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 "interpreter/bytecode.hpp"
39 #include "interpreter/interpreter.hpp"
40 #include "interpreter/oopMapCache.hpp"
41 #include "jvm.h"
42 #include "logging/log.hpp"
43 #include "logging/logLevel.hpp"
44 #include "logging/logMessage.hpp"
45 #include "logging/logStream.hpp"
46 #include "memory/allocation.inline.hpp"
47 #include "memory/oopFactory.hpp"
48 #include "memory/resourceArea.hpp"
49 #include "memory/universe.hpp"
50 #include "oops/constantPool.hpp"
51 #include "oops/fieldStreams.inline.hpp"
52 #include "oops/method.hpp"
53 #include "oops/objArrayKlass.hpp"
54 #include "oops/objArrayOop.inline.hpp"
55 #include "oops/oop.inline.hpp"
56 #include "oops/typeArrayOop.inline.hpp"
57 #include "oops/verifyOopClosure.hpp"
58 #include "prims/jvmtiDeferredUpdates.hpp"
59 #include "prims/jvmtiExport.hpp"
60 #include "prims/jvmtiThreadState.hpp"
61 #include "prims/methodHandles.hpp"
62 #include "prims/vectorSupport.hpp"
63 #include "runtime/atomic.hpp"
64 #include "runtime/continuation.hpp"
65 #include "runtime/continuationEntry.inline.hpp"
66 #include "runtime/deoptimization.hpp"
67 #include "runtime/escapeBarrier.hpp"
68 #include "runtime/fieldDescriptor.hpp"
69 #include "runtime/fieldDescriptor.inline.hpp"
70 #include "runtime/frame.inline.hpp"
71 #include "runtime/handles.inline.hpp"
72 #include "runtime/interfaceSupport.inline.hpp"
73 #include "runtime/javaThread.hpp"
74 #include "runtime/jniHandles.inline.hpp"
75 #include "runtime/keepStackGCProcessed.hpp"
76 #include "runtime/objectMonitor.inline.hpp"
77 #include "runtime/osThread.hpp"
78 #include "runtime/safepointVerifiers.hpp"
79 #include "runtime/sharedRuntime.hpp"
80 #include "runtime/signature.hpp"
81 #include "runtime/stackFrameStream.inline.hpp"
82 #include "runtime/stackValue.hpp"
83 #include "runtime/stackWatermarkSet.hpp"
84 #include "runtime/stubRoutines.hpp"
85 #include "runtime/synchronizer.hpp"
86 #include "runtime/threadSMR.hpp"
87 #include "runtime/threadWXSetters.inline.hpp"
88 #include "runtime/vframe.hpp"
89 #include "runtime/vframeArray.hpp"
90 #include "runtime/vframe_hp.hpp"
91 #include "runtime/vmOperations.hpp"
92 #include "utilities/events.hpp"
93 #include "utilities/growableArray.hpp"
94 #include "utilities/macros.hpp"
95 #include "utilities/preserveException.hpp"
96 #include "utilities/xmlstream.hpp"
97 #if INCLUDE_JFR
98 #include "jfr/jfrEvents.hpp"
99 #include "jfr/metadata/jfrSerializer.hpp"
100 #endif
101
102 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
103 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
104 bool DeoptimizationScope::_committing_in_progress = false;
105
106 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
107 DEBUG_ONLY(_deopted = false;)
108
109 MutexLocker ml(CompiledMethod_lock, Mutex::_no_safepoint_check_flag);
110 // If there is nothing to deopt _required_gen is the same as comitted.
111 _required_gen = DeoptimizationScope::_committed_deopt_gen;
112 }
113
114 DeoptimizationScope::~DeoptimizationScope() {
115 assert(_deopted, "Deopt not executed");
116 }
117
118 void DeoptimizationScope::mark(CompiledMethod* cm, bool inc_recompile_counts) {
119 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock,
120 Mutex::_no_safepoint_check_flag);
121
122 // If it's already marked but we still need it to be deopted.
123 if (cm->is_marked_for_deoptimization()) {
124 dependent(cm);
125 return;
126 }
127
128 CompiledMethod::DeoptimizationStatus status =
129 inc_recompile_counts ? CompiledMethod::deoptimize : CompiledMethod::deoptimize_noupdate;
130 Atomic::store(&cm->_deoptimization_status, status);
131
132 // Make sure active is not committed
133 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
134 assert(cm->_deoptimization_generation == 0, "Is already marked");
135
136 cm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
137 _required_gen = DeoptimizationScope::_active_deopt_gen;
138 }
139
140 void DeoptimizationScope::dependent(CompiledMethod* cm) {
141 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock,
142 Mutex::_no_safepoint_check_flag);
143 // A method marked by someone else may have a _required_gen lower than what we marked with.
144 // Therefore only store it if it's higher than _required_gen.
145 if (_required_gen < cm->_deoptimization_generation) {
146 _required_gen = cm->_deoptimization_generation;
147 }
148 }
149
150 void DeoptimizationScope::deoptimize_marked() {
151 assert(!_deopted, "Already deopted");
152
153 // We are not alive yet.
154 if (!Universe::is_fully_initialized()) {
155 DEBUG_ONLY(_deopted = true;)
156 return;
157 }
158
159 // Safepoints are a special case, handled here.
160 if (SafepointSynchronize::is_at_safepoint()) {
161 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
162 DeoptimizationScope::_active_deopt_gen++;
163 Deoptimization::deoptimize_all_marked();
164 DEBUG_ONLY(_deopted = true;)
165 return;
166 }
167
168 uint64_t comitting = 0;
169 bool wait = false;
170 while (true) {
171 {
172 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock,
173 Mutex::_no_safepoint_check_flag);
174 // First we check if we or someone else already deopted the gen we want.
175 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
176 DEBUG_ONLY(_deopted = true;)
177 return;
178 }
179 if (!_committing_in_progress) {
180 // The version we are about to commit.
181 comitting = DeoptimizationScope::_active_deopt_gen;
182 // Make sure new marks use a higher gen.
183 DeoptimizationScope::_active_deopt_gen++;
184 _committing_in_progress = true;
185 wait = false;
186 } else {
187 // Another thread is handshaking and committing a gen.
188 wait = true;
189 }
190 }
191 if (wait) {
192 // Wait and let the concurrent handshake be performed.
193 ThreadBlockInVM tbivm(JavaThread::current());
194 os::naked_yield();
195 } else {
196 // Performs the handshake.
197 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
198 DEBUG_ONLY(_deopted = true;)
199 {
200 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock,
201 Mutex::_no_safepoint_check_flag);
202 // Make sure that committed doesn't go backwards.
203 // Should only happen if we did a deopt during a safepoint above.
204 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
205 DeoptimizationScope::_committed_deopt_gen = comitting;
206 }
207 _committing_in_progress = false;
208
209 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
210
211 return;
212 }
213 }
214 }
215 }
216
217 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
218 int caller_adjustment,
219 int caller_actual_parameters,
220 int number_of_frames,
221 intptr_t* frame_sizes,
222 address* frame_pcs,
223 BasicType return_type,
224 int exec_mode) {
225 _size_of_deoptimized_frame = size_of_deoptimized_frame;
226 _caller_adjustment = caller_adjustment;
227 _caller_actual_parameters = caller_actual_parameters;
228 _number_of_frames = number_of_frames;
229 _frame_sizes = frame_sizes;
230 _frame_pcs = frame_pcs;
231 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
232 _return_type = return_type;
233 _initial_info = 0;
234 // PD (x86 only)
235 _counter_temp = 0;
236 _unpack_kind = exec_mode;
237 _sender_sp_temp = 0;
238
239 _total_frame_sizes = size_of_frames();
240 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
241 }
242
243 Deoptimization::UnrollBlock::~UnrollBlock() {
244 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
245 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
246 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
247 }
248
249 int Deoptimization::UnrollBlock::size_of_frames() const {
250 // Account first for the adjustment of the initial frame
251 int result = _caller_adjustment;
252 for (int index = 0; index < number_of_frames(); index++) {
253 result += frame_sizes()[index];
254 }
255 return result;
256 }
257
258 void Deoptimization::UnrollBlock::print() {
259 ResourceMark rm;
260 stringStream st;
261 st.print_cr("UnrollBlock");
262 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
263 st.print( " frame_sizes: ");
264 for (int index = 0; index < number_of_frames(); index++) {
265 st.print(INTX_FORMAT " ", frame_sizes()[index]);
266 }
267 st.cr();
268 tty->print_raw(st.freeze());
269 }
270
271 // In order to make fetch_unroll_info work properly with escape
272 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
273 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
274 // which is called from the method fetch_unroll_info_helper below.
275 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
276 // fetch_unroll_info() is called at the beginning of the deoptimization
277 // handler. Note this fact before we start generating temporary frames
278 // that can confuse an asynchronous stack walker. This counter is
279 // decremented at the end of unpack_frames().
280 current->inc_in_deopt_handler();
281
282 if (exec_mode == Unpack_exception) {
283 // When we get here, a callee has thrown an exception into a deoptimized
284 // frame. That throw might have deferred stack watermark checking until
285 // after unwinding. So we deal with such deferred requests here.
286 StackWatermarkSet::after_unwind(current);
287 }
288
289 return fetch_unroll_info_helper(current, exec_mode);
290 JRT_END
291
292 #if COMPILER2_OR_JVMCI
293 // print information about reallocated objects
294 static void print_objects(JavaThread* deoptee_thread,
295 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
296 ResourceMark rm;
297 stringStream st; // change to logStream with logging
298 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
299 fieldDescriptor fd;
300
301 for (int i = 0; i < objects->length(); i++) {
302 ObjectValue* sv = (ObjectValue*) objects->at(i);
303 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
304 Handle obj = sv->value();
305
306 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
307 k->print_value_on(&st);
308 assert(obj.not_null() || realloc_failures, "reallocation was missed");
309 if (obj.is_null()) {
310 st.print(" allocation failed");
311 } else {
312 st.print(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize);
313 }
314 st.cr();
315
316 if (Verbose && !obj.is_null()) {
317 k->oop_print_on(obj(), &st);
318 }
319 }
320 tty->print_raw(st.freeze());
321 }
322
323 static bool rematerialize_objects(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method,
324 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
325 bool& deoptimized_objects) {
326 bool realloc_failures = false;
327 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
328
329 JavaThread* deoptee_thread = chunk->at(0)->thread();
330 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
331 "a frame can only be deoptimized by the owner thread");
332
333 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
334
335 // The flag return_oop() indicates call sites which return oop
336 // in compiled code. Such sites include java method calls,
337 // runtime calls (for example, used to allocate new objects/arrays
338 // on slow code path) and any other calls generated in compiled code.
339 // It is not guaranteed that we can get such information here only
340 // by analyzing bytecode in deoptimized frames. This is why this flag
341 // is set during method compilation (see Compile::Process_OopMap_Node()).
342 // If the previous frame was popped or if we are dispatching an exception,
343 // we don't have an oop result.
344 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
345 Handle return_value;
346 if (save_oop_result) {
347 // Reallocation may trigger GC. If deoptimization happened on return from
348 // call which returns oop we need to save it since it is not in oopmap.
