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 (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1245 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1246 } else {
1247 obj = ik->allocate_instance(THREAD);
1248 }
1249 }
1250 } else if (k->is_typeArray_klass()) {
1251 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1252 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1253 int len = sv->field_size() / type2size[ak->element_type()];
1254 InternalOOMEMark iom(THREAD);
1255 obj = ak->allocate(len, THREAD);
1256 } else if (k->is_objArray_klass()) {
1257 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1258 InternalOOMEMark iom(THREAD);
1259 obj = ak->allocate(sv->field_size(), THREAD);
1260 }
1261
1262 if (obj == nullptr) {
1263 failures = true;
1264 }
1265
1266 assert(sv->value().is_null(), "redundant reallocation");
1267 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1268 CLEAR_PENDING_EXCEPTION;
1269 sv->set_value(obj);
1270 }
1271
1272 if (failures) {
1273 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1274 } else if (pending_exception.not_null()) {
1275 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1276 }
1277
1278 return failures;
1279 }
1280
1281 #if INCLUDE_JVMCI
1282 /**
1283 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1284 * we need to somehow be able to recover the actual kind to be able to write the correct
1285 * amount of bytes.
1286 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1287 * the entries at index n + 1 to n + i are 'markers'.
1288 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1289 * expected form of the array would be:
1290 *
1291 * {b0, b1, b2, b3, INT, marker, b6, b7}
1292 *
1293 * Thus, in order to get back the size of the entry, we simply need to count the number
1294 * of marked entries
1295 *
1296 * @param virtualArray the virtualized byte array
1297 * @param i index of the virtual entry we are recovering
1298 * @return The number of bytes the entry spans
1299 */
1300 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1301 int index = i;
1302 while (++index < virtualArray->field_size() &&
1303 virtualArray->field_at(index)->is_marker()) {}
1304 return index - i;
1305 }
1306
1307 /**
1308 * If there was a guarantee for byte array to always start aligned to a long, we could
1309 * do a simple check on the parity of the index. Unfortunately, that is not always the
1310 * case. Thus, we check alignment of the actual address we are writing to.
1311 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1312 * write a long to index 3.
1313 */
1314 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1315 jbyte* res = obj->byte_at_addr(index);
1316 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1317 return res;
1318 }
1319
1320 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1321 switch (byte_count) {
1322 case 1:
1323 obj->byte_at_put(index, (jbyte) value->get_jint());
1324 break;
1325 case 2:
1326 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1327 break;
1328 case 4:
1329 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1330 break;
1331 case 8:
1332 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1333 break;
1334 default:
1335 ShouldNotReachHere();
1336 }
1337 }
1338 #endif // INCLUDE_JVMCI
1339
1340
1341 // restore elements of an eliminated type array
1342 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1343 int index = 0;
1344
1345 for (int i = 0; i < sv->field_size(); i++) {
1346 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1347 switch(type) {
1348 case T_LONG: case T_DOUBLE: {
1349 assert(value->type() == T_INT, "Agreement.");
1350 StackValue* low =
1351 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1352 #ifdef _LP64
1353 jlong res = (jlong)low->get_intptr();
1354 #else
1355 jlong res = jlong_from(value->get_jint(), low->get_jint());
1356 #endif
1357 obj->long_at_put(index, res);
1358 break;
1359 }
1360
1361 case T_INT: case T_FLOAT: { // 4 bytes.
1362 assert(value->type() == T_INT, "Agreement.");
1363 bool big_value = false;
1364 if (i + 1 < sv->field_size() && type == T_INT) {
1365 if (sv->field_at(i)->is_location()) {
1366 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1367 if (type == Location::dbl || type == Location::lng) {
1368 big_value = true;
1369 }
1370 } else if (sv->field_at(i)->is_constant_int()) {
1371 ScopeValue* next_scope_field = sv->field_at(i + 1);
1372 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1373 big_value = true;
1374 }
1375 }
1376 }
1377
1378 if (big_value) {
1379 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1380 #ifdef _LP64
1381 jlong res = (jlong)low->get_intptr();
1382 #else
1383 jlong res = jlong_from(value->get_jint(), low->get_jint());
1384 #endif
1385 obj->int_at_put(index, *(jint*)&res);
1386 obj->int_at_put(++index, *((jint*)&res + 1));
1387 } else {
1388 obj->int_at_put(index, value->get_jint());
1389 }
1390 break;
1391 }
1392
1393 case T_SHORT:
1394 assert(value->type() == T_INT, "Agreement.");
1395 obj->short_at_put(index, (jshort)value->get_jint());
1396 break;
1397
1398 case T_CHAR:
1399 assert(value->type() == T_INT, "Agreement.");
1400 obj->char_at_put(index, (jchar)value->get_jint());
1401 break;
1402
1403 case T_BYTE: {
1404 assert(value->type() == T_INT, "Agreement.");
1405 #if INCLUDE_JVMCI
1406 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1407 int byte_count = count_number_of_bytes_for_entry(sv, i);
1408 byte_array_put(obj, value, index, byte_count);
1409 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1410 i += byte_count - 1; // Balance the loop counter.
1411 index += byte_count;
1412 // index has been updated so continue at top of loop
1413 continue;
1414 #else
1415 obj->byte_at_put(index, (jbyte)value->get_jint());
1416 break;
1417 #endif // INCLUDE_JVMCI
1418 }
1419
1420 case T_BOOLEAN: {
1421 assert(value->type() == T_INT, "Agreement.");
1422 obj->bool_at_put(index, (jboolean)value->get_jint());
1423 break;
1424 }
1425
1426 default:
1427 ShouldNotReachHere();
1428 }
1429 index++;
1430 }
1431 }
1432
1433 // restore fields of an eliminated object array
1434 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1435 for (int i = 0; i < sv->field_size(); i++) {
1436 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1437 assert(value->type() == T_OBJECT, "object element expected");
1438 obj->obj_at_put(i, value->get_obj()());
1439 }
1440 }
1441
1442 class ReassignedField {
1443 public:
1444 int _offset;
1445 BasicType _type;
1446 public:
1447 ReassignedField() {
1448 _offset = 0;
1449 _type = T_ILLEGAL;
1450 }
1451 };
1452
1453 static int compare(ReassignedField* left, ReassignedField* right) {
1454 return left->_offset - right->_offset;
1455 }
1456
1457 // Restore fields of an eliminated instance object using the same field order
1458 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1459 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1460 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1461 InstanceKlass* ik = klass;
1462 while (ik != nullptr) {
1463 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1464 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) {
1465 ReassignedField field;
1466 field._offset = fs.offset();
1467 field._type = Signature::basic_type(fs.signature());
1468 fields->append(field);
1469 }
1470 }
1471 ik = ik->superklass();
1472 }
1473 fields->sort(compare);
1474 for (int i = 0; i < fields->length(); i++) {
1475 ScopeValue* scope_field = sv->field_at(svIndex);
1476 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1477 int offset = fields->at(i)._offset;
1478 BasicType type = fields->at(i)._type;
1479 switch (type) {
1480 case T_OBJECT: case T_ARRAY:
1481 assert(value->type() == T_OBJECT, "Agreement.");
1482 obj->obj_field_put(offset, value->get_obj()());
1483 break;
1484
1485 case T_INT: case T_FLOAT: { // 4 bytes.
