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/flatArrayKlass.hpp" 53 #include "oops/flatArrayOop.hpp" 54 #include "oops/fieldStreams.inline.hpp" 55 #include "oops/method.hpp" 56 #include "oops/objArrayKlass.hpp" 57 #include "oops/objArrayOop.inline.hpp" 58 #include "oops/oop.inline.hpp" 59 #include "oops/inlineKlass.inline.hpp" 60 #include "oops/typeArrayOop.inline.hpp" 61 #include "oops/verifyOopClosure.hpp" 62 #include "prims/jvmtiDeferredUpdates.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "prims/jvmtiThreadState.hpp" 65 #include "prims/methodHandles.hpp" 66 #include "prims/vectorSupport.hpp" 67 #include "runtime/atomic.hpp" 68 #include "runtime/continuation.hpp" 69 #include "runtime/continuationEntry.inline.hpp" 70 #include "runtime/deoptimization.hpp" 71 #include "runtime/escapeBarrier.hpp" 72 #include "runtime/fieldDescriptor.hpp" 73 #include "runtime/fieldDescriptor.inline.hpp" 74 #include "runtime/frame.inline.hpp" 75 #include "runtime/handles.inline.hpp" 76 #include "runtime/interfaceSupport.inline.hpp" 77 #include "runtime/javaThread.hpp" 78 #include "runtime/jniHandles.inline.hpp" 79 #include "runtime/keepStackGCProcessed.hpp" 80 #include "runtime/objectMonitor.inline.hpp" 81 #include "runtime/osThread.hpp" 82 #include "runtime/safepointVerifiers.hpp" 83 #include "runtime/sharedRuntime.hpp" 84 #include "runtime/signature.hpp" 85 #include "runtime/stackFrameStream.inline.hpp" 86 #include "runtime/stackValue.hpp" 87 #include "runtime/stackWatermarkSet.hpp" 88 #include "runtime/stubRoutines.hpp" 89 #include "runtime/synchronizer.hpp" 90 #include "runtime/threadSMR.hpp" 91 #include "runtime/threadWXSetters.inline.hpp" 92 #include "runtime/vframe.hpp" 93 #include "runtime/vframeArray.hpp" 94 #include "runtime/vframe_hp.hpp" 95 #include "runtime/vmOperations.hpp" 96 #include "utilities/checkedCast.hpp" 97 #include "utilities/events.hpp" 98 #include "utilities/growableArray.hpp" 99 #include "utilities/macros.hpp" 100 #include "utilities/preserveException.hpp" 101 #include "utilities/xmlstream.hpp" 102 #if INCLUDE_JFR 103 #include "jfr/jfrEvents.hpp" 104 #include "jfr/metadata/jfrSerializer.hpp" 105 #endif 106 107 uint64_t DeoptimizationScope::_committed_deopt_gen = 0; 108 uint64_t DeoptimizationScope::_active_deopt_gen = 1; 109 bool DeoptimizationScope::_committing_in_progress = false; 110 111 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) { 112 DEBUG_ONLY(_deopted = false;) 113 114 MutexLocker ml(CompiledMethod_lock, Mutex::_no_safepoint_check_flag); 115 // If there is nothing to deopt _required_gen is the same as comitted. 116 _required_gen = DeoptimizationScope::_committed_deopt_gen; 117 } 118 119 DeoptimizationScope::~DeoptimizationScope() { 120 assert(_deopted, "Deopt not executed"); 121 } 122 123 void DeoptimizationScope::mark(CompiledMethod* cm, bool inc_recompile_counts) { 124 ConditionalMutexLocker ml(CompiledMethod_lock, !CompiledMethod_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 125 126 // If it's already marked but we still need it to be deopted. 127 if (cm->is_marked_for_deoptimization()) { 128 dependent(cm); 129 return; 130 } 131 132 CompiledMethod::DeoptimizationStatus status = 133 inc_recompile_counts ? CompiledMethod::deoptimize : CompiledMethod::deoptimize_noupdate; 134 Atomic::store(&cm->_deoptimization_status, status); 135 136 // Make sure active is not committed 137 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be"); 138 assert(cm->_deoptimization_generation == 0, "Is already marked"); 139 140 cm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen; 141 _required_gen = DeoptimizationScope::_active_deopt_gen; 142 } 143 144 void DeoptimizationScope::dependent(CompiledMethod* cm) { 145 ConditionalMutexLocker ml(CompiledMethod_lock, !CompiledMethod_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 146 147 // A method marked by someone else may have a _required_gen lower than what we marked with. 148 // Therefore only store it if it's higher than _required_gen. 149 if (_required_gen < cm->_deoptimization_generation) { 150 _required_gen = cm->_deoptimization_generation; 151 } 152 } 153 154 void DeoptimizationScope::deoptimize_marked() { 155 assert(!_deopted, "Already deopted"); 156 157 // We are not alive yet. 158 if (!Universe::is_fully_initialized()) { 159 DEBUG_ONLY(_deopted = true;) 160 return; 161 } 162 163 // Safepoints are a special case, handled here. 164 if (SafepointSynchronize::is_at_safepoint()) { 165 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen; 166 DeoptimizationScope::_active_deopt_gen++; 167 Deoptimization::deoptimize_all_marked(); 168 DEBUG_ONLY(_deopted = true;) 169 return; 170 } 171 172 uint64_t comitting = 0; 173 bool wait = false; 174 while (true) { 175 { 176 ConditionalMutexLocker ml(CompiledMethod_lock, !CompiledMethod_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 177 178 // First we check if we or someone else already deopted the gen we want. 179 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) { 180 DEBUG_ONLY(_deopted = true;) 181 return; 182 } 183 if (!_committing_in_progress) { 184 // The version we are about to commit. 185 comitting = DeoptimizationScope::_active_deopt_gen; 186 // Make sure new marks use a higher gen. 187 DeoptimizationScope::_active_deopt_gen++; 188 _committing_in_progress = true; 189 wait = false; 190 } else { 191 // Another thread is handshaking and committing a gen. 192 wait = true; 193 } 194 } 195 if (wait) { 196 // Wait and let the concurrent handshake be performed. 197 ThreadBlockInVM tbivm(JavaThread::current()); 198 os::naked_yield(); 199 } else { 200 // Performs the handshake. 201 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred. 202 DEBUG_ONLY(_deopted = true;) 203 { 204 ConditionalMutexLocker ml(CompiledMethod_lock, !CompiledMethod_lock->owned_by_self(), Mutex::_no_safepoint_check_flag); 205 206 // Make sure that committed doesn't go backwards. 207 // Should only happen if we did a deopt during a safepoint above. 