349 oop result = deoptee.saved_oop_result(&map);
350 assert(oopDesc::is_oop_or_null(result), "must be oop");
351 return_value = Handle(thread, result);
352 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
353 if (TraceDeoptimization) {
354 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
355 tty->cr();
356 }
357 }
358 if (objects != nullptr) {
359 if (exec_mode == Deoptimization::Unpack_none) {
360 assert(thread->thread_state() == _thread_in_vm, "assumption");
361 JavaThread* THREAD = thread; // For exception macros.
362 // Clear pending OOM if reallocation fails and return true indicating allocation failure
363 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
364 deoptimized_objects = true;
365 } else {
366 JavaThread* current = thread; // For JRT_BLOCK
367 JRT_BLOCK
368 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
369 JRT_END
370 }
371 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci();
372 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
373 if (TraceDeoptimization) {
374 print_objects(deoptee_thread, objects, realloc_failures);
375 }
376 }
377 if (save_oop_result) {
378 // Restore result.
379 deoptee.set_saved_oop_result(&map, return_value());
380 }
381 return realloc_failures;
382 }
383
384 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
385 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
386 JavaThread* deoptee_thread = chunk->at(0)->thread();
387 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
388 assert(thread == Thread::current(), "should be");
389 HandleMark hm(thread);
390 #ifndef PRODUCT
391 bool first = true;
392 #endif // !PRODUCT
393 for (int i = 0; i < chunk->length(); i++) {
394 compiledVFrame* cvf = chunk->at(i);
395 assert (cvf->scope() != nullptr,"expect only compiled java frames");
396 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
397 if (monitors->is_nonempty()) {
398 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
399 exec_mode, realloc_failures);
400 deoptimized_objects = deoptimized_objects || relocked;
401 #ifndef PRODUCT
402 if (PrintDeoptimizationDetails) {
403 ResourceMark rm;
404 stringStream st;
405 for (int j = 0; j < monitors->length(); j++) {
406 MonitorInfo* mi = monitors->at(j);
407 if (mi->eliminated()) {
408 if (first) {
409 first = false;
410 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
411 }
412 if (exec_mode == Deoptimization::Unpack_none) {
413 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
414 if (monitor != nullptr && monitor->object() == mi->owner()) {
415 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
416 continue;
417 }
418 }
419 if (mi->owner_is_scalar_replaced()) {
420 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
421 st.print_cr(" failed reallocation for klass %s", k->external_name());
422 } else {
423 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
424 }
425 }
426 }
427 tty->print_raw(st.freeze());
428 }
429 #endif // !PRODUCT
430 }
431 }
432 }
433
434 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
435 // The given vframes cover one physical frame.
436 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
437 bool& realloc_failures) {
438 frame deoptee = chunk->at(0)->fr();
439 JavaThread* deoptee_thread = chunk->at(0)->thread();
440 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
441 RegisterMap map(chunk->at(0)->register_map());
442 bool deoptimized_objects = false;
443
444 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false);
445
446 // Reallocate the non-escaping objects and restore their fields.
447 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
448 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
449 realloc_failures = rematerialize_objects(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects);
450 }
451
452 // MonitorInfo structures used in eliminate_locks are not GC safe.
453 NoSafepointVerifier no_safepoint;
454
455 // Now relock objects if synchronization on them was eliminated.
456 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
457 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
458 }
459 return deoptimized_objects;
460 }
461 #endif // COMPILER2_OR_JVMCI
462
463 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
464 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
465 // When we get here we are about to unwind the deoptee frame. In order to
466 // catch not yet safe to use frames, the following stack watermark barrier
467 // poll will make such frames safe to use.
468 StackWatermarkSet::before_unwind(current);
469
470 // Note: there is a safepoint safety issue here. No matter whether we enter
471 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
472 // the vframeArray is created.
473 //
474
475 // Allocate our special deoptimization ResourceMark
476 DeoptResourceMark* dmark = new DeoptResourceMark(current);
477 assert(current->deopt_mark() == nullptr, "Pending deopt!");
478 current->set_deopt_mark(dmark);
479
480 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
481 RegisterMap map(current,
482 RegisterMap::UpdateMap::include,
483 RegisterMap::ProcessFrames::include,
484 RegisterMap::WalkContinuation::skip);
485 RegisterMap dummy_map(current,
486 RegisterMap::UpdateMap::skip,
487 RegisterMap::ProcessFrames::include,
488 RegisterMap::WalkContinuation::skip);
489 // Now get the deoptee with a valid map
490 frame deoptee = stub_frame.sender(&map);
491 // Set the deoptee nmethod
492 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
493 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
494 current->set_deopt_compiled_method(cm);
495
496 if (VerifyStack) {
497 current->validate_frame_layout();
498 }
499
500 // Create a growable array of VFrames where each VFrame represents an inlined
501 // Java frame. This storage is allocated with the usual system arena.
502 assert(deoptee.is_compiled_frame(), "Wrong frame type");
503 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
504 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
505 while (!vf->is_top()) {
506 assert(vf->is_compiled_frame(), "Wrong frame type");
507 chunk->push(compiledVFrame::cast(vf));
508 vf = vf->sender();
509 }
510 assert(vf->is_compiled_frame(), "Wrong frame type");
511 chunk->push(compiledVFrame::cast(vf));
512
513 bool realloc_failures = false;
514
515 #if COMPILER2_OR_JVMCI
516 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
517
518 // Reallocate the non-escaping objects and restore their fields. Then
519 // relock objects if synchronization on them was eliminated.
520 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
521 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
522 bool unused;
523 realloc_failures = rematerialize_objects(current, exec_mode, cm, deoptee, map, chunk, unused);
524 }
525 #endif // COMPILER2_OR_JVMCI
526
527 // Ensure that no safepoint is taken after pointers have been stored
528 // in fields of rematerialized objects. If a safepoint occurs from here on
529 // out the java state residing in the vframeArray will be missed.
530 // Locks may be rebaised in a safepoint.
531 NoSafepointVerifier no_safepoint;
532
533 #if COMPILER2_OR_JVMCI
534 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
535 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
536 bool unused;
537 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
538 }
539 #endif // COMPILER2_OR_JVMCI
540
541 ScopeDesc* trap_scope = chunk->at(0)->scope();
542 Handle exceptionObject;
543 if (trap_scope->rethrow_exception()) {
544 #ifndef PRODUCT
545 if (PrintDeoptimizationDetails) {
546 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());
547 }
548 #endif // !PRODUCT
549
550 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
551 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
552 ScopeValue* topOfStack = expressions->top();
553 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
554 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
555 }
556
557 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
558 #if COMPILER2_OR_JVMCI
559 if (realloc_failures) {
560 // This destroys all ScopedValue bindings.
561 current->clear_scopedValueBindings();
562 pop_frames_failed_reallocs(current, array);
563 }
564 #endif
565
566 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
567 current->set_vframe_array_head(array);
568
569 // Now that the vframeArray has been created if we have any deferred local writes
570 // added by jvmti then we can free up that structure as the data is now in the
571 // vframeArray
572
573 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
574
575 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
576 CodeBlob* cb = stub_frame.cb();
577 // Verify we have the right vframeArray
578 assert(cb->frame_size() >= 0, "Unexpected frame size");
579 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
580
581 // If the deopt call site is a MethodHandle invoke call site we have
582 // to adjust the unpack_sp.
583 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
584 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
585 unpack_sp = deoptee.unextended_sp();
586
587 #ifdef ASSERT
588 assert(cb->is_deoptimization_stub() ||
589 cb->is_uncommon_trap_stub() ||
590 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
591 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
592 "unexpected code blob: %s", cb->name());
593 #endif
594
595 // This is a guarantee instead of an assert because if vframe doesn't match
596 // we will unpack the wrong deoptimized frame and wind up in strange places
597 // where it will be very difficult to figure out what went wrong. Better
598 // to die an early death here than some very obscure death later when the
599 // trail is cold.
600 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
601 // in that it will fail to detect a problem when there is one. This needs
602 // more work in tiger timeframe.
603 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
604
605 int number_of_frames = array->frames();
606
607 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
608 // virtual activation, which is the reverse of the elements in the vframes array.
609 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
610 // +1 because we always have an interpreter return address for the final slot.
611 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
612 int popframe_extra_args = 0;
613 // Create an interpreter return address for the stub to use as its return
614 // address so the skeletal frames are perfectly walkable
615 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
616
617 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
618 // activation be put back on the expression stack of the caller for reexecution
619 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
620 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
621 }
622
623 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
624 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
625 // than simply use array->sender.pc(). This requires us to walk the current set of frames
626 //
627 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
628 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
629
630 // It's possible that the number of parameters at the call site is
631 // different than number of arguments in the callee when method
632 // handles are used. If the caller is interpreted get the real
633 // value so that the proper amount of space can be added to it's
634 // frame.
635 bool caller_was_method_handle = false;
636 if (deopt_sender.is_interpreted_frame()) {
637 methodHandle method(current, deopt_sender.interpreter_frame_method());
638 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
639 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
640 // Method handle invokes may involve fairly arbitrary chains of
641 // calls so it's impossible to know how much actual space the
642 // caller has for locals.
643 caller_was_method_handle = true;
644 }
645 }
646
647 //
648 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
649 // frame_sizes/frame_pcs[1] next oldest frame (int)
650 // frame_sizes/frame_pcs[n] youngest frame (int)
651 //
652 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
653 // owns the space for the return address to it's caller). Confusing ain't it.
654 //
655 // The vframe array can address vframes with indices running from
656 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
657 // When we create the skeletal frames we need the oldest frame to be in the zero slot
658 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
659 // so things look a little strange in this loop.
660 //
661 int callee_parameters = 0;
662 int callee_locals = 0;
663 for (int index = 0; index < array->frames(); index++ ) {
664 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
665 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
666 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
667 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
668 callee_locals,
669 index == 0,
670 popframe_extra_args);
671 // This pc doesn't have to be perfect just good enough to identify the frame
672 // as interpreted so the skeleton frame will be walkable
673 // The correct pc will be set when the skeleton frame is completely filled out
674 // The final pc we store in the loop is wrong and will be overwritten below
675 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
676
677 callee_parameters = array->element(index)->method()->size_of_parameters();
678 callee_locals = array->element(index)->method()->max_locals();
679 popframe_extra_args = 0;
680 }
681
682 // Compute whether the root vframe returns a float or double value.
683 BasicType return_type;
684 {
685 methodHandle method(current, array->element(0)->method());
686 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
687 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
688 }
689
690 // Compute information for handling adapters and adjusting the frame size of the caller.
691 int caller_adjustment = 0;
692
693 // Compute the amount the oldest interpreter frame will have to adjust
694 // its caller's stack by. If the caller is a compiled frame then
695 // we pretend that the callee has no parameters so that the
696 // extension counts for the full amount of locals and not just
697 // locals-parms. This is because without a c2i adapter the parm
698 // area as created by the compiled frame will not be usable by
699 // the interpreter. (Depending on the calling convention there
700 // may not even be enough space).
701
702 // QQQ I'd rather see this pushed down into last_frame_adjust
703 // and have it take the sender (aka caller).
704
705 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
706 caller_adjustment = last_frame_adjust(0, callee_locals);
707 } else if (callee_locals > callee_parameters) {
708 // The caller frame may need extending to accommodate
709 // non-parameter locals of the first unpacked interpreted frame.
710 // Compute that adjustment.
711 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
712 }
713
714 // If the sender is deoptimized the we must retrieve the address of the handler
715 // since the frame will "magically" show the original pc before the deopt
716 // and we'd undo the deopt.
