1 /* 2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/javaClasses.inline.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "classfile/vmClasses.hpp" 30 #include "code/codeCache.hpp" 31 #include "code/debugInfoRec.hpp" 32 #include "code/nmethod.hpp" 33 #include "code/pcDesc.hpp" 34 #include "code/scopeDesc.hpp" 35 #include "compiler/compilationPolicy.hpp" 36 #include "compiler/compilerDefinitions.inline.hpp" 37 #include "gc/shared/collectedHeap.hpp" 38 #include "interpreter/bytecode.hpp" 39 #include "interpreter/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/atomic.hpp" 65 #include "runtime/continuation.hpp" 66 #include "runtime/continuationEntry.inline.hpp" 67 #include "runtime/deoptimization.hpp" 68 #include "runtime/escapeBarrier.hpp" 69 #include "runtime/fieldDescriptor.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/safepointVerifiers.hpp" 82 #include "runtime/sharedRuntime.hpp" 83 #include "runtime/signature.hpp" 84 #include "runtime/stackFrameStream.inline.hpp" 85 #include "runtime/stackValue.hpp" 86 #include "runtime/stackWatermarkSet.hpp" 87 #include "runtime/stubRoutines.hpp" 88 #include "runtime/synchronizer.hpp" 89 #include "runtime/threadSMR.hpp" 90 #include "runtime/threadWXSetters.inline.hpp" 91 #include "runtime/vframe.hpp" 92 #include "runtime/vframeArray.hpp" 93 #include "runtime/vframe_hp.hpp" 94 #include "runtime/vmOperations.hpp" 95 #include "utilities/checkedCast.hpp" 96 #include "utilities/events.hpp" 97 #include "utilities/growableArray.hpp" 98 #include "utilities/macros.hpp" 99 #include "utilities/preserveException.hpp" 100 #include "utilities/xmlstream.hpp" 101 #if INCLUDE_JFR 102 #include "jfr/jfrEvents.hpp" 103 #include "jfr/metadata/jfrSerializer.hpp" 104 #endif 105 106 uint64_t DeoptimizationScope::_committed_deopt_gen = 0; 107 uint64_t DeoptimizationScope::_active_deopt_gen = 1; 108 bool DeoptimizationScope::_committing_in_progress = false; 109 110 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) { 111 DEBUG_ONLY(_deopted = false;) 112 113 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag); 114 // If there is nothing to deopt _required_gen is the same as comitted. 115 _required_gen = DeoptimizationScope::_committed_deopt_gen; 116 } 117 118 DeoptimizationScope::~DeoptimizationScope() { 119 assert(_deopted, "Deopt not executed"); 120 } 121 122 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) { 123 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 124 125 // If it's already marked but we still need it to be deopted. 126 if (nm->is_marked_for_deoptimization()) { 127 dependent(nm); 128 return; 129 } 130 131 nmethod::DeoptimizationStatus status = 132 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate; 133 Atomic::store(&nm->_deoptimization_status, status); 134 135 // Make sure active is not committed 136 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be"); 137 assert(nm->_deoptimization_generation == 0, "Is already marked"); 138 139 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen; 140 _required_gen = DeoptimizationScope::_active_deopt_gen; 141 } 142 143 void DeoptimizationScope::dependent(nmethod* nm) { 144 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 145 146 // A method marked by someone else may have a _required_gen lower than what we marked with. 147 // Therefore only store it if it's higher than _required_gen. 148 if (_required_gen < nm->_deoptimization_generation) { 149 _required_gen = nm->_deoptimization_generation; 150 } 151 } 152 153 void DeoptimizationScope::deoptimize_marked() { 154 assert(!_deopted, "Already deopted"); 155 156 // We are not alive yet. 157 if (!Universe::is_fully_initialized()) { 158 DEBUG_ONLY(_deopted = true;) 159 return; 160 } 161 162 // Safepoints are a special case, handled here. 163 if (SafepointSynchronize::is_at_safepoint()) { 164 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen; 165 DeoptimizationScope::_active_deopt_gen++; 166 Deoptimization::deoptimize_all_marked(); 167 DEBUG_ONLY(_deopted = true;) 168 return; 169 } 170 171 uint64_t comitting = 0; 172 bool wait = false; 173 while (true) { 174 { 175 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 176 177 // First we check if we or someone else already deopted the gen we want. 178 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) { 179 DEBUG_ONLY(_deopted = true;) 180 return; 181 } 182 if (!_committing_in_progress) { 183 // The version we are about to commit. 184 comitting = DeoptimizationScope::_active_deopt_gen; 185 // Make sure new marks use a higher gen. 186 DeoptimizationScope::_active_deopt_gen++; 187 _committing_in_progress = true; 188 wait = false; 189 } else { 190 // Another thread is handshaking and committing a gen. 191 wait = true; 192 } 193 } 194 if (wait) { 195 // Wait and let the concurrent handshake be performed. 196 ThreadBlockInVM tbivm(JavaThread::current()); 197 os::naked_yield(); 198 } else { 199 // Performs the handshake. 200 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred. 201 DEBUG_ONLY(_deopted = true;) 202 { 203 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 204 205 // Make sure that committed doesn't go backwards. 206 // Should only happen if we did a deopt during a safepoint above. 207 if (DeoptimizationScope::_committed_deopt_gen < comitting) { 208 DeoptimizationScope::_committed_deopt_gen = comitting; 209 } 210 _committing_in_progress = false; 211 212 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be"); 213 214 return; 215 } 216 } 217 } 218 } 219 220 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 221 int caller_adjustment, 222 int caller_actual_parameters, 223 int number_of_frames, 224 intptr_t* frame_sizes, 225 address* frame_pcs, 226 BasicType return_type, 227 int exec_mode) { 228 _size_of_deoptimized_frame = size_of_deoptimized_frame; 229 _caller_adjustment = caller_adjustment; 230 _caller_actual_parameters = caller_actual_parameters; 231 _number_of_frames = number_of_frames; 232 _frame_sizes = frame_sizes; 233 _frame_pcs = frame_pcs; 234 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 235 _return_type = return_type; 236 _initial_info = 0; 237 // PD (x86 only) 238 _counter_temp = 0; 239 _unpack_kind = exec_mode; 240 _sender_sp_temp = 0; 241 242 _total_frame_sizes = size_of_frames(); 243 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 244 } 245 246 Deoptimization::UnrollBlock::~UnrollBlock() { 247 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 248 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 249 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 250 } 251 252 int Deoptimization::UnrollBlock::size_of_frames() const { 253 // Account first for the adjustment of the initial frame 254 intptr_t result = _caller_adjustment; 255 for (int index = 0; index < number_of_frames(); index++) { 256 result += frame_sizes()[index]; 257 } 258 return checked_cast<int>(result); 259 } 260 261 void Deoptimization::UnrollBlock::print() { 262 ResourceMark rm; 263 stringStream st; 264 st.print_cr("UnrollBlock"); 265 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 266 st.print( " frame_sizes: "); 267 for (int index = 0; index < number_of_frames(); index++) { 268 st.print(INTX_FORMAT " ", frame_sizes()[index]); 269 } 270 st.cr(); 271 tty->print_raw(st.freeze()); 272 } 273 274 // In order to make fetch_unroll_info work properly with escape 275 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 276 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 277 // which is called from the method fetch_unroll_info_helper below. 278 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 279 // fetch_unroll_info() is called at the beginning of the deoptimization 280 // handler. Note this fact before we start generating temporary frames 281 // that can confuse an asynchronous stack walker. This counter is 282 // decremented at the end of unpack_frames(). 283 current->inc_in_deopt_handler(); 284 285 if (exec_mode == Unpack_exception) { 286 // When we get here, a callee has thrown an exception into a deoptimized 287 // frame. That throw might have deferred stack watermark checking until 288 // after unwinding. So we deal with such deferred requests here. 289 StackWatermarkSet::after_unwind(current); 290 } 291 292 return fetch_unroll_info_helper(current, exec_mode); 293 JRT_END 294 295 #if COMPILER2_OR_JVMCI 296 // print information about reallocated objects 297 static void print_objects(JavaThread* deoptee_thread, 298 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 299 ResourceMark rm; 300 stringStream st; // change to logStream with logging 301 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 302 fieldDescriptor fd; 303 304 for (int i = 0; i < objects->length(); i++) { 305 ObjectValue* sv = (ObjectValue*) objects->at(i); 306 Handle obj = sv->value(); 307 308 if (obj.is_null()) { 309 st.print_cr(" nullptr"); 310 continue; 311 } 312 313 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 314 315 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 316 k->print_value_on(&st); 317 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize); 318 319 if (Verbose && k != nullptr) { 320 k->oop_print_on(obj(), &st); 321 } 322 } 323 tty->print_raw(st.freeze()); 324 } 325 326 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method, 327 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 328 bool& deoptimized_objects) { 329 bool realloc_failures = false; 330 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames"); 331 332 JavaThread* deoptee_thread = chunk->at(0)->thread(); 333 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 334 "a frame can only be deoptimized by the owner thread"); 335 336 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map); 337 338 // The flag return_oop() indicates call sites which return oop 339 // in compiled code. Such sites include java method calls, 340 // runtime calls (for example, used to allocate new objects/arrays 341 // on slow code path) and any other calls generated in compiled code. 342 // It is not guaranteed that we can get such information here only 343 // by analyzing bytecode in deoptimized frames. This is why this flag 344 // is set during method compilation (see Compile::Process_OopMap_Node()). 345 // If the previous frame was popped or if we are dispatching an exception, 346 // we don't have an oop result. 347 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 348 Handle return_value; 349 if (save_oop_result) { 350 // Reallocation may trigger GC. If deoptimization happened on return from 351 // call which returns oop we need to save it since it is not in oopmap. 352 oop result = deoptee.saved_oop_result(&map); 353 assert(oopDesc::is_oop_or_null(result), "must be oop"); 354 return_value = Handle(thread, result); 355 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 356 if (TraceDeoptimization) { 357 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 358 tty->cr(); 359 } 360 } 361 if (objects != nullptr) { 362 if (exec_mode == Deoptimization::Unpack_none) { 363 assert(thread->thread_state() == _thread_in_vm, "assumption"); 364 JavaThread* THREAD = thread; // For exception macros. 365 // Clear pending OOM if reallocation fails and return true indicating allocation failure 366 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 367 deoptimized_objects = true; 368 } else { 369 JavaThread* current = thread; // For JRT_BLOCK 370 JRT_BLOCK 371 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 372 JRT_END 373 } 374 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 375 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 376 if (TraceDeoptimization) { 377 print_objects(deoptee_thread, objects, realloc_failures); 378 } 379 } 380 if (save_oop_result) { 381 // Restore result. 382 deoptee.set_saved_oop_result(&map, return_value()); 383 } 384 return realloc_failures; 385 } 386 387 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 388 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 389 JavaThread* deoptee_thread = chunk->at(0)->thread(); 390 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 391 assert(thread == Thread::current(), "should be"); 392 HandleMark hm(thread); 393 #ifndef PRODUCT 394 bool first = true; 395 #endif // !PRODUCT 396 DEBUG_ONLY(GrowableArray<oop> lock_order{0};) 397 // Start locking from outermost/oldest frame 398 for (int i = (chunk->length() - 1); i >= 0; i--) { 399 compiledVFrame* cvf = chunk->at(i); 400 assert (cvf->scope() != nullptr,"expect only compiled java frames"); 401 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 402 if (monitors->is_nonempty()) { 403 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 404 exec_mode, realloc_failures); 405 deoptimized_objects = deoptimized_objects || relocked; 406 #ifdef ASSERT 407 if (LockingMode == LM_LIGHTWEIGHT && !realloc_failures) { 408 for (MonitorInfo* mi : *monitors) { 409 lock_order.push(mi->owner()); 410 } 411 } 412 #endif // ASSERT 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 #ifdef ASSERT 445 if (LockingMode == LM_LIGHTWEIGHT && !realloc_failures) { 446 deoptee_thread->lock_stack().verify_consistent_lock_order(lock_order, exec_mode != Deoptimization::Unpack_none); 447 } 448 #endif // ASSERT 449 } 450 451 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 452 // The given vframes cover one physical frame. 