {
1120 cache_init_error = true;
1121 return nullptr;
1122 }
1123 // Have to cast to avoid little/big-endian problems.
1124 if (sizeof(PrimitiveType) > sizeof(jint)) {
1125 jlong value = (jlong)raw_value;
1126 return lookup(value);
1127 }
1128 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1129 return lookup(value);
1130 }
1131 };
1132
1133 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1134 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1135 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1136 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1137 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1138
1139 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1140 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1141 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1142 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1143 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1144
1145 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1146 jobject _true_cache;
1147 jobject _false_cache;
1148 protected:
1149 static BooleanBoxCache *_singleton;
1150 BooleanBoxCache(Thread *thread) {
1151 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1152 if (ik->is_in_error_state()) {
1153 _true_cache = nullptr;
1154 _false_cache = nullptr;
1155 } else {
1156 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1157 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1158 }
1159 }
1160 ~BooleanBoxCache() {
1161 JNIHandles::destroy_global(_true_cache);
1162 JNIHandles::destroy_global(_false_cache);
1163 }
1164 public:
1165 static BooleanBoxCache* singleton(Thread* thread) {
1166 if (_singleton == nullptr) {
1167 BooleanBoxCache* s = new BooleanBoxCache(thread);
1168 if (!Atomic::replace_if_null(&_singleton, s)) {
1169 delete s;
1170 }
1171 }
1172 return _singleton;
1173 }
1174 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1175 if (_true_cache == nullptr) {
1176 cache_in_error = true;
1177 return nullptr;
1178 }
1179 // Have to cast to avoid little/big-endian problems.
1180 jboolean value = (jboolean)*((jint*)&raw_value);
1181 return lookup(value);
1182 }
1183 oop lookup(jboolean value) {
1184 if (value != 0) {
1185 return JNIHandles::resolve_non_null(_true_cache);
1186 }
1187 return JNIHandles::resolve_non_null(_false_cache);
1188 }
1189 };
1190
1191 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1192
1193 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1194 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1195 BasicType box_type = vmClasses::box_klass_type(k);
1196 if (box_type != T_OBJECT) {
1197 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1198 switch(box_type) {
1199 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1200 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1201 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1202 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1203 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1204 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1205 default:;
1206 }
1207 }
1208 return nullptr;
1209 }
1210 #endif // INCLUDE_JVMCI
1211
1212 #if COMPILER2_OR_JVMCI
1213 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1214 Handle pending_exception(THREAD, thread->pending_exception());
1215 const char* exception_file = thread->exception_file();
1216 int exception_line = thread->exception_line();
1217 thread->clear_pending_exception();
1218
1219 bool failures = false;
1220
1221 for (int i = 0; i < objects->length(); i++) {
1222 assert(objects->at(i)->is_object(), "invalid debug information");
1223 ObjectValue* sv = (ObjectValue*) objects->at(i);
1224
1225 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1226 oop obj = nullptr;
1227
1228 bool cache_init_error = false;
1229 if (k->is_instance_klass()) {
1230 #if INCLUDE_JVMCI
1231 nmethod* nm = fr->cb()->as_nmethod_or_null();
1232 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1233 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1234 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1235 if (obj != nullptr) {
1236 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1237 abv->set_cached(true);
1238 } else if (cache_init_error) {
1239 // Results in an OOME which is valid (as opposed to a class initialization error)
1240 // and is fine for the rare case a cache initialization failing.
1241 failures = true;
1242 }
1243 }
1244 #endif // INCLUDE_JVMCI
1245
1246 InstanceKlass* ik = InstanceKlass::cast(k);
1247 if (obj == nullptr && !cache_init_error) {
1248 InternalOOMEMark iom(THREAD);
1249 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1250 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1251 } else {
1252 obj = ik->allocate_instance(THREAD);
1253 }
1254 }
1255 } else if (k->is_typeArray_klass()) {
1256 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1257 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1258 int len = sv->field_size() / type2size[ak->element_type()];
1259 InternalOOMEMark iom(THREAD);
1260 obj = ak->allocate(len, THREAD);
1261 } else if (k->is_objArray_klass()) {
1262 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1263 InternalOOMEMark iom(THREAD);
1264 obj = ak->allocate(sv->field_size(), THREAD);
1265 }
1266
1267 if (obj == nullptr) {
1268 failures = true;
1269 }
1270
1271 assert(sv->value().is_null(), "redundant reallocation");
1272 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1273 CLEAR_PENDING_EXCEPTION;
1274 sv->set_value(obj);
1275 }
1276
1277 if (failures) {
1278 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1279 } else if (pending_exception.not_null()) {
1280 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1281 }
1282
1283 return failures;
1284 }
1285
1286 #if INCLUDE_JVMCI
1287 /**
1288 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1289 * we need to somehow be able to recover the actual kind to be able to write the correct
1290 * amount of bytes.
1291 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1292 * the entries at index n + 1 to n + i are 'markers'.
1293 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1294 * expected form of the array would be:
1295 *
1296 * {b0, b1, b2, b3, INT, marker, b6, b7}
1297 *
1298 * Thus, in order to get back the size of the entry, we simply need to count the number
1299 * of marked entries
1300 *
1301 * @param virtualArray the virtualized byte array
1302 * @param i index of the virtual entry we are recovering
1303 * @return The number of bytes the entry spans
1304 */
1305 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1306 int index = i;
1307 while (++index < virtualArray->field_size() &&
1308 virtualArray->field_at(index)->is_marker()) {}
1309 return index - i;
1310 }
1311
1312 /**
1313 * If there was a guarantee for byte array to always start aligned to a long, we could
1314 * do a simple check on the parity of the index. Unfortunately, that is not always the
1315 * case. Thus, we check alignment of the actual address we are writing to.
1316 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1317 * write a long to index 3.
1318 */
1319 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1320 jbyte* res = obj->byte_at_addr(index);
1321 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1322 return res;
1323 }
1324
1325 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1326 switch (byte_count) {
1327 case 1:
1328 obj->byte_at_put(index, (jbyte) value->get_jint());
1329 break;
1330 case 2:
1331 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1332 break;
1333 case 4:
1334 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1335 break;
1336 case 8:
1337 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1338 break;
1339 default:
1340 ShouldNotReachHere();
1341 }
1342 }
1343 #endif // INCLUDE_JVMCI
1344
1345
1346 // restore elements of an eliminated type array
1347 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1348 int index = 0;
1349
1350 for (int i = 0; i < sv->field_size(); i++) {
1351 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1352 switch(type) {
1353 case T_LONG: case T_DOUBLE: {
1354 assert(value->type() == T_INT, "Agreement.");
1355 StackValue* low =
1356 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1357 #ifdef _LP64
1358 jlong res = (jlong)low->get_intptr();
1359 #else
1360 jlong res = jlong_from(value->get_jint(), low->get_jint());
1361 #endif
1362 obj->long_at_put(index, res);
1363 break;
1364 }
1365
1366 case T_INT: case T_FLOAT: { // 4 bytes.
