1 /* 2 * Copyright (c) 1997, 2025, 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 "classfile/javaClasses.inline.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmClasses.hpp" 29 #include "code/codeCache.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "code/nmethod.hpp" 32 #include "code/pcDesc.hpp" 33 #include "code/scopeDesc.hpp" 34 #include "compiler/compilationPolicy.hpp" 35 #include "compiler/compilerDefinitions.inline.hpp" 36 #include "gc/shared/collectedHeap.hpp" 37 #include "gc/shared/memAllocator.hpp" 38 #include "interpreter/bytecode.inline.hpp" 39 #include "interpreter/bytecodeStream.hpp" 40 #include "interpreter/interpreter.hpp" 41 #include "interpreter/oopMapCache.hpp" 42 #include "jvm.h" 43 #include "logging/log.hpp" 44 #include "logging/logLevel.hpp" 45 #include "logging/logMessage.hpp" 46 #include "logging/logStream.hpp" 47 #include "memory/allocation.inline.hpp" 48 #include "memory/oopFactory.hpp" 49 #include "memory/resourceArea.hpp" 50 #include "memory/universe.hpp" 51 #include "oops/constantPool.hpp" 52 #include "oops/fieldStreams.inline.hpp" 53 #include "oops/method.hpp" 54 #include "oops/objArrayKlass.hpp" 55 #include "oops/objArrayOop.inline.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "oops/typeArrayOop.inline.hpp" 58 #include "oops/verifyOopClosure.hpp" 59 #include "prims/jvmtiDeferredUpdates.hpp" 60 #include "prims/jvmtiExport.hpp" 61 #include "prims/jvmtiThreadState.hpp" 62 #include "prims/methodHandles.hpp" 63 #include "prims/vectorSupport.hpp" 64 #include "runtime/atomicAccess.hpp" 65 #include "runtime/basicLock.inline.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.inline.hpp" 71 #include "runtime/frame.inline.hpp" 72 #include "runtime/handles.inline.hpp" 73 #include "runtime/interfaceSupport.inline.hpp" 74 #include "runtime/javaThread.hpp" 75 #include "runtime/jniHandles.inline.hpp" 76 #include "runtime/keepStackGCProcessed.hpp" 77 #include "runtime/lightweightSynchronizer.hpp" 78 #include "runtime/lockStack.inline.hpp" 79 #include "runtime/objectMonitor.inline.hpp" 80 #include "runtime/osThread.hpp" 81 #include "runtime/perfData.inline.hpp" 82 #include "runtime/safepointVerifiers.hpp" 83 #include "runtime/sharedRuntime.hpp" 84 #include "runtime/signature.hpp" 85 #include "runtime/stackFrameStream.inline.hpp" 86 #include "runtime/stackValue.hpp" 87 #include "runtime/stackWatermarkSet.hpp" 88 #include "runtime/stubRoutines.hpp" 89 #include "runtime/synchronizer.inline.hpp" 90 #include "runtime/threadSMR.hpp" 91 #include "runtime/threadWXSetters.inline.hpp" 92 #include "runtime/vframe.hpp" 93 #include "runtime/vframe_hp.hpp" 94 #include "runtime/vframeArray.hpp" 95 #include "runtime/vmOperations.hpp" 96 #include "services/management.hpp" 97 #include "utilities/checkedCast.hpp" 98 #include "utilities/events.hpp" 99 #include "utilities/growableArray.hpp" 100 #include "utilities/macros.hpp" 101 #include "utilities/preserveException.hpp" 102 #include "utilities/xmlstream.hpp" 103 #if INCLUDE_JFR 104 #include "jfr/jfr.inline.hpp" 105 #include "jfr/jfrEvents.hpp" 106 #include "jfr/metadata/jfrSerializer.hpp" 107 #endif 108 109 uint64_t DeoptimizationScope::_committed_deopt_gen = 0; 110 uint64_t DeoptimizationScope::_active_deopt_gen = 1; 111 bool DeoptimizationScope::_committing_in_progress = false; 112 113 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) { 114 DEBUG_ONLY(_deopted = false;) 115 116 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag); 117 // If there is nothing to deopt _required_gen is the same as comitted. 118 _required_gen = DeoptimizationScope::_committed_deopt_gen; 119 } 120 121 DeoptimizationScope::~DeoptimizationScope() { 122 assert(_deopted, "Deopt not executed"); 123 } 124 125 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) { 126 if (!nm->can_be_deoptimized()) { 127 return; 128 } 129 130 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 131 132 // If it's already marked but we still need it to be deopted. 133 if (nm->is_marked_for_deoptimization()) { 134 dependent(nm); 135 return; 136 } 137 138 nmethod::DeoptimizationStatus status = 139 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate; 140 AtomicAccess::store(&nm->_deoptimization_status, status); 141 142 // Make sure active is not committed 143 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be"); 144 assert(nm->_deoptimization_generation == 0, "Is already marked"); 145 146 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen; 147 _required_gen = DeoptimizationScope::_active_deopt_gen; 148 } 149 150 void DeoptimizationScope::dependent(nmethod* nm) { 151 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 152 153 // A method marked by someone else may have a _required_gen lower than what we marked with. 154 // Therefore only store it if it's higher than _required_gen. 155 if (_required_gen < nm->_deoptimization_generation) { 156 _required_gen = nm->_deoptimization_generation; 157 } 158 } 159 160 void DeoptimizationScope::deoptimize_marked() { 161 assert(!_deopted, "Already deopted"); 162 163 // We are not alive yet. 164 if (!Universe::is_fully_initialized()) { 165 DEBUG_ONLY(_deopted = true;) 166 return; 167 } 168 169 // Safepoints are a special case, handled here. 170 if (SafepointSynchronize::is_at_safepoint()) { 171 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen; 172 DeoptimizationScope::_active_deopt_gen++; 173 Deoptimization::deoptimize_all_marked(); 174 DEBUG_ONLY(_deopted = true;) 175 return; 176 } 177 178 uint64_t comitting = 0; 179 bool wait = false; 180 while (true) { 181 { 182 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 183 184 // First we check if we or someone else already deopted the gen we want. 185 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) { 186 DEBUG_ONLY(_deopted = true;) 187 return; 188 } 189 if (!_committing_in_progress) { 190 // The version we are about to commit. 191 comitting = DeoptimizationScope::_active_deopt_gen; 192 // Make sure new marks use a higher gen. 193 DeoptimizationScope::_active_deopt_gen++; 194 _committing_in_progress = true; 195 wait = false; 196 } else { 197 // Another thread is handshaking and committing a gen. 198 wait = true; 199 } 200 } 201 if (wait) { 202 // Wait and let the concurrent handshake be performed. 203 ThreadBlockInVM tbivm(JavaThread::current()); 204 os::naked_yield(); 205 } else { 206 // Performs the handshake. 207 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred. 208 DEBUG_ONLY(_deopted = true;) 209 { 210 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 211 212 // Make sure that committed doesn't go backwards. 213 // Should only happen if we did a deopt during a safepoint above. 214 if (DeoptimizationScope::_committed_deopt_gen < comitting) { 215 DeoptimizationScope::_committed_deopt_gen = comitting; 216 } 217 _committing_in_progress = false; 218 219 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be"); 220 221 return; 222 } 223 } 224 } 225 } 226 227 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 228 int caller_adjustment, 229 int caller_actual_parameters, 230 int number_of_frames, 231 intptr_t* frame_sizes, 232 address* frame_pcs, 233 BasicType return_type, 234 int exec_mode) { 235 _size_of_deoptimized_frame = size_of_deoptimized_frame; 236 _caller_adjustment = caller_adjustment; 237 _caller_actual_parameters = caller_actual_parameters; 238 _number_of_frames = number_of_frames; 239 _frame_sizes = frame_sizes; 240 _frame_pcs = frame_pcs; 241 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 242 _return_type = return_type; 243 _initial_info = 0; 244 // PD (x86 only) 245 _counter_temp = 0; 246 _unpack_kind = exec_mode; 247 _sender_sp_temp = 0; 248 249 _total_frame_sizes = size_of_frames(); 250 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 251 } 252 253 Deoptimization::UnrollBlock::~UnrollBlock() { 254 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 255 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 256 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 257 } 258 259 int Deoptimization::UnrollBlock::size_of_frames() const { 260 // Account first for the adjustment of the initial frame 261 intptr_t result = _caller_adjustment; 262 for (int index = 0; index < number_of_frames(); index++) { 263 result += frame_sizes()[index]; 264 } 265 return checked_cast<int>(result); 266 } 267 268 void Deoptimization::UnrollBlock::print() { 269 ResourceMark rm; 270 stringStream st; 271 st.print_cr("UnrollBlock"); 272 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 273 st.print( " frame_sizes: "); 274 for (int index = 0; index < number_of_frames(); index++) { 275 st.print("%zd ", frame_sizes()[index]); 276 } 277 st.cr(); 278 tty->print_raw(st.freeze()); 279 } 280 281 // In order to make fetch_unroll_info work properly with escape 282 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 283 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 284 // which is called from the method fetch_unroll_info_helper below. 285 JRT_BLOCK_ENTRY_PROF(Deoptimization::UnrollBlock*, Deoptimization, fetch_unroll_info, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 286 // fetch_unroll_info() is called at the beginning of the deoptimization 287 // handler. Note this fact before we start generating temporary frames 288 // that can confuse an asynchronous stack walker. This counter is 289 // decremented at the end of unpack_frames(). 290 current->inc_in_deopt_handler(); 291 292 if (exec_mode == Unpack_exception) { 293 // When we get here, a callee has thrown an exception into a deoptimized 294 // frame. That throw might have deferred stack watermark checking until 295 // after unwinding. So we deal with such deferred requests here. 296 StackWatermarkSet::after_unwind(current); 297 } 298 299 return fetch_unroll_info_helper(current, exec_mode); 300 JRT_END 301 302 #if COMPILER2_OR_JVMCI 303 // print information about reallocated objects 304 static void print_objects(JavaThread* deoptee_thread, 305 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 306 ResourceMark rm; 307 stringStream st; // change to logStream with logging 308 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 309 fieldDescriptor fd; 310 311 for (int i = 0; i < objects->length(); i++) { 312 ObjectValue* sv = (ObjectValue*) objects->at(i); 313 Handle obj = sv->value(); 314 315 if (obj.is_null()) { 316 st.print_cr(" nullptr"); 317 continue; 318 } 319 320 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 321 322 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 323 k->print_value_on(&st); 324 st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize); 325 326 if (Verbose && k != nullptr) { 327 k->oop_print_on(obj(), &st); 328 } 329 } 330 tty->print_raw(st.freeze()); 331 } 332 333 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method, 334 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 335 bool& deoptimized_objects) { 336 bool realloc_failures = false; 337 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames"); 338 339 JavaThread* deoptee_thread = chunk->at(0)->thread(); 340 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 341 "a frame can only be deoptimized by the owner thread"); 342 343 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map); 344 345 // The flag return_oop() indicates call sites which return oop 346 // in compiled code. Such sites include java method calls, 347 // runtime calls (for example, used to allocate new objects/arrays 348 // on slow code path) and any other calls generated in compiled code. 349 // It is not guaranteed that we can get such information here only 350 // by analyzing bytecode in deoptimized frames. This is why this flag 351 // is set during method compilation (see Compile::Process_OopMap_Node()). 352 // If the previous frame was popped or if we are dispatching an exception, 353 // we don't have an oop result. 354 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 355 Handle return_value; 356 if (save_oop_result) { 357 // Reallocation may trigger GC. If deoptimization happened on return from 358 // call which returns oop we need to save it since it is not in oopmap. 359 oop result = deoptee.saved_oop_result(&map); 360 assert(oopDesc::is_oop_or_null(result), "must be oop"); 361 return_value = Handle(thread, result); 362 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 363 if (TraceDeoptimization) { 364 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 365 tty->cr(); 366 } 367 } 368 if (objects != nullptr) { 369 if (exec_mode == Deoptimization::Unpack_none) { 370 assert(thread->thread_state() == _thread_in_vm, "assumption"); 371 JavaThread* THREAD = thread; // For exception macros. 372 // Clear pending OOM if reallocation fails and return true indicating allocation failure 373 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 374 deoptimized_objects = true; 375 } else { 376 JavaThread* current = thread; // For JRT_BLOCK 377 JRT_BLOCK 378 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 379 JRT_END 380 } 381 guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod"); 382 bool is_jvmci = compiled_method->is_compiled_by_jvmci(); 383 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci); 384 if (TraceDeoptimization) { 385 print_objects(deoptee_thread, objects, realloc_failures); 386 } 387 } 388 if (save_oop_result) { 389 // Restore result. 