717
718 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
719 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
720 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
721 }
722
723 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
724
725 #if INCLUDE_JVMCI
726 if (exceptionObject() != nullptr) {
727 current->set_exception_oop(exceptionObject());
728 exec_mode = Unpack_exception;
729 }
730 #endif
731
732 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
733 assert(current->has_pending_exception(), "should have thrown OOME");
734 current->set_exception_oop(current->pending_exception());
735 current->clear_pending_exception();
736 exec_mode = Unpack_exception;
737 }
738
739 #if INCLUDE_JVMCI
740 if (current->frames_to_pop_failed_realloc() > 0) {
741 current->set_pending_monitorenter(false);
742 }
743 #endif
744
745 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
746 caller_adjustment * BytesPerWord,
747 caller_was_method_handle ? 0 : callee_parameters,
748 number_of_frames,
749 frame_sizes,
750 frame_pcs,
751 return_type,
752 exec_mode);
753 // On some platforms, we need a way to pass some platform dependent
754 // information to the unpacking code so the skeletal frames come out
755 // correct (initial fp value, unextended sp, ...)
756 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
757
758 if (array->frames() > 1) {
759 if (VerifyStack && TraceDeoptimization) {
760 tty->print_cr("Deoptimizing method containing inlining");
761 }
762 }
763
764 array->set_unroll_block(info);
765 return info;
766 }
767
768 // Called to cleanup deoptimization data structures in normal case
769 // after unpacking to stack and when stack overflow error occurs
770 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
771 vframeArray *array) {
772
773 // Get array if coming from exception
774 if (array == nullptr) {
775 array = thread->vframe_array_head();
776 }
777 thread->set_vframe_array_head(nullptr);
778
779 // Free the previous UnrollBlock
780 vframeArray* old_array = thread->vframe_array_last();
781 thread->set_vframe_array_last(array);
782
783 if (old_array != nullptr) {
784 UnrollBlock* old_info = old_array->unroll_block();
785 old_array->set_unroll_block(nullptr);
786 delete old_info;
787 delete old_array;
788 }
789
790 // Deallocate any resource creating in this routine and any ResourceObjs allocated
791 // inside the vframeArray (StackValueCollections)
792
793 delete thread->deopt_mark();
794 thread->set_deopt_mark(nullptr);
795 thread->set_deopt_compiled_method(nullptr);
796
797
798 if (JvmtiExport::can_pop_frame()) {
799 // Regardless of whether we entered this routine with the pending
800 // popframe condition bit set, we should always clear it now
801 thread->clear_popframe_condition();
802 }
803
804 // unpack_frames() is called at the end of the deoptimization handler
805 // and (in C2) at the end of the uncommon trap handler. Note this fact
806 // so that an asynchronous stack walker can work again. This counter is
807 // incremented at the beginning of fetch_unroll_info() and (in C2) at
808 // the beginning of uncommon_trap().
809 thread->dec_in_deopt_handler();
810 }
811
812 // Moved from cpu directories because none of the cpus has callee save values.
813 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
814 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
815
816 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
817 // the days we had adapter frames. When we deoptimize a situation where a
818 // compiled caller calls a compiled caller will have registers it expects
819 // to survive the call to the callee. If we deoptimize the callee the only
820 // way we can restore these registers is to have the oldest interpreter
821 // frame that we create restore these values. That is what this routine
822 // will accomplish.
823
824 // At the moment we have modified c2 to not have any callee save registers
825 // so this problem does not exist and this routine is just a place holder.
826
827 assert(f->is_interpreted_frame(), "must be interpreted");
828 }
829
830 #ifndef PRODUCT
831 static bool falls_through(Bytecodes::Code bc) {
832 switch (bc) {
833 // List may be incomplete. Here we really only care about bytecodes where compiled code
834 // can deoptimize.
835 case Bytecodes::_goto:
836 case Bytecodes::_goto_w:
837 case Bytecodes::_athrow:
838 return false;
839 default:
840 return true;
841 }
842 }
843 #endif
844
845 // Return BasicType of value being returned
846 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
847 assert(thread == JavaThread::current(), "pre-condition");
848
849 // We are already active in the special DeoptResourceMark any ResourceObj's we
850 // allocate will be freed at the end of the routine.
851
852 // JRT_LEAF methods don't normally allocate handles and there is a
853 // NoHandleMark to enforce that. It is actually safe to use Handles
854 // in a JRT_LEAF method, and sometimes desirable, but to do so we
855 // must use ResetNoHandleMark to bypass the NoHandleMark, and
856 // then use a HandleMark to ensure any Handles we do create are
857 // cleaned up in this scope.
858 ResetNoHandleMark rnhm;
859 HandleMark hm(thread);
860
861 frame stub_frame = thread->last_frame();
862
863 Continuation::notify_deopt(thread, stub_frame.sp());
864
865 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
866 // must point to the vframeArray for the unpack frame.
867 vframeArray* array = thread->vframe_array_head();
868 UnrollBlock* info = array->unroll_block();
869
870 // We set the last_Java frame. But the stack isn't really parsable here. So we
871 // clear it to make sure JFR understands not to try and walk stacks from events
872 // in here.
873 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
874 thread->frame_anchor()->set_last_Java_sp(nullptr);
875
876 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
877 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
878
879 thread->frame_anchor()->set_last_Java_sp(sp);
880
881 BasicType bt = info->return_type();
882
883 // If we have an exception pending, claim that the return type is an oop
884 // so the deopt_blob does not overwrite the exception_oop.
885
886 if (exec_mode == Unpack_exception)
887 bt = T_OBJECT;
888
889 // Cleanup thread deopt data
890 cleanup_deopt_info(thread, array);
891
892 #ifndef PRODUCT
893 if (VerifyStack) {
894 ResourceMark res_mark;
895 // Clear pending exception to not break verification code (restored afterwards)
896 PreserveExceptionMark pm(thread);
897
898 thread->validate_frame_layout();
899
900 // Verify that the just-unpacked frames match the interpreter's
901 // notions of expression stack and locals
902 vframeArray* cur_array = thread->vframe_array_last();
903 RegisterMap rm(thread,
904 RegisterMap::UpdateMap::skip,
905 RegisterMap::ProcessFrames::include,
906 RegisterMap::WalkContinuation::skip);
907 rm.set_include_argument_oops(false);
908 bool is_top_frame = true;
909 int callee_size_of_parameters = 0;
910 int callee_max_locals = 0;
911 for (int i = 0; i < cur_array->frames(); i++) {
912 vframeArrayElement* el = cur_array->element(i);
913 frame* iframe = el->iframe();
914 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
915
916 // Get the oop map for this bci
917 InterpreterOopMap mask;
918 int cur_invoke_parameter_size = 0;
919 bool try_next_mask = false;
920 int next_mask_expression_stack_size = -1;
921 int top_frame_expression_stack_adjustment = 0;
922 methodHandle mh(thread, iframe->interpreter_frame_method());
923 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
924 BytecodeStream str(mh, iframe->interpreter_frame_bci());
925 int max_bci = mh->code_size();
926 // Get to the next bytecode if possible
927 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
928 // Check to see if we can grab the number of outgoing arguments
929 // at an uncommon trap for an invoke (where the compiler
930 // generates debug info before the invoke has executed)
931 Bytecodes::Code cur_code = str.next();
932 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
933 if (Bytecodes::is_invoke(cur_code)) {
934 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
935 cur_invoke_parameter_size = invoke.size_of_parameters();
936 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
937 callee_size_of_parameters++;
938 }
939 }
940 if (str.bci() < max_bci) {
941 next_code = str.next();
942 if (next_code >= 0) {
943 // The interpreter oop map generator reports results before
944 // the current bytecode has executed except in the case of
945 // calls. It seems to be hard to tell whether the compiler
946 // has emitted debug information matching the "state before"
947 // a given bytecode or the state after, so we try both
948 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
949 // Get expression stack size for the next bytecode
950 InterpreterOopMap next_mask;
951 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
952 next_mask_expression_stack_size = next_mask.expression_stack_size();
953 if (Bytecodes::is_invoke(next_code)) {
954 Bytecode_invoke invoke(mh, str.bci());
955 next_mask_expression_stack_size += invoke.size_of_parameters();
956 }
957 // Need to subtract off the size of the result type of
958 // the bytecode because this is not described in the
959 // debug info but returned to the interpreter in the TOS
960 // caching register
961 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
962 if (bytecode_result_type != T_ILLEGAL) {
963 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
964 }
965 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
966 try_next_mask = true;
967 }
968 }
969 }
970
971 // Verify stack depth and oops in frame
972 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
973 if (!(
974 /* SPARC */
975 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
976 /* x86 */
977 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
978 (try_next_mask &&
979 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
980 top_frame_expression_stack_adjustment))) ||
981 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
982 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
983 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
984 )) {
985 {
986 // Print out some information that will help us debug the problem
987 tty->print_cr("Wrong number of expression stack elements during deoptimization");
988 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
989 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
990 if (try_next_mask) {
991 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
992 }
993 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
994 iframe->interpreter_frame_expression_stack_size());
995 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
996 tty->print_cr(" try_next_mask = %d", try_next_mask);
997 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
998 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
999 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1000 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1001 tty->print_cr(" exec_mode = %d", exec_mode);
1002 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1003 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1004 tty->print_cr(" Interpreted frames:");
1005 for (int k = 0; k < cur_array->frames(); k++) {
1006 vframeArrayElement* el = cur_array->element(k);
1007 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1008 }
1009 cur_array->print_on_2(tty);
1010 }
1011 guarantee(false, "wrong number of expression stack elements during deopt");
1012 }
1013 VerifyOopClosure verify;
1014 iframe->oops_interpreted_do(&verify, &rm, false);
1015 callee_size_of_parameters = mh->size_of_parameters();
1016 callee_max_locals = mh->max_locals();
1017 is_top_frame = false;
1018 }
1019 }
1020 #endif // !PRODUCT
1021
1022 return bt;
1023 JRT_END
1024
1025 class DeoptimizeMarkedClosure : public HandshakeClosure {
1026 public:
1027 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1028 void do_thread(Thread* thread) {
1029 JavaThread* jt = JavaThread::cast(thread);
1030 jt->deoptimize_marked_methods();
1031 }
1032 };
1033
1034 void Deoptimization::deoptimize_all_marked() {
1035 ResourceMark rm;
1036
1037 // Make the dependent methods not entrant
1038 CodeCache::make_marked_nmethods_deoptimized();
1039
1040 DeoptimizeMarkedClosure deopt;
1041 if (SafepointSynchronize::is_at_safepoint()) {
1042 Threads::java_threads_do(&deopt);
1043 } else {
1044 Handshake::execute(&deopt);
1045 }
1046 }
1047
1048 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1049 = Deoptimization::Action_reinterpret;
1050
1051 #if INCLUDE_JVMCI
1052 template<typename CacheType>
1053 class BoxCacheBase : public CHeapObj<mtCompiler> {
1054 protected:
1055 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1056 ResourceMark rm(thread);
1057 char* klass_name_str = klass_name->as_C_string();
1058 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle());
1059 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1060 if (!ik->is_in_error_state()) {
1061 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1062 CacheType::compute_offsets(ik);
1063 }
1064 return ik;
1065 }
1066 };
1067
1068 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1069 PrimitiveType _low;
1070 PrimitiveType _high;
1071 jobject _cache;
1072 protected:
1073 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1074 BoxCache(Thread* thread) {
1075 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1076 if (ik->is_in_error_state()) {
1077 _low = 1;
1078 _high = 0;
1079 _cache = nullptr;
1080 } else {
1081 objArrayOop cache = CacheType::cache(ik);
1082 assert(cache->length() > 0, "Empty cache");
1083 _low = BoxType::value(cache->obj_at(0));
1084 _high = _low + cache->length() - 1;
1085 _cache = JNIHandles::make_global(Handle(thread, cache));
1086 }
1087 }
1088 ~BoxCache() {
1089 JNIHandles::destroy_global(_cache);
1090 }
1091 public:
1092 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1093 if (_singleton == nullptr) {
1094 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1095 if (!Atomic::replace_if_null(&_singleton, s)) {
1096 delete s;
1097 }
1098 }
1099 return _singleton;
1100 }
1101 oop lookup(PrimitiveType value) {
1102 if (_low <= value && value <= _high) {
1103 int offset = value - _low;
1104 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1105 }
1106 return nullptr;
1107 }
1108 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1109 if (_cache == nullptr) {
1110 cache_init_error = true;
1111 return nullptr;
1112 }
1113 // Have to cast to avoid little/big-endian problems.
