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/objectMonitor.inline.hpp" 78 #include "runtime/osThread.hpp" 79 #include "runtime/safepointVerifiers.hpp" 80 #include "runtime/sharedRuntime.hpp" 81 #include "runtime/signature.hpp" 82 #include "runtime/stackFrameStream.inline.hpp" 83 #include "runtime/stackValue.hpp" 84 #include "runtime/stackWatermarkSet.hpp" 85 #include "runtime/stubRoutines.hpp" 86 #include "runtime/synchronizer.hpp" 87 #include "runtime/threadSMR.hpp" 88 #include "runtime/threadWXSetters.inline.hpp" 89 #include "runtime/vframe.hpp" 90 #include "runtime/vframeArray.hpp" 91 #include "runtime/vframe_hp.hpp" 92 #include "runtime/vmOperations.hpp" 93 #include "utilities/checkedCast.hpp" 94 #include "utilities/events.hpp" 95 #include "utilities/growableArray.hpp" 96 #include "utilities/macros.hpp" 97 #include "utilities/preserveException.hpp" 98 #include "utilities/xmlstream.hpp" 99 #if INCLUDE_JFR 100 #include "jfr/jfrEvents.hpp" 101 #include "jfr/metadata/jfrSerializer.hpp" 102 #endif 103 104 uint64_t DeoptimizationScope::_committed_deopt_gen = 0; 105 uint64_t DeoptimizationScope::_active_deopt_gen = 1; 106 bool DeoptimizationScope::_committing_in_progress = false; 107 108 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) { 109 DEBUG_ONLY(_deopted = false;) 110 111 MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag); 112 // If there is nothing to deopt _required_gen is the same as comitted. 113 _required_gen = DeoptimizationScope::_committed_deopt_gen; 114 } 115 116 DeoptimizationScope::~DeoptimizationScope() { 117 assert(_deopted, "Deopt not executed"); 118 } 119 120 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) { 121 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 122 123 // If it's already marked but we still need it to be deopted. 124 if (nm->is_marked_for_deoptimization()) { 125 dependent(nm); 126 return; 127 } 128 129 nmethod::DeoptimizationStatus status = 130 inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate; 131 Atomic::store(&nm->_deoptimization_status, status); 132 133 // Make sure active is not committed 134 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be"); 135 assert(nm->_deoptimization_generation == 0, "Is already marked"); 136 137 nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen; 138 _required_gen = DeoptimizationScope::_active_deopt_gen; 139 } 140 141 void DeoptimizationScope::dependent(nmethod* nm) { 142 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 143 144 // A method marked by someone else may have a _required_gen lower than what we marked with. 145 // Therefore only store it if it's higher than _required_gen. 146 if (_required_gen < nm->_deoptimization_generation) { 147 _required_gen = nm->_deoptimization_generation; 148 } 149 } 150 151 void DeoptimizationScope::deoptimize_marked() { 152 assert(!_deopted, "Already deopted"); 153 154 // We are not alive yet. 155 if (!Universe::is_fully_initialized()) { 156 DEBUG_ONLY(_deopted = true;) 157 return; 158 } 159 160 // Safepoints are a special case, handled here. 161 if (SafepointSynchronize::is_at_safepoint()) { 162 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen; 163 DeoptimizationScope::_active_deopt_gen++; 164 Deoptimization::deoptimize_all_marked(); 165 DEBUG_ONLY(_deopted = true;) 166 return; 167 } 168 169 uint64_t comitting = 0; 170 bool wait = false; 171 while (true) { 172 { 173 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 174 175 // First we check if we or someone else already deopted the gen we want. 176 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) { 177 DEBUG_ONLY(_deopted = true;) 178 return; 179 } 180 if (!_committing_in_progress) { 181 // The version we are about to commit. 182 comitting = DeoptimizationScope::_active_deopt_gen; 183 // Make sure new marks use a higher gen. 184 DeoptimizationScope::_active_deopt_gen++; 185 _committing_in_progress = true; 186 wait = false; 187 } else { 188 // Another thread is handshaking and committing a gen. 189 wait = true; 190 } 191 } 192 if (wait) { 193 // Wait and let the concurrent handshake be performed. 194 ThreadBlockInVM tbivm(JavaThread::current()); 195 os::naked_yield(); 196 } else { 197 // Performs the handshake. 198 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred. 199 DEBUG_ONLY(_deopted = true;) 200 { 201 ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 202 203 // Make sure that committed doesn't go backwards. 204 // Should only happen if we did a deopt during a safepoint above. 205 if (DeoptimizationScope::_committed_deopt_gen < comitting) { 206 DeoptimizationScope::_committed_deopt_gen = comitting; 207 } 208 _committing_in_progress = false; 209 210 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be"); 211 212 return; 213 } 214 } 215 } 216 } 217 218 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 219 int caller_adjustment, 220 int caller_actual_parameters, 221 int number_of_frames, 222 intptr_t* frame_sizes, 223 address* frame_pcs, 224 BasicType return_type, 225 int exec_mode) { 226 _size_of_deoptimized_frame = size_of_deoptimized_frame; 227 _caller_adjustment = caller_adjustment; 228 _caller_actual_parameters = caller_actual_parameters; 229 _number_of_frames = number_of_frames; 230 _frame_sizes = frame_sizes; 231 _frame_pcs = frame_pcs; 232 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 233 _return_type = return_type; 234 _initial_info = 0; 235 // PD (x86 only) 236 _counter_temp = 0; 237 _unpack_kind = exec_mode; 238 _sender_sp_temp = 0; 239 240 _total_frame_sizes = size_of_frames(); 241 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 242 } 243 244 Deoptimization::UnrollBlock::~UnrollBlock() { 245 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 246 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 247 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 248 } 249 250 int Deoptimization::UnrollBlock::size_of_frames() const { 251 // Account first for the adjustment of the initial frame 252 intptr_t result = _caller_adjustment; 253 for (int index = 0; index < number_of_frames(); index++) { 254 result += frame_sizes()[index]; 255 } 256 return checked_cast<int>(result); 257 } 258 259 void Deoptimization::UnrollBlock::print() { 260 ResourceMark rm; 261 stringStream st; 262 st.print_cr("UnrollBlock"); 263 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 264 st.print( " frame_sizes: "); 265 for (int index = 0; index < number_of_frames(); index++) { 266 st.print(INTX_FORMAT " ", frame_sizes()[index]); 267 } 268 st.cr(); 269 tty->print_raw(st.freeze()); 270 } 271 272 // In order to make fetch_unroll_info work properly with escape 273 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 274 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 275 // which is called from the method fetch_unroll_info_helper below. 276 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 277 // fetch_unroll_info() is called at the beginning of the deoptimization 278 // handler. Note this fact before we start generating temporary frames 279 // that can confuse an asynchronous stack walker. This counter is 280 // decremented at the end of unpack_frames(). 281 current->inc_in_deopt_handler(); 282 283 if (exec_mode == Unpack_exception) { 284 // When we get here, a callee has thrown an exception into a deoptimized 285 // frame. That throw might have deferred stack watermark checking until 286 // after unwinding. So we deal with such deferred requests here. 287 StackWatermarkSet::after_unwind(current); 288 } 289 290 return fetch_unroll_info_helper(current, exec_mode); 291 JRT_END 292 293 #if COMPILER2_OR_JVMCI 294 // print information about reallocated objects 295 static void print_objects(JavaThread* deoptee_thread, 296 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 297 ResourceMark rm; 298 stringStream st; // change to logStream with logging 299 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 300 fieldDescriptor fd; 301 302 for (int i = 0; i < objects->length(); i++) { 303 ObjectValue* sv = (ObjectValue*) objects->at(i); 304 Handle obj = sv->value(); 305 306 if (obj.is_null()) { 307 st.print_cr(" nullptr"); 308 continue; 309 } 310 311 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 312 313 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 314 k->print_value_on(&st); 315 st.print_cr(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize); 316 317 if (Verbose && k != nullptr) { 318 k->oop_print_on(obj(), &st); 319 } 320 } 321 tty->print_raw(st.freeze()); 322 } 323 324 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method, 325 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 326 bool& deoptimized_objects) { 327 bool realloc_failures = false; 328 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames"); 329 330 JavaThread* deoptee_thread = chunk->at(0)->thread(); 331 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 332 "a frame can only be deoptimized by the owner thread"); 333 334 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map); 335 336 // The flag return_oop() indicates call sites which return oop 337 // in compiled code. Such sites include java method calls, 338 // runtime calls (for example, used to allocate new objects/arrays 339 // on slow code path) and any other calls generated in compiled code. 340 // It is not guaranteed that we can get such information here only 341 // by analyzing bytecode in deoptimized frames. This is why this flag 342 // is set during method compilation (see Compile::Process_OopMap_Node()). 343 // If the previous frame was popped or if we are dispatching an exception, 344 // we don't have an oop result. 345 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 346 Handle return_value; 347 if (save_oop_result) { 348 // Reallocation may trigger GC. If deoptimization happened on return from 349 // call which returns oop we need to save it since it is not in oopmap. 350 oop result = deoptee.saved_oop_result(&map); 351 assert(oopDesc::is_oop_or_null(result), "must be oop"); 352 return_value = Handle(thread, result); 353 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 354 if (TraceDeoptimization) { 355 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 356 tty->cr(); 357 } 358 } 359 if (objects != nullptr) { 360 if (exec_mode == Deoptimization::Unpack_none) { 361 assert(thread->thread_state() == _thread_in_vm, "assumption"); 362 JavaThread* THREAD = thread; // For exception macros. 363 // Clear pending OOM if reallocation fails and return true indicating allocation failure 364 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 365 deoptimized_objects = true; 366 } else { 367 JavaThread* current = thread; // For JRT_BLOCK 368 JRT_BLOCK 369 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 370 JRT_END 371 } 372 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 373 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 374 if (TraceDeoptimization) { 375 print_objects(deoptee_thread, objects, realloc_failures); 376 } 377 } 378 if (save_oop_result) { 379 // Restore result. 380 deoptee.set_saved_oop_result(&map, return_value()); 381 } 382 return realloc_failures; 383 } 384 385 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 386 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 387 JavaThread* deoptee_thread = chunk->at(0)->thread(); 388 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 389 assert(thread == Thread::current(), "should be"); 390 HandleMark hm(thread); 391 #ifndef PRODUCT 392 bool first = true; 393 #endif // !PRODUCT 394 DEBUG_ONLY(GrowableArray<oop> lock_order{0};) 395 // Start locking from outermost/oldest frame 396 for (int i = (chunk->length() - 1); i >= 0; i--) { 397 compiledVFrame* cvf = chunk->at(i); 398 assert (cvf->scope() != nullptr,"expect only compiled java frames"); 399 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 400 if (monitors->is_nonempty()) { 401 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 402 exec_mode, realloc_failures); 403 deoptimized_objects = deoptimized_objects || relocked; 404 #ifdef ASSERT 405 if (LockingMode == LM_LIGHTWEIGHT && !realloc_failures) { 406 for (MonitorInfo* mi : *monitors) { 407 lock_order.push(mi->owner()); 408 } 409 } 410 #endif // ASSERT 411 #ifndef PRODUCT 412 if (PrintDeoptimizationDetails) { 413 ResourceMark rm; 414 stringStream st; 415 for (int j = 0; j < monitors->length(); j++) { 416 MonitorInfo* mi = monitors->at(j); 417 if (mi->eliminated()) { 418 if (first) { 419 first = false; 420 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 421 } 422 if (exec_mode == Deoptimization::Unpack_none) { 423 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 424 if (monitor != nullptr && monitor->object() == mi->owner()) { 425 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 426 continue; 427 } 428 } 429 if (mi->owner_is_scalar_replaced()) { 430 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 431 st.print_cr(" failed reallocation for klass %s", k->external_name()); 432 } else { 433 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 434 } 435 } 436 } 437 tty->print_raw(st.freeze()); 438 } 439 #endif // !PRODUCT 440 } 441 } 442 #ifdef ASSERT 443 if (LockingMode == LM_LIGHTWEIGHT && !realloc_failures) { 444 deoptee_thread->lock_stack().verify_consistent_lock_order(lock_order, exec_mode != Deoptimization::Unpack_none); 445 } 446 #endif // ASSERT 447 } 448 449 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 450 // The given vframes cover one physical frame. 