1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/javaClasses.inline.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmClasses.hpp"
30 #include "code/codeCache.hpp"
31 #include "code/debugInfoRec.hpp"
32 #include "code/nmethod.hpp"
33 #include "code/pcDesc.hpp"
34 #include "code/scopeDesc.hpp"
35 #include "compiler/compilationPolicy.hpp"
36 #include "compiler/compilerDefinitions.inline.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "gc/shared/memAllocator.hpp"
39 #include "interpreter/bytecode.hpp"
40 #include "interpreter/bytecodeStream.hpp"
41 #include "interpreter/interpreter.hpp"
42 #include "interpreter/oopMapCache.hpp"
43 #include "jvm.h"
44 #include "logging/log.hpp"
45 #include "logging/logLevel.hpp"
46 #include "logging/logMessage.hpp"
47 #include "logging/logStream.hpp"
48 #include "memory/allocation.inline.hpp"
49 #include "memory/oopFactory.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "memory/universe.hpp"
52 #include "oops/constantPool.hpp"
53 #include "oops/fieldStreams.inline.hpp"
54 #include "oops/method.hpp"
55 #include "oops/objArrayKlass.hpp"
56 #include "oops/objArrayOop.inline.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "oops/typeArrayOop.inline.hpp"
59 #include "oops/verifyOopClosure.hpp"
60 #include "prims/jvmtiDeferredUpdates.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "prims/jvmtiThreadState.hpp"
63 #include "prims/methodHandles.hpp"
64 #include "prims/vectorSupport.hpp"
65 #include "runtime/atomic.hpp"
66 #include "runtime/continuation.hpp"
67 #include "runtime/continuationEntry.inline.hpp"
68 #include "runtime/deoptimization.hpp"
69 #include "runtime/escapeBarrier.hpp"
70 #include "runtime/fieldDescriptor.hpp"
71 #include "runtime/fieldDescriptor.inline.hpp"
72 #include "runtime/frame.inline.hpp"
73 #include "runtime/handles.inline.hpp"
74 #include "runtime/interfaceSupport.inline.hpp"
75 #include "runtime/javaThread.hpp"
76 #include "runtime/jniHandles.inline.hpp"
77 #include "runtime/keepStackGCProcessed.hpp"
78 #include "runtime/objectMonitor.inline.hpp"
79 #include "runtime/osThread.hpp"
80 #include "runtime/perfData.inline.hpp"
81 #include "runtime/safepointVerifiers.hpp"
82 #include "runtime/sharedRuntime.hpp"
83 #include "runtime/signature.hpp"
84 #include "runtime/stackFrameStream.inline.hpp"
85 #include "runtime/stackValue.hpp"
86 #include "runtime/stackWatermarkSet.hpp"
87 #include "runtime/stubRoutines.hpp"
88 #include "runtime/synchronizer.hpp"
89 #include "runtime/threadSMR.hpp"
90 #include "runtime/threadWXSetters.inline.hpp"
91 #include "runtime/vframe.hpp"
92 #include "runtime/vframeArray.hpp"
93 #include "runtime/vframe_hp.hpp"
94 #include "runtime/vmOperations.hpp"
95 #include "services/management.hpp"
96 #include "utilities/checkedCast.hpp"
97 #include "utilities/events.hpp"
98 #include "utilities/growableArray.hpp"
99 #include "utilities/macros.hpp"
100 #include "utilities/preserveException.hpp"
101 #include "utilities/xmlstream.hpp"
102 #if INCLUDE_JFR
103 #include "jfr/jfrEvents.hpp"
104 #include "jfr/metadata/jfrSerializer.hpp"
105 #endif
106
107 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
108 uint64_t DeoptimizationScope::_active_deopt_gen = 1;
109 bool DeoptimizationScope::_committing_in_progress = false;
110
111 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
112 DEBUG_ONLY(_deopted = false;)
113
114 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
115 // If there is nothing to deopt _required_gen is the same as comitted.
116 _required_gen = DeoptimizationScope::_committed_deopt_gen;
117 }
118
119 DeoptimizationScope::~DeoptimizationScope() {
120 assert(_deopted, "Deopt not executed");
121 }
122
123 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
124 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
125
126 // If it's already marked but we still need it to be deopted.
127 if (nm->is_marked_for_deoptimization()) {
128 dependent(nm);
129 return;
130 }
131
132 nmethod::DeoptimizationStatus status =
133 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
134 Atomic::store(&nm->_deoptimization_status, status);
135
136 // Make sure active is not committed
137 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
138 assert(nm->_deoptimization_generation == 0, "Is already marked");
139
140 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
141 _required_gen = DeoptimizationScope::_active_deopt_gen;
142 }
143
144 void DeoptimizationScope::dependent(nmethod* nm) {
145 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
146
147 // A method marked by someone else may have a _required_gen lower than what we marked with.
148 // Therefore only store it if it's higher than _required_gen.
149 if (_required_gen < nm->_deoptimization_generation) {
150 _required_gen = nm->_deoptimization_generation;
151 }
152 }
153
154 void DeoptimizationScope::deoptimize_marked() {
155 assert(!_deopted, "Already deopted");
156
157 // We are not alive yet.
158 if (!Universe::is_fully_initialized()) {
159 DEBUG_ONLY(_deopted = true;)
160 return;
161 }
162
163 // Safepoints are a special case, handled here.
164 if (SafepointSynchronize::is_at_safepoint()) {
165 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
166 DeoptimizationScope::_active_deopt_gen++;
167 Deoptimization::deoptimize_all_marked();
168 DEBUG_ONLY(_deopted = true;)
169 return;
170 }
171
172 uint64_t comitting = 0;
173 bool wait = false;
174 while (true) {
175 {
176 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
177
178 // First we check if we or someone else already deopted the gen we want.
179 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
180 DEBUG_ONLY(_deopted = true;)
181 return;
182 }
183 if (!_committing_in_progress) {
184 // The version we are about to commit.
185 comitting = DeoptimizationScope::_active_deopt_gen;
186 // Make sure new marks use a higher gen.
187 DeoptimizationScope::_active_deopt_gen++;
188 _committing_in_progress = true;
189 wait = false;
190 } else {
191 // Another thread is handshaking and committing a gen.
192 wait = true;
193 }
194 }
195 if (wait) {
196 // Wait and let the concurrent handshake be performed.
197 ThreadBlockInVM tbivm(JavaThread::current());
198 os::naked_yield();
199 } else {
200 // Performs the handshake.
201 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
202 DEBUG_ONLY(_deopted = true;)
203 {
204 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
205
206 // Make sure that committed doesn't go backwards.
207 // Should only happen if we did a deopt during a safepoint above.
208 if (DeoptimizationScope::_committed_deopt_gen < comitting) {
209 DeoptimizationScope::_committed_deopt_gen = comitting;
210 }
211 _committing_in_progress = false;
212
213 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
214
215 return;
216 }
217 }
218 }
219 }
220
221 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
222 int caller_adjustment,
223 int caller_actual_parameters,
224 int number_of_frames,
225 intptr_t* frame_sizes,
226 address* frame_pcs,
227 BasicType return_type,
228 int exec_mode) {
229 _size_of_deoptimized_frame = size_of_deoptimized_frame;
230 _caller_adjustment = caller_adjustment;
231 _caller_actual_parameters = caller_actual_parameters;
232 _number_of_frames = number_of_frames;
233 _frame_sizes = frame_sizes;
234 _frame_pcs = frame_pcs;
235 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
236 _return_type = return_type;
237 _initial_info = 0;
238 // PD (x86 only)
239 _counter_temp = 0;
240 _unpack_kind = exec_mode;
241 _sender_sp_temp = 0;
242
243 _total_frame_sizes = size_of_frames();
244 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
245 }
246
247 Deoptimization::UnrollBlock::~UnrollBlock() {
248 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
249 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
250 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
251 }
252
253 int Deoptimization::UnrollBlock::size_of_frames() const {
254 // Account first for the adjustment of the initial frame
255 intptr_t result = _caller_adjustment;
256 for (int index = 0; index < number_of_frames(); index++) {
257 result += frame_sizes()[index];
258 }
259 return checked_cast<int>(result);
260 }
261
262 void Deoptimization::UnrollBlock::print() {
263 ResourceMark rm;
264 stringStream st;
265 st.print_cr("UnrollBlock");
266 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
267 st.print( " frame_sizes: ");
268 for (int index = 0; index < number_of_frames(); index++) {
269 st.print(INTX_FORMAT " ", frame_sizes()[index]);
270 }
271 st.cr();
272 tty->print_raw(st.freeze());
273 }
274
275 // In order to make fetch_unroll_info work properly with escape
276 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
277 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
278 // which is called from the method fetch_unroll_info_helper below.
279 JRT_BLOCK_ENTRY_PROF(Deoptimization::UnrollBlock*, Deoptimization, fetch_unroll_info, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
280 // fetch_unroll_info() is called at the beginning of the deoptimization
281 // handler. Note this fact before we start generating temporary frames
282 // that can confuse an asynchronous stack walker. This counter is
283 // decremented at the end of unpack_frames().
284 current->inc_in_deopt_handler();
285
286 if (exec_mode == Unpack_exception) {
287 // When we get here, a callee has thrown an exception into a deoptimized
288 // frame. That throw might have deferred stack watermark checking until
289 // after unwinding. So we deal with such deferred requests here.
290 StackWatermarkSet::after_unwind(current);
291 }
292
293 return fetch_unroll_info_helper(current, exec_mode);
294 JRT_END
295
296 #if COMPILER2_OR_JVMCI
297 // print information about reallocated objects
298 static void print_objects(JavaThread* deoptee_thread,
299 GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
300 ResourceMark rm;
301 stringStream st; // change to logStream with logging
302 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
303 fieldDescriptor fd;
304
305 for (int i = 0; i < objects->length(); i++) {
306 ObjectValue* sv = (ObjectValue*) objects->at(i);
307 Handle obj = sv->value();
308
309 if (obj.is_null()) {
310 st.print_cr(" nullptr");
311 continue;
312 }
313
314 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
315
316 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
317 k->print_value_on(&st);
318 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize);
319
320 if (Verbose && k != nullptr) {
321 k->oop_print_on(obj(), &st);
322 }
323 }
324 tty->print_raw(st.freeze());
325 }
326
327 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
328 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
329 bool& deoptimized_objects) {
330 bool realloc_failures = false;
331 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
332
333 JavaThread* deoptee_thread = chunk->at(0)->thread();
334 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
335 "a frame can only be deoptimized by the owner thread");
336
337 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
338
339 // The flag return_oop() indicates call sites which return oop
340 // in compiled code. Such sites include java method calls,
341 // runtime calls (for example, used to allocate new objects/arrays
342 // on slow code path) and any other calls generated in compiled code.
343 // It is not guaranteed that we can get such information here only
344 // by analyzing bytecode in deoptimized frames. This is why this flag
345 // is set during method compilation (see Compile::Process_OopMap_Node()).
346 // If the previous frame was popped or if we are dispatching an exception,
347 // we don't have an oop result.
348 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
349 Handle return_value;
350 if (save_oop_result) {
351 // Reallocation may trigger GC. If deoptimization happened on return from
352 // call which returns oop we need to save it since it is not in oopmap.
353 oop result = deoptee.saved_oop_result(&map);
354 assert(oopDesc::is_oop_or_null(result), "must be oop");
355 return_value = Handle(thread, result);
356 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
357 if (TraceDeoptimization) {
358 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
359 tty->cr();
360 }
361 }
362 if (objects != nullptr) {
363 if (exec_mode == Deoptimization::Unpack_none) {
364 assert(thread->thread_state() == _thread_in_vm, "assumption");
365 JavaThread* THREAD = thread; // For exception macros.
366 // Clear pending OOM if reallocation fails and return true indicating allocation failure
367 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
368 deoptimized_objects = true;
369 } else {
370 JavaThread* current = thread; // For JRT_BLOCK
371 JRT_BLOCK
372 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
373 JRT_END
374 }
375 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci();
376 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
377 if (TraceDeoptimization) {
378 print_objects(deoptee_thread, objects, realloc_failures);
379 }
380 }
381 if (save_oop_result) {
382 // Restore result.