1114 if (sizeof(PrimitiveType) > sizeof(jint)) {
1115 jlong value = (jlong)raw_value;
1116 return lookup(value);
1117 }
1118 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1119 return lookup(value);
1120 }
1121 };
1122
1123 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1124 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1125 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1126 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1127 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1128
1129 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1130 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1131 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1132 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1133 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1134
1135 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1136 jobject _true_cache;
1137 jobject _false_cache;
1138 protected:
1139 static BooleanBoxCache *_singleton;
1140 BooleanBoxCache(Thread *thread) {
1141 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1142 if (ik->is_in_error_state()) {
1143 _true_cache = nullptr;
1144 _false_cache = nullptr;
1145 } else {
1146 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1147 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1148 }
1149 }
1150 ~BooleanBoxCache() {
1151 JNIHandles::destroy_global(_true_cache);
1152 JNIHandles::destroy_global(_false_cache);
1153 }
1154 public:
1155 static BooleanBoxCache* singleton(Thread* thread) {
1156 if (_singleton == nullptr) {
1157 BooleanBoxCache* s = new BooleanBoxCache(thread);
1158 if (!Atomic::replace_if_null(&_singleton, s)) {
1159 delete s;
1160 }
1161 }
1162 return _singleton;
1163 }
1164 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1165 if (_true_cache == nullptr) {
1166 cache_in_error = true;
1167 return nullptr;
1168 }
1169 // Have to cast to avoid little/big-endian problems.
1170 jboolean value = (jboolean)*((jint*)&raw_value);
1171 return lookup(value);
1172 }
1173 oop lookup(jboolean value) {
1174 if (value != 0) {
1175 return JNIHandles::resolve_non_null(_true_cache);
1176 }
1177 return JNIHandles::resolve_non_null(_false_cache);
1178 }
1179 };
1180
1181 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1182
1183 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1184 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1185 BasicType box_type = vmClasses::box_klass_type(k);
1186 if (box_type != T_OBJECT) {
1187 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1188 switch(box_type) {
1189 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1190 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1191 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1192 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1193 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1194 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error);
1195 default:;
1196 }
1197 }
1198 return nullptr;
1199 }
1200 #endif // INCLUDE_JVMCI
1201
1202 #if COMPILER2_OR_JVMCI
1203 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1204 Handle pending_exception(THREAD, thread->pending_exception());
1205 const char* exception_file = thread->exception_file();
1206 int exception_line = thread->exception_line();
1207 thread->clear_pending_exception();
1208
1209 bool failures = false;
1210
1211 for (int i = 0; i < objects->length(); i++) {
1212 assert(objects->at(i)->is_object(), "invalid debug information");
1213 ObjectValue* sv = (ObjectValue*) objects->at(i);
1214
1215 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1216 oop obj = nullptr;
1217
1218 bool cache_init_error = false;
1219 if (k->is_instance_klass()) {
1220 #if INCLUDE_JVMCI
1221 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null();
1222 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1223 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1224 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1225 if (obj != nullptr) {
1226 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1227 abv->set_cached(true);
1228 } else if (cache_init_error) {
1229 // Results in an OOME which is valid (as opposed to a class initialization error)
1230 // and is fine for the rare case a cache initialization failing.
1231 failures = true;
1232 }
1233 }
1234 #endif // INCLUDE_JVMCI
1235
1236 InstanceKlass* ik = InstanceKlass::cast(k);
1237 if (obj == nullptr && !cache_init_error) {
1238 #ifdef COMPILER2
1239 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1240 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1241 } else {
1242 obj = ik->allocate_instance(THREAD);
1243 }
1244 #else
1245 obj = ik->allocate_instance(THREAD);
1246 #endif // COMPILER2
1247 }
1248 } else if (k->is_typeArray_klass()) {
1249 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1250 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1251 int len = sv->field_size() / type2size[ak->element_type()];
1252 obj = ak->allocate(len, THREAD);
1253 } else if (k->is_objArray_klass()) {
1254 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1255 obj = ak->allocate(sv->field_size(), THREAD);
1256 }
1257
1258 if (obj == nullptr) {
1259 failures = true;
1260 }
1261
1262 assert(sv->value().is_null(), "redundant reallocation");
1263 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1264 CLEAR_PENDING_EXCEPTION;
1265 sv->set_value(obj);
1266 }
1267
1268 if (failures) {
1269 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1270 } else if (pending_exception.not_null()) {
1271 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1272 }
1273
1274 return failures;
1275 }
1276
1277 #if INCLUDE_JVMCI
1278 /**
1279 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1280 * we need to somehow be able to recover the actual kind to be able to write the correct
1281 * amount of bytes.
1282 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1283 * the entries at index n + 1 to n + i are 'markers'.
1284 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1285 * expected form of the array would be:
1286 *
1287 * {b0, b1, b2, b3, INT, marker, b6, b7}
1288 *
1289 * Thus, in order to get back the size of the entry, we simply need to count the number
1290 * of marked entries
1291 *
1292 * @param virtualArray the virtualized byte array
1293 * @param i index of the virtual entry we are recovering
1294 * @return The number of bytes the entry spans
1295 */
1296 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1297 int index = i;
1298 while (++index < virtualArray->field_size() &&
1299 virtualArray->field_at(index)->is_marker()) {}
1300 return index - i;
1301 }
1302
1303 /**
1304 * If there was a guarantee for byte array to always start aligned to a long, we could
1305 * do a simple check on the parity of the index. Unfortunately, that is not always the
1306 * case. Thus, we check alignment of the actual address we are writing to.
1307 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1308 * write a long to index 3.
1309 */
1310 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1311 jbyte* res = obj->byte_at_addr(index);
1312 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1313 return res;
1314 }
1315
1316 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) {
1317 switch (byte_count) {
1318 case 1:
1319 obj->byte_at_put(index, (jbyte) *((jint *) &val));
1320 break;
1321 case 2:
1322 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val);
1323 break;
1324 case 4:
1325 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val);
1326 break;
1327 case 8:
1328 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val);
1329 break;
1330 default:
1331 ShouldNotReachHere();
1332 }
1333 }
1334 #endif // INCLUDE_JVMCI
1335
1336
1337 // restore elements of an eliminated type array
1338 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1339 int index = 0;
1340 intptr_t val;
1341
1342 for (int i = 0; i < sv->field_size(); i++) {
1343 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1344 switch(type) {
1345 case T_LONG: case T_DOUBLE: {
1346 assert(value->type() == T_INT, "Agreement.");
1347 StackValue* low =
1348 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1349 #ifdef _LP64
1350 jlong res = (jlong)low->get_int();
1351 #else
1352 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1353 #endif
1354 obj->long_at_put(index, res);
1355 break;
1356 }
1357
1358 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1359 case T_INT: case T_FLOAT: { // 4 bytes.
1360 assert(value->type() == T_INT, "Agreement.");
1361 bool big_value = false;
1362 if (i + 1 < sv->field_size() && type == T_INT) {
1363 if (sv->field_at(i)->is_location()) {
1364 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1365 if (type == Location::dbl || type == Location::lng) {
1366 big_value = true;
1367 }
1368 } else if (sv->field_at(i)->is_constant_int()) {
1369 ScopeValue* next_scope_field = sv->field_at(i + 1);
1370 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1371 big_value = true;
1372 }
1373 }
1374 }
1375
1376 if (big_value) {
1377 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1378 #ifdef _LP64
1379 jlong res = (jlong)low->get_int();
1380 #else
1381 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1382 #endif
1383 obj->int_at_put(index, (jint)*((jint*)&res));
1384 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1385 } else {
1386 val = value->get_int();
1387 obj->int_at_put(index, (jint)*((jint*)&val));
1388 }
1389 break;
1390 }
1391
1392 case T_SHORT:
1393 assert(value->type() == T_INT, "Agreement.");
1394 val = value->get_int();
1395 obj->short_at_put(index, (jshort)*((jint*)&val));
1396 break;
1397
1398 case T_CHAR:
1399 assert(value->type() == T_INT, "Agreement.");
1400 val = value->get_int();
1401 obj->char_at_put(index, (jchar)*((jint*)&val));
1402 break;
1403
1404 case T_BYTE: {
1405 assert(value->type() == T_INT, "Agreement.");
1406 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1407 val = value->get_int();
1408 #if INCLUDE_JVMCI
1409 int byte_count = count_number_of_bytes_for_entry(sv, i);
1410 byte_array_put(obj, val, index, byte_count);
1411 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1412 i += byte_count - 1; // Balance the loop counter.