390 deoptee.set_saved_oop_result(&map, return_value()); 391 } 392 return realloc_failures; 393 } 394 395 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 396 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 397 JavaThread* deoptee_thread = chunk->at(0)->thread(); 398 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 399 assert(thread == Thread::current(), "should be"); 400 HandleMark hm(thread); 401 #ifndef PRODUCT 402 bool first = true; 403 #endif // !PRODUCT 404 // Start locking from outermost/oldest frame 405 for (int i = (chunk->length() - 1); i >= 0; i--) { 406 compiledVFrame* cvf = chunk->at(i); 407 assert (cvf->scope() != nullptr,"expect only compiled java frames"); 408 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 409 if (monitors->is_nonempty()) { 410 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 411 exec_mode, realloc_failures); 412 deoptimized_objects = deoptimized_objects || relocked; 413 #ifndef PRODUCT 414 if (PrintDeoptimizationDetails) { 415 ResourceMark rm; 416 stringStream st; 417 for (int j = 0; j < monitors->length(); j++) { 418 MonitorInfo* mi = monitors->at(j); 419 if (mi->eliminated()) { 420 if (first) { 421 first = false; 422 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 423 } 424 if (exec_mode == Deoptimization::Unpack_none) { 425 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 426 if (monitor != nullptr && monitor->object() == mi->owner()) { 427 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 428 continue; 429 } 430 } 431 if (mi->owner_is_scalar_replaced()) { 432 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 433 st.print_cr(" failed reallocation for klass %s", k->external_name()); 434 } else { 435 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 436 } 437 } 438 } 439 tty->print_raw(st.freeze()); 440 } 441 #endif // !PRODUCT 442 } 443 } 444 } 445 446 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 447 // The given vframes cover one physical frame. 448 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 449 bool& realloc_failures) { 450 frame deoptee = chunk->at(0)->fr(); 451 JavaThread* deoptee_thread = chunk->at(0)->thread(); 452 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 453 RegisterMap map(chunk->at(0)->register_map()); 454 bool deoptimized_objects = false; 455 456 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 457 458 // Reallocate the non-escaping objects and restore their fields. 459 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 460 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 461 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects); 462 } 463 464 // MonitorInfo structures used in eliminate_locks are not GC safe. 465 NoSafepointVerifier no_safepoint; 466 467 // Now relock objects if synchronization on them was eliminated. 468 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 469 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 470 } 471 return deoptimized_objects; 472 } 473 #endif // COMPILER2_OR_JVMCI 474 475 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 476 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 477 JFR_ONLY(Jfr::check_and_process_sample_request(current);) 478 // When we get here we are about to unwind the deoptee frame. In order to 479 // catch not yet safe to use frames, the following stack watermark barrier 480 // poll will make such frames safe to use. 481 StackWatermarkSet::before_unwind(current); 482 483 // Note: there is a safepoint safety issue here. No matter whether we enter 484 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 485 // the vframeArray is created. 486 // 487 488 // Allocate our special deoptimization ResourceMark 489 DeoptResourceMark* dmark = new DeoptResourceMark(current); 490 assert(current->deopt_mark() == nullptr, "Pending deopt!"); 491 current->set_deopt_mark(dmark); 492 493 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 494 RegisterMap map(current, 495 RegisterMap::UpdateMap::include, 496 RegisterMap::ProcessFrames::include, 497 RegisterMap::WalkContinuation::skip); 498 RegisterMap dummy_map(current, 499 RegisterMap::UpdateMap::skip, 500 RegisterMap::ProcessFrames::include, 501 RegisterMap::WalkContinuation::skip); 502 // Now get the deoptee with a valid map 503 frame deoptee = stub_frame.sender(&map); 504 // Set the deoptee nmethod 505 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!"); 506 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 507 current->set_deopt_compiled_method(nm); 508 509 if (VerifyStack) { 510 current->validate_frame_layout(); 511 } 512 513 // Create a growable array of VFrames where each VFrame represents an inlined 514 // Java frame. This storage is allocated with the usual system arena. 515 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 516 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 517 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 518 while (!vf->is_top()) { 519 assert(vf->is_compiled_frame(), "Wrong frame type"); 520 chunk->push(compiledVFrame::cast(vf)); 521 vf = vf->sender(); 522 } 523 assert(vf->is_compiled_frame(), "Wrong frame type"); 524 chunk->push(compiledVFrame::cast(vf)); 525 526 bool realloc_failures = false; 527 528 #if COMPILER2_OR_JVMCI 529 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 530 531 // Reallocate the non-escaping objects and restore their fields. Then 532 // relock objects if synchronization on them was eliminated. 533 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 534 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 535 bool unused; 536 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused); 537 } 538 #endif // COMPILER2_OR_JVMCI 539 540 // Ensure that no safepoint is taken after pointers have been stored 541 // in fields of rematerialized objects. If a safepoint occurs from here on 542 // out the java state residing in the vframeArray will be missed. 543 // Locks may be rebaised in a safepoint. 544 NoSafepointVerifier no_safepoint; 545 546 #if COMPILER2_OR_JVMCI 547 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 548 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 549 bool unused = false; 550 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 551 } 552 #endif // COMPILER2_OR_JVMCI 553 554 ScopeDesc* trap_scope = chunk->at(0)->scope(); 555 Handle exceptionObject; 556 if (trap_scope->rethrow_exception()) { 557 #ifndef PRODUCT 558 if (PrintDeoptimizationDetails) { 559 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()); 560 } 561 #endif // !PRODUCT 562 563 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 564 guarantee(expressions != nullptr && expressions->length() == 1, "should have only exception on stack"); 565 guarantee(exec_mode != Unpack_exception, "rethrow_exception set with Unpack_exception"); 566 ScopeValue* topOfStack = expressions->top(); 567 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 568 guarantee(exceptionObject() != nullptr, "exception oop can not be null"); 569 } 570 571 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 572 #if COMPILER2_OR_JVMCI 573 if (realloc_failures) { 574 // This destroys all ScopedValue bindings. 575 current->clear_scopedValueBindings(); 576 pop_frames_failed_reallocs(current, array); 577 } 578 #endif 579 580 assert(current->vframe_array_head() == nullptr, "Pending deopt!"); 581 current->set_vframe_array_head(array); 582 583 // Now that the vframeArray has been created if we have any deferred local writes 584 // added by jvmti then we can free up that structure as the data is now in the 585 // vframeArray 586 587 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 588 589 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 590 CodeBlob* cb = stub_frame.cb(); 591 // Verify we have the right vframeArray 592 assert(cb->frame_size() >= 0, "Unexpected frame size"); 593 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 594 assert(unpack_sp == deoptee.unextended_sp(), "must be"); 595 596 #ifdef ASSERT 597 assert(cb->is_deoptimization_stub() || 598 cb->is_uncommon_trap_stub() || 599 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 600 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 601 "unexpected code blob: %s", cb->name()); 602 #endif 603 604 // This is a guarantee instead of an assert because if vframe doesn't match 605 // we will unpack the wrong deoptimized frame and wind up in strange places 606 // where it will be very difficult to figure out what went wrong. Better 607 // to die an early death here than some very obscure death later when the 608 // trail is cold. 609 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 610 611 int number_of_frames = array->frames(); 612 613 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 614 // virtual activation, which is the reverse of the elements in the vframes array. 615 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 616 // +1 because we always have an interpreter return address for the final slot. 617 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 618 int popframe_extra_args = 0; 619 // Create an interpreter return address for the stub to use as its return 620 // address so the skeletal frames are perfectly walkable 621 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 622 623 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 624 // activation be put back on the expression stack of the caller for reexecution 625 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 626 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 627 } 628 629 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 630 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 631 // than simply use array->sender.pc(). This requires us to walk the current set of frames 632 // 633 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 634 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 635 636 // It's possible that the number of parameters at the call site is 637 // different than number of arguments in the callee when method 638 // handles are used. If the caller is interpreted get the real 639 // value so that the proper amount of space can be added to it's 640 // frame. 641 bool caller_was_method_handle = false; 642 if (deopt_sender.is_interpreted_frame()) { 643 methodHandle method(current, deopt_sender.interpreter_frame_method()); 644 Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci()); 645 if (cur.has_member_arg()) { 646 // This should cover all real-world cases. One exception is a pathological chain of 647 // MH.linkToXXX() linker calls, which only trusted code could do anyway. To handle that case, we 648 // would need to get the size from the resolved method entry. Another exception would 649 // be an invokedynamic with an adapter that is really a MethodHandle linker. 650 caller_was_method_handle = true; 651 } 652 } 653 654 // 655 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 656 // frame_sizes/frame_pcs[1] next oldest frame (int) 657 // frame_sizes/frame_pcs[n] youngest frame (int) 658 // 659 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 660 // owns the space for the return address to it's caller). Confusing ain't it. 661 // 662 // The vframe array can address vframes with indices running from 663 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 664 // When we create the skeletal frames we need the oldest frame to be in the zero slot 665 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 666 // so things look a little strange in this loop. 667 // 668 int callee_parameters = 0; 669 int callee_locals = 0; 670 for (int index = 0; index < array->frames(); index++ ) { 671 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 672 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 673 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 674 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 675 callee_locals, 676 index == 0, 677 popframe_extra_args); 678 // This pc doesn't have to be perfect just good enough to identify the frame 679 // as interpreted so the skeleton frame will be walkable 680 // The correct pc will be set when the skeleton frame is completely filled out 681 // The final pc we store in the loop is wrong and will be overwritten below 682 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 683 684 callee_parameters = array->element(index)->method()->size_of_parameters(); 685 callee_locals = array->element(index)->method()->max_locals(); 686 popframe_extra_args = 0; 687 } 688 689 // Compute whether the root vframe returns a float or double value. 690 BasicType return_type; 691 { 692 methodHandle method(current, array->element(0)->method()); 693 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 694 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 695 } 696 697 // Compute information for handling adapters and adjusting the frame size of the caller. 698 int caller_adjustment = 0; 699 700 // Compute the amount the oldest interpreter frame will have to adjust 701 // its caller's stack by. If the caller is a compiled frame then 702 // we pretend that the callee has no parameters so that the 703 // extension counts for the full amount of locals and not just 704 // locals-parms. This is because without a c2i adapter the parm 705 // area as created by the compiled frame will not be usable by 706 // the interpreter. (Depending on the calling convention there 707 // may not even be enough space). 708 709 // QQQ I'd rather see this pushed down into last_frame_adjust 710 // and have it take the sender (aka caller). 711 712 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 713 caller_adjustment = last_frame_adjust(0, callee_locals); 714 } else if (callee_locals > callee_parameters) { 715 // The caller frame may need extending to accommodate 716 // non-parameter locals of the first unpacked interpreted frame. 717 // Compute that adjustment. 718 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 719 } 720 721 // If the sender is deoptimized the we must retrieve the address of the handler 722 // since the frame will "magically" show the original pc before the deopt 723 // and we'd undo the deopt. 