1367 assert(value->type() == T_INT, "Agreement.");
1368 bool big_value = false;
1369 if (i + 1 < sv->field_size() && type == T_INT) {
1370 if (sv->field_at(i)->is_location()) {
1371 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1372 if (type == Location::dbl || type == Location::lng) {
1373 big_value = true;
1374 }
1375 } else if (sv->field_at(i)->is_constant_int()) {
1376 ScopeValue* next_scope_field = sv->field_at(i + 1);
1377 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1378 big_value = true;
1379 }
1380 }
1381 }
1382
1383 if (big_value) {
1384 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1385 #ifdef _LP64
1386 jlong res = (jlong)low->get_intptr();
1387 #else
1388 jlong res = jlong_from(value->get_jint(), low->get_jint());
1389 #endif
1390 obj->int_at_put(index, *(jint*)&res);
1391 obj->int_at_put(++index, *((jint*)&res + 1));
1392 } else {
1393 obj->int_at_put(index, value->get_jint());
1394 }
1395 break;
1396 }
1397
1398 case T_SHORT:
1399 assert(value->type() == T_INT, "Agreement.");
1400 obj->short_at_put(index, (jshort)value->get_jint());
1401 break;
1402
1403 case T_CHAR:
1404 assert(value->type() == T_INT, "Agreement.");
1405 obj->char_at_put(index, (jchar)value->get_jint());
1406 break;
1407
1408 case T_BYTE: {
1409 assert(value->type() == T_INT, "Agreement.");
1410 #if INCLUDE_JVMCI
1411 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1412 int byte_count = count_number_of_bytes_for_entry(sv, i);
1413 byte_array_put(obj, value, index, byte_count);
1414 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1415 i += byte_count - 1; // Balance the loop counter.
1416 index += byte_count;
1417 // index has been updated so continue at top of loop
1418 continue;
1419 #else
1420 obj->byte_at_put(index, (jbyte)value->get_jint());
1421 break;
1422 #endif // INCLUDE_JVMCI
1423 }
1424
1425 case T_BOOLEAN: {
1426 assert(value->type() == T_INT, "Agreement.");
1427 obj->bool_at_put(index, (jboolean)value->get_jint());
1428 break;
1429 }
1430
1431 default:
1432 ShouldNotReachHere();
1433 }
1434 index++;
1435 }
1436 }
1437
1438 // restore fields of an eliminated object array
1439 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1440 for (int i = 0; i < sv->field_size(); i++) {
1441 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1442 assert(value->type() == T_OBJECT, "object element expected");
1443 obj->obj_at_put(i, value->get_obj()());
1444 }
1445 }
1446
1447 class ReassignedField {
1448 public:
1449 int _offset;
1450 BasicType _type;
1451 public:
1452 ReassignedField() {
1453 _offset = 0;
1454 _type = T_ILLEGAL;
1455 }
1456 };
1457
1458 static int compare(ReassignedField* left, ReassignedField* right) {
1459 return left->_offset - right->_offset;
1460 }
1461
1462 // Restore fields of an eliminated instance object using the same field order
1463 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1464 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1465 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1466 InstanceKlass* ik = klass;
1467 while (ik != nullptr) {
1468 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1469 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1470 ReassignedField field;
1471 field._offset = fs.offset();
1472 field._type = Signature::basic_type(fs.signature());
1473 fields->append(field);
1474 }
1475 }
1476 ik = ik->superklass();
1477 }
1478 fields->sort(compare);
1479 for (int i = 0; i < fields->length(); i++) {
1480 ScopeValue* scope_field = sv->field_at(svIndex);
1481 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1482 int offset = fields->at(i)._offset;
1483 BasicType type = fields->at(i)._type;
1484 switch (type) {
1485 case T_OBJECT: case T_ARRAY:
1486 assert(value->type() == T_OBJECT, "Agreement.");
1487 obj->obj_field_put(offset, value->get_obj()());
1488 break;
1489
1490 case T_INT: case T_FLOAT: { // 4 bytes.
1491 assert(value->type() == T_INT, "Agreement.");
1492 bool big_value = false;
1493 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1494 if (scope_field->is_location()) {
1495 Location::Type type = ((LocationValue*) scope_field)->location().type();
1496 if (type == Location::dbl || type == Location::lng) {
1497 big_value = true;
1498 }
1499 }
1500 if (scope_field->is_constant_int()) {
1501 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1502 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1503 big_value = true;
1504 }
1505 }
1506 }
1507
1508 if (big_value) {
1509 i++;
1510 assert(i < fields->length(), "second T_INT field needed");
1511 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1512 } else {
1513 obj->int_field_put(offset, value->get_jint());
1514 break;
1515 }
1516 }
1517 /* no break */
1518
1519 case T_LONG: case T_DOUBLE: {
1520 assert(value->type() == T_INT, "Agreement.");
1521 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1522 #ifdef _LP64
1523 jlong res = (jlong)low->get_intptr();
1524 #else
1525 jlong res = jlong_from(value->get_jint(), low->get_jint());
1526 #endif
1527 obj->long_field_put(offset, res);
1528 break;
1529 }
1530
1531 case T_SHORT:
1532 assert(value->type() == T_INT, "Agreement.");
1533 obj->short_field_put(offset, (jshort)value->get_jint());
1534 break;
1535
1536 case T_CHAR:
1537 assert(value->type() == T_INT, "Agreement.");
1538 obj->char_field_put(offset, (jchar)value->get_jint());
1539 break;
1540
1541 case T_BYTE:
1542 assert(value->type() == T_INT, "Agreement.");
1543 obj->byte_field_put(offset, (jbyte)value->get_jint());
1544 break;
1545
1546 case T_BOOLEAN:
1547 assert(value->type() == T_INT, "Agreement.");
1548 obj->bool_field_put(offset, (jboolean)value->get_jint());
1549 break;
1550
1551 default:
1552 ShouldNotReachHere();
1553 }
1554 svIndex++;
1555 }
1556 return svIndex;
1557 }
1558
1559 // restore fields of all eliminated objects and arrays
1560 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1561 for (int i = 0; i < objects->length(); i++) {
1562 assert(objects->at(i)->is_object(), "invalid debug information");
1563 ObjectValue* sv = (ObjectValue*) objects->at(i);
1564 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1565 Handle obj = sv->value();
1566 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1567 #ifndef PRODUCT
1568 if (PrintDeoptimizationDetails) {
1569 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1570 }
1571 #endif // !PRODUCT
1572
1573 if (obj.is_null()) {
1574 continue;
1575 }
1576
1577 #if INCLUDE_JVMCI
1578 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1579 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1580 continue;
1581 }
1582 #endif // INCLUDE_JVMCI
1583 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1584 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1585 ScopeValue* payload = sv->field_at(0);
1586 if (payload->is_location() &&
1587 payload->as_LocationValue()->location().type() == Location::vector) {
1588 #ifndef PRODUCT
1589 if (PrintDeoptimizationDetails) {
1590 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1591 if (Verbose) {
1592 Handle obj = sv->value();
1593 k->oop_print_on(obj(), tty);
1594 }
1595 }
1596 #endif // !PRODUCT
1597 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1598 }
1599 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1600 // which could be restored after vector object allocation.
1601 }
1602 if (k->is_instance_klass()) {
1603 InstanceKlass* ik = InstanceKlass::cast(k);
1604 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1605 } else if (k->is_typeArray_klass()) {
1606 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1607 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1608 } else if (k->is_objArray_klass()) {
1609 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1610 }
1611 }
1612 // These objects may escape when we return to Interpreter after deoptimization.
1613 // We need barrier so that stores that initialize these objects can't be reordered
1614 // with subsequent stores that make these objects accessible by other threads.