383 deoptee.set_saved_oop_result(&map, return_value());
384 }
385 return realloc_failures;
386 }
387
388 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
389 frame& deoptee, int exec_mode, bool& deoptimized_objects) {
390 JavaThread* deoptee_thread = chunk->at(0)->thread();
391 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
392 assert(thread == Thread::current(), "should be");
393 HandleMark hm(thread);
394 #ifndef PRODUCT
395 bool first = true;
396 #endif // !PRODUCT
397 // Start locking from outermost/oldest frame
398 for (int i = (chunk->length() - 1); i >= 0; i--) {
399 compiledVFrame* cvf = chunk->at(i);
400 assert (cvf->scope() != nullptr,"expect only compiled java frames");
401 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
402 if (monitors->is_nonempty()) {
403 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
404 exec_mode, realloc_failures);
405 deoptimized_objects = deoptimized_objects || relocked;
406 #ifndef PRODUCT
407 if (PrintDeoptimizationDetails) {
408 ResourceMark rm;
409 stringStream st;
410 for (int j = 0; j < monitors->length(); j++) {
411 MonitorInfo* mi = monitors->at(j);
412 if (mi->eliminated()) {
413 if (first) {
414 first = false;
415 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
416 }
417 if (exec_mode == Deoptimization::Unpack_none) {
418 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
419 if (monitor != nullptr && monitor->object() == mi->owner()) {
420 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
421 continue;
422 }
423 }
424 if (mi->owner_is_scalar_replaced()) {
425 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
426 st.print_cr(" failed reallocation for klass %s", k->external_name());
427 } else {
428 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
429 }
430 }
431 }
432 tty->print_raw(st.freeze());
433 }
434 #endif // !PRODUCT
435 }
436 }
437 }
438
439 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
440 // The given vframes cover one physical frame.
441 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
442 bool& realloc_failures) {
443 frame deoptee = chunk->at(0)->fr();
444 JavaThread* deoptee_thread = chunk->at(0)->thread();
445 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
446 RegisterMap map(chunk->at(0)->register_map());
447 bool deoptimized_objects = false;
448
449 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false);
450
451 // Reallocate the non-escaping objects and restore their fields.
452 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
453 || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
454 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
455 }
456
457 // MonitorInfo structures used in eliminate_locks are not GC safe.
458 NoSafepointVerifier no_safepoint;
459
460 // Now relock objects if synchronization on them was eliminated.
461 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
462 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
463 }
464 return deoptimized_objects;
465 }
466 #endif // COMPILER2_OR_JVMCI
467
468 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
469 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
470 // When we get here we are about to unwind the deoptee frame. In order to
471 // catch not yet safe to use frames, the following stack watermark barrier
472 // poll will make such frames safe to use.
473 StackWatermarkSet::before_unwind(current);
474
475 // Note: there is a safepoint safety issue here. No matter whether we enter
476 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
477 // the vframeArray is created.
478 //
479
480 // Allocate our special deoptimization ResourceMark
481 DeoptResourceMark* dmark = new DeoptResourceMark(current);
482 assert(current->deopt_mark() == nullptr, "Pending deopt!");
483 current->set_deopt_mark(dmark);
484
485 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
486 RegisterMap map(current,
487 RegisterMap::UpdateMap::include,
488 RegisterMap::ProcessFrames::include,
489 RegisterMap::WalkContinuation::skip);
490 RegisterMap dummy_map(current,
491 RegisterMap::UpdateMap::skip,
492 RegisterMap::ProcessFrames::include,
493 RegisterMap::WalkContinuation::skip);
494 // Now get the deoptee with a valid map
495 frame deoptee = stub_frame.sender(&map);
496 // Set the deoptee nmethod
497 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
498 nmethod* nm = deoptee.cb()->as_nmethod_or_null();
499 current->set_deopt_compiled_method(nm);
500
501 if (VerifyStack) {
502 current->validate_frame_layout();
503 }
504
505 // Create a growable array of VFrames where each VFrame represents an inlined
506 // Java frame. This storage is allocated with the usual system arena.
507 assert(deoptee.is_compiled_frame(), "Wrong frame type");
508 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
509 vframe* vf = vframe::new_vframe(&deoptee, &map, current);
510 while (!vf->is_top()) {
511 assert(vf->is_compiled_frame(), "Wrong frame type");
512 chunk->push(compiledVFrame::cast(vf));
513 vf = vf->sender();
514 }
515 assert(vf->is_compiled_frame(), "Wrong frame type");
516 chunk->push(compiledVFrame::cast(vf));
517
518 bool realloc_failures = false;
519
520 #if COMPILER2_OR_JVMCI
521 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
522
523 // Reallocate the non-escaping objects and restore their fields. Then
524 // relock objects if synchronization on them was eliminated.
525 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
526 || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
527 bool unused;
528 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
529 }
530 #endif // COMPILER2_OR_JVMCI
531
532 // Ensure that no safepoint is taken after pointers have been stored
533 // in fields of rematerialized objects. If a safepoint occurs from here on
534 // out the java state residing in the vframeArray will be missed.
535 // Locks may be rebaised in a safepoint.
536 NoSafepointVerifier no_safepoint;
537
538 #if COMPILER2_OR_JVMCI
539 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
540 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
541 bool unused = false;
542 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
543 }
544 #endif // COMPILER2_OR_JVMCI
545
546 ScopeDesc* trap_scope = chunk->at(0)->scope();
547 Handle exceptionObject;
548 if (trap_scope->rethrow_exception()) {
549 #ifndef PRODUCT
550 if (PrintDeoptimizationDetails) {
551 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());
552 }
553 #endif // !PRODUCT
554
555 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
556 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw");
557 ScopeValue* topOfStack = expressions->top();
558 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
559 guarantee(exceptionObject() != nullptr, "exception oop can not be null");
560 }
561
562 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
563 #if COMPILER2_OR_JVMCI
564 if (realloc_failures) {
565 // This destroys all ScopedValue bindings.
566 current->clear_scopedValueBindings();
567 pop_frames_failed_reallocs(current, array);
568 }
569 #endif
570
571 assert(current->vframe_array_head() == nullptr, "Pending deopt!");
572 current->set_vframe_array_head(array);
573
574 // Now that the vframeArray has been created if we have any deferred local writes
575 // added by jvmti then we can free up that structure as the data is now in the
576 // vframeArray
577
578 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
579
580 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
581 CodeBlob* cb = stub_frame.cb();
582 // Verify we have the right vframeArray
583 assert(cb->frame_size() >= 0, "Unexpected frame size");
584 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
585
586 // If the deopt call site is a MethodHandle invoke call site we have
587 // to adjust the unpack_sp.
588 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
589 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
590 unpack_sp = deoptee.unextended_sp();
591
592 #ifdef ASSERT
593 assert(cb->is_deoptimization_stub() ||
594 cb->is_uncommon_trap_stub() ||
595 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
596 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
597 "unexpected code blob: %s", cb->name());
598 #endif
599
600 // This is a guarantee instead of an assert because if vframe doesn't match
601 // we will unpack the wrong deoptimized frame and wind up in strange places
602 // where it will be very difficult to figure out what went wrong. Better
603 // to die an early death here than some very obscure death later when the
604 // trail is cold.
605 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
606 // in that it will fail to detect a problem when there is one. This needs
607 // more work in tiger timeframe.
608 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
609
610 int number_of_frames = array->frames();
611
612 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
613 // virtual activation, which is the reverse of the elements in the vframes array.
614 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
615 // +1 because we always have an interpreter return address for the final slot.
616 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
617 int popframe_extra_args = 0;
618 // Create an interpreter return address for the stub to use as its return
619 // address so the skeletal frames are perfectly walkable
620 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
621
622 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
623 // activation be put back on the expression stack of the caller for reexecution
624 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
625 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
626 }
627
628 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
629 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
630 // than simply use array->sender.pc(). This requires us to walk the current set of frames
631 //
632 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
633 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
634
635 // It's possible that the number of parameters at the call site is
636 // different than number of arguments in the callee when method
637 // handles are used. If the caller is interpreted get the real
638 // value so that the proper amount of space can be added to it's
639 // frame.
640 bool caller_was_method_handle = false;
641 if (deopt_sender.is_interpreted_frame()) {
642 methodHandle method(current, deopt_sender.interpreter_frame_method());
643 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
644 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
645 // Method handle invokes may involve fairly arbitrary chains of
646 // calls so it's impossible to know how much actual space the
647 // caller has for locals.
648 caller_was_method_handle = true;
649 }
650 }
651
652 //
653 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
654 // frame_sizes/frame_pcs[1] next oldest frame (int)
655 // frame_sizes/frame_pcs[n] youngest frame (int)
656 //
657 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
658 // owns the space for the return address to it's caller). Confusing ain't it.
659 //
660 // The vframe array can address vframes with indices running from
661 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
662 // When we create the skeletal frames we need the oldest frame to be in the zero slot
663 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
664 // so things look a little strange in this loop.
665 //
666 int callee_parameters = 0;
667 int callee_locals = 0;
668 for (int index = 0; index < array->frames(); index++ ) {
669 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
670 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
671 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
672 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
673 callee_locals,
674 index == 0,
675 popframe_extra_args);
676 // This pc doesn't have to be perfect just good enough to identify the frame
677 // as interpreted so the skeleton frame will be walkable
678 // The correct pc will be set when the skeleton frame is completely filled out
679 // The final pc we store in the loop is wrong and will be overwritten below
680 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
681
682 callee_parameters = array->element(index)->method()->size_of_parameters();
683 callee_locals = array->element(index)->method()->max_locals();
684 popframe_extra_args = 0;
685 }
686
687 // Compute whether the root vframe returns a float or double value.
688 BasicType return_type;
689 {
690 methodHandle method(current, array->element(0)->method());
691 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
692 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
693 }
694
695 // Compute information for handling adapters and adjusting the frame size of the caller.
696 int caller_adjustment = 0;
697
698 // Compute the amount the oldest interpreter frame will have to adjust
699 // its caller's stack by. If the caller is a compiled frame then
700 // we pretend that the callee has no parameters so that the
701 // extension counts for the full amount of locals and not just
702 // locals-parms. This is because without a c2i adapter the parm
703 // area as created by the compiled frame will not be usable by
704 // the interpreter. (Depending on the calling convention there
705 // may not even be enough space).
706
707 // QQQ I'd rather see this pushed down into last_frame_adjust
708 // and have it take the sender (aka caller).
709
710 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
711 caller_adjustment = last_frame_adjust(0, callee_locals);
712 } else if (callee_locals > callee_parameters) {
713 // The caller frame may need extending to accommodate
714 // non-parameter locals of the first unpacked interpreted frame.
715 // Compute that adjustment.
716 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
717 }
718
719 // If the sender is deoptimized the we must retrieve the address of the handler
720 // since the frame will "magically" show the original pc before the deopt
721 // and we'd undo the deopt.
722
723 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
724 if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
725 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
726 }
727
728 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
729
730 #if INCLUDE_JVMCI
731 if (exceptionObject() != nullptr) {
732 current->set_exception_oop(exceptionObject());
733 exec_mode = Unpack_exception;
734 }
735 #endif
736
737 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
738 assert(current->has_pending_exception(), "should have thrown OOME");
739 current->set_exception_oop(current->pending_exception());
740 current->clear_pending_exception();
741 exec_mode = Unpack_exception;
742 }
743
744 #if INCLUDE_JVMCI
745 if (current->frames_to_pop_failed_realloc() > 0) {
746 current->set_pending_monitorenter(false);
747 }
748 #endif
749
750 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
751 caller_adjustment * BytesPerWord,
752 caller_was_method_handle ? 0 : callee_parameters,
753 number_of_frames,
754 frame_sizes,
755 frame_pcs,
756 return_type,
757 exec_mode);
758 // On some platforms, we need a way to pass some platform dependent
759 // information to the unpacking code so the skeletal frames come out
760 // correct (initial fp value, unextended sp, ...)