208 if (DeoptimizationScope::_committed_deopt_gen < comitting) { 209 DeoptimizationScope::_committed_deopt_gen = comitting; 210 } 211 _committing_in_progress = false; 212 213 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be"); 214 215 return; 216 } 217 } 218 } 219 } 220 221 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 222 int caller_adjustment, 223 int caller_actual_parameters, 224 int number_of_frames, 225 intptr_t* frame_sizes, 226 address* frame_pcs, 227 BasicType return_type, 228 int exec_mode) { 229 _size_of_deoptimized_frame = size_of_deoptimized_frame; 230 _caller_adjustment = caller_adjustment; 231 _caller_actual_parameters = caller_actual_parameters; 232 _number_of_frames = number_of_frames; 233 _frame_sizes = frame_sizes; 234 _frame_pcs = frame_pcs; 235 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 236 _return_type = return_type; 237 _initial_info = 0; 238 // PD (x86 only) 239 _counter_temp = 0; 240 _unpack_kind = exec_mode; 241 _sender_sp_temp = 0; 242 243 _total_frame_sizes = size_of_frames(); 244 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 245 } 246 247 Deoptimization::UnrollBlock::~UnrollBlock() { 248 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 249 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 250 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 251 } 252 253 int Deoptimization::UnrollBlock::size_of_frames() const { 254 // Account first for the adjustment of the initial frame 255 intptr_t result = _caller_adjustment; 256 for (int index = 0; index < number_of_frames(); index++) { 257 result += frame_sizes()[index]; 258 } 259 return checked_cast<int>(result); 260 } 261 262 void Deoptimization::UnrollBlock::print() { 263 ResourceMark rm; 264 stringStream st; 265 st.print_cr("UnrollBlock"); 266 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 267 st.print( " frame_sizes: "); 268 for (int index = 0; index < number_of_frames(); index++) { 269 st.print(INTX_FORMAT " ", frame_sizes()[index]); 270 } 271 st.cr(); 272 tty->print_raw(st.freeze()); 273 } 274 275 // In order to make fetch_unroll_info work properly with escape 276 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 277 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 278 // which is called from the method fetch_unroll_info_helper below. 279 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 280 // fetch_unroll_info() is called at the beginning of the deoptimization 281 // handler. Note this fact before we start generating temporary frames 282 // that can confuse an asynchronous stack walker. This counter is 283 // decremented at the end of unpack_frames(). 284 current->inc_in_deopt_handler(); 285 286 if (exec_mode == Unpack_exception) { 287 // When we get here, a callee has thrown an exception into a deoptimized 288 // frame. That throw might have deferred stack watermark checking until 289 // after unwinding. So we deal with such deferred requests here. 290 StackWatermarkSet::after_unwind(current); 291 } 292 293 return fetch_unroll_info_helper(current, exec_mode); 294 JRT_END 295 296 #if COMPILER2_OR_JVMCI 297 // print information about reallocated objects 298 static void print_objects(JavaThread* deoptee_thread, 299 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 300 ResourceMark rm; 301 stringStream st; // change to logStream with logging 302 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 303 fieldDescriptor fd; 304 305 for (int i = 0; i < objects->length(); i++) { 306 ObjectValue* sv = (ObjectValue*) objects->at(i); 307 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 308 Handle obj = sv->value(); 309 310 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 311 k->print_value_on(&st); 312 assert(obj.not_null() || k->is_inline_klass() || realloc_failures, "reallocation was missed"); 313 if (obj.is_null()) { 314 if (k->is_inline_klass()) { 315 st.print(" is null"); 316 } else { 317 st.print(" allocation failed"); 318 } 319 } else { 320 st.print(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize); 321 } 322 st.cr(); 323 324 if (Verbose && !obj.is_null()) { 325 k->oop_print_on(obj(), &st); 326 } 327 } 328 tty->print_raw(st.freeze()); 329 } 330 331 static bool rematerialize_objects(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 332 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 333 bool& deoptimized_objects) { 334 bool realloc_failures = false; 335 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames"); 336 337 JavaThread* deoptee_thread = chunk->at(0)->thread(); 338 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 339 "a frame can only be deoptimized by the owner thread"); 340 341 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map); 342 343 // The flag return_oop() indicates call sites which return oop 344 // in compiled code. Such sites include java method calls, 345 // runtime calls (for example, used to allocate new objects/arrays 346 // on slow code path) and any other calls generated in compiled code. 347 // It is not guaranteed that we can get such information here only 348 // by analyzing bytecode in deoptimized frames. This is why this flag 349 // is set during method compilation (see Compile::Process_OopMap_Node()). 350 // If the previous frame was popped or if we are dispatching an exception, 351 // we don't have an oop result. 352 ScopeDesc* scope = chunk->at(0)->scope(); 353 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 354 // In case of the return of multiple values, we must take care 355 // of all oop return values. 356 GrowableArray<Handle> return_oops; 357 InlineKlass* vk = nullptr; 358 if (save_oop_result && scope->return_scalarized()) { 359 vk = InlineKlass::returned_inline_klass(map); 360 if (vk != nullptr) { 361 vk->save_oop_fields(map, return_oops); 362 save_oop_result = false; 363 } 364 } 365 if (save_oop_result) { 366 // Reallocation may trigger GC. If deoptimization happened on return from 367 // call which returns oop we need to save it since it is not in oopmap. 368 oop result = deoptee.saved_oop_result(&map); 369 assert(oopDesc::is_oop_or_null(result), "must be oop"); 370 return_oops.push(Handle(thread, result)); 371 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 372 if (TraceDeoptimization) { 373 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 374 tty->cr(); 375 } 376 } 377 if (objects != nullptr || vk != nullptr) { 378 if (exec_mode == Deoptimization::Unpack_none) { 379 assert(thread->thread_state() == _thread_in_vm, "assumption"); 380 JavaThread* THREAD = thread; // For exception macros. 381 // Clear pending OOM if reallocation fails and return true indicating allocation failure 382 if (vk != nullptr) { 383 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, CHECK_AND_CLEAR_(true)); 384 } 385 if (objects != nullptr) { 386 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 387 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 388 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, CHECK_AND_CLEAR_(true)); 389 } 390 deoptimized_objects = true; 391 } else { 392 JavaThread* current = thread; // For JRT_BLOCK 393 JRT_BLOCK 394 if (vk != nullptr) { 395 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, THREAD); 396 } 397 if (objects != nullptr) { 398 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 399 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 400 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 401 } 402 JRT_END 403 } 404 if (TraceDeoptimization && objects != nullptr) { 405 print_objects(deoptee_thread, objects, realloc_failures); 406 } 407 } 408 if (save_oop_result || vk != nullptr) { 409 // Restore result. 