451 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 452 bool& realloc_failures) { 453 frame deoptee = chunk->at(0)->fr(); 454 JavaThread* deoptee_thread = chunk->at(0)->thread(); 455 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 456 RegisterMap map(chunk->at(0)->register_map()); 457 bool deoptimized_objects = false; 458 459 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false); 460 461 // Reallocate the non-escaping objects and restore their fields. 462 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 463 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 464 realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects); 465 } 466 467 // MonitorInfo structures used in eliminate_locks are not GC safe. 468 NoSafepointVerifier no_safepoint; 469 470 // Now relock objects if synchronization on them was eliminated. 471 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 472 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 473 } 474 return deoptimized_objects; 475 } 476 #endif // COMPILER2_OR_JVMCI 477 478 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 479 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 480 // When we get here we are about to unwind the deoptee frame. In order to 481 // catch not yet safe to use frames, the following stack watermark barrier 482 // poll will make such frames safe to use. 483 StackWatermarkSet::before_unwind(current); 484 485 // Note: there is a safepoint safety issue here. No matter whether we enter 486 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 487 // the vframeArray is created. 488 // 489 490 // Allocate our special deoptimization ResourceMark 491 DeoptResourceMark* dmark = new DeoptResourceMark(current); 492 assert(current->deopt_mark() == nullptr, "Pending deopt!"); 493 current->set_deopt_mark(dmark); 494 495 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 496 RegisterMap map(current, 497 RegisterMap::UpdateMap::include, 498 RegisterMap::ProcessFrames::include, 499 RegisterMap::WalkContinuation::skip); 500 RegisterMap dummy_map(current, 501 RegisterMap::UpdateMap::skip, 502 RegisterMap::ProcessFrames::include, 503 RegisterMap::WalkContinuation::skip); 504 // Now get the deoptee with a valid map 505 frame deoptee = stub_frame.sender(&map); 506 // Set the deoptee nmethod 507 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!"); 508 nmethod* nm = deoptee.cb()->as_nmethod_or_null(); 509 current->set_deopt_compiled_method(nm); 510 511 if (VerifyStack) { 512 current->validate_frame_layout(); 513 } 514 515 // Create a growable array of VFrames where each VFrame represents an inlined 516 // Java frame. This storage is allocated with the usual system arena. 517 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 518 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 519 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 520 while (!vf->is_top()) { 521 assert(vf->is_compiled_frame(), "Wrong frame type"); 522 chunk->push(compiledVFrame::cast(vf)); 523 vf = vf->sender(); 524 } 525 assert(vf->is_compiled_frame(), "Wrong frame type"); 526 chunk->push(compiledVFrame::cast(vf)); 527 528 bool realloc_failures = false; 529 530 #if COMPILER2_OR_JVMCI 531 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 532 533 // Reallocate the non-escaping objects and restore their fields. Then 534 // relock objects if synchronization on them was eliminated. 535 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 536 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 537 bool unused; 538 realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused); 539 } 540 #endif // COMPILER2_OR_JVMCI 541 542 // Ensure that no safepoint is taken after pointers have been stored 543 // in fields of rematerialized objects. If a safepoint occurs from here on 544 // out the java state residing in the vframeArray will be missed. 545 // Locks may be rebaised in a safepoint. 546 NoSafepointVerifier no_safepoint; 547 548 #if COMPILER2_OR_JVMCI 549 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 550 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 551 bool unused; 552 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 553 } 554 #endif // COMPILER2_OR_JVMCI 555 556 ScopeDesc* trap_scope = chunk->at(0)->scope(); 557 Handle exceptionObject; 558 if (trap_scope->rethrow_exception()) { 559 #ifndef PRODUCT 560 if (PrintDeoptimizationDetails) { 561 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()); 562 } 563 #endif // !PRODUCT 564 565 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 566 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw"); 567 ScopeValue* topOfStack = expressions->top(); 568 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 569 guarantee(exceptionObject() != nullptr, "exception oop can not be null"); 570 } 571 572 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 573 #if COMPILER2_OR_JVMCI 574 if (realloc_failures) { 575 // This destroys all ScopedValue bindings. 576 current->clear_scopedValueBindings(); 577 pop_frames_failed_reallocs(current, array); 578 } 579 #endif 580 581 assert(current->vframe_array_head() == nullptr, "Pending deopt!"); 582 current->set_vframe_array_head(array); 583 584 // Now that the vframeArray has been created if we have any deferred local writes 585 // added by jvmti then we can free up that structure as the data is now in the 586 // vframeArray 587 588 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 589 590 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 591 CodeBlob* cb = stub_frame.cb(); 592 // Verify we have the right vframeArray 593 assert(cb->frame_size() >= 0, "Unexpected frame size"); 594 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 595 596 // If the deopt call site is a MethodHandle invoke call site we have 597 // to adjust the unpack_sp. 598 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 599 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc())) 600 unpack_sp = deoptee.unextended_sp(); 601 602 #ifdef ASSERT 603 assert(cb->is_deoptimization_stub() || 604 cb->is_uncommon_trap_stub() || 605 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 606 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 607 "unexpected code blob: %s", cb->name()); 608 #endif 609 610 // This is a guarantee instead of an assert because if vframe doesn't match 611 // we will unpack the wrong deoptimized frame and wind up in strange places 612 // where it will be very difficult to figure out what went wrong. Better 613 // to die an early death here than some very obscure death later when the 614 // trail is cold. 615 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 616 // in that it will fail to detect a problem when there is one. This needs 617 // more work in tiger timeframe. 618 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 619 620 int number_of_frames = array->frames(); 621 622 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 623 // virtual activation, which is the reverse of the elements in the vframes array. 624 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 625 // +1 because we always have an interpreter return address for the final slot. 626 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 627 int popframe_extra_args = 0; 628 // Create an interpreter return address for the stub to use as its return 629 // address so the skeletal frames are perfectly walkable 630 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 631 632 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 633 // activation be put back on the expression stack of the caller for reexecution 634 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 635 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 636 } 637 638 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 639 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 640 // than simply use array->sender.pc(). This requires us to walk the current set of frames 641 // 642 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 643 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 644 645 // It's possible that the number of parameters at the call site is 646 // different than number of arguments in the callee when method 647 // handles are used. If the caller is interpreted get the real 648 // value so that the proper amount of space can be added to it's 649 // frame. 650 bool caller_was_method_handle = false; 651 if (deopt_sender.is_interpreted_frame()) { 652 methodHandle method(current, deopt_sender.interpreter_frame_method()); 653 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 654 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 655 // Method handle invokes may involve fairly arbitrary chains of 656 // calls so it's impossible to know how much actual space the 657 // caller has for locals. 658 caller_was_method_handle = true; 659 } 660 } 661 662 // 663 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 664 // frame_sizes/frame_pcs[1] next oldest frame (int) 665 // frame_sizes/frame_pcs[n] youngest frame (int) 666 // 667 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 668 // owns the space for the return address to it's caller). Confusing ain't it. 669 // 670 // The vframe array can address vframes with indices running from 671 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 672 // When we create the skeletal frames we need the oldest frame to be in the zero slot 673 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 674 // so things look a little strange in this loop. 675 // 676 int callee_parameters = 0; 677 int callee_locals = 0; 678 for (int index = 0; index < array->frames(); index++ ) { 679 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 680 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 681 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 682 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 683 callee_locals, 684 index == 0, 685 popframe_extra_args); 686 // This pc doesn't have to be perfect just good enough to identify the frame 687 // as interpreted so the skeleton frame will be walkable 688 // The correct pc will be set when the skeleton frame is completely filled out 689 // The final pc we store in the loop is wrong and will be overwritten below 690 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 691 692 callee_parameters = array->element(index)->method()->size_of_parameters(); 693 callee_locals = array->element(index)->method()->max_locals(); 694 popframe_extra_args = 0; 695 } 696 697 // Compute whether the root vframe returns a float or double value. 698 BasicType return_type; 699 { 700 methodHandle method(current, array->element(0)->method()); 701 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 702 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 703 } 704 705 // Compute information for handling adapters and adjusting the frame size of the caller. 706 int caller_adjustment = 0; 707 708 // Compute the amount the oldest interpreter frame will have to adjust 709 // its caller's stack by. If the caller is a compiled frame then 710 // we pretend that the callee has no parameters so that the 711 // extension counts for the full amount of locals and not just 712 // locals-parms. This is because without a c2i adapter the parm 713 // area as created by the compiled frame will not be usable by 714 // the interpreter. (Depending on the calling convention there 715 // may not even be enough space). 716 717 // QQQ I'd rather see this pushed down into last_frame_adjust 718 // and have it take the sender (aka caller). 719 720 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 721 caller_adjustment = last_frame_adjust(0, callee_locals); 722 } else if (callee_locals > callee_parameters) { 723 // The caller frame may need extending to accommodate 724 // non-parameter locals of the first unpacked interpreted frame. 725 // Compute that adjustment. 726 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 727 } 728 729 // If the sender is deoptimized the we must retrieve the address of the handler 730 // since the frame will "magically" show the original pc before the deopt 731 // and we'd undo the deopt. 732 733 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 734 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 735 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 736 } 737 738 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc"); 739 740 #if INCLUDE_JVMCI 741 if (exceptionObject() != nullptr) { 742 current->set_exception_oop(exceptionObject()); 743 exec_mode = Unpack_exception; 744 } 745 #endif 746 747 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 748 assert(current->has_pending_exception(), "should have thrown OOME"); 749 current->set_exception_oop(current->pending_exception()); 750 current->clear_pending_exception(); 751 exec_mode = Unpack_exception; 752 } 753 754 #if INCLUDE_JVMCI 755 if (current->frames_to_pop_failed_realloc() > 0) { 756 current->set_pending_monitorenter(false); 757 } 758 #endif 759 760 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 761 caller_adjustment * BytesPerWord, 762 caller_was_method_handle ? 