1486 assert(value->type() == T_INT, "Agreement.");
1487 bool big_value = false;
1488 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1489 if (scope_field->is_location()) {
1490 Location::Type type = ((LocationValue*) scope_field)->location().type();
1491 if (type == Location::dbl || type == Location::lng) {
1492 big_value = true;
1493 }
1494 }
1495 if (scope_field->is_constant_int()) {
1496 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1497 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1498 big_value = true;
1499 }
1500 }
1501 }
1502
1503 if (big_value) {
1504 i++;
1505 assert(i < fields->length(), "second T_INT field needed");
1506 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1507 } else {
1508 obj->int_field_put(offset, value->get_jint());
1509 break;
1510 }
1511 }
1512 /* no break */
1513
1514 case T_LONG: case T_DOUBLE: {
1515 assert(value->type() == T_INT, "Agreement.");
1516 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1517 #ifdef _LP64
1518 jlong res = (jlong)low->get_intptr();
1519 #else
1520 jlong res = jlong_from(value->get_jint(), low->get_jint());
1521 #endif
1522 obj->long_field_put(offset, res);
1523 break;
1524 }
1525
1526 case T_SHORT:
1527 assert(value->type() == T_INT, "Agreement.");
1528 obj->short_field_put(offset, (jshort)value->get_jint());
1529 break;
1530
1531 case T_CHAR:
1532 assert(value->type() == T_INT, "Agreement.");
1533 obj->char_field_put(offset, (jchar)value->get_jint());
1534 break;
1535
1536 case T_BYTE:
1537 assert(value->type() == T_INT, "Agreement.");
1538 obj->byte_field_put(offset, (jbyte)value->get_jint());
1539 break;
1540
1541 case T_BOOLEAN:
1542 assert(value->type() == T_INT, "Agreement.");
1543 obj->bool_field_put(offset, (jboolean)value->get_jint());
1544 break;
1545
1546 default:
1547 ShouldNotReachHere();
1548 }
1549 svIndex++;
1550 }
1551 return svIndex;
1552 }
1553
1554 // restore fields of all eliminated objects and arrays
1555 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1556 for (int i = 0; i < objects->length(); i++) {
1557 assert(objects->at(i)->is_object(), "invalid debug information");
1558 ObjectValue* sv = (ObjectValue*) objects->at(i);
1559 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1560 Handle obj = sv->value();
1561 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1562 #ifndef PRODUCT
1563 if (PrintDeoptimizationDetails) {
1564 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1565 }
1566 #endif // !PRODUCT
1567
1568 if (obj.is_null()) {
1569 continue;
1570 }
1571
1572 #if INCLUDE_JVMCI
1573 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1574 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1575 continue;
1576 }
1577 #endif // INCLUDE_JVMCI
1578 if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1579 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1580 ScopeValue* payload = sv->field_at(0);
1581 if (payload->is_location() &&
1582 payload->as_LocationValue()->location().type() == Location::vector) {
1583 #ifndef PRODUCT
1584 if (PrintDeoptimizationDetails) {
1585 tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1586 if (Verbose) {
1587 Handle obj = sv->value();
1588 k->oop_print_on(obj(), tty);
1589 }
1590 }
1591 #endif // !PRODUCT
1592 continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1593 }
1594 // Else fall-through to do assignment for scalar-replaced boxed vector representation
1595 // which could be restored after vector object allocation.
1596 }
1597 if (k->is_instance_klass()) {
1598 InstanceKlass* ik = InstanceKlass::cast(k);
1599 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1600 } else if (k->is_typeArray_klass()) {
1601 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1602 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1603 } else if (k->is_objArray_klass()) {
1604 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1605 }
1606 }
1607 // These objects may escape when we return to Interpreter after deoptimization.
1608 // We need barrier so that stores that initialize these objects can't be reordered
1609 // with subsequent stores that make these objects accessible by other threads.
1610 OrderAccess::storestore();
1611 }
1612
1613
1614 // relock objects for which synchronization was eliminated
1615 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1616 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1617 bool relocked_objects = false;
1618 for (int i = 0; i < monitors->length(); i++) {
1619 MonitorInfo* mon_info = monitors->at(i);
1620 if (mon_info->eliminated()) {
1621 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1622 relocked_objects = true;
1623 if (!mon_info->owner_is_scalar_replaced()) {
1624 Handle obj(thread, mon_info->owner());
1625 markWord mark = obj->mark();
1626 if (exec_mode == Unpack_none) {
1627 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1628 // With exec_mode == Unpack_none obj may be thread local and locked in
1629 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1630 markWord dmw = mark.displaced_mark_helper();
1631 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1632 obj->set_mark(dmw);
1633 }
1634 if (mark.has_monitor()) {
1635 // defer relocking if the deoptee thread is currently waiting for obj
1636 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1637 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1638 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1639 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1640 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1641 continue;
1642 }
1643 }
1644 }
1645 if (LockingMode == LM_LIGHTWEIGHT) {
1646 // We have lost information about the correct state of the lock stack.
1647 // Inflate the locks instead. Enter then inflate to avoid races with
1648 // deflation.
1649 ObjectSynchronizer::enter_for(obj, nullptr, deoptee_thread);
1650 assert(mon_info->owner()->is_locked(), "object must be locked now");
1651 ObjectMonitor* mon = ObjectSynchronizer::inflate_for(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal);
1652 assert(mon->owner() == deoptee_thread, "must be");
1653 } else {
1654 BasicLock* lock = mon_info->lock();
1655 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1656 assert(mon_info->owner()->is_locked(), "object must be locked now");
1657 }
1658 }
1659 }
1660 }
1661 return relocked_objects;
1662 }
1663 #endif // COMPILER2_OR_JVMCI
1664
1665 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1666 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1667
1668 // Register map for next frame (used for stack crawl). We capture
1669 // the state of the deopt'ing frame's caller. Thus if we need to
1670 // stuff a C2I adapter we can properly fill in the callee-save
1671 // register locations.
1672 frame caller = fr.sender(reg_map);
1673 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1674
1675 frame sender = caller;
1676
1677 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1678 // the vframeArray containing the unpacking information is allocated in the C heap.
1679 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1680 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1681
1682 // Compare the vframeArray to the collected vframes
1683 assert(array->structural_compare(thread, chunk), "just checking");
1684
1685 if (TraceDeoptimization) {
1686 ResourceMark rm;
1687 stringStream st;
1688 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1689 st.print(" ");
1690 fr.print_on(&st);
1691 st.print_cr(" Virtual frames (innermost/newest first):");
1692 for (int index = 0; index < chunk->length(); index++) {
1693 compiledVFrame* vf = chunk->at(index);
1694 int bci = vf->raw_bci();
1695 const char* code_name;
1696 if (bci == SynchronizationEntryBCI) {
1697 code_name = "sync entry";
1698 } else {
1699 Bytecodes::Code code = vf->method()->code_at(bci);
1700 code_name = Bytecodes::name(code);
1701 }
1702
1703 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1704 st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1705 st.print(" - %s", code_name);
1706 st.print_cr(" @ bci=%d ", bci);
1707 }
1708 tty->print_raw(st.freeze());
1709 tty->cr();
1710 }
1711
1712 return array;
1713 }
1714
1715 #if COMPILER2_OR_JVMCI
1716 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1717 // Reallocation of some scalar replaced objects failed. Record
1718 // that we need to pop all the interpreter frames for the
1719 // deoptimized compiled frame.