410 assert(return_oops.length() == 1, "no inline type"); 411 deoptee.set_saved_oop_result(&map, return_oops.pop()()); 412 } 413 return realloc_failures; 414 } 415 416 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 417 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 418 JavaThread* deoptee_thread = chunk->at(0)->thread(); 419 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 420 assert(thread == Thread::current(), "should be"); 421 HandleMark hm(thread); 422 #ifndef PRODUCT 423 bool first = true; 424 #endif // !PRODUCT 425 // Start locking from outermost/oldest frame 426 for (int i = (chunk->length() - 1); i >= 0; i--) { 427 compiledVFrame* cvf = chunk->at(i); 428 assert (cvf->scope() != nullptr,"expect only compiled java frames"); 429 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 430 if (monitors->is_nonempty()) { 431 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 432 exec_mode, realloc_failures); 433 deoptimized_objects = deoptimized_objects || relocked; 434 #ifndef PRODUCT 435 if (PrintDeoptimizationDetails) { 436 ResourceMark rm; 437 stringStream st; 438 for (int j = 0; j < monitors->length(); j++) { 439 MonitorInfo* mi = monitors->at(j); 440 if (mi->eliminated()) { 441 if (first) { 442 first = false; 443 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 444 } 445 if (exec_mode == Deoptimization::Unpack_none) { 446 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 447 if (monitor != nullptr && monitor->object() == mi->owner()) { 448 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 449 continue; 450 } 451 } 452 if (mi->owner_is_scalar_replaced()) { 453 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 454 st.print_cr(" failed reallocation for klass %s", k->external_name()); 455 } else { 456 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 457 } 458 } 459 } 460 tty->print_raw(st.freeze()); 461 } 462 #endif // !PRODUCT 463 } 464 } 465 } 466 467 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 468 // The given vframes cover one physical frame. 469 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 470 bool& realloc_failures) { 471 frame deoptee = chunk->at(0)->fr(); 472 JavaThread* deoptee_thread = chunk->at(0)->thread(); 473 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 474 RegisterMap map(chunk->at(0)->register_map()); 475 bool deoptimized_objects = false; 476 477 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false); 478 479 // Reallocate the non-escaping objects and restore their fields. 480 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 481 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 482 realloc_failures = rematerialize_objects(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects); 483 } 484 485 // MonitorInfo structures used in eliminate_locks are not GC safe. 486 NoSafepointVerifier no_safepoint; 487 488 // Now relock objects if synchronization on them was eliminated. 489 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 490 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 491 } 492 return deoptimized_objects; 493 } 494 #endif // COMPILER2_OR_JVMCI 495 496 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 497 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 498 // When we get here we are about to unwind the deoptee frame. In order to 499 // catch not yet safe to use frames, the following stack watermark barrier 500 // poll will make such frames safe to use. 501 StackWatermarkSet::before_unwind(current); 502 503 // Note: there is a safepoint safety issue here. No matter whether we enter 504 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 505 // the vframeArray is created. 506 // 507 508 // Allocate our special deoptimization ResourceMark 509 DeoptResourceMark* dmark = new DeoptResourceMark(current); 510 assert(current->deopt_mark() == nullptr, "Pending deopt!"); 511 current->set_deopt_mark(dmark); 512 513 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 514 RegisterMap map(current, 515 RegisterMap::UpdateMap::include, 516 RegisterMap::ProcessFrames::include, 517 RegisterMap::WalkContinuation::skip); 518 RegisterMap dummy_map(current, 519 RegisterMap::UpdateMap::skip, 520 RegisterMap::ProcessFrames::include, 521 RegisterMap::WalkContinuation::skip); 522 // Now get the deoptee with a valid map 523 frame deoptee = stub_frame.sender(&map); 524 // Set the deoptee nmethod 525 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!"); 526 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 527 current->set_deopt_compiled_method(cm); 528 529 if (VerifyStack) { 530 current->validate_frame_layout(); 531 } 532 533 // Create a growable array of VFrames where each VFrame represents an inlined 534 // Java frame. This storage is allocated with the usual system arena. 535 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 536 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 537 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 538 while (!vf->is_top()) { 539 assert(vf->is_compiled_frame(), "Wrong frame type"); 540 chunk->push(compiledVFrame::cast(vf)); 541 vf = vf->sender(); 542 } 543 assert(vf->is_compiled_frame(), "Wrong frame type"); 544 chunk->push(compiledVFrame::cast(vf)); 545 546 bool realloc_failures = false; 547 548 #if COMPILER2_OR_JVMCI 549 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 550 551 // Reallocate the non-escaping objects and restore their fields. Then 552 // relock objects if synchronization on them was eliminated. 553 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 554 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 555 bool unused; 556 realloc_failures = rematerialize_objects(current, exec_mode, cm, deoptee, map, chunk, unused); 557 } 558 #endif // COMPILER2_OR_JVMCI 559 560 // Ensure that no safepoint is taken after pointers have been stored 561 // in fields of rematerialized objects. If a safepoint occurs from here on 562 // out the java state residing in the vframeArray will be missed. 563 // Locks may be rebaised in a safepoint. 564 NoSafepointVerifier no_safepoint; 565 566 #if COMPILER2_OR_JVMCI 567 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 568 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 569 bool unused; 570 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 571 } 572 #endif // COMPILER2_OR_JVMCI 573 574 ScopeDesc* trap_scope = chunk->at(0)->scope(); 575 Handle exceptionObject; 576 if (trap_scope->rethrow_exception()) { 577 #ifndef PRODUCT 578 if (PrintDeoptimizationDetails) { 579 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()); 580 } 581 #endif // !PRODUCT 582 583 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 584 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw"); 585 ScopeValue* topOfStack = expressions->top(); 586 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 587 guarantee(exceptionObject() != nullptr, "exception oop can not be null"); 588 } 589 590 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 591 #if COMPILER2_OR_JVMCI 592 if (realloc_failures) { 593 // This destroys all ScopedValue bindings. 