724 725 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 726 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 727 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 728 } 729 730 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc"); 731 732 #if INCLUDE_JVMCI 733 if (exceptionObject() != nullptr) { 734 current->set_exception_oop(exceptionObject()); 735 exec_mode = Unpack_exception; 736 assert(array->element(0)->rethrow_exception(), "must be"); 737 } 738 #endif 739 740 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 741 assert(current->has_pending_exception(), "should have thrown OOME"); 742 current->set_exception_oop(current->pending_exception()); 743 current->clear_pending_exception(); 744 exec_mode = Unpack_exception; 745 } 746 747 #if INCLUDE_JVMCI 748 if (current->frames_to_pop_failed_realloc() > 0) { 749 current->set_pending_monitorenter(false); 750 } 751 #endif 752 753 int caller_actual_parameters = -1; // value not used except for interpreted frames, see below 754 if (deopt_sender.is_interpreted_frame()) { 755 caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0); 756 } 757 758 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 759 caller_adjustment * BytesPerWord, 760 caller_actual_parameters, 761 number_of_frames, 762 frame_sizes, 763 frame_pcs, 764 return_type, 765 exec_mode); 766 // On some platforms, we need a way to pass some platform dependent 767 // information to the unpacking code so the skeletal frames come out 768 // correct (initial fp value, unextended sp, ...) 769 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 770 771 if (array->frames() > 1) { 772 if (VerifyStack && TraceDeoptimization) { 773 tty->print_cr("Deoptimizing method containing inlining"); 774 } 775 } 776 777 array->set_unroll_block(info); 778 return info; 779 } 780 781 // Called to cleanup deoptimization data structures in normal case 782 // after unpacking to stack and when stack overflow error occurs 783 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 784 vframeArray *array) { 785 786 // Get array if coming from exception 787 if (array == nullptr) { 788 array = thread->vframe_array_head(); 789 } 790 thread->set_vframe_array_head(nullptr); 791 792 // Free the previous UnrollBlock 793 vframeArray* old_array = thread->vframe_array_last(); 794 thread->set_vframe_array_last(array); 795 796 if (old_array != nullptr) { 797 UnrollBlock* old_info = old_array->unroll_block(); 798 old_array->set_unroll_block(nullptr); 799 delete old_info; 800 delete old_array; 801 } 802 803 // Deallocate any resource creating in this routine and any ResourceObjs allocated 804 // inside the vframeArray (StackValueCollections) 805 806 delete thread->deopt_mark(); 807 thread->set_deopt_mark(nullptr); 808 thread->set_deopt_compiled_method(nullptr); 809 810 811 if (JvmtiExport::can_pop_frame()) { 812 // Regardless of whether we entered this routine with the pending 813 // popframe condition bit set, we should always clear it now 814 thread->clear_popframe_condition(); 815 } 816 817 // unpack_frames() is called at the end of the deoptimization handler 818 // and (in C2) at the end of the uncommon trap handler. Note this fact 819 // so that an asynchronous stack walker can work again. This counter is 820 // incremented at the beginning of fetch_unroll_info() and (in C2) at 821 // the beginning of uncommon_trap(). 822 thread->dec_in_deopt_handler(); 823 } 824 825 // Moved from cpu directories because none of the cpus has callee save values. 826 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 827 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 828 829 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 830 // the days we had adapter frames. When we deoptimize a situation where a 831 // compiled caller calls a compiled caller will have registers it expects 832 // to survive the call to the callee. If we deoptimize the callee the only 833 // way we can restore these registers is to have the oldest interpreter 834 // frame that we create restore these values. That is what this routine 835 // will accomplish. 836 837 // At the moment we have modified c2 to not have any callee save registers 838 // so this problem does not exist and this routine is just a place holder. 839 840 assert(f->is_interpreted_frame(), "must be interpreted"); 841 } 842 843 #ifndef PRODUCT 844 #ifdef ASSERT 845 // Return true if the execution after the provided bytecode continues at the 846 // next bytecode in the code. This is not the case for gotos, returns, and 847 // throws. 848 static bool falls_through(Bytecodes::Code bc) { 849 switch (bc) { 850 case Bytecodes::_goto: 851 case Bytecodes::_goto_w: 852 case Bytecodes::_athrow: 853 case Bytecodes::_areturn: 854 case Bytecodes::_dreturn: 855 case Bytecodes::_freturn: 856 case Bytecodes::_ireturn: 857 case Bytecodes::_lreturn: 858 case Bytecodes::_jsr: 859 case Bytecodes::_ret: 860 case Bytecodes::_return: 861 case Bytecodes::_lookupswitch: 862 case Bytecodes::_tableswitch: 863 return false; 864 default: 865 return true; 866 } 867 } 868 #endif 869 #endif 870 871 // Return BasicType of value being returned 872 JRT_LEAF_PROF_NO_THREAD(BasicType, Deoptimization, unpack_frames, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 873 assert(thread == JavaThread::current(), "pre-condition"); 874 875 // We are already active in the special DeoptResourceMark any ResourceObj's we 876 // allocate will be freed at the end of the routine. 877 878 // JRT_LEAF methods don't normally allocate handles and there is a 879 // NoHandleMark to enforce that. It is actually safe to use Handles 880 // in a JRT_LEAF method, and sometimes desirable, but to do so we 881 // must use ResetNoHandleMark to bypass the NoHandleMark, and 882 // then use a HandleMark to ensure any Handles we do create are 883 // cleaned up in this scope. 884 ResetNoHandleMark rnhm; 885 HandleMark hm(thread); 886 887 frame stub_frame = thread->last_frame(); 888 889 Continuation::notify_deopt(thread, stub_frame.sp()); 890 891 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 892 // must point to the vframeArray for the unpack frame. 893 vframeArray* array = thread->vframe_array_head(); 894 UnrollBlock* info = array->unroll_block(); 895 896 // We set the last_Java frame. But the stack isn't really parsable here. So we 897 // clear it to make sure JFR understands not to try and walk stacks from events 898 // in here. 899 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 900 thread->frame_anchor()->set_last_Java_sp(nullptr); 901 902 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 903 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 904 905 thread->frame_anchor()->set_last_Java_sp(sp); 906 907 BasicType bt = info->return_type(); 908 909 // If we have an exception pending, claim that the return type is an oop 910 // so the deopt_blob does not overwrite the exception_oop. 911 912 if (exec_mode == Unpack_exception) 913 bt = T_OBJECT; 914 915 // Cleanup thread deopt data 916 cleanup_deopt_info(thread, array); 917 918 #ifndef PRODUCT 919 if (VerifyStack) { 920 ResourceMark res_mark; 921 // Clear pending exception to not break verification code (restored afterwards) 922 PreserveExceptionMark pm(thread); 923 924 thread->validate_frame_layout(); 925 926 // Verify that the just-unpacked frames match the interpreter's 927 // notions of expression stack and locals 928 vframeArray* cur_array = thread->vframe_array_last(); 929 RegisterMap rm(thread, 930 RegisterMap::UpdateMap::skip, 931 RegisterMap::ProcessFrames::include, 932 RegisterMap::WalkContinuation::skip); 933 rm.set_include_argument_oops(false); 934 int callee_size_of_parameters = 0; 935 for (int frame_idx = 0; frame_idx < cur_array->frames(); frame_idx++) { 936 bool is_top_frame = (frame_idx == 0); 937 vframeArrayElement* el = cur_array->element(frame_idx); 938 frame* iframe = el->iframe(); 939 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 940 methodHandle mh(thread, iframe->interpreter_frame_method()); 941 bool reexecute = el->should_reexecute(); 942 943 int cur_invoke_parameter_size = 0; 944 int top_frame_expression_stack_adjustment = 0; 945 int max_bci = mh->code_size(); 946 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 947 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 948 Bytecodes::Code cur_code = str.next(); 949 950 if (!reexecute && !Bytecodes::is_invoke(cur_code)) { 951 // We can only compute OopMaps for the before state, so we need to roll forward 952 // to the next bytecode. 953 assert(is_top_frame, "must be"); 954 assert(falls_through(cur_code), "must be"); 955 assert(cur_code != Bytecodes::_illegal, "illegal bytecode"); 956 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 957 958 // Need to subtract off the size of the result type of 959 // the bytecode because this is not described in the 960 // debug info but returned to the interpreter in the TOS 961 // caching register 962 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 963 if (bytecode_result_type != T_ILLEGAL) { 964 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 965 } 966 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 967 968 cur_code = str.next(); 969 // Reflect the fact that we have rolled forward and now need 970 // top_frame_expression_stack_adjustment 971 reexecute = true; 972 } 973 974 assert(cur_code != Bytecodes::_illegal, "illegal bytecode"); 975 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 976 977 // Get the oop map for this bci 978 InterpreterOopMap mask; 979 OopMapCache::compute_one_oop_map(mh, str.bci(), &mask); 980 // Check to see if we can grab the number of outgoing arguments 981 // at an uncommon trap for an invoke (where the compiler 982 // generates debug info before the invoke has executed) 983 if (Bytecodes::is_invoke(cur_code)) { 984 Bytecode_invoke invoke(mh, str.bci()); 985 cur_invoke_parameter_size = invoke.size_of_parameters(); 986 if (!is_top_frame && invoke.has_member_arg()) { 987 callee_size_of_parameters++; 988 } 989 } 990 991 // Verify stack depth and oops in frame 992 auto match = [&]() { 993 int iframe_expr_ssize = iframe->interpreter_frame_expression_stack_size(); 994 #if INCLUDE_JVMCI 995 if (is_top_frame && el->rethrow_exception()) { 996 return iframe_expr_ssize == 1; 997 } 998 #endif 999 // This should only be needed for C1 1000 if (is_top_frame && exec_mode == Unpack_exception && iframe_expr_ssize == 0) { 1001 return true; 1002 } 1003 if (reexecute) { 1004 int expr_ssize_before = iframe_expr_ssize + top_frame_expression_stack_adjustment; 1005 int oopmap_expr_invoke_ssize = mask.expression_stack_size() + cur_invoke_parameter_size; 1006 return expr_ssize_before == oopmap_expr_invoke_ssize; 1007 } else { 1008 int oopmap_expr_callee_ssize = mask.expression_stack_size() + callee_size_of_parameters; 1009 return iframe_expr_ssize == oopmap_expr_callee_ssize; 1010 } 1011 }; 1012 if (!match()) { 1013 // Print out some information that will help us debug the problem 1014 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1015 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", frame_idx, cur_array->frames() - 1); 1016 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 1017 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1018 iframe->interpreter_frame_expression_stack_size()); 1019 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1020 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1021 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1022 tty->print_cr(" exec_mode = %d", exec_mode); 1023 tty->print_cr(" original should_reexecute = %s", el->should_reexecute() ? "true" : "false"); 1024 tty->print_cr(" reexecute = %s%s", reexecute ? "true" : "false", 1025 (reexecute != el->should_reexecute()) ? " (changed)" : ""); 1026 #if INCLUDE_JVMCI 1027 tty->print_cr(" rethrow_exception = %s", el->rethrow_exception() ? "true" : "false"); 1028 #endif 1029 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1030 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 1031 tty->print_cr(" Interpreted frames:"); 1032 for (int k = 0; k < cur_array->frames(); k++) { 1033 vframeArrayElement* el = cur_array->element(k); 1034 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1035 } 1036 cur_array->print_on_2(tty); 1037 guarantee(false, "wrong number of expression stack elements during deopt"); 1038 } 1039 VerifyOopClosure verify; 1040 iframe->oops_interpreted_do(&verify, &rm, false); 1041 callee_size_of_parameters = mh->size_of_parameters(); 1042 } 1043 } 1044 #endif // !PRODUCT 1045 1046 return bt; 1047 JRT_END 1048 1049 class DeoptimizeMarkedHandshakeClosure : public HandshakeClosure { 1050 public: 1051 DeoptimizeMarkedHandshakeClosure() : HandshakeClosure("Deoptimize") {} 1052 void do_thread(Thread* thread) { 1053 JavaThread* jt = JavaThread::cast(thread); 1054 jt->deoptimize_marked_methods(); 1055 } 1056 }; 1057 1058 void Deoptimization::deoptimize_all_marked() { 1059 ResourceMark rm; 1060 1061 // Make the dependent methods not entrant 1062 CodeCache::make_marked_nmethods_deoptimized(); 1063 1064 DeoptimizeMarkedHandshakeClosure deopt; 1065 if (SafepointSynchronize::is_at_safepoint()) { 1066 Threads::java_threads_do(&deopt); 1067 } else { 1068 Handshake::execute(&deopt); 1069 } 1070 } 1071 1072 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1073 = Deoptimization::Action_reinterpret; 1074 1075 #if INCLUDE_JVMCI 1076 template<typename CacheType> 1077 class BoxCacheBase : public CHeapObj<mtCompiler> { 1078 protected: 1079 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 1080 ResourceMark rm(thread); 1081 char* klass_name_str = klass_name->as_C_string(); 1082 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle()); 1083 guarantee(ik != nullptr, "%s must be loaded", klass_name_str); 1084 if (!ik->is_in_error_state()) { 1085 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 1086 CacheType::compute_offsets(ik); 1087 } 1088 return ik; 1089 } 1090 }; 1091 1092 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 1093 PrimitiveType _low; 1094 PrimitiveType _high; 1095 jobject _cache; 1096 protected: 1097 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 1098 BoxCache(Thread* thread) { 1099 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 1100 if (ik->is_in_error_state()) { 1101 _low = 1; 1102 _high = 0; 1103 _cache = nullptr; 1104 } else { 1105 objArrayOop cache = CacheType::cache(ik); 1106 assert(cache->length() > 0, "Empty cache"); 1107 _low = BoxType::value(cache->obj_at(0)); 1108 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1); 1109 _cache = JNIHandles::make_global(Handle(thread, cache)); 1110 } 1111 } 1112 ~BoxCache() { 1113 JNIHandles::destroy_global(_cache); 1114 } 1115 public: 1116 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1117 if (_singleton == nullptr) { 1118 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1119 if (!AtomicAccess::replace_if_null(&_singleton, s)) { 1120 delete s; 1121 } 1122 } 1123 return _singleton; 1124 } 1125 oop lookup(PrimitiveType value) { 1126 if (_low <= value && value <= _high) { 1127 int offset = checked_cast<int>(value - _low); 1128 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1129 } 1130 return nullptr; 1131 } 1132 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) { 1133 if (_cache == nullptr) { 1134 cache_init_error = true; 1135 return nullptr; 1136 } 1137 // Have to cast to avoid little/big-endian problems. 