1720 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1721 thread->set_frames_to_pop_failed_realloc(array->frames());
1722 // Unlock all monitors here otherwise the interpreter will see a
1723 // mix of locked and unlocked monitors (because of failed
1724 // reallocations of synchronized objects) and be confused.
1725 for (int i = 0; i < array->frames(); i++) {
1726 MonitorChunk* monitors = array->element(i)->monitors();
1727 if (monitors != nullptr) {
1728 // Unlock in reverse order starting from most nested monitor.
1729 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1730 BasicObjectLock* src = monitors->at(j);
1731 if (src->obj() != nullptr) {
1732 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1733 }
1734 }
1735 array->element(i)->free_monitors();
1736 #ifdef ASSERT
1737 array->element(i)->set_removed_monitors();
1738 #endif
1739 }
1740 }
1741 }
1742 #endif
1743
1744 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1745 assert(fr.can_be_deoptimized(), "checking frame type");
1746
1747 nmethod* nm = fr.cb()->as_nmethod_or_null();
1748 assert(nm != nullptr, "only compiled methods can deopt");
1749 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none);
1750 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc());
1751 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry
1752 : cur_sd->method()->java_code_at(cur_sd->bci()));
1753 gather_statistics(nm, reason, action, bc);
1754
1755 if (LogCompilation && xtty != nullptr) {
1756 ttyLocker ttyl;
1757 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1758 nm->log_identity(xtty);
1759 xtty->end_head();
1760 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1761 xtty->begin_elem("jvms bci='%d'", sd->bci());
1762 xtty->method(sd->method());
1763 xtty->end_elem();
1764 if (sd->is_top()) break;
1765 }
1766 xtty->tail("deoptimized");
1767 }
1768
1769 Continuation::notify_deopt(thread, fr.sp());
1770
1771 // Patch the compiled method so that when execution returns to it we will
1772 // deopt the execution state and return to the interpreter.
1773 fr.deoptimize(thread);
1774 }
1775
1776 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1777 // Deoptimize only if the frame comes from compile code.
1778 // Do not deoptimize the frame which is already patched
1779 // during the execution of the loops below.
1780 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1781 return;
1782 }
1783 ResourceMark rm;
1784 deoptimize_single_frame(thread, fr, reason);
1785 }
1786
1787 #if INCLUDE_JVMCI
1788 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1789 // there is no exception handler for this pc => deoptimize
1790 nm->make_not_entrant();
1791
1792 // Use Deoptimization::deoptimize for all of its side-effects:
1793 // gathering traps statistics, logging...
1794 // it also patches the return pc but we do not care about that
1795 // since we return a continuation to the deopt_blob below.
1796 JavaThread* thread = JavaThread::current();
1797 RegisterMap reg_map(thread,
1798 RegisterMap::UpdateMap::skip,
1799 RegisterMap::ProcessFrames::include,
1800 RegisterMap::WalkContinuation::skip);
1801 frame runtime_frame = thread->last_frame();
1802 frame caller_frame = runtime_frame.sender(®_map);
1803 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1804 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread);
1805 compiledVFrame* cvf = compiledVFrame::cast(vf);
1806 ScopeDesc* imm_scope = cvf->scope();
1807 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1808 if (imm_mdo != nullptr) {
1809 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1810 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1811
1812 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1813 if (pdata != nullptr && pdata->is_BitData()) {
1814 BitData* bit_data = (BitData*) pdata;
1815 bit_data->set_exception_seen();
1816 }
1817 }
1818
1819 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1820
1821 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1822 if (trap_mdo != nullptr) {
1823 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1824 }
1825
1826 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1827 }
1828 #endif
1829
1830 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1831 assert(thread == Thread::current() ||
1832 thread->is_handshake_safe_for(Thread::current()) ||
1833 SafepointSynchronize::is_at_safepoint(),
1834 "can only deoptimize other thread at a safepoint/handshake");
1835 // Compute frame and register map based on thread and sp.
1836 RegisterMap reg_map(thread,
1837 RegisterMap::UpdateMap::skip,
1838 RegisterMap::ProcessFrames::include,
1839 RegisterMap::WalkContinuation::skip);
1840 frame fr = thread->last_frame();
1841 while (fr.id() != id) {
1842 fr = fr.sender(®_map);
1843 }
1844 deoptimize(thread, fr, reason);
1845 }
1846
1847
1848 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1849 Thread* current = Thread::current();
1850 if (thread == current || thread->is_handshake_safe_for(current)) {
1851 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1852 } else {
1853 VM_DeoptimizeFrame deopt(thread, id, reason);
1854 VMThread::execute(&deopt);
1855 }
1856 }
1857
1858 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1859 deoptimize_frame(thread, id, Reason_constraint);
1860 }
1861
1862 // JVMTI PopFrame support
1863 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1864 {
1865 assert(thread == JavaThread::current(), "pre-condition");
1866 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1867 }
1868 JRT_END
1869
1870 MethodData*
1871 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1872 bool create_if_missing) {
1873 JavaThread* THREAD = thread; // For exception macros.
1874 MethodData* mdo = m()->method_data();
1875 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1876 // Build an MDO. Ignore errors like OutOfMemory;
1877 // that simply means we won't have an MDO to update.
1878 Method::build_profiling_method_data(m, THREAD);
1879 if (HAS_PENDING_EXCEPTION) {
1880 // Only metaspace OOM is expected. No Java code executed.
1881 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1882 CLEAR_PENDING_EXCEPTION;
1883 }
1884 mdo = m()->method_data();
1885 }
1886 return mdo;
1887 }
1888
1889 #if COMPILER2_OR_JVMCI
1890 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1891 // In case of an unresolved klass entry, load the class.
1892 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1893 // and probably nowhere else.
1894 // Even that case would benefit from simply re-interpreting the
1895 // bytecode, without paying special attention to the class index.
1896 // So this whole "class index" feature should probably be removed.
1897
1898 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1899 Klass* tk = constant_pool->klass_at(index, THREAD);
1900 if (HAS_PENDING_EXCEPTION) {
1901 // Exception happened during classloading. We ignore the exception here, since it
1902 // is going to be rethrown since the current activation is going to be deoptimized and
1903 // the interpreter will re-execute the bytecode.
1904 // Do not clear probable Async Exceptions.
1905 CLEAR_PENDING_NONASYNC_EXCEPTION;
1906 // Class loading called java code which may have caused a stack
1907 // overflow. If the exception was thrown right before the return
1908 // to the runtime the stack is no longer guarded. Reguard the
1909 // stack otherwise if we return to the uncommon trap blob and the
1910 // stack bang causes a stack overflow we crash.