1615 OrderAccess::storestore();
1616 }
1617
1618
1619 // relock objects for which synchronization was eliminated
1620 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1621 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1622 bool relocked_objects = false;
1623 for (int i = 0; i < monitors->length(); i++) {
1624 MonitorInfo* mon_info = monitors->at(i);
1625 if (mon_info->eliminated()) {
1626 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1627 relocked_objects = true;
1628 if (!mon_info->owner_is_scalar_replaced()) {
1629 Handle obj(thread, mon_info->owner());
1630 markWord mark = obj->mark();
1631 if (exec_mode == Unpack_none) {
1632 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1633 // With exec_mode == Unpack_none obj may be thread local and locked in
1634 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1635 markWord dmw = mark.displaced_mark_helper();
1636 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1637 obj->set_mark(dmw);
1638 }
1639 if (mark.has_monitor()) {
1640 // defer relocking if the deoptee thread is currently waiting for obj
1641 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1642 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1643 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1644 if (LockingMode == LM_LEGACY) {
1645 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1646 } else if (UseObjectMonitorTable) {
1647 mon_info->lock()->clear_object_monitor_cache();
1648 }
1649 #ifdef ASSERT
1650 else {
1651 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1652 mon_info->lock()->set_bad_metadata_deopt();
1653 }
1654 #endif
1655 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1656 continue;
1657 }
1658 }
1659 }
1660 BasicLock* lock = mon_info->lock();
1661 if (LockingMode == LM_LIGHTWEIGHT) {
1662 // We have lost information about the correct state of the lock stack.
1663 // Entering may create an invalid lock stack. Inflate the lock if it
1664 // was fast_locked to restore the valid lock stack.
1665 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1666 if (deoptee_thread->lock_stack().contains(obj())) {
1667 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1668 deoptee_thread, thread);
1669 }
1670 assert(mon_info->owner()->is_locked(), "object must be locked now");
1671 assert(obj->mark().has_monitor(), "must be");
1672 assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1673 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1674 } else {
1675 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1676 assert(mon_info->owner()->is_locked(), "object must be locked now");
1677 }
1678 }
1679 }
1680 }
1681 return relocked_objects;
1682 }
1683 #endif // COMPILER2_OR_JVMCI
1684
1685 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1686 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1687
1688 // Register map for next frame (used for stack crawl). We capture
1689 // the state of the deopt'ing frame's caller. Thus if we need to
1690 // stuff a C2I adapter we can properly fill in the callee-save
1691 // register locations.
1692 frame caller = fr.sender(reg_map);
1693 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1694
1695 frame sender = caller;
1696
1697 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1698 // the vframeArray containing the unpacking information is allocated in the C heap.
1699 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1700 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1701
1702 // Compare the vframeArray to the collected vframes
1703 assert(array->structural_compare(thread, chunk), "just checking");
1704
1705 if (TraceDeoptimization) {
1706 ResourceMark rm;
1707 stringStream st;
1708 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1709 st.print(" ");
1710 fr.print_on(&st);
1711 st.print_cr(" Virtual frames (innermost/newest first):");
1712 for (int index = 0; index < chunk->length(); index++) {
1713 compiledVFrame* vf = chunk->at(index);
1714 int bci = vf->raw_bci();
1715 const char* code_name;
1716 if (bci == SynchronizationEntryBCI) {
1717 code_name = "sync entry";
1718 } else {
1719 Bytecodes::Code code = vf->method()->code_at(bci);
1720 code_name = Bytecodes::name(code);
1721 }
1722
1723 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1724 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1725 st.print(" - %s", code_name);
1726 st.print_cr(" @ bci=%d ", bci);
1727 }
1728 tty->print_raw(st.freeze());
1729 tty->cr();
1730 }
1731
1732 return array;
1733 }
1734
1735 #if COMPILER2_OR_JVMCI
1736 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1737 // Reallocation of some scalar replaced objects failed. Record
1738 // that we need to pop all the interpreter frames for the
1739 // deoptimized compiled frame.
1740 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1741 thread->set_frames_to_pop_failed_realloc(array->frames());
1742 // Unlock all monitors here otherwise the interpreter will see a
1743 // mix of locked and unlocked monitors (because of failed
1744 // reallocations of synchronized objects) and be confused.
1745 for (int i = 0; i < array->frames(); i++) {
1746 MonitorChunk* monitors = array->element(i)->monitors();
1747 if (monitors != nullptr) {
1748 // Unlock in reverse order starting from most nested monitor.
1749 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1750 BasicObjectLock* src = monitors->at(j);
1751 if (src->obj() != nullptr) {
1752 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1753 }
1754 }
1755 array->element(i)->free_monitors();
1756 #ifdef ASSERT
1757 array->element(i)->set_removed_monitors();
1758 #endif
1759 }
1760 }
1761 }
1762 #endif
1763
1764 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1765 assert(fr.can_be_deoptimized(), "checking frame type");
1766
1767 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1768
1769 if (LogCompilation && xtty != nullptr) {
1770 nmethod* nm = fr.cb()->as_nmethod_or_null();
1771 assert(nm != nullptr, "only compiled methods can deopt");
1772
1773 ttyLocker ttyl;
1774 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1775 nm->log_identity(xtty);
1776 xtty->end_head();
1777 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1778 xtty->begin_elem("jvms bci='%d'", sd->bci());
1779 xtty->method(sd->method());
1780 xtty->end_elem();
1781 if (sd->is_top()) break;
1782 }
1783 xtty->tail("deoptimized");
1784 }
1785
1786 Continuation::notify_deopt(thread, fr.sp());
1787
1788 // Patch the compiled method so that when execution returns to it we will
1789 // deopt the execution state and return to the interpreter.
1790 fr.deoptimize(thread);
1791 }
1792
1793 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1794 // Deoptimize only if the frame comes from compile code.
1795 // Do not deoptimize the frame which is already patched
1796 // during the execution of the loops below.
1797 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1798 return;
1799 }
1800 ResourceMark rm;
1801 deoptimize_single_frame(thread, fr, reason);
1802 }
1803
1804 #if INCLUDE_JVMCI
1805 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1806 // there is no exception handler for this pc => deoptimize
1807 nm->make_not_entrant();
1808
1809 // Use Deoptimization::deoptimize for all of its side-effects:
1810 // gathering traps statistics, logging...
1811 // it also patches the return pc but we do not care about that
1812 // since we return a continuation to the deopt_blob below.
1813 JavaThread* thread = JavaThread::current();
1814 RegisterMap reg_map(thread,
1815 RegisterMap::UpdateMap::skip,
1816 RegisterMap::ProcessFrames::include,
1817 RegisterMap::WalkContinuation::skip);
1818 frame runtime_frame = thread->last_frame();
1819 frame caller_frame = runtime_frame.sender(®_map);
1820 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1821 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1822 compiledVFrame* cvf = compiledVFrame::cast(vf);
1823 ScopeDesc* imm_scope = cvf->scope();
1824 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1825 if (imm_mdo != nullptr) {
1826 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1827 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1828
1829 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1830 if (pdata != nullptr && pdata->is_BitData()) {
1831 BitData* bit_data = (BitData*) pdata;
1832 bit_data->set_exception_seen();
1833 }
1834 }
1835
1836 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1837
1838 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1839 if (trap_mdo != nullptr) {
1840 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1841 }
1842
1843 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1844 }
1845 #endif
1846
1847 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1848 assert(thread == Thread::current() ||
1849 thread->is_handshake_safe_for(Thread::current()) ||
1850 SafepointSynchronize::is_at_safepoint(),
1851 "can only deoptimize other thread at a safepoint/handshake");
1852 // Compute frame and register map based on thread and sp.