1138 if (sizeof(PrimitiveType) > sizeof(jint)) { 1139 jlong value = (jlong)raw_value; 1140 return lookup(value); 1141 } 1142 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1143 return lookup(value); 1144 } 1145 }; 1146 1147 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1148 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1149 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1150 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1151 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1152 1153 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr; 1154 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr; 1155 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr; 1156 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr; 1157 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr; 1158 1159 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1160 jobject _true_cache; 1161 jobject _false_cache; 1162 protected: 1163 static BooleanBoxCache *_singleton; 1164 BooleanBoxCache(Thread *thread) { 1165 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1166 if (ik->is_in_error_state()) { 1167 _true_cache = nullptr; 1168 _false_cache = nullptr; 1169 } else { 1170 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1171 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1172 } 1173 } 1174 ~BooleanBoxCache() { 1175 JNIHandles::destroy_global(_true_cache); 1176 JNIHandles::destroy_global(_false_cache); 1177 } 1178 public: 1179 static BooleanBoxCache* singleton(Thread* thread) { 1180 if (_singleton == nullptr) { 1181 BooleanBoxCache* s = new BooleanBoxCache(thread); 1182 if (!AtomicAccess::replace_if_null(&_singleton, s)) { 1183 delete s; 1184 } 1185 } 1186 return _singleton; 1187 } 1188 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) { 1189 if (_true_cache == nullptr) { 1190 cache_in_error = true; 1191 return nullptr; 1192 } 1193 // Have to cast to avoid little/big-endian problems. 1194 jboolean value = (jboolean)*((jint*)&raw_value); 1195 return lookup(value); 1196 } 1197 oop lookup(jboolean value) { 1198 if (value != 0) { 1199 return JNIHandles::resolve_non_null(_true_cache); 1200 } 1201 return JNIHandles::resolve_non_null(_false_cache); 1202 } 1203 }; 1204 1205 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr; 1206 1207 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) { 1208 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1209 BasicType box_type = vmClasses::box_klass_type(k); 1210 if (box_type != T_OBJECT) { 1211 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1212 switch(box_type) { 1213 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1214 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1215 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1216 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1217 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1218 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1219 default:; 1220 } 1221 } 1222 return nullptr; 1223 } 1224 #endif // INCLUDE_JVMCI 1225 1226 #if COMPILER2_OR_JVMCI 1227 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1228 Handle pending_exception(THREAD, thread->pending_exception()); 1229 const char* exception_file = thread->exception_file(); 1230 int exception_line = thread->exception_line(); 1231 thread->clear_pending_exception(); 1232 1233 bool failures = false; 1234 1235 for (int i = 0; i < objects->length(); i++) { 1236 assert(objects->at(i)->is_object(), "invalid debug information"); 1237 ObjectValue* sv = (ObjectValue*) objects->at(i); 1238 1239 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1240 oop obj = nullptr; 1241 1242 bool cache_init_error = false; 1243 if (k->is_instance_klass()) { 1244 #if INCLUDE_JVMCI 1245 nmethod* nm = fr->cb()->as_nmethod_or_null(); 1246 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1247 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1248 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD); 1249 if (obj != nullptr) { 1250 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1251 abv->set_cached(true); 1252 } else if (cache_init_error) { 1253 // Results in an OOME which is valid (as opposed to a class initialization error) 1254 // and is fine for the rare case a cache initialization failing. 1255 failures = true; 1256 } 1257 } 1258 #endif // INCLUDE_JVMCI 1259 1260 InstanceKlass* ik = InstanceKlass::cast(k); 1261 if (obj == nullptr && !cache_init_error) { 1262 InternalOOMEMark iom(THREAD); 1263 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1264 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1265 } else { 1266 obj = ik->allocate_instance(THREAD); 1267 } 1268 } 1269 } else if (k->is_typeArray_klass()) { 1270 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1271 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1272 int len = sv->field_size() / type2size[ak->element_type()]; 1273 InternalOOMEMark iom(THREAD); 1274 obj = ak->allocate_instance(len, THREAD); 1275 } else if (k->is_objArray_klass()) { 1276 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1277 InternalOOMEMark iom(THREAD); 1278 obj = ak->allocate_instance(sv->field_size(), THREAD); 1279 } 1280 1281 if (obj == nullptr) { 1282 failures = true; 1283 } 1284 1285 assert(sv->value().is_null(), "redundant reallocation"); 1286 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception"); 1287 CLEAR_PENDING_EXCEPTION; 1288 sv->set_value(obj); 1289 } 1290 1291 if (failures) { 1292 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1293 } else if (pending_exception.not_null()) { 1294 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1295 } 1296 1297 return failures; 1298 } 1299 1300 #if INCLUDE_JVMCI 1301 /** 1302 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1303 * we need to somehow be able to recover the actual kind to be able to write the correct 1304 * amount of bytes. 1305 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1306 * the entries at index n + 1 to n + i are 'markers'. 1307 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1308 * expected form of the array would be: 1309 * 1310 * {b0, b1, b2, b3, INT, marker, b6, b7} 1311 * 1312 * Thus, in order to get back the size of the entry, we simply need to count the number 1313 * of marked entries 1314 * 1315 * @param virtualArray the virtualized byte array 1316 * @param i index of the virtual entry we are recovering 1317 * @return The number of bytes the entry spans 1318 */ 1319 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1320 int index = i; 1321 while (++index < virtualArray->field_size() && 1322 virtualArray->field_at(index)->is_marker()) {} 1323 return index - i; 1324 } 1325 1326 /** 1327 * If there was a guarantee for byte array to always start aligned to a long, we could 1328 * do a simple check on the parity of the index. Unfortunately, that is not always the 1329 * case. Thus, we check alignment of the actual address we are writing to. 1330 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1331 * write a long to index 3. 1332 */ 1333 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1334 jbyte* res = obj->byte_at_addr(index); 1335 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1336 return res; 1337 } 1338 1339 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) { 1340 switch (byte_count) { 1341 case 1: 1342 obj->byte_at_put(index, (jbyte) value->get_jint()); 1343 break; 1344 case 2: 1345 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint(); 1346 break; 1347 case 4: 1348 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint(); 1349 break; 1350 case 8: 1351 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr(); 1352 break; 1353 default: 1354 ShouldNotReachHere(); 1355 } 1356 } 1357 #endif // INCLUDE_JVMCI 1358 1359 1360 // restore elements of an eliminated type array 1361 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1362 int index = 0; 1363 1364 for (int i = 0; i < sv->field_size(); i++) { 1365 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1366 switch(type) { 1367 case T_LONG: case T_DOUBLE: { 1368 assert(value->type() == T_INT, "Agreement."); 1369 StackValue* low = 1370 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1371 #ifdef _LP64 1372 jlong res = (jlong)low->get_intptr(); 1373 #else 1374 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1375 #endif 1376 obj->long_at_put(index, res); 1377 break; 1378 } 1379 1380 case T_INT: case T_FLOAT: { // 4 bytes. 1381 assert(value->type() == T_INT, "Agreement."); 1382 bool big_value = false; 1383 if (i + 1 < sv->field_size() && type == T_INT) { 1384 if (sv->field_at(i)->is_location()) { 1385 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1386 if (type == Location::dbl || type == Location::lng) { 1387 big_value = true; 1388 } 1389 } else if (sv->field_at(i)->is_constant_int()) { 1390 ScopeValue* next_scope_field = sv->field_at(i + 1); 1391 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1392 big_value = true; 1393 } 1394 } 1395 } 1396 1397 if (big_value) { 1398 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1399 #ifdef _LP64 1400 jlong res = (jlong)low->get_intptr(); 1401 #else 1402 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1403 #endif 1404 obj->int_at_put(index, *(jint*)&res); 1405 obj->int_at_put(++index, *((jint*)&res + 1)); 1406 } else { 1407 obj->int_at_put(index, value->get_jint()); 1408 } 1409 break; 1410 } 1411 1412 case T_SHORT: 1413 assert(value->type() == T_INT, "Agreement."); 1414 obj->short_at_put(index, (jshort)value->get_jint()); 1415 break; 1416 1417 case T_CHAR: 1418 assert(value->type() == T_INT, "Agreement."); 1419 obj->char_at_put(index, (jchar)value->get_jint()); 1420 break; 1421 1422 case T_BYTE: { 1423 assert(value->type() == T_INT, "Agreement."); 1424 #if INCLUDE_JVMCI 1425 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1426 int byte_count = count_number_of_bytes_for_entry(sv, i); 1427 byte_array_put(obj, value, index, byte_count); 1428 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1429 i += byte_count - 1; // Balance the loop counter. 1430 index += byte_count; 1431 // index has been updated so continue at top of loop 1432 continue; 1433 #else 1434 obj->byte_at_put(index, (jbyte)value->get_jint()); 1435 break; 1436 #endif // INCLUDE_JVMCI 1437 } 1438 1439 case T_BOOLEAN: { 1440 assert(value->type() == T_INT, "Agreement."); 1441 obj->bool_at_put(index, (jboolean)value->get_jint()); 1442 break; 1443 } 1444 1445 default: 1446 ShouldNotReachHere(); 1447 } 1448 index++; 1449 } 1450 } 1451 1452 // restore fields of an eliminated object array 1453 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1454 for (int i = 0; i < sv->field_size(); i++) { 1455 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1456 assert(value->type() == T_OBJECT, "object element expected"); 1457 obj->obj_at_put(i, value->get_obj()()); 1458 } 1459 } 1460 1461 class ReassignedField { 1462 public: 1463 int _offset; 1464 BasicType _type; 1465 public: 1466 ReassignedField() { 1467 _offset = 0; 1468 _type = T_ILLEGAL; 1469 } 1470 }; 1471 1472 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned 1473 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) { 1474 InstanceKlass* super = klass->super(); 1475 if (super != nullptr) { 1476 get_reassigned_fields(super, fields, is_jvmci); 1477 } 1478 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1479 if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) { 1480 ReassignedField field; 1481 field._offset = fs.offset(); 1482 field._type = Signature::basic_type(fs.signature()); 1483 fields->append(field); 1484 } 1485 } 1486 return fields; 1487 } 1488 1489 // Restore fields of an eliminated instance object employing the same field order used by the compiler. 1490 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool is_jvmci) { 1491 GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci); 1492 for (int i = 0; i < fields->length(); i++) { 1493 ScopeValue* scope_field = sv->field_at(svIndex); 1494 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1495 int offset = fields->at(i)._offset; 1496 BasicType type = fields->at(i)._type; 1497 switch (type) { 1498 case T_OBJECT: case T_ARRAY: 1499 assert(value->type() == T_OBJECT, "Agreement."); 1500 obj->obj_field_put(offset, value->get_obj()()); 1501 break; 1502 1503 case T_INT: case T_FLOAT: { // 4 bytes. 