453 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 454 bool& realloc_failures) { 455 frame deoptee = chunk->at(0)->fr(); 456 JavaThread* deoptee_thread = chunk->at(0)->thread(); 457 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 458 RegisterMap map(chunk->at(0)->register_map()); 459 bool deoptimized_objects = false; 460 461 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false); 462 463 // Reallocate the non-escaping objects and restore their fields. 464 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 465 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 466 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects); 467 } 468 469 // MonitorInfo structures used in eliminate_locks are not GC safe. 470 NoSafepointVerifier no_safepoint; 471 472 // Now relock objects if synchronization on them was eliminated. 473 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 474 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 475 } 476 return deoptimized_objects; 477 } 478 #endif // COMPILER2_OR_JVMCI 479 480 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 481 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 482 // When we get here we are about to unwind the deoptee frame. In order to 483 // catch not yet safe to use frames, the following stack watermark barrier 484 // poll will make such frames safe to use. 485 StackWatermarkSet::before_unwind(current); 486 487 // Note: there is a safepoint safety issue here. No matter whether we enter 488 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 489 // the vframeArray is created. 490 // 491 492 // Allocate our special deoptimization ResourceMark 493 DeoptResourceMark* dmark = new DeoptResourceMark(current); 494 assert(current->deopt_mark() == nullptr, "Pending deopt!"); 495 current->set_deopt_mark(dmark); 496 497 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 498 RegisterMap map(current, 499 RegisterMap::UpdateMap::include, 500 RegisterMap::ProcessFrames::include, 501 RegisterMap::WalkContinuation::skip); 502 RegisterMap dummy_map(current, 503 RegisterMap::UpdateMap::skip, 504 RegisterMap::ProcessFrames::include, 505 RegisterMap::WalkContinuation::skip); 506 // Now get the deoptee with a valid map 507 frame deoptee = stub_frame.sender(&map); 508 // Set the deoptee nmethod 509 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!"); 510 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 511 current->set_deopt_compiled_method(nm); 512 513 if (VerifyStack) { 514 current->validate_frame_layout(); 515 } 516 517 // Create a growable array of VFrames where each VFrame represents an inlined 518 // Java frame. This storage is allocated with the usual system arena. 519 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 520 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 521 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 522 while (!vf->is_top()) { 523 assert(vf->is_compiled_frame(), "Wrong frame type"); 524 chunk->push(compiledVFrame::cast(vf)); 525 vf = vf->sender(); 526 } 527 assert(vf->is_compiled_frame(), "Wrong frame type"); 528 chunk->push(compiledVFrame::cast(vf)); 529 530 bool realloc_failures = false; 531 532 #if COMPILER2_OR_JVMCI 533 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 534 535 // Reallocate the non-escaping objects and restore their fields. Then 536 // relock objects if synchronization on them was eliminated. 537 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 538 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 539 bool unused; 540 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused); 541 } 542 #endif // COMPILER2_OR_JVMCI 543 544 // Ensure that no safepoint is taken after pointers have been stored 545 // in fields of rematerialized objects. If a safepoint occurs from here on 546 // out the java state residing in the vframeArray will be missed. 547 // Locks may be rebaised in a safepoint. 548 NoSafepointVerifier no_safepoint; 549 550 #if COMPILER2_OR_JVMCI 551 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 552 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 553 bool unused; 554 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 555 } 556 #endif // COMPILER2_OR_JVMCI 557 558 ScopeDesc* trap_scope = chunk->at(0)->scope(); 559 Handle exceptionObject; 560 if (trap_scope->rethrow_exception()) { 561 #ifndef PRODUCT 562 if (PrintDeoptimizationDetails) { 563 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()); 564 } 565 #endif // !PRODUCT 566 567 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 568 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw"); 569 ScopeValue* topOfStack = expressions->top(); 570 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 571 guarantee(exceptionObject() != nullptr, "exception oop can not be null"); 572 } 573 574 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 575 #if COMPILER2_OR_JVMCI 576 if (realloc_failures) { 577 // This destroys all ScopedValue bindings. 578 current->clear_scopedValueBindings(); 579 pop_frames_failed_reallocs(current, array); 580 } 581 #endif 582 583 assert(current->vframe_array_head() == nullptr, "Pending deopt!"); 584 current->set_vframe_array_head(array); 585 586 // Now that the vframeArray has been created if we have any deferred local writes 587 // added by jvmti then we can free up that structure as the data is now in the 588 // vframeArray 589 590 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 591 592 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 593 CodeBlob* cb = stub_frame.cb(); 594 // Verify we have the right vframeArray 595 assert(cb->frame_size() >= 0, "Unexpected frame size"); 596 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 597 598 // If the deopt call site is a MethodHandle invoke call site we have 599 // to adjust the unpack_sp. 600 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 601 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc())) 602 unpack_sp = deoptee.unextended_sp(); 603 604 #ifdef ASSERT 605 assert(cb->is_deoptimization_stub() || 606 cb->is_uncommon_trap_stub() || 607 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 608 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 609 "unexpected code blob: %s", cb->name()); 610 #endif 611 612 // This is a guarantee instead of an assert because if vframe doesn't match 613 // we will unpack the wrong deoptimized frame and wind up in strange places 614 // where it will be very difficult to figure out what went wrong. Better 615 // to die an early death here than some very obscure death later when the 616 // trail is cold. 617 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 618 // in that it will fail to detect a problem when there is one. This needs 619 // more work in tiger timeframe. 620 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 621 622 int number_of_frames = array->frames(); 623 624 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 625 // virtual activation, which is the reverse of the elements in the vframes array. 626 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 627 // +1 because we always have an interpreter return address for the final slot. 628 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 629 int popframe_extra_args = 0; 630 // Create an interpreter return address for the stub to use as its return 631 // address so the skeletal frames are perfectly walkable 632 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 633 634 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 635 // activation be put back on the expression stack of the caller for reexecution 636 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 637 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 638 } 639 640 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 641 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 642 // than simply use array->sender.pc(). This requires us to walk the current set of frames 643 // 644 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 645 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 646 647 // It's possible that the number of parameters at the call site is 648 // different than number of arguments in the callee when method 649 // handles are used. If the caller is interpreted get the real 650 // value so that the proper amount of space can be added to it's 651 // frame. 652 bool caller_was_method_handle = false; 653 if (deopt_sender.is_interpreted_frame()) { 654 methodHandle method(current, deopt_sender.interpreter_frame_method()); 655 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 656 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 657 // Method handle invokes may involve fairly arbitrary chains of 658 // calls so it's impossible to know how much actual space the 659 // caller has for locals. 660 caller_was_method_handle = true; 661 } 662 } 663 664 // 665 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 666 // frame_sizes/frame_pcs[1] next oldest frame (int) 667 // frame_sizes/frame_pcs[n] youngest frame (int) 668 // 669 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 670 // owns the space for the return address to it's caller). Confusing ain't it. 671 // 672 // The vframe array can address vframes with indices running from 673 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 674 // When we create the skeletal frames we need the oldest frame to be in the zero slot 675 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 676 // so things look a little strange in this loop. 677 // 678 int callee_parameters = 0; 679 int callee_locals = 0; 680 for (int index = 0; index < array->frames(); index++ ) { 681 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 682 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 683 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 684 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 685 callee_locals, 686 index == 0, 687 popframe_extra_args); 688 // This pc doesn't have to be perfect just good enough to identify the frame 689 // as interpreted so the skeleton frame will be walkable 690 // The correct pc will be set when the skeleton frame is completely filled out 691 // The final pc we store in the loop is wrong and will be overwritten below 692 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 693 694 callee_parameters = array->element(index)->method()->size_of_parameters(); 695 callee_locals = array->element(index)->method()->max_locals(); 696 popframe_extra_args = 0; 697 } 698 699 // Compute whether the root vframe returns a float or double value. 700 BasicType return_type; 701 { 702 methodHandle method(current, array->element(0)->method()); 703 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 704 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 705 } 706 707 // Compute information for handling adapters and adjusting the frame size of the caller. 708 int caller_adjustment = 0; 709 710 // Compute the amount the oldest interpreter frame will have to adjust 711 // its caller's stack by. If the caller is a compiled frame then 712 // we pretend that the callee has no parameters so that the 713 // extension counts for the full amount of locals and not just 714 // locals-parms. This is because without a c2i adapter the parm 715 // area as created by the compiled frame will not be usable by 716 // the interpreter. (Depending on the calling convention there 717 // may not even be enough space). 718 719 // QQQ I'd rather see this pushed down into last_frame_adjust 720 // and have it take the sender (aka caller). 721 722 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 723 caller_adjustment = last_frame_adjust(0, callee_locals); 724 } else if (callee_locals > callee_parameters) { 725 // The caller frame may need extending to accommodate 726 // non-parameter locals of the first unpacked interpreted frame. 727 // Compute that adjustment. 728 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 729 } 730 731 // If the sender is deoptimized the we must retrieve the address of the handler 732 // since the frame will "magically" show the original pc before the deopt 733 // and we'd undo the deopt. 734 735 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 736 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 737 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 738 } 739 740 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc"); 741 742 #if INCLUDE_JVMCI 743 if (exceptionObject() != nullptr) { 744 current->set_exception_oop(exceptionObject()); 745 exec_mode = Unpack_exception; 746 } 747 #endif 748 749 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 750 assert(current->has_pending_exception(), "should have thrown OOME"); 751 current->set_exception_oop(current->pending_exception()); 752 current->clear_pending_exception(); 753 exec_mode = Unpack_exception; 754 } 755 756 #if INCLUDE_JVMCI 757 if (current->frames_to_pop_failed_realloc() > 0) { 758 current->set_pending_monitorenter(false); 759 } 760 #endif 761 762 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 763 caller_adjustment * BytesPerWord, 764 caller_was_method_handle ? 0 : callee_parameters, 765 number_of_frames, 766 frame_sizes, 767 frame_pcs, 768 return_type, 769 exec_mode); 770 // On some platforms, we need a way to pass some platform dependent 771 // information to the unpacking code so the skeletal frames come out 772 // correct (initial fp value, unextended sp, ...) 