1413 index += byte_count;
1414 // index has been updated so continue at top of loop
1415 continue;
1416 #else
1417 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1418 break;
1419 #endif // INCLUDE_JVMCI
1420 }
1421
1422 case T_BOOLEAN: {
1423 assert(value->type() == T_INT, "Agreement.");
1424 val = value->get_int();
1425 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1426 break;
1427 }
1428
1429 default:
1430 ShouldNotReachHere();
1431 }
1432 index++;
1433 }
1434 }
1435
1436 // restore fields of an eliminated object array
1437 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1438 for (int i = 0; i < sv->field_size(); i++) {
1439 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1440 assert(value->type() == T_OBJECT, "object element expected");
1441 obj->obj_at_put(i, value->get_obj()());
1442 }
1443 }
1444
1445 class ReassignedField {
1446 public:
1447 int _offset;
1448 BasicType _type;
1449 public:
1450 ReassignedField() {
1451 _offset = 0;
1452 _type = T_ILLEGAL;
1453 }
1454 };
1455
1456 int compare(ReassignedField* left, ReassignedField* right) {
1457 return left->_offset - right->_offset;
1458 }
1459
1460 // Restore fields of an eliminated instance object using the same field order
1461 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1462 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1463 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1464 InstanceKlass* ik = klass;
1465 while (ik != nullptr) {
1466 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1467 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1468 ReassignedField field;
1469 field._offset = fs.offset();
1470 field._type = Signature::basic_type(fs.signature());
1471 fields->append(field);
1472 }
1473 }
1474 ik = ik->superklass();
1475 }
1476 fields->sort(compare);
1477 for (int i = 0; i < fields->length(); i++) {
1478 intptr_t val;
1479 ScopeValue* scope_field = sv->field_at(svIndex);
1480 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1481 int offset = fields->at(i)._offset;
1482 BasicType type = fields->at(i)._type;
1483 switch (type) {
1484 case T_OBJECT: case T_ARRAY:
1485 assert(value->type() == T_OBJECT, "Agreement.");
1486 obj->obj_field_put(offset, value->get_obj()());
1487 break;
1488
1489 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
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 val = value->get_int();
1514 obj->int_field_put(offset, (jint)*((jint*)&val));
1515 break;
1516 }
1517 }
1518 /* no break */
1519
1520 case T_LONG: case T_DOUBLE: {
1521 assert(value->type() == T_INT, "Agreement.");
1522 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1523 #ifdef _LP64
1524 jlong res = (jlong)low->get_int();
1525 #else
1526 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1527 #endif
1528 obj->long_field_put(offset, res);
1529 break;
1530 }
1531
1532 case T_SHORT:
1533 assert(value->type() == T_INT, "Agreement.");
1534 val = value->get_int();
1535 obj->short_field_put(offset, (jshort)*((jint*)&val));
1536 break;
1537
1538 case T_CHAR:
1539 assert(value->type() == T_INT, "Agreement.");
1540 val = value->get_int();
1541 obj->char_field_put(offset, (jchar)*((jint*)&val));
1542 break;
1543
1544 case T_BYTE:
1545 assert(value->type() == T_INT, "Agreement.");
1546 val = value->get_int();
1547 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1548 break;
1549
1550 case T_BOOLEAN:
1551 assert(value->type() == T_INT, "Agreement.");
1552 val = value->get_int();
1553 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1554 break;
1555
1556 default:
1557 ShouldNotReachHere();
1558 }
1559 svIndex++;
1560 }
1561 return svIndex;
1562 }
1563
1564 // restore fields of all eliminated objects and arrays
1565 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1566 for (int i = 0; i < objects->length(); i++) {
1567 ObjectValue* sv = (ObjectValue*) objects->at(i);
1568 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1569 Handle obj = sv->value();
1570 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1571 #ifndef PRODUCT
1572 if (PrintDeoptimizationDetails) {
1573 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1574 }
1575 #endif // !PRODUCT
1576
1577 if (obj.is_null()) {
1578 continue;
1579 }
1580
1581 #if INCLUDE_JVMCI
1582 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1583 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1584 continue;
1585 }
1586 #endif // INCLUDE_JVMCI
1587 #ifdef COMPILER2
1588 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1589 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1590 ScopeValue* payload = sv->field_at(0);
1591 if (payload->is_location() &&
1592 payload->as_LocationValue()->location().type() == Location::vector) {
1593 #ifndef PRODUCT
1594 if (PrintDeoptimizationDetails) {
1595 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1596 if (Verbose) {
1597 Handle obj = sv->value();
1598 k->oop_print_on(obj(), tty);
1599 }
1600 }
1601 #endif // !PRODUCT
1602 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1603 }
1604 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1605 // which could be restored after vector object allocation.
1606 }
1607 #endif /* !COMPILER2 */
1608 if (k->is_instance_klass()) {
1609 InstanceKlass* ik = InstanceKlass::cast(k);
1610 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1611 } else if (k->is_typeArray_klass()) {
1612 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1613 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1614 } else if (k->is_objArray_klass()) {
1615 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1616 }
1617 }
1618 // These objects may escape when we return to Interpreter after deoptimization.
1619 // We need barrier so that stores that initialize these objects can't be reordered
1620 // with subsequent stores that make these objects accessible by other threads.
1621 OrderAccess::storestore();
1622 }
1623
1624
1625 // relock objects for which synchronization was eliminated
1626 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1627 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1628 bool relocked_objects = false;
1629 for (int i = 0; i < monitors->length(); i++) {
1630 MonitorInfo* mon_info = monitors->at(i);
1631 if (mon_info->eliminated()) {
1632 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1633 relocked_objects = true;
1634 if (!mon_info->owner_is_scalar_replaced()) {
1635 Handle obj(thread, mon_info->owner());
1636 markWord mark = obj->mark();
1637 if (exec_mode == Unpack_none) {
1638 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1639 // With exec_mode == Unpack_none obj may be thread local and locked in
1640 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1641 markWord dmw = mark.displaced_mark_helper();
1642 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1643 obj->set_mark(dmw);
1644 }
1645 if (mark.has_monitor()) {
1646 // defer relocking if the deoptee thread is currently waiting for obj
1647 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1648 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1649 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1650 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1651 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1652 continue;
1653 }
1654 }
1655 }
1656 if (LockingMode == LM_LIGHTWEIGHT && exec_mode == Unpack_none) {
1657 // We have lost information about the correct state of the lock stack.
1658 // Inflate the locks instead. Enter then inflate to avoid races with
1659 // deflation.
1660 ObjectSynchronizer::enter(obj, nullptr, deoptee_thread);
1661 assert(mon_info->owner()->is_locked(), "object must be locked now");
1662 ObjectMonitor* mon = ObjectSynchronizer::inflate(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal);
1663 assert(mon->owner() == deoptee_thread, "must be");
1664 } else {
1665 BasicLock* lock = mon_info->lock();
1666 ObjectSynchronizer::enter(obj, lock, deoptee_thread);
1667 assert(mon_info->owner()->is_locked(), "object must be locked now");
1668 }
1669 }
1670 }
1671 }
1672 return relocked_objects;
1673 }
1674 #endif // COMPILER2_OR_JVMCI
1675
1676 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1677 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1678
1679 // Register map for next frame (used for stack crawl). We capture
1680 // the state of the deopt'ing frame's caller. Thus if we need to
1681 // stuff a C2I adapter we can properly fill in the callee-save
1682 // register locations.
1683 frame caller = fr.sender(reg_map);
1684 int frame_size = caller.sp() - fr.sp();
1685
1686 frame sender = caller;
1687
1688 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1689 // the vframeArray containing the unpacking information is allocated in the C heap.
1690 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1691 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1692
1693 // Compare the vframeArray to the collected vframes
1694 assert(array->structural_compare(thread, chunk), "just checking");
1695
1696 if (TraceDeoptimization) {
1697 ResourceMark rm;
1698 stringStream st;
1699 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1700 st.print(" ");
1701 fr.print_on(&st);
1702 st.print_cr(" Virtual frames (innermost/newest first):");
1703 for (int index = 0; index < chunk->length(); index++) {
1704 compiledVFrame* vf = chunk->at(index);
1705 int bci = vf->raw_bci();
1706 const char* code_name;
1707 if (bci == SynchronizationEntryBCI) {
1708 code_name = "sync entry";
1709 } else {
1710 Bytecodes::Code code = vf->method()->code_at(bci);
1711 code_name = Bytecodes::name(code);
1712 }
1713
1714 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1715 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1716 st.print(" - %s", code_name);
1717 st.print_cr(" @ bci=%d ", bci);
1718 }
1719 tty->print_raw(st.freeze());
1720 tty->cr();
1721 }
1722
1723 return array;
1724 }
1725
1726 #if COMPILER2_OR_JVMCI
1727 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1728 // Reallocation of some scalar replaced objects failed. Record
1729 // that we need to pop all the interpreter frames for the
1730 // deoptimized compiled frame.
1731 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1732 thread->set_frames_to_pop_failed_realloc(array->frames());
1733 // Unlock all monitors here otherwise the interpreter will see a
1734 // mix of locked and unlocked monitors (because of failed
1735 // reallocations of synchronized objects) and be confused.
1736 for (int i = 0; i < array->frames(); i++) {
1737 MonitorChunk* monitors = array->element(i)->monitors();
1738 if (monitors != nullptr) {
1739 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1740 BasicObjectLock* src = monitors->at(j);
1741 if (src->obj() != nullptr) {
1742 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1743 }
1744 }
1745 array->element(i)->free_monitors(thread);
1746 #ifdef ASSERT
1747 array->element(i)->set_removed_monitors();
1748 #endif
1749 }
1750 }
1751 }
1752 #endif
1753
1754 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1755 assert(fr.can_be_deoptimized(), "checking frame type");
1756
1757 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1758
1759 if (LogCompilation && xtty != nullptr) {
1760 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1761 assert(cm != nullptr, "only compiled methods can deopt");
1762
1763 ttyLocker ttyl;
1764 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1765 cm->log_identity(xtty);
1766 xtty->end_head();
1767 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1768 xtty->begin_elem("jvms bci='%d'", sd->bci());
1769 xtty->method(sd->method());
1770 xtty->end_elem();
1771 if (sd->is_top()) break;
1772 }
1773 xtty->tail("deoptimized");
1774 }
1775
1776 Continuation::notify_deopt(thread, fr.sp());
1777
1778 // Patch the compiled method so that when execution returns to it we will
1779 // deopt the execution state and return to the interpreter.
1780 fr.deoptimize(thread);
1781 }
1782
1783 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1784 // Deoptimize only if the frame comes from compile code.
1785 // Do not deoptimize the frame which is already patched
1786 // during the execution of the loops below.
1787 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1788 return;
1789 }
1790 ResourceMark rm;
1791 deoptimize_single_frame(thread, fr, reason);
1792 }
1793
1794 #if INCLUDE_JVMCI
1795 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1796 // there is no exception handler for this pc => deoptimize
1797 cm->make_not_entrant();
1798
1799 // Use Deoptimization::deoptimize for all of its side-effects:
1800 // gathering traps statistics, logging...
1801 // it also patches the return pc but we do not care about that
1802 // since we return a continuation to the deopt_blob below.
1803 JavaThread* thread = JavaThread::current();
1804 RegisterMap reg_map(thread,
1805 RegisterMap::UpdateMap::skip,
1806 RegisterMap::ProcessFrames::include,
1807 RegisterMap::WalkContinuation::skip);
1808 frame runtime_frame = thread->last_frame();
1809 frame caller_frame = runtime_frame.sender(®_map);
1810 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1811 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1812 compiledVFrame* cvf = compiledVFrame::cast(vf);
1813 ScopeDesc* imm_scope = cvf->scope();
1814 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1815 if (imm_mdo != nullptr) {
1816 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1817 if (pdata != nullptr && pdata->is_BitData()) {
1818 BitData* bit_data = (BitData*) pdata;
1819 bit_data->set_exception_seen();
1820 }
1821 }
1822
1823 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1824
1825 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1826 if (trap_mdo != nullptr) {
1827 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1828 }
1829
1830 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1831 }
1832 #endif
1833
1834 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1835 assert(thread == Thread::current() ||
1836 thread->is_handshake_safe_for(Thread::current()) ||
1837 SafepointSynchronize::is_at_safepoint(),
1838 "can only deoptimize other thread at a safepoint/handshake");
1839 // Compute frame and register map based on thread and sp.
1840 RegisterMap reg_map(thread,
1841 RegisterMap::UpdateMap::skip,
1842 RegisterMap::ProcessFrames::include,
1843 RegisterMap::WalkContinuation::skip);
1844 frame fr = thread->last_frame();
1845 while (fr.id() != id) {
1846 fr = fr.sender(®_map);
1847 }
1848 deoptimize(thread, fr, reason);
1849 }
1850
1851
1852 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1853 Thread* current = Thread::current();
1854 if (thread == current || thread->is_handshake_safe_for(current)) {
1855 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1856 } else {
1857 VM_DeoptimizeFrame deopt(thread, id, reason);
1858 VMThread::execute(&deopt);
1859 }
1860 }
1861
1862 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1863 deoptimize_frame(thread, id, Reason_constraint);
1864 }
1865
1866 // JVMTI PopFrame support
1867 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1868 {
1869 assert(thread == JavaThread::current(), "pre-condition");
1870 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1871 }
1872 JRT_END
1873
1874 MethodData*
1875 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1876 bool create_if_missing) {
1877 JavaThread* THREAD = thread; // For exception macros.
1878 MethodData* mdo = m()->method_data();
1879 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1880 // Build an MDO. Ignore errors like OutOfMemory;
1881 // that simply means we won't have an MDO to update.
1882 Method::build_profiling_method_data(m, THREAD);
1883 if (HAS_PENDING_EXCEPTION) {
1884 // Only metaspace OOM is expected. No Java code executed.