1504 assert(value->type() == T_INT, "Agreement."); 1505 bool big_value = false; 1506 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1507 if (scope_field->is_location()) { 1508 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1509 if (type == Location::dbl || type == Location::lng) { 1510 big_value = true; 1511 } 1512 } 1513 if (scope_field->is_constant_int()) { 1514 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1515 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1516 big_value = true; 1517 } 1518 } 1519 } 1520 1521 if (big_value) { 1522 i++; 1523 assert(i < fields->length(), "second T_INT field needed"); 1524 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1525 } else { 1526 obj->int_field_put(offset, value->get_jint()); 1527 break; 1528 } 1529 } 1530 /* no break */ 1531 1532 case T_LONG: case T_DOUBLE: { 1533 assert(value->type() == T_INT, "Agreement."); 1534 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1535 #ifdef _LP64 1536 jlong res = (jlong)low->get_intptr(); 1537 #else 1538 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1539 #endif 1540 obj->long_field_put(offset, res); 1541 break; 1542 } 1543 1544 case T_SHORT: 1545 assert(value->type() == T_INT, "Agreement."); 1546 obj->short_field_put(offset, (jshort)value->get_jint()); 1547 break; 1548 1549 case T_CHAR: 1550 assert(value->type() == T_INT, "Agreement."); 1551 obj->char_field_put(offset, (jchar)value->get_jint()); 1552 break; 1553 1554 case T_BYTE: 1555 assert(value->type() == T_INT, "Agreement."); 1556 obj->byte_field_put(offset, (jbyte)value->get_jint()); 1557 break; 1558 1559 case T_BOOLEAN: 1560 assert(value->type() == T_INT, "Agreement."); 1561 obj->bool_field_put(offset, (jboolean)value->get_jint()); 1562 break; 1563 1564 default: 1565 ShouldNotReachHere(); 1566 } 1567 svIndex++; 1568 } 1569 return svIndex; 1570 } 1571 1572 // restore fields of all eliminated objects and arrays 1573 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci) { 1574 for (int i = 0; i < objects->length(); i++) { 1575 assert(objects->at(i)->is_object(), "invalid debug information"); 1576 ObjectValue* sv = (ObjectValue*) objects->at(i); 1577 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1578 Handle obj = sv->value(); 1579 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1580 #ifndef PRODUCT 1581 if (PrintDeoptimizationDetails) { 1582 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1583 } 1584 #endif // !PRODUCT 1585 1586 if (obj.is_null()) { 1587 continue; 1588 } 1589 1590 #if INCLUDE_JVMCI 1591 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1592 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1593 continue; 1594 } 1595 #endif // INCLUDE_JVMCI 1596 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1597 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1598 ScopeValue* payload = sv->field_at(0); 1599 if (payload->is_location() && 1600 payload->as_LocationValue()->location().type() == Location::vector) { 1601 #ifndef PRODUCT 1602 if (PrintDeoptimizationDetails) { 1603 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1604 if (Verbose) { 1605 Handle obj = sv->value(); 1606 k->oop_print_on(obj(), tty); 1607 } 1608 } 1609 #endif // !PRODUCT 1610 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1611 } 1612 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1613 // which could be restored after vector object allocation. 1614 } 1615 if (k->is_instance_klass()) { 1616 InstanceKlass* ik = InstanceKlass::cast(k); 1617 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci); 1618 } else if (k->is_typeArray_klass()) { 1619 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1620 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1621 } else if (k->is_objArray_klass()) { 1622 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1623 } 1624 } 1625 // These objects may escape when we return to Interpreter after deoptimization. 1626 // We need barrier so that stores that initialize these objects can't be reordered 1627 // with subsequent stores that make these objects accessible by other threads. 1628 OrderAccess::storestore(); 1629 } 1630 1631 1632 // relock objects for which synchronization was eliminated 1633 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1634 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1635 bool relocked_objects = false; 1636 for (int i = 0; i < monitors->length(); i++) { 1637 MonitorInfo* mon_info = monitors->at(i); 1638 if (mon_info->eliminated()) { 1639 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1640 relocked_objects = true; 1641 if (!mon_info->owner_is_scalar_replaced()) { 1642 Handle obj(thread, mon_info->owner()); 1643 markWord mark = obj->mark(); 1644 if (exec_mode == Unpack_none) { 1645 if (mark.has_monitor()) { 1646 // defer relocking if the deoptee thread is currently waiting for obj 1647 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1648 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) { 1649 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1650 if (UseObjectMonitorTable) { 1651 mon_info->lock()->clear_object_monitor_cache(); 1652 } 1653 #ifdef ASSERT 1654 else { 1655 assert(!UseObjectMonitorTable, "must be"); 1656 mon_info->lock()->set_bad_monitor_deopt(); 1657 } 1658 #endif 1659 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1660 continue; 1661 } 1662 } 1663 } 1664 BasicLock* lock = mon_info->lock(); 1665 // We have lost information about the correct state of the lock stack. 1666 // Entering may create an invalid lock stack. Inflate the lock if it 1667 // was fast_locked to restore the valid lock stack. 1668 if (UseObjectMonitorTable) { 1669 // UseObjectMonitorTable expects the BasicLock cache to be either a 1670 // valid ObjectMonitor* or nullptr. Right now it is garbage, set it 1671 // to nullptr. 1672 lock->clear_object_monitor_cache(); 1673 } 1674 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread); 1675 if (deoptee_thread->lock_stack().contains(obj())) { 1676 LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal, 1677 deoptee_thread, thread); 1678 } 1679 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1680 assert(obj->mark().has_monitor(), "must be"); 1681 assert(!deoptee_thread->lock_stack().contains(obj()), "must be"); 1682 assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be"); 1683 } 1684 } 1685 } 1686 return relocked_objects; 1687 } 1688 #endif // COMPILER2_OR_JVMCI 1689 1690 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1691 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1692 1693 // Register map for next frame (used for stack crawl). We capture 1694 // the state of the deopt'ing frame's caller. Thus if we need to 1695 // stuff a C2I adapter we can properly fill in the callee-save 1696 // register locations. 1697 frame caller = fr.sender(reg_map); 1698 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp()); 1699 1700 frame sender = caller; 1701 1702 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1703 // the vframeArray containing the unpacking information is allocated in the C heap. 1704 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1705 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1706 1707 // Compare the vframeArray to the collected vframes 1708 assert(array->structural_compare(thread, chunk), "just checking"); 1709 1710 if (TraceDeoptimization) { 1711 ResourceMark rm; 1712 stringStream st; 1713 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1714 st.print(" "); 1715 fr.print_on(&st); 1716 st.print_cr(" Virtual frames (innermost/newest first):"); 1717 for (int index = 0; index < chunk->length(); index++) { 1718 compiledVFrame* vf = chunk->at(index); 1719 int bci = vf->raw_bci(); 1720 const char* code_name; 1721 if (bci == SynchronizationEntryBCI) { 1722 code_name = "sync entry"; 1723 } else { 1724 Bytecodes::Code code = vf->method()->code_at(bci); 1725 code_name = Bytecodes::name(code); 1726 } 1727 1728 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1729 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1730 st.print(" - %s", code_name); 1731 st.print_cr(" @ bci=%d ", bci); 1732 } 1733 tty->print_raw(st.freeze()); 1734 tty->cr(); 1735 } 1736 1737 return array; 1738 } 1739 1740 #if COMPILER2_OR_JVMCI 1741 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1742 // Reallocation of some scalar replaced objects failed. Record 1743 // that we need to pop all the interpreter frames for the 1744 // deoptimized compiled frame. 1745 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1746 thread->set_frames_to_pop_failed_realloc(array->frames()); 1747 // Unlock all monitors here otherwise the interpreter will see a 1748 // mix of locked and unlocked monitors (because of failed 1749 // reallocations of synchronized objects) and be confused. 1750 for (int i = 0; i < array->frames(); i++) { 1751 MonitorChunk* monitors = array->element(i)->monitors(); 1752 if (monitors != nullptr) { 1753 // Unlock in reverse order starting from most nested monitor. 1754 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) { 1755 BasicObjectLock* src = monitors->at(j); 1756 if (src->obj() != nullptr) { 1757 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1758 } 1759 } 1760 array->element(i)->free_monitors(); 1761 #ifdef ASSERT 1762 array->element(i)->set_removed_monitors(); 1763 #endif 1764 } 1765 } 1766 } 1767 #endif 1768 1769 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1770 assert(fr.can_be_deoptimized(), "checking frame type"); 1771 1772 nmethod* nm = fr.cb()->as_nmethod_or_null(); 1773 assert(nm != nullptr, "only compiled methods can deopt"); 1774 DeoptAction action = (nm->is_not_entrant() ? Action_make_not_entrant : Action_none); 1775 ScopeDesc* cur_sd = nm->scope_desc_at(fr.pc()); 1776 Bytecodes::Code bc = (cur_sd->bci() == -1 ? Bytecodes::_nop // deopt on method entry 1777 : cur_sd->method()->java_code_at(cur_sd->bci())); 1778 gather_statistics(nm, reason, action, bc); 1779 1780 if (LogCompilation && xtty != nullptr) { 1781 ttyLocker ttyl; 1782 xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1783 nm->log_identity(xtty); 1784 xtty->end_head(); 1785 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1786 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1787 xtty->method(sd->method()); 1788 xtty->end_elem(); 1789 if (sd->is_top()) break; 1790 } 1791 xtty->tail("deoptimized"); 1792 } 1793 1794 Continuation::notify_deopt(thread, fr.sp()); 1795 1796 // Patch the compiled method so that when execution returns to it we will 1797 // deopt the execution state and return to the interpreter. 1798 fr.deoptimize(thread); 1799 } 1800 1801 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1802 // Deoptimize only if the frame comes from compile code. 1803 // Do not deoptimize the frame which is already patched 1804 // during the execution of the loops below. 1805 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1806 return; 1807 } 1808 ResourceMark rm; 1809 deoptimize_single_frame(thread, fr, reason); 1810 } 1811 1812 #if INCLUDE_JVMCI 1813 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) { 1814 // there is no exception handler for this pc => deoptimize 1815 nm->make_not_entrant(nmethod::InvalidationReason::MISSING_EXCEPTION_HANDLER); 1816 1817 // Use Deoptimization::deoptimize for all of its side-effects: 1818 // gathering traps statistics, logging... 1819 // it also patches the return pc but we do not care about that 1820 // since we return a continuation to the deopt_blob below. 1821 JavaThread* thread = JavaThread::current(); 1822 RegisterMap reg_map(thread, 1823 RegisterMap::UpdateMap::skip, 1824 RegisterMap::ProcessFrames::include, 1825 RegisterMap::WalkContinuation::skip); 1826 frame runtime_frame = thread->last_frame(); 1827 frame caller_frame = runtime_frame.sender(®_map); 1828 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method"); 1829 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1830 compiledVFrame* cvf = compiledVFrame::cast(vf); 1831 ScopeDesc* imm_scope = cvf->scope(); 1832 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1833 if (imm_mdo != nullptr) { 1834 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 1835 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 1836 1837 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr); 1838 if (pdata != nullptr && pdata->is_BitData()) { 1839 BitData* bit_data = (BitData*) pdata; 1840 bit_data->set_exception_seen(); 1841 } 1842 } 1843 1844 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1845 1846 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true); 1847 if (trap_mdo != nullptr) { 1848 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1849 } 1850 1851 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1852 } 1853 #endif 1854 1855 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1856 assert(thread == Thread::current() || 1857 thread->is_handshake_safe_for(Thread::current()) || 1858 SafepointSynchronize::is_at_safepoint(), 1859 "can only deoptimize other thread at a safepoint/handshake"); 1860 // Compute frame and register map based on thread and sp. 1861 RegisterMap reg_map(thread, 1862 RegisterMap::UpdateMap::skip, 1863 RegisterMap::ProcessFrames::include, 1864 RegisterMap::WalkContinuation::skip); 1865 frame fr = thread->last_frame(); 1866 while (fr.id() != id) { 1867 fr = fr.sender(®_map); 1868 } 1869 deoptimize(thread, fr, reason); 1870 } 1871 1872 1873 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1874 Thread* current = Thread::current(); 1875 if (thread == current || thread->is_handshake_safe_for(current)) { 1876 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1877 } else { 1878 VM_DeoptimizeFrame deopt(thread, id, reason); 1879 VMThread::execute(&deopt); 1880 } 1881 } 1882 1883 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1884 deoptimize_frame(thread, id, Reason_constraint); 1885 } 1886 1887 // JVMTI PopFrame support 1888 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1889 { 1890 assert(thread == JavaThread::current(), "pre-condition"); 1891 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1892 } 1893 JRT_END 1894 1895 MethodData* 1896 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1897 bool create_if_missing) { 1898 JavaThread* THREAD = thread; // For exception macros. 1899 MethodData* mdo = m()->method_data(); 1900 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) { 1901 // Build an MDO. Ignore errors like OutOfMemory; 1902 // that simply means we won't have an MDO to update. 1903 Method::build_profiling_method_data(m, THREAD); 1904 if (HAS_PENDING_EXCEPTION) { 1905 // Only metaspace OOM is expected. No Java code executed. 1906 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1907 CLEAR_PENDING_EXCEPTION; 1908 } 1909 mdo = m()->method_data(); 1910 } 1911 return mdo; 1912 } 1913 1914 #if COMPILER2_OR_JVMCI 1915 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1916 // In case of an unresolved klass entry, load the class. 1917 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1918 // and probably nowhere else. 1919 // Even that case would benefit from simply re-interpreting the 1920 // bytecode, without paying special attention to the class index. 1921 // So this whole "class index" feature should probably be removed. 