1853 RegisterMap reg_map(thread,
1854 RegisterMap::UpdateMap::skip,
1855 RegisterMap::ProcessFrames::include,
1856 RegisterMap::WalkContinuation::skip);
1857 frame fr = thread->last_frame();
1858 while (fr.id() != id) {
1859 fr = fr.sender(®_map);
1860 }
1861 deoptimize(thread, fr, reason);
1862 }
1863
1864
1865 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1866 Thread* current = Thread::current();
1867 if (thread == current || thread->is_handshake_safe_for(current)) {
1868 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1869 } else {
1870 VM_DeoptimizeFrame deopt(thread, id, reason);
1871 VMThread::execute(&deopt);
1872 }
1873 }
1874
1875 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1876 deoptimize_frame(thread, id, Reason_constraint);
1877 }
1878
1879 // JVMTI PopFrame support
1880 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1881 {
1882 assert(thread == JavaThread::current(), "pre-condition");
1883 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1884 }
1885 JRT_END
1886
1887 MethodData*
1888 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1889 bool create_if_missing) {
1890 JavaThread* THREAD = thread; // For exception macros.
1891 MethodData* mdo = m()->method_data();
1892 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1893 // Build an MDO. Ignore errors like OutOfMemory;
1894 // that simply means we won't have an MDO to update.
1895 Method::build_profiling_method_data(m, THREAD);
1896 if (HAS_PENDING_EXCEPTION) {
1897 // Only metaspace OOM is expected. No Java code executed.
1898 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1899 CLEAR_PENDING_EXCEPTION;
1900 }
1901 mdo = m()->method_data();
1902 }
1903 return mdo;
1904 }
1905
1906 #if COMPILER2_OR_JVMCI
1907 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1908 // In case of an unresolved klass entry, load the class.
1909 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1910 // and probably nowhere else.
1911 // Even that case would benefit from simply re-interpreting the
1912 // bytecode, without paying special attention to the class index.
1913 // So this whole "class index" feature should probably be removed.
1914
1915 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1916 Klass* tk = constant_pool->klass_at(index, THREAD);
1917 if (HAS_PENDING_EXCEPTION) {
1918 // Exception happened during classloading. We ignore the exception here, since it
1919 // is going to be rethrown since the current activation is going to be deoptimized and
1920 // the interpreter will re-execute the bytecode.
1921 // Do not clear probable Async Exceptions.
1922 CLEAR_PENDING_NONASYNC_EXCEPTION;
1923 // Class loading called java code which may have caused a stack
1924 // overflow. If the exception was thrown right before the return
1925 // to the runtime the stack is no longer guarded. Reguard the
1926 // stack otherwise if we return to the uncommon trap blob and the
1927 // stack bang causes a stack overflow we crash.
1928 JavaThread* jt = THREAD;
1929 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1930 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1931 }
1932 return;
1933 }
1934
1935 assert(!constant_pool->tag_at(index).is_symbol(),
1936 "no symbolic names here, please");
1937 }
1938
1939 #if INCLUDE_JFR
1940
1941 class DeoptReasonSerializer : public JfrSerializer {
1942 public:
1943 void serialize(JfrCheckpointWriter& writer) {
1944 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1945 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1946 writer.write_key((u8)i);
1947 writer.write(Deoptimization::trap_reason_name(i));
1948 }
1949 }
1950 };
1951
1952 class DeoptActionSerializer : public JfrSerializer {
1953 public:
1954 void serialize(JfrCheckpointWriter& writer) {
1955 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1956 writer.write_count(nof_actions);
1957 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1958 writer.write_key(i);
1959 writer.write(Deoptimization::trap_action_name((int)i));
1960 }
1961 }
1962 };
1963
1964 static void register_serializers() {
1965 static int critical_section = 0;
1966 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1967 return;
1968 }
1969 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1970 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1971 }
1972
1973 static void post_deoptimization_event(nmethod* nm,
1974 const Method* method,
1975 int trap_bci,
1976 int instruction,
1977 Deoptimization::DeoptReason reason,
1978 Deoptimization::DeoptAction action) {
1979 assert(nm != nullptr, "invariant");
1980 assert(method != nullptr, "invariant");
1981 if (EventDeoptimization::is_enabled()) {
1982 static bool serializers_registered = false;
1983 if (!serializers_registered) {
1984 register_serializers();
1985 serializers_registered = true;
1986 }
1987 EventDeoptimization event;
1988 event.set_compileId(nm->compile_id());
1989 event.set_compiler(nm->compiler_type());
1990 event.set_method(method);
1991 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1992 event.set_bci(trap_bci);
1993 event.set_instruction(instruction);
1994 event.set_reason(reason);
1995 event.set_action(action);
1996 event.commit();
1997 }
1998 }
1999
2000 #endif // INCLUDE_JFR
2001
2002 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2003 const char* reason_name, const char* reason_action) {
2004 LogTarget(Debug, deoptimization) lt;
2005 if (lt.is_enabled()) {
2006 LogStream ls(lt);
2007 bool is_osr = nm->is_osr_method();
2008 ls.print("cid=%4d %s level=%d",
2009 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
2010 ls.print(" %s", tm->name_and_sig_as_C_string());
2011 ls.print(" trap_bci=%d ", trap_bci);
2012 if (is_osr) {
2013 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2014 }
2015 ls.print("%s ", reason_name);
2016 ls.print("%s ", reason_action);
2017 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2018 pc, fr.pc() - nm->code_begin());
2019 }
2020 }
2021
2022 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2023 HandleMark hm(current);
2024
2025 // uncommon_trap() is called at the beginning of the uncommon trap
2026 // handler. Note this fact before we start generating temporary frames
2027 // that can confuse an asynchronous stack walker. This counter is
2028 // decremented at the end of unpack_frames().
2029
2030 current->inc_in_deopt_handler();
2031
2032 #if INCLUDE_JVMCI
2033 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2034 RegisterMap reg_map(current,
2035 RegisterMap::UpdateMap::include,
2036 RegisterMap::ProcessFrames::include,
2037 RegisterMap::WalkContinuation::skip);
2038 #else
2039 RegisterMap reg_map(current,
2040 RegisterMap::UpdateMap::skip,
2041 RegisterMap::ProcessFrames::include,
2042 RegisterMap::WalkContinuation::skip);
2043 #endif
2044 frame stub_frame = current->last_frame();
2045 frame fr = stub_frame.sender(®_map);
2046
2047 // Log a message
2048 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2049 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2050
2051 {
2052 ResourceMark rm;
2053
2054 DeoptReason reason = trap_request_reason(trap_request);
2055 DeoptAction action = trap_request_action(trap_request);
2056 #if INCLUDE_JVMCI
2057 int debug_id = trap_request_debug_id(trap_request);
2058 #endif
2059 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2060
2061 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2062 compiledVFrame* cvf = compiledVFrame::cast(vf);
2063
2064 nmethod* nm = cvf->code();
2065
2066 ScopeDesc* trap_scope = cvf->scope();
2067
2068 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2069
2070 if (is_receiver_constraint_failure) {
2071 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2072 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2073 JVMCI_ONLY(COMMA debug_id));
2074 }
2075
2076 methodHandle trap_method(current, trap_scope->method());
2077 int trap_bci = trap_scope->bci();
2078 #if INCLUDE_JVMCI
2079 jlong speculation = current->pending_failed_speculation();
2080 if (nm->is_compiled_by_jvmci()) {
2081 nm->update_speculation(current);
2082 } else {
2083 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2084 }
2085
2086 if (trap_bci == SynchronizationEntryBCI) {
2087 trap_bci = 0;
2088 current->set_pending_monitorenter(true);
2089 }
2090
2091 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2092 current->set_pending_transfer_to_interpreter(true);
2093 }
2094 #endif
2095
2096 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2097 // Record this event in the histogram.