1911 JavaThread* jt = THREAD;
1912 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1913 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1914 }
1915 return;
1916 }
1917
1918 assert(!constant_pool->tag_at(index).is_symbol(),
1919 "no symbolic names here, please");
1920 }
1921
1922 #if INCLUDE_JFR
1923
1924 class DeoptReasonSerializer : public JfrSerializer {
1925 public:
1926 void serialize(JfrCheckpointWriter& writer) {
1927 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1928 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1929 writer.write_key((u8)i);
1930 writer.write(Deoptimization::trap_reason_name(i));
1931 }
1932 }
1933 };
1934
1935 class DeoptActionSerializer : public JfrSerializer {
1936 public:
1937 void serialize(JfrCheckpointWriter& writer) {
1938 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1939 writer.write_count(nof_actions);
1940 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1941 writer.write_key(i);
1942 writer.write(Deoptimization::trap_action_name((int)i));
1943 }
1944 }
1945 };
1946
1947 static void register_serializers() {
1948 static int critical_section = 0;
1949 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1950 return;
1951 }
1952 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1953 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1954 }
1955
1956 static void post_deoptimization_event(nmethod* nm,
1957 const Method* method,
1958 int trap_bci,
1959 int instruction,
1960 Deoptimization::DeoptReason reason,
1961 Deoptimization::DeoptAction action) {
1962 assert(nm != nullptr, "invariant");
1963 assert(method != nullptr, "invariant");
1964 if (EventDeoptimization::is_enabled()) {
1965 static bool serializers_registered = false;
1966 if (!serializers_registered) {
1967 register_serializers();
1968 serializers_registered = true;
1969 }
1970 EventDeoptimization event;
1971 event.set_compileId(nm->compile_id());
1972 event.set_compiler(nm->compiler_type());
1973 event.set_method(method);
1974 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1975 event.set_bci(trap_bci);
1976 event.set_instruction(instruction);
1977 event.set_reason(reason);
1978 event.set_action(action);
1979 event.commit();
1980 }
1981 }
1982
1983 #endif // INCLUDE_JFR
1984
1985 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
1986 const char* reason_name, const char* reason_action) {
1987 LogTarget(Debug, deoptimization) lt;
1988 if (lt.is_enabled()) {
1989 LogStream ls(lt);
1990 bool is_osr = nm->is_osr_method();
1991 ls.print("cid=%4d %s level=%d",
1992 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level());
1993 ls.print(" %s", tm->name_and_sig_as_C_string());
1994 ls.print(" trap_bci=%d ", trap_bci);
1995 if (is_osr) {
1996 ls.print("osr_bci=%d ", nm->osr_entry_bci());
1997 }
1998 ls.print("%s ", reason_name);
1999 ls.print("%s ", reason_action);
2000 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2001 pc, fr.pc() - nm->code_begin());
2002 }
2003 }
2004
2005 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2006 HandleMark hm(current);
2007
2008 // uncommon_trap() is called at the beginning of the uncommon trap
2009 // handler. Note this fact before we start generating temporary frames
2010 // that can confuse an asynchronous stack walker. This counter is
2011 // decremented at the end of unpack_frames().
2012
2013 current->inc_in_deopt_handler();
2014
2015 #if INCLUDE_JVMCI
2016 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2017 RegisterMap reg_map(current,
2018 RegisterMap::UpdateMap::include,
2019 RegisterMap::ProcessFrames::include,
2020 RegisterMap::WalkContinuation::skip);
2021 #else
2022 RegisterMap reg_map(current,
2023 RegisterMap::UpdateMap::skip,
2024 RegisterMap::ProcessFrames::include,
2025 RegisterMap::WalkContinuation::skip);
2026 #endif
2027 frame stub_frame = current->last_frame();
2028 frame fr = stub_frame.sender(®_map);
2029
2030 // Log a message
2031 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2032 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2033
2034 {
2035 ResourceMark rm;
2036
2037 DeoptReason reason = trap_request_reason(trap_request);
2038 DeoptAction action = trap_request_action(trap_request);
2039 #if INCLUDE_JVMCI
2040 int debug_id = trap_request_debug_id(trap_request);
2041 #endif
2042 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2043
2044 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
2045 compiledVFrame* cvf = compiledVFrame::cast(vf);
2046
2047 nmethod* nm = cvf->code();
2048
2049 ScopeDesc* trap_scope = cvf->scope();
2050
2051 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2052
2053 if (is_receiver_constraint_failure) {
2054 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2055 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2056 JVMCI_ONLY(COMMA debug_id));
2057 }
2058
2059 methodHandle trap_method(current, trap_scope->method());
2060 int trap_bci = trap_scope->bci();
2061 #if INCLUDE_JVMCI
2062 jlong speculation = current->pending_failed_speculation();
2063 if (nm->is_compiled_by_jvmci()) {
2064 nm->update_speculation(current);
2065 } else {
2066 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2067 }
2068
2069 if (trap_bci == SynchronizationEntryBCI) {
2070 trap_bci = 0;
2071 current->set_pending_monitorenter(true);
2072 }
2073
2074 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2075 current->set_pending_transfer_to_interpreter(true);
2076 }
2077 #endif
2078
2079 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
2080 // Record this event in the histogram.
2081 gather_statistics(nm, reason, action, trap_bc);
2082
2083 // Ensure that we can record deopt. history:
2084 bool create_if_missing = ProfileTraps;
2085
2086 methodHandle profiled_method;
2087 #if INCLUDE_JVMCI
2088 if (nm->is_compiled_by_jvmci()) {
2089 profiled_method = methodHandle(current, nm->method());
2090 } else {
2091 profiled_method = trap_method;
2092 }
2093 #else
2094 profiled_method = trap_method;
2095 #endif
2096
2097 MethodData* trap_mdo =
2098 get_method_data(current, profiled_method, create_if_missing);
2099
2100 { // Log Deoptimization event for JFR, UL and event system
2101 Method* tm = trap_method();
2102 const char* reason_name = trap_reason_name(reason);
2103 const char* reason_action = trap_action_name(action);
2104 intptr_t pc = p2i(fr.pc());
2105
2106 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2107 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2108 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2109 reason_name, reason_action, pc,
2110 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2111 }
2112
2113 // Print a bunch of diagnostics, if requested.
2114 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2115 ResourceMark rm;
2116
2117 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2118 // We must do this already now, since we cannot acquire this lock while
2119 // holding the tty lock (lock ordering by rank).
2120 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2121
2122 ttyLocker ttyl;
2123
2124 char buf[100];
2125 if (xtty != nullptr) {
2126 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
2127 os::current_thread_id(),
2128 format_trap_request(buf, sizeof(buf), trap_request));
2129 #if INCLUDE_JVMCI
2130 if (speculation != 0) {
2131 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2132 }
2133 #endif
2134 nm->log_identity(xtty);
2135 }
2136 Symbol* class_name = nullptr;
2137 bool unresolved = false;
2138 if (unloaded_class_index >= 0) {
2139 constantPoolHandle constants (current, trap_method->constants());
2140 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2141 class_name = constants->klass_name_at(unloaded_class_index);
2142 unresolved = true;
2143 if (xtty != nullptr)
2144 xtty->print(" unresolved='1'");
2145 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2146 class_name = constants->symbol_at(unloaded_class_index);
2147 }
2148 if (xtty != nullptr)
2149 xtty->name(class_name);
2150 }
2151 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2152 // Dump the relevant MDO state.
2153 // This is the deopt count for the current reason, any previous
2154 // reasons or recompiles seen at this point.
2155 int dcnt = trap_mdo->trap_count(reason);
2156 if (dcnt != 0)
2157 xtty->print(" count='%d'", dcnt);
2158
2159 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2160 // lock extra_data_lock before the tty lock.