594 current->clear_scopedValueBindings(); 595 pop_frames_failed_reallocs(current, array); 596 } 597 #endif 598 599 assert(current->vframe_array_head() == nullptr, "Pending deopt!"); 600 current->set_vframe_array_head(array); 601 602 // Now that the vframeArray has been created if we have any deferred local writes 603 // added by jvmti then we can free up that structure as the data is now in the 604 // vframeArray 605 606 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 607 608 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 609 CodeBlob* cb = stub_frame.cb(); 610 // Verify we have the right vframeArray 611 assert(cb->frame_size() >= 0, "Unexpected frame size"); 612 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 613 614 // If the deopt call site is a MethodHandle invoke call site we have 615 // to adjust the unpack_sp. 616 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 617 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc())) 618 unpack_sp = deoptee.unextended_sp(); 619 620 #ifdef ASSERT 621 assert(cb->is_deoptimization_stub() || 622 cb->is_uncommon_trap_stub() || 623 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 624 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 625 "unexpected code blob: %s", cb->name()); 626 #endif 627 628 // This is a guarantee instead of an assert because if vframe doesn't match 629 // we will unpack the wrong deoptimized frame and wind up in strange places 630 // where it will be very difficult to figure out what went wrong. Better 631 // to die an early death here than some very obscure death later when the 632 // trail is cold. 633 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 634 // in that it will fail to detect a problem when there is one. This needs 635 // more work in tiger timeframe. 636 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 637 638 int number_of_frames = array->frames(); 639 640 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 641 // virtual activation, which is the reverse of the elements in the vframes array. 642 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 643 // +1 because we always have an interpreter return address for the final slot. 644 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 645 int popframe_extra_args = 0; 646 // Create an interpreter return address for the stub to use as its return 647 // address so the skeletal frames are perfectly walkable 648 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 649 650 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 651 // activation be put back on the expression stack of the caller for reexecution 652 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 653 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 654 } 655 656 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 657 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 658 // than simply use array->sender.pc(). This requires us to walk the current set of frames 659 // 660 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 661 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 662 663 // It's possible that the number of parameters at the call site is 664 // different than number of arguments in the callee when method 665 // handles are used. If the caller is interpreted get the real 666 // value so that the proper amount of space can be added to it's 667 // frame. 668 bool caller_was_method_handle = false; 669 if (deopt_sender.is_interpreted_frame()) { 670 methodHandle method(current, deopt_sender.interpreter_frame_method()); 671 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 672 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 673 // Method handle invokes may involve fairly arbitrary chains of 674 // calls so it's impossible to know how much actual space the 675 // caller has for locals. 676 caller_was_method_handle = true; 677 } 678 } 679 680 // 681 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 682 // frame_sizes/frame_pcs[1] next oldest frame (int) 683 // frame_sizes/frame_pcs[n] youngest frame (int) 684 // 685 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 686 // owns the space for the return address to it's caller). Confusing ain't it. 687 // 688 // The vframe array can address vframes with indices running from 689 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 690 // When we create the skeletal frames we need the oldest frame to be in the zero slot 691 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 692 // so things look a little strange in this loop. 693 // 694 int callee_parameters = 0; 695 int callee_locals = 0; 696 for (int index = 0; index < array->frames(); index++ ) { 697 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 698 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 699 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 700 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 701 callee_locals, 702 index == 0, 703 popframe_extra_args); 704 // This pc doesn't have to be perfect just good enough to identify the frame 705 // as interpreted so the skeleton frame will be walkable 706 // The correct pc will be set when the skeleton frame is completely filled out 707 // The final pc we store in the loop is wrong and will be overwritten below 708 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 709 710 callee_parameters = array->element(index)->method()->size_of_parameters(); 711 callee_locals = array->element(index)->method()->max_locals(); 712 popframe_extra_args = 0; 713 } 714 715 // Compute whether the root vframe returns a float or double value. 716 BasicType return_type; 717 { 718 methodHandle method(current, array->element(0)->method()); 719 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 720 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 721 } 722 723 // Compute information for handling adapters and adjusting the frame size of the caller. 724 int caller_adjustment = 0; 725 726 // Compute the amount the oldest interpreter frame will have to adjust 727 // its caller's stack by. If the caller is a compiled frame then 728 // we pretend that the callee has no parameters so that the 729 // extension counts for the full amount of locals and not just 730 // locals-parms. This is because without a c2i adapter the parm 731 // area as created by the compiled frame will not be usable by 732 // the interpreter. (Depending on the calling convention there 733 // may not even be enough space). 734 735 // QQQ I'd rather see this pushed down into last_frame_adjust 736 // and have it take the sender (aka caller). 737 738 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 739 caller_adjustment = last_frame_adjust(0, callee_locals); 740 } else if (callee_locals > callee_parameters) { 741 // The caller frame may need extending to accommodate 742 // non-parameter locals of the first unpacked interpreted frame. 743 // Compute that adjustment. 744 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 745 } 746 747 // If the sender is deoptimized we must retrieve the address of the handler 748 // since the frame will "magically" show the original pc before the deopt 749 // and we'd undo the deopt. 