2098 gather_statistics(reason, action, trap_bc);
2099
2100 // Ensure that we can record deopt. history:
2101 bool create_if_missing = ProfileTraps;
2102
2103 methodHandle profiled_method;
2104 #if INCLUDE_JVMCI
2105 if (nm->is_compiled_by_jvmci()) {
2106 profiled_method = methodHandle(current, nm->method());
2107 } else {
2108 profiled_method = trap_method;
2109 }
2110 #else
2111 profiled_method = trap_method;
2112 #endif
2113
2114 MethodData* trap_mdo =
2115 get_method_data(current, profiled_method, create_if_missing);
2116
2117 { // Log Deoptimization event for JFR, UL and event system
2118 Method* tm = trap_method();
2119 const char* reason_name = trap_reason_name(reason);
2120 const char* reason_action = trap_action_name(action);
2121 intptr_t pc = p2i(fr.pc());
2122
2123 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2124 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2125 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2126 reason_name, reason_action, pc,
2127 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2128 }
2129
2130 // Print a bunch of diagnostics, if requested.
2131 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2132 ResourceMark rm;
2133
2134 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2135 // We must do this already now, since we cannot acquire this lock while
2136 // holding the tty lock (lock ordering by rank).
2137 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2138
2139 ttyLocker ttyl;
2140
2141 char buf[100];
2142 if (xtty != nullptr) {
2143 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2144 os::current_thread_id(),
2145 format_trap_request(buf, sizeof(buf), trap_request));
2146 #if INCLUDE_JVMCI
2147 if (speculation != 0) {
2148 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2149 }
2150 #endif
2151 nm->log_identity(xtty);
2152 }
2153 Symbol* class_name = nullptr;
2154 bool unresolved = false;
2155 if (unloaded_class_index >= 0) {
2156 constantPoolHandle constants (current, trap_method->constants());
2157 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2158 class_name = constants->klass_name_at(unloaded_class_index);
2159 unresolved = true;
2160 if (xtty != nullptr)
2161 xtty->print(" unresolved='1'");
2162 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2163 class_name = constants->symbol_at(unloaded_class_index);
2164 }
2165 if (xtty != nullptr)
2166 xtty->name(class_name);
2167 }
2168 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2169 // Dump the relevant MDO state.
2170 // This is the deopt count for the current reason, any previous
2171 // reasons or recompiles seen at this point.
2172 int dcnt = trap_mdo->trap_count(reason);
2173 if (dcnt != 0)
2174 xtty->print(" count='%d'", dcnt);
2175
2176 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2177 // lock extra_data_lock before the tty lock.
2178 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2179 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2180 if (dos != 0) {
2181 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2182 if (trap_state_is_recompiled(dos)) {
2183 int recnt2 = trap_mdo->overflow_recompile_count();
2184 if (recnt2 != 0)
2185 xtty->print(" recompiles2='%d'", recnt2);
2186 }
2187 }
2188 }
2189 if (xtty != nullptr) {
2190 xtty->stamp();
2191 xtty->end_head();
2192 }
2193 if (TraceDeoptimization) { // make noise on the tty
2194 stringStream st;
2195 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2196 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2197 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2198 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2199 #if INCLUDE_JVMCI
2200 if (nm->is_compiled_by_jvmci()) {
2201 const char* installed_code_name = nm->jvmci_name();
2202 if (installed_code_name != nullptr) {
2203 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2204 }
2205 }
2206 #endif
2207 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2208 p2i(fr.pc()),
2209 os::current_thread_id(),
2210 trap_reason_name(reason),
2211 trap_action_name(action),
2212 unloaded_class_index
2213 #if INCLUDE_JVMCI
2214 , debug_id
2215 #endif
2216 );
2217 if (class_name != nullptr) {
2218 st.print(unresolved ? " unresolved class: " : " symbol: ");
2219 class_name->print_symbol_on(&st);
2220 }
2221 st.cr();
2222 tty->print_raw(st.freeze());
2223 }
2224 if (xtty != nullptr) {
2225 // Log the precise location of the trap.
2226 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2227 xtty->begin_elem("jvms bci='%d'", sd->bci());
2228 xtty->method(sd->method());
2229 xtty->end_elem();
2230 if (sd->is_top()) break;
2231 }
2232 xtty->tail("uncommon_trap");
2233 }
2234 }
2235 // (End diagnostic printout.)
2236
2237 if (is_receiver_constraint_failure) {
2238 fatal("missing receiver type check");
2239 }
2240
2241 // Load class if necessary
2242 if (unloaded_class_index >= 0) {
2243 constantPoolHandle constants(current, trap_method->constants());
2244 load_class_by_index(constants, unloaded_class_index, THREAD);
2245 }
2246
2247 // Flush the nmethod if necessary and desirable.
2248 //
2249 // We need to avoid situations where we are re-flushing the nmethod
2250 // because of a hot deoptimization site. Repeated flushes at the same
2251 // point need to be detected by the compiler and avoided. If the compiler
2252 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2253 // module must take measures to avoid an infinite cycle of recompilation
2254 // and deoptimization. There are several such measures:
2255 //
2256 // 1. If a recompilation is ordered a second time at some site X
2257 // and for the same reason R, the action is adjusted to 'reinterpret',
2258 // to give the interpreter time to exercise the method more thoroughly.
2259 // If this happens, the method's overflow_recompile_count is incremented.
2260 //
2261 // 2. If the compiler fails to reduce the deoptimization rate, then
2262 // the method's overflow_recompile_count will begin to exceed the set
2263 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2264 // is adjusted to 'make_not_compilable', and the method is abandoned
2265 // to the interpreter. This is a performance hit for hot methods,
2266 // but is better than a disastrous infinite cycle of recompilations.
2267 // (Actually, only the method containing the site X is abandoned.)
2268 //
2269 // 3. In parallel with the previous measures, if the total number of
2270 // recompilations of a method exceeds the much larger set limit
2271 // PerMethodRecompilationCutoff, the method is abandoned.
2272 // This should only happen if the method is very large and has
2273 // many "lukewarm" deoptimizations. The code which enforces this
2274 // limit is elsewhere (class nmethod, class Method).
2275 //
2276 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2277 // to recompile at each bytecode independently of the per-BCI cutoff.
2278 //
2279 // The decision to update code is up to the compiler, and is encoded
2280 // in the Action_xxx code. If the compiler requests Action_none
2281 // no trap state is changed, no compiled code is changed, and the
2282 // computation suffers along in the interpreter.
2283 //
2284 // The other action codes specify various tactics for decompilation
2285 // and recompilation. Action_maybe_recompile is the loosest, and
2286 // allows the compiled code to stay around until enough traps are seen,
2287 // and until the compiler gets around to recompiling the trapping method.