0 : callee_parameters, 763 number_of_frames, 764 frame_sizes, 765 frame_pcs, 766 return_type, 767 exec_mode); 768 // On some platforms, we need a way to pass some platform dependent 769 // information to the unpacking code so the skeletal frames come out 770 // correct (initial fp value, unextended sp, ...) 771 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 772 773 if (array->frames() > 1) { 774 if (VerifyStack && TraceDeoptimization) { 775 tty->print_cr("Deoptimizing method containing inlining"); 776 } 777 } 778 779 array->set_unroll_block(info); 780 return info; 781 } 782 783 // Called to cleanup deoptimization data structures in normal case 784 // after unpacking to stack and when stack overflow error occurs 785 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 786 vframeArray *array) { 787 788 // Get array if coming from exception 789 if (array == nullptr) { 790 array = thread->vframe_array_head(); 791 } 792 thread->set_vframe_array_head(nullptr); 793 794 // Free the previous UnrollBlock 795 vframeArray* old_array = thread->vframe_array_last(); 796 thread->set_vframe_array_last(array); 797 798 if (old_array != nullptr) { 799 UnrollBlock* old_info = old_array->unroll_block(); 800 old_array->set_unroll_block(nullptr); 801 delete old_info; 802 delete old_array; 803 } 804 805 // Deallocate any resource creating in this routine and any ResourceObjs allocated 806 // inside the vframeArray (StackValueCollections) 807 808 delete thread->deopt_mark(); 809 thread->set_deopt_mark(nullptr); 810 thread->set_deopt_compiled_method(nullptr); 811 812 813 if (JvmtiExport::can_pop_frame()) { 814 // Regardless of whether we entered this routine with the pending 815 // popframe condition bit set, we should always clear it now 816 thread->clear_popframe_condition(); 817 } 818 819 // unpack_frames() is called at the end of the deoptimization handler 820 // and (in C2) at the end of the uncommon trap handler. Note this fact 821 // so that an asynchronous stack walker can work again. This counter is 822 // incremented at the beginning of fetch_unroll_info() and (in C2) at 823 // the beginning of uncommon_trap(). 824 thread->dec_in_deopt_handler(); 825 } 826 827 // Moved from cpu directories because none of the cpus has callee save values. 828 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 829 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 830 831 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 832 // the days we had adapter frames. When we deoptimize a situation where a 833 // compiled caller calls a compiled caller will have registers it expects 834 // to survive the call to the callee. If we deoptimize the callee the only 835 // way we can restore these registers is to have the oldest interpreter 836 // frame that we create restore these values. That is what this routine 837 // will accomplish. 838 839 // At the moment we have modified c2 to not have any callee save registers 840 // so this problem does not exist and this routine is just a place holder. 841 842 assert(f->is_interpreted_frame(), "must be interpreted"); 843 } 844 845 #ifndef PRODUCT 846 static bool falls_through(Bytecodes::Code bc) { 847 switch (bc) { 848 // List may be incomplete. Here we really only care about bytecodes where compiled code 849 // can deoptimize. 850 case Bytecodes::_goto: 851 case Bytecodes::_goto_w: 852 case Bytecodes::_athrow: 853 return false; 854 default: 855 return true; 856 } 857 } 858 #endif 859 860 // Return BasicType of value being returned 861 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 862 assert(thread == JavaThread::current(), "pre-condition"); 863 864 // We are already active in the special DeoptResourceMark any ResourceObj's we 865 // allocate will be freed at the end of the routine. 866 867 // JRT_LEAF methods don't normally allocate handles and there is a 868 // NoHandleMark to enforce that. It is actually safe to use Handles 869 // in a JRT_LEAF method, and sometimes desirable, but to do so we 870 // must use ResetNoHandleMark to bypass the NoHandleMark, and 871 // then use a HandleMark to ensure any Handles we do create are 872 // cleaned up in this scope. 873 ResetNoHandleMark rnhm; 874 HandleMark hm(thread); 875 876 frame stub_frame = thread->last_frame(); 877 878 Continuation::notify_deopt(thread, stub_frame.sp()); 879 880 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 881 // must point to the vframeArray for the unpack frame. 882 vframeArray* array = thread->vframe_array_head(); 883 UnrollBlock* info = array->unroll_block(); 884 885 // We set the last_Java frame. But the stack isn't really parsable here. So we 886 // clear it to make sure JFR understands not to try and walk stacks from events 887 // in here. 888 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 889 thread->frame_anchor()->set_last_Java_sp(nullptr); 890 891 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 892 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 893 894 thread->frame_anchor()->set_last_Java_sp(sp); 895 896 BasicType bt = info->return_type(); 897 898 // If we have an exception pending, claim that the return type is an oop 899 // so the deopt_blob does not overwrite the exception_oop. 900 901 if (exec_mode == Unpack_exception) 902 bt = T_OBJECT; 903 904 // Cleanup thread deopt data 905 cleanup_deopt_info(thread, array); 906 907 #ifndef PRODUCT 908 if (VerifyStack) { 909 ResourceMark res_mark; 910 // Clear pending exception to not break verification code (restored afterwards) 911 PreserveExceptionMark pm(thread); 912 913 thread->validate_frame_layout(); 914 915 // Verify that the just-unpacked frames match the interpreter's 916 // notions of expression stack and locals 917 vframeArray* cur_array = thread->vframe_array_last(); 918 RegisterMap rm(thread, 919 RegisterMap::UpdateMap::skip, 920 RegisterMap::ProcessFrames::include, 921 RegisterMap::WalkContinuation::skip); 922 rm.set_include_argument_oops(false); 923 bool is_top_frame = true; 924 int callee_size_of_parameters = 0; 925 int callee_max_locals = 0; 926 for (int i = 0; i < cur_array->frames(); i++) { 927 vframeArrayElement* el = cur_array->element(i); 928 frame* iframe = el->iframe(); 929 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 930 931 // Get the oop map for this bci 932 InterpreterOopMap mask; 933 int cur_invoke_parameter_size = 0; 934 bool try_next_mask = false; 935 int next_mask_expression_stack_size = -1; 936 int top_frame_expression_stack_adjustment = 0; 937 methodHandle mh(thread, iframe->interpreter_frame_method()); 938 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 939 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 940 int max_bci = mh->code_size(); 941 // Get to the next bytecode if possible 942 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 943 // Check to see if we can grab the number of outgoing arguments 944 // at an uncommon trap for an invoke (where the compiler 945 // generates debug info before the invoke has executed) 946 Bytecodes::Code cur_code = str.next(); 947 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere; 948 if (Bytecodes::is_invoke(cur_code)) { 949 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 950 cur_invoke_parameter_size = invoke.size_of_parameters(); 951 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 952 callee_size_of_parameters++; 953 } 954 } 955 if (str.bci() < max_bci) { 956 next_code = str.next(); 957 if (next_code >= 0) { 958 // The interpreter oop map generator reports results before 959 // the current bytecode has executed except in the case of 960 // calls. It seems to be hard to tell whether the compiler 961 // has emitted debug information matching the "state before" 962 // a given bytecode or the state after, so we try both 963 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) { 964 // Get expression stack size for the next bytecode 965 InterpreterOopMap next_mask; 966 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 967 next_mask_expression_stack_size = next_mask.expression_stack_size(); 968 if (Bytecodes::is_invoke(next_code)) { 969 Bytecode_invoke invoke(mh, str.bci()); 970 next_mask_expression_stack_size += invoke.size_of_parameters(); 971 } 972 // Need to subtract off the size of the result type of 973 // the bytecode because this is not described in the 974 // debug info but returned to the interpreter in the TOS 975 // caching register 976 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 977 if (bytecode_result_type != T_ILLEGAL) { 978 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 979 } 980 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 981 try_next_mask = true; 982 } 983 } 984 } 985 986 // Verify stack depth and oops in frame 987 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 988 if (!( 989 /* SPARC */ 990 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 991 /* x86 */ 992 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 993 (try_next_mask && 994 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 995 top_frame_expression_stack_adjustment))) || 996 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 997 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 998 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 999 )) { 1000 { 1001 // Print out some information that will help us debug the problem 1002 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1003 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 1004 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 1005 if (try_next_mask) { 1006 tty->print_cr(" Next code %s", Bytecodes::name(next_code)); 1007 } 1008 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1009 iframe->interpreter_frame_expression_stack_size()); 1010 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1011 tty->print_cr(" try_next_mask = %d", try_next_mask); 1012 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 1013 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1014 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 1015 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1016 tty->print_cr(" exec_mode = %d", exec_mode); 1017 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1018 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 1019 tty->print_cr(" Interpreted frames:"); 1020 for (int k = 0; k < cur_array->frames(); k++) { 1021 vframeArrayElement* el = cur_array->element(k); 1022 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1023 } 1024 cur_array->print_on_2(tty); 1025 } 1026 guarantee(false, "wrong number of expression stack elements during deopt"); 1027 } 1028 VerifyOopClosure verify; 1029 iframe->oops_interpreted_do(&verify, &rm, false); 1030 callee_size_of_parameters = mh->size_of_parameters(); 1031 callee_max_locals = mh->max_locals(); 1032 is_top_frame = false; 1033 } 1034 } 1035 #endif // !PRODUCT 1036 1037 return bt; 1038 JRT_END 1039 1040 class DeoptimizeMarkedClosure : public HandshakeClosure { 1041 public: 1042 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 1043 void do_thread(Thread* thread) { 1044 JavaThread* jt = JavaThread::cast(thread); 1045 jt->deoptimize_marked_methods(); 1046 } 1047 }; 1048 1049 void Deoptimization::deoptimize_all_marked() { 1050 ResourceMark rm; 1051 1052 // Make the dependent methods not entrant 1053 CodeCache::make_marked_nmethods_deoptimized(); 1054 1055 DeoptimizeMarkedClosure deopt; 1056 if (SafepointSynchronize::is_at_safepoint()) { 1057 Threads::java_threads_do(&deopt); 1058 } else { 1059 Handshake::execute(&deopt); 1060 } 1061 } 1062 1063 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1064 = Deoptimization::Action_reinterpret; 1065 1066 #if INCLUDE_JVMCI 1067 template<typename CacheType> 1068 class BoxCacheBase : public CHeapObj<mtCompiler> { 1069 protected: 1070 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 1071 ResourceMark rm(thread); 1072 char* klass_name_str = klass_name->as_C_string(); 1073 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle()); 1074 guarantee(ik != nullptr, "%s must be loaded", klass_name_str); 1075 if (!ik->is_in_error_state()) { 1076 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 1077 CacheType::compute_offsets(ik); 1078 } 1079 return ik; 1080 } 1081 }; 1082 1083 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 1084 PrimitiveType _low; 1085 PrimitiveType _high; 1086 jobject _cache; 1087 protected: 1088 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 1089 BoxCache(Thread* thread) { 1090 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 1091 if (ik->is_in_error_state()) { 1092 _low = 1; 1093 _high = 0; 1094 _cache = nullptr; 1095 } else { 1096 objArrayOop cache = CacheType::cache(ik); 1097 assert(cache->length() > 0, "Empty cache"); 1098 _low = BoxType::value(cache->obj_at(0)); 1099 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1); 1100 _cache = JNIHandles::make_global(Handle(thread, cache)); 1101 } 1102 } 1103 ~BoxCache() { 1104 JNIHandles::destroy_global(_cache); 1105 } 1106 public: 1107 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1108 if (_singleton == nullptr) { 1109 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1110 if (!Atomic::replace_if_null(&_singleton, s)) { 1111 delete s; 1112 } 1113 } 1114 return _singleton; 1115 } 1116 oop lookup(PrimitiveType value) { 1117 if (_low <= value && value <= _high) { 1118 int offset = checked_cast<int>(value - _low); 1119 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1120 } 1121 return nullptr; 1122 } 1123 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) { 1124 if (_cache == nullptr) { 1125 cache_init_error = true; 1126 return nullptr; 1127 } 1128 // Have to cast to avoid little/big-endian problems. 