761 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
762
763 if (array->frames() > 1) {
764 if (VerifyStack && TraceDeoptimization) {
765 tty->print_cr("Deoptimizing method containing inlining");
766 }
767 }
768
769 array->set_unroll_block(info);
770 return info;
771 }
772
773 // Called to cleanup deoptimization data structures in normal case
774 // after unpacking to stack and when stack overflow error occurs
775 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
776 vframeArray *array) {
777
778 // Get array if coming from exception
779 if (array == nullptr) {
780 array = thread->vframe_array_head();
781 }
782 thread->set_vframe_array_head(nullptr);
783
784 // Free the previous UnrollBlock
785 vframeArray* old_array = thread->vframe_array_last();
786 thread->set_vframe_array_last(array);
787
788 if (old_array != nullptr) {
789 UnrollBlock* old_info = old_array->unroll_block();
790 old_array->set_unroll_block(nullptr);
791 delete old_info;
792 delete old_array;
793 }
794
795 // Deallocate any resource creating in this routine and any ResourceObjs allocated
796 // inside the vframeArray (StackValueCollections)
797
798 delete thread->deopt_mark();
799 thread->set_deopt_mark(nullptr);
800 thread->set_deopt_compiled_method(nullptr);
801
802
803 if (JvmtiExport::can_pop_frame()) {
804 // Regardless of whether we entered this routine with the pending
805 // popframe condition bit set, we should always clear it now
806 thread->clear_popframe_condition();
807 }
808
809 // unpack_frames() is called at the end of the deoptimization handler
810 // and (in C2) at the end of the uncommon trap handler. Note this fact
811 // so that an asynchronous stack walker can work again. This counter is
812 // incremented at the beginning of fetch_unroll_info() and (in C2) at
813 // the beginning of uncommon_trap().
814 thread->dec_in_deopt_handler();
815 }
816
817 // Moved from cpu directories because none of the cpus has callee save values.
818 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
819 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
820
821 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
822 // the days we had adapter frames. When we deoptimize a situation where a
823 // compiled caller calls a compiled caller will have registers it expects
824 // to survive the call to the callee. If we deoptimize the callee the only
825 // way we can restore these registers is to have the oldest interpreter
826 // frame that we create restore these values. That is what this routine
827 // will accomplish.
828
829 // At the moment we have modified c2 to not have any callee save registers
830 // so this problem does not exist and this routine is just a place holder.
831
832 assert(f->is_interpreted_frame(), "must be interpreted");
833 }
834
835 #ifndef PRODUCT
836 static bool falls_through(Bytecodes::Code bc) {
837 switch (bc) {
838 // List may be incomplete. Here we really only care about bytecodes where compiled code
839 // can deoptimize.
840 case Bytecodes::_goto:
841 case Bytecodes::_goto_w:
842 case Bytecodes::_athrow:
843 return false;
844 default:
845 return true;
846 }
847 }
848 #endif
849
850 // Return BasicType of value being returned
851 JRT_LEAF_PROF_NO_THREAD(BasicType, Deoptimization, unpack_frames, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
852 assert(thread == JavaThread::current(), "pre-condition");
853
854 // We are already active in the special DeoptResourceMark any ResourceObj's we
855 // allocate will be freed at the end of the routine.
856
857 // JRT_LEAF methods don't normally allocate handles and there is a
858 // NoHandleMark to enforce that. It is actually safe to use Handles
859 // in a JRT_LEAF method, and sometimes desirable, but to do so we
860 // must use ResetNoHandleMark to bypass the NoHandleMark, and
861 // then use a HandleMark to ensure any Handles we do create are
862 // cleaned up in this scope.
863 ResetNoHandleMark rnhm;
864 HandleMark hm(thread);
865
866 frame stub_frame = thread->last_frame();
867
868 Continuation::notify_deopt(thread, stub_frame.sp());
869
870 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
871 // must point to the vframeArray for the unpack frame.
872 vframeArray* array = thread->vframe_array_head();
873 UnrollBlock* info = array->unroll_block();
874
875 // We set the last_Java frame. But the stack isn't really parsable here. So we
876 // clear it to make sure JFR understands not to try and walk stacks from events
877 // in here.
878 intptr_t* sp = thread->frame_anchor()->last_Java_sp();
879 thread->frame_anchor()->set_last_Java_sp(nullptr);
880
881 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
882 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
883
884 thread->frame_anchor()->set_last_Java_sp(sp);
885
886 BasicType bt = info->return_type();
887
888 // If we have an exception pending, claim that the return type is an oop
889 // so the deopt_blob does not overwrite the exception_oop.
890
891 if (exec_mode == Unpack_exception)
892 bt = T_OBJECT;
893
894 // Cleanup thread deopt data
895 cleanup_deopt_info(thread, array);
896
897 #ifndef PRODUCT
898 if (VerifyStack) {
899 ResourceMark res_mark;
900 // Clear pending exception to not break verification code (restored afterwards)
901 PreserveExceptionMark pm(thread);
902
903 thread->validate_frame_layout();
904
905 // Verify that the just-unpacked frames match the interpreter's
906 // notions of expression stack and locals
907 vframeArray* cur_array = thread->vframe_array_last();
908 RegisterMap rm(thread,
909 RegisterMap::UpdateMap::skip,
910 RegisterMap::ProcessFrames::include,
911 RegisterMap::WalkContinuation::skip);
912 rm.set_include_argument_oops(false);
913 bool is_top_frame = true;
914 int callee_size_of_parameters = 0;
915 int callee_max_locals = 0;
916 for (int i = 0; i < cur_array->frames(); i++) {
917 vframeArrayElement* el = cur_array->element(i);
918 frame* iframe = el->iframe();
919 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
920
921 // Get the oop map for this bci
922 InterpreterOopMap mask;
923 int cur_invoke_parameter_size = 0;
924 bool try_next_mask = false;
925 int next_mask_expression_stack_size = -1;
926 int top_frame_expression_stack_adjustment = 0;
927 methodHandle mh(thread, iframe->interpreter_frame_method());
928 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
929 BytecodeStream str(mh, iframe->interpreter_frame_bci());
930 int max_bci = mh->code_size();
931 // Get to the next bytecode if possible
932 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
933 // Check to see if we can grab the number of outgoing arguments
934 // at an uncommon trap for an invoke (where the compiler
935 // generates debug info before the invoke has executed)
936 Bytecodes::Code cur_code = str.next();
937 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere;
938 if (Bytecodes::is_invoke(cur_code)) {
939 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
940 cur_invoke_parameter_size = invoke.size_of_parameters();
941 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
942 callee_size_of_parameters++;
943 }
944 }
945 if (str.bci() < max_bci) {
946 next_code = str.next();
947 if (next_code >= 0) {
948 // The interpreter oop map generator reports results before
949 // the current bytecode has executed except in the case of
950 // calls. It seems to be hard to tell whether the compiler
951 // has emitted debug information matching the "state before"
952 // a given bytecode or the state after, so we try both
953 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) {
954 // Get expression stack size for the next bytecode
955 InterpreterOopMap next_mask;
956 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
957 next_mask_expression_stack_size = next_mask.expression_stack_size();
958 if (Bytecodes::is_invoke(next_code)) {
959 Bytecode_invoke invoke(mh, str.bci());
960 next_mask_expression_stack_size += invoke.size_of_parameters();
961 }
962 // Need to subtract off the size of the result type of
963 // the bytecode because this is not described in the
964 // debug info but returned to the interpreter in the TOS
965 // caching register
966 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
967 if (bytecode_result_type != T_ILLEGAL) {
968 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
969 }
970 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
971 try_next_mask = true;
972 }
973 }
974 }
975
976 // Verify stack depth and oops in frame
977 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
978 if (!(
979 /* SPARC */
980 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
981 /* x86 */
982 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
983 (try_next_mask &&
984 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
985 top_frame_expression_stack_adjustment))) ||
986 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
987 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
988 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
989 )) {
990 {
991 // Print out some information that will help us debug the problem
992 tty->print_cr("Wrong number of expression stack elements during deoptimization");
993 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
994 tty->print_cr(" Current code %s", Bytecodes::name(cur_code));
995 if (try_next_mask) {
996 tty->print_cr(" Next code %s", Bytecodes::name(next_code));
997 }
998 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
999 iframe->interpreter_frame_expression_stack_size());
1000 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1001 tty->print_cr(" try_next_mask = %d", try_next_mask);
1002 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
1003 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
1004 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
1005 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1006 tty->print_cr(" exec_mode = %d", exec_mode);
1007 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1008 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1009 tty->print_cr(" Interpreted frames:");
1010 for (int k = 0; k < cur_array->frames(); k++) {
1011 vframeArrayElement* el = cur_array->element(k);
1012 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1013 }
1014 cur_array->print_on_2(tty);
1015 }
1016 guarantee(false, "wrong number of expression stack elements during deopt");
1017 }
1018 VerifyOopClosure verify;
1019 iframe->oops_interpreted_do(&verify, &rm, false);
1020 callee_size_of_parameters = mh->size_of_parameters();
1021 callee_max_locals = mh->max_locals();
1022 is_top_frame = false;
1023 }
1024 }
1025 #endif // !PRODUCT
1026
1027 return bt;
1028 JRT_END
1029
1030 class DeoptimizeMarkedClosure : public HandshakeClosure {
1031 public:
1032 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
1033 void do_thread(Thread* thread) {
1034 JavaThread* jt = JavaThread::cast(thread);
1035 jt->deoptimize_marked_methods();
1036 }
1037 };
1038
1039 void Deoptimization::deoptimize_all_marked() {
1040 ResourceMark rm;
1041
1042 // Make the dependent methods not entrant
1043 CodeCache::make_marked_nmethods_deoptimized();
1044
1045 DeoptimizeMarkedClosure deopt;
1046 if (SafepointSynchronize::is_at_safepoint()) {
1047 Threads::java_threads_do(&deopt);
1048 } else {
1049 Handshake::execute(&deopt);
1050 }
1051 }
1052
1053 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1054 = Deoptimization::Action_reinterpret;
1055
1056 #if INCLUDE_JVMCI
1057 template<typename CacheType>
1058 class BoxCacheBase : public CHeapObj<mtCompiler> {
1059 protected:
1060 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1061 ResourceMark rm(thread);
1062 char* klass_name_str = klass_name->as_C_string();
1063 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle());
1064 guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1065 if (!ik->is_in_error_state()) {
1066 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1067 CacheType::compute_offsets(ik);
1068 }
1069 return ik;
1070 }
1071 };
1072
1073 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
1074 PrimitiveType _low;
1075 PrimitiveType _high;
1076 jobject _cache;
1077 protected:
1078 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1079 BoxCache(Thread* thread) {
1080 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1081 if (ik->is_in_error_state()) {
1082 _low = 1;
1083 _high = 0;
1084 _cache = nullptr;
1085 } else {
1086 objArrayOop cache = CacheType::cache(ik);
1087 assert(cache->length() > 0, "Empty cache");
1088 _low = BoxType::value(cache->obj_at(0));
1089 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1090 _cache = JNIHandles::make_global(Handle(thread, cache));
1091 }
1092 }
1093 ~BoxCache() {
1094 JNIHandles::destroy_global(_cache);
1095 }
1096 public:
1097 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1098 if (_singleton == nullptr) {
1099 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1100 if (!Atomic::replace_if_null(&_singleton, s)) {
1101 delete s;
1102 }
1103 }
1104 return _singleton;
1105 }
1106 oop lookup(PrimitiveType value) {
1107 if (_low <= value && value <= _high) {
1108 int offset = checked_cast<int>(value - _low);
1109 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1110 }
1111 return nullptr;
1112 }
1113 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1114 if (_cache == nullptr) {
1115 cache_init_error = true;
1116 return nullptr;
1117 }
1118 // Have to cast to avoid little/big-endian problems.
1119 if (sizeof(PrimitiveType) > sizeof(jint)) {
1120 jlong value = (jlong)raw_value;
1121 return lookup(value);
1122 }
1123 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1124 return lookup(value);
1125 }
1126 };
1127
1128 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1129 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1130 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1131 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1132 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1133
1134 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1135 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1136 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1137 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1138 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1139
1140 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1141 jobject _true_cache;
1142 jobject _false_cache;
1143 protected:
1144 static BooleanBoxCache *_singleton;
1145 BooleanBoxCache(Thread *thread) {
1146 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1147 if (ik->is_in_error_state()) {
1148 _true_cache = nullptr;
1149 _false_cache = nullptr;
1150 } else {
1151 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1152 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1153 }
1154 }
1155 ~BooleanBoxCache() {
1156 JNIHandles::destroy_global(_true_cache);
1157 JNIHandles::destroy_global(_false_cache);
1158 }
1159 public:
1160 static BooleanBoxCache* singleton(Thread* thread) {
1161 if (_singleton == nullptr) {
1162 BooleanBoxCache* s = new BooleanBoxCache(thread);
1163 if (!Atomic::replace_if_null(&_singleton, s)) {
1164 delete s;
1165 }
1166 }
1167 return _singleton;
1168 }
1169 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1170 if (_true_cache == nullptr) {
1171 cache_in_error = true;
1172 return nullptr;
1173 }
1174 // Have to cast to avoid little/big-endian problems.