2288 //
2289 // The other actions cause immediate removal of the present code.
2290
2291 // Traps caused by injected profile shouldn't pollute trap counts.
2292 bool injected_profile_trap = trap_method->has_injected_profile() &&
2293 (reason == Reason_intrinsic || reason == Reason_unreached);
2294
2295 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2296 bool make_not_entrant = false;
2297 bool make_not_compilable = false;
2298 bool reprofile = false;
2299 switch (action) {
2300 case Action_none:
2301 // Keep the old code.
2302 update_trap_state = false;
2303 break;
2304 case Action_maybe_recompile:
2305 // Do not need to invalidate the present code, but we can
2306 // initiate another
2307 // Start compiler without (necessarily) invalidating the nmethod.
2308 // The system will tolerate the old code, but new code should be
2309 // generated when possible.
2310 break;
2311 case Action_reinterpret:
2312 // Go back into the interpreter for a while, and then consider
2313 // recompiling form scratch.
2314 make_not_entrant = true;
2315 // Reset invocation counter for outer most method.
2316 // This will allow the interpreter to exercise the bytecodes
2317 // for a while before recompiling.
2318 // By contrast, Action_make_not_entrant is immediate.
2319 //
2320 // Note that the compiler will track null_check, null_assert,
2321 // range_check, and class_check events and log them as if they
2322 // had been traps taken from compiled code. This will update
2323 // the MDO trap history so that the next compilation will
2324 // properly detect hot trap sites.
2325 reprofile = true;
2326 break;
2327 case Action_make_not_entrant:
2328 // Request immediate recompilation, and get rid of the old code.
2329 // Make them not entrant, so next time they are called they get
2330 // recompiled. Unloaded classes are loaded now so recompile before next
2331 // time they are called. Same for uninitialized. The interpreter will
2332 // link the missing class, if any.
2333 make_not_entrant = true;
2334 break;
2335 case Action_make_not_compilable:
2336 // Give up on compiling this method at all.
2337 make_not_entrant = true;
2338 make_not_compilable = true;
2339 break;
2340 default:
2341 ShouldNotReachHere();
2342 }
2343
2344 // Setting +ProfileTraps fixes the following, on all platforms:
2345 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2346 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2347 // recompile relies on a MethodData* to record heroic opt failures.
2348
2349 // Whether the interpreter is producing MDO data or not, we also need
2350 // to use the MDO to detect hot deoptimization points and control
2351 // aggressive optimization.
2352 bool inc_recompile_count = false;
2353
2354 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2355 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2356 (trap_mdo != nullptr),
2357 Mutex::_no_safepoint_check_flag);
2358 ProfileData* pdata = nullptr;
2359 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2360 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2361 uint this_trap_count = 0;
2362 bool maybe_prior_trap = false;
2363 bool maybe_prior_recompile = false;
2364
2365 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2366 #if INCLUDE_JVMCI
2367 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2368 #endif
2369 nm->method(),
2370 //outputs:
2371 this_trap_count,
2372 maybe_prior_trap,
2373 maybe_prior_recompile);
2374 // Because the interpreter also counts null, div0, range, and class
2375 // checks, these traps from compiled code are double-counted.
2376 // This is harmless; it just means that the PerXTrapLimit values
2377 // are in effect a little smaller than they look.
2378
2379 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2380 if (per_bc_reason != Reason_none) {
2381 // Now take action based on the partially known per-BCI history.
2382 if (maybe_prior_trap
2383 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2384 // If there are too many traps at this BCI, force a recompile.
2385 // This will allow the compiler to see the limit overflow, and
2386 // take corrective action, if possible. The compiler generally
2387 // does not use the exact PerBytecodeTrapLimit value, but instead
2388 // changes its tactics if it sees any traps at all. This provides
2389 // a little hysteresis, delaying a recompile until a trap happens
2390 // several times.
2391 //
2392 // Actually, since there is only one bit of counter per BCI,
2393 // the possible per-BCI counts are {0,1,(per-method count)}.
2394 // This produces accurate results if in fact there is only
2395 // one hot trap site, but begins to get fuzzy if there are
2396 // many sites. For example, if there are ten sites each
2397 // trapping two or more times, they each get the blame for
2398 // all of their traps.
2399 make_not_entrant = true;
2400 }
2401
2402 // Detect repeated recompilation at the same BCI, and enforce a limit.
2403 if (make_not_entrant && maybe_prior_recompile) {
2404 // More than one recompile at this point.
2405 inc_recompile_count = maybe_prior_trap;
2406 }
2407 } else {
2408 // For reasons which are not recorded per-bytecode, we simply
2409 // force recompiles unconditionally.
2410 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2411 make_not_entrant = true;
2412 }
2413
2414 // Go back to the compiler if there are too many traps in this method.
2415 if (this_trap_count >= per_method_trap_limit(reason)) {
2416 // If there are too many traps in this method, force a recompile.
2417 // This will allow the compiler to see the limit overflow, and
2418 // take corrective action, if possible.
2419 // (This condition is an unlikely backstop only, because the
2420 // PerBytecodeTrapLimit is more likely to take effect first,
2421 // if it is applicable.)
2422 make_not_entrant = true;
2423 }
2424
2425 // Here's more hysteresis: If there has been a recompile at
2426 // this trap point already, run the method in the interpreter
2427 // for a while to exercise it more thoroughly.
2428 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2429 reprofile = true;
2430 }
2431 }
2432
2433 // Take requested actions on the method:
2434
2435 // Recompile
2436 if (make_not_entrant) {
2437 if (!nm->make_not_entrant()) {
2438 return; // the call did not change nmethod's state
2439 }
2440
2441 if (pdata != nullptr) {
2442 // Record the recompilation event, if any.
2443 int tstate0 = pdata->trap_state();
2444 int tstate1 = trap_state_set_recompiled(tstate0, true);
2445 if (tstate1 != tstate0)
2446 pdata->set_trap_state(tstate1);
2447 }
2448
2449 // For code aging we count traps separately here, using make_not_entrant()
2450 // as a guard against simultaneous deopts in multiple threads.
2451 if (reason == Reason_tenured && trap_mdo != nullptr) {
2452 trap_mdo->inc_tenure_traps();
2453 }
2454 }
2455
2456 if (inc_recompile_count) {
2457 trap_mdo->inc_overflow_recompile_count();
2458 if ((uint)trap_mdo->overflow_recompile_count() >
2459 (uint)PerBytecodeRecompilationCutoff) {
2460 // Give up on the method containing the bad BCI.
2461 if (trap_method() == nm->method()) {
2462 make_not_compilable = true;
2463 } else {
2464 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2465 // But give grace to the enclosing nm->method().
2466 }
2467 }
2468 }
2469
2470 // Reprofile
2471 if (reprofile) {
2472 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2473 }
2474
2475 // Give up compiling
2476 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2477 assert(make_not_entrant, "consistent");
2478 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2479 }
2480
2481 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2482 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2483 if (exception_handler_data != nullptr) {
2484 // uncommon trap at the start of an exception handler.
2485 // C2 generates these for un-entered exception handlers.