1129 if (sizeof(PrimitiveType) > sizeof(jint)) { 1130 jlong value = (jlong)raw_value; 1131 return lookup(value); 1132 } 1133 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1134 return lookup(value); 1135 } 1136 }; 1137 1138 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1139 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1140 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1141 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1142 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1143 1144 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr; 1145 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr; 1146 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr; 1147 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr; 1148 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr; 1149 1150 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1151 jobject _true_cache; 1152 jobject _false_cache; 1153 protected: 1154 static BooleanBoxCache *_singleton; 1155 BooleanBoxCache(Thread *thread) { 1156 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1157 if (ik->is_in_error_state()) { 1158 _true_cache = nullptr; 1159 _false_cache = nullptr; 1160 } else { 1161 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1162 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1163 } 1164 } 1165 ~BooleanBoxCache() { 1166 JNIHandles::destroy_global(_true_cache); 1167 JNIHandles::destroy_global(_false_cache); 1168 } 1169 public: 1170 static BooleanBoxCache* singleton(Thread* thread) { 1171 if (_singleton == nullptr) { 1172 BooleanBoxCache* s = new BooleanBoxCache(thread); 1173 if (!Atomic::replace_if_null(&_singleton, s)) { 1174 delete s; 1175 } 1176 } 1177 return _singleton; 1178 } 1179 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) { 1180 if (_true_cache == nullptr) { 1181 cache_in_error = true; 1182 return nullptr; 1183 } 1184 // Have to cast to avoid little/big-endian problems. 1185 jboolean value = (jboolean)*((jint*)&raw_value); 1186 return lookup(value); 1187 } 1188 oop lookup(jboolean value) { 1189 if (value != 0) { 1190 return JNIHandles::resolve_non_null(_true_cache); 1191 } 1192 return JNIHandles::resolve_non_null(_false_cache); 1193 } 1194 }; 1195 1196 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr; 1197 1198 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) { 1199 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1200 BasicType box_type = vmClasses::box_klass_type(k); 1201 if (box_type != T_OBJECT) { 1202 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1203 switch(box_type) { 1204 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1205 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1206 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1207 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1208 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1209 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1210 default:; 1211 } 1212 } 1213 return nullptr; 1214 } 1215 #endif // INCLUDE_JVMCI 1216 1217 #if COMPILER2_OR_JVMCI 1218 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1219 Handle pending_exception(THREAD, thread->pending_exception()); 1220 const char* exception_file = thread->exception_file(); 1221 int exception_line = thread->exception_line(); 1222 thread->clear_pending_exception(); 1223 1224 bool failures = false; 1225 1226 for (int i = 0; i < objects->length(); i++) { 1227 assert(objects->at(i)->is_object(), "invalid debug information"); 1228 ObjectValue* sv = (ObjectValue*) objects->at(i); 1229 1230 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1231 oop obj = nullptr; 1232 1233 bool cache_init_error = false; 1234 if (k->is_instance_klass()) { 1235 #if INCLUDE_JVMCI 1236 nmethod* nm = fr->cb()->as_nmethod_or_null(); 1237 if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1238 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1239 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD); 1240 if (obj != nullptr) { 1241 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1242 abv->set_cached(true); 1243 } else if (cache_init_error) { 1244 // Results in an OOME which is valid (as opposed to a class initialization error) 1245 // and is fine for the rare case a cache initialization failing. 1246 failures = true; 1247 } 1248 } 1249 #endif // INCLUDE_JVMCI 1250 1251 InstanceKlass* ik = InstanceKlass::cast(k); 1252 if (obj == nullptr && !cache_init_error) { 1253 #if COMPILER2_OR_JVMCI 1254 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1255 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1256 } else { 1257 obj = ik->allocate_instance(THREAD); 1258 } 1259 #else 1260 obj = ik->allocate_instance(THREAD); 1261 #endif // COMPILER2_OR_JVMCI 1262 } 1263 } else if (k->is_typeArray_klass()) { 1264 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1265 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1266 int len = sv->field_size() / type2size[ak->element_type()]; 1267 obj = ak->allocate(len, THREAD); 1268 } else if (k->is_objArray_klass()) { 1269 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1270 obj = ak->allocate(sv->field_size(), THREAD); 1271 } 1272 1273 if (obj == nullptr) { 1274 failures = true; 1275 } 1276 1277 assert(sv->value().is_null(), "redundant reallocation"); 1278 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception"); 1279 CLEAR_PENDING_EXCEPTION; 1280 sv->set_value(obj); 1281 } 1282 1283 if (failures) { 1284 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1285 } else if (pending_exception.not_null()) { 1286 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1287 } 1288 1289 return failures; 1290 } 1291 1292 #if INCLUDE_JVMCI 1293 /** 1294 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1295 * we need to somehow be able to recover the actual kind to be able to write the correct 1296 * amount of bytes. 1297 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1298 * the entries at index n + 1 to n + i are 'markers'. 1299 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1300 * expected form of the array would be: 1301 * 1302 * {b0, b1, b2, b3, INT, marker, b6, b7} 1303 * 1304 * Thus, in order to get back the size of the entry, we simply need to count the number 1305 * of marked entries 1306 * 1307 * @param virtualArray the virtualized byte array 1308 * @param i index of the virtual entry we are recovering 1309 * @return The number of bytes the entry spans 1310 */ 1311 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1312 int index = i; 1313 while (++index < virtualArray->field_size() && 1314 virtualArray->field_at(index)->is_marker()) {} 1315 return index - i; 1316 } 1317 1318 /** 1319 * If there was a guarantee for byte array to always start aligned to a long, we could 1320 * do a simple check on the parity of the index. Unfortunately, that is not always the 1321 * case. Thus, we check alignment of the actual address we are writing to. 1322 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1323 * write a long to index 3. 1324 */ 1325 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1326 jbyte* res = obj->byte_at_addr(index); 1327 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1328 return res; 1329 } 1330 1331 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) { 1332 switch (byte_count) { 1333 case 1: 1334 obj->byte_at_put(index, (jbyte) value->get_jint()); 1335 break; 1336 case 2: 1337 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint(); 1338 break; 1339 case 4: 1340 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint(); 1341 break; 1342 case 8: 1343 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr(); 1344 break; 1345 default: 1346 ShouldNotReachHere(); 1347 } 1348 } 1349 #endif // INCLUDE_JVMCI 1350 1351 1352 // restore elements of an eliminated type array 1353 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1354 int index = 0; 1355 1356 for (int i = 0; i < sv->field_size(); i++) { 1357 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1358 switch(type) { 1359 case T_LONG: case T_DOUBLE: { 1360 assert(value->type() == T_INT, "Agreement."); 1361 StackValue* low = 1362 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1363 #ifdef _LP64 1364 jlong res = (jlong)low->get_intptr(); 1365 #else 1366 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1367 #endif 1368 obj->long_at_put(index, res); 1369 break; 1370 } 1371 1372 case T_INT: case T_FLOAT: { // 4 bytes. 1373 assert(value->type() == T_INT, "Agreement."); 1374 bool big_value = false; 1375 if (i + 1 < sv->field_size() && type == T_INT) { 1376 if (sv->field_at(i)->is_location()) { 1377 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1378 if (type == Location::dbl || type == Location::lng) { 1379 big_value = true; 1380 } 1381 } else if (sv->field_at(i)->is_constant_int()) { 1382 ScopeValue* next_scope_field = sv->field_at(i + 1); 1383 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1384 big_value = true; 1385 } 1386 } 1387 } 1388 1389 if (big_value) { 1390 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1391 #ifdef _LP64 1392 jlong res = (jlong)low->get_intptr(); 1393 #else 1394 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1395 #endif 1396 obj->int_at_put(index, *(jint*)&res); 1397 obj->int_at_put(++index, *((jint*)&res + 1)); 1398 } else { 1399 obj->int_at_put(index, value->get_jint()); 1400 } 1401 break; 1402 } 1403 1404 case T_SHORT: 1405 assert(value->type() == T_INT, "Agreement."); 1406 obj->short_at_put(index, (jshort)value->get_jint()); 1407 break; 1408 1409 case T_CHAR: 1410 assert(value->type() == T_INT, "Agreement."); 1411 obj->char_at_put(index, (jchar)value->get_jint()); 1412 break; 1413 1414 case T_BYTE: { 1415 assert(value->type() == T_INT, "Agreement."); 1416 #if INCLUDE_JVMCI 1417 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1418 int byte_count = count_number_of_bytes_for_entry(sv, i); 1419 byte_array_put(obj, value, index, byte_count); 1420 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1421 i += byte_count - 1; // Balance the loop counter. 1422 index += byte_count; 1423 // index has been updated so continue at top of loop 1424 continue; 1425 #else 1426 obj->byte_at_put(index, (jbyte)value->get_jint()); 1427 break; 1428 #endif // INCLUDE_JVMCI 1429 } 1430 1431 case T_BOOLEAN: { 1432 assert(value->type() == T_INT, "Agreement."); 1433 obj->bool_at_put(index, (jboolean)value->get_jint()); 1434 break; 1435 } 1436 1437 default: 1438 ShouldNotReachHere(); 1439 } 1440 index++; 1441 } 1442 } 1443 1444 // restore fields of an eliminated object array 1445 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1446 for (int i = 0; i < sv->field_size(); i++) { 1447 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1448 assert(value->type() == T_OBJECT, "object element expected"); 1449 obj->obj_at_put(i, value->get_obj()()); 1450 } 1451 } 1452 1453 class ReassignedField { 1454 public: 1455 int _offset; 1456 BasicType _type; 1457 public: 1458 ReassignedField() { 1459 _offset = 0; 1460 _type = T_ILLEGAL; 1461 } 1462 }; 1463 1464 static int compare(ReassignedField* left, ReassignedField* right) { 1465 return left->_offset - right->_offset; 1466 } 1467 1468 // Restore fields of an eliminated instance object using the same field order 1469 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1470 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1471 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1472 InstanceKlass* ik = klass; 1473 while (ik != nullptr) { 1474 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1475 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) { 1476 ReassignedField field; 1477 field._offset = fs.offset(); 1478 field._type = Signature::basic_type(fs.signature()); 1479 fields->append(field); 1480 } 1481 } 1482 ik = ik->superklass(); 1483 } 1484 fields->sort(compare); 1485 for (int i = 0; i < fields->length(); i++) { 1486 ScopeValue* scope_field = sv->field_at(svIndex); 1487 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1488 int offset = fields->at(i)._offset; 1489 BasicType type = fields->at(i)._type; 1490 switch (type) { 1491 case T_OBJECT: case T_ARRAY: 1492 assert(value->type() == T_OBJECT, "Agreement."); 1493 obj->obj_field_put(offset, value->get_obj()()); 1494 break; 1495 1496 case T_INT: case T_FLOAT: { // 4 bytes. 