1885 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1886 CLEAR_PENDING_EXCEPTION;
1887 }
1888 mdo = m()->method_data();
1889 }
1890 return mdo;
1891 }
1892
1893 #if COMPILER2_OR_JVMCI
1894 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1895 // In case of an unresolved klass entry, load the class.
1896 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1897 // and probably nowhere else.
1898 // Even that case would benefit from simply re-interpreting the
1899 // bytecode, without paying special attention to the class index.
1900 // So this whole "class index" feature should probably be removed.
1901
1902 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1903 Klass* tk = constant_pool->klass_at(index, THREAD);
1904 if (HAS_PENDING_EXCEPTION) {
1905 // Exception happened during classloading. We ignore the exception here, since it
1906 // is going to be rethrown since the current activation is going to be deoptimized and
1907 // the interpreter will re-execute the bytecode.
1908 // Do not clear probable Async Exceptions.
1909 CLEAR_PENDING_NONASYNC_EXCEPTION;
1910 // Class loading called java code which may have caused a stack
1911 // overflow. If the exception was thrown right before the return
1912 // to the runtime the stack is no longer guarded. Reguard the
1913 // stack otherwise if we return to the uncommon trap blob and the
1914 // stack bang causes a stack overflow we crash.
1915 JavaThread* jt = THREAD;
1916 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1917 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1918 }
1919 return;
1920 }
1921
1922 assert(!constant_pool->tag_at(index).is_symbol(),
1923 "no symbolic names here, please");
1924 }
1925
1926 #if INCLUDE_JFR
1927
1928 class DeoptReasonSerializer : public JfrSerializer {
1929 public:
1930 void serialize(JfrCheckpointWriter& writer) {
1931 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1932 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1933 writer.write_key((u8)i);
1934 writer.write(Deoptimization::trap_reason_name(i));
1935 }
1936 }
1937 };
1938
1939 class DeoptActionSerializer : public JfrSerializer {
1940 public:
1941 void serialize(JfrCheckpointWriter& writer) {
1942 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1943 writer.write_count(nof_actions);
1944 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1945 writer.write_key(i);
1946 writer.write(Deoptimization::trap_action_name((int)i));
1947 }
1948 }
1949 };
1950
1951 static void register_serializers() {
1952 static int critical_section = 0;
1953 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1954 return;
1955 }
1956 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1957 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1958 }
1959
1960 static void post_deoptimization_event(CompiledMethod* nm,
1961 const Method* method,
1962 int trap_bci,
1963 int instruction,
1964 Deoptimization::DeoptReason reason,
1965 Deoptimization::DeoptAction action) {
1966 assert(nm != nullptr, "invariant");
1967 assert(method != nullptr, "invariant");
1968 if (EventDeoptimization::is_enabled()) {
1969 static bool serializers_registered = false;
1970 if (!serializers_registered) {
1971 register_serializers();
1972 serializers_registered = true;
1973 }
1974 EventDeoptimization event;
1975 event.set_compileId(nm->compile_id());
1976 event.set_compiler(nm->compiler_type());
1977 event.set_method(method);
1978 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1979 event.set_bci(trap_bci);
1980 event.set_instruction(instruction);
1981 event.set_reason(reason);
1982 event.set_action(action);
1983 event.commit();
1984 }
1985 }
1986
1987 #endif // INCLUDE_JFR
1988
1989 static void log_deopt(CompiledMethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
1990 const char* reason_name, const char* reason_action) {
1991 LogTarget(Debug, deoptimization) lt;
1992 if (lt.is_enabled()) {
1993 LogStream ls(lt);
1994 bool is_osr = nm->is_osr_method();
1995 ls.print("cid=%4d %s level=%d",
1996 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
1997 ls.print(" %s", tm->name_and_sig_as_C_string());
1998 ls.print(" trap_bci=%d ", trap_bci);
1999 if (is_osr) {
2000 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2001 }
2002 ls.print("%s ", reason_name);
2003 ls.print("%s ", reason_action);
2004 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2005 pc, fr.pc() - nm->code_begin());
2006 }
2007 }
2008
2009 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2010 HandleMark hm(current);
2011
2012 // uncommon_trap() is called at the beginning of the uncommon trap
2013 // handler. Note this fact before we start generating temporary frames
2014 // that can confuse an asynchronous stack walker. This counter is
2015 // decremented at the end of unpack_frames().
2016
2017 current->inc_in_deopt_handler();
2018
2019 #if INCLUDE_JVMCI
2020 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2021 RegisterMap reg_map(current,
2022 RegisterMap::UpdateMap::include,
2023 RegisterMap::ProcessFrames::include,
2024 RegisterMap::WalkContinuation::skip);
2025 #else
2026 RegisterMap reg_map(current,
2027 RegisterMap::UpdateMap::skip,
2028 RegisterMap::ProcessFrames::include,
2029 RegisterMap::WalkContinuation::skip);
2030 #endif
2031 frame stub_frame = current->last_frame();
2032 frame fr = stub_frame.sender(®_map);
2033
2034 // Log a message
2035 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2036 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2037
2038 {
2039 ResourceMark rm;
2040
2041 DeoptReason reason = trap_request_reason(trap_request);
2042 DeoptAction action = trap_request_action(trap_request);
2043 #if INCLUDE_JVMCI
2044 int debug_id = trap_request_debug_id(trap_request);
2045 #endif
2046 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2047
2048 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2049 compiledVFrame* cvf = compiledVFrame::cast(vf);
2050
2051 CompiledMethod* nm = cvf->code();
2052
2053 ScopeDesc* trap_scope = cvf->scope();
2054
2055 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2056
2057 if (is_receiver_constraint_failure) {
2058 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2059 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2060 JVMCI_ONLY(COMMA debug_id));
2061 }
2062
2063 methodHandle trap_method(current, trap_scope->method());
2064 int trap_bci = trap_scope->bci();
2065 #if INCLUDE_JVMCI
2066 jlong speculation = current->pending_failed_speculation();
2067 if (nm->is_compiled_by_jvmci()) {
2068 nm->as_nmethod()->update_speculation(current);
2069 } else {
2070 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2071 }
2072
2073 if (trap_bci == SynchronizationEntryBCI) {
2074 trap_bci = 0;
2075 current->set_pending_monitorenter(true);
2076 }
2077
2078 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2079 current->set_pending_transfer_to_interpreter(true);
2080 }
2081 #endif
2082
2083 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2084 // Record this event in the histogram.
2085 gather_statistics(reason, action, trap_bc);
2086
2087 // Ensure that we can record deopt. history:
2088 // Need MDO to record RTM code generation state.
2089 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking );
2090
2091 methodHandle profiled_method;
2092 #if INCLUDE_JVMCI
2093 if (nm->is_compiled_by_jvmci()) {
2094 profiled_method = methodHandle(current, nm->method());
2095 } else {
2096 profiled_method = trap_method;
2097 }
2098 #else
2099 profiled_method = trap_method;
2100 #endif
2101
2102 MethodData* trap_mdo =
2103 get_method_data(current, profiled_method, create_if_missing);
2104
2105 { // Log Deoptimization event for JFR, UL and event system
2106 Method* tm = trap_method();
2107 const char* reason_name = trap_reason_name(reason);
2108 const char* reason_action = trap_action_name(action);
2109 intptr_t pc = p2i(fr.pc());
2110
2111 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2112 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2113 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2114 reason_name, reason_action, pc,
2115 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2116 }
2117
2118 // Print a bunch of diagnostics, if requested.
2119 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2120 ResourceMark rm;
2121 ttyLocker ttyl;
2122 char buf[100];
2123 if (xtty != nullptr) {
2124 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2125 os::current_thread_id(),
2126 format_trap_request(buf, sizeof(buf), trap_request));
2127 #if INCLUDE_JVMCI
2128 if (speculation != 0) {
2129 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2130 }
2131 #endif
2132 nm->log_identity(xtty);
2133 }
2134 Symbol* class_name = nullptr;
2135 bool unresolved = false;
2136 if (unloaded_class_index >= 0) {
2137 constantPoolHandle constants (current, trap_method->constants());
2138 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2139 class_name = constants->klass_name_at(unloaded_class_index);
2140 unresolved = true;
2141 if (xtty != nullptr)
2142 xtty->print(" unresolved='1'");
2143 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2144 class_name = constants->symbol_at(unloaded_class_index);
2145 }
2146 if (xtty != nullptr)
2147 xtty->name(class_name);
2148 }
2149 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2150 // Dump the relevant MDO state.
2151 // This is the deopt count for the current reason, any previous
2152 // reasons or recompiles seen at this point.
2153 int dcnt = trap_mdo->trap_count(reason);
2154 if (dcnt != 0)
2155 xtty->print(" count='%d'", dcnt);
2156 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2157 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2158 if (dos != 0) {
2159 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2160 if (trap_state_is_recompiled(dos)) {
2161 int recnt2 = trap_mdo->overflow_recompile_count();
2162 if (recnt2 != 0)
2163 xtty->print(" recompiles2='%d'", recnt2);
2164 }
2165 }
2166 }
2167 if (xtty != nullptr) {
2168 xtty->stamp();
2169 xtty->end_head();
2170 }
2171 if (TraceDeoptimization) { // make noise on the tty
2172 stringStream st;
2173 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2174 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2175 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2176 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2177 #if INCLUDE_JVMCI
2178 if (nm->is_nmethod()) {
2179 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
2180 if (installed_code_name != nullptr) {
2181 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2182 }
2183 }
2184 #endif
2185 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2186 p2i(fr.pc()),
2187 os::current_thread_id(),
2188 trap_reason_name(reason),
2189 trap_action_name(action),
2190 unloaded_class_index
2191 #if INCLUDE_JVMCI
2192 , debug_id
2193 #endif
2194 );
2195 if (class_name != nullptr) {
2196 st.print(unresolved ? " unresolved class: " : " symbol: ");
2197 class_name->print_symbol_on(&st);
2198 }
2199 st.cr();
2200 tty->print_raw(st.freeze());
2201 }
2202 if (xtty != nullptr) {
2203 // Log the precise location of the trap.
2204 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2205 xtty->begin_elem("jvms bci='%d'", sd->bci());
2206 xtty->method(sd->method());
2207 xtty->end_elem();
2208 if (sd->is_top()) break;
2209 }
2210 xtty->tail("uncommon_trap");
2211 }
2212 }
2213 // (End diagnostic printout.)
2214
2215 if (is_receiver_constraint_failure) {
2216 fatal("missing receiver type check");
2217 }
2218
2219 // Load class if necessary
2220 if (unloaded_class_index >= 0) {
2221 constantPoolHandle constants(current, trap_method->constants());
2222 load_class_by_index(constants, unloaded_class_index, THREAD);
2223 }
2224
2225 // Flush the nmethod if necessary and desirable.
2226 //
2227 // We need to avoid situations where we are re-flushing the nmethod
2228 // because of a hot deoptimization site. Repeated flushes at the same
2229 // point need to be detected by the compiler and avoided. If the compiler
2230 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2231 // module must take measures to avoid an infinite cycle of recompilation
2232 // and deoptimization. There are several such measures:
2233 //
2234 // 1. If a recompilation is ordered a second time at some site X
2235 // and for the same reason R, the action is adjusted to 'reinterpret',
2236 // to give the interpreter time to exercise the method more thoroughly.
2237 // If this happens, the method's overflow_recompile_count is incremented.