2161 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2162 int dos = (pdata == nullptr)? 0: pdata->trap_state();
2163 if (dos != 0) {
2164 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2165 if (trap_state_is_recompiled(dos)) {
2166 int recnt2 = trap_mdo->overflow_recompile_count();
2167 if (recnt2 != 0)
2168 xtty->print(" recompiles2='%d'", recnt2);
2169 }
2170 }
2171 }
2172 if (xtty != nullptr) {
2173 xtty->stamp();
2174 xtty->end_head();
2175 }
2176 if (TraceDeoptimization) { // make noise on the tty
2177 stringStream st;
2178 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2179 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2180 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2181 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2182 #if INCLUDE_JVMCI
2183 if (nm->is_compiled_by_jvmci()) {
2184 const char* installed_code_name = nm->jvmci_name();
2185 if (installed_code_name != nullptr) {
2186 st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2187 }
2188 }
2189 #endif
2190 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2191 p2i(fr.pc()),
2192 os::current_thread_id(),
2193 trap_reason_name(reason),
2194 trap_action_name(action),
2195 unloaded_class_index
2196 #if INCLUDE_JVMCI
2197 , debug_id
2198 #endif
2199 );
2200 if (class_name != nullptr) {
2201 st.print(unresolved ? " unresolved class: " : " symbol: ");
2202 class_name->print_symbol_on(&st);
2203 }
2204 st.cr();
2205 tty->print_raw(st.freeze());
2206 }
2207 if (xtty != nullptr) {
2208 // Log the precise location of the trap.
2209 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2210 xtty->begin_elem("jvms bci='%d'", sd->bci());
2211 xtty->method(sd->method());
2212 xtty->end_elem();
2213 if (sd->is_top()) break;
2214 }
2215 xtty->tail("uncommon_trap");
2216 }
2217 }
2218 // (End diagnostic printout.)
2219
2220 if (is_receiver_constraint_failure) {
2221 fatal("missing receiver type check");
2222 }
2223
2224 // Load class if necessary
2225 if (unloaded_class_index >= 0) {
2226 constantPoolHandle constants(current, trap_method->constants());
2227 load_class_by_index(constants, unloaded_class_index, THREAD);
2228 }
2229
2230 // Flush the nmethod if necessary and desirable.
2231 //
2232 // We need to avoid situations where we are re-flushing the nmethod
2233 // because of a hot deoptimization site. Repeated flushes at the same
2234 // point need to be detected by the compiler and avoided. If the compiler
2235 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2236 // module must take measures to avoid an infinite cycle of recompilation
2237 // and deoptimization. There are several such measures:
2238 //
2239 // 1. If a recompilation is ordered a second time at some site X
2240 // and for the same reason R, the action is adjusted to 'reinterpret',
2241 // to give the interpreter time to exercise the method more thoroughly.
2242 // If this happens, the method's overflow_recompile_count is incremented.
2243 //
2244 // 2. If the compiler fails to reduce the deoptimization rate, then
2245 // the method's overflow_recompile_count will begin to exceed the set
2246 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2247 // is adjusted to 'make_not_compilable', and the method is abandoned
2248 // to the interpreter. This is a performance hit for hot methods,
2249 // but is better than a disastrous infinite cycle of recompilations.
2250 // (Actually, only the method containing the site X is abandoned.)
2251 //
2252 // 3. In parallel with the previous measures, if the total number of
2253 // recompilations of a method exceeds the much larger set limit
2254 // PerMethodRecompilationCutoff, the method is abandoned.
2255 // This should only happen if the method is very large and has
2256 // many "lukewarm" deoptimizations. The code which enforces this
2257 // limit is elsewhere (class nmethod, class Method).
2258 //
2259 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2260 // to recompile at each bytecode independently of the per-BCI cutoff.
2261 //
2262 // The decision to update code is up to the compiler, and is encoded
2263 // in the Action_xxx code. If the compiler requests Action_none
2264 // no trap state is changed, no compiled code is changed, and the
2265 // computation suffers along in the interpreter.
2266 //
2267 // The other action codes specify various tactics for decompilation
2268 // and recompilation. Action_maybe_recompile is the loosest, and
2269 // allows the compiled code to stay around until enough traps are seen,
2270 // and until the compiler gets around to recompiling the trapping method.
2271 //
2272 // The other actions cause immediate removal of the present code.
2273
2274 // Traps caused by injected profile shouldn't pollute trap counts.
2275 bool injected_profile_trap = trap_method->has_injected_profile() &&
2276 (reason == Reason_intrinsic || reason == Reason_unreached);
2277
2278 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2279 bool make_not_entrant = false;
2280 bool make_not_compilable = false;
2281 bool reprofile = false;
2282 switch (action) {
2283 case Action_none:
2284 // Keep the old code.
2285 update_trap_state = false;
2286 break;
2287 case Action_maybe_recompile:
2288 // Do not need to invalidate the present code, but we can
2289 // initiate another
2290 // Start compiler without (necessarily) invalidating the nmethod.
2291 // The system will tolerate the old code, but new code should be
2292 // generated when possible.
2293 break;
2294 case Action_reinterpret:
2295 // Go back into the interpreter for a while, and then consider
2296 // recompiling form scratch.
2297 make_not_entrant = true;
2298 // Reset invocation counter for outer most method.
2299 // This will allow the interpreter to exercise the bytecodes
2300 // for a while before recompiling.
2301 // By contrast, Action_make_not_entrant is immediate.
2302 //
2303 // Note that the compiler will track null_check, null_assert,
2304 // range_check, and class_check events and log them as if they
2305 // had been traps taken from compiled code. This will update
2306 // the MDO trap history so that the next compilation will
2307 // properly detect hot trap sites.
2308 reprofile = true;
2309 break;
2310 case Action_make_not_entrant:
2311 // Request immediate recompilation, and get rid of the old code.
2312 // Make them not entrant, so next time they are called they get
2313 // recompiled. Unloaded classes are loaded now so recompile before next
2314 // time they are called. Same for uninitialized. The interpreter will
2315 // link the missing class, if any.
2316 make_not_entrant = true;
2317 break;
2318 case Action_make_not_compilable:
2319 // Give up on compiling this method at all.
2320 make_not_entrant = true;
2321 make_not_compilable = true;
2322 break;
2323 default:
2324 ShouldNotReachHere();
2325 }
2326
2327 // Setting +ProfileTraps fixes the following, on all platforms:
2328 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2329 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2330 // recompile relies on a MethodData* to record heroic opt failures.
2331
2332 // Whether the interpreter is producing MDO data or not, we also need
2333 // to use the MDO to detect hot deoptimization points and control
2334 // aggressive optimization.
2335 bool inc_recompile_count = false;
2336
2337 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2338 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2339 (trap_mdo != nullptr),
2340 Mutex::_no_safepoint_check_flag);
2341 ProfileData* pdata = nullptr;
2342 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2343 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2344 uint this_trap_count = 0;
2345 bool maybe_prior_trap = false;
2346 bool maybe_prior_recompile = false;
2347
2348 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2349 #if INCLUDE_JVMCI
2350 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2351 #endif
2352 nm->method(),
2353 //outputs:
2354 this_trap_count,
2355 maybe_prior_trap,
2356 maybe_prior_recompile);
2357 // Because the interpreter also counts null, div0, range, and class
2358 // checks, these traps from compiled code are double-counted.
2359 // This is harmless; it just means that the PerXTrapLimit values
2360 // are in effect a little smaller than they look.
2361
2362 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2363 if (per_bc_reason != Reason_none) {
2364 // Now take action based on the partially known per-BCI history.