1497 assert(value->type() == T_INT, "Agreement."); 1498 bool big_value = false; 1499 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1500 if (scope_field->is_location()) { 1501 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1502 if (type == Location::dbl || type == Location::lng) { 1503 big_value = true; 1504 } 1505 } 1506 if (scope_field->is_constant_int()) { 1507 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1508 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1509 big_value = true; 1510 } 1511 } 1512 } 1513 1514 if (big_value) { 1515 i++; 1516 assert(i < fields->length(), "second T_INT field needed"); 1517 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1518 } else { 1519 obj->int_field_put(offset, value->get_jint()); 1520 break; 1521 } 1522 } 1523 /* no break */ 1524 1525 case T_LONG: case T_DOUBLE: { 1526 assert(value->type() == T_INT, "Agreement."); 1527 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1528 #ifdef _LP64 1529 jlong res = (jlong)low->get_intptr(); 1530 #else 1531 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1532 #endif 1533 obj->long_field_put(offset, res); 1534 break; 1535 } 1536 1537 case T_SHORT: 1538 assert(value->type() == T_INT, "Agreement."); 1539 obj->short_field_put(offset, (jshort)value->get_jint()); 1540 break; 1541 1542 case T_CHAR: 1543 assert(value->type() == T_INT, "Agreement."); 1544 obj->char_field_put(offset, (jchar)value->get_jint()); 1545 break; 1546 1547 case T_BYTE: 1548 assert(value->type() == T_INT, "Agreement."); 1549 obj->byte_field_put(offset, (jbyte)value->get_jint()); 1550 break; 1551 1552 case T_BOOLEAN: 1553 assert(value->type() == T_INT, "Agreement."); 1554 obj->bool_field_put(offset, (jboolean)value->get_jint()); 1555 break; 1556 1557 default: 1558 ShouldNotReachHere(); 1559 } 1560 svIndex++; 1561 } 1562 return svIndex; 1563 } 1564 1565 // restore fields of all eliminated objects and arrays 1566 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1567 for (int i = 0; i < objects->length(); i++) { 1568 assert(objects->at(i)->is_object(), "invalid debug information"); 1569 ObjectValue* sv = (ObjectValue*) objects->at(i); 1570 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1571 Handle obj = sv->value(); 1572 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1573 #ifndef PRODUCT 1574 if (PrintDeoptimizationDetails) { 1575 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1576 } 1577 #endif // !PRODUCT 1578 1579 if (obj.is_null()) { 1580 continue; 1581 } 1582 1583 #if INCLUDE_JVMCI 1584 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1585 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1586 continue; 1587 } 1588 #endif // INCLUDE_JVMCI 1589 #if COMPILER2_OR_JVMCI 1590 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1591 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1592 ScopeValue* payload = sv->field_at(0); 1593 if (payload->is_location() && 1594 payload->as_LocationValue()->location().type() == Location::vector) { 1595 #ifndef PRODUCT 1596 if (PrintDeoptimizationDetails) { 1597 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1598 if (Verbose) { 1599 Handle obj = sv->value(); 1600 k->oop_print_on(obj(), tty); 1601 } 1602 } 1603 #endif // !PRODUCT 1604 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1605 } 1606 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1607 // which could be restored after vector object allocation. 1608 } 1609 #endif /* !COMPILER2_OR_JVMCI */ 1610 if (k->is_instance_klass()) { 1611 InstanceKlass* ik = InstanceKlass::cast(k); 1612 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1613 } else if (k->is_typeArray_klass()) { 1614 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1615 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1616 } else if (k->is_objArray_klass()) { 1617 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1618 } 1619 } 1620 // These objects may escape when we return to Interpreter after deoptimization. 1621 // We need barrier so that stores that initialize these objects can't be reordered 1622 // with subsequent stores that make these objects accessible by other threads. 1623 OrderAccess::storestore(); 1624 } 1625 1626 1627 // relock objects for which synchronization was eliminated 1628 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1629 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1630 bool relocked_objects = false; 1631 for (int i = 0; i < monitors->length(); i++) { 1632 MonitorInfo* mon_info = monitors->at(i); 1633 if (mon_info->eliminated()) { 1634 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1635 relocked_objects = true; 1636 if (!mon_info->owner_is_scalar_replaced()) { 1637 Handle obj(thread, mon_info->owner()); 1638 markWord mark = obj->mark(); 1639 if (exec_mode == Unpack_none) { 1640 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) { 1641 // With exec_mode == Unpack_none obj may be thread local and locked in 1642 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header. 1643 markWord dmw = mark.displaced_mark_helper(); 1644 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr)); 1645 obj->set_mark(dmw); 1646 } 1647 if (mark.has_monitor()) { 1648 // defer relocking if the deoptee thread is currently waiting for obj 1649 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1650 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) { 1651 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1652 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1653 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1654 continue; 1655 } 1656 } 1657 } 1658 if (LockingMode == LM_LIGHTWEIGHT) { 1659 // We have lost information about the correct state of the lock stack. 1660 // Inflate the locks instead. Enter then inflate to avoid races with 1661 // deflation. 1662 ObjectSynchronizer::enter_for(obj, nullptr, deoptee_thread); 1663 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1664 ObjectMonitor* mon = ObjectSynchronizer::inflate_for(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal); 1665 assert(mon->owner() == deoptee_thread, "must be"); 1666 } else { 1667 BasicLock* lock = mon_info->lock(); 1668 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread); 1669 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1670 } 1671 } 1672 } 1673 } 1674 return relocked_objects; 1675 } 1676 #endif // COMPILER2_OR_JVMCI 1677 1678 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1679 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1680 1681 // Register map for next frame (used for stack crawl). We capture 1682 // the state of the deopt'ing frame's caller. Thus if we need to 1683 // stuff a C2I adapter we can properly fill in the callee-save 1684 // register locations. 1685 frame caller = fr.sender(reg_map); 1686 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp()); 1687 1688 frame sender = caller; 1689 1690 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1691 // the vframeArray containing the unpacking information is allocated in the C heap. 1692 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1693 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1694 1695 // Compare the vframeArray to the collected vframes 1696 assert(array->structural_compare(thread, chunk), "just checking"); 1697 1698 if (TraceDeoptimization) { 1699 ResourceMark rm; 1700 stringStream st; 1701 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1702 st.print(" "); 1703 fr.print_on(&st); 1704 st.print_cr(" Virtual frames (innermost/newest first):"); 1705 for (int index = 0; index < chunk->length(); index++) { 1706 compiledVFrame* vf = chunk->at(index); 1707 int bci = vf->raw_bci(); 1708 const char* code_name; 1709 if (bci == SynchronizationEntryBCI) { 1710 code_name = "sync entry"; 1711 } else { 1712 Bytecodes::Code code = vf->method()->code_at(bci); 1713 code_name = Bytecodes::name(code); 1714 } 1715 1716 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1717 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1718 st.print(" - %s", code_name); 1719 st.print_cr(" @ bci=%d ", bci); 1720 } 1721 tty->print_raw(st.freeze()); 1722 tty->cr(); 1723 } 1724 1725 return array; 1726 } 1727 1728 #if COMPILER2_OR_JVMCI 1729 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1730 // Reallocation of some scalar replaced objects failed. Record 1731 // that we need to pop all the interpreter frames for the 1732 // deoptimized compiled frame. 1733 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1734 thread->set_frames_to_pop_failed_realloc(array->frames()); 1735 // Unlock all monitors here otherwise the interpreter will see a 1736 // mix of locked and unlocked monitors (because of failed 1737 // reallocations of synchronized objects) and be confused. 1738 for (int i = 0; i < array->frames(); i++) { 1739 MonitorChunk* monitors = array->element(i)->monitors(); 1740 if (monitors != nullptr) { 1741 // Unlock in reverse order starting from most nested monitor. 1742 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) { 1743 BasicObjectLock* src = monitors->at(j); 1744 if (src->obj() != nullptr) { 1745 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1746 } 1747 } 1748 array->element(i)->free_monitors(thread); 1749 #ifdef ASSERT 1750 array->element(i)->set_removed_monitors(); 1751 #endif 1752 } 1753 } 1754 } 1755 #endif 1756 1757 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1758 assert(fr.can_be_deoptimized(), "checking frame type"); 1759 1760 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1761 1762 if (LogCompilation && xtty != nullptr) { 1763 nmethod* nm = fr.cb()->as_nmethod_or_null(); 1764 assert(nm != nullptr, "only compiled methods can deopt"); 1765 1766 ttyLocker ttyl; 1767 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1768 nm->log_identity(xtty); 1769 xtty->end_head(); 1770 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1771 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1772 xtty->method(sd->method()); 1773 xtty->end_elem(); 1774 if (sd->is_top()) break; 1775 } 1776 xtty->tail("deoptimized"); 1777 } 1778 1779 Continuation::notify_deopt(thread, fr.sp()); 1780 1781 // Patch the compiled method so that when execution returns to it we will 1782 // deopt the execution state and return to the interpreter. 1783 fr.deoptimize(thread); 1784 } 1785 1786 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1787 // Deoptimize only if the frame comes from compile code. 1788 // Do not deoptimize the frame which is already patched 1789 // during the execution of the loops below. 1790 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1791 return; 1792 } 1793 ResourceMark rm; 1794 deoptimize_single_frame(thread, fr, reason); 1795 } 1796 1797 #if INCLUDE_JVMCI 1798 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) { 1799 // there is no exception handler for this pc => deoptimize 1800 nm->make_not_entrant(); 1801 1802 // Use Deoptimization::deoptimize for all of its side-effects: 1803 // gathering traps statistics, logging... 1804 // it also patches the return pc but we do not care about that 1805 // since we return a continuation to the deopt_blob below. 1806 JavaThread* thread = JavaThread::current(); 1807 RegisterMap reg_map(thread, 1808 RegisterMap::UpdateMap::skip, 1809 RegisterMap::ProcessFrames::include, 1810 RegisterMap::WalkContinuation::skip); 1811 frame runtime_frame = thread->last_frame(); 1812 frame caller_frame = runtime_frame.sender(®_map); 1813 assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method"); 1814 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1815 compiledVFrame* cvf = compiledVFrame::cast(vf); 1816 ScopeDesc* imm_scope = cvf->scope(); 1817 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1818 if (imm_mdo != nullptr) { 1819 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 1820 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 1821 1822 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr); 1823 if (pdata != nullptr && pdata->is_BitData()) { 1824 BitData* bit_data = (BitData*) pdata; 1825 bit_data->set_exception_seen(); 1826 } 1827 } 1828 1829 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1830 1831 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true); 1832 if (trap_mdo != nullptr) { 1833 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1834 } 1835 1836 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1837 } 1838 #endif 1839 1840 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1841 assert(thread == Thread::current() || 1842 thread->is_handshake_safe_for(Thread::current()) || 1843 SafepointSynchronize::is_at_safepoint(), 1844 "can only deoptimize other thread at a safepoint/handshake"); 1845 // Compute frame and register map based on thread and sp. 1846 RegisterMap reg_map(thread, 1847 RegisterMap::UpdateMap::skip, 1848 RegisterMap::ProcessFrames::include, 1849 RegisterMap::WalkContinuation::skip); 1850 frame fr = thread->last_frame(); 1851 while (fr.id() != id) { 1852 fr = fr.sender(®_map); 1853 } 1854 deoptimize(thread, fr, reason); 1855 } 1856 1857 1858 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1859 Thread* current = Thread::current(); 1860 if (thread == current || thread->is_handshake_safe_for(current)) { 1861 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1862 } else { 1863 VM_DeoptimizeFrame deopt(thread, id, reason); 1864 VMThread::execute(&deopt); 1865 } 1866 } 1867 1868 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1869 deoptimize_frame(thread, id, Reason_constraint); 1870 } 1871 1872 // JVMTI PopFrame support 1873 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1874 { 1875 assert(thread == JavaThread::current(), "pre-condition"); 1876 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1877 } 1878 JRT_END 1879 1880 MethodData* 1881 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1882 bool create_if_missing) { 1883 JavaThread* THREAD = thread; // For exception macros. 