773 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 774 775 if (array->frames() > 1) { 776 if (VerifyStack && TraceDeoptimization) { 777 tty->print_cr("Deoptimizing method containing inlining"); 778 } 779 } 780 781 array->set_unroll_block(info); 782 return info; 783 } 784 785 // Called to cleanup deoptimization data structures in normal case 786 // after unpacking to stack and when stack overflow error occurs 787 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 788 vframeArray *array) { 789 790 // Get array if coming from exception 791 if (array == nullptr) { 792 array = thread->vframe_array_head(); 793 } 794 thread->set_vframe_array_head(nullptr); 795 796 // Free the previous UnrollBlock 797 vframeArray* old_array = thread->vframe_array_last(); 798 thread->set_vframe_array_last(array); 799 800 if (old_array != nullptr) { 801 UnrollBlock* old_info = old_array->unroll_block(); 802 old_array->set_unroll_block(nullptr); 803 delete old_info; 804 delete old_array; 805 } 806 807 // Deallocate any resource creating in this routine and any ResourceObjs allocated 808 // inside the vframeArray (StackValueCollections) 809 810 delete thread->deopt_mark(); 811 thread->set_deopt_mark(nullptr); 812 thread->set_deopt_compiled_method(nullptr); 813 814 815 if (JvmtiExport::can_pop_frame()) { 816 // Regardless of whether we entered this routine with the pending 817 // popframe condition bit set, we should always clear it now 818 thread->clear_popframe_condition(); 819 } 820 821 // unpack_frames() is called at the end of the deoptimization handler 822 // and (in C2) at the end of the uncommon trap handler. Note this fact 823 // so that an asynchronous stack walker can work again. This counter is 824 // incremented at the beginning of fetch_unroll_info() and (in C2) at 825 // the beginning of uncommon_trap(). 826 thread->dec_in_deopt_handler(); 827 } 828 829 // Moved from cpu directories because none of the cpus has callee save values. 830 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 831 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 832 833 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 834 // the days we had adapter frames. When we deoptimize a situation where a 835 // compiled caller calls a compiled caller will have registers it expects 836 // to survive the call to the callee. If we deoptimize the callee the only 837 // way we can restore these registers is to have the oldest interpreter 838 // frame that we create restore these values. That is what this routine 839 // will accomplish. 840 841 // At the moment we have modified c2 to not have any callee save registers 842 // so this problem does not exist and this routine is just a place holder. 843 844 assert(f->is_interpreted_frame(), "must be interpreted"); 845 } 846 847 #ifndef PRODUCT 848 static bool falls_through(Bytecodes::Code bc) { 849 switch (bc) { 850 // List may be incomplete. Here we really only care about bytecodes where compiled code 851 // can deoptimize. 852 case Bytecodes::_goto: 853 case Bytecodes::_goto_w: 854 case Bytecodes::_athrow: 855 return false; 856 default: 857 return true; 858 } 859 } 860 #endif 861 862 // Return BasicType of value being returned 863 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 864 assert(thread == JavaThread::current(), "pre-condition"); 865 866 // We are already active in the special DeoptResourceMark any ResourceObj's we 867 // allocate will be freed at the end of the routine. 868 869 // JRT_LEAF methods don't normally allocate handles and there is a 870 // NoHandleMark to enforce that. It is actually safe to use Handles 871 // in a JRT_LEAF method, and sometimes desirable, but to do so we 872 // must use ResetNoHandleMark to bypass the NoHandleMark, and 873 // then use a HandleMark to ensure any Handles we do create are 874 // cleaned up in this scope. 875 ResetNoHandleMark rnhm; 876 HandleMark hm(thread); 877 878 frame stub_frame = thread->last_frame(); 879 880 Continuation::notify_deopt(thread, stub_frame.sp()); 881 882 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 883 // must point to the vframeArray for the unpack frame. 884 vframeArray* array = thread->vframe_array_head(); 885 UnrollBlock* info = array->unroll_block(); 886 887 // We set the last_Java frame. But the stack isn't really parsable here. So we 888 // clear it to make sure JFR understands not to try and walk stacks from events 889 // in here. 890 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 891 thread->frame_anchor()->set_last_Java_sp(nullptr); 892 893 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 894 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 895 896 thread->frame_anchor()->set_last_Java_sp(sp); 897 898 BasicType bt = info->return_type(); 899 900 // If we have an exception pending, claim that the return type is an oop 901 // so the deopt_blob does not overwrite the exception_oop. 902 903 if (exec_mode == Unpack_exception) 904 bt = T_OBJECT; 905 906 // Cleanup thread deopt data 907 cleanup_deopt_info(thread, array); 908 909 #ifndef PRODUCT 910 if (VerifyStack) { 911 ResourceMark res_mark; 912 // Clear pending exception to not break verification code (restored afterwards) 913 PreserveExceptionMark pm(thread); 914 915 thread->validate_frame_layout(); 916 917 // Verify that the just-unpacked frames match the interpreter's 918 // notions of expression stack and locals 919 vframeArray* cur_array = thread->vframe_array_last(); 920 RegisterMap rm(thread, 921 RegisterMap::UpdateMap::skip, 922 RegisterMap::ProcessFrames::include, 923 RegisterMap::WalkContinuation::skip); 924 rm.set_include_argument_oops(false); 925 bool is_top_frame = true; 926 int callee_size_of_parameters = 0; 927 int callee_max_locals = 0; 928 for (int i = 0; i < cur_array->frames(); i++) { 929 vframeArrayElement* el = cur_array->element(i); 930 frame* iframe = el->iframe(); 931 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 932 933 // Get the oop map for this bci 934 InterpreterOopMap mask; 935 int cur_invoke_parameter_size = 0; 936 bool try_next_mask = false; 937 int next_mask_expression_stack_size = -1; 938 int top_frame_expression_stack_adjustment = 0; 939 methodHandle mh(thread, iframe->interpreter_frame_method()); 940 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 941 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 942 int max_bci = mh->code_size(); 943 // Get to the next bytecode if possible 944 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 945 // Check to see if we can grab the number of outgoing arguments 946 // at an uncommon trap for an invoke (where the compiler 947 // generates debug info before the invoke has executed) 948 Bytecodes::Code cur_code = str.next(); 949 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere; 950 if (Bytecodes::is_invoke(cur_code)) { 951 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 952 cur_invoke_parameter_size = invoke.size_of_parameters(); 953 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 954 callee_size_of_parameters++; 955 } 956 } 957 if (str.bci() < max_bci) { 958 next_code = str.next(); 959 if (next_code >= 0) { 960 // The interpreter oop map generator reports results before 961 // the current bytecode has executed except in the case of 962 // calls. It seems to be hard to tell whether the compiler 963 // has emitted debug information matching the "state before" 964 // a given bytecode or the state after, so we try both 965 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) { 966 // Get expression stack size for the next bytecode 967 InterpreterOopMap next_mask; 968 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 969 next_mask_expression_stack_size = next_mask.expression_stack_size(); 970 if (Bytecodes::is_invoke(next_code)) { 971 Bytecode_invoke invoke(mh, str.bci()); 972 next_mask_expression_stack_size += invoke.size_of_parameters(); 973 } 974 // Need to subtract off the size of the result type of 975 // the bytecode because this is not described in the 976 // debug info but returned to the interpreter in the TOS 977 // caching register 978 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 979 if (bytecode_result_type != T_ILLEGAL) { 980 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 981 } 982 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 983 try_next_mask = true; 984 } 985 } 986 } 987 988 // Verify stack depth and oops in frame 989 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 990 if (!( 991 /* SPARC */ 992 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 993 /* x86 */ 994 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 995 (try_next_mask && 996 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 997 top_frame_expression_stack_adjustment))) || 998 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 999 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 1000 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 1001 )) { 1002 { 1003 // Print out some information that will help us debug the problem 1004 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1005 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 1006 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 1007 if (try_next_mask) { 1008 tty->print_cr(" Next code %s", Bytecodes::name(next_code)); 1009 } 1010 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1011 iframe->interpreter_frame_expression_stack_size()); 1012 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1013 tty->print_cr(" try_next_mask = %d", try_next_mask); 1014 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 1015 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1016 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 1017 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1018 tty->print_cr(" exec_mode = %d", exec_mode); 1019 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1020 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 1021 tty->print_cr(" Interpreted frames:"); 1022 for (int k = 0; k < cur_array->frames(); k++) { 1023 vframeArrayElement* el = cur_array->element(k); 1024 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1025 } 1026 cur_array->print_on_2(tty); 1027 } 1028 guarantee(false, "wrong number of expression stack elements during deopt"); 1029 } 1030 VerifyOopClosure verify; 1031 iframe->oops_interpreted_do(&verify, &rm, false); 1032 callee_size_of_parameters = mh->size_of_parameters(); 1033 callee_max_locals = mh->max_locals(); 1034 is_top_frame = false; 1035 } 1036 } 1037 #endif // !PRODUCT 1038 1039 return bt; 1040 JRT_END 1041 1042 class DeoptimizeMarkedClosure : public HandshakeClosure { 1043 public: 1044 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 1045 void do_thread(Thread* thread) { 1046 JavaThread* jt = JavaThread::cast(thread); 1047 jt->deoptimize_marked_methods(); 1048 } 1049 }; 1050 1051 void Deoptimization::deoptimize_all_marked() { 1052 ResourceMark rm; 1053 1054 // Make the dependent methods not entrant 1055 CodeCache::make_marked_nmethods_deoptimized(); 1056 1057 DeoptimizeMarkedClosure deopt; 1058 if (SafepointSynchronize::is_at_safepoint()) { 1059 Threads::java_threads_do(&deopt); 1060 } else { 1061 Handshake::execute(&deopt); 1062 } 1063 } 1064 1065 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1066 = Deoptimization::Action_reinterpret; 1067 1068 #if INCLUDE_JVMCI 1069 template<typename CacheType> 1070 class BoxCacheBase : public CHeapObj<mtCompiler> { 1071 protected: 1072 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 1073 ResourceMark rm(thread); 1074 char* klass_name_str = klass_name->as_C_string(); 1075 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle()); 1076 guarantee(ik != nullptr, "%s must be loaded", klass_name_str); 1077 if (!ik->is_in_error_state()) { 1078 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 1079 CacheType::compute_offsets(ik); 1080 } 1081 return ik; 1082 } 1083 }; 1084 1085 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 1086 PrimitiveType _low; 1087 PrimitiveType _high; 1088 jobject _cache; 1089 protected: 1090 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 1091 BoxCache(Thread* thread) { 1092 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 1093 if (ik->is_in_error_state()) { 1094 _low = 1; 1095 _high = 0; 1096 _cache = nullptr; 1097 } else { 1098 objArrayOop cache = CacheType::cache(ik); 1099 assert(cache->length() > 0, "Empty cache"); 1100 _low = BoxType::value(cache->obj_at(0)); 1101 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1); 1102 _cache = JNIHandles::make_global(Handle(thread, cache)); 1103 } 1104 } 1105 ~BoxCache() { 1106 JNIHandles::destroy_global(_cache); 1107 } 1108 public: 1109 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1110 if (_singleton == nullptr) { 1111 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1112 if (!Atomic::replace_if_null(&_singleton, s)) { 1113 delete s; 1114 } 1115 } 1116 return _singleton; 1117 } 1118 oop lookup(PrimitiveType value) { 1119 if (_low <= value && value <= _high) { 1120 int offset = checked_cast<int>(value - _low); 1121 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1122 } 1123 return nullptr; 1124 } 1125 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) { 1126 if (_cache == nullptr) { 1127 cache_init_error = true; 1128 return nullptr; 1129 } 1130 // Have to cast to avoid little/big-endian problems. 