2365 if (maybe_prior_trap
2366 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2367 // If there are too many traps at this BCI, force a recompile.
2368 // This will allow the compiler to see the limit overflow, and
2369 // take corrective action, if possible. The compiler generally
2370 // does not use the exact PerBytecodeTrapLimit value, but instead
2371 // changes its tactics if it sees any traps at all. This provides
2372 // a little hysteresis, delaying a recompile until a trap happens
2373 // several times.
2374 //
2375 // Actually, since there is only one bit of counter per BCI,
2376 // the possible per-BCI counts are {0,1,(per-method count)}.
2377 // This produces accurate results if in fact there is only
2378 // one hot trap site, but begins to get fuzzy if there are
2379 // many sites. For example, if there are ten sites each
2380 // trapping two or more times, they each get the blame for
2381 // all of their traps.
2382 make_not_entrant = true;
2383 }
2384
2385 // Detect repeated recompilation at the same BCI, and enforce a limit.
2386 if (make_not_entrant && maybe_prior_recompile) {
2387 // More than one recompile at this point.
2388 inc_recompile_count = maybe_prior_trap;
2389 }
2390 } else {
2391 // For reasons which are not recorded per-bytecode, we simply
2392 // force recompiles unconditionally.
2393 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2394 make_not_entrant = true;
2395 }
2396
2397 // Go back to the compiler if there are too many traps in this method.
2398 if (this_trap_count >= per_method_trap_limit(reason)) {
2399 // If there are too many traps in this method, force a recompile.
2400 // This will allow the compiler to see the limit overflow, and
2401 // take corrective action, if possible.
2402 // (This condition is an unlikely backstop only, because the
2403 // PerBytecodeTrapLimit is more likely to take effect first,
2404 // if it is applicable.)
2405 make_not_entrant = true;
2406 }
2407
2408 // Here's more hysteresis: If there has been a recompile at
2409 // this trap point already, run the method in the interpreter
2410 // for a while to exercise it more thoroughly.
2411 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2412 reprofile = true;
2413 }
2414 }
2415
2416 // Take requested actions on the method:
2417
2418 // Recompile
2419 if (make_not_entrant) {
2420 if (!nm->make_not_entrant()) {
2421 return; // the call did not change nmethod's state
2422 }
2423
2424 if (pdata != nullptr) {
2425 // Record the recompilation event, if any.
2426 int tstate0 = pdata->trap_state();
2427 int tstate1 = trap_state_set_recompiled(tstate0, true);
2428 if (tstate1 != tstate0)
2429 pdata->set_trap_state(tstate1);
2430 }
2431
2432 // For code aging we count traps separately here, using make_not_entrant()
2433 // as a guard against simultaneous deopts in multiple threads.
2434 if (reason == Reason_tenured && trap_mdo != nullptr) {
2435 trap_mdo->inc_tenure_traps();
2436 }
2437 }
2438
2439 if (inc_recompile_count) {
2440 trap_mdo->inc_overflow_recompile_count();
2441 if ((uint)trap_mdo->overflow_recompile_count() >
2442 (uint)PerBytecodeRecompilationCutoff) {
2443 // Give up on the method containing the bad BCI.
2444 if (trap_method() == nm->method()) {
2445 make_not_compilable = true;
2446 } else {
2447 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2448 // But give grace to the enclosing nm->method().
2449 }
2450 }
2451 }
2452
2453 // Reprofile
2454 if (reprofile) {
2455 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2456 }
2457
2458 // Give up compiling
2459 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2460 assert(make_not_entrant, "consistent");
2461 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2462 }
2463
2464 if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2465 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2466 if (exception_handler_data != nullptr) {
2467 // uncommon trap at the start of an exception handler.
2468 // C2 generates these for un-entered exception handlers.
2469 // mark the handler as entered to avoid generating
2470 // another uncommon trap the next time the handler is compiled
2471 exception_handler_data->set_exception_handler_entered();
2472 }
2473 }
2474
2475 } // Free marked resources
2476
2477 }
2478 JRT_END
2479
2480 ProfileData*
2481 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2482 int trap_bci,
2483 Deoptimization::DeoptReason reason,
2484 bool update_total_trap_count,
2485 #if INCLUDE_JVMCI
2486 bool is_osr,
2487 #endif
2488 Method* compiled_method,
2489 //outputs:
2490 uint& ret_this_trap_count,
2491 bool& ret_maybe_prior_trap,
2492 bool& ret_maybe_prior_recompile) {
2493 trap_mdo->check_extra_data_locked();
2494
2495 bool maybe_prior_trap = false;
2496 bool maybe_prior_recompile = false;
2497 uint this_trap_count = 0;
2498 if (update_total_trap_count) {
2499 uint idx = reason;
2500 #if INCLUDE_JVMCI
2501 if (is_osr) {
2502 // Upper half of history array used for traps in OSR compilations
2503 idx += Reason_TRAP_HISTORY_LENGTH;
2504 }
2505 #endif
2506 uint prior_trap_count = trap_mdo->trap_count(idx);
2507 this_trap_count = trap_mdo->inc_trap_count(idx);
2508
2509 // If the runtime cannot find a place to store trap history,
2510 // it is estimated based on the general condition of the method.
2511 // If the method has ever been recompiled, or has ever incurred
2512 // a trap with the present reason , then this BCI is assumed
2513 // (pessimistically) to be the culprit.
2514 maybe_prior_trap = (prior_trap_count != 0);
2515 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2516 }
2517 ProfileData* pdata = nullptr;
2518
2519
2520 // For reasons which are recorded per bytecode, we check per-BCI data.
2521 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2522 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2523 if (per_bc_reason != Reason_none) {
2524 // Find the profile data for this BCI. If there isn't one,
2525 // try to allocate one from the MDO's set of spares.
2526 // This will let us detect a repeated trap at this point.
2527 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2528
2529 if (pdata != nullptr) {
2530 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2531 if (LogCompilation && xtty != nullptr) {
2532 ttyLocker ttyl;
2533 // no more room for speculative traps in this MDO
2534 xtty->elem("speculative_traps_oom");
2535 }
2536 }
2537 // Query the trap state of this profile datum.
2538 int tstate0 = pdata->trap_state();
2539 if (!trap_state_has_reason(tstate0, per_bc_reason))
2540 maybe_prior_trap = false;
2541 if (!trap_state_is_recompiled(tstate0))
2542 maybe_prior_recompile = false;
2543
2544 // Update the trap state of this profile datum.
2545 int tstate1 = tstate0;
2546 // Record the reason.
2547 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2548 // Store the updated state on the MDO, for next time.
2549 if (tstate1 != tstate0)
2550 pdata->set_trap_state(tstate1);
2551 } else {
2552 if (LogCompilation && xtty != nullptr) {
2553 ttyLocker ttyl;
2554 // Missing MDP? Leave a small complaint in the log.