1922 1923 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1924 Klass* tk = constant_pool->klass_at(index, THREAD); 1925 if (HAS_PENDING_EXCEPTION) { 1926 // Exception happened during classloading. We ignore the exception here, since it 1927 // is going to be rethrown since the current activation is going to be deoptimized and 1928 // the interpreter will re-execute the bytecode. 1929 // Do not clear probable Async Exceptions. 1930 CLEAR_PENDING_NONASYNC_EXCEPTION; 1931 // Class loading called java code which may have caused a stack 1932 // overflow. If the exception was thrown right before the return 1933 // to the runtime the stack is no longer guarded. Reguard the 1934 // stack otherwise if we return to the uncommon trap blob and the 1935 // stack bang causes a stack overflow we crash. 1936 JavaThread* jt = THREAD; 1937 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 1938 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1939 } 1940 return; 1941 } 1942 1943 assert(!constant_pool->tag_at(index).is_symbol(), 1944 "no symbolic names here, please"); 1945 } 1946 1947 #if INCLUDE_JFR 1948 1949 class DeoptReasonSerializer : public JfrSerializer { 1950 public: 1951 void serialize(JfrCheckpointWriter& writer) { 1952 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 1953 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 1954 writer.write_key((u8)i); 1955 writer.write(Deoptimization::trap_reason_name(i)); 1956 } 1957 } 1958 }; 1959 1960 class DeoptActionSerializer : public JfrSerializer { 1961 public: 1962 void serialize(JfrCheckpointWriter& writer) { 1963 static const u4 nof_actions = Deoptimization::Action_LIMIT; 1964 writer.write_count(nof_actions); 1965 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 1966 writer.write_key(i); 1967 writer.write(Deoptimization::trap_action_name((int)i)); 1968 } 1969 } 1970 }; 1971 1972 static void register_serializers() { 1973 static int critical_section = 0; 1974 if (1 == critical_section || AtomicAccess::cmpxchg(&critical_section, 0, 1) == 1) { 1975 return; 1976 } 1977 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 1978 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 1979 } 1980 1981 static void post_deoptimization_event(nmethod* nm, 1982 const Method* method, 1983 int trap_bci, 1984 int instruction, 1985 Deoptimization::DeoptReason reason, 1986 Deoptimization::DeoptAction action) { 1987 assert(nm != nullptr, "invariant"); 1988 assert(method != nullptr, "invariant"); 1989 if (EventDeoptimization::is_enabled()) { 1990 static bool serializers_registered = false; 1991 if (!serializers_registered) { 1992 register_serializers(); 1993 serializers_registered = true; 1994 } 1995 EventDeoptimization event; 1996 event.set_compileId(nm->compile_id()); 1997 event.set_compiler(nm->compiler_type()); 1998 event.set_method(method); 1999 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 2000 event.set_bci(trap_bci); 2001 event.set_instruction(instruction); 2002 event.set_reason(reason); 2003 event.set_action(action); 2004 event.commit(); 2005 } 2006 } 2007 2008 #endif // INCLUDE_JFR 2009 2010 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 2011 const char* reason_name, const char* reason_action) { 2012 LogTarget(Debug, deoptimization) lt; 2013 if (lt.is_enabled()) { 2014 LogStream ls(lt); 2015 bool is_osr = nm->is_osr_method(); 2016 ls.print("cid=%4d %s%s%s level=%d", 2017 nm->compile_id(), (is_osr ? "osr" : " "), 2018 (nm->is_aot() ? "aot " : ""), 2019 (nm->preloaded() ? "preload " : ""), 2020 nm->comp_level()); 2021 ls.print(" %s", tm->name_and_sig_as_C_string()); 2022 ls.print(" trap_bci=%d ", trap_bci); 2023 if (is_osr) { 2024 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 2025 } 2026 ls.print("%s ", reason_name); 2027 ls.print("%s ", reason_action); 2028 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 2029 pc, fr.pc() - nm->code_begin()); 2030 } 2031 } 2032 2033 JRT_ENTRY_PROF(void, Deoptimization, uncommon_trap_inner, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 2034 HandleMark hm(current); 2035 2036 // uncommon_trap() is called at the beginning of the uncommon trap 2037 // handler. Note this fact before we start generating temporary frames 2038 // that can confuse an asynchronous stack walker. This counter is 2039 // decremented at the end of unpack_frames(). 2040 2041 current->inc_in_deopt_handler(); 2042 2043 #if INCLUDE_JVMCI 2044 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 2045 RegisterMap reg_map(current, 2046 RegisterMap::UpdateMap::include, 2047 RegisterMap::ProcessFrames::include, 2048 RegisterMap::WalkContinuation::skip); 2049 #else 2050 RegisterMap reg_map(current, 2051 RegisterMap::UpdateMap::skip, 2052 RegisterMap::ProcessFrames::include, 2053 RegisterMap::WalkContinuation::skip); 2054 #endif 2055 frame stub_frame = current->last_frame(); 2056 frame fr = stub_frame.sender(®_map); 2057 2058 // Log a message 2059 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 2060 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 2061 2062 { 2063 ResourceMark rm; 2064 2065 DeoptReason reason = trap_request_reason(trap_request); 2066 DeoptAction action = trap_request_action(trap_request); 2067 #if INCLUDE_JVMCI 2068 int debug_id = trap_request_debug_id(trap_request); 2069 #endif 2070 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 2071 2072 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 2073 compiledVFrame* cvf = compiledVFrame::cast(vf); 2074 2075 nmethod* nm = cvf->code(); 2076 2077 ScopeDesc* trap_scope = cvf->scope(); 2078 2079 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2080 2081 if (is_receiver_constraint_failure) { 2082 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2083 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2084 JVMCI_ONLY(COMMA debug_id)); 2085 } 2086 2087 methodHandle trap_method(current, trap_scope->method()); 2088 int trap_bci = trap_scope->bci(); 2089 #if INCLUDE_JVMCI 2090 jlong speculation = current->pending_failed_speculation(); 2091 if (nm->is_compiled_by_jvmci()) { 2092 nm->update_speculation(current); 2093 } else { 2094 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2095 } 2096 2097 if (trap_bci == SynchronizationEntryBCI) { 2098 trap_bci = 0; 2099 current->set_pending_monitorenter(true); 2100 } 2101 2102 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2103 current->set_pending_transfer_to_interpreter(true); 2104 } 2105 #endif 2106 2107 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2108 // Record this event in the histogram. 2109 gather_statistics(nm, reason, action, trap_bc); 2110 2111 // Ensure that we can record deopt. history: 2112 bool create_if_missing = ProfileTraps; 2113 2114 methodHandle profiled_method; 2115 #if INCLUDE_JVMCI 2116 if (nm->is_compiled_by_jvmci()) { 2117 profiled_method = methodHandle(current, nm->method()); 2118 } else { 2119 profiled_method = trap_method; 2120 } 2121 #else 2122 profiled_method = trap_method; 2123 #endif 2124 2125 const char* nm_kind = nm->compile_kind(); 2126 MethodData* trap_mdo = 2127 get_method_data(current, profiled_method, create_if_missing); 2128 2129 { // Log Deoptimization event for JFR, UL and event system 2130 Method* tm = trap_method(); 2131 const char* reason_name = trap_reason_name(reason); 2132 const char* reason_action = trap_action_name(action); 2133 intptr_t pc = p2i(fr.pc()); 2134 2135 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2136 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2137 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s %s", 2138 reason_name, reason_action, pc, 2139 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name(), nm_kind); 2140 } 2141 2142 // Print a bunch of diagnostics, if requested. 2143 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2144 ResourceMark rm; 2145 2146 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2147 // We must do this already now, since we cannot acquire this lock while 2148 // holding the tty lock (lock ordering by rank). 2149 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2150 2151 ttyLocker ttyl; 2152 2153 char buf[100]; 2154 if (xtty != nullptr) { 2155 xtty->begin_head("uncommon_trap thread='%zu' %s", 2156 os::current_thread_id(), 2157 format_trap_request(buf, sizeof(buf), trap_request)); 2158 #if INCLUDE_JVMCI 2159 if (speculation != 0) { 2160 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2161 } 2162 #endif 2163 nm->log_identity(xtty); 2164 } 2165 Symbol* class_name = nullptr; 2166 bool unresolved = false; 2167 if (unloaded_class_index >= 0) { 2168 constantPoolHandle constants (current, trap_method->constants()); 2169 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2170 class_name = constants->klass_name_at(unloaded_class_index); 2171 unresolved = true; 2172 if (xtty != nullptr) 2173 xtty->print(" unresolved='1'"); 2174 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2175 class_name = constants->symbol_at(unloaded_class_index); 2176 } 2177 if (xtty != nullptr) 2178 xtty->name(class_name); 2179 } 2180 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) { 2181 // Dump the relevant MDO state. 2182 // This is the deopt count for the current reason, any previous 2183 // reasons or recompiles seen at this point. 2184 int dcnt = trap_mdo->trap_count(reason); 2185 if (dcnt != 0) 2186 xtty->print(" count='%d'", dcnt); 2187 2188 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to 2189 // lock extra_data_lock before the tty lock. 2190 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2191 int dos = (pdata == nullptr)? 0: pdata->trap_state(); 2192 if (dos != 0) { 2193 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2194 if (trap_state_is_recompiled(dos)) { 2195 int recnt2 = trap_mdo->overflow_recompile_count(); 2196 if (recnt2 != 0) 2197 xtty->print(" recompiles2='%d'", recnt2); 2198 } 2199 } 2200 } 2201 if (xtty != nullptr) { 2202 xtty->stamp(); 2203 xtty->end_head(); 2204 } 2205 if (TraceDeoptimization) { // make noise on the tty 2206 stringStream st; 2207 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2208 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2209 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2210 st.print(" compiler=%s compile_id=%d kind=%s", nm->compiler_name(), nm->compile_id(), nm_kind); 2211 #if INCLUDE_JVMCI 2212 if (nm->is_compiled_by_jvmci()) { 2213 const char* installed_code_name = nm->jvmci_name(); 2214 if (installed_code_name != nullptr) { 2215 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2216 } 2217 } 2218 #endif 2219 st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2220 p2i(fr.pc()), 2221 os::current_thread_id(), 2222 trap_reason_name(reason), 2223 trap_action_name(action), 2224 unloaded_class_index 2225 #if INCLUDE_JVMCI 2226 , debug_id 2227 #endif 2228 ); 2229 if (class_name != nullptr) { 2230 st.print(unresolved ? " unresolved class: " : " symbol: "); 2231 class_name->print_symbol_on(&st); 2232 } 2233 st.cr(); 2234 tty->print_raw(st.freeze()); 2235 } 2236 if (xtty != nullptr) { 2237 // Log the precise location of the trap. 2238 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2239 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2240 xtty->method(sd->method()); 2241 xtty->end_elem(); 2242 if (sd->is_top()) break; 2243 } 2244 xtty->tail("uncommon_trap"); 2245 } 2246 } 2247 // (End diagnostic printout.) 2248 2249 if (is_receiver_constraint_failure) { 2250 fatal("missing receiver type check"); 2251 } 2252 2253 // Load class if necessary 2254 if (unloaded_class_index >= 0) { 2255 constantPoolHandle constants(current, trap_method->constants()); 2256 load_class_by_index(constants, unloaded_class_index, THREAD); 2257 } 2258 2259 // Flush the nmethod if necessary and desirable. 2260 // 2261 // We need to avoid situations where we are re-flushing the nmethod 2262 // because of a hot deoptimization site. Repeated flushes at the same 2263 // point need to be detected by the compiler and avoided. If the compiler 2264 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2265 // module must take measures to avoid an infinite cycle of recompilation 2266 // and deoptimization. There are several such measures: 2267 // 2268 // 1. If a recompilation is ordered a second time at some site X 2269 // and for the same reason R, the action is adjusted to 'reinterpret', 2270 // to give the interpreter time to exercise the method more thoroughly. 2271 // If this happens, the method's overflow_recompile_count is incremented. 2272 // 2273 // 2. If the compiler fails to reduce the deoptimization rate, then 2274 // the method's overflow_recompile_count will begin to exceed the set 2275 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2276 // is adjusted to 'make_not_compilable', and the method is abandoned 2277 // to the interpreter. This is a performance hit for hot methods, 2278 // but is better than a disastrous infinite cycle of recompilations. 2279 // (Actually, only the method containing the site X is abandoned.) 2280 // 2281 // 3. In parallel with the previous measures, if the total number of 2282 // recompilations of a method exceeds the much larger set limit 2283 // PerMethodRecompilationCutoff, the method is abandoned. 2284 // This should only happen if the method is very large and has 2285 // many "lukewarm" deoptimizations. The code which enforces this 2286 // limit is elsewhere (class nmethod, class Method). 2287 // 2288 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2289 // to recompile at each bytecode independently of the per-BCI cutoff. 2290 // 2291 // The decision to update code is up to the compiler, and is encoded 2292 // in the Action_xxx code. If the compiler requests Action_none 2293 // no trap state is changed, no compiled code is changed, and the 2294 // computation suffers along in the interpreter. 2295 // 2296 // The other action codes specify various tactics for decompilation 2297 // and recompilation. Action_maybe_recompile is the loosest, and 2298 // allows the compiled code to stay around until enough traps are seen, 2299 // and until the compiler gets around to recompiling the trapping method. 