1131 if (sizeof(PrimitiveType) > sizeof(jint)) { 1132 jlong value = (jlong)raw_value; 1133 return lookup(value); 1134 } 1135 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1136 return lookup(value); 1137 } 1138 }; 1139 1140 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1141 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1142 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1143 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1144 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1145 1146 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr; 1147 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr; 1148 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr; 1149 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr; 1150 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr; 1151 1152 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1153 jobject _true_cache; 1154 jobject _false_cache; 1155 protected: 1156 static BooleanBoxCache *_singleton; 1157 BooleanBoxCache(Thread *thread) { 1158 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1159 if (ik->is_in_error_state()) { 1160 _true_cache = nullptr; 1161 _false_cache = nullptr; 1162 } else { 1163 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1164 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1165 } 1166 } 1167 ~BooleanBoxCache() { 1168 JNIHandles::destroy_global(_true_cache); 1169 JNIHandles::destroy_global(_false_cache); 1170 } 1171 public: 1172 static BooleanBoxCache* singleton(Thread* thread) { 1173 if (_singleton == nullptr) { 1174 BooleanBoxCache* s = new BooleanBoxCache(thread); 1175 if (!Atomic::replace_if_null(&_singleton, s)) { 1176 delete s; 1177 } 1178 } 1179 return _singleton; 1180 } 1181 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) { 1182 if (_true_cache == nullptr) { 1183 cache_in_error = true; 1184 return nullptr; 1185 } 1186 // Have to cast to avoid little/big-endian problems. 1187 jboolean value = (jboolean)*((jint*)&raw_value); 1188 return lookup(value); 1189 } 1190 oop lookup(jboolean value) { 1191 if (value != 0) { 1192 return JNIHandles::resolve_non_null(_true_cache); 1193 } 1194 return JNIHandles::resolve_non_null(_false_cache); 1195 } 1196 }; 1197 1198 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr; 1199 1200 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) { 1201 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1202 BasicType box_type = vmClasses::box_klass_type(k); 1203 if (box_type != T_OBJECT) { 1204 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1205 switch(box_type) { 1206 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1207 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1208 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1209 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1210 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1211 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1212 default:; 1213 } 1214 } 1215 return nullptr; 1216 } 1217 #endif // INCLUDE_JVMCI 1218 1219 #if COMPILER2_OR_JVMCI 1220 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1221 Handle pending_exception(THREAD, thread->pending_exception()); 1222 const char* exception_file = thread->exception_file(); 1223 int exception_line = thread->exception_line(); 1224 thread->clear_pending_exception(); 1225 1226 bool failures = false; 1227 1228 for (int i = 0; i < objects->length(); i++) { 1229 assert(objects->at(i)->is_object(), "invalid debug information"); 1230 ObjectValue* sv = (ObjectValue*) objects->at(i); 1231 1232 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1233 oop obj = nullptr; 1234 1235 bool cache_init_error = false; 1236 if (k->is_instance_klass()) { 1237 #if INCLUDE_JVMCI 1238 nmethod* nm = fr->cb()->as_nmethod_or_null(); 1239 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1240 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1241 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD); 1242 if (obj != nullptr) { 1243 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1244 abv->set_cached(true); 1245 } else if (cache_init_error) { 1246 // Results in an OOME which is valid (as opposed to a class initialization error) 1247 // and is fine for the rare case a cache initialization failing. 1248 failures = true; 1249 } 1250 } 1251 #endif // INCLUDE_JVMCI 1252 1253 InstanceKlass* ik = InstanceKlass::cast(k); 1254 if (obj == nullptr && !cache_init_error) { 1255 #if COMPILER2_OR_JVMCI 1256 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1257 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1258 } else { 1259 obj = ik->allocate_instance(THREAD); 1260 } 1261 #else 1262 obj = ik->allocate_instance(THREAD); 1263 #endif // COMPILER2_OR_JVMCI 1264 } 1265 } else if (k->is_typeArray_klass()) { 1266 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1267 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1268 int len = sv->field_size() / type2size[ak->element_type()]; 1269 obj = ak->allocate(len, THREAD); 1270 } else if (k->is_objArray_klass()) { 1271 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1272 obj = ak->allocate(sv->field_size(), THREAD); 1273 } 1274 1275 if (obj == nullptr) { 1276 failures = true; 1277 } 1278 1279 assert(sv->value().is_null(), "redundant reallocation"); 1280 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception"); 1281 CLEAR_PENDING_EXCEPTION; 1282 sv->set_value(obj); 1283 } 1284 1285 if (failures) { 1286 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1287 } else if (pending_exception.not_null()) { 1288 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1289 } 1290 1291 return failures; 1292 } 1293 1294 #if INCLUDE_JVMCI 1295 /** 1296 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1297 * we need to somehow be able to recover the actual kind to be able to write the correct 1298 * amount of bytes. 1299 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1300 * the entries at index n + 1 to n + i are 'markers'. 1301 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1302 * expected form of the array would be: 1303 * 1304 * {b0, b1, b2, b3, INT, marker, b6, b7} 1305 * 1306 * Thus, in order to get back the size of the entry, we simply need to count the number 1307 * of marked entries 1308 * 1309 * @param virtualArray the virtualized byte array 1310 * @param i index of the virtual entry we are recovering 1311 * @return The number of bytes the entry spans 1312 */ 1313 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1314 int index = i; 1315 while (++index < virtualArray->field_size() && 1316 virtualArray->field_at(index)->is_marker()) {} 1317 return index - i; 1318 } 1319 1320 /** 1321 * If there was a guarantee for byte array to always start aligned to a long, we could 1322 * do a simple check on the parity of the index. Unfortunately, that is not always the 1323 * case. Thus, we check alignment of the actual address we are writing to. 1324 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1325 * write a long to index 3. 1326 */ 1327 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1328 jbyte* res = obj->byte_at_addr(index); 1329 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1330 return res; 1331 } 1332 1333 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) { 1334 switch (byte_count) { 1335 case 1: 1336 obj->byte_at_put(index, (jbyte) value->get_jint()); 1337 break; 1338 case 2: 1339 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint(); 1340 break; 1341 case 4: 1342 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint(); 1343 break; 1344 case 8: 1345 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr(); 1346 break; 1347 default: 1348 ShouldNotReachHere(); 1349 } 1350 } 1351 #endif // INCLUDE_JVMCI 1352 1353 1354 // restore elements of an eliminated type array 1355 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1356 int index = 0; 1357 1358 for (int i = 0; i < sv->field_size(); i++) { 1359 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1360 switch(type) { 1361 case T_LONG: case T_DOUBLE: { 1362 assert(value->type() == T_INT, "Agreement."); 1363 StackValue* low = 1364 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1365 #ifdef _LP64 1366 jlong res = (jlong)low->get_intptr(); 1367 #else 1368 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1369 #endif 1370 obj->long_at_put(index, res); 1371 break; 1372 } 1373 1374 case T_INT: case T_FLOAT: { // 4 bytes. 1375 assert(value->type() == T_INT, "Agreement."); 1376 bool big_value = false; 1377 if (i + 1 < sv->field_size() && type == T_INT) { 1378 if (sv->field_at(i)->is_location()) { 1379 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1380 if (type == Location::dbl || type == Location::lng) { 1381 big_value = true; 1382 } 1383 } else if (sv->field_at(i)->is_constant_int()) { 1384 ScopeValue* next_scope_field = sv->field_at(i + 1); 1385 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1386 big_value = true; 1387 } 1388 } 1389 } 1390 1391 if (big_value) { 1392 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1393 #ifdef _LP64 1394 jlong res = (jlong)low->get_intptr(); 1395 #else 1396 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1397 #endif 1398 obj->int_at_put(index, *(jint*)&res); 1399 obj->int_at_put(++index, *((jint*)&res + 1)); 1400 } else { 1401 obj->int_at_put(index, value->get_jint()); 1402 } 1403 break; 1404 } 1405 1406 case T_SHORT: 1407 assert(value->type() == T_INT, "Agreement."); 1408 obj->short_at_put(index, (jshort)value->get_jint()); 1409 break; 1410 1411 case T_CHAR: 1412 assert(value->type() == T_INT, "Agreement."); 1413 obj->char_at_put(index, (jchar)value->get_jint()); 1414 break; 1415 1416 case T_BYTE: { 1417 assert(value->type() == T_INT, "Agreement."); 1418 #if INCLUDE_JVMCI 1419 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1420 int byte_count = count_number_of_bytes_for_entry(sv, i); 1421 byte_array_put(obj, value, index, byte_count); 1422 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1423 i += byte_count - 1; // Balance the loop counter. 1424 index += byte_count; 1425 // index has been updated so continue at top of loop 1426 continue; 1427 #else 1428 obj->byte_at_put(index, (jbyte)value->get_jint()); 1429 break; 1430 #endif // INCLUDE_JVMCI 1431 } 1432 1433 case T_BOOLEAN: { 1434 assert(value->type() == T_INT, "Agreement."); 1435 obj->bool_at_put(index, (jboolean)value->get_jint()); 1436 break; 1437 } 1438 1439 default: 1440 ShouldNotReachHere(); 1441 } 1442 index++; 1443 } 1444 } 1445 1446 // restore fields of an eliminated object array 1447 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1448 for (int i = 0; i < sv->field_size(); i++) { 1449 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1450 assert(value->type() == T_OBJECT, "object element expected"); 1451 obj->obj_at_put(i, value->get_obj()()); 1452 } 1453 } 1454 1455 class ReassignedField { 1456 public: 1457 int _offset; 1458 BasicType _type; 1459 public: 1460 ReassignedField() { 1461 _offset = 0; 1462 _type = T_ILLEGAL; 1463 } 1464 }; 1465 1466 static int compare(ReassignedField* left, ReassignedField* right) { 1467 return left->_offset - right->_offset; 1468 } 1469 1470 // Restore fields of an eliminated instance object using the same field order 1471 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1472 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1473 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1474 InstanceKlass* ik = klass; 1475 while (ik != nullptr) { 1476 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1477 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) { 1478 ReassignedField field; 1479 field._offset = fs.offset(); 1480 field._type = Signature::basic_type(fs.signature()); 1481 fields->append(field); 1482 } 1483 } 1484 ik = ik->superklass(); 1485 } 1486 fields->sort(compare); 1487 for (int i = 0; i < fields->length(); i++) { 1488 ScopeValue* scope_field = sv->field_at(svIndex); 1489 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1490 int offset = fields->at(i)._offset; 1491 BasicType type = fields->at(i)._type; 1492 switch (type) { 1493 case T_OBJECT: case T_ARRAY: 1494 assert(value->type() == T_OBJECT, "Agreement."); 1495 obj->obj_field_put(offset, value->get_obj()()); 1496 break; 1497 1498 case T_INT: case T_FLOAT: { // 4 bytes. 