750 751 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 752 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 753 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 754 } 755 756 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc"); 757 758 #if INCLUDE_JVMCI 759 if (exceptionObject() != nullptr) { 760 current->set_exception_oop(exceptionObject()); 761 exec_mode = Unpack_exception; 762 } 763 #endif 764 765 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 766 assert(current->has_pending_exception(), "should have thrown OOME"); 767 current->set_exception_oop(current->pending_exception()); 768 current->clear_pending_exception(); 769 exec_mode = Unpack_exception; 770 } 771 772 #if INCLUDE_JVMCI 773 if (current->frames_to_pop_failed_realloc() > 0) { 774 current->set_pending_monitorenter(false); 775 } 776 #endif 777 778 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 779 caller_adjustment * BytesPerWord, 780 caller_was_method_handle ? 0 : callee_parameters, 781 number_of_frames, 782 frame_sizes, 783 frame_pcs, 784 return_type, 785 exec_mode); 786 // On some platforms, we need a way to pass some platform dependent 787 // information to the unpacking code so the skeletal frames come out 788 // correct (initial fp value, unextended sp, ...) 789 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 790 791 if (array->frames() > 1) { 792 if (VerifyStack && TraceDeoptimization) { 793 tty->print_cr("Deoptimizing method containing inlining"); 794 } 795 } 796 797 array->set_unroll_block(info); 798 return info; 799 } 800 801 // Called to cleanup deoptimization data structures in normal case 802 // after unpacking to stack and when stack overflow error occurs 803 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 804 vframeArray *array) { 805 806 // Get array if coming from exception 807 if (array == nullptr) { 808 array = thread->vframe_array_head(); 809 } 810 thread->set_vframe_array_head(nullptr); 811 812 // Free the previous UnrollBlock 813 vframeArray* old_array = thread->vframe_array_last(); 814 thread->set_vframe_array_last(array); 815 816 if (old_array != nullptr) { 817 UnrollBlock* old_info = old_array->unroll_block(); 818 old_array->set_unroll_block(nullptr); 819 delete old_info; 820 delete old_array; 821 } 822 823 // Deallocate any resource creating in this routine and any ResourceObjs allocated 824 // inside the vframeArray (StackValueCollections) 825 826 delete thread->deopt_mark(); 827 thread->set_deopt_mark(nullptr); 828 thread->set_deopt_compiled_method(nullptr); 829 830 831 if (JvmtiExport::can_pop_frame()) { 832 // Regardless of whether we entered this routine with the pending 833 // popframe condition bit set, we should always clear it now 834 thread->clear_popframe_condition(); 835 } 836 837 // unpack_frames() is called at the end of the deoptimization handler 838 // and (in C2) at the end of the uncommon trap handler. Note this fact 839 // so that an asynchronous stack walker can work again. This counter is 840 // incremented at the beginning of fetch_unroll_info() and (in C2) at 841 // the beginning of uncommon_trap(). 842 thread->dec_in_deopt_handler(); 843 } 844 845 // Moved from cpu directories because none of the cpus has callee save values. 846 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 847 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 848 849 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 850 // the days we had adapter frames. When we deoptimize a situation where a 851 // compiled caller calls a compiled caller will have registers it expects 852 // to survive the call to the callee. If we deoptimize the callee the only 853 // way we can restore these registers is to have the oldest interpreter 854 // frame that we create restore these values. That is what this routine 855 // will accomplish. 856 857 // At the moment we have modified c2 to not have any callee save registers 858 // so this problem does not exist and this routine is just a place holder. 859 860 assert(f->is_interpreted_frame(), "must be interpreted"); 861 } 862 863 #ifndef PRODUCT 864 static bool falls_through(Bytecodes::Code bc) { 865 switch (bc) { 866 // List may be incomplete. Here we really only care about bytecodes where compiled code 867 // can deoptimize. 868 case Bytecodes::_goto: 869 case Bytecodes::_goto_w: 870 case Bytecodes::_athrow: 871 return false; 872 default: 873 return true; 874 } 875 } 876 #endif 877 878 // Return BasicType of value being returned 879 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 880 assert(thread == JavaThread::current(), "pre-condition"); 881 882 // We are already active in the special DeoptResourceMark any ResourceObj's we 883 // allocate will be freed at the end of the routine. 884 885 // JRT_LEAF methods don't normally allocate handles and there is a 886 // NoHandleMark to enforce that. It is actually safe to use Handles 887 // in a JRT_LEAF method, and sometimes desirable, but to do so we 888 // must use ResetNoHandleMark to bypass the NoHandleMark, and 889 // then use a HandleMark to ensure any Handles we do create are 890 // cleaned up in this scope. 891 ResetNoHandleMark rnhm; 892 HandleMark hm(thread); 893 894 frame stub_frame = thread->last_frame(); 895 896 Continuation::notify_deopt(thread, stub_frame.sp()); 897 898 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 899 // must point to the vframeArray for the unpack frame. 900 vframeArray* array = thread->vframe_array_head(); 901 UnrollBlock* info = array->unroll_block(); 902 903 // We set the last_Java frame. But the stack isn't really parsable here. So we 904 // clear it to make sure JFR understands not to try and walk stacks from events 905 // in here. 906 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 907 thread->frame_anchor()->set_last_Java_sp(nullptr); 908 909 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 910 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 911 912 thread->frame_anchor()->set_last_Java_sp(sp); 913 914 BasicType bt = info->return_type(); 915 916 // If we have an exception pending, claim that the return type is an oop 917 // so the deopt_blob does not overwrite the exception_oop. 918 919 if (exec_mode == Unpack_exception) 920 bt = T_OBJECT; 921 922 // Cleanup thread deopt data 923 cleanup_deopt_info(thread, array); 924 925 #ifndef PRODUCT 926 if (VerifyStack) { 927 ResourceMark res_mark; 928 // Clear pending exception to not break verification code (restored afterwards) 929 PreserveExceptionMark pm(thread); 930 931 thread->validate_frame_layout(); 932 933 // Verify that the just-unpacked frames match the interpreter's 934 // notions of expression stack and locals 935 vframeArray* cur_array = thread->vframe_array_last(); 936 RegisterMap rm(thread, 937 RegisterMap::UpdateMap::skip, 938 RegisterMap::ProcessFrames::include, 939 RegisterMap::WalkContinuation::skip); 940 rm.set_include_argument_oops(false); 941 bool is_top_frame = true; 942 int callee_size_of_parameters = 0; 943 int callee_max_locals = 0; 944 for (int i = 0; i < cur_array->frames(); i++) { 945 vframeArrayElement* el = cur_array->element(i); 946 frame* iframe = el->iframe(); 947 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 948 949 // Get the oop map for this bci 950 InterpreterOopMap mask; 951 int