2300 // 2301 // The other actions cause immediate removal of the present code. 2302 2303 // Traps caused by injected profile shouldn't pollute trap counts. 2304 bool injected_profile_trap = trap_method->has_injected_profile() && 2305 (reason == Reason_intrinsic || reason == Reason_unreached); 2306 2307 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2308 bool make_not_entrant = false; 2309 bool make_not_compilable = false; 2310 bool reprofile = false; 2311 switch (action) { 2312 case Action_none: 2313 // Keep the old code. 2314 update_trap_state = false; 2315 break; 2316 case Action_maybe_recompile: 2317 // Do not need to invalidate the present code, but we can 2318 // initiate another 2319 // Start compiler without (necessarily) invalidating the nmethod. 2320 // The system will tolerate the old code, but new code should be 2321 // generated when possible. 2322 break; 2323 case Action_reinterpret: 2324 // Go back into the interpreter for a while, and then consider 2325 // recompiling form scratch. 2326 make_not_entrant = true; 2327 // Reset invocation counter for outer most method. 2328 // This will allow the interpreter to exercise the bytecodes 2329 // for a while before recompiling. 2330 // By contrast, Action_make_not_entrant is immediate. 2331 // 2332 // Note that the compiler will track null_check, null_assert, 2333 // range_check, and class_check events and log them as if they 2334 // had been traps taken from compiled code. This will update 2335 // the MDO trap history so that the next compilation will 2336 // properly detect hot trap sites. 2337 reprofile = true; 2338 break; 2339 case Action_make_not_entrant: 2340 // Request immediate recompilation, and get rid of the old code. 2341 // Make them not entrant, so next time they are called they get 2342 // recompiled. Unloaded classes are loaded now so recompile before next 2343 // time they are called. Same for uninitialized. The interpreter will 2344 // link the missing class, if any. 2345 make_not_entrant = true; 2346 break; 2347 case Action_make_not_compilable: 2348 // Give up on compiling this method at all. 2349 make_not_entrant = true; 2350 make_not_compilable = true; 2351 break; 2352 default: 2353 ShouldNotReachHere(); 2354 } 2355 2356 #if INCLUDE_JVMCI 2357 // Deoptimization count is used by the CompileBroker to reason about compilations 2358 // it requests so do not pollute the count for deoptimizations in non-default (i.e. 2359 // non-CompilerBroker) compilations. 2360 if (nm->jvmci_skip_profile_deopt()) { 2361 update_trap_state = false; 2362 } 2363 #endif 2364 // Setting +ProfileTraps fixes the following, on all platforms: 2365 // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2366 // recompile relies on a MethodData* to record heroic opt failures. 2367 2368 // Whether the interpreter is producing MDO data or not, we also need 2369 // to use the MDO to detect hot deoptimization points and control 2370 // aggressive optimization. 2371 bool inc_recompile_count = false; 2372 2373 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2374 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr, 2375 (trap_mdo != nullptr), 2376 Mutex::_no_safepoint_check_flag); 2377 ProfileData* pdata = nullptr; 2378 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) { 2379 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2380 uint this_trap_count = 0; 2381 bool maybe_prior_trap = false; 2382 bool maybe_prior_recompile = false; 2383 2384 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2385 #if INCLUDE_JVMCI 2386 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2387 #endif 2388 nm->method(), 2389 //outputs: 2390 this_trap_count, 2391 maybe_prior_trap, 2392 maybe_prior_recompile); 2393 // Because the interpreter also counts null, div0, range, and class 2394 // checks, these traps from compiled code are double-counted. 2395 // This is harmless; it just means that the PerXTrapLimit values 2396 // are in effect a little smaller than they look. 2397 2398 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2399 if (per_bc_reason != Reason_none) { 2400 // Now take action based on the partially known per-BCI history. 2401 if (maybe_prior_trap 2402 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2403 // If there are too many traps at this BCI, force a recompile. 2404 // This will allow the compiler to see the limit overflow, and 2405 // take corrective action, if possible. The compiler generally 2406 // does not use the exact PerBytecodeTrapLimit value, but instead 2407 // changes its tactics if it sees any traps at all. This provides 2408 // a little hysteresis, delaying a recompile until a trap happens 2409 // several times. 2410 // 2411 // Actually, since there is only one bit of counter per BCI, 2412 // the possible per-BCI counts are {0,1,(per-method count)}. 2413 // This produces accurate results if in fact there is only 2414 // one hot trap site, but begins to get fuzzy if there are 2415 // many sites. For example, if there are ten sites each 2416 // trapping two or more times, they each get the blame for 2417 // all of their traps. 2418 make_not_entrant = true; 2419 } 2420 2421 // Detect repeated recompilation at the same BCI, and enforce a limit. 2422 if (make_not_entrant && maybe_prior_recompile) { 2423 // More than one recompile at this point. 2424 inc_recompile_count = maybe_prior_trap; 2425 } 2426 } else { 2427 // For reasons which are not recorded per-bytecode, we simply 2428 // force recompiles unconditionally. 2429 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2430 make_not_entrant = true; 2431 } 2432 2433 // Go back to the compiler if there are too many traps in this method. 2434 if (this_trap_count >= per_method_trap_limit(reason)) { 2435 // If there are too many traps in this method, force a recompile. 2436 // This will allow the compiler to see the limit overflow, and 2437 // take corrective action, if possible. 2438 // (This condition is an unlikely backstop only, because the 2439 // PerBytecodeTrapLimit is more likely to take effect first, 2440 // if it is applicable.) 2441 make_not_entrant = true; 2442 } 2443 2444 // Here's more hysteresis: If there has been a recompile at 2445 // this trap point already, run the method in the interpreter 2446 // for a while to exercise it more thoroughly. 2447 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2448 reprofile = true; 2449 } 2450 } 2451 2452 // Take requested actions on the method: 2453 2454 // Recompile 2455 if (make_not_entrant) { 2456 if (!nm->make_not_entrant(nmethod::InvalidationReason::UNCOMMON_TRAP)) { 2457 return; // the call did not change nmethod's state 2458 } 2459 2460 if (pdata != nullptr) { 2461 // Record the recompilation event, if any. 2462 int tstate0 = pdata->trap_state(); 2463 int tstate1 = trap_state_set_recompiled(tstate0, true); 2464 if (tstate1 != tstate0) 2465 pdata->set_trap_state(tstate1); 2466 } 2467 2468 // For code aging we count traps separately here, using make_not_entrant() 2469 // as a guard against simultaneous deopts in multiple threads. 2470 if (reason == Reason_tenured && trap_mdo != nullptr) { 2471 trap_mdo->inc_tenure_traps(); 2472 } 2473 } 2474 if (inc_recompile_count) { 2475 trap_mdo->inc_overflow_recompile_count(); 2476 if ((uint)trap_mdo->overflow_recompile_count() > 2477 (uint)PerBytecodeRecompilationCutoff) { 2478 // Give up on the method containing the bad BCI. 2479 if (trap_method() == nm->method()) { 2480 make_not_compilable = true; 2481 } else { 2482 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2483 // But give grace to the enclosing nm->method(). 2484 } 2485 } 2486 } 2487 2488 // Reprofile 2489 if (reprofile) { 2490 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2491 } 2492 2493 // Give up compiling 2494 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2495 assert(make_not_entrant, "consistent"); 2496 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2497 } 2498 2499 if (ProfileExceptionHandlers && trap_mdo != nullptr) { 2500 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci); 2501 if (exception_handler_data != nullptr) { 2502 // uncommon trap at the start of an exception handler. 2503 // C2 generates these for un-entered exception handlers. 2504 // mark the handler as entered to avoid generating 2505 // another uncommon trap the next time the handler is compiled 2506 exception_handler_data->set_exception_handler_entered(); 2507 } 2508 } 2509 2510 } // Free marked resources 2511 2512 } 2513 JRT_END 2514 2515 ProfileData* 2516 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2517 int trap_bci, 2518 Deoptimization::DeoptReason reason, 2519 bool update_total_trap_count, 2520 #if INCLUDE_JVMCI 2521 bool is_osr, 2522 #endif 2523 Method* compiled_method, 2524 //outputs: 2525 uint& ret_this_trap_count, 2526 bool& ret_maybe_prior_trap, 2527 bool& ret_maybe_prior_recompile) { 2528 trap_mdo->check_extra_data_locked(); 2529 2530 bool maybe_prior_trap = false; 2531 bool maybe_prior_recompile = false; 2532 uint this_trap_count = 0; 2533 if (update_total_trap_count) { 2534 uint idx = reason; 2535 #if INCLUDE_JVMCI 2536 if (is_osr) { 2537 // Upper half of history array used for traps in OSR compilations 2538 idx += Reason_TRAP_HISTORY_LENGTH; 2539 } 2540 #endif 2541 uint prior_trap_count = trap_mdo->trap_count(idx); 2542 this_trap_count = trap_mdo->inc_trap_count(idx); 2543 2544 // If the runtime cannot find a place to store trap history, 2545 // it is estimated based on the general condition of the method. 2546 // If the method has ever been recompiled, or has ever incurred 2547 // a trap with the present reason , then this BCI is assumed 2548 // (pessimistically) to be the culprit. 2549 maybe_prior_trap = (prior_trap_count != 0); 2550 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2551 } 2552 ProfileData* pdata = nullptr; 2553 2554 2555 // For reasons which are recorded per bytecode, we check per-BCI data. 2556 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2557 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2558 if (per_bc_reason != Reason_none) { 2559 // Find the profile data for this BCI. If there isn't one, 2560 // try to allocate one from the MDO's set of spares. 2561 // This will let us detect a repeated trap at this point. 2562 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr); 2563 2564 if (pdata != nullptr) { 2565 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2566 if (LogCompilation && xtty != nullptr) { 2567 ttyLocker ttyl; 2568 // no more room for speculative traps in this MDO 2569 xtty->elem("speculative_traps_oom"); 2570 } 2571 } 2572 // Query the trap state of this profile datum. 2573 int tstate0 = pdata->trap_state(); 2574 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2575 maybe_prior_trap = false; 2576 if (!trap_state_is_recompiled(tstate0)) 2577 maybe_prior_recompile = false; 2578 2579 // Update the trap state of this profile datum. 2580 int tstate1 = tstate0; 2581 // Record the reason. 2582 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2583 // Store the updated state on the MDO, for next time. 2584 if (tstate1 != tstate0) 2585 pdata->set_trap_state(tstate1); 2586 } else { 2587 if (LogCompilation && xtty != nullptr) { 2588 ttyLocker ttyl; 2589 // Missing MDP? Leave a small complaint in the log. 2590 xtty->elem("missing_mdp bci='%d'", trap_bci); 2591 } 2592 } 2593 } 2594 2595 // Return results: 2596 ret_this_trap_count = this_trap_count; 2597 ret_maybe_prior_trap = maybe_prior_trap; 2598 ret_maybe_prior_recompile = maybe_prior_recompile; 2599 return pdata; 2600 } 2601 2602 void 2603 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2604 ResourceMark rm; 2605 // Ignored outputs: 2606 uint ignore_this_trap_count; 2607 bool ignore_maybe_prior_trap; 2608 bool ignore_maybe_prior_recompile; 2609 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2610 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2611 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2612 2613 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2614 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2615 2616 query_update_method_data(trap_mdo, trap_bci, 2617 (DeoptReason)reason, 2618 update_total_counts, 2619 #if INCLUDE_JVMCI 2620 false, 2621 #endif 2622 nullptr, 2623 ignore_this_trap_count, 2624 ignore_maybe_prior_trap, 2625 ignore_maybe_prior_recompile); 2626 } 2627 2628 PROF_ENTRY(Deoptimization::UnrollBlock*, Deoptimization, uncommon_trap, Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode)) 2629 // Enable WXWrite: current function is called from methods compiled by C2 directly 2630 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2631 2632 // Still in Java no safepoints 2633 { 2634 // This enters VM and may safepoint 2635 uncommon_trap_inner(current, trap_request); 2636 } 2637 HandleMark hm(current); 2638 return fetch_unroll_info_helper(current, exec_mode); 2639 PROF_END 2640 2641 // Local derived constants. 2642 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2643 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2644 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2645 2646 //---------------------------trap_state_reason--------------------------------- 2647 Deoptimization::DeoptReason 2648 Deoptimization::trap_state_reason(int trap_state) { 2649 // This assert provides the link between the width of DataLayout::trap_bits 2650 // and the encoding of "recorded" reasons. It ensures there are enough 2651 // bits to store all needed reasons in the per-BCI MDO profile. 