1499 assert(value->type() == T_INT, "Agreement."); 1500 bool big_value = false; 1501 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1502 if (scope_field->is_location()) { 1503 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1504 if (type == Location::dbl || type == Location::lng) { 1505 big_value = true; 1506 } 1507 } 1508 if (scope_field->is_constant_int()) { 1509 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1510 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1511 big_value = true; 1512 } 1513 } 1514 } 1515 1516 if (big_value) { 1517 i++; 1518 assert(i < fields->length(), "second T_INT field needed"); 1519 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1520 } else { 1521 obj->int_field_put(offset, value->get_jint()); 1522 break; 1523 } 1524 } 1525 /* no break */ 1526 1527 case T_LONG: case T_DOUBLE: { 1528 assert(value->type() == T_INT, "Agreement."); 1529 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1530 #ifdef _LP64 1531 jlong res = (jlong)low->get_intptr(); 1532 #else 1533 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1534 #endif 1535 obj->long_field_put(offset, res); 1536 break; 1537 } 1538 1539 case T_SHORT: 1540 assert(value->type() == T_INT, "Agreement."); 1541 obj->short_field_put(offset, (jshort)value->get_jint()); 1542 break; 1543 1544 case T_CHAR: 1545 assert(value->type() == T_INT, "Agreement."); 1546 obj->char_field_put(offset, (jchar)value->get_jint()); 1547 break; 1548 1549 case T_BYTE: 1550 assert(value->type() == T_INT, "Agreement."); 1551 obj->byte_field_put(offset, (jbyte)value->get_jint()); 1552 break; 1553 1554 case T_BOOLEAN: 1555 assert(value->type() == T_INT, "Agreement."); 1556 obj->bool_field_put(offset, (jboolean)value->get_jint()); 1557 break; 1558 1559 default: 1560 ShouldNotReachHere(); 1561 } 1562 svIndex++; 1563 } 1564 return svIndex; 1565 } 1566 1567 // restore fields of all eliminated objects and arrays 1568 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1569 for (int i = 0; i < objects->length(); i++) { 1570 assert(objects->at(i)->is_object(), "invalid debug information"); 1571 ObjectValue* sv = (ObjectValue*) objects->at(i); 1572 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1573 Handle obj = sv->value(); 1574 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1575 #ifndef PRODUCT 1576 if (PrintDeoptimizationDetails) { 1577 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1578 } 1579 #endif // !PRODUCT 1580 1581 if (obj.is_null()) { 1582 continue; 1583 } 1584 1585 #if INCLUDE_JVMCI 1586 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1587 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1588 continue; 1589 } 1590 #endif // INCLUDE_JVMCI 1591 #if COMPILER2_OR_JVMCI 1592 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1593 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1594 ScopeValue* payload = sv->field_at(0); 1595 if (payload->is_location() && 1596 payload->as_LocationValue()->location().type() == Location::vector) { 1597 #ifndef PRODUCT 1598 if (PrintDeoptimizationDetails) { 1599 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1600 if (Verbose) { 1601 Handle obj = sv->value(); 1602 k->oop_print_on(obj(), tty); 1603 } 1604 } 1605 #endif // !PRODUCT 1606 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1607 } 1608 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1609 // which could be restored after vector object allocation. 1610 } 1611 #endif /* !COMPILER2_OR_JVMCI */ 1612 if (k->is_instance_klass()) { 1613 InstanceKlass* ik = InstanceKlass::cast(k); 1614 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1615 } else if (k->is_typeArray_klass()) { 1616 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1617 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1618 } else if (k->is_objArray_klass()) { 1619 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1620 } 1621 } 1622 // These objects may escape when we return to Interpreter after deoptimization. 1623 // We need barrier so that stores that initialize these objects can't be reordered 1624 // with subsequent stores that make these objects accessible by other threads. 1625 OrderAccess::storestore(); 1626 } 1627 1628 1629 // relock objects for which synchronization was eliminated 1630 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1631 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1632 bool relocked_objects = false; 1633 for (int i = 0; i < monitors->length(); i++) { 1634 MonitorInfo* mon_info = monitors->at(i); 1635 if (mon_info->eliminated()) { 1636 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1637 relocked_objects = true; 1638 if (!mon_info->owner_is_scalar_replaced()) { 1639 Handle obj(thread, mon_info->owner()); 1640 markWord mark = obj->mark(); 1641 if (exec_mode == Unpack_none) { 1642 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) { 1643 // With exec_mode == Unpack_none obj may be thread local and locked in 1644 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header. 1645 markWord dmw = mark.displaced_mark_helper(); 1646 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr)); 1647 obj->set_mark(dmw); 1648 } 1649 if (mark.has_monitor()) { 1650 // defer relocking if the deoptee thread is currently waiting for obj 1651 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1652 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) { 1653 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1654 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1655 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1656 continue; 1657 } 1658 } 1659 } 1660 BasicLock* lock = mon_info->lock(); 1661 if (LockingMode == LM_LIGHTWEIGHT) { 1662 // We have lost information about the correct state of the lock stack. 1663 // Entering may create an invalid lock stack. Inflate the lock if it 1664 // was fast_locked to restore the valid lock stack. 1665 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread); 1666 if (deoptee_thread->lock_stack().contains(obj())) { 1667 LightweightSynchronizer::inflate_fast_locked_object(obj(), deoptee_thread, thread, 1668 ObjectSynchronizer::InflateCause::inflate_cause_vm_internal); 1669 } 1670 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1671 assert(obj->mark().has_monitor(), "must be"); 1672 assert(!deoptee_thread->lock_stack().contains(obj()), "must be"); 1673 assert(LightweightSynchronizer::read_monitor(thread, obj())->owner() == deoptee_thread, "must be"); 1674 } else { 1675 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread); 1676 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1677 } 1678 } 1679 } 1680 } 1681 return relocked_objects; 1682 } 1683 #endif // COMPILER2_OR_JVMCI 1684 1685 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1686 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1687 1688 // Register map for next frame (used for stack crawl). We capture 1689 // the state of the deopt'ing frame's caller. Thus if we need to 1690 // stuff a C2I adapter we can properly fill in the callee-save 1691 // register locations. 1692 frame caller = fr.sender(reg_map); 1693 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp()); 1694 1695 frame sender = caller; 1696 1697 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1698 // the vframeArray containing the unpacking information is allocated in the C heap. 1699 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1700 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1701 1702 // Compare the vframeArray to the collected vframes 1703 assert(array->structural_compare(thread, chunk), "just checking"); 1704 1705 if (TraceDeoptimization) { 1706 ResourceMark rm; 1707 stringStream st; 1708 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1709 st.print(" "); 1710 fr.print_on(&st); 1711 st.print_cr(" Virtual frames (innermost/newest first):"); 1712 for (int index = 0; index < chunk->length(); index++) { 1713 compiledVFrame* vf = chunk->at(index); 1714 int bci = vf->raw_bci(); 1715 const char* code_name; 1716 if (bci == SynchronizationEntryBCI) { 1717 code_name = "sync entry"; 1718 } else { 1719 Bytecodes::Code code = vf->method()->code_at(bci); 1720 code_name = Bytecodes::name(code); 1721 } 1722 1723 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1724 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1725 st.print(" - %s", code_name); 1726 st.print_cr(" @ bci=%d ", bci); 1727 } 1728 tty->print_raw(st.freeze()); 1729 tty->cr(); 1730 } 1731 1732 return array; 1733 } 1734 1735 #if COMPILER2_OR_JVMCI 1736 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1737 // Reallocation of some scalar replaced objects failed. Record 1738 // that we need to pop all the interpreter frames for the 1739 // deoptimized compiled frame. 1740 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1741 thread->set_frames_to_pop_failed_realloc(array->frames()); 1742 // Unlock all monitors here otherwise the interpreter will see a 1743 // mix of locked and unlocked monitors (because of failed 1744 // reallocations of synchronized objects) and be confused. 1745 for (int i = 0; i < array->frames(); i++) { 1746 MonitorChunk* monitors = array->element(i)->monitors(); 1747 if (monitors != nullptr) { 1748 // Unlock in reverse order starting from most nested monitor. 1749 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) { 1750 BasicObjectLock* src = monitors->at(j); 1751 if (src->obj() != nullptr) { 1752 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1753 } 1754 } 1755 array->element(i)->free_monitors(thread); 1756 #ifdef ASSERT 1757 array->element(i)->set_removed_monitors(); 1758 #endif 1759 } 1760 } 1761 } 1762 #endif 1763 1764 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1765 assert(fr.can_be_deoptimized(), "checking frame type"); 1766 1767 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1768 1769 if (LogCompilation && xtty != nullptr) { 1770 nmethod* nm = fr.cb()->as_nmethod_or_null(); 1771 assert(nm != nullptr, "only compiled methods can deopt"); 1772 1773 ttyLocker ttyl; 1774 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1775 nm->log_identity(xtty); 1776 xtty->end_head(); 1777 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1778 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1779 xtty->method(sd->method()); 1780 xtty->end_elem(); 1781 if (sd->is_top()) break; 1782 } 1783 xtty->tail("deoptimized"); 1784 } 1785 1786 Continuation::notify_deopt(thread, fr.sp()); 1787 1788 // Patch the compiled method so that when execution returns to it we will 1789 // deopt the execution state and return to the interpreter. 1790 fr.deoptimize(thread); 1791 } 1792 1793 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1794 // Deoptimize only if the frame comes from compile code. 1795 // Do not deoptimize the frame which is already patched 1796 // during the execution of the loops below. 1797 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1798 return; 1799 } 1800 ResourceMark rm; 1801 deoptimize_single_frame(thread, fr, reason); 1802 } 1803 1804 #if INCLUDE_JVMCI 1805 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) { 1806 // there is no exception handler for this pc => deoptimize 1807 nm->make_not_entrant(); 1808 1809 // Use Deoptimization::deoptimize for all of its side-effects: 1810 // gathering traps statistics, logging... 1811 // it also patches the return pc but we do not care about that 1812 // since we return a continuation to the deopt_blob below. 1813 JavaThread* thread = JavaThread::current(); 1814 RegisterMap reg_map(thread, 1815 RegisterMap::UpdateMap::skip, 1816 RegisterMap::ProcessFrames::include, 1817 RegisterMap::WalkContinuation::skip); 1818 frame runtime_frame = thread->last_frame(); 1819 frame caller_frame = runtime_frame.sender(®_map); 1820 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method"); 1821 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1822 compiledVFrame* cvf = compiledVFrame::cast(vf); 1823 ScopeDesc* imm_scope = cvf->scope(); 1824 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1825 if (imm_mdo != nullptr) { 1826 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 1827 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 1828 1829 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr); 1830 if (pdata != nullptr && pdata->is_BitData()) { 1831 BitData* bit_data = (BitData*) pdata; 1832 bit_data->set_exception_seen(); 1833 } 1834 } 1835 1836 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1837 1838 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true); 1839 if (trap_mdo != nullptr) { 1840 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1841 } 1842 1843 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1844 } 1845 #endif 1846 1847 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1848 assert(thread == Thread::current() || 1849 thread->is_handshake_safe_for(Thread::current()) || 1850 SafepointSynchronize::is_at_safepoint(), 1851 "can only deoptimize other thread at a safepoint/handshake"); 1852 // Compute frame and register map based on thread and sp. 1853 RegisterMap reg_map(thread, 1854 RegisterMap::UpdateMap::skip, 1855 RegisterMap::ProcessFrames::include, 1856 RegisterMap::WalkContinuation::skip); 1857 frame fr = thread->last_frame(); 1858 while (fr.id() != id) { 1859 fr = fr.sender(®_map); 1860 } 1861 deoptimize(thread, fr, reason); 1862 } 1863 1864 1865 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1866 Thread* current = Thread::current(); 1867 if (thread == current || thread->is_handshake_safe_for(current)) { 1868 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1869 } else { 1870 VM_DeoptimizeFrame deopt(thread, id, reason); 1871 VMThread::execute(&deopt); 1872 } 1873 } 1874 1875 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1876 deoptimize_frame(thread, id, Reason_constraint); 1877 } 1878 1879 // JVMTI PopFrame support 1880 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1881 { 1882 assert(thread == JavaThread::current(), "pre-condition"); 1883 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1884 } 1885 JRT_END 1886 1887 MethodData* 1888 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1889 bool create_if_missing) { 1890 JavaThread* THREAD = thread; // For exception macros. 