1175 jboolean value = (jboolean)*((jint*)&raw_value);
1176 return lookup(value);
1177 }
1178 oop lookup(jboolean value) {
1179 if (value != 0) {
1180 return JNIHandles::resolve_non_null(_true_cache);
1181 }
1182 return JNIHandles::resolve_non_null(_false_cache);
1183 }
1184 };
1185
1186 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1187
1188 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1189 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1190 BasicType box_type = vmClasses::box_klass_type(k);
1191 if (box_type != T_OBJECT) {
1192 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1193 switch(box_type) {
1194 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1195 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1196 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1197 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1198 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1199 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1200 default:;
1201 }
1202 }
1203 return nullptr;
1204 }
1205 #endif // INCLUDE_JVMCI
1206
1207 #if COMPILER2_OR_JVMCI
1208 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1209 Handle pending_exception(THREAD, thread->pending_exception());
1210 const char* exception_file = thread->exception_file();
1211 int exception_line = thread->exception_line();
1212 thread->clear_pending_exception();
1213
1214 bool failures = false;
1215
1216 for (int i = 0; i < objects->length(); i++) {
1217 assert(objects->at(i)->is_object(), "invalid debug information");
1218 ObjectValue* sv = (ObjectValue*) objects->at(i);
1219
1220 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1221 oop obj = nullptr;
1222
1223 bool cache_init_error = false;
1224 if (k->is_instance_klass()) {
1225 #if INCLUDE_JVMCI
1226 nmethod* nm = fr->cb()->as_nmethod_or_null();
1227 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1228 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1229 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1230 if (obj != nullptr) {
1231 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1232 abv->set_cached(true);
1233 } else if (cache_init_error) {
1234 // Results in an OOME which is valid (as opposed to a class initialization error)
1235 // and is fine for the rare case a cache initialization failing.
1236 failures = true;
1237 }
1238 }
1239 #endif // INCLUDE_JVMCI
1240
1241 InstanceKlass* ik = InstanceKlass::cast(k);
1242 if (obj == nullptr && !cache_init_error) {
1243 InternalOOMEMark iom(THREAD);
1244 #if COMPILER2_OR_JVMCI
1245 if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1246 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1247 } else {
1248 obj = ik->allocate_instance(THREAD);
1249 }
1250 #else
1251 obj = ik->allocate_instance(THREAD);
1252 #endif // COMPILER2_OR_JVMCI
1253 }
1254 } else if (k->is_typeArray_klass()) {
1255 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1256 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1257 int len = sv->field_size() / type2size[ak->element_type()];
1258 InternalOOMEMark iom(THREAD);
1259 obj = ak->allocate(len, THREAD);
1260 } else if (k->is_objArray_klass()) {
1261 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1262 InternalOOMEMark iom(THREAD);
1263 obj = ak->allocate(sv->field_size(), THREAD);
1264 }
1265
1266 if (obj == nullptr) {
1267 failures = true;
1268 }
1269
1270 assert(sv->value().is_null(), "redundant reallocation");
1271 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1272 CLEAR_PENDING_EXCEPTION;
1273 sv->set_value(obj);
1274 }
1275
1276 if (failures) {
1277 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1278 } else if (pending_exception.not_null()) {
1279 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1280 }
1281
1282 return failures;
1283 }
1284
1285 #if INCLUDE_JVMCI
1286 /**
1287 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1288 * we need to somehow be able to recover the actual kind to be able to write the correct
1289 * amount of bytes.
1290 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1291 * the entries at index n + 1 to n + i are 'markers'.
1292 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1293 * expected form of the array would be:
1294 *
1295 * {b0, b1, b2, b3, INT, marker, b6, b7}
1296 *
1297 * Thus, in order to get back the size of the entry, we simply need to count the number
1298 * of marked entries
1299 *
1300 * @param virtualArray the virtualized byte array
1301 * @param i index of the virtual entry we are recovering
1302 * @return The number of bytes the entry spans
1303 */
1304 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1305 int index = i;
1306 while (++index < virtualArray->field_size() &&
1307 virtualArray->field_at(index)->is_marker()) {}
1308 return index - i;
1309 }
1310
1311 /**
1312 * If there was a guarantee for byte array to always start aligned to a long, we could
1313 * do a simple check on the parity of the index. Unfortunately, that is not always the
1314 * case. Thus, we check alignment of the actual address we are writing to.
1315 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1316 * write a long to index 3.
1317 */
1318 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1319 jbyte* res = obj->byte_at_addr(index);
1320 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1321 return res;
1322 }
1323
1324 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1325 switch (byte_count) {
1326 case 1:
1327 obj->byte_at_put(index, (jbyte) value->get_jint());
1328 break;
1329 case 2:
1330 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1331 break;
1332 case 4:
1333 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1334 break;
1335 case 8:
1336 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1337 break;
1338 default:
1339 ShouldNotReachHere();
1340 }
1341 }
1342 #endif // INCLUDE_JVMCI
1343
1344
1345 // restore elements of an eliminated type array
1346 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1347 int index = 0;
1348
1349 for (int i = 0; i < sv->field_size(); i++) {
1350 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1351 switch(type) {
1352 case T_LONG: case T_DOUBLE: {
1353 assert(value->type() == T_INT, "Agreement.");
1354 StackValue* low =
1355 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1356 #ifdef _LP64
1357 jlong res = (jlong)low->get_intptr();
1358 #else
1359 jlong res = jlong_from(value->get_jint(), low->get_jint());
1360 #endif
1361 obj->long_at_put(index, res);
1362 break;
1363 }
1364
1365 case T_INT: case T_FLOAT: { // 4 bytes.
1366 assert(value->type() == T_INT, "Agreement.");
1367 bool big_value = false;
1368 if (i + 1 < sv->field_size() && type == T_INT) {
1369 if (sv->field_at(i)->is_location()) {
1370 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1371 if (type == Location::dbl || type == Location::lng) {
1372 big_value = true;
1373 }
1374 } else if (sv->field_at(i)->is_constant_int()) {
1375 ScopeValue* next_scope_field = sv->field_at(i + 1);
1376 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1377 big_value = true;
1378 }
1379 }
1380 }
1381
1382 if (big_value) {
1383 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1384 #ifdef _LP64
1385 jlong res = (jlong)low->get_intptr();
1386 #else
1387 jlong res = jlong_from(value->get_jint(), low->get_jint());
1388 #endif
1389 obj->int_at_put(index, *(jint*)&res);
1390 obj->int_at_put(++index, *((jint*)&res + 1));
1391 } else {
1392 obj->int_at_put(index, value->get_jint());
1393 }
1394 break;
1395 }
1396
1397 case T_SHORT:
1398 assert(value->type() == T_INT, "Agreement.");
1399 obj->short_at_put(index, (jshort)value->get_jint());
1400 break;
1401
1402 case T_CHAR:
1403 assert(value->type() == T_INT, "Agreement.");
1404 obj->char_at_put(index, (jchar)value->get_jint());
1405 break;
1406
1407 case T_BYTE: {
1408 assert(value->type() == T_INT, "Agreement.");
1409 #if INCLUDE_JVMCI
1410 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1411 int byte_count = count_number_of_bytes_for_entry(sv, i);
1412 byte_array_put(obj, value, index, byte_count);
1413 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1414 i += byte_count - 1; // Balance the loop counter.
1415 index += byte_count;
1416 // index has been updated so continue at top of loop
1417 continue;
1418 #else
1419 obj->byte_at_put(index, (jbyte)value->get_jint());
1420 break;
1421 #endif // INCLUDE_JVMCI
1422 }
1423
1424 case T_BOOLEAN: {
1425 assert(value->type() == T_INT, "Agreement.");
1426 obj->bool_at_put(index, (jboolean)value->get_jint());
1427 break;
1428 }
1429
1430 default:
1431 ShouldNotReachHere();
1432 }
1433 index++;
1434 }
1435 }
1436
1437 // restore fields of an eliminated object array
1438 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1439 for (int i = 0; i < sv->field_size(); i++) {
1440 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1441 assert(value->type() == T_OBJECT, "object element expected");
1442 obj->obj_at_put(i, value->get_obj()());
1443 }
1444 }
1445
1446 class ReassignedField {
1447 public:
1448 int _offset;
1449 BasicType _type;
1450 public:
1451 ReassignedField() {
1452 _offset = 0;
1453 _type = T_ILLEGAL;
1454 }
1455 };
1456
1457 static int compare(ReassignedField* left, ReassignedField* right) {
1458 return left->_offset - right->_offset;
1459 }
1460
1461 // Restore fields of an eliminated instance object using the same field order
1462 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1463 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1464 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1465 InstanceKlass* ik = klass;
1466 while (ik != nullptr) {
1467 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1468 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1469 ReassignedField field;
1470 field._offset = fs.offset();
1471 field._type = Signature::basic_type(fs.signature());
1472 fields->append(field);
1473 }
1474 }
1475 ik = ik->superklass();
1476 }
1477 fields->sort(compare);
1478 for (int i = 0; i < fields->length(); i++) {
1479 ScopeValue* scope_field = sv->field_at(svIndex);
1480 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1481 int offset = fields->at(i)._offset;
1482 BasicType type = fields->at(i)._type;
1483 switch (type) {
1484 case T_OBJECT: case T_ARRAY:
1485 assert(value->type() == T_OBJECT, "Agreement.");
1486 obj->obj_field_put(offset, value->get_obj()());
1487 break;
1488
1489 case T_INT: case T_FLOAT: { // 4 bytes.
1490 assert(value->type() == T_INT, "Agreement.");
1491 bool big_value = false;
1492 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1493 if (scope_field->is_location()) {
1494 Location::Type type = ((LocationValue*) scope_field)->location().type();
1495 if (type == Location::dbl || type == Location::lng) {
1496 big_value = true;
1497 }
1498 }
1499 if (scope_field->is_constant_int()) {
1500 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1501 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1502 big_value = true;
1503 }
1504 }
1505 }
1506
1507 if (big_value) {
1508 i++;
1509 assert(i < fields->length(), "second T_INT field needed");
1510 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1511 } else {
1512 obj->int_field_put(offset, value->get_jint());
1513 break;
1514 }
1515 }
1516 /* no break */
1517
1518 case T_LONG: case T_DOUBLE: {
1519 assert(value->type() == T_INT, "Agreement.");
1520 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1521 #ifdef _LP64
1522 jlong res = (jlong)low->get_intptr();
1523 #else
1524 jlong res = jlong_from(value->get_jint(), low->get_jint());
1525 #endif
1526 obj->long_field_put(offset, res);
1527 break;
1528 }
1529
1530 case T_SHORT:
1531 assert(value->type() == T_INT, "Agreement.");
1532 obj->short_field_put(offset, (jshort)value->get_jint());
1533 break;
1534
1535 case T_CHAR:
1536 assert(value->type() == T_INT, "Agreement.");
1537 obj->char_field_put(offset, (jchar)value->get_jint());
1538 break;
1539
1540 case T_BYTE:
1541 assert(value->type() == T_INT, "Agreement.");
1542 obj->byte_field_put(offset, (jbyte)value->get_jint());
1543 break;
1544
1545 case T_BOOLEAN:
1546 assert(value->type() == T_INT, "Agreement.");
1547 obj->bool_field_put(offset, (jboolean)value->get_jint());
1548 break;
1549
1550 default:
1551 ShouldNotReachHere();
1552 }
1553 svIndex++;
1554 }
1555 return svIndex;
1556 }
1557
1558 // restore fields of all eliminated objects and arrays
1559 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1560 for (int i = 0; i < objects->length(); i++) {
1561 assert(objects->at(i)->is_object(), "invalid debug information");
1562 ObjectValue* sv = (ObjectValue*) objects->at(i);
1563 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1564 Handle obj = sv->value();
1565 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1566 #ifndef PRODUCT
1567 if (PrintDeoptimizationDetails) {
1568 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1569 }
1570 #endif // !PRODUCT
1571
1572 if (obj.is_null()) {
1573 continue;
1574 }
1575
1576 #if INCLUDE_JVMCI
1577 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1578 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1579 continue;
1580 }
1581 #endif // INCLUDE_JVMCI
1582 #if COMPILER2_OR_JVMCI
1583 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1584 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1585 ScopeValue* payload = sv->field_at(0);
1586 if (payload->is_location() &&
1587 payload->as_LocationValue()->location().type() == Location::vector) {
1588 #ifndef PRODUCT
1589 if (PrintDeoptimizationDetails) {
1590 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1591 if (Verbose) {
1592 Handle obj = sv->value();
1593 k->oop_print_on(obj(), tty);
1594 }
1595 }
1596 #endif // !PRODUCT
1597 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1598 }
1599 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1600 // which could be restored after vector object allocation.