1884 MethodData* mdo = m()->method_data(); 1885 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) { 1886 // Build an MDO. Ignore errors like OutOfMemory; 1887 // that simply means we won't have an MDO to update. 1888 Method::build_profiling_method_data(m, THREAD); 1889 if (HAS_PENDING_EXCEPTION) { 1890 // Only metaspace OOM is expected. No Java code executed. 1891 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1892 CLEAR_PENDING_EXCEPTION; 1893 } 1894 mdo = m()->method_data(); 1895 } 1896 return mdo; 1897 } 1898 1899 #if COMPILER2_OR_JVMCI 1900 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1901 // In case of an unresolved klass entry, load the class. 1902 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1903 // and probably nowhere else. 1904 // Even that case would benefit from simply re-interpreting the 1905 // bytecode, without paying special attention to the class index. 1906 // So this whole "class index" feature should probably be removed. 1907 1908 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1909 Klass* tk = constant_pool->klass_at(index, THREAD); 1910 if (HAS_PENDING_EXCEPTION) { 1911 // Exception happened during classloading. We ignore the exception here, since it 1912 // is going to be rethrown since the current activation is going to be deoptimized and 1913 // the interpreter will re-execute the bytecode. 1914 // Do not clear probable Async Exceptions. 1915 CLEAR_PENDING_NONASYNC_EXCEPTION; 1916 // Class loading called java code which may have caused a stack 1917 // overflow. If the exception was thrown right before the return 1918 // to the runtime the stack is no longer guarded. Reguard the 1919 // stack otherwise if we return to the uncommon trap blob and the 1920 // stack bang causes a stack overflow we crash. 1921 JavaThread* jt = THREAD; 1922 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 1923 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1924 } 1925 return; 1926 } 1927 1928 assert(!constant_pool->tag_at(index).is_symbol(), 1929 "no symbolic names here, please"); 1930 } 1931 1932 #if INCLUDE_JFR 1933 1934 class DeoptReasonSerializer : public JfrSerializer { 1935 public: 1936 void serialize(JfrCheckpointWriter& writer) { 1937 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 1938 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 1939 writer.write_key((u8)i); 1940 writer.write(Deoptimization::trap_reason_name(i)); 1941 } 1942 } 1943 }; 1944 1945 class DeoptActionSerializer : public JfrSerializer { 1946 public: 1947 void serialize(JfrCheckpointWriter& writer) { 1948 static const u4 nof_actions = Deoptimization::Action_LIMIT; 1949 writer.write_count(nof_actions); 1950 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 1951 writer.write_key(i); 1952 writer.write(Deoptimization::trap_action_name((int)i)); 1953 } 1954 } 1955 }; 1956 1957 static void register_serializers() { 1958 static int critical_section = 0; 1959 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 1960 return; 1961 } 1962 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 1963 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 1964 } 1965 1966 static void post_deoptimization_event(nmethod* nm, 1967 const Method* method, 1968 int trap_bci, 1969 int instruction, 1970 Deoptimization::DeoptReason reason, 1971 Deoptimization::DeoptAction action) { 1972 assert(nm != nullptr, "invariant"); 1973 assert(method != nullptr, "invariant"); 1974 if (EventDeoptimization::is_enabled()) { 1975 static bool serializers_registered = false; 1976 if (!serializers_registered) { 1977 register_serializers(); 1978 serializers_registered = true; 1979 } 1980 EventDeoptimization event; 1981 event.set_compileId(nm->compile_id()); 1982 event.set_compiler(nm->compiler_type()); 1983 event.set_method(method); 1984 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 1985 event.set_bci(trap_bci); 1986 event.set_instruction(instruction); 1987 event.set_reason(reason); 1988 event.set_action(action); 1989 event.commit(); 1990 } 1991 } 1992 1993 #endif // INCLUDE_JFR 1994 1995 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 1996 const char* reason_name, const char* reason_action) { 1997 LogTarget(Debug, deoptimization) lt; 1998 if (lt.is_enabled()) { 1999 LogStream ls(lt); 2000 bool is_osr = nm->is_osr_method(); 2001 ls.print("cid=%4d %s level=%d", 2002 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level()); 2003 ls.print(" %s", tm->name_and_sig_as_C_string()); 2004 ls.print(" trap_bci=%d ", trap_bci); 2005 if (is_osr) { 2006 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 2007 } 2008 ls.print("%s ", reason_name); 2009 ls.print("%s ", reason_action); 2010 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 2011 pc, fr.pc() - nm->code_begin()); 2012 } 2013 } 2014 2015 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 2016 HandleMark hm(current); 2017 2018 // uncommon_trap() is called at the beginning of the uncommon trap 2019 // handler. Note this fact before we start generating temporary frames 2020 // that can confuse an asynchronous stack walker. This counter is 2021 // decremented at the end of unpack_frames(). 2022 2023 current->inc_in_deopt_handler(); 2024 2025 #if INCLUDE_JVMCI 2026 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 2027 RegisterMap reg_map(current, 2028 RegisterMap::UpdateMap::include, 2029 RegisterMap::ProcessFrames::include, 2030 RegisterMap::WalkContinuation::skip); 2031 #else 2032 RegisterMap reg_map(current, 2033 RegisterMap::UpdateMap::skip, 2034 RegisterMap::ProcessFrames::include, 2035 RegisterMap::WalkContinuation::skip); 2036 #endif 2037 frame stub_frame = current->last_frame(); 2038 frame fr = stub_frame.sender(®_map); 2039 2040 // Log a message 2041 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 2042 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 2043 2044 { 2045 ResourceMark rm; 2046 2047 DeoptReason reason = trap_request_reason(trap_request); 2048 DeoptAction action = trap_request_action(trap_request); 2049 #if INCLUDE_JVMCI 2050 int debug_id = trap_request_debug_id(trap_request); 2051 #endif 2052 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 2053 2054 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 2055 compiledVFrame* cvf = compiledVFrame::cast(vf); 2056 2057 nmethod* nm = cvf->code(); 2058 2059 ScopeDesc* trap_scope = cvf->scope(); 2060 2061 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2062 2063 if (is_receiver_constraint_failure) { 2064 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2065 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2066 JVMCI_ONLY(COMMA debug_id)); 2067 } 2068 2069 methodHandle trap_method(current, trap_scope->method()); 2070 int trap_bci = trap_scope->bci(); 2071 #if INCLUDE_JVMCI 2072 jlong speculation = current->pending_failed_speculation(); 2073 if (nm->is_compiled_by_jvmci()) { 2074 nm->update_speculation(current); 2075 } else { 2076 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2077 } 2078 2079 if (trap_bci == SynchronizationEntryBCI) { 2080 trap_bci = 0; 2081 current->set_pending_monitorenter(true); 2082 } 2083 2084 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2085 current->set_pending_transfer_to_interpreter(true); 2086 } 2087 #endif 2088 2089 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2090 // Record this event in the histogram. 2091 gather_statistics(reason, action, trap_bc); 2092 2093 // Ensure that we can record deopt. history: 2094 // Need MDO to record RTM code generation state. 2095 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 2096 2097 methodHandle profiled_method; 2098 #if INCLUDE_JVMCI 2099 if (nm->is_compiled_by_jvmci()) { 2100 profiled_method = methodHandle(current, nm->method()); 2101 } else { 2102 profiled_method = trap_method; 2103 } 2104 #else 2105 profiled_method = trap_method; 2106 #endif 2107 2108 MethodData* trap_mdo = 2109 get_method_data(current, profiled_method, create_if_missing); 2110 2111 { // Log Deoptimization event for JFR, UL and event system 2112 Method* tm = trap_method(); 2113 const char* reason_name = trap_reason_name(reason); 2114 const char* reason_action = trap_action_name(action); 2115 intptr_t pc = p2i(fr.pc()); 2116 2117 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2118 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2119 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 2120 reason_name, reason_action, pc, 2121 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 2122 } 2123 2124 // Print a bunch of diagnostics, if requested. 2125 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2126 ResourceMark rm; 2127 2128 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2129 // We must do this already now, since we cannot acquire this lock while 2130 // holding the tty lock (lock ordering by rank). 2131 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2132 2133 ttyLocker ttyl; 2134 2135 char buf[100]; 2136 if (xtty != nullptr) { 2137 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 2138 os::current_thread_id(), 2139 format_trap_request(buf, sizeof(buf), trap_request)); 2140 #if INCLUDE_JVMCI 2141 if (speculation != 0) { 2142 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2143 } 2144 #endif 2145 nm->log_identity(xtty); 2146 } 2147 Symbol* class_name = nullptr; 2148 bool unresolved = false; 2149 if (unloaded_class_index >= 0) { 2150 constantPoolHandle constants (current, trap_method->constants()); 2151 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2152 class_name = constants->klass_name_at(unloaded_class_index); 2153 unresolved = true; 2154 if (xtty != nullptr) 2155 xtty->print(" unresolved='1'"); 2156 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2157 class_name = constants->symbol_at(unloaded_class_index); 2158 } 2159 if (xtty != nullptr) 2160 xtty->name(class_name); 2161 } 2162 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) { 2163 // Dump the relevant MDO state. 2164 // This is the deopt count for the current reason, any previous 2165 // reasons or recompiles seen at this point. 2166 int dcnt = trap_mdo->trap_count(reason); 2167 if (dcnt != 0) 2168 xtty->print(" count='%d'", dcnt); 2169 2170 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to 2171 // lock extra_data_lock before the tty lock. 2172 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2173 int dos = (pdata == nullptr)? 0: pdata->trap_state(); 2174 if (dos != 0) { 2175 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2176 if (trap_state_is_recompiled(dos)) { 2177 int recnt2 = trap_mdo->overflow_recompile_count(); 2178 if (recnt2 != 0) 2179 xtty->print(" recompiles2='%d'", recnt2); 2180 } 2181 } 2182 } 2183 if (xtty != nullptr) { 2184 xtty->stamp(); 2185 xtty->end_head(); 2186 } 2187 if (TraceDeoptimization) { // make noise on the tty 2188 stringStream st; 2189 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2190 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2191 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2192 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 2193 #if INCLUDE_JVMCI 2194 if (nm->is_compiled_by_jvmci()) { 2195 const char* installed_code_name = nm->jvmci_name(); 2196 if (installed_code_name != nullptr) { 2197 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2198 } 2199 } 2200 #endif 2201 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2202 p2i(fr.pc()), 2203 os::current_thread_id(), 2204 trap_reason_name(reason), 2205 trap_action_name(action), 2206 unloaded_class_index 2207 #if INCLUDE_JVMCI 2208 , debug_id 2209 #endif 2210 ); 2211 if (class_name != nullptr) { 2212 st.print(unresolved ? " unresolved class: " : " symbol: "); 2213 class_name->print_symbol_on(&st); 2214 } 2215 st.cr(); 2216 tty->print_raw(st.freeze()); 2217 } 2218 if (xtty != nullptr) { 2219 // Log the precise location of the trap. 2220 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2221 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2222 xtty->method(sd->method()); 2223 xtty->end_elem(); 2224 if (sd->is_top()) break; 2225 } 2226 xtty->tail("uncommon_trap"); 2227 } 2228 } 2229 // (End diagnostic printout.) 2230 2231 if (is_receiver_constraint_failure) { 2232 fatal("missing receiver type check"); 2233 } 2234 2235 // Load class if necessary 2236 if (unloaded_class_index >= 0) { 2237 constantPoolHandle constants(current, trap_method->constants()); 2238 load_class_by_index(constants, unloaded_class_index, THREAD); 2239 } 2240 2241 // Flush the nmethod if necessary and desirable. 2242 // 2243 // We need to avoid situations where we are re-flushing the nmethod 2244 // because of a hot deoptimization site. Repeated flushes at the same 2245 // point need to be detected by the compiler and avoided. If the compiler 2246 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2247 // module must take measures to avoid an infinite cycle of recompilation 2248 // and deoptimization. There are several such measures: 2249 // 2250 // 1. If a recompilation is ordered a second time at some site X 2251 // and for the same reason R, the action is adjusted to 'reinterpret', 2252 // to give the interpreter time to exercise the method more thoroughly. 2253 // If this happens, the method's overflow_recompile_count is incremented. 