1891 MethodData* mdo = m()->method_data(); 1892 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) { 1893 // Build an MDO. Ignore errors like OutOfMemory; 1894 // that simply means we won't have an MDO to update. 1895 Method::build_profiling_method_data(m, THREAD); 1896 if (HAS_PENDING_EXCEPTION) { 1897 // Only metaspace OOM is expected. No Java code executed. 1898 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1899 CLEAR_PENDING_EXCEPTION; 1900 } 1901 mdo = m()->method_data(); 1902 } 1903 return mdo; 1904 } 1905 1906 #if COMPILER2_OR_JVMCI 1907 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1908 // In case of an unresolved klass entry, load the class. 1909 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1910 // and probably nowhere else. 1911 // Even that case would benefit from simply re-interpreting the 1912 // bytecode, without paying special attention to the class index. 1913 // So this whole "class index" feature should probably be removed. 1914 1915 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1916 Klass* tk = constant_pool->klass_at(index, THREAD); 1917 if (HAS_PENDING_EXCEPTION) { 1918 // Exception happened during classloading. We ignore the exception here, since it 1919 // is going to be rethrown since the current activation is going to be deoptimized and 1920 // the interpreter will re-execute the bytecode. 1921 // Do not clear probable Async Exceptions. 1922 CLEAR_PENDING_NONASYNC_EXCEPTION; 1923 // Class loading called java code which may have caused a stack 1924 // overflow. If the exception was thrown right before the return 1925 // to the runtime the stack is no longer guarded. Reguard the 1926 // stack otherwise if we return to the uncommon trap blob and the 1927 // stack bang causes a stack overflow we crash. 1928 JavaThread* jt = THREAD; 1929 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 1930 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1931 } 1932 return; 1933 } 1934 1935 assert(!constant_pool->tag_at(index).is_symbol(), 1936 "no symbolic names here, please"); 1937 } 1938 1939 #if INCLUDE_JFR 1940 1941 class DeoptReasonSerializer : public JfrSerializer { 1942 public: 1943 void serialize(JfrCheckpointWriter& writer) { 1944 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 1945 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 1946 writer.write_key((u8)i); 1947 writer.write(Deoptimization::trap_reason_name(i)); 1948 } 1949 } 1950 }; 1951 1952 class DeoptActionSerializer : public JfrSerializer { 1953 public: 1954 void serialize(JfrCheckpointWriter& writer) { 1955 static const u4 nof_actions = Deoptimization::Action_LIMIT; 1956 writer.write_count(nof_actions); 1957 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 1958 writer.write_key(i); 1959 writer.write(Deoptimization::trap_action_name((int)i)); 1960 } 1961 } 1962 }; 1963 1964 static void register_serializers() { 1965 static int critical_section = 0; 1966 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 1967 return; 1968 } 1969 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 1970 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 1971 } 1972 1973 static void post_deoptimization_event(nmethod* nm, 1974 const Method* method, 1975 int trap_bci, 1976 int instruction, 1977 Deoptimization::DeoptReason reason, 1978 Deoptimization::DeoptAction action) { 1979 assert(nm != nullptr, "invariant"); 1980 assert(method != nullptr, "invariant"); 1981 if (EventDeoptimization::is_enabled()) { 1982 static bool serializers_registered = false; 1983 if (!serializers_registered) { 1984 register_serializers(); 1985 serializers_registered = true; 1986 } 1987 EventDeoptimization event; 1988 event.set_compileId(nm->compile_id()); 1989 event.set_compiler(nm->compiler_type()); 1990 event.set_method(method); 1991 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 1992 event.set_bci(trap_bci); 1993 event.set_instruction(instruction); 1994 event.set_reason(reason); 1995 event.set_action(action); 1996 event.commit(); 1997 } 1998 } 1999 2000 #endif // INCLUDE_JFR 2001 2002 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 2003 const char* reason_name, const char* reason_action) { 2004 LogTarget(Debug, deoptimization) lt; 2005 if (lt.is_enabled()) { 2006 LogStream ls(lt); 2007 bool is_osr = nm->is_osr_method(); 2008 ls.print("cid=%4d %s level=%d", 2009 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level()); 2010 ls.print(" %s", tm->name_and_sig_as_C_string()); 2011 ls.print(" trap_bci=%d ", trap_bci); 2012 if (is_osr) { 2013 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 2014 } 2015 ls.print("%s ", reason_name); 2016 ls.print("%s ", reason_action); 2017 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 2018 pc, fr.pc() - nm->code_begin()); 2019 } 2020 } 2021 2022 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 2023 HandleMark hm(current); 2024 2025 // uncommon_trap() is called at the beginning of the uncommon trap 2026 // handler. Note this fact before we start generating temporary frames 2027 // that can confuse an asynchronous stack walker. This counter is 2028 // decremented at the end of unpack_frames(). 2029 2030 current->inc_in_deopt_handler(); 2031 2032 #if INCLUDE_JVMCI 2033 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 2034 RegisterMap reg_map(current, 2035 RegisterMap::UpdateMap::include, 2036 RegisterMap::ProcessFrames::include, 2037 RegisterMap::WalkContinuation::skip); 2038 #else 2039 RegisterMap reg_map(current, 2040 RegisterMap::UpdateMap::skip, 2041 RegisterMap::ProcessFrames::include, 2042 RegisterMap::WalkContinuation::skip); 2043 #endif 2044 frame stub_frame = current->last_frame(); 2045 frame fr = stub_frame.sender(®_map); 2046 2047 // Log a message 2048 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 2049 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 2050 2051 { 2052 ResourceMark rm; 2053 2054 DeoptReason reason = trap_request_reason(trap_request); 2055 DeoptAction action = trap_request_action(trap_request); 2056 #if INCLUDE_JVMCI 2057 int debug_id = trap_request_debug_id(trap_request); 2058 #endif 2059 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 2060 2061 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 2062 compiledVFrame* cvf = compiledVFrame::cast(vf); 2063 2064 nmethod* nm = cvf->code(); 2065 2066 ScopeDesc* trap_scope = cvf->scope(); 2067 2068 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2069 2070 if (is_receiver_constraint_failure) { 2071 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2072 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2073 JVMCI_ONLY(COMMA debug_id)); 2074 } 2075 2076 methodHandle trap_method(current, trap_scope->method()); 2077 int trap_bci = trap_scope->bci(); 2078 #if INCLUDE_JVMCI 2079 jlong speculation = current->pending_failed_speculation(); 2080 if (nm->is_compiled_by_jvmci()) { 2081 nm->update_speculation(current); 2082 } else { 2083 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2084 } 2085 2086 if (trap_bci == SynchronizationEntryBCI) { 2087 trap_bci = 0; 2088 current->set_pending_monitorenter(true); 2089 } 2090 2091 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2092 current->set_pending_transfer_to_interpreter(true); 2093 } 2094 #endif 2095 2096 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2097 // Record this event in the histogram. 2098 gather_statistics(reason, action, trap_bc); 2099 2100 // Ensure that we can record deopt. history: 2101 // Need MDO to record RTM code generation state. 2102 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 2103 2104 methodHandle profiled_method; 2105 #if INCLUDE_JVMCI 2106 if (nm->is_compiled_by_jvmci()) { 2107 profiled_method = methodHandle(current, nm->method()); 2108 } else { 2109 profiled_method = trap_method; 2110 } 2111 #else 2112 profiled_method = trap_method; 2113 #endif 2114 2115 MethodData* trap_mdo = 2116 get_method_data(current, profiled_method, create_if_missing); 2117 2118 { // Log Deoptimization event for JFR, UL and event system 2119 Method* tm = trap_method(); 2120 const char* reason_name = trap_reason_name(reason); 2121 const char* reason_action = trap_action_name(action); 2122 intptr_t pc = p2i(fr.pc()); 2123 2124 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2125 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2126 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 2127 reason_name, reason_action, pc, 2128 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 2129 } 2130 2131 // Print a bunch of diagnostics, if requested. 2132 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2133 ResourceMark rm; 2134 2135 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2136 // We must do this already now, since we cannot acquire this lock while 2137 // holding the tty lock (lock ordering by rank). 2138 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2139 2140 ttyLocker ttyl; 2141 2142 char buf[100]; 2143 if (xtty != nullptr) { 2144 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 2145 os::current_thread_id(), 2146 format_trap_request(buf, sizeof(buf), trap_request)); 2147 #if INCLUDE_JVMCI 2148 if (speculation != 0) { 2149 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2150 } 2151 #endif 2152 nm->log_identity(xtty); 2153 } 2154 Symbol* class_name = nullptr; 2155 bool unresolved = false; 2156 if (unloaded_class_index >= 0) { 2157 constantPoolHandle constants (current, trap_method->constants()); 2158 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2159 class_name = constants->klass_name_at(unloaded_class_index); 2160 unresolved = true; 2161 if (xtty != nullptr) 2162 xtty->print(" unresolved='1'"); 2163 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2164 class_name = constants->symbol_at(unloaded_class_index); 2165 } 2166 if (xtty != nullptr) 2167 xtty->name(class_name); 2168 } 2169 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) { 2170 // Dump the relevant MDO state. 2171 // This is the deopt count for the current reason, any previous 2172 // reasons or recompiles seen at this point. 2173 int dcnt = trap_mdo->trap_count(reason); 2174 if (dcnt != 0) 2175 xtty->print(" count='%d'", dcnt); 2176 2177 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to 2178 // lock extra_data_lock before the tty lock. 2179 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2180 int dos = (pdata == nullptr)? 0: pdata->trap_state(); 2181 if (dos != 0) { 2182 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2183 if (trap_state_is_recompiled(dos)) { 2184 int recnt2 = trap_mdo->overflow_recompile_count(); 2185 if (recnt2 != 0) 2186 xtty->print(" recompiles2='%d'", recnt2); 2187 } 2188 } 2189 } 2190 if (xtty != nullptr) { 2191 xtty->stamp(); 2192 xtty->end_head(); 2193 } 2194 if (TraceDeoptimization) { // make noise on the tty 2195 stringStream st; 2196 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2197 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2198 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2199 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 2200 #if INCLUDE_JVMCI 2201 if (nm->is_compiled_by_jvmci()) { 2202 const char* installed_code_name = nm->jvmci_name(); 2203 if (installed_code_name != nullptr) { 2204 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2205 } 2206 } 2207 #endif 2208 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2209 p2i(fr.pc()), 2210 os::current_thread_id(), 2211 trap_reason_name(reason), 2212 trap_action_name(action), 2213 unloaded_class_index 2214 #if INCLUDE_JVMCI 2215 , debug_id 2216 #endif 2217 ); 2218 if (class_name != nullptr) { 2219 st.print(unresolved ? " unresolved class: " : " symbol: "); 2220 class_name->print_symbol_on(&st); 2221 } 2222 st.cr(); 2223 tty->print_raw(st.freeze()); 2224 } 2225 if (xtty != nullptr) { 2226 // Log the precise location of the trap. 2227 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2228 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2229 xtty->method(sd->method()); 2230 xtty->end_elem(); 2231 if (sd->is_top()) break; 2232 } 2233 xtty->tail("uncommon_trap"); 2234 } 2235 } 2236 // (End diagnostic printout.) 2237 2238 if (is_receiver_constraint_failure) { 2239 fatal("missing receiver type check"); 2240 } 2241 2242 // Load class if necessary 2243 if (unloaded_class_index >= 0) { 2244 constantPoolHandle constants(current, trap_method->constants()); 2245 load_class_by_index(constants, unloaded_class_index, THREAD); 2246 } 2247 2248 // Flush the nmethod if necessary and desirable. 2249 // 2250 // We need to avoid situations where we are re-flushing the nmethod 2251 // because of a hot deoptimization site. Repeated flushes at the same 2252 // point need to be detected by the compiler and avoided. If the compiler 2253 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2254 // module must take measures to avoid an infinite cycle of recompilation 2255 // and deoptimization. There are several such measures: 2256 // 2257 // 1. If a recompilation is ordered a second time at some site X 2258 // and for the same reason R, the action is adjusted to 'reinterpret', 2259 // to give the interpreter time to exercise the method more thoroughly. 2260 // If this happens, the method's overflow_recompile_count is incremented. 