1601 }
1602 #endif /* !COMPILER2_OR_JVMCI */
1603 if (k->is_instance_klass()) {
1604 InstanceKlass* ik = InstanceKlass::cast(k);
1605 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1606 } else if (k->is_typeArray_klass()) {
1607 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1608 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1609 } else if (k->is_objArray_klass()) {
1610 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1611 }
1612 }
1613 // These objects may escape when we return to Interpreter after deoptimization.
1614 // We need barrier so that stores that initialize these objects can't be reordered
1615 // with subsequent stores that make these objects accessible by other threads.
1616 OrderAccess::storestore();
1617 }
1618
1619
1620 // relock objects for which synchronization was eliminated
1621 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1622 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1623 bool relocked_objects = false;
1624 for (int i = 0; i < monitors->length(); i++) {
1625 MonitorInfo* mon_info = monitors->at(i);
1626 if (mon_info->eliminated()) {
1627 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1628 relocked_objects = true;
1629 if (!mon_info->owner_is_scalar_replaced()) {
1630 Handle obj(thread, mon_info->owner());
1631 markWord mark = obj->mark();
1632 if (exec_mode == Unpack_none) {
1633 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1634 // With exec_mode == Unpack_none obj may be thread local and locked in
1635 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1636 markWord dmw = mark.displaced_mark_helper();
1637 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1638 obj->set_mark(dmw);
1639 }
1640 if (mark.has_monitor()) {
1641 // defer relocking if the deoptee thread is currently waiting for obj
1642 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1643 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1644 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1645 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1646 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1647 continue;
1648 }
1649 }
1650 }
1651 if (LockingMode == LM_LIGHTWEIGHT) {
1652 // We have lost information about the correct state of the lock stack.
1653 // Inflate the locks instead. Enter then inflate to avoid races with
1654 // deflation.
1655 ObjectSynchronizer::enter_for(obj, nullptr, deoptee_thread);
1656 assert(mon_info->owner()->is_locked(), "object must be locked now");
1657 ObjectMonitor* mon = ObjectSynchronizer::inflate_for(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal);
1658 assert(mon->owner() == deoptee_thread, "must be");
1659 } else {
1660 BasicLock* lock = mon_info->lock();
1661 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1662 assert(mon_info->owner()->is_locked(), "object must be locked now");
1663 }
1664 }
1665 }
1666 }
1667 return relocked_objects;
1668 }
1669 #endif // COMPILER2_OR_JVMCI
1670
1671 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1672 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1673
1674 // Register map for next frame (used for stack crawl). We capture
1675 // the state of the deopt'ing frame's caller. Thus if we need to
1676 // stuff a C2I adapter we can properly fill in the callee-save
1677 // register locations.
1678 frame caller = fr.sender(reg_map);
1679 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1680
1681 frame sender = caller;
1682
1683 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1684 // the vframeArray containing the unpacking information is allocated in the C heap.
1685 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1686 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1687
1688 // Compare the vframeArray to the collected vframes
1689 assert(array->structural_compare(thread, chunk), "just checking");
1690
1691 if (TraceDeoptimization) {
1692 ResourceMark rm;
1693 stringStream st;
1694 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1695 st.print(" ");
1696 fr.print_on(&st);
1697 st.print_cr(" Virtual frames (innermost/newest first):");
1698 for (int index = 0; index < chunk->length(); index++) {
1699 compiledVFrame* vf = chunk->at(index);
1700 int bci = vf->raw_bci();
1701 const char* code_name;
1702 if (bci == SynchronizationEntryBCI) {
1703 code_name = "sync entry";
1704 } else {
1705 Bytecodes::Code code = vf->method()->code_at(bci);
1706 code_name = Bytecodes::name(code);
1707 }
1708
1709 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1710 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1711 st.print(" - %s", code_name);
1712 st.print_cr(" @ bci=%d ", bci);
1713 }
1714 tty->print_raw(st.freeze());
1715 tty->cr();
1716 }
1717
1718 return array;
1719 }
1720
1721 #if COMPILER2_OR_JVMCI
1722 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1723 // Reallocation of some scalar replaced objects failed. Record
1724 // that we need to pop all the interpreter frames for the
1725 // deoptimized compiled frame.
1726 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1727 thread->set_frames_to_pop_failed_realloc(array->frames());
1728 // Unlock all monitors here otherwise the interpreter will see a
1729 // mix of locked and unlocked monitors (because of failed
1730 // reallocations of synchronized objects) and be confused.
1731 for (int i = 0; i < array->frames(); i++) {
1732 MonitorChunk* monitors = array->element(i)->monitors();
1733 if (monitors != nullptr) {
1734 // Unlock in reverse order starting from most nested monitor.
1735 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1736 BasicObjectLock* src = monitors->at(j);
1737 if (src->obj() != nullptr) {
1738 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1739 }
1740 }
1741 array->element(i)->free_monitors();
1742 #ifdef ASSERT
1743 array->element(i)->set_removed_monitors();
1744 #endif
1745 }
1746 }
1747 }
1748 #endif
1749
1750 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1751 assert(fr.can_be_deoptimized(), "checking frame type");
1752
1753 nmethod* nm = fr.cb()->as_nmethod_or_null();
1754 assert(nm != nullptr, "only compiled methods can deopt");
1755 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none);
1756 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc());
1757 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry
1758 : cur_sd->method()->java_code_at(cur_sd->bci()));
1759 gather_statistics(nm, reason, action, bc);
1760
1761 if (LogCompilation && xtty != nullptr) {
1762 ttyLocker ttyl;
1763 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1764 nm->log_identity(xtty);
1765 xtty->end_head();
1766 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1767 xtty->begin_elem("jvms bci='%d'", sd->bci());
1768 xtty->method(sd->method());
1769 xtty->end_elem();
1770 if (sd->is_top()) break;
1771 }
1772 xtty->tail("deoptimized");
1773 }
1774
1775 Continuation::notify_deopt(thread, fr.sp());
1776
1777 // Patch the compiled method so that when execution returns to it we will
1778 // deopt the execution state and return to the interpreter.
1779 fr.deoptimize(thread);
1780 }
1781
1782 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1783 // Deoptimize only if the frame comes from compile code.
1784 // Do not deoptimize the frame which is already patched
1785 // during the execution of the loops below.
1786 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1787 return;
1788 }
1789 ResourceMark rm;
1790 deoptimize_single_frame(thread, fr, reason);
1791 }
1792
1793 #if INCLUDE_JVMCI
1794 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1795 // there is no exception handler for this pc => deoptimize
1796 nm->make_not_entrant();
1797
1798 // Use Deoptimization::deoptimize for all of its side-effects:
1799 // gathering traps statistics, logging...
1800 // it also patches the return pc but we do not care about that
1801 // since we return a continuation to the deopt_blob below.
1802 JavaThread* thread = JavaThread::current();
1803 RegisterMap reg_map(thread,
1804 RegisterMap::UpdateMap::skip,
1805 RegisterMap::ProcessFrames::include,
1806 RegisterMap::WalkContinuation::skip);
1807 frame runtime_frame = thread->last_frame();
1808 frame caller_frame = runtime_frame.sender(®_map);
1809 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1810 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1811 compiledVFrame* cvf = compiledVFrame::cast(vf);
1812 ScopeDesc* imm_scope = cvf->scope();
1813 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1814 if (imm_mdo != nullptr) {
1815 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1816 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1817
1818 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1819 if (pdata != nullptr && pdata->is_BitData()) {
1820 BitData* bit_data = (BitData*) pdata;
1821 bit_data->set_exception_seen();
1822 }
1823 }
1824
1825 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1826
1827 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1828 if (trap_mdo != nullptr) {
1829 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1830 }
1831
1832 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1833 }
1834 #endif
1835
1836 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1837 assert(thread == Thread::current() ||
1838 thread->is_handshake_safe_for(Thread::current()) ||
1839 SafepointSynchronize::is_at_safepoint(),
1840 "can only deoptimize other thread at a safepoint/handshake");
1841 // Compute frame and register map based on thread and sp.
1842 RegisterMap reg_map(thread,
1843 RegisterMap::UpdateMap::skip,
1844 RegisterMap::ProcessFrames::include,
1845 RegisterMap::WalkContinuation::skip);
1846 frame fr = thread->last_frame();
1847 while (fr.id() != id) {
1848 fr = fr.sender(®_map);
1849 }
1850 deoptimize(thread, fr, reason);
1851 }
1852
1853
1854 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1855 Thread* current = Thread::current();
1856 if (thread == current || thread->is_handshake_safe_for(current)) {
1857 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1858 } else {
1859 VM_DeoptimizeFrame deopt(thread, id, reason);
1860 VMThread::execute(&deopt);
1861 }
1862 }
1863
1864 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1865 deoptimize_frame(thread, id, Reason_constraint);
1866 }
1867
1868 // JVMTI PopFrame support
1869 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1870 {
1871 assert(thread == JavaThread::current(), "pre-condition");
1872 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1873 }
1874 JRT_END
1875
1876 MethodData*
1877 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1878 bool create_if_missing) {
1879 JavaThread* THREAD = thread; // For exception macros.
1880 MethodData* mdo = m()->method_data();
1881 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1882 // Build an MDO. Ignore errors like OutOfMemory;
1883 // that simply means we won't have an MDO to update.
1884 Method::build_profiling_method_data(m, THREAD);
1885 if (HAS_PENDING_EXCEPTION) {
1886 // Only metaspace OOM is expected. No Java code executed.
1887 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1888 CLEAR_PENDING_EXCEPTION;
1889 }
1890 mdo = m()->method_data();
1891 }
1892 return mdo;
1893 }
1894
1895 #if COMPILER2_OR_JVMCI
1896 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1897 // In case of an unresolved klass entry, load the class.
1898 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1899 // and probably nowhere else.
1900 // Even that case would benefit from simply re-interpreting the
1901 // bytecode, without paying special attention to the class index.
1902 // So this whole "class index" feature should probably be removed.
1903
1904 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1905 Klass* tk = constant_pool->klass_at(index, THREAD);
1906 if (HAS_PENDING_EXCEPTION) {
1907 // Exception happened during classloading. We ignore the exception here, since it
1908 // is going to be rethrown since the current activation is going to be deoptimized and
1909 // the interpreter will re-execute the bytecode.
1910 // Do not clear probable Async Exceptions.
1911 CLEAR_PENDING_NONASYNC_EXCEPTION;
1912 // Class loading called java code which may have caused a stack
1913 // overflow. If the exception was thrown right before the return
1914 // to the runtime the stack is no longer guarded. Reguard the
1915 // stack otherwise if we return to the uncommon trap blob and the
1916 // stack bang causes a stack overflow we crash.