2652 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2653 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2654 trap_state -= recompile_bit; 2655 if (trap_state == DS_REASON_MASK) { 2656 return Reason_many; 2657 } else { 2658 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2659 return (DeoptReason)trap_state; 2660 } 2661 } 2662 //-------------------------trap_state_has_reason------------------------------- 2663 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2664 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2665 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2666 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2667 trap_state -= recompile_bit; 2668 if (trap_state == DS_REASON_MASK) { 2669 return -1; // true, unspecifically (bottom of state lattice) 2670 } else if (trap_state == reason) { 2671 return 1; // true, definitely 2672 } else if (trap_state == 0) { 2673 return 0; // false, definitely (top of state lattice) 2674 } else { 2675 return 0; // false, definitely 2676 } 2677 } 2678 //-------------------------trap_state_add_reason------------------------------- 2679 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2680 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2681 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2682 trap_state -= recompile_bit; 2683 if (trap_state == DS_REASON_MASK) { 2684 return trap_state + recompile_bit; // already at state lattice bottom 2685 } else if (trap_state == reason) { 2686 return trap_state + recompile_bit; // the condition is already true 2687 } else if (trap_state == 0) { 2688 return reason + recompile_bit; // no condition has yet been true 2689 } else { 2690 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2691 } 2692 } 2693 //-----------------------trap_state_is_recompiled------------------------------ 2694 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2695 return (trap_state & DS_RECOMPILE_BIT) != 0; 2696 } 2697 //-----------------------trap_state_set_recompiled----------------------------- 2698 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2699 if (z) return trap_state | DS_RECOMPILE_BIT; 2700 else return trap_state & ~DS_RECOMPILE_BIT; 2701 } 2702 //---------------------------format_trap_state--------------------------------- 2703 // This is used for debugging and diagnostics, including LogFile output. 2704 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2705 int trap_state) { 2706 assert(buflen > 0, "sanity"); 2707 DeoptReason reason = trap_state_reason(trap_state); 2708 bool recomp_flag = trap_state_is_recompiled(trap_state); 2709 // Re-encode the state from its decoded components. 2710 int decoded_state = 0; 2711 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2712 decoded_state = trap_state_add_reason(decoded_state, reason); 2713 if (recomp_flag) 2714 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2715 // If the state re-encodes properly, format it symbolically. 2716 // Because this routine is used for debugging and diagnostics, 2717 // be robust even if the state is a strange value. 2718 size_t len; 2719 if (decoded_state != trap_state) { 2720 // Random buggy state that doesn't decode?? 2721 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2722 } else { 2723 len = jio_snprintf(buf, buflen, "%s%s", 2724 trap_reason_name(reason), 2725 recomp_flag ? " recompiled" : ""); 2726 } 2727 return buf; 2728 } 2729 2730 2731 //--------------------------------statics-------------------------------------- 2732 const char* Deoptimization::_trap_reason_name[] = { 2733 // Note: Keep this in sync. with enum DeoptReason. 2734 "none", 2735 "null_check", 2736 "null_assert" JVMCI_ONLY("_or_unreached0"), 2737 "range_check", 2738 "class_check", 2739 "array_check", 2740 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2741 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2742 "profile_predicate", 2743 "auto_vectorization_check", 2744 "unloaded", 2745 "uninitialized", 2746 "initialized", 2747 "unreached", 2748 "unhandled", 2749 "constraint", 2750 "div0_check", 2751 "age", 2752 "predicate", 2753 "loop_limit_check", 2754 "speculate_class_check", 2755 "speculate_null_check", 2756 "speculate_null_assert", 2757 "unstable_if", 2758 "unstable_fused_if", 2759 "receiver_constraint", 2760 "short_running_loop" JVMCI_ONLY("_or_aliasing"), 2761 #if INCLUDE_JVMCI 2762 "transfer_to_interpreter", 2763 "not_compiled_exception_handler", 2764 "unresolved", 2765 "jsr_mismatch", 2766 #endif 2767 "tenured" 2768 }; 2769 const char* Deoptimization::_trap_action_name[] = { 2770 // Note: Keep this in sync. with enum DeoptAction. 2771 "none", 2772 "maybe_recompile", 2773 "reinterpret", 2774 "make_not_entrant", 2775 "make_not_compilable" 2776 }; 2777 2778 const char* Deoptimization::trap_reason_name(int reason) { 2779 // Check that every reason has a name 2780 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2781 2782 if (reason == Reason_many) return "many"; 2783 if ((uint)reason < Reason_LIMIT) 2784 return _trap_reason_name[reason]; 2785 static char buf[20]; 2786 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2787 return buf; 2788 } 2789 const char* Deoptimization::trap_action_name(int action) { 2790 // Check that every action has a name 2791 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2792 2793 if ((uint)action < Action_LIMIT) 2794 return _trap_action_name[action]; 2795 static char buf[20]; 2796 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2797 return buf; 2798 } 2799 2800 // This is used for debugging and diagnostics, including LogFile output. 2801 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2802 int trap_request) { 2803 jint unloaded_class_index = trap_request_index(trap_request); 2804 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2805 const char* action = trap_action_name(trap_request_action(trap_request)); 2806 #if INCLUDE_JVMCI 2807 int debug_id = trap_request_debug_id(trap_request); 2808 #endif 2809 size_t len; 2810 if (unloaded_class_index < 0) { 2811 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2812 reason, action 2813 #if INCLUDE_JVMCI 2814 ,debug_id 2815 #endif 2816 ); 2817 } else { 2818 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2819 reason, action, unloaded_class_index 2820 #if INCLUDE_JVMCI 2821 ,debug_id 2822 #endif 2823 ); 2824 } 2825 return buf; 2826 } 2827 2828 juint Deoptimization::_deoptimization_hist 2829 [1 + 4 + 5] // total + online + archived 2830 [Deoptimization::Reason_LIMIT] 2831 [1 + Deoptimization::Action_LIMIT] 2832 [Deoptimization::BC_CASE_LIMIT] 2833 = {0}; 2834 2835 enum { 2836 LSB_BITS = 8, 2837 LSB_MASK = right_n_bits(LSB_BITS) 2838 }; 2839 2840 static void update(juint* cases, Bytecodes::Code bc) { 2841 juint* bc_counter_addr = nullptr; 2842 juint bc_counter = 0; 2843 // Look for an unused counter, or an exact match to this BC. 2844 if (bc != Bytecodes::_illegal) { 2845 for (int bc_case = 0; bc_case < Deoptimization::BC_CASE_LIMIT; bc_case++) { 2846 juint* counter_addr = &cases[bc_case]; 2847 juint counter = *counter_addr; 2848 if ((counter == 0 && bc_counter_addr == nullptr) 2849 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2850 // this counter is either free or is already devoted to this BC 2851 bc_counter_addr = counter_addr; 2852 bc_counter = counter | bc; 2853 } 2854 } 2855 } 2856 if (bc_counter_addr == nullptr) { 2857 // Overflow, or no given bytecode. 2858 bc_counter_addr = &cases[Deoptimization::BC_CASE_LIMIT-1]; 2859 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2860 } 2861 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2862 } 2863 2864 2865 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action, 2866 Bytecodes::Code bc) { 2867 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2868 assert(action >= 0 && action < Action_LIMIT, "oob"); 2869 _deoptimization_hist[0][Reason_none][0][0] += 1; // total 2870 _deoptimization_hist[0][reason][0][0] += 1; // per-reason total 2871 2872 update(_deoptimization_hist[0][reason][1+action], bc); 2873 2874 uint lvl = nm->comp_level() + (nm->is_aot() ? 4 : 0) + (nm->preloaded() ? 1 : 0); 2875 _deoptimization_hist[lvl][Reason_none][0][0] += 1; // total 2876 _deoptimization_hist[lvl][reason][0][0] += 1; // per-reason total 2877 update(_deoptimization_hist[lvl][reason][1+action], bc); 2878 } 2879 2880 jint Deoptimization::total_deoptimization_count() { 2881 return _deoptimization_hist[0][Reason_none][0][0]; 2882 } 2883 2884 // Get the deopt count for a specific reason and a specific action. If either 2885 // one of 'reason' or 'action' is null, the method returns the sum of all 2886 // deoptimizations with the specific 'action' or 'reason' respectively. 2887 // If both arguments are null, the method returns the total deopt count. 2888 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2889 if (reason_str == nullptr && action_str == nullptr) { 2890 return total_deoptimization_count(); 2891 } 2892 juint counter = 0; 2893 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2894 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) { 2895 for (int action = 0; action < Action_LIMIT; action++) { 2896 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) { 2897 juint* cases = _deoptimization_hist[0][reason][1+action]; 2898 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2899 counter += cases[bc_case] >> LSB_BITS; 2900 } 2901 } 2902 } 2903 } 2904 } 2905 return counter; 2906 } 2907 2908 void Deoptimization::print_statistics() { 2909 ttyLocker ttyl; 2910 if (xtty != nullptr) xtty->head("statistics type='deoptimization'"); 2911 tty->print_cr("Deoptimization traps recorded:"); 2912 print_statistics_on(tty); 2913 if (xtty != nullptr) xtty->tail("statistics"); 2914 } 2915 2916 void Deoptimization::print_statistics_on(const char* title, int lvl, outputStream* st) { 2917 juint total = _deoptimization_hist[lvl][Reason_none][0][0]; 2918 juint account = total; 2919 #define PRINT_STAT_LINE(name, r) \ 2920 st->print_cr(" %d (%4.1f%%) %s", (int)(r), ((r) == total ? 100.0 : (((r) * 100.0) / total)), name); 2921 if (total > 0) { 2922 st->print(" %s: ", title); 2923 PRINT_STAT_LINE("total", total); 2924 // For each non-zero entry in the histogram, print the reason, 2925 // the action, and (if specifically known) the type of bytecode. 2926 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2927 for (int action = 0; action < Action_LIMIT; action++) { 2928 juint* cases = Deoptimization::_deoptimization_hist[lvl][reason][1+action]; 2929 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2930 juint counter = cases[bc_case]; 2931 if (counter != 0) { 2932 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2933 const char* bc_name = "other"; 2934 if (bc_case == (BC_CASE_LIMIT-1) && bc == Bytecodes::_nop) { 2935 // overwritten 2936 } else if (Bytecodes::is_defined(bc)) { 2937 bc_name = Bytecodes::name(bc); 2938 } 2939 juint r = counter >> LSB_BITS; 2940 st->print_cr(" %-34s %16s %16s: " UINT32_FORMAT_W(5) " (%4.1f%%)", 2941 trap_reason_name(reason), trap_action_name(action), bc_name, 2942 r, (r * 100.0) / total); 2943 account -= r; 2944 } 2945 } 2946 } 2947 } 2948 if (account != 0) { 2949 PRINT_STAT_LINE("unaccounted", account); 2950 } 2951 #undef PRINT_STAT_LINE 2952 } 2953 } 2954 2955 void Deoptimization::print_statistics_on(outputStream* st) { 2956 // print_statistics_on("Total", 0, st); 2957 print_statistics_on("Tier1", 1, st); 2958 print_statistics_on("Tier2", 2, st); 2959 print_statistics_on("Tier3", 3, st); 2960 print_statistics_on("Tier4", 4, st); 2961 2962 print_statistics_on("AOT Code Tier1", 5, st); 2963 print_statistics_on("AOT Code Tier2", 6, st); 2964 print_statistics_on("AOT Code Tier4", 8, st); 2965 print_statistics_on("AOT Code Tier5 (preloaded)", 9, st); 2966 } 2967 2968 #define DO_COUNTERS(macro) \ 2969 macro(Deoptimization, fetch_unroll_info) \ 2970 macro(Deoptimization, unpack_frames) \ 2971 macro(Deoptimization, uncommon_trap_inner) \ 2972 macro(Deoptimization, uncommon_trap) 2973 2974 #define INIT_COUNTER(sub, name) \ 2975 NEWPERFTICKCOUNTERS(_perf_##sub##_##name##_timer, SUN_RT, #sub "::" #name) \ 2976 NEWPERFEVENTCOUNTER(_perf_##sub##_##name##_count, SUN_RT, #sub "::" #name "_count"); 2977 2978 void Deoptimization::init_counters() { 2979 if (ProfileRuntimeCalls && UsePerfData) { 2980 EXCEPTION_MARK; 2981 2982 DO_COUNTERS(INIT_COUNTER) 2983 2984 if (HAS_PENDING_EXCEPTION) { 2985 vm_exit_during_initialization("jvm_perf_init failed unexpectedly"); 2986 } 2987 } 2988 } 2989 #undef INIT_COUNTER 2990 2991 #define PRINT_COUNTER(sub, name) { \ 2992 jlong count = _perf_##sub##_##name##_count->get_value(); \ 2993 if (count > 0) { \ 2994 st->print_cr(" %-50s = " JLONG_FORMAT_W(6) "us (elapsed) " JLONG_FORMAT_W(6) "us (thread) (" JLONG_FORMAT_W(5) " events)", #sub "::" #name, \ 2995 _perf_##sub##_##name##_timer->elapsed_counter_value_us(), \ 2996 _perf_##sub##_##name##_timer->thread_counter_value_us(), \ 2997 count); \ 2998 }} 2999 3000 void Deoptimization::print_counters_on(outputStream* st) { 3001 if (ProfileRuntimeCalls && UsePerfData) { 3002 DO_COUNTERS(PRINT_COUNTER) 3003 } else { 3004 st->print_cr(" Deoptimization: no info (%s is disabled)", (UsePerfData ? "ProfileRuntimeCalls" : "UsePerfData")); 3005 } 3006 } 3007 3008 #undef PRINT_COUNTER 3009 #undef DO_COUNTERS 3010 3011 #else // COMPILER2_OR_JVMCI 3012 3013 3014 // Stubs for C1 only system. 3015 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 3016 return false; 3017 } 3018 3019 const char* Deoptimization::trap_reason_name(int reason) { 3020 return "unknown"; 3021 } 3022 3023 jint Deoptimization::total_deoptimization_count() { 3024 return 0; 3025 } 3026 3027 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 3028 return 0; 3029 } 3030 3031 void Deoptimization::print_statistics() { 3032 // no output 3033 } 3034 3035 void 3036 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 3037 // no update 3038 } 3039 3040 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 3041 return 0; 3042 } 3043 3044 void Deoptimization::gather_statistics(nmethod* nm, DeoptReason reason, DeoptAction action, 3045 Bytecodes::Code bc) { 3046 // no update 3047 } 3048 3049 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 3050 int trap_state) { 3051 jio_snprintf(buf, buflen, "#%d", trap_state); 3052 return buf; 3053 } 3054 3055 void Deoptimization::init_counters() { 3056 // nothing to do 3057 } 3058 3059 void Deoptimization::print_counters_on(outputStream* st) { 3060 // no output 3061 } 3062 3063 void Deoptimization::print_statistics_on(outputStream* st) { 3064 // no output 3065 } 3066 3067 #endif // COMPILER2_OR_JVMCI