2555 xtty->elem("missing_mdp bci='%d'", trap_bci);
2556 }
2557 }
2558 }
2559
2560 // Return results:
2561 ret_this_trap_count = this_trap_count;
2562 ret_maybe_prior_trap = maybe_prior_trap;
2563 ret_maybe_prior_recompile = maybe_prior_recompile;
2564 return pdata;
2565 }
2566
2567 void
2568 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2569 ResourceMark rm;
2570 // Ignored outputs:
2571 uint ignore_this_trap_count;
2572 bool ignore_maybe_prior_trap;
2573 bool ignore_maybe_prior_recompile;
2574 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2575 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2576 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2577
2578 // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2579 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2580
2581 query_update_method_data(trap_mdo, trap_bci,
2582 (DeoptReason)reason,
2583 update_total_counts,
2584 #if INCLUDE_JVMCI
2585 false,
2586 #endif
2587 nullptr,
2588 ignore_this_trap_count,
2589 ignore_maybe_prior_trap,
2590 ignore_maybe_prior_recompile);
2591 }
2592
2593 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode))
2594 // Enable WXWrite: current function is called from methods compiled by C2 directly
2595 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2596
2597 // Still in Java no safepoints
2598 {
2599 // This enters VM and may safepoint
2600 uncommon_trap_inner(current, trap_request);
2601 }
2602 HandleMark hm(current);
2603 return fetch_unroll_info_helper(current, exec_mode);
2604 PROF_END
2605
2606 // Local derived constants.
2607 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2608 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2609 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2610
2611 //---------------------------trap_state_reason---------------------------------
2612 Deoptimization::DeoptReason
2613 Deoptimization::trap_state_reason(int trap_state) {
2614 // This assert provides the link between the width of DataLayout::trap_bits
2615 // and the encoding of "recorded" reasons. It ensures there are enough
2616 // bits to store all needed reasons in the per-BCI MDO profile.
2617 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2618 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2619 trap_state -= recompile_bit;
2620 if (trap_state == DS_REASON_MASK) {
2621 return Reason_many;
2622 } else {
2623 assert((int)Reason_none == 0, "state=0 => Reason_none");
2624 return (DeoptReason)trap_state;
2625 }
2626 }
2627 //-------------------------trap_state_has_reason-------------------------------
2628 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2629 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2630 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2631 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2632 trap_state -= recompile_bit;
2633 if (trap_state == DS_REASON_MASK) {
2634 return -1; // true, unspecifically (bottom of state lattice)
2635 } else if (trap_state == reason) {
2636 return 1; // true, definitely
2637 } else if (trap_state == 0) {
2638 return 0; // false, definitely (top of state lattice)
2639 } else {
2640 return 0; // false, definitely
2641 }
2642 }
2643 //-------------------------trap_state_add_reason-------------------------------
2644 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2645 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2646 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2647 trap_state -= recompile_bit;
2648 if (trap_state == DS_REASON_MASK) {
2649 return trap_state + recompile_bit; // already at state lattice bottom
2650 } else if (trap_state == reason) {
2651 return trap_state + recompile_bit; // the condition is already true
2652 } else if (trap_state == 0) {
2653 return reason + recompile_bit; // no condition has yet been true
2654 } else {
2655 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2656 }
2657 }
2658 //-----------------------trap_state_is_recompiled------------------------------
2659 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2660 return (trap_state & DS_RECOMPILE_BIT) != 0;
2661 }
2662 //-----------------------trap_state_set_recompiled-----------------------------
2663 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2664 if (z) return trap_state | DS_RECOMPILE_BIT;
2665 else return trap_state & ~DS_RECOMPILE_BIT;
2666 }
2667 //---------------------------format_trap_state---------------------------------
2668 // This is used for debugging and diagnostics, including LogFile output.
2669 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2670 int trap_state) {
2671 assert(buflen > 0, "sanity");
2672 DeoptReason reason = trap_state_reason(trap_state);
2673 bool recomp_flag = trap_state_is_recompiled(trap_state);
2674 // Re-encode the state from its decoded components.
2675 int decoded_state = 0;
2676 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2677 decoded_state = trap_state_add_reason(decoded_state, reason);
2678 if (recomp_flag)
2679 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2680 // If the state re-encodes properly, format it symbolically.
2681 // Because this routine is used for debugging and diagnostics,
2682 // be robust even if the state is a strange value.
2683 size_t len;
2684 if (decoded_state != trap_state) {
2685 // Random buggy state that doesn't decode??
2686 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2687 } else {
2688 len = jio_snprintf(buf, buflen, "%s%s",
2689 trap_reason_name(reason),
2690 recomp_flag ? " recompiled" : "");
2691 }
2692 return buf;
2693 }
2694
2695
2696 //--------------------------------statics--------------------------------------
2697 const char* Deoptimization::_trap_reason_name[] = {
2698 // Note: Keep this in sync. with enum DeoptReason.
2699 "none",
2700 "null_check",
2701 "null_assert" JVMCI_ONLY("_or_unreached0"),
2702 "range_check",
2703 "class_check",
2704 "array_check",
2705 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2706 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2707 "profile_predicate",
2708 "unloaded",
2709 "uninitialized",
2710 "initialized",
2711 "unreached",
2712 "unhandled",
2713 "constraint",
2714 "div0_check",
2715 "age",
2716 "predicate",
2717 "loop_limit_check",
2718 "speculate_class_check",
2719 "speculate_null_check",
2720 "speculate_null_assert",
2721 "unstable_if",
2722 "unstable_fused_if",
2723 "receiver_constraint",
2724 #if INCLUDE_JVMCI
2725 "aliasing",
2726 "transfer_to_interpreter",
2727 "not_compiled_exception_handler",
2728 "unresolved",
2729 "jsr_mismatch",
2730 #endif
2731 "tenured"
2732 };
2733 const char* Deoptimization::_trap_action_name[] = {
2734 // Note: Keep this in sync. with enum DeoptAction.
2735 "none",
2736 "maybe_recompile",
2737 "reinterpret",
2738 "make_not_entrant",
2739 "make_not_compilable"
2740 };
2741
2742 const char* Deoptimization::trap_reason_name(int reason) {
2743 // Check that every reason has a name
2744 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2745
2746 if (reason == Reason_many) return "many";
2747 if ((uint)reason < Reason_LIMIT)
2748 return _trap_reason_name[reason];
2749 static char buf[20];
2750 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2751 return buf;
2752 }
2753 const char* Deoptimization::trap_action_name(int action) {
2754 // Check that every action has a name
2755 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2756
2757 if ((uint)action < Action_LIMIT)
2758 return _trap_action_name[action];
2759 static char buf[20];
2760 os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2761 return buf;
2762 }
2763
2764 // This is used for debugging and diagnostics, including LogFile output.
2765 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2766 int trap_request) {
2767 jint unloaded_class_index = trap_request_index(trap_request);
2768 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2769 const char* action = trap_action_name(trap_request_action(trap_request));
2770 #if INCLUDE_JVMCI
2771 int debug_id = trap_request_debug_id(trap_request);
2772 #endif
2773 size_t len;
2774 if (unloaded_class_index < 0) {
2775 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2776 reason, action
2777 #if INCLUDE_JVMCI
2778 ,debug_id
2779 #endif
2780 );
2781 } else {
2782 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2783 reason, action, unloaded_class_index
2784 #if INCLUDE_JVMCI
2785 ,debug_id
2786 #endif
2787 );
2788 }
2789 return buf;
2790 }
2791
2792 juint Deoptimization::_deoptimization_hist
2793 [1 + 4 + 5] // total + online + archived
2794 [Deoptimization::Reason_LIMIT]
2795 [1 + Deoptimization::Action_LIMIT]
2796 [Deoptimization::BC_CASE_LIMIT]
2797 = {0};
2798
2799 enum {
2800 LSB_BITS = 8,
2801 LSB_MASK = right_n_bits(LSB_BITS)
2802 };
2803
2804 static void update(juint* cases, Bytecodes::Code bc) {
2805 juint* bc_counter_addr = nullptr;
2806 juint bc_counter = 0;
2807 // Look for an unused counter, or an exact match to this BC.