2486 // mark the handler as entered to avoid generating
2487 // another uncommon trap the next time the handler is compiled
2488 exception_handler_data->set_exception_handler_entered();
2489 }
2490 }
2491
2492 } // Free marked resources
2493
2494 }
2495 JRT_END
2496
2497 ProfileData*
2498 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2499 int trap_bci,
2500 Deoptimization::DeoptReason reason,
2501 bool update_total_trap_count,
2502 #if INCLUDE_JVMCI
2503 bool is_osr,
2504 #endif
2505 Method* compiled_method,
2506 //outputs:
2507 uint& ret_this_trap_count,
2508 bool& ret_maybe_prior_trap,
2509 bool& ret_maybe_prior_recompile) {
2510 trap_mdo->check_extra_data_locked();
2511
2512 bool maybe_prior_trap = false;
2513 bool maybe_prior_recompile = false;
2514 uint this_trap_count = 0;
2515 if (update_total_trap_count) {
2516 uint idx = reason;
2517 #if INCLUDE_JVMCI
2518 if (is_osr) {
2519 // Upper half of history array used for traps in OSR compilations
2520 idx += Reason_TRAP_HISTORY_LENGTH;
2521 }
2522 #endif
2523 uint prior_trap_count = trap_mdo->trap_count(idx);
2524 this_trap_count = trap_mdo->inc_trap_count(idx);
2525
2526 // If the runtime cannot find a place to store trap history,
2527 // it is estimated based on the general condition of the method.
2528 // If the method has ever been recompiled, or has ever incurred
2529 // a trap with the present reason , then this BCI is assumed
2530 // (pessimistically) to be the culprit.
2531 maybe_prior_trap = (prior_trap_count != 0);
2532 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2533 }
2534 ProfileData* pdata = nullptr;
2535
2536
2537 // For reasons which are recorded per bytecode, we check per-BCI data.
2538 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2539 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2540 if (per_bc_reason != Reason_none) {
2541 // Find the profile data for this BCI. If there isn't one,
2542 // try to allocate one from the MDO's set of spares.
2543 // This will let us detect a repeated trap at this point.
2544 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2545
2546 if (pdata != nullptr) {
2547 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2548 if (LogCompilation && xtty != nullptr) {
2549 ttyLocker ttyl;
2550 // no more room for speculative traps in this MDO
2551 xtty->elem("speculative_traps_oom");
2552 }
2553 }
2554 // Query the trap state of this profile datum.
2555 int tstate0 = pdata->trap_state();
2556 if (!trap_state_has_reason(tstate0, per_bc_reason))
2557 maybe_prior_trap = false;
2558 if (!trap_state_is_recompiled(tstate0))
2559 maybe_prior_recompile = false;
2560
2561 // Update the trap state of this profile datum.
2562 int tstate1 = tstate0;
2563 // Record the reason.
2564 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2565 // Store the updated state on the MDO, for next time.
2566 if (tstate1 != tstate0)
2567 pdata->set_trap_state(tstate1);
2568 } else {
2569 if (LogCompilation && xtty != nullptr) {
2570 ttyLocker ttyl;
2571 // Missing MDP? Leave a small complaint in the log.
2572 xtty->elem("missing_mdp bci='%d'", trap_bci);
2573 }
2574 }
2575 }
2576
2577 // Return results:
2578 ret_this_trap_count = this_trap_count;
2579 ret_maybe_prior_trap = maybe_prior_trap;
2580 ret_maybe_prior_recompile = maybe_prior_recompile;
2581 return pdata;
2582 }
2583
2584 void
2585 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2586 ResourceMark rm;
2587 // Ignored outputs:
2588 uint ignore_this_trap_count;
2589 bool ignore_maybe_prior_trap;
2590 bool ignore_maybe_prior_recompile;
2591 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2592 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2593 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2594
2595 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2596 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2597
2598 query_update_method_data(trap_mdo, trap_bci,
2599 (DeoptReason)reason,
2600 update_total_counts,
2601 #if INCLUDE_JVMCI
2602 false,
2603 #endif
2604 nullptr,
2605 ignore_this_trap_count,
2606 ignore_maybe_prior_trap,
2607 ignore_maybe_prior_recompile);
2608 }
2609
2610 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2611 // Enable WXWrite: current function is called from methods compiled by C2 directly
2612 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2613
2614 // Still in Java no safepoints
2615 {
2616 // This enters VM and may safepoint
2617 uncommon_trap_inner(current, trap_request);
2618 }
2619 HandleMark hm(current);
2620 return fetch_unroll_info_helper(current, exec_mode);
2621 }
2622
2623 // Local derived constants.
2624 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2625 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2626 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2627
2628 //---------------------------trap_state_reason---------------------------------
2629 Deoptimization::DeoptReason
2630 Deoptimization::trap_state_reason(int trap_state) {
2631 // This assert provides the link between the width of DataLayout::trap_bits
2632 // and the encoding of "recorded" reasons. It ensures there are enough
2633 // bits to store all needed reasons in the per-BCI MDO profile.
2634 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2635 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2636 trap_state -= recompile_bit;
2637 if (trap_state == DS_REASON_MASK) {
2638 return Reason_many;
2639 } else {
2640 assert((int)Reason_none == 0, "state=0 => Reason_none");
2641 return (DeoptReason)trap_state;
2642 }
2643 }
2644 //-------------------------trap_state_has_reason-------------------------------
2645 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2646 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2647 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2648 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2649 trap_state -= recompile_bit;
2650 if (trap_state == DS_REASON_MASK) {
2651 return -1; // true, unspecifically (bottom of state lattice)
2652 } else if (trap_state == reason) {
2653 return 1; // true, definitely
2654 } else if (trap_state == 0) {
2655 return 0; // false, definitely (top of state lattice)
2656 } else {
2657 return 0; // false, definitely
2658 }
2659 }
2660 //-------------------------trap_state_add_reason-------------------------------
2661 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2662 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2663 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2664 trap_state -= recompile_bit;
2665 if (trap_state == DS_REASON_MASK) {
2666 return trap_state + recompile_bit; // already at state lattice bottom
2667 } else if (trap_state == reason) {
2668 return trap_state + recompile_bit; // the condition is already true
2669 } else if (trap_state == 0) {
2670 return reason + recompile_bit; // no condition has yet been true
2671 } else {
2672 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2673 }
2674 }
2675 //-----------------------trap_state_is_recompiled------------------------------
2676 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2677 return (trap_state & DS_RECOMPILE_BIT) != 0;
2678 }
2679 //-----------------------trap_state_set_recompiled-----------------------------
2680 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2681 if (z) return trap_state | DS_RECOMPILE_BIT;
2682 else return trap_state & ~DS_RECOMPILE_BIT;
2683 }
2684 //---------------------------format_trap_state---------------------------------
2685 // This is used for debugging and diagnostics, including LogFile output.
2686 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2687 int trap_state) {
2688 assert(buflen > 0, "sanity");
2689 DeoptReason reason = trap_state_reason(trap_state);
2690 bool recomp_flag = trap_state_is_recompiled(trap_state);
2691 // Re-encode the state from its decoded components.
2692 int decoded_state = 0;
2693 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2694 decoded_state = trap_state_add_reason(decoded_state, reason);
2695 if (recomp_flag)
2696 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2697 // If the state re-encodes properly, format it symbolically.
2698 // Because this routine is used for debugging and diagnostics,
2699 // be robust even if the state is a strange value.
2700 size_t len;
2701 if (decoded_state != trap_state) {
2702 // Random buggy state that doesn't decode??