2261 // 2262 // 2. If the compiler fails to reduce the deoptimization rate, then 2263 // the method's overflow_recompile_count will begin to exceed the set 2264 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2265 // is adjusted to 'make_not_compilable', and the method is abandoned 2266 // to the interpreter. This is a performance hit for hot methods, 2267 // but is better than a disastrous infinite cycle of recompilations. 2268 // (Actually, only the method containing the site X is abandoned.) 2269 // 2270 // 3. In parallel with the previous measures, if the total number of 2271 // recompilations of a method exceeds the much larger set limit 2272 // PerMethodRecompilationCutoff, the method is abandoned. 2273 // This should only happen if the method is very large and has 2274 // many "lukewarm" deoptimizations. The code which enforces this 2275 // limit is elsewhere (class nmethod, class Method). 2276 // 2277 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2278 // to recompile at each bytecode independently of the per-BCI cutoff. 2279 // 2280 // The decision to update code is up to the compiler, and is encoded 2281 // in the Action_xxx code. If the compiler requests Action_none 2282 // no trap state is changed, no compiled code is changed, and the 2283 // computation suffers along in the interpreter. 2284 // 2285 // The other action codes specify various tactics for decompilation 2286 // and recompilation. Action_maybe_recompile is the loosest, and 2287 // allows the compiled code to stay around until enough traps are seen, 2288 // and until the compiler gets around to recompiling the trapping method. 2289 // 2290 // The other actions cause immediate removal of the present code. 2291 2292 // Traps caused by injected profile shouldn't pollute trap counts. 2293 bool injected_profile_trap = trap_method->has_injected_profile() && 2294 (reason == Reason_intrinsic || reason == Reason_unreached); 2295 2296 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2297 bool make_not_entrant = false; 2298 bool make_not_compilable = false; 2299 bool reprofile = false; 2300 switch (action) { 2301 case Action_none: 2302 // Keep the old code. 2303 update_trap_state = false; 2304 break; 2305 case Action_maybe_recompile: 2306 // Do not need to invalidate the present code, but we can 2307 // initiate another 2308 // Start compiler without (necessarily) invalidating the nmethod. 2309 // The system will tolerate the old code, but new code should be 2310 // generated when possible. 2311 break; 2312 case Action_reinterpret: 2313 // Go back into the interpreter for a while, and then consider 2314 // recompiling form scratch. 2315 make_not_entrant = true; 2316 // Reset invocation counter for outer most method. 2317 // This will allow the interpreter to exercise the bytecodes 2318 // for a while before recompiling. 2319 // By contrast, Action_make_not_entrant is immediate. 2320 // 2321 // Note that the compiler will track null_check, null_assert, 2322 // range_check, and class_check events and log them as if they 2323 // had been traps taken from compiled code. This will update 2324 // the MDO trap history so that the next compilation will 2325 // properly detect hot trap sites. 2326 reprofile = true; 2327 break; 2328 case Action_make_not_entrant: 2329 // Request immediate recompilation, and get rid of the old code. 2330 // Make them not entrant, so next time they are called they get 2331 // recompiled. Unloaded classes are loaded now so recompile before next 2332 // time they are called. Same for uninitialized. The interpreter will 2333 // link the missing class, if any. 2334 make_not_entrant = true; 2335 break; 2336 case Action_make_not_compilable: 2337 // Give up on compiling this method at all. 2338 make_not_entrant = true; 2339 make_not_compilable = true; 2340 break; 2341 default: 2342 ShouldNotReachHere(); 2343 } 2344 2345 // Setting +ProfileTraps fixes the following, on all platforms: 2346 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2347 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2348 // recompile relies on a MethodData* to record heroic opt failures. 2349 2350 // Whether the interpreter is producing MDO data or not, we also need 2351 // to use the MDO to detect hot deoptimization points and control 2352 // aggressive optimization. 2353 bool inc_recompile_count = false; 2354 2355 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2356 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr, 2357 (trap_mdo != nullptr), 2358 Mutex::_no_safepoint_check_flag); 2359 ProfileData* pdata = nullptr; 2360 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) { 2361 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2362 uint this_trap_count = 0; 2363 bool maybe_prior_trap = false; 2364 bool maybe_prior_recompile = false; 2365 2366 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2367 #if INCLUDE_JVMCI 2368 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2369 #endif 2370 nm->method(), 2371 //outputs: 2372 this_trap_count, 2373 maybe_prior_trap, 2374 maybe_prior_recompile); 2375 // Because the interpreter also counts null, div0, range, and class 2376 // checks, these traps from compiled code are double-counted. 2377 // This is harmless; it just means that the PerXTrapLimit values 2378 // are in effect a little smaller than they look. 2379 2380 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2381 if (per_bc_reason != Reason_none) { 2382 // Now take action based on the partially known per-BCI history. 2383 if (maybe_prior_trap 2384 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2385 // If there are too many traps at this BCI, force a recompile. 2386 // This will allow the compiler to see the limit overflow, and 2387 // take corrective action, if possible. The compiler generally 2388 // does not use the exact PerBytecodeTrapLimit value, but instead 2389 // changes its tactics if it sees any traps at all. This provides 2390 // a little hysteresis, delaying a recompile until a trap happens 2391 // several times. 2392 // 2393 // Actually, since there is only one bit of counter per BCI, 2394 // the possible per-BCI counts are {0,1,(per-method count)}. 2395 // This produces accurate results if in fact there is only 2396 // one hot trap site, but begins to get fuzzy if there are 2397 // many sites. For example, if there are ten sites each 2398 // trapping two or more times, they each get the blame for 2399 // all of their traps. 2400 make_not_entrant = true; 2401 } 2402 2403 // Detect repeated recompilation at the same BCI, and enforce a limit. 2404 if (make_not_entrant && maybe_prior_recompile) { 2405 // More than one recompile at this point. 2406 inc_recompile_count = maybe_prior_trap; 2407 } 2408 } else { 2409 // For reasons which are not recorded per-bytecode, we simply 2410 // force recompiles unconditionally. 2411 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2412 make_not_entrant = true; 2413 } 2414 2415 // Go back to the compiler if there are too many traps in this method. 2416 if (this_trap_count >= per_method_trap_limit(reason)) { 2417 // If there are too many traps in this method, force a recompile. 2418 // This will allow the compiler to see the limit overflow, and 2419 // take corrective action, if possible. 2420 // (This condition is an unlikely backstop only, because the 2421 // PerBytecodeTrapLimit is more likely to take effect first, 2422 // if it is applicable.) 2423 make_not_entrant = true; 2424 } 2425 2426 // Here's more hysteresis: If there has been a recompile at 2427 // this trap point already, run the method in the interpreter 2428 // for a while to exercise it more thoroughly. 2429 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2430 reprofile = true; 2431 } 2432 } 2433 2434 // Take requested actions on the method: 2435 2436 // Recompile 2437 if (make_not_entrant) { 2438 if (!nm->make_not_entrant()) { 2439 return; // the call did not change nmethod's state 2440 } 2441 2442 if (pdata != nullptr) { 2443 // Record the recompilation event, if any. 2444 int tstate0 = pdata->trap_state(); 2445 int tstate1 = trap_state_set_recompiled(tstate0, true); 2446 if (tstate1 != tstate0) 2447 pdata->set_trap_state(tstate1); 2448 } 2449 2450 #if INCLUDE_RTM_OPT 2451 // Restart collecting RTM locking abort statistic if the method 2452 // is recompiled for a reason other than RTM state change. 2453 // Assume that in new recompiled code the statistic could be different, 2454 // for example, due to different inlining. 2455 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) && 2456 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 2457 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2458 } 2459 #endif 2460 // For code aging we count traps separately here, using make_not_entrant() 2461 // as a guard against simultaneous deopts in multiple threads. 2462 if (reason == Reason_tenured && trap_mdo != nullptr) { 2463 trap_mdo->inc_tenure_traps(); 2464 } 2465 } 2466 2467 if (inc_recompile_count) { 2468 trap_mdo->inc_overflow_recompile_count(); 2469 if ((uint)trap_mdo->overflow_recompile_count() > 2470 (uint)PerBytecodeRecompilationCutoff) { 2471 // Give up on the method containing the bad BCI. 2472 if (trap_method() == nm->method()) { 2473 make_not_compilable = true; 2474 } else { 2475 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2476 // But give grace to the enclosing nm->method(). 2477 } 2478 } 2479 } 2480 2481 // Reprofile 2482 if (reprofile) { 2483 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2484 } 2485 2486 // Give up compiling 2487 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2488 assert(make_not_entrant, "consistent"); 2489 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2490 } 2491 2492 if (ProfileExceptionHandlers && trap_mdo != nullptr) { 2493 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci); 2494 if (exception_handler_data != nullptr) { 2495 // uncommon trap at the start of an exception handler. 2496 // C2 generates these for un-entered exception handlers. 2497 // mark the handler as entered to avoid generating 2498 // another uncommon trap the next time the handler is compiled 2499 exception_handler_data->set_exception_handler_entered(); 2500 } 2501 } 2502 2503 } // Free marked resources 2504 2505 } 2506 JRT_END 2507 2508 ProfileData* 2509 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2510 int trap_bci, 2511 Deoptimization::DeoptReason reason, 2512 bool update_total_trap_count, 2513 #if INCLUDE_JVMCI 2514 bool is_osr, 2515 #endif 2516 Method* compiled_method, 2517 //outputs: 2518 uint& ret_this_trap_count, 2519 bool& ret_maybe_prior_trap, 2520 bool& ret_maybe_prior_recompile) { 2521 trap_mdo->check_extra_data_locked(); 2522 2523 bool maybe_prior_trap = false; 2524 bool maybe_prior_recompile = false; 2525 uint this_trap_count = 0; 2526 if (update_total_trap_count) { 2527 uint idx = reason; 2528 #if INCLUDE_JVMCI 2529 if (is_osr) { 2530 // Upper half of history array used for traps in OSR compilations 2531 idx += Reason_TRAP_HISTORY_LENGTH; 2532 } 2533 #endif 2534 uint prior_trap_count = trap_mdo->trap_count(idx); 2535 this_trap_count = trap_mdo->inc_trap_count(idx); 2536 2537 // If the runtime cannot find a place to store trap history, 2538 // it is estimated based on the general condition of the method. 2539 // If the method has ever been recompiled, or has ever incurred 2540 // a trap with the present reason , then this BCI is assumed 2541 // (pessimistically) to be the culprit. 2542 maybe_prior_trap = (prior_trap_count != 0); 2543 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2544 } 2545 ProfileData* pdata = nullptr; 2546 2547 2548 // For reasons which are recorded per bytecode, we check per-BCI data. 2549 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2550 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2551 if (per_bc_reason != Reason_none) { 2552 // Find the profile data for this BCI. If there isn't one, 2553 // try to allocate one from the MDO's set of spares. 2554 // This will let us detect a repeated trap at this point. 2555 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr); 2556 2557 if (pdata != nullptr) { 2558 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2559 if (LogCompilation && xtty != nullptr) { 2560 ttyLocker ttyl; 2561 // no more room for speculative traps in this MDO 2562 xtty->elem("speculative_traps_oom"); 2563 } 2564 } 2565 // Query the trap state of this profile datum. 2566 int tstate0 = pdata->trap_state(); 2567 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2568 maybe_prior_trap = false; 2569 if (!trap_state_is_recompiled(tstate0)) 2570 maybe_prior_recompile = false; 2571 2572 // Update the trap state of this profile datum. 2573 int tstate1 = tstate0; 2574 // Record the reason. 