2808 if (bc != Bytecodes::_illegal) {
2809 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) {
2810 juint* counter_addr = &cases[bc_case];
2811 juint counter = *counter_addr;
2812 if ((counter == 0 && bc_counter_addr == nullptr)
2813 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2814 // this counter is either free or is already devoted to this BC
2815 bc_counter_addr = counter_addr;
2816 bc_counter = counter | bc;
2817 }
2818 }
2819 }
2820 if (bc_counter_addr == nullptr) {
2821 // Overflow, or no given bytecode.
2822 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1];
2823 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2824 }
2825 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2826 }
2827
2828
2829 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action,
2830 Bytecodes::Code bc) {
2831 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2832 assert(action >= 0 && action < Action_LIMIT, "oob");
2833 _deoptimization_hist[0][Reason_none][0][0] += 1; // total
2834 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total
2835
2836 update(_deoptimization_hist[0][reason][1+action], bc);
2837
2838 uint lvl = nm->comp_level() + (nm->is_scc() ? 4 : 0) + (nm->preloaded() ? 1 : 0);
2839 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total
2840 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total
2841 update(_deoptimization_hist[lvl][reason][1+action], bc);
2842 }
2843
2844 jint Deoptimization::total_deoptimization_count() {
2845 return _deoptimization_hist[0][Reason_none][0][0];
2846 }
2847
2848 // Get the deopt count for a specific reason and a specific action. If either
2849 // one of 'reason' or 'action' is null, the method returns the sum of all
2850 // deoptimizations with the specific 'action' or 'reason' respectively.
2851 // If both arguments are null, the method returns the total deopt count.
2852 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2853 if (reason_str == nullptr && action_str == nullptr) {
2854 return total_deoptimization_count();
2855 }
2856 juint counter = 0;
2857 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2858 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2859 for (int action = 0; action < Action_LIMIT; action++) {
2860 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2861 juint* cases = _deoptimization_hist[0][reason][1+action];
2862 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2863 counter += cases[bc_case] >> LSB_BITS;
2864 }
2865 }
2866 }
2867 }
2868 }
2869 return counter;
2870 }
2871
2872 void Deoptimization::print_statistics() {
2873 ttyLocker ttyl;
2874 if (xtty != nullptr) xtty->head("statistics type='deoptimization'");
2875 tty->print_cr("Deoptimization traps recorded:");
2876 print_statistics_on(tty);
2877 if (xtty != nullptr) xtty->tail("statistics");
2878 }
2879
2880 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) {
2881 juint total = _deoptimization_hist[lvl][Reason_none][0][0];
2882 juint account = total;
2883 #define PRINT_STAT_LINE(name, r) \
2884 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name);
2885 if (total > 0) {
2886 st->print(" %s: ", title);
2887 PRINT_STAT_LINE("total", total);
2888 // For each non-zero entry in the histogram, print the reason,
2889 // the action, and (if specifically known) the type of bytecode.
2890 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2891 for (int action = 0; action < Action_LIMIT; action++) {
2892 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action];
2893 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2894 juint counter = cases[bc_case];
2895 if (counter != 0) {
2896 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2897 const char* bc_name = "other";
2898 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) {
2899 // overwritten
2900 } else if (Bytecodes::is_defined(bc)) {
2901 bc_name = Bytecodes::name(bc);
2902 }
2903 juint r = counter >> LSB_BITS;
2904 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)",
2905 trap_reason_name(reason), trap_action_name(action), bc_name,
2906 r, (r * 100.0) / total);
2907 account -= r;
2908 }
2909 }
2910 }
2911 }
2912 if (account != 0) {
2913 PRINT_STAT_LINE("unaccounted", account);
2914 }
2915 #undef PRINT_STAT_LINE
2916 }
2917 }
2918
2919 void Deoptimization::print_statistics_on(outputStream* st) {
2920 // print_statistics_on("Total", 0, st);
2921 print_statistics_on("Tier1", 1, st);
2922 print_statistics_on("Tier2", 2, st);
2923 print_statistics_on("Tier3", 3, st);
2924 print_statistics_on("Tier4", 4, st);
2925
2926 print_statistics_on("SC Tier1", 5, st);
2927 print_statistics_on("SC Tier2", 6, st);
2928 print_statistics_on("SC Tier4", 8, st);
2929 print_statistics_on("SC Tier5 (preloaded)", 9, st);
2930 }
2931
2932 #else // COMPILER2_OR_JVMCI
2933
2934
2935 // Stubs for C1 only system.
2936 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2937 return false;
2938 }
2939
2940 const char* Deoptimization::trap_reason_name(int reason) {
2941 return "unknown";
2942 }
2943
2944 jint Deoptimization::total_deoptimization_count() {
2945 return 0;
2946 }
2947
2948 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2949 return 0;
2950 }
2951
2952 void Deoptimization::print_statistics() {
2953 // no output
2954 }
2955
2956 void
2957 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2958 // no update
2959 }
2960
2961 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2962 return 0;
2963 }
2964
2965 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2966 Bytecodes::Code bc) {
2967 // no update
2968 }
2969
2970 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2971 int trap_state) {
2972 jio_snprintf(buf, buflen, "#%d", trap_state);
2973 return buf;
2974 }
2975
2976 #endif // COMPILER2_OR_JVMCI
2977
2978 #define DO_COUNTERS(macro) \
2979 macro(Deoptimization, fetch_unroll_info) \
2980 macro(Deoptimization, unpack_frames) \
2981 macro(Deoptimization, uncommon_trap_inner) \
2982 macro(Deoptimization, uncommon_trap)
2983
2984 #define INIT_COUNTER(sub, name) \
2985 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \
2986 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count");
2987
2988 void Deoptimization::init_counters() {
2989 if (ProfileRuntimeCalls && UsePerfData) {
2990 EXCEPTION_MARK;
2991
2992 DO_COUNTERS(INIT_COUNTER)
2993
2994 if (HAS_PENDING_EXCEPTION) {
2995 vm_exit_during_initialization("jvm_perf_init failed unexpectedly");
2996 }
2997 }
2998 }
2999 #undef INIT_COUNTER
3000
3001 #define PRINT_COUNTER(sub, name) { \
3002 jlong count = _perf_##sub##_##name##_count->get_value(); \
3003 if (count > 0) { \
3004 st->print_cr(" %-50s = %4ldms (elapsed) %4ldms (thread) (%5ld events)", #sub "::" #name, \
3005 _perf_##sub##_##name##_timer->elapsed_counter_value_ms(), \
3006 _perf_##sub##_##name##_timer->thread_counter_value_ms(), \
3007 count); \
3008 }}
3009
3010 void Deoptimization::print_counters_on(outputStream* st) {
3011 if (ProfileRuntimeCalls && UsePerfData) {
3012 DO_COUNTERS(PRINT_COUNTER)
3013 } else {
3014 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData"));
3015 }
3016 }
3017
3018 #undef PRINT_COUNTER
3019 #undef DO_COUNTERS