2238 //
2239 // 2. If the compiler fails to reduce the deoptimization rate, then
2240 // the method's overflow_recompile_count will begin to exceed the set
2241 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2242 // is adjusted to 'make_not_compilable', and the method is abandoned
2243 // to the interpreter. This is a performance hit for hot methods,
2244 // but is better than a disastrous infinite cycle of recompilations.
2245 // (Actually, only the method containing the site X is abandoned.)
2246 //
2247 // 3. In parallel with the previous measures, if the total number of
2248 // recompilations of a method exceeds the much larger set limit
2249 // PerMethodRecompilationCutoff, the method is abandoned.
2250 // This should only happen if the method is very large and has
2251 // many "lukewarm" deoptimizations. The code which enforces this
2252 // limit is elsewhere (class nmethod, class Method).
2253 //
2254 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2255 // to recompile at each bytecode independently of the per-BCI cutoff.
2256 //
2257 // The decision to update code is up to the compiler, and is encoded
2258 // in the Action_xxx code. If the compiler requests Action_none
2259 // no trap state is changed, no compiled code is changed, and the
2260 // computation suffers along in the interpreter.
2261 //
2262 // The other action codes specify various tactics for decompilation
2263 // and recompilation. Action_maybe_recompile is the loosest, and
2264 // allows the compiled code to stay around until enough traps are seen,
2265 // and until the compiler gets around to recompiling the trapping method.
2266 //
2267 // The other actions cause immediate removal of the present code.
2268
2269 // Traps caused by injected profile shouldn't pollute trap counts.
2270 bool injected_profile_trap = trap_method->has_injected_profile() &&
2271 (reason == Reason_intrinsic || reason == Reason_unreached);
2272
2273 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2274 bool make_not_entrant = false;
2275 bool make_not_compilable = false;
2276 bool reprofile = false;
2277 switch (action) {
2278 case Action_none:
2279 // Keep the old code.
2280 update_trap_state = false;
2281 break;
2282 case Action_maybe_recompile:
2283 // Do not need to invalidate the present code, but we can
2284 // initiate another
2285 // Start compiler without (necessarily) invalidating the nmethod.
2286 // The system will tolerate the old code, but new code should be
2287 // generated when possible.
2288 break;
2289 case Action_reinterpret:
2290 // Go back into the interpreter for a while, and then consider
2291 // recompiling form scratch.
2292 make_not_entrant = true;
2293 // Reset invocation counter for outer most method.
2294 // This will allow the interpreter to exercise the bytecodes
2295 // for a while before recompiling.
2296 // By contrast, Action_make_not_entrant is immediate.
2297 //
2298 // Note that the compiler will track null_check, null_assert,
2299 // range_check, and class_check events and log them as if they
2300 // had been traps taken from compiled code. This will update
2301 // the MDO trap history so that the next compilation will
2302 // properly detect hot trap sites.
2303 reprofile = true;
2304 break;
2305 case Action_make_not_entrant:
2306 // Request immediate recompilation, and get rid of the old code.
2307 // Make them not entrant, so next time they are called they get
2308 // recompiled. Unloaded classes are loaded now so recompile before next
2309 // time they are called. Same for uninitialized. The interpreter will
2310 // link the missing class, if any.
2311 make_not_entrant = true;
2312 break;
2313 case Action_make_not_compilable:
2314 // Give up on compiling this method at all.
2315 make_not_entrant = true;
2316 make_not_compilable = true;
2317 break;
2318 default:
2319 ShouldNotReachHere();
2320 }
2321
2322 // Setting +ProfileTraps fixes the following, on all platforms:
2323 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2324 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2325 // recompile relies on a MethodData* to record heroic opt failures.
2326
2327 // Whether the interpreter is producing MDO data or not, we also need
2328 // to use the MDO to detect hot deoptimization points and control
2329 // aggressive optimization.
2330 bool inc_recompile_count = false;
2331 ProfileData* pdata = nullptr;
2332 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2333 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2334 uint this_trap_count = 0;
2335 bool maybe_prior_trap = false;
2336 bool maybe_prior_recompile = false;
2337 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2338 #if INCLUDE_JVMCI
2339 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2340 #endif
2341 nm->method(),
2342 //outputs:
2343 this_trap_count,
2344 maybe_prior_trap,
2345 maybe_prior_recompile);
2346 // Because the interpreter also counts null, div0, range, and class
2347 // checks, these traps from compiled code are double-counted.
2348 // This is harmless; it just means that the PerXTrapLimit values
2349 // are in effect a little smaller than they look.
2350
2351 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2352 if (per_bc_reason != Reason_none) {
2353 // Now take action based on the partially known per-BCI history.
2354 if (maybe_prior_trap
2355 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2356 // If there are too many traps at this BCI, force a recompile.
2357 // This will allow the compiler to see the limit overflow, and
2358 // take corrective action, if possible. The compiler generally
2359 // does not use the exact PerBytecodeTrapLimit value, but instead
2360 // changes its tactics if it sees any traps at all. This provides
2361 // a little hysteresis, delaying a recompile until a trap happens
2362 // several times.
2363 //
2364 // Actually, since there is only one bit of counter per BCI,
2365 // the possible per-BCI counts are {0,1,(per-method count)}.
2366 // This produces accurate results if in fact there is only
2367 // one hot trap site, but begins to get fuzzy if there are
2368 // many sites. For example, if there are ten sites each
2369 // trapping two or more times, they each get the blame for
2370 // all of their traps.
2371 make_not_entrant = true;
2372 }
2373
2374 // Detect repeated recompilation at the same BCI, and enforce a limit.
2375 if (make_not_entrant && maybe_prior_recompile) {
2376 // More than one recompile at this point.
2377 inc_recompile_count = maybe_prior_trap;
2378 }
2379 } else {
2380 // For reasons which are not recorded per-bytecode, we simply
2381 // force recompiles unconditionally.
2382 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2383 make_not_entrant = true;
2384 }
2385
2386 // Go back to the compiler if there are too many traps in this method.
2387 if (this_trap_count >= per_method_trap_limit(reason)) {
2388 // If there are too many traps in this method, force a recompile.
2389 // This will allow the compiler to see the limit overflow, and
2390 // take corrective action, if possible.
2391 // (This condition is an unlikely backstop only, because the
2392 // PerBytecodeTrapLimit is more likely to take effect first,
2393 // if it is applicable.)
2394 make_not_entrant = true;
2395 }
2396
2397 // Here's more hysteresis: If there has been a recompile at
2398 // this trap point already, run the method in the interpreter
2399 // for a while to exercise it more thoroughly.
2400 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2401 reprofile = true;
2402 }
2403 }
2404
2405 // Take requested actions on the method:
2406
2407 // Recompile
2408 if (make_not_entrant) {
2409 if (!nm->make_not_entrant()) {
2410 return; // the call did not change nmethod's state
2411 }
2412
2413 if (pdata != nullptr) {
2414 // Record the recompilation event, if any.
2415 int tstate0 = pdata->trap_state();
2416 int tstate1 = trap_state_set_recompiled(tstate0, true);
2417 if (tstate1 != tstate0)
2418 pdata->set_trap_state(tstate1);
2419 }
2420
2421 #if INCLUDE_RTM_OPT
2422 // Restart collecting RTM locking abort statistic if the method
2423 // is recompiled for a reason other than RTM state change.
2424 // Assume that in new recompiled code the statistic could be different,
2425 // for example, due to different inlining.
2426 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) &&
2427 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2428 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2429 }
2430 #endif
2431 // For code aging we count traps separately here, using make_not_entrant()
2432 // as a guard against simultaneous deopts in multiple threads.
2433 if (reason == Reason_tenured && trap_mdo != nullptr) {
2434 trap_mdo->inc_tenure_traps();
2435 }
2436 }
2437
2438 if (inc_recompile_count) {
2439 trap_mdo->inc_overflow_recompile_count();
2440 if ((uint)trap_mdo->overflow_recompile_count() >
2441 (uint)PerBytecodeRecompilationCutoff) {
2442 // Give up on the method containing the bad BCI.
2443 if (trap_method() == nm->method()) {
2444 make_not_compilable = true;
2445 } else {
2446 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2447 // But give grace to the enclosing nm->method().
2448 }
2449 }
2450 }
2451
2452 // Reprofile
2453 if (reprofile) {
2454 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2455 }
2456
2457 // Give up compiling
2458 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2459 assert(make_not_entrant, "consistent");
2460 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2461 }
2462
2463 } // Free marked resources
2464
2465 }
2466 JRT_END
2467
2468 ProfileData*
2469 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2470 int trap_bci,
2471 Deoptimization::DeoptReason reason,
2472 bool update_total_trap_count,
2473 #if INCLUDE_JVMCI
2474 bool is_osr,
2475 #endif
2476 Method* compiled_method,
2477 //outputs:
2478 uint& ret_this_trap_count,
2479 bool& ret_maybe_prior_trap,
2480 bool& ret_maybe_prior_recompile) {
2481 bool maybe_prior_trap = false;
2482 bool maybe_prior_recompile = false;
2483 uint this_trap_count = 0;
2484 if (update_total_trap_count) {
2485 uint idx = reason;
2486 #if INCLUDE_JVMCI
2487 if (is_osr) {
2488 // Upper half of history array used for traps in OSR compilations
2489 idx += Reason_TRAP_HISTORY_LENGTH;
2490 }
2491 #endif
2492 uint prior_trap_count = trap_mdo->trap_count(idx);
2493 this_trap_count = trap_mdo->inc_trap_count(idx);
2494
2495 // If the runtime cannot find a place to store trap history,
2496 // it is estimated based on the general condition of the method.
2497 // If the method has ever been recompiled, or has ever incurred
2498 // a trap with the present reason , then this BCI is assumed
2499 // (pessimistically) to be the culprit.
2500 maybe_prior_trap = (prior_trap_count != 0);
2501 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2502 }
2503 ProfileData* pdata = nullptr;
2504
2505
2506 // For reasons which are recorded per bytecode, we check per-BCI data.
2507 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2508 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2509 if (per_bc_reason != Reason_none) {
2510 // Find the profile data for this BCI. If there isn't one,
2511 // try to allocate one from the MDO's set of spares.
2512 // This will let us detect a repeated trap at this point.
2513 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2514
2515 if (pdata != nullptr) {
2516 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2517 if (LogCompilation && xtty != nullptr) {
2518 ttyLocker ttyl;
2519 // no more room for speculative traps in this MDO
2520 xtty->elem("speculative_traps_oom");
2521 }
2522 }
2523 // Query the trap state of this profile datum.
2524 int tstate0 = pdata->trap_state();
2525 if (!trap_state_has_reason(tstate0, per_bc_reason))
2526 maybe_prior_trap = false;
2527 if (!trap_state_is_recompiled(tstate0))
2528 maybe_prior_recompile = false;
2529
2530 // Update the trap state of this profile datum.
2531 int tstate1 = tstate0;
2532 // Record the reason.
2533 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2534 // Store the updated state on the MDO, for next time.
2535 if (tstate1 != tstate0)
2536 pdata->set_trap_state(tstate1);
2537 } else {
2538 if (LogCompilation && xtty != nullptr) {
2539 ttyLocker ttyl;
2540 // Missing MDP? Leave a small complaint in the log.