2575 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2576 // Store the updated state on the MDO, for next time. 2577 if (tstate1 != tstate0) 2578 pdata->set_trap_state(tstate1); 2579 } else { 2580 if (LogCompilation && xtty != nullptr) { 2581 ttyLocker ttyl; 2582 // Missing MDP? Leave a small complaint in the log. 2583 xtty->elem("missing_mdp bci='%d'", trap_bci); 2584 } 2585 } 2586 } 2587 2588 // Return results: 2589 ret_this_trap_count = this_trap_count; 2590 ret_maybe_prior_trap = maybe_prior_trap; 2591 ret_maybe_prior_recompile = maybe_prior_recompile; 2592 return pdata; 2593 } 2594 2595 void 2596 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2597 ResourceMark rm; 2598 // Ignored outputs: 2599 uint ignore_this_trap_count; 2600 bool ignore_maybe_prior_trap; 2601 bool ignore_maybe_prior_recompile; 2602 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2603 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2604 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2605 2606 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2607 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2608 2609 query_update_method_data(trap_mdo, trap_bci, 2610 (DeoptReason)reason, 2611 update_total_counts, 2612 #if INCLUDE_JVMCI 2613 false, 2614 #endif 2615 nullptr, 2616 ignore_this_trap_count, 2617 ignore_maybe_prior_trap, 2618 ignore_maybe_prior_recompile); 2619 } 2620 2621 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) { 2622 // Enable WXWrite: current function is called from methods compiled by C2 directly 2623 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2624 2625 // Still in Java no safepoints 2626 { 2627 // This enters VM and may safepoint 2628 uncommon_trap_inner(current, trap_request); 2629 } 2630 HandleMark hm(current); 2631 return fetch_unroll_info_helper(current, exec_mode); 2632 } 2633 2634 // Local derived constants. 2635 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2636 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2637 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2638 2639 //---------------------------trap_state_reason--------------------------------- 2640 Deoptimization::DeoptReason 2641 Deoptimization::trap_state_reason(int trap_state) { 2642 // This assert provides the link between the width of DataLayout::trap_bits 2643 // and the encoding of "recorded" reasons. It ensures there are enough 2644 // bits to store all needed reasons in the per-BCI MDO profile. 2645 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2646 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2647 trap_state -= recompile_bit; 2648 if (trap_state == DS_REASON_MASK) { 2649 return Reason_many; 2650 } else { 2651 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2652 return (DeoptReason)trap_state; 2653 } 2654 } 2655 //-------------------------trap_state_has_reason------------------------------- 2656 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2657 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2658 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2659 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2660 trap_state -= recompile_bit; 2661 if (trap_state == DS_REASON_MASK) { 2662 return -1; // true, unspecifically (bottom of state lattice) 2663 } else if (trap_state == reason) { 2664 return 1; // true, definitely 2665 } else if (trap_state == 0) { 2666 return 0; // false, definitely (top of state lattice) 2667 } else { 2668 return 0; // false, definitely 2669 } 2670 } 2671 //-------------------------trap_state_add_reason------------------------------- 2672 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2673 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2674 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2675 trap_state -= recompile_bit; 2676 if (trap_state == DS_REASON_MASK) { 2677 return trap_state + recompile_bit; // already at state lattice bottom 2678 } else if (trap_state == reason) { 2679 return trap_state + recompile_bit; // the condition is already true 2680 } else if (trap_state == 0) { 2681 return reason + recompile_bit; // no condition has yet been true 2682 } else { 2683 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2684 } 2685 } 2686 //-----------------------trap_state_is_recompiled------------------------------ 2687 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2688 return (trap_state & DS_RECOMPILE_BIT) != 0; 2689 } 2690 //-----------------------trap_state_set_recompiled----------------------------- 2691 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2692 if (z) return trap_state | DS_RECOMPILE_BIT; 2693 else return trap_state & ~DS_RECOMPILE_BIT; 2694 } 2695 //---------------------------format_trap_state--------------------------------- 2696 // This is used for debugging and diagnostics, including LogFile output. 2697 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2698 int trap_state) { 2699 assert(buflen > 0, "sanity"); 2700 DeoptReason reason = trap_state_reason(trap_state); 2701 bool recomp_flag = trap_state_is_recompiled(trap_state); 2702 // Re-encode the state from its decoded components. 2703 int decoded_state = 0; 2704 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2705 decoded_state = trap_state_add_reason(decoded_state, reason); 2706 if (recomp_flag) 2707 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2708 // If the state re-encodes properly, format it symbolically. 2709 // Because this routine is used for debugging and diagnostics, 2710 // be robust even if the state is a strange value. 2711 size_t len; 2712 if (decoded_state != trap_state) { 2713 // Random buggy state that doesn't decode?? 2714 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2715 } else { 2716 len = jio_snprintf(buf, buflen, "%s%s", 2717 trap_reason_name(reason), 2718 recomp_flag ? " recompiled" : ""); 2719 } 2720 return buf; 2721 } 2722 2723 2724 //--------------------------------statics-------------------------------------- 2725 const char* Deoptimization::_trap_reason_name[] = { 2726 // Note: Keep this in sync. with enum DeoptReason. 2727 "none", 2728 "null_check", 2729 "null_assert" JVMCI_ONLY("_or_unreached0"), 2730 "range_check", 2731 "class_check", 2732 "array_check", 2733 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2734 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2735 "profile_predicate", 2736 "unloaded", 2737 "uninitialized", 2738 "initialized", 2739 "unreached", 2740 "unhandled", 2741 "constraint", 2742 "div0_check", 2743 "age", 2744 "predicate", 2745 "loop_limit_check", 2746 "speculate_class_check", 2747 "speculate_null_check", 2748 "speculate_null_assert", 2749 "rtm_state_change", 2750 "unstable_if", 2751 "unstable_fused_if", 2752 "receiver_constraint", 2753 #if INCLUDE_JVMCI 2754 "aliasing", 2755 "transfer_to_interpreter", 2756 "not_compiled_exception_handler", 2757 "unresolved", 2758 "jsr_mismatch", 2759 #endif 2760 "tenured" 2761 }; 2762 const char* Deoptimization::_trap_action_name[] = { 2763 // Note: Keep this in sync. with enum DeoptAction. 2764 "none", 2765 "maybe_recompile", 2766 "reinterpret", 2767 "make_not_entrant", 2768 "make_not_compilable" 2769 }; 2770 2771 const char* Deoptimization::trap_reason_name(int reason) { 2772 // Check that every reason has a name 2773 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2774 2775 if (reason == Reason_many) return "many"; 2776 if ((uint)reason < Reason_LIMIT) 2777 return _trap_reason_name[reason]; 2778 static char buf[20]; 2779 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2780 return buf; 2781 } 2782 const char* Deoptimization::trap_action_name(int action) { 2783 // Check that every action has a name 2784 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2785 2786 if ((uint)action < Action_LIMIT) 2787 return _trap_action_name[action]; 2788 static char buf[20]; 2789 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2790 return buf; 2791 } 2792 2793 // This is used for debugging and diagnostics, including LogFile output. 2794 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2795 int trap_request) { 2796 jint unloaded_class_index = trap_request_index(trap_request); 2797 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2798 const char* action = trap_action_name(trap_request_action(trap_request)); 2799 #if INCLUDE_JVMCI 2800 int debug_id = trap_request_debug_id(trap_request); 2801 #endif 2802 size_t len; 2803 if (unloaded_class_index < 0) { 2804 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2805 reason, action 2806 #if INCLUDE_JVMCI 2807 ,debug_id 2808 #endif 2809 ); 2810 } else { 2811 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2812 reason, action, unloaded_class_index 2813 #if INCLUDE_JVMCI 2814 ,debug_id 2815 #endif 2816 ); 2817 } 2818 return buf; 2819 } 2820 2821 juint Deoptimization::_deoptimization_hist 2822 [Deoptimization::Reason_LIMIT] 2823 [1 + Deoptimization::Action_LIMIT] 2824 [Deoptimization::BC_CASE_LIMIT] 2825 = {0}; 2826 2827 enum { 2828 LSB_BITS = 8, 2829 LSB_MASK = right_n_bits(LSB_BITS) 2830 }; 2831 2832 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2833 Bytecodes::Code bc) { 2834 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2835 assert(action >= 0 && action < Action_LIMIT, "oob"); 2836 _deoptimization_hist[Reason_none][0][0] += 1; // total 2837 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2838 juint* cases = _deoptimization_hist[reason][1+action]; 2839 juint* bc_counter_addr = nullptr; 2840 juint bc_counter = 0; 2841 // Look for an unused counter, or an exact match to this BC. 2842 if (bc != Bytecodes::_illegal) { 2843 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2844 juint* counter_addr = &cases[bc_case]; 2845 juint counter = *counter_addr; 2846 if ((counter == 0 && bc_counter_addr == nullptr) 2847 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2848 // this counter is either free or is already devoted to this BC 2849 bc_counter_addr = counter_addr; 2850 bc_counter = counter | bc; 2851 } 2852 } 2853 } 2854 if (bc_counter_addr == nullptr) { 2855 // Overflow, or no given bytecode. 2856 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2857 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2858 } 2859 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2860 } 2861 2862 jint Deoptimization::total_deoptimization_count() { 2863 return _deoptimization_hist[Reason_none][0][0]; 2864 } 2865 2866 // Get the deopt count for a specific reason and a specific action. If either 2867 // one of 'reason' or 'action' is null, the method returns the sum of all 2868 // deoptimizations with the specific 'action' or 'reason' respectively. 2869 // If both arguments are null, the method returns the total deopt count. 2870 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2871 if (reason_str == nullptr && action_str == nullptr) { 2872 return total_deoptimization_count(); 2873 } 2874 juint counter = 0; 2875 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2876 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) { 2877 for (int action = 0; action < Action_LIMIT; action++) { 2878 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) { 2879 juint* cases = _deoptimization_hist[reason][1+action]; 2880 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2881 counter += cases[bc_case] >> LSB_BITS; 2882 } 2883 } 2884 } 2885 } 2886 } 2887 return counter; 2888 } 2889 2890 void Deoptimization::print_statistics() { 2891 juint total = total_deoptimization_count(); 2892 juint account = total; 2893 if (total != 0) { 2894 ttyLocker ttyl; 2895 if (xtty != nullptr) xtty->head("statistics type='deoptimization'"); 2896 tty->print_cr("Deoptimization traps recorded:"); 2897 #define PRINT_STAT_LINE(name, r) \ 2898 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2899 PRINT_STAT_LINE("total", total); 2900 // For each non-zero entry in the histogram, print the reason, 2901 // the action, and (if specifically known) the type of bytecode. 2902 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2903 for (int action = 0; action < Action_LIMIT; action++) { 2904 juint* cases = _deoptimization_hist[reason][1+action]; 2905 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2906 juint counter = cases[bc_case]; 2907 if (counter != 0) { 2908 char name[1*K]; 2909 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2910 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2911 bc = Bytecodes::_illegal; 2912 os::snprintf_checked(name, sizeof(name), "%s/%s/%s", 2913 trap_reason_name(reason), 2914 trap_action_name(action), 2915 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2916 juint r = counter >> LSB_BITS; 2917 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2918 account -= r; 2919 } 2920 } 2921 } 2922 } 2923 if (account != 0) { 2924 PRINT_STAT_LINE("unaccounted", account); 2925 } 2926 #undef PRINT_STAT_LINE 2927 if (xtty != nullptr) xtty->tail("statistics"); 2928 } 2929 } 2930 2931 #else // COMPILER2_OR_JVMCI 2932 2933 2934 // Stubs for C1 only system. 2935 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2936 return false; 2937 } 2938 2939 const char* Deoptimization::trap_reason_name(int reason) { 2940 return "unknown"; 2941 } 2942 2943 jint Deoptimization::total_deoptimization_count() { 2944 return 0; 2945 } 2946 2947 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2948 return 0; 2949 } 2950 2951 void Deoptimization::print_statistics() { 2952 // no output 2953 } 2954 2955 void 2956 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2957 // no update 2958 } 2959 2960 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2961 return 0; 2962 } 2963 2964 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2965 Bytecodes::Code bc) { 2966 // no update 2967 } 2968 2969 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2970 int trap_state) { 2971 jio_snprintf(buf, buflen, "#%d", trap_state); 2972 return buf; 2973 } 2974 2975 #endif // COMPILER2_OR_JVMCI