cur_invoke_parameter_size = 0; 952 bool try_next_mask = false; 953 int next_mask_expression_stack_size = -1; 954 int top_frame_expression_stack_adjustment = 0; 955 methodHandle mh(thread, iframe->interpreter_frame_method()); 956 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 957 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 958 int max_bci = mh->code_size(); 959 // Get to the next bytecode if possible 960 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 961 // Check to see if we can grab the number of outgoing arguments 962 // at an uncommon trap for an invoke (where the compiler 963 // generates debug info before the invoke has executed) 964 Bytecodes::Code cur_code = str.next(); 965 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere; 966 if (Bytecodes::is_invoke(cur_code)) { 967 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 968 cur_invoke_parameter_size = invoke.size_of_parameters(); 969 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 970 callee_size_of_parameters++; 971 } 972 } 973 if (str.bci() < max_bci) { 974 next_code = str.next(); 975 if (next_code >= 0) { 976 // The interpreter oop map generator reports results before 977 // the current bytecode has executed except in the case of 978 // calls. It seems to be hard to tell whether the compiler 979 // has emitted debug information matching the "state before" 980 // a given bytecode or the state after, so we try both 981 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) { 982 // Get expression stack size for the next bytecode 983 InterpreterOopMap next_mask; 984 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 985 next_mask_expression_stack_size = next_mask.expression_stack_size(); 986 if (Bytecodes::is_invoke(next_code)) { 987 Bytecode_invoke invoke(mh, str.bci()); 988 next_mask_expression_stack_size += invoke.size_of_parameters(); 989 } 990 // Need to subtract off the size of the result type of 991 // the bytecode because this is not described in the 992 // debug info but returned to the interpreter in the TOS 993 // caching register 994 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 995 if (bytecode_result_type != T_ILLEGAL) { 996 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 997 } 998 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 999 try_next_mask = true; 1000 } 1001 } 1002 } 1003 1004 // Verify stack depth and oops in frame 1005 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 1006 if (!( 1007 /* SPARC */ 1008 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 1009 /* x86 */ 1010 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 1011 (try_next_mask && 1012 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 1013 top_frame_expression_stack_adjustment))) || 1014 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 1015 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 1016 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 1017 )) { 1018 { 1019 // Print out some information that will help us debug the problem 1020 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1021 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 1022 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 1023 if (try_next_mask) { 1024 tty->print_cr(" Next code %s", Bytecodes::name(next_code)); 1025 } 1026 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1027 iframe->interpreter_frame_expression_stack_size()); 1028 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1029 tty->print_cr(" try_next_mask = %d", try_next_mask); 1030 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 1031 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1032 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 1033 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1034 tty->print_cr(" exec_mode = %d", exec_mode); 1035 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1036 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 1037 tty->print_cr(" Interpreted frames:"); 1038 for (int k = 0; k < cur_array->frames(); k++) { 1039 vframeArrayElement* el = cur_array->element(k); 1040 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1041 } 1042 cur_array->print_on_2(tty); 1043 } 1044 guarantee(false, "wrong number of expression stack elements during deopt"); 1045 } 1046 VerifyOopClosure verify; 1047 iframe->oops_interpreted_do(&verify, &rm, false); 1048 callee_size_of_parameters = mh->size_of_parameters(); 1049 callee_max_locals = mh->max_locals(); 1050 is_top_frame = false; 1051 } 1052 } 1053 #endif // !PRODUCT 1054 1055 return bt; 1056 JRT_END 1057 1058 class DeoptimizeMarkedClosure : public HandshakeClosure { 1059 public: 1060 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 1061 void do_thread(Thread* thread) { 1062 JavaThread* jt = JavaThread::cast(thread); 1063 jt->deoptimize_marked_methods(); 1064 } 1065 }; 1066 1067 void Deoptimization::deoptimize_all_marked() { 1068 ResourceMark rm; 1069 1070 // Make the dependent methods not entrant 1071 CodeCache::make_marked_nmethods_deoptimized(); 1072 1073 DeoptimizeMarkedClosure deopt; 1074 if (SafepointSynchronize::is_at_safepoint()) { 1075 Threads::java_threads_do(&deopt); 1076 } else { 1077 Handshake::execute(&deopt); 1078 } 1079 } 1080 1081 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1082 = Deoptimization::Action_reinterpret; 1083 1084 #if INCLUDE_JVMCI 1085 template<typename CacheType> 1086 class BoxCacheBase : public CHeapObj<mtCompiler> { 1087 protected: 1088 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 1089 ResourceMark rm(thread); 1090 char* klass_name_str = klass_name->as_C_string(); 1091 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle()); 1092 guarantee(ik != nullptr, "%s must be loaded", klass_name_str); 1093 if (!ik->is_in_error_state()) { 1094 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 1095 CacheType::compute_offsets(ik); 1096 } 1097 return ik; 1098 } 1099 }; 1100 1101 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 1102 PrimitiveType _low; 1103 PrimitiveType _high; 1104 jobject _cache; 1105 protected: 1106 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 1107 BoxCache(Thread* thread) { 1108 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 1109 if (ik->is_in_error_state()) { 1110 _low = 1; 1111 _high = 0; 1112 _cache = nullptr; 1113 } else { 1114 objArrayOop cache = CacheType::cache(ik); 1115 assert(cache->length() > 0, "Empty cache"); 1116 _low = BoxType::value(cache->obj_at(0)); 1117 _high = checked_cast<PrimitiveType>(_low + cache->length() - 1); 1118 _cache = JNIHandles::make_global(Handle(thread, cache)); 1119 } 1120 } 1121 ~BoxCache() { 1122 JNIHandles::destroy_global(_cache); 1123 } 1124 public: 1125 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1126 if (_singleton == nullptr) { 1127 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1128 if (!Atomic::replace_if_null(&_singleton, s)) { 1129 delete s; 1130 } 1131 } 1132 return _singleton; 1133 } 1134 oop lookup(PrimitiveType value) { 1135 if (_low <= value && value <= _high) { 1136 int offset = checked_cast<int>(value - _low); 1137 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1138 } 1139 return nullptr; 1140 } 1141 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) { 1142 if (_cache == nullptr) { 1143 cache_init_error = true; 1144 return nullptr; 1145 } 1146 // Have to cast to avoid little/big-endian problems. 