1917 JavaThread* jt = THREAD;
1918 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1919 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1920 }
1921 return;
1922 }
1923
1924 assert(!constant_pool->tag_at(index).is_symbol(),
1925 "no symbolic names here, please");
1926 }
1927
1928 #if INCLUDE_JFR
1929
1930 class DeoptReasonSerializer : public JfrSerializer {
1931 public:
1932 void serialize(JfrCheckpointWriter& writer) {
1933 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1934 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1935 writer.write_key((u8)i);
1936 writer.write(Deoptimization::trap_reason_name(i));
1937 }
1938 }
1939 };
1940
1941 class DeoptActionSerializer : public JfrSerializer {
1942 public:
1943 void serialize(JfrCheckpointWriter& writer) {
1944 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1945 writer.write_count(nof_actions);
1946 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1947 writer.write_key(i);
1948 writer.write(Deoptimization::trap_action_name((int)i));
1949 }
1950 }
1951 };
1952
1953 static void register_serializers() {
1954 static int critical_section = 0;
1955 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1956 return;
1957 }
1958 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1959 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1960 }
1961
1962 static void post_deoptimization_event(nmethod* nm,
1963 const Method* method,
1964 int trap_bci,
1965 int instruction,
1966 Deoptimization::DeoptReason reason,
1967 Deoptimization::DeoptAction action) {
1968 assert(nm != nullptr, "invariant");
1969 assert(method != nullptr, "invariant");
1970 if (EventDeoptimization::is_enabled()) {
1971 static bool serializers_registered = false;
1972 if (!serializers_registered) {
1973 register_serializers();
1974 serializers_registered = true;
1975 }
1976 EventDeoptimization event;
1977 event.set_compileId(nm->compile_id());
1978 event.set_compiler(nm->compiler_type());
1979 event.set_method(method);
1980 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1981 event.set_bci(trap_bci);
1982 event.set_instruction(instruction);
1983 event.set_reason(reason);
1984 event.set_action(action);
1985 event.commit();
1986 }
1987 }
1988
1989 #endif // INCLUDE_JFR
1990
1991 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
1992 const char* reason_name, const char* reason_action) {
1993 LogTarget(Debug, deoptimization) lt;
1994 if (lt.is_enabled()) {
1995 LogStream ls(lt);
1996 bool is_osr = nm->is_osr_method();
1997 ls.print("cid=%4d %s level=%d",
1998 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
1999 ls.print(" %s", tm->name_and_sig_as_C_string());
2000 ls.print(" trap_bci=%d ", trap_bci);
2001 if (is_osr) {
2002 ls.print("osr_bci=%d ", nm->osr_entry_bci());
2003 }
2004 ls.print("%s ", reason_name);
2005 ls.print("%s ", reason_action);
2006 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2007 pc, fr.pc() - nm->code_begin());
2008 }
2009 }
2010
2011 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2012 HandleMark hm(current);
2013
2014 // uncommon_trap() is called at the beginning of the uncommon trap
2015 // handler. Note this fact before we start generating temporary frames
2016 // that can confuse an asynchronous stack walker. This counter is
2017 // decremented at the end of unpack_frames().
2018
2019 current->inc_in_deopt_handler();
2020
2021 #if INCLUDE_JVMCI
2022 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2023 RegisterMap reg_map(current,
2024 RegisterMap::UpdateMap::include,
2025 RegisterMap::ProcessFrames::include,
2026 RegisterMap::WalkContinuation::skip);
2027 #else
2028 RegisterMap reg_map(current,
2029 RegisterMap::UpdateMap::skip,
2030 RegisterMap::ProcessFrames::include,
2031 RegisterMap::WalkContinuation::skip);
2032 #endif
2033 frame stub_frame = current->last_frame();
2034 frame fr = stub_frame.sender(®_map);
2035
2036 // Log a message
2037 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2038 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2039
2040 {
2041 ResourceMark rm;
2042
2043 DeoptReason reason = trap_request_reason(trap_request);
2044 DeoptAction action = trap_request_action(trap_request);
2045 #if INCLUDE_JVMCI
2046 int debug_id = trap_request_debug_id(trap_request);
2047 #endif
2048 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2049
2050 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2051 compiledVFrame* cvf = compiledVFrame::cast(vf);
2052
2053 nmethod* nm = cvf->code();
2054
2055 ScopeDesc* trap_scope = cvf->scope();
2056
2057 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2058
2059 if (is_receiver_constraint_failure) {
2060 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2061 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2062 JVMCI_ONLY(COMMA debug_id));
2063 }
2064
2065 methodHandle trap_method(current, trap_scope->method());
2066 int trap_bci = trap_scope->bci();
2067 #if INCLUDE_JVMCI
2068 jlong speculation = current->pending_failed_speculation();
2069 if (nm->is_compiled_by_jvmci()) {
2070 nm->update_speculation(current);
2071 } else {
2072 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2073 }
2074
2075 if (trap_bci == SynchronizationEntryBCI) {
2076 trap_bci = 0;
2077 current->set_pending_monitorenter(true);
2078 }
2079
2080 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2081 current->set_pending_transfer_to_interpreter(true);
2082 }
2083 #endif
2084
2085 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2086 // Record this event in the histogram.
2087 gather_statistics(nm, reason, action, trap_bc);
2088
2089 // Ensure that we can record deopt. history:
2090 // Need MDO to record RTM code generation state.
2091 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking );
2092
2093 methodHandle profiled_method;
2094 #if INCLUDE_JVMCI
2095 if (nm->is_compiled_by_jvmci()) {
2096 profiled_method = methodHandle(current, nm->method());
2097 } else {
2098 profiled_method = trap_method;
2099 }
2100 #else
2101 profiled_method = trap_method;
2102 #endif
2103
2104 MethodData* trap_mdo =
2105 get_method_data(current, profiled_method, create_if_missing);
2106
2107 { // Log Deoptimization event for JFR, UL and event system
2108 Method* tm = trap_method();
2109 const char* reason_name = trap_reason_name(reason);
2110 const char* reason_action = trap_action_name(action);
2111 intptr_t pc = p2i(fr.pc());
2112
2113 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2114 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2115 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2116 reason_name, reason_action, pc,
2117 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2118 }
2119
2120 // Print a bunch of diagnostics, if requested.
2121 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2122 ResourceMark rm;
2123
2124 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2125 // We must do this already now, since we cannot acquire this lock while
2126 // holding the tty lock (lock ordering by rank).
2127 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2128
2129 ttyLocker ttyl;
2130
2131 char buf[100];
2132 if (xtty != nullptr) {
2133 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2134 os::current_thread_id(),
2135 format_trap_request(buf, sizeof(buf), trap_request));
2136 #if INCLUDE_JVMCI
2137 if (speculation != 0) {
2138 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2139 }
2140 #endif
2141 nm->log_identity(xtty);
2142 }
2143 Symbol* class_name = nullptr;
2144 bool unresolved = false;
2145 if (unloaded_class_index >= 0) {
2146 constantPoolHandle constants (current, trap_method->constants());
2147 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2148 class_name = constants->klass_name_at(unloaded_class_index);
2149 unresolved = true;
2150 if (xtty != nullptr)
2151 xtty->print(" unresolved='1'");
2152 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2153 class_name = constants->symbol_at(unloaded_class_index);
2154 }
2155 if (xtty != nullptr)
2156 xtty->name(class_name);
2157 }
2158 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2159 // Dump the relevant MDO state.
2160 // This is the deopt count for the current reason, any previous
2161 // reasons or recompiles seen at this point.
2162 int dcnt = trap_mdo->trap_count(reason);
2163 if (dcnt != 0)
2164 xtty->print(" count='%d'", dcnt);
2165
2166 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2167 // lock extra_data_lock before the tty lock.
2168 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2169 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2170 if (dos != 0) {
2171 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2172 if (trap_state_is_recompiled(dos)) {
2173 int recnt2 = trap_mdo->overflow_recompile_count();
2174 if (recnt2 != 0)
2175 xtty->print(" recompiles2='%d'", recnt2);
2176 }
2177 }
2178 }
2179 if (xtty != nullptr) {
2180 xtty->stamp();
2181 xtty->end_head();
2182 }
2183 if (TraceDeoptimization) { // make noise on the tty
2184 stringStream st;
2185 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2186 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2187 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2188 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2189 #if INCLUDE_JVMCI
2190 if (nm->is_compiled_by_jvmci()) {
2191 const char* installed_code_name = nm->jvmci_name();
2192 if (installed_code_name != nullptr) {
2193 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2194 }
2195 }
2196 #endif
2197 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2198 p2i(fr.pc()),
2199 os::current_thread_id(),
2200 trap_reason_name(reason),
2201 trap_action_name(action),
2202 unloaded_class_index
2203 #if INCLUDE_JVMCI
2204 , debug_id
2205 #endif
2206 );
2207 if (class_name != nullptr) {
2208 st.print(unresolved ? " unresolved class: " : " symbol: ");
2209 class_name->print_symbol_on(&st);
2210 }
2211 st.cr();
2212 tty->print_raw(st.freeze());
2213 }
2214 if (xtty != nullptr) {
2215 // Log the precise location of the trap.
2216 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2217 xtty->begin_elem("jvms bci='%d'", sd->bci());
2218 xtty->method(sd->method());
2219 xtty->end_elem();
2220 if (sd->is_top()) break;
2221 }
2222 xtty->tail("uncommon_trap");
2223 }
2224 }
2225 // (End diagnostic printout.)
2226
2227 if (is_receiver_constraint_failure) {
2228 fatal("missing receiver type check");
2229 }
2230
2231 // Load class if necessary
2232 if (unloaded_class_index >= 0) {
2233 constantPoolHandle constants(current, trap_method->constants());
2234 load_class_by_index(constants, unloaded_class_index, THREAD);
2235 }
2236
2237 // Flush the nmethod if necessary and desirable.
2238 //
2239 // We need to avoid situations where we are re-flushing the nmethod
2240 // because of a hot deoptimization site. Repeated flushes at the same
2241 // point need to be detected by the compiler and avoided. If the compiler
2242 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2243 // module must take measures to avoid an infinite cycle of recompilation
2244 // and deoptimization. There are several such measures:
2245 //
2246 // 1. If a recompilation is ordered a second time at some site X
2247 // and for the same reason R, the action is adjusted to 'reinterpret',
2248 // to give the interpreter time to exercise the method more thoroughly.
2249 // If this happens, the method's overflow_recompile_count is incremented.
2250 //
2251 // 2. If the compiler fails to reduce the deoptimization rate, then
2252 // the method's overflow_recompile_count will begin to exceed the set
2253 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2254 // is adjusted to 'make_not_compilable', and the method is abandoned
2255 // to the interpreter. This is a performance hit for hot methods,
2256 // but is better than a disastrous infinite cycle of recompilations.
2257 // (Actually, only the method containing the site X is abandoned.)
2258 //
2259 // 3. In parallel with the previous measures, if the total number of
2260 // recompilations of a method exceeds the much larger set limit
2261 // PerMethodRecompilationCutoff, the method is abandoned.
2262 // This should only happen if the method is very large and has
2263 // many "lukewarm" deoptimizations. The code which enforces this
2264 // limit is elsewhere (class nmethod, class Method).
2265 //
2266 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2267 // to recompile at each bytecode independently of the per-BCI cutoff.
2268 //
2269 // The decision to update code is up to the compiler, and is encoded
2270 // in the Action_xxx code. If the compiler requests Action_none
2271 // no trap state is changed, no compiled code is changed, and the
2272 // computation suffers along in the interpreter.
2273 //
2274 // The other action codes specify various tactics for decompilation
2275 // and recompilation. Action_maybe_recompile is the loosest, and
2276 // allows the compiled code to stay around until enough traps are seen,
2277 // and until the compiler gets around to recompiling the trapping method.
2278 //
2279 // The other actions cause immediate removal of the present code.
2280
2281 // Traps caused by injected profile shouldn't pollute trap counts.
2282 bool injected_profile_trap = trap_method->has_injected_profile() &&
2283 (reason == Reason_intrinsic || reason == Reason_unreached);
2284
2285 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2286 bool make_not_entrant = false;
2287 bool make_not_compilable = false;
2288 bool reprofile = false;
2289 switch (action) {
2290 case Action_none:
2291 // Keep the old code.
2292 update_trap_state = false;
2293 break;
2294 case Action_maybe_recompile:
2295 // Do not need to invalidate the present code, but we can
2296 // initiate another
2297 // Start compiler without (necessarily) invalidating the nmethod.
2298 // The system will tolerate the old code, but new code should be
2299 // generated when possible.
2300 break;
2301 case Action_reinterpret:
2302 // Go back into the interpreter for a while, and then consider
2303 // recompiling form scratch.