2254 // 2255 // 2. If the compiler fails to reduce the deoptimization rate, then 2256 // the method's overflow_recompile_count will begin to exceed the set 2257 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2258 // is adjusted to 'make_not_compilable', and the method is abandoned 2259 // to the interpreter. This is a performance hit for hot methods, 2260 // but is better than a disastrous infinite cycle of recompilations. 2261 // (Actually, only the method containing the site X is abandoned.) 2262 // 2263 // 3. In parallel with the previous measures, if the total number of 2264 // recompilations of a method exceeds the much larger set limit 2265 // PerMethodRecompilationCutoff, the method is abandoned. 2266 // This should only happen if the method is very large and has 2267 // many "lukewarm" deoptimizations. The code which enforces this 2268 // limit is elsewhere (class nmethod, class Method). 2269 // 2270 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2271 // to recompile at each bytecode independently of the per-BCI cutoff. 2272 // 2273 // The decision to update code is up to the compiler, and is encoded 2274 // in the Action_xxx code. If the compiler requests Action_none 2275 // no trap state is changed, no compiled code is changed, and the 2276 // computation suffers along in the interpreter. 2277 // 2278 // The other action codes specify various tactics for decompilation 2279 // and recompilation. Action_maybe_recompile is the loosest, and 2280 // allows the compiled code to stay around until enough traps are seen, 2281 // and until the compiler gets around to recompiling the trapping method. 2282 // 2283 // The other actions cause immediate removal of the present code. 2284 2285 // Traps caused by injected profile shouldn't pollute trap counts. 2286 bool injected_profile_trap = trap_method->has_injected_profile() && 2287 (reason == Reason_intrinsic || reason == Reason_unreached); 2288 2289 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2290 bool make_not_entrant = false; 2291 bool make_not_compilable = false; 2292 bool reprofile = false; 2293 switch (action) { 2294 case Action_none: 2295 // Keep the old code. 2296 update_trap_state = false; 2297 break; 2298 case Action_maybe_recompile: 2299 // Do not need to invalidate the present code, but we can 2300 // initiate another 2301 // Start compiler without (necessarily) invalidating the nmethod. 2302 // The system will tolerate the old code, but new code should be 2303 // generated when possible. 2304 break; 2305 case Action_reinterpret: 2306 // Go back into the interpreter for a while, and then consider 2307 // recompiling form scratch. 2308 make_not_entrant = true; 2309 // Reset invocation counter for outer most method. 2310 // This will allow the interpreter to exercise the bytecodes 2311 // for a while before recompiling. 2312 // By contrast, Action_make_not_entrant is immediate. 2313 // 2314 // Note that the compiler will track null_check, null_assert, 2315 // range_check, and class_check events and log them as if they 2316 // had been traps taken from compiled code. This will update 2317 // the MDO trap history so that the next compilation will 2318 // properly detect hot trap sites. 2319 reprofile = true; 2320 break; 2321 case Action_make_not_entrant: 2322 // Request immediate recompilation, and get rid of the old code. 2323 // Make them not entrant, so next time they are called they get 2324 // recompiled. Unloaded classes are loaded now so recompile before next 2325 // time they are called. Same for uninitialized. The interpreter will 2326 // link the missing class, if any. 2327 make_not_entrant = true; 2328 break; 2329 case Action_make_not_compilable: 2330 // Give up on compiling this method at all. 2331 make_not_entrant = true; 2332 make_not_compilable = true; 2333 break; 2334 default: 2335 ShouldNotReachHere(); 2336 } 2337 2338 // Setting +ProfileTraps fixes the following, on all platforms: 2339 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2340 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2341 // recompile relies on a MethodData* to record heroic opt failures. 2342 2343 // Whether the interpreter is producing MDO data or not, we also need 2344 // to use the MDO to detect hot deoptimization points and control 2345 // aggressive optimization. 2346 bool inc_recompile_count = false; 2347 2348 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2349 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr, 2350 (trap_mdo != nullptr), 2351 Mutex::_no_safepoint_check_flag); 2352 ProfileData* pdata = nullptr; 2353 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) { 2354 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2355 uint this_trap_count = 0; 2356 bool maybe_prior_trap = false; 2357 bool maybe_prior_recompile = false; 2358 2359 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2360 #if INCLUDE_JVMCI 2361 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2362 #endif 2363 nm->method(), 2364 //outputs: 2365 this_trap_count, 2366 maybe_prior_trap, 2367 maybe_prior_recompile); 2368 // Because the interpreter also counts null, div0, range, and class 2369 // checks, these traps from compiled code are double-counted. 2370 // This is harmless; it just means that the PerXTrapLimit values 2371 // are in effect a little smaller than they look. 2372 2373 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2374 if (per_bc_reason != Reason_none) { 2375 // Now take action based on the partially known per-BCI history. 2376 if (maybe_prior_trap 2377 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2378 // If there are too many traps at this BCI, force a recompile. 2379 // This will allow the compiler to see the limit overflow, and 2380 // take corrective action, if possible. The compiler generally 2381 // does not use the exact PerBytecodeTrapLimit value, but instead 2382 // changes its tactics if it sees any traps at all. This provides 2383 // a little hysteresis, delaying a recompile until a trap happens 2384 // several times. 2385 // 2386 // Actually, since there is only one bit of counter per BCI, 2387 // the possible per-BCI counts are {0,1,(per-method count)}. 2388 // This produces accurate results if in fact there is only 2389 // one hot trap site, but begins to get fuzzy if there are 2390 // many sites. For example, if there are ten sites each 2391 // trapping two or more times, they each get the blame for 2392 // all of their traps. 2393 make_not_entrant = true; 2394 } 2395 2396 // Detect repeated recompilation at the same BCI, and enforce a limit. 2397 if (make_not_entrant && maybe_prior_recompile) { 2398 // More than one recompile at this point. 2399 inc_recompile_count = maybe_prior_trap; 2400 } 2401 } else { 2402 // For reasons which are not recorded per-bytecode, we simply 2403 // force recompiles unconditionally. 2404 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2405 make_not_entrant = true; 2406 } 2407 2408 // Go back to the compiler if there are too many traps in this method. 2409 if (this_trap_count >= per_method_trap_limit(reason)) { 2410 // If there are too many traps in this method, force a recompile. 2411 // This will allow the compiler to see the limit overflow, and 2412 // take corrective action, if possible. 2413 // (This condition is an unlikely backstop only, because the 2414 // PerBytecodeTrapLimit is more likely to take effect first, 2415 // if it is applicable.) 2416 make_not_entrant = true; 2417 } 2418 2419 // Here's more hysteresis: If there has been a recompile at 2420 // this trap point already, run the method in the interpreter 2421 // for a while to exercise it more thoroughly. 2422 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2423 reprofile = true; 2424 } 2425 } 2426 2427 // Take requested actions on the method: 2428 2429 // Recompile 2430 if (make_not_entrant) { 2431 if (!nm->make_not_entrant()) { 2432 return; // the call did not change nmethod's state 2433 } 2434 2435 if (pdata != nullptr) { 2436 // Record the recompilation event, if any. 2437 int tstate0 = pdata->trap_state(); 2438 int tstate1 = trap_state_set_recompiled(tstate0, true); 2439 if (tstate1 != tstate0) 2440 pdata->set_trap_state(tstate1); 2441 } 2442 2443 #if INCLUDE_RTM_OPT 2444 // Restart collecting RTM locking abort statistic if the method 2445 // is recompiled for a reason other than RTM state change. 2446 // Assume that in new recompiled code the statistic could be different, 2447 // for example, due to different inlining. 2448 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) && 2449 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 2450 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2451 } 2452 #endif 2453 // For code aging we count traps separately here, using make_not_entrant() 2454 // as a guard against simultaneous deopts in multiple threads. 2455 if (reason == Reason_tenured && trap_mdo != nullptr) { 2456 trap_mdo->inc_tenure_traps(); 2457 } 2458 } 2459 2460 if (inc_recompile_count) { 2461 trap_mdo->inc_overflow_recompile_count(); 2462 if ((uint)trap_mdo->overflow_recompile_count() > 2463 (uint)PerBytecodeRecompilationCutoff) { 2464 // Give up on the method containing the bad BCI. 2465 if (trap_method() == nm->method()) { 2466 make_not_compilable = true; 2467 } else { 2468 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2469 // But give grace to the enclosing nm->method(). 2470 } 2471 } 2472 } 2473 2474 // Reprofile 2475 if (reprofile) { 2476 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2477 } 2478 2479 // Give up compiling 2480 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2481 assert(make_not_entrant, "consistent"); 2482 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2483 } 2484 2485 if (ProfileExceptionHandlers && trap_mdo != nullptr) { 2486 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci); 2487 if (exception_handler_data != nullptr) { 2488 // uncommon trap at the start of an exception handler. 2489 // C2 generates these for un-entered exception handlers. 2490 // mark the handler as entered to avoid generating 2491 // another uncommon trap the next time the handler is compiled 2492 exception_handler_data->set_exception_handler_entered(); 2493 } 2494 } 2495 2496 } // Free marked resources 2497 2498 } 2499 JRT_END 2500 2501 ProfileData* 2502 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2503 int trap_bci, 2504 Deoptimization::DeoptReason reason, 2505 bool update_total_trap_count, 2506 #if INCLUDE_JVMCI 2507 bool is_osr, 2508 #endif 2509 Method* compiled_method, 2510 //outputs: 2511 uint& ret_this_trap_count, 2512 bool& ret_maybe_prior_trap, 2513 bool& ret_maybe_prior_recompile) { 2514 trap_mdo->check_extra_data_locked(); 2515 2516 bool maybe_prior_trap = false; 2517 bool maybe_prior_recompile = false; 2518 uint this_trap_count = 0; 2519 if (update_total_trap_count) { 2520 uint idx = reason; 2521 #if INCLUDE_JVMCI 2522 if (is_osr) { 2523 // Upper half of history array used for traps in OSR compilations 2524 idx += Reason_TRAP_HISTORY_LENGTH; 2525 } 2526 #endif 2527 uint prior_trap_count = trap_mdo->trap_count(idx); 2528 this_trap_count = trap_mdo->inc_trap_count(idx); 2529 2530 // If the runtime cannot find a place to store trap history, 2531 // it is estimated based on the general condition of the method. 2532 // If the method has ever been recompiled, or has ever incurred 2533 // a trap with the present reason , then this BCI is assumed 2534 // (pessimistically) to be the culprit. 2535 maybe_prior_trap = (prior_trap_count != 0); 2536 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2537 } 2538 ProfileData* pdata = nullptr; 2539 2540 2541 // For reasons which are recorded per bytecode, we check per-BCI data. 2542 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2543 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2544 if (per_bc_reason != Reason_none) { 2545 // Find the profile data for this BCI. If there isn't one, 2546 // try to allocate one from the MDO's set of spares. 2547 // This will let us detect a repeated trap at this point. 2548 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr); 2549 2550 if (pdata != nullptr) { 2551 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2552 if (LogCompilation && xtty != nullptr) { 2553 ttyLocker ttyl; 2554 // no more room for speculative traps in this MDO 2555 xtty->elem("speculative_traps_oom"); 2556 } 2557 } 2558 // Query the trap state of this profile datum. 2559 int tstate0 = pdata->trap_state(); 2560 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2561 maybe_prior_trap = false; 2562 if (!trap_state_is_recompiled(tstate0)) 2563 maybe_prior_recompile = false; 2564 2565 // Update the trap state of this profile datum. 2566 int tstate1 = tstate0; 2567 // Record the reason. 