2703 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2704 } else {
2705 len = jio_snprintf(buf, buflen, "%s%s",
2706 trap_reason_name(reason),
2707 recomp_flag ? " recompiled" : "");
2708 }
2709 return buf;
2710 }
2711
2712
2713 //--------------------------------statics--------------------------------------
2714 const char* Deoptimization::_trap_reason_name[] = {
2715 // Note: Keep this in sync. with enum DeoptReason.
2716 "none",
2717 "null_check",
2718 "null_assert" JVMCI_ONLY("_or_unreached0"),
2719 "range_check",
2720 "class_check",
2721 "array_check",
2722 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2723 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2724 "profile_predicate",
2725 "unloaded",
2726 "uninitialized",
2727 "initialized",
2728 "unreached",
2729 "unhandled",
2730 "constraint",
2731 "div0_check",
2732 "age",
2733 "predicate",
2734 "loop_limit_check",
2735 "speculate_class_check",
2736 "speculate_null_check",
2737 "speculate_null_assert",
2738 "unstable_if",
2739 "unstable_fused_if",
2740 "receiver_constraint",
2741 #if INCLUDE_JVMCI
2742 "aliasing",
2743 "transfer_to_interpreter",
2744 "not_compiled_exception_handler",
2745 "unresolved",
2746 "jsr_mismatch",
2747 #endif
2748 "tenured"
2749 };
2750 const char* Deoptimization::_trap_action_name[] = {
2751 // Note: Keep this in sync. with enum DeoptAction.
2752 "none",
2753 "maybe_recompile",
2754 "reinterpret",
2755 "make_not_entrant",
2756 "make_not_compilable"
2757 };
2758
2759 const char* Deoptimization::trap_reason_name(int reason) {
2760 // Check that every reason has a name
2761 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2762
2763 if (reason == Reason_many) return "many";
2764 if ((uint)reason < Reason_LIMIT)
2765 return _trap_reason_name[reason];
2766 static char buf[20];
2767 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2768 return buf;
2769 }
2770 const char* Deoptimization::trap_action_name(int action) {
2771 // Check that every action has a name
2772 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2773
2774 if ((uint)action < Action_LIMIT)
2775 return _trap_action_name[action];
2776 static char buf[20];
2777 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2778 return buf;
2779 }
2780
2781 // This is used for debugging and diagnostics, including LogFile output.
2782 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2783 int trap_request) {
2784 jint unloaded_class_index = trap_request_index(trap_request);
2785 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2786 const char* action = trap_action_name(trap_request_action(trap_request));
2787 #if INCLUDE_JVMCI
2788 int debug_id = trap_request_debug_id(trap_request);
2789 #endif
2790 size_t len;
2791 if (unloaded_class_index < 0) {
2792 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2793 reason, action
2794 #if INCLUDE_JVMCI
2795 ,debug_id
2796 #endif
2797 );
2798 } else {
2799 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2800 reason, action, unloaded_class_index
2801 #if INCLUDE_JVMCI
2802 ,debug_id
2803 #endif
2804 );
2805 }
2806 return buf;
2807 }
2808
2809 juint Deoptimization::_deoptimization_hist
2810 [Deoptimization::Reason_LIMIT]
2811 [1 + Deoptimization::Action_LIMIT]
2812 [Deoptimization::BC_CASE_LIMIT]
2813 = {0};
2814
2815 enum {
2816 LSB_BITS = 8,
2817 LSB_MASK = right_n_bits(LSB_BITS)
2818 };
2819
2820 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2821 Bytecodes::Code bc) {
2822 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2823 assert(action >= 0 && action < Action_LIMIT, "oob");
2824 _deoptimization_hist[Reason_none][0][0] += 1; // total
2825 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2826 juint* cases = _deoptimization_hist[reason][1+action];
2827 juint* bc_counter_addr = nullptr;
2828 juint bc_counter = 0;
2829 // Look for an unused counter, or an exact match to this BC.
2830 if (bc != Bytecodes::_illegal) {
2831 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2832 juint* counter_addr = &cases[bc_case];
2833 juint counter = *counter_addr;
2834 if ((counter == 0 && bc_counter_addr == nullptr)
2835 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2836 // this counter is either free or is already devoted to this BC
2837 bc_counter_addr = counter_addr;
2838 bc_counter = counter | bc;
2839 }
2840 }
2841 }
2842 if (bc_counter_addr == nullptr) {
2843 // Overflow, or no given bytecode.
2844 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2845 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2846 }
2847 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2848 }
2849
2850 jint Deoptimization::total_deoptimization_count() {
2851 return _deoptimization_hist[Reason_none][0][0];
2852 }
2853
2854 // Get the deopt count for a specific reason and a specific action. If either
2855 // one of 'reason' or 'action' is null, the method returns the sum of all
2856 // deoptimizations with the specific 'action' or 'reason' respectively.
2857 // If both arguments are null, the method returns the total deopt count.
2858 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2859 if (reason_str == nullptr && action_str == nullptr) {
2860 return total_deoptimization_count();
2861 }
2862 juint counter = 0;
2863 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2864 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2865 for (int action = 0; action < Action_LIMIT; action++) {
2866 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2867 juint* cases = _deoptimization_hist[reason][1+action];
2868 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2869 counter += cases[bc_case] >> LSB_BITS;
2870 }
2871 }
2872 }
2873 }
2874 }
2875 return counter;
2876 }
2877
2878 void Deoptimization::print_statistics() {
2879 juint total = total_deoptimization_count();
2880 juint account = total;
2881 if (total != 0) {
2882 ttyLocker ttyl;
2883 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2884 tty->print_cr("Deoptimization traps recorded:");
2885 #define PRINT_STAT_LINE(name, r) \
2886 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2887 PRINT_STAT_LINE("total", total);
2888 // For each non-zero entry in the histogram, print the reason,
2889 // the action, and (if specifically known) the type of bytecode.
2890 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2891 for (int action = 0; action < Action_LIMIT; action++) {
2892 juint* cases = _deoptimization_hist[reason][1+action];
2893 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2894 juint counter = cases[bc_case];
2895 if (counter != 0) {
2896 char name[1*K];
2897 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2898 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2899 bc = Bytecodes::_illegal;
2900 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
2901 trap_reason_name(reason),
2902 trap_action_name(action),
2903 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2904 juint r = counter >> LSB_BITS;
2905 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2906 account -= r;
2907 }
2908 }
2909 }
2910 }
2911 if (account != 0) {
2912 PRINT_STAT_LINE("unaccounted", account);
2913 }
2914 #undef PRINT_STAT_LINE
2915 if (xtty != nullptr) xtty->tail("statistics");
2916 }
2917 }
2918
2919 #else // COMPILER2_OR_JVMCI
2920
2921
2922 // Stubs for C1 only system.
2923 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2924 return false;
2925 }
2926
2927 const char* Deoptimization::trap_reason_name(int reason) {
2928 return "unknown";
2929 }
2930
2931 jint Deoptimization::total_deoptimization_count() {
2932 return 0;
2933 }
2934
2935 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2936 return 0;
2937 }
2938
2939 void Deoptimization::print_statistics() {
2940 // no output
2941 }
2942
2943 void
2944 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2945 // no update
2946 }
2947
2948 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2949 return 0;
2950 }
2951
2952 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2953 Bytecodes::Code bc) {
2954 // no update
2955 }
2956
2957 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2958 int trap_state) {
2959 jio_snprintf(buf, buflen, "#%d", trap_state);
2960 return buf;
2961 }
2962
2963 #endif // COMPILER2_OR_JVMCI