2304 make_not_entrant = true;
2305 // Reset invocation counter for outer most method.
2306 // This will allow the interpreter to exercise the bytecodes
2307 // for a while before recompiling.
2308 // By contrast, Action_make_not_entrant is immediate.
2309 //
2310 // Note that the compiler will track null_check, null_assert,
2311 // range_check, and class_check events and log them as if they
2312 // had been traps taken from compiled code. This will update
2313 // the MDO trap history so that the next compilation will
2314 // properly detect hot trap sites.
2315 reprofile = true;
2316 break;
2317 case Action_make_not_entrant:
2318 // Request immediate recompilation, and get rid of the old code.
2319 // Make them not entrant, so next time they are called they get
2320 // recompiled. Unloaded classes are loaded now so recompile before next
2321 // time they are called. Same for uninitialized. The interpreter will
2322 // link the missing class, if any.
2323 make_not_entrant = true;
2324 break;
2325 case Action_make_not_compilable:
2326 // Give up on compiling this method at all.
2327 make_not_entrant = true;
2328 make_not_compilable = true;
2329 break;
2330 default:
2331 ShouldNotReachHere();
2332 }
2333
2334 // Setting +ProfileTraps fixes the following, on all platforms:
2335 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2336 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2337 // recompile relies on a MethodData* to record heroic opt failures.
2338
2339 // Whether the interpreter is producing MDO data or not, we also need
2340 // to use the MDO to detect hot deoptimization points and control
2341 // aggressive optimization.
2342 bool inc_recompile_count = false;
2343
2344 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2345 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2346 (trap_mdo != nullptr),
2347 Mutex::_no_safepoint_check_flag);
2348 ProfileData* pdata = nullptr;
2349 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2350 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2351 uint this_trap_count = 0;
2352 bool maybe_prior_trap = false;
2353 bool maybe_prior_recompile = false;
2354
2355 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2356 #if INCLUDE_JVMCI
2357 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2358 #endif
2359 nm->method(),
2360 //outputs:
2361 this_trap_count,
2362 maybe_prior_trap,
2363 maybe_prior_recompile);
2364 // Because the interpreter also counts null, div0, range, and class
2365 // checks, these traps from compiled code are double-counted.
2366 // This is harmless; it just means that the PerXTrapLimit values
2367 // are in effect a little smaller than they look.
2368
2369 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2370 if (per_bc_reason != Reason_none) {
2371 // Now take action based on the partially known per-BCI history.
2372 if (maybe_prior_trap
2373 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2374 // If there are too many traps at this BCI, force a recompile.
2375 // This will allow the compiler to see the limit overflow, and
2376 // take corrective action, if possible. The compiler generally
2377 // does not use the exact PerBytecodeTrapLimit value, but instead
2378 // changes its tactics if it sees any traps at all. This provides
2379 // a little hysteresis, delaying a recompile until a trap happens
2380 // several times.
2381 //
2382 // Actually, since there is only one bit of counter per BCI,
2383 // the possible per-BCI counts are {0,1,(per-method count)}.
2384 // This produces accurate results if in fact there is only
2385 // one hot trap site, but begins to get fuzzy if there are
2386 // many sites. For example, if there are ten sites each
2387 // trapping two or more times, they each get the blame for
2388 // all of their traps.
2389 make_not_entrant = true;
2390 }
2391
2392 // Detect repeated recompilation at the same BCI, and enforce a limit.
2393 if (make_not_entrant && maybe_prior_recompile) {
2394 // More than one recompile at this point.
2395 inc_recompile_count = maybe_prior_trap;
2396 }
2397 } else {
2398 // For reasons which are not recorded per-bytecode, we simply
2399 // force recompiles unconditionally.
2400 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2401 make_not_entrant = true;
2402 }
2403
2404 // Go back to the compiler if there are too many traps in this method.
2405 if (this_trap_count >= per_method_trap_limit(reason)) {
2406 // If there are too many traps in this method, force a recompile.
2407 // This will allow the compiler to see the limit overflow, and
2408 // take corrective action, if possible.
2409 // (This condition is an unlikely backstop only, because the
2410 // PerBytecodeTrapLimit is more likely to take effect first,
2411 // if it is applicable.)
2412 make_not_entrant = true;
2413 }
2414
2415 // Here's more hysteresis: If there has been a recompile at
2416 // this trap point already, run the method in the interpreter
2417 // for a while to exercise it more thoroughly.
2418 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2419 reprofile = true;
2420 }
2421 }
2422
2423 // Take requested actions on the method:
2424
2425 // Recompile
2426 if (make_not_entrant) {
2427 if (!nm->make_not_entrant()) {
2428 return; // the call did not change nmethod's state
2429 }
2430
2431 if (pdata != nullptr) {
2432 // Record the recompilation event, if any.
2433 int tstate0 = pdata->trap_state();
2434 int tstate1 = trap_state_set_recompiled(tstate0, true);
2435 if (tstate1 != tstate0)
2436 pdata->set_trap_state(tstate1);
2437 }
2438
2439 #if INCLUDE_RTM_OPT
2440 // Restart collecting RTM locking abort statistic if the method
2441 // is recompiled for a reason other than RTM state change.
2442 // Assume that in new recompiled code the statistic could be different,
2443 // for example, due to different inlining.
2444 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) &&
2445 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) {
2446 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2447 }
2448 #endif
2449 // For code aging we count traps separately here, using make_not_entrant()
2450 // as a guard against simultaneous deopts in multiple threads.
2451 if (reason == Reason_tenured && trap_mdo != nullptr) {
2452 trap_mdo->inc_tenure_traps();
2453 }
2454 }
2455
2456 if (inc_recompile_count) {
2457 trap_mdo->inc_overflow_recompile_count();
2458 if ((uint)trap_mdo->overflow_recompile_count() >
2459 (uint)PerBytecodeRecompilationCutoff) {
2460 // Give up on the method containing the bad BCI.
2461 if (trap_method() == nm->method()) {
2462 make_not_compilable = true;
2463 } else {
2464 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2465 // But give grace to the enclosing nm->method().
2466 }
2467 }
2468 }
2469
2470 // Reprofile
2471 if (reprofile) {
2472 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2473 }
2474
2475 // Give up compiling
2476 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2477 assert(make_not_entrant, "consistent");
2478 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2479 }
2480
2481 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2482 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2483 if (exception_handler_data != nullptr) {
2484 // uncommon trap at the start of an exception handler.
2485 // C2 generates these for un-entered exception handlers.
2486 // mark the handler as entered to avoid generating
2487 // another uncommon trap the next time the handler is compiled
2488 exception_handler_data->set_exception_handler_entered();
2489 }
2490 }
2491
2492 } // Free marked resources
2493
2494 }
2495 JRT_END
2496
2497 ProfileData*
2498 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2499 int trap_bci,
2500 Deoptimization::DeoptReason reason,
2501 bool update_total_trap_count,
2502 #if INCLUDE_JVMCI
2503 bool is_osr,
2504 #endif
2505 Method* compiled_method,
2506 //outputs:
2507 uint& ret_this_trap_count,
2508 bool& ret_maybe_prior_trap,
2509 bool& ret_maybe_prior_recompile) {
2510 trap_mdo->check_extra_data_locked();
2511
2512 bool maybe_prior_trap = false;
2513 bool maybe_prior_recompile = false;
2514 uint this_trap_count = 0;
2515 if (update_total_trap_count) {
2516 uint idx = reason;
2517 #if INCLUDE_JVMCI
2518 if (is_osr) {
2519 // Upper half of history array used for traps in OSR compilations
2520 idx += Reason_TRAP_HISTORY_LENGTH;
2521 }
2522 #endif
2523 uint prior_trap_count = trap_mdo->trap_count(idx);
2524 this_trap_count = trap_mdo->inc_trap_count(idx);
2525
2526 // If the runtime cannot find a place to store trap history,
2527 // it is estimated based on the general condition of the method.
2528 // If the method has ever been recompiled, or has ever incurred
2529 // a trap with the present reason , then this BCI is assumed
2530 // (pessimistically) to be the culprit.
2531 maybe_prior_trap = (prior_trap_count != 0);
2532 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2533 }
2534 ProfileData* pdata = nullptr;
2535
2536
2537 // For reasons which are recorded per bytecode, we check per-BCI data.
2538 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2539 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2540 if (per_bc_reason != Reason_none) {
2541 // Find the profile data for this BCI. If there isn't one,
2542 // try to allocate one from the MDO's set of spares.
2543 // This will let us detect a repeated trap at this point.
2544 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2545
2546 if (pdata != nullptr) {
2547 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2548 if (LogCompilation && xtty != nullptr) {
2549 ttyLocker ttyl;
2550 // no more room for speculative traps in this MDO
2551 xtty->elem("speculative_traps_oom");
2552 }
2553 }
2554 // Query the trap state of this profile datum.
2555 int tstate0 = pdata->trap_state();
2556 if (!trap_state_has_reason(tstate0, per_bc_reason))
2557 maybe_prior_trap = false;
2558 if (!trap_state_is_recompiled(tstate0))
2559 maybe_prior_recompile = false;
2560
2561 // Update the trap state of this profile datum.
2562 int tstate1 = tstate0;
2563 // Record the reason.
2564 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2565 // Store the updated state on the MDO, for next time.
2566 if (tstate1 != tstate0)
2567 pdata->set_trap_state(tstate1);
2568 } else {
2569 if (LogCompilation && xtty != nullptr) {
2570 ttyLocker ttyl;
2571 // Missing MDP? Leave a small complaint in the log.
2572 xtty->elem("missing_mdp bci='%d'", trap_bci);
2573 }
2574 }
2575 }
2576
2577 // Return results:
2578 ret_this_trap_count = this_trap_count;
2579 ret_maybe_prior_trap = maybe_prior_trap;
2580 ret_maybe_prior_recompile = maybe_prior_recompile;
2581 return pdata;
2582 }
2583
2584 void
2585 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2586 ResourceMark rm;
2587 // Ignored outputs:
2588 uint ignore_this_trap_count;
2589 bool ignore_maybe_prior_trap;
2590 bool ignore_maybe_prior_recompile;
2591 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2592 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2593 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2594
2595 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2596 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2597
2598 query_update_method_data(trap_mdo, trap_bci,
2599 (DeoptReason)reason,
2600 update_total_counts,
2601 #if INCLUDE_JVMCI
2602 false,
2603 #endif
2604 nullptr,
2605 ignore_this_trap_count,
2606 ignore_maybe_prior_trap,
2607 ignore_maybe_prior_recompile);
2608 }
2609
2610 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode))
2611 // Enable WXWrite: current function is called from methods compiled by C2 directly
2612 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2613
2614 // Still in Java no safepoints
2615 {
2616 // This enters VM and may safepoint
2617 uncommon_trap_inner(current, trap_request);
2618 }
2619 HandleMark hm(current);
2620 return fetch_unroll_info_helper(current, exec_mode);
2621 PROF_END
2622
2623 // Local derived constants.
2624 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2625 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2626 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2627
2628 //---------------------------trap_state_reason---------------------------------
2629 Deoptimization::DeoptReason
2630 Deoptimization::trap_state_reason(int trap_state) {
2631 // This assert provides the link between the width of DataLayout::trap_bits
2632 // and the encoding of "recorded" reasons. It ensures there are enough
2633 // bits to store all needed reasons in the per-BCI MDO profile.