1147 if (sizeof(PrimitiveType) > sizeof(jint)) { 1148 jlong value = (jlong)raw_value; 1149 return lookup(value); 1150 } 1151 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1152 return lookup(value); 1153 } 1154 }; 1155 1156 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1157 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1158 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1159 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1160 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1161 1162 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr; 1163 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr; 1164 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr; 1165 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr; 1166 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr; 1167 1168 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1169 jobject _true_cache; 1170 jobject _false_cache; 1171 protected: 1172 static BooleanBoxCache *_singleton; 1173 BooleanBoxCache(Thread *thread) { 1174 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1175 if (ik->is_in_error_state()) { 1176 _true_cache = nullptr; 1177 _false_cache = nullptr; 1178 } else { 1179 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1180 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1181 } 1182 } 1183 ~BooleanBoxCache() { 1184 JNIHandles::destroy_global(_true_cache); 1185 JNIHandles::destroy_global(_false_cache); 1186 } 1187 public: 1188 static BooleanBoxCache* singleton(Thread* thread) { 1189 if (_singleton == nullptr) { 1190 BooleanBoxCache* s = new BooleanBoxCache(thread); 1191 if (!Atomic::replace_if_null(&_singleton, s)) { 1192 delete s; 1193 } 1194 } 1195 return _singleton; 1196 } 1197 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) { 1198 if (_true_cache == nullptr) { 1199 cache_in_error = true; 1200 return nullptr; 1201 } 1202 // Have to cast to avoid little/big-endian problems. 1203 jboolean value = (jboolean)*((jint*)&raw_value); 1204 return lookup(value); 1205 } 1206 oop lookup(jboolean value) { 1207 if (value != 0) { 1208 return JNIHandles::resolve_non_null(_true_cache); 1209 } 1210 return JNIHandles::resolve_non_null(_false_cache); 1211 } 1212 }; 1213 1214 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr; 1215 1216 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) { 1217 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1218 BasicType box_type = vmClasses::box_klass_type(k); 1219 if (box_type != T_OBJECT) { 1220 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1221 switch(box_type) { 1222 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1223 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1224 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1225 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1226 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1227 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error); 1228 default:; 1229 } 1230 } 1231 return nullptr; 1232 } 1233 #endif // INCLUDE_JVMCI 1234 1235 #if COMPILER2_OR_JVMCI 1236 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1237 Handle pending_exception(THREAD, thread->pending_exception()); 1238 const char* exception_file = thread->exception_file(); 1239 int exception_line = thread->exception_line(); 1240 thread->clear_pending_exception(); 1241 1242 bool failures = false; 1243 1244 for (int i = 0; i < objects->length(); i++) { 1245 assert(objects->at(i)->is_object(), "invalid debug information"); 1246 ObjectValue* sv = (ObjectValue*) objects->at(i); 1247 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1248 1249 // Check if the object may be null and has an additional is_init input that needs 1250 // to be checked before using the field values. Skip re-allocation if it is null. 1251 if (sv->maybe_null()) { 1252 assert(k->is_inline_klass(), "must be an inline klass"); 1253 jint is_init = StackValue::create_stack_value(fr, reg_map, sv->is_init())->get_jint(); 1254 if (is_init == 0) { 1255 continue; 1256 } 1257 } 1258 1259 oop obj = nullptr; 1260 bool cache_init_error = false; 1261 if (k->is_instance_klass()) { 1262 #if INCLUDE_JVMCI 1263 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 1264 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1265 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1266 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD); 1267 if (obj != nullptr) { 1268 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1269 abv->set_cached(true); 1270 } else if (cache_init_error) { 1271 // Results in an OOME which is valid (as opposed to a class initialization error) 1272 // and is fine for the rare case a cache initialization failing. 1273 failures = true; 1274 } 1275 } 1276 #endif // INCLUDE_JVMCI 1277 1278 InstanceKlass* ik = InstanceKlass::cast(k); 1279 if (obj == nullptr && !cache_init_error) { 1280 #if COMPILER2_OR_JVMCI 1281 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1282 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1283 } else { 1284 obj = ik->allocate_instance(THREAD); 1285 } 1286 #else 1287 obj = ik->allocate_instance(THREAD); 1288 #endif // COMPILER2_OR_JVMCI 1289 } 1290 } else if (k->is_flatArray_klass()) { 1291 FlatArrayKlass* ak = FlatArrayKlass::cast(k); 1292 // Inline type array must be zeroed because not all memory is reassigned 1293 obj = ak->allocate(sv->field_size(), THREAD); 1294 } else if (k->is_typeArray_klass()) { 1295 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1296 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1297 int len = sv->field_size() / type2size[ak->element_type()]; 1298 obj = ak->allocate(len, THREAD); 1299 } else if (k->is_objArray_klass()) { 1300 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1301 obj = ak->allocate(sv->field_size(), THREAD); 1302 } 1303 1304 if (obj == nullptr) { 1305 failures = true; 1306 } 1307 1308 assert(sv->value().is_null(), "redundant reallocation"); 1309 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception"); 1310 CLEAR_PENDING_EXCEPTION; 1311 sv->set_value(obj); 1312 } 1313 1314 if (failures) { 1315 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1316 } else if (pending_exception.not_null()) { 1317 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1318 } 1319 1320 return failures; 1321 } 1322 1323 // We're deoptimizing at the return of a call, inline type fields are 1324 // in registers. When we go back to the interpreter, it will expect a 1325 // reference to an inline type instance. Allocate and initialize it from 1326 // the register values here. 1327 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 1328 oop new_vt = vk->realloc_result(map, return_oops, THREAD); 1329 if (new_vt == nullptr) { 1330 CLEAR_PENDING_EXCEPTION; 1331 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 1332 } 1333 return_oops.clear(); 1334 return_oops.push(Handle(THREAD, new_vt)); 1335 return false; 1336 } 1337 1338 #if INCLUDE_JVMCI 1339 /** 1340 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1341 * we need to somehow be able to recover the actual kind to be able to write the correct 1342 * amount of bytes. 