2568 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2569 // Store the updated state on the MDO, for next time. 2570 if (tstate1 != tstate0) 2571 pdata->set_trap_state(tstate1); 2572 } else { 2573 if (LogCompilation && xtty != nullptr) { 2574 ttyLocker ttyl; 2575 // Missing MDP? Leave a small complaint in the log. 2576 xtty->elem("missing_mdp bci='%d'", trap_bci); 2577 } 2578 } 2579 } 2580 2581 // Return results: 2582 ret_this_trap_count = this_trap_count; 2583 ret_maybe_prior_trap = maybe_prior_trap; 2584 ret_maybe_prior_recompile = maybe_prior_recompile; 2585 return pdata; 2586 } 2587 2588 void 2589 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2590 ResourceMark rm; 2591 // Ignored outputs: 2592 uint ignore_this_trap_count; 2593 bool ignore_maybe_prior_trap; 2594 bool ignore_maybe_prior_recompile; 2595 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2596 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2597 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2598 2599 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2600 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2601 2602 query_update_method_data(trap_mdo, trap_bci, 2603 (DeoptReason)reason, 2604 update_total_counts, 2605 #if INCLUDE_JVMCI 2606 false, 2607 #endif 2608 nullptr, 2609 ignore_this_trap_count, 2610 ignore_maybe_prior_trap, 2611 ignore_maybe_prior_recompile); 2612 } 2613 2614 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) { 2615 // Enable WXWrite: current function is called from methods compiled by C2 directly 2616 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2617 2618 // Still in Java no safepoints 2619 { 2620 // This enters VM and may safepoint 2621 uncommon_trap_inner(current, trap_request); 2622 } 2623 HandleMark hm(current); 2624 return fetch_unroll_info_helper(current, exec_mode); 2625 } 2626 2627 // Local derived constants. 2628 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2629 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2630 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2631 2632 //---------------------------trap_state_reason--------------------------------- 2633 Deoptimization::DeoptReason 2634 Deoptimization::trap_state_reason(int trap_state) { 2635 // This assert provides the link between the width of DataLayout::trap_bits 2636 // and the encoding of "recorded" reasons. It ensures there are enough 2637 // bits to store all needed reasons in the per-BCI MDO profile. 2638 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2639 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2640 trap_state -= recompile_bit; 2641 if (trap_state == DS_REASON_MASK) { 2642 return Reason_many; 2643 } else { 2644 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2645 return (DeoptReason)trap_state; 2646 } 2647 } 2648 //-------------------------trap_state_has_reason------------------------------- 2649 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2650 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2651 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2652 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2653 trap_state -= recompile_bit; 2654 if (trap_state == DS_REASON_MASK) { 2655 return -1; // true, unspecifically (bottom of state lattice) 2656 } else if (trap_state == reason) { 2657 return 1; // true, definitely 2658 } else if (trap_state == 0) { 2659 return 0; // false, definitely (top of state lattice) 2660 } else { 2661 return 0; // false, definitely 2662 } 2663 } 2664 //-------------------------trap_state_add_reason------------------------------- 2665 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2666 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2667 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2668 trap_state -= recompile_bit; 2669 if (trap_state == DS_REASON_MASK) { 2670 return trap_state + recompile_bit; // already at state lattice bottom 2671 } else if (trap_state == reason) { 2672 return trap_state + recompile_bit; // the condition is already true 2673 } else if (trap_state == 0) { 2674 return reason + recompile_bit; // no condition has yet been true 2675 } else { 2676 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2677 } 2678 } 2679 //-----------------------trap_state_is_recompiled------------------------------ 2680 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2681 return (trap_state & DS_RECOMPILE_BIT) != 0; 2682 } 2683 //-----------------------trap_state_set_recompiled----------------------------- 2684 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2685 if (z) return trap_state | DS_RECOMPILE_BIT; 2686 else return trap_state & ~DS_RECOMPILE_BIT; 2687 } 2688 //---------------------------format_trap_state--------------------------------- 2689 // This is used for debugging and diagnostics, including LogFile output. 2690 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2691 int trap_state) { 2692 assert(buflen > 0, "sanity"); 2693 DeoptReason reason = trap_state_reason(trap_state); 2694 bool recomp_flag = trap_state_is_recompiled(trap_state); 2695 // Re-encode the state from its decoded components. 2696 int decoded_state = 0; 2697 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2698 decoded_state = trap_state_add_reason(decoded_state, reason); 2699 if (recomp_flag) 2700 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2701 // If the state re-encodes properly, format it symbolically. 2702 // Because this routine is used for debugging and diagnostics, 2703 // be robust even if the state is a strange value. 2704 size_t len; 2705 if (decoded_state != trap_state) { 2706 // Random buggy state that doesn't decode?? 2707 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2708 } else { 2709 len = jio_snprintf(buf, buflen, "%s%s", 2710 trap_reason_name(reason), 2711 recomp_flag ? " recompiled" : ""); 2712 } 2713 return buf; 2714 } 2715 2716 2717 //--------------------------------statics-------------------------------------- 2718 const char* Deoptimization::_trap_reason_name[] = { 2719 // Note: Keep this in sync. with enum DeoptReason. 2720 "none", 2721 "null_check", 2722 "null_assert" JVMCI_ONLY("_or_unreached0"), 2723 "range_check", 2724 "class_check", 2725 "array_check", 2726 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2727 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2728 "profile_predicate", 2729 "unloaded", 2730 "uninitialized", 2731 "initialized", 2732 "unreached", 2733 "unhandled", 2734 "constraint", 2735 "div0_check", 2736 "age", 2737 "predicate", 2738 "loop_limit_check", 2739 "speculate_class_check", 2740 "speculate_null_check", 2741 "speculate_null_assert", 2742 "rtm_state_change", 2743 "unstable_if", 2744 "unstable_fused_if", 2745 "receiver_constraint", 2746 #if INCLUDE_JVMCI 2747 "aliasing", 2748 "transfer_to_interpreter", 2749 "not_compiled_exception_handler", 2750 "unresolved", 2751 "jsr_mismatch", 2752 #endif 2753 "tenured" 2754 }; 2755 const char* Deoptimization::_trap_action_name[] = { 2756 // Note: Keep this in sync. with enum DeoptAction. 2757 "none", 2758 "maybe_recompile", 2759 "reinterpret", 2760 "make_not_entrant", 2761 "make_not_compilable" 2762 }; 2763 2764 const char* Deoptimization::trap_reason_name(int reason) { 2765 // Check that every reason has a name 2766 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2767 2768 if (reason == Reason_many) return "many"; 2769 if ((uint)reason < Reason_LIMIT) 2770 return _trap_reason_name[reason]; 2771 static char buf[20]; 2772 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2773 return buf; 2774 } 2775 const char* Deoptimization::trap_action_name(int action) { 2776 // Check that every action has a name 2777 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2778 2779 if ((uint)action < Action_LIMIT) 2780 return _trap_action_name[action]; 2781 static char buf[20]; 2782 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2783 return buf; 2784 } 2785 2786 // This is used for debugging and diagnostics, including LogFile output. 2787 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2788 int trap_request) { 2789 jint unloaded_class_index = trap_request_index(trap_request); 2790 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2791 const char* action = trap_action_name(trap_request_action(trap_request)); 2792 #if INCLUDE_JVMCI 2793 int debug_id = trap_request_debug_id(trap_request); 2794 #endif 2795 size_t len; 2796 if (unloaded_class_index < 0) { 2797 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2798 reason, action 2799 #if INCLUDE_JVMCI 2800 ,debug_id 2801 #endif 2802 ); 2803 } else { 2804 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2805 reason, action, unloaded_class_index 2806 #if INCLUDE_JVMCI 2807 ,debug_id 2808 #endif 2809 ); 2810 } 2811 return buf; 2812 } 2813 2814 juint Deoptimization::_deoptimization_hist 2815 [Deoptimization::Reason_LIMIT] 2816 [1 + Deoptimization::Action_LIMIT] 2817 [Deoptimization::BC_CASE_LIMIT] 2818 = {0}; 2819 2820 enum { 2821 LSB_BITS = 8, 2822 LSB_MASK = right_n_bits(LSB_BITS) 2823 }; 2824 2825 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2826 Bytecodes::Code bc) { 2827 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2828 assert(action >= 0 && action < Action_LIMIT, "oob"); 2829 _deoptimization_hist[Reason_none][0][0] += 1; // total 2830 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2831 juint* cases = _deoptimization_hist[reason][1+action]; 2832 juint* bc_counter_addr = nullptr; 2833 juint bc_counter = 0; 2834 // Look for an unused counter, or an exact match to this BC. 2835 if (bc != Bytecodes::_illegal) { 2836 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2837 juint* counter_addr = &cases[bc_case]; 2838 juint counter = *counter_addr; 2839 if ((counter == 0 && bc_counter_addr == nullptr) 2840 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2841 // this counter is either free or is already devoted to this BC 2842 bc_counter_addr = counter_addr; 2843 bc_counter = counter | bc; 2844 } 2845 } 2846 } 2847 if (bc_counter_addr == nullptr) { 2848 // Overflow, or no given bytecode. 2849 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2850 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2851 } 2852 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2853 } 2854 2855 jint Deoptimization::total_deoptimization_count() { 2856 return _deoptimization_hist[Reason_none][0][0]; 2857 } 2858 2859 // Get the deopt count for a specific reason and a specific action. If either 2860 // one of 'reason' or 'action' is null, the method returns the sum of all 2861 // deoptimizations with the specific 'action' or 'reason' respectively. 2862 // If both arguments are null, the method returns the total deopt count. 2863 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2864 if (reason_str == nullptr && action_str == nullptr) { 2865 return total_deoptimization_count(); 2866 } 2867 juint counter = 0; 2868 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2869 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) { 2870 for (int action = 0; action < Action_LIMIT; action++) { 2871 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) { 2872 juint* cases = _deoptimization_hist[reason][1+action]; 2873 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2874 counter += cases[bc_case] >> LSB_BITS; 2875 } 2876 } 2877 } 2878 } 2879 } 2880 return counter; 2881 } 2882 2883 void Deoptimization::print_statistics() { 2884 juint total = total_deoptimization_count(); 2885 juint account = total; 2886 if (total != 0) { 2887 ttyLocker ttyl; 2888 if (xtty != nullptr) xtty->head("statistics type='deoptimization'"); 2889 tty->print_cr("Deoptimization traps recorded:"); 2890 #define PRINT_STAT_LINE(name, r) \ 2891 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2892 PRINT_STAT_LINE("total", total); 2893 // For each non-zero entry in the histogram, print the reason, 2894 // the action, and (if specifically known) the type of bytecode. 2895 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2896 for (int action = 0; action < Action_LIMIT; action++) { 2897 juint* cases = _deoptimization_hist[reason][1+action]; 2898 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2899 juint counter = cases[bc_case]; 2900 if (counter != 0) { 2901 char name[1*K]; 2902 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2903 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2904 bc = Bytecodes::_illegal; 2905 os::snprintf_checked(name, sizeof(name), "%s/%s/%s", 2906 trap_reason_name(reason), 2907 trap_action_name(action), 2908 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2909 juint r = counter >> LSB_BITS; 2910 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2911 account -= r; 2912 } 2913 } 2914 } 2915 } 2916 if (account != 0) { 2917 PRINT_STAT_LINE("unaccounted", account); 2918 } 2919 #undef PRINT_STAT_LINE 2920 if (xtty != nullptr) xtty->tail("statistics"); 2921 } 2922 } 2923 2924 #else // COMPILER2_OR_JVMCI 2925 2926 2927 // Stubs for C1 only system. 2928 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2929 return false; 2930 } 2931 2932 const char* Deoptimization::trap_reason_name(int reason) { 2933 return "unknown"; 2934 } 2935 2936 jint Deoptimization::total_deoptimization_count() { 2937 return 0; 2938 } 2939 2940 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2941 return 0; 2942 } 2943 2944 void Deoptimization::print_statistics() { 2945 // no output 2946 } 2947 2948 void 2949 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2950 // no update 2951 } 2952 2953 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2954 return 0; 2955 } 2956 2957 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2958 Bytecodes::Code bc) { 2959 // no update 2960 } 2961 2962 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2963 int trap_state) { 2964 jio_snprintf(buf, buflen, "#%d", trap_state); 2965 return buf; 2966 } 2967 2968 #endif // COMPILER2_OR_JVMCI