2634 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2635 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2636 trap_state -= recompile_bit;
2637 if (trap_state == DS_REASON_MASK) {
2638 return Reason_many;
2639 } else {
2640 assert((int)Reason_none == 0, "state=0 => Reason_none");
2641 return (DeoptReason)trap_state;
2642 }
2643 }
2644 //-------------------------trap_state_has_reason-------------------------------
2645 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2646 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2647 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2648 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2649 trap_state -= recompile_bit;
2650 if (trap_state == DS_REASON_MASK) {
2651 return -1; // true, unspecifically (bottom of state lattice)
2652 } else if (trap_state == reason) {
2653 return 1; // true, definitely
2654 } else if (trap_state == 0) {
2655 return 0; // false, definitely (top of state lattice)
2656 } else {
2657 return 0; // false, definitely
2658 }
2659 }
2660 //-------------------------trap_state_add_reason-------------------------------
2661 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2662 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2663 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2664 trap_state -= recompile_bit;
2665 if (trap_state == DS_REASON_MASK) {
2666 return trap_state + recompile_bit; // already at state lattice bottom
2667 } else if (trap_state == reason) {
2668 return trap_state + recompile_bit; // the condition is already true
2669 } else if (trap_state == 0) {
2670 return reason + recompile_bit; // no condition has yet been true
2671 } else {
2672 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2673 }
2674 }
2675 //-----------------------trap_state_is_recompiled------------------------------
2676 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2677 return (trap_state & DS_RECOMPILE_BIT) != 0;
2678 }
2679 //-----------------------trap_state_set_recompiled-----------------------------
2680 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2681 if (z) return trap_state | DS_RECOMPILE_BIT;
2682 else return trap_state & ~DS_RECOMPILE_BIT;
2683 }
2684 //---------------------------format_trap_state---------------------------------
2685 // This is used for debugging and diagnostics, including LogFile output.
2686 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2687 int trap_state) {
2688 assert(buflen > 0, "sanity");
2689 DeoptReason reason = trap_state_reason(trap_state);
2690 bool recomp_flag = trap_state_is_recompiled(trap_state);
2691 // Re-encode the state from its decoded components.
2692 int decoded_state = 0;
2693 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2694 decoded_state = trap_state_add_reason(decoded_state, reason);
2695 if (recomp_flag)
2696 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2697 // If the state re-encodes properly, format it symbolically.
2698 // Because this routine is used for debugging and diagnostics,
2699 // be robust even if the state is a strange value.
2700 size_t len;
2701 if (decoded_state != trap_state) {
2702 // Random buggy state that doesn't decode??
2703 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2704 } else {
2705 len = jio_snprintf(buf, buflen, "%s%s",
2706 trap_reason_name(reason),
2707 recomp_flag ? " recompiled" : "");
2708 }
2709 return buf;
2710 }
2711
2712
2713 //--------------------------------statics--------------------------------------
2714 const char* Deoptimization::_trap_reason_name[] = {
2715 // Note: Keep this in sync. with enum DeoptReason.
2716 "none",
2717 "null_check",
2718 "null_assert" JVMCI_ONLY("_or_unreached0"),
2719 "range_check",
2720 "class_check",
2721 "array_check",
2722 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2723 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2724 "profile_predicate",
2725 "unloaded",
2726 "uninitialized",
2727 "initialized",
2728 "unreached",
2729 "unhandled",
2730 "constraint",
2731 "div0_check",
2732 "age",
2733 "predicate",
2734 "loop_limit_check",
2735 "speculate_class_check",
2736 "speculate_null_check",
2737 "speculate_null_assert",
2738 "rtm_state_change",
2739 "unstable_if",
2740 "unstable_fused_if",
2741 "receiver_constraint",
2742 #if INCLUDE_JVMCI
2743 "aliasing",
2744 "transfer_to_interpreter",
2745 "not_compiled_exception_handler",
2746 "unresolved",
2747 "jsr_mismatch",
2748 #endif
2749 "tenured"
2750 };
2751 const char* Deoptimization::_trap_action_name[] = {
2752 // Note: Keep this in sync. with enum DeoptAction.
2753 "none",
2754 "maybe_recompile",
2755 "reinterpret",
2756 "make_not_entrant",
2757 "make_not_compilable"
2758 };
2759
2760 const char* Deoptimization::trap_reason_name(int reason) {
2761 // Check that every reason has a name
2762 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2763
2764 if (reason == Reason_many) return "many";
2765 if ((uint)reason < Reason_LIMIT)
2766 return _trap_reason_name[reason];
2767 static char buf[20];
2768 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2769 return buf;
2770 }
2771 const char* Deoptimization::trap_action_name(int action) {
2772 // Check that every action has a name
2773 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2774
2775 if ((uint)action < Action_LIMIT)
2776 return _trap_action_name[action];
2777 static char buf[20];
2778 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2779 return buf;
2780 }
2781
2782 // This is used for debugging and diagnostics, including LogFile output.
2783 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2784 int trap_request) {
2785 jint unloaded_class_index = trap_request_index(trap_request);
2786 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2787 const char* action = trap_action_name(trap_request_action(trap_request));
2788 #if INCLUDE_JVMCI
2789 int debug_id = trap_request_debug_id(trap_request);
2790 #endif
2791 size_t len;
2792 if (unloaded_class_index < 0) {
2793 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2794 reason, action
2795 #if INCLUDE_JVMCI
2796 ,debug_id
2797 #endif
2798 );
2799 } else {
2800 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2801 reason, action, unloaded_class_index
2802 #if INCLUDE_JVMCI
2803 ,debug_id
2804 #endif
2805 );
2806 }
2807 return buf;
2808 }
2809
2810 juint Deoptimization::_deoptimization_hist
2811 [1 + 4 + 5] // total + online + archived
2812 [Deoptimization::Reason_LIMIT]
2813 [1 + Deoptimization::Action_LIMIT]
2814 [Deoptimization::BC_CASE_LIMIT]
2815 = {0};
2816
2817 enum {
2818 LSB_BITS = 8,
2819 LSB_MASK = right_n_bits(LSB_BITS)
2820 };
2821
2822 static void update(juint* cases, Bytecodes::Code bc) {
2823 juint* bc_counter_addr = nullptr;
2824 juint bc_counter = 0;
2825 // Look for an unused counter, or an exact match to this BC.
2826 if (bc != Bytecodes::_illegal) {
2827 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) {
2828 juint* counter_addr = &cases[bc_case];
2829 juint counter = *counter_addr;
2830 if ((counter == 0 && bc_counter_addr == nullptr)
2831 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2832 // this counter is either free or is already devoted to this BC
2833 bc_counter_addr = counter_addr;
2834 bc_counter = counter | bc;
2835 }
2836 }
2837 }
2838 if (bc_counter_addr == nullptr) {
2839 // Overflow, or no given bytecode.
2840 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1];
2841 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2842 }
2843 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2844 }
2845
2846
2847 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
2848 Bytecodes::Code bc) {
2849 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2850 assert(action >= 0 && action < Action_LIMIT, "oob");
2851 _deoptimization_hist[0][Reason_none][0][0] += 1; // total
2852 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total
2853
2854 update(_deoptimization_hist[0][reason][1+action], bc);
2855
2856 uint lvl = nm->comp_level() + (nm->is_scc() ? 4 : 0) + (nm->preloaded() ? 1 : 0);
2857 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total
2858 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total
2859 update(_deoptimization_hist[lvl][reason][1+action], bc);
2860 }
2861
2862 jint Deoptimization::total_deoptimization_count() {
2863 return _deoptimization_hist[0][Reason_none][0][0];
2864 }
2865
2866 // Get the deopt count for a specific reason and a specific action. If either
2867 // one of 'reason' or 'action' is null, the method returns the sum of all
2868 // deoptimizations with the specific 'action' or 'reason' respectively.
2869 // If both arguments are null, the method returns the total deopt count.
2870 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2871 if (reason_str == nullptr && action_str == nullptr) {
2872 return total_deoptimization_count();
2873 }
2874 juint counter = 0;
2875 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2876 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2877 for (int action = 0; action < Action_LIMIT; action++) {
2878 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2879 juint* cases = _deoptimization_hist[0][reason][1+action];
2880 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2881 counter += cases[bc_case] >> LSB_BITS;
2882 }
2883 }
2884 }
2885 }
2886 }
2887 return counter;
2888 }
2889
2890 void Deoptimization::print_statistics() {
2891 ttyLocker ttyl;
2892 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2893 tty->print_cr("Deoptimization traps recorded:");
2894 print_statistics_on(tty);
2895 if (xtty != nullptr) xtty->tail("statistics");
2896 }
2897
2898 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) {
2899 juint total = _deoptimization_hist[lvl][Reason_none][0][0];
2900 juint account = total;
2901 #define PRINT_STAT_LINE(name, r) \
2902 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name);
2903 if (total > 0) {
2904 st->print(" %s: ", title);
2905 PRINT_STAT_LINE("total", total);
2906 // For each non-zero entry in the histogram, print the reason,
2907 // the action, and (if specifically known) the type of bytecode.
2908 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2909 for (int action = 0; action < Action_LIMIT; action++) {
2910 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action];
2911 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2912 juint counter = cases[bc_case];
2913 if (counter != 0) {
2914 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2915 const char* bc_name = "other";
2916 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) {
2917 // overwritten
2918 } else if (Bytecodes::is_defined(bc)) {
2919 bc_name = Bytecodes::name(bc);
2920 }
2921 juint r = counter >> LSB_BITS;
2922 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)",
2923 trap_reason_name(reason), trap_action_name(action), bc_name,
2924 r, (r * 100.0) / total);
2925 account -= r;
2926 }
2927 }
2928 }
2929 }
2930 if (account != 0) {
2931 PRINT_STAT_LINE("unaccounted", account);
2932 }
2933 #undef PRINT_STAT_LINE
2934 }
2935 }
2936
2937 void Deoptimization::print_statistics_on(outputStream* st) {
2938 // print_statistics_on("Total", 0, st);
2939 print_statistics_on("Tier1", 1, st);
2940 print_statistics_on("Tier2", 2, st);
2941 print_statistics_on("Tier3", 3, st);
2942 print_statistics_on("Tier4", 4, st);
2943
2944 print_statistics_on("SC Tier1", 5, st);
2945 print_statistics_on("SC Tier2", 6, st);
2946 print_statistics_on("SC Tier4", 8, st);
2947 print_statistics_on("SC Tier5 (preloaded)", 9, st);
2948 }
2949
2950 #else // COMPILER2_OR_JVMCI
2951
2952
2953 // Stubs for C1 only system.
2954 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2955 return false;
2956 }
2957
2958 const char* Deoptimization::trap_reason_name(int reason) {
2959 return "unknown";
2960 }
2961
2962 jint Deoptimization::total_deoptimization_count() {
2963 return 0;
2964 }
2965
2966 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2967 return 0;
2968 }
2969
2970 void Deoptimization::print_statistics() {
2971 // no output
2972 }
2973
2974 void
2975 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2976 // no update
2977 }
2978
2979 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2980 return 0;
2981 }
2982
2983 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2984 Bytecodes::Code bc) {
2985 // no update
2986 }
2987
2988 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2989 int trap_state) {
2990 jio_snprintf(buf, buflen, "#%d", trap_state);
2991 return buf;
2992 }
2993
2994 #endif // COMPILER2_OR_JVMCI
2995
2996 #define DO_COUNTERS(macro) \
2997 macro(Deoptimization, fetch_unroll_info) \
2998 macro(Deoptimization, unpack_frames) \
2999 macro(Deoptimization, uncommon_trap_inner) \
3000 macro(Deoptimization, uncommon_trap)
3001
3002 #define INIT_COUNTER(sub, name) \
3003 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \
3004 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count");
3005
3006 void Deoptimization::init_counters() {
3007 if (ProfileRuntimeCalls && UsePerfData) {
3008 EXCEPTION_MARK;
3009
3010 DO_COUNTERS(INIT_COUNTER)
3011
3012 if (HAS_PENDING_EXCEPTION) {
3013 vm_exit_during_initialization("jvm_perf_init failed unexpectedly");
3014 }
3015 }
3016 }
3017 #undef INIT_COUNTER
3018
3019 #define PRINT_COUNTER(sub, name) { \
3020 jlong count = _perf_##sub##_##name##_count->get_value(); \
3021 if (count > 0) { \
3022 st->print_cr(" %-50s = %4ldms (elapsed) %4ldms (thread) (%5ld events)", #sub "::" #name, \
3023 _perf_##sub##_##name##_timer->elapsed_counter_value_ms(), \
3024 _perf_##sub##_##name##_timer->thread_counter_value_ms(), \
3025 count); \
3026 }}
3027
3028 void Deoptimization::print_counters_on(outputStream* st) {
3029 if (ProfileRuntimeCalls && UsePerfData) {
3030 DO_COUNTERS(PRINT_COUNTER)
3031 } else {
3032 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData"));
3033 }
3034 }
3035
3036 #undef PRINT_COUNTER
3037 #undef DO_COUNTERS