2541 xtty->elem("missing_mdp bci='%d'", trap_bci);
2542 }
2543 }
2544 }
2545
2546 // Return results:
2547 ret_this_trap_count = this_trap_count;
2548 ret_maybe_prior_trap = maybe_prior_trap;
2549 ret_maybe_prior_recompile = maybe_prior_recompile;
2550 return pdata;
2551 }
2552
2553 void
2554 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2555 ResourceMark rm;
2556 // Ignored outputs:
2557 uint ignore_this_trap_count;
2558 bool ignore_maybe_prior_trap;
2559 bool ignore_maybe_prior_recompile;
2560 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2561 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2562 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2563 query_update_method_data(trap_mdo, trap_bci,
2564 (DeoptReason)reason,
2565 update_total_counts,
2566 #if INCLUDE_JVMCI
2567 false,
2568 #endif
2569 nullptr,
2570 ignore_this_trap_count,
2571 ignore_maybe_prior_trap,
2572 ignore_maybe_prior_recompile);
2573 }
2574
2575 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2576 // Enable WXWrite: current function is called from methods compiled by C2 directly
2577 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2578
2579 // Still in Java no safepoints
2580 {
2581 // This enters VM and may safepoint
2582 uncommon_trap_inner(current, trap_request);
2583 }
2584 HandleMark hm(current);
2585 return fetch_unroll_info_helper(current, exec_mode);
2586 }
2587
2588 // Local derived constants.
2589 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2590 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2591 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2592
2593 //---------------------------trap_state_reason---------------------------------
2594 Deoptimization::DeoptReason
2595 Deoptimization::trap_state_reason(int trap_state) {
2596 // This assert provides the link between the width of DataLayout::trap_bits
2597 // and the encoding of "recorded" reasons. It ensures there are enough
2598 // bits to store all needed reasons in the per-BCI MDO profile.
2599 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2600 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2601 trap_state -= recompile_bit;
2602 if (trap_state == DS_REASON_MASK) {
2603 return Reason_many;
2604 } else {
2605 assert((int)Reason_none == 0, "state=0 => Reason_none");
2606 return (DeoptReason)trap_state;
2607 }
2608 }
2609 //-------------------------trap_state_has_reason-------------------------------
2610 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2611 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2612 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2613 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2614 trap_state -= recompile_bit;
2615 if (trap_state == DS_REASON_MASK) {
2616 return -1; // true, unspecifically (bottom of state lattice)
2617 } else if (trap_state == reason) {
2618 return 1; // true, definitely
2619 } else if (trap_state == 0) {
2620 return 0; // false, definitely (top of state lattice)
2621 } else {
2622 return 0; // false, definitely
2623 }
2624 }
2625 //-------------------------trap_state_add_reason-------------------------------
2626 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2627 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2628 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2629 trap_state -= recompile_bit;
2630 if (trap_state == DS_REASON_MASK) {
2631 return trap_state + recompile_bit; // already at state lattice bottom
2632 } else if (trap_state == reason) {
2633 return trap_state + recompile_bit; // the condition is already true
2634 } else if (trap_state == 0) {
2635 return reason + recompile_bit; // no condition has yet been true
2636 } else {
2637 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2638 }
2639 }
2640 //-----------------------trap_state_is_recompiled------------------------------
2641 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2642 return (trap_state & DS_RECOMPILE_BIT) != 0;
2643 }
2644 //-----------------------trap_state_set_recompiled-----------------------------
2645 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2646 if (z) return trap_state | DS_RECOMPILE_BIT;
2647 else return trap_state & ~DS_RECOMPILE_BIT;
2648 }
2649 //---------------------------format_trap_state---------------------------------
2650 // This is used for debugging and diagnostics, including LogFile output.
2651 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2652 int trap_state) {
2653 assert(buflen > 0, "sanity");
2654 DeoptReason reason = trap_state_reason(trap_state);
2655 bool recomp_flag = trap_state_is_recompiled(trap_state);
2656 // Re-encode the state from its decoded components.
2657 int decoded_state = 0;
2658 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2659 decoded_state = trap_state_add_reason(decoded_state, reason);
2660 if (recomp_flag)
2661 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2662 // If the state re-encodes properly, format it symbolically.
2663 // Because this routine is used for debugging and diagnostics,
2664 // be robust even if the state is a strange value.
2665 size_t len;
2666 if (decoded_state != trap_state) {
2667 // Random buggy state that doesn't decode??
2668 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2669 } else {
2670 len = jio_snprintf(buf, buflen, "%s%s",
2671 trap_reason_name(reason),
2672 recomp_flag ? " recompiled" : "");
2673 }
2674 return buf;
2675 }
2676
2677
2678 //--------------------------------statics--------------------------------------
2679 const char* Deoptimization::_trap_reason_name[] = {
2680 // Note: Keep this in sync. with enum DeoptReason.
2681 "none",
2682 "null_check",
2683 "null_assert" JVMCI_ONLY("_or_unreached0"),
2684 "range_check",
2685 "class_check",
2686 "array_check",
2687 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2688 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2689 "profile_predicate",
2690 "unloaded",
2691 "uninitialized",
2692 "initialized",
2693 "unreached",
2694 "unhandled",
2695 "constraint",
2696 "div0_check",
2697 "age",
2698 "predicate",
2699 "loop_limit_check",
2700 "speculate_class_check",
2701 "speculate_null_check",
2702 "speculate_null_assert",
2703 "rtm_state_change",
2704 "unstable_if",
2705 "unstable_fused_if",
2706 "receiver_constraint",
2707 #if INCLUDE_JVMCI
2708 "aliasing",
2709 "transfer_to_interpreter",
2710 "not_compiled_exception_handler",
2711 "unresolved",
2712 "jsr_mismatch",
2713 #endif
2714 "tenured"
2715 };
2716 const char* Deoptimization::_trap_action_name[] = {
2717 // Note: Keep this in sync. with enum DeoptAction.
2718 "none",
2719 "maybe_recompile",
2720 "reinterpret",
2721 "make_not_entrant",
2722 "make_not_compilable"
2723 };
2724
2725 const char* Deoptimization::trap_reason_name(int reason) {
2726 // Check that every reason has a name
2727 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2728
2729 if (reason == Reason_many) return "many";
2730 if ((uint)reason < Reason_LIMIT)
2731 return _trap_reason_name[reason];
2732 static char buf[20];
2733 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2734 return buf;
2735 }
2736 const char* Deoptimization::trap_action_name(int action) {
2737 // Check that every action has a name
2738 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2739
2740 if ((uint)action < Action_LIMIT)
2741 return _trap_action_name[action];
2742 static char buf[20];
2743 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2744 return buf;
2745 }
2746
2747 // This is used for debugging and diagnostics, including LogFile output.
2748 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2749 int trap_request) {
2750 jint unloaded_class_index = trap_request_index(trap_request);
2751 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2752 const char* action = trap_action_name(trap_request_action(trap_request));
2753 #if INCLUDE_JVMCI
2754 int debug_id = trap_request_debug_id(trap_request);
2755 #endif
2756 size_t len;
2757 if (unloaded_class_index < 0) {
2758 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2759 reason, action
2760 #if INCLUDE_JVMCI
2761 ,debug_id
2762 #endif
2763 );
2764 } else {
2765 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2766 reason, action, unloaded_class_index
2767 #if INCLUDE_JVMCI
2768 ,debug_id
2769 #endif
2770 );
2771 }
2772 return buf;
2773 }
2774
2775 juint Deoptimization::_deoptimization_hist
2776 [Deoptimization::Reason_LIMIT]
2777 [1 + Deoptimization::Action_LIMIT]
2778 [Deoptimization::BC_CASE_LIMIT]
2779 = {0};
2780
2781 enum {
2782 LSB_BITS = 8,
2783 LSB_MASK = right_n_bits(LSB_BITS)
2784 };
2785
2786 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2787 Bytecodes::Code bc) {
2788 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2789 assert(action >= 0 && action < Action_LIMIT, "oob");
2790 _deoptimization_hist[Reason_none][0][0] += 1; // total
2791 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2792 juint* cases = _deoptimization_hist[reason][1+action];
2793 juint* bc_counter_addr = nullptr;
2794 juint bc_counter = 0;
2795 // Look for an unused counter, or an exact match to this BC.
2796 if (bc != Bytecodes::_illegal) {
2797 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2798 juint* counter_addr = &cases[bc_case];
2799 juint counter = *counter_addr;
2800 if ((counter == 0 && bc_counter_addr == nullptr)
2801 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2802 // this counter is either free or is already devoted to this BC
2803 bc_counter_addr = counter_addr;
2804 bc_counter = counter | bc;
2805 }
2806 }
2807 }
2808 if (bc_counter_addr == nullptr) {
2809 // Overflow, or no given bytecode.
2810 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2811 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2812 }
2813 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2814 }
2815
2816 jint Deoptimization::total_deoptimization_count() {
2817 return _deoptimization_hist[Reason_none][0][0];
2818 }
2819
2820 // Get the deopt count for a specific reason and a specific action. If either
2821 // one of 'reason' or 'action' is null, the method returns the sum of all
2822 // deoptimizations with the specific 'action' or 'reason' respectively.
2823 // If both arguments are null, the method returns the total deopt count.
2824 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2825 if (reason_str == nullptr && action_str == nullptr) {
2826 return total_deoptimization_count();
2827 }
2828 juint counter = 0;
2829 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2830 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2831 for (int action = 0; action < Action_LIMIT; action++) {
2832 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2833 juint* cases = _deoptimization_hist[reason][1+action];
2834 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2835 counter += cases[bc_case] >> LSB_BITS;
2836 }
2837 }
2838 }
2839 }
2840 }
2841 return counter;
2842 }
2843
2844 void Deoptimization::print_statistics() {
2845 juint total = total_deoptimization_count();
2846 juint account = total;
2847 if (total != 0) {
2848 ttyLocker ttyl;
2849 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2850 tty->print_cr("Deoptimization traps recorded:");
2851 #define PRINT_STAT_LINE(name, r) \
2852 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2853 PRINT_STAT_LINE("total", total);
2854 // For each non-zero entry in the histogram, print the reason,
2855 // the action, and (if specifically known) the type of bytecode.
2856 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2857 for (int action = 0; action < Action_LIMIT; action++) {
2858 juint* cases = _deoptimization_hist[reason][1+action];
2859 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2860 juint counter = cases[bc_case];
2861 if (counter != 0) {
2862 char name[1*K];
2863 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2864 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2865 bc = Bytecodes::_illegal;
2866 os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
2867 trap_reason_name(reason),
2868 trap_action_name(action),
2869 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2870 juint r = counter >> LSB_BITS;
2871 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2872 account -= r;
2873 }
2874 }
2875 }
2876 }
2877 if (account != 0) {
2878 PRINT_STAT_LINE("unaccounted", account);
2879 }
2880 #undef PRINT_STAT_LINE
2881 if (xtty != nullptr) xtty->tail("statistics");
2882 }
2883 }
2884
2885 #else // COMPILER2_OR_JVMCI
2886
2887
2888 // Stubs for C1 only system.
2889 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2890 return false;
2891 }
2892
2893 const char* Deoptimization::trap_reason_name(int reason) {
2894 return "unknown";
2895 }
2896
2897 jint Deoptimization::total_deoptimization_count() {
2898 return 0;
2899 }
2900
2901 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2902 return 0;
2903 }
2904
2905 void Deoptimization::print_statistics() {
2906 // no output
2907 }
2908
2909 void
2910 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2911 // no update
2912 }
2913
2914 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2915 return 0;
2916 }
2917
2918 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2919 Bytecodes::Code bc) {
2920 // no update
2921 }
2922
2923 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2924 int trap_state) {
2925 jio_snprintf(buf, buflen, "#%d", trap_state);
2926 return buf;
2927 }
2928
2929 #endif // COMPILER2_OR_JVMCI