1343 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1344 * the entries at index n + 1 to n + i are 'markers'. 1345 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1346 * expected form of the array would be: 1347 * 1348 * {b0, b1, b2, b3, INT, marker, b6, b7} 1349 * 1350 * Thus, in order to get back the size of the entry, we simply need to count the number 1351 * of marked entries 1352 * 1353 * @param virtualArray the virtualized byte array 1354 * @param i index of the virtual entry we are recovering 1355 * @return The number of bytes the entry spans 1356 */ 1357 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1358 int index = i; 1359 while (++index < virtualArray->field_size() && 1360 virtualArray->field_at(index)->is_marker()) {} 1361 return index - i; 1362 } 1363 1364 /** 1365 * If there was a guarantee for byte array to always start aligned to a long, we could 1366 * do a simple check on the parity of the index. Unfortunately, that is not always the 1367 * case. Thus, we check alignment of the actual address we are writing to. 1368 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1369 * write a long to index 3. 1370 */ 1371 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1372 jbyte* res = obj->byte_at_addr(index); 1373 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1374 return res; 1375 } 1376 1377 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) { 1378 switch (byte_count) { 1379 case 1: 1380 obj->byte_at_put(index, (jbyte) value->get_jint()); 1381 break; 1382 case 2: 1383 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint(); 1384 break; 1385 case 4: 1386 *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint(); 1387 break; 1388 case 8: 1389 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr(); 1390 break; 1391 default: 1392 ShouldNotReachHere(); 1393 } 1394 } 1395 #endif // INCLUDE_JVMCI 1396 1397 1398 // restore elements of an eliminated type array 1399 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1400 int index = 0; 1401 1402 for (int i = 0; i < sv->field_size(); i++) { 1403 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1404 switch(type) { 1405 case T_LONG: case T_DOUBLE: { 1406 assert(value->type() == T_INT, "Agreement."); 1407 StackValue* low = 1408 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1409 #ifdef _LP64 1410 jlong res = (jlong)low->get_intptr(); 1411 #else 1412 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1413 #endif 1414 obj->long_at_put(index, res); 1415 break; 1416 } 1417 1418 case T_INT: case T_FLOAT: { // 4 bytes. 1419 assert(value->type() == T_INT, "Agreement."); 1420 bool big_value = false; 1421 if (i + 1 < sv->field_size() && type == T_INT) { 1422 if (sv->field_at(i)->is_location()) { 1423 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1424 if (type == Location::dbl || type == Location::lng) { 1425 big_value = true; 1426 } 1427 } else if (sv->field_at(i)->is_constant_int()) { 1428 ScopeValue* next_scope_field = sv->field_at(i + 1); 1429 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1430 big_value = true; 1431 } 1432 } 1433 } 1434 1435 if (big_value) { 1436 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1437 #ifdef _LP64 1438 jlong res = (jlong)low->get_intptr(); 1439 #else 1440 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1441 #endif 1442 obj->int_at_put(index, *(jint*)&res); 1443 obj->int_at_put(++index, *((jint*)&res + 1)); 1444 } else { 1445 obj->int_at_put(index, value->get_jint()); 1446 } 1447 break; 1448 } 1449 1450 case T_SHORT: 1451 assert(value->type() == T_INT, "Agreement."); 1452 obj->short_at_put(index, (jshort)value->get_jint()); 1453 break; 1454 1455 case T_CHAR: 1456 assert(value->type() == T_INT, "Agreement."); 1457 obj->char_at_put(index, (jchar)value->get_jint()); 1458 break; 1459 1460 case T_BYTE: { 1461 assert(value->type() == T_INT, "Agreement."); 1462 #if INCLUDE_JVMCI 1463 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1464 int byte_count = count_number_of_bytes_for_entry(sv, i); 1465 byte_array_put(obj, value, index, byte_count); 1466 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1467 i += byte_count - 1; // Balance the loop counter. 1468 index += byte_count; 1469 // index has been updated so continue at top of loop 1470 continue; 1471 #else 1472 obj->byte_at_put(index, (jbyte)value->get_jint()); 1473 break; 1474 #endif // INCLUDE_JVMCI 1475 } 1476 1477 case T_BOOLEAN: { 1478 assert(value->type() == T_INT, "Agreement."); 1479 obj->bool_at_put(index, (jboolean)value->get_jint()); 1480 break; 1481 } 1482 1483 default: 1484 ShouldNotReachHere(); 1485 } 1486 index++; 1487 } 1488 } 1489 1490 // restore fields of an eliminated object array 1491 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1492 for (int i = 0; i < sv->field_size(); i++) { 1493 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1494 assert(value->type() == T_OBJECT, "object element expected"); 1495 obj->obj_at_put(i, value->get_obj()()); 1496 } 1497 } 1498 1499 class ReassignedField { 1500 public: 1501 int _offset; 1502 BasicType _type; 1503 InstanceKlass* _klass; 1504 bool _is_flat; 1505 public: 1506 ReassignedField() : _offset(0), _type(T_ILLEGAL), _klass(nullptr), _is_flat(false) { } 1507 }; 1508 1509 static int compare(ReassignedField* left, ReassignedField* right) { 1510 return left->_offset - right->_offset; 1511 } 1512 1513 // Restore fields of an eliminated instance object using the same field order 1514 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1515 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) { 1516 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1517 InstanceKlass* ik = klass; 1518 while (ik != nullptr) { 1519 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1520 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) { 1521 ReassignedField field; 1522 field._offset = fs.offset(); 1523 field._type = Signature::basic_type(fs.signature()); 1524 if (fs.is_null_free_inline_type()) { 1525 if (fs.is_flat()) { 1526 field._is_flat = true; 1527 // Resolve klass of flat inline type field 1528 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index())); 1529 } else { 1530 field._type = T_OBJECT; // Can be removed once Q-descriptors have been removed. 1531 } 1532 } 1533 fields->append(field); 1534 } 1535 } 1536 ik = ik->superklass(); 1537 } 1538 fields->sort(compare); 1539 for (int i = 0; i < fields->length(); i++) { 1540 BasicType type = fields->at(i)._type; 1541 int offset = base_offset + fields->at(i)._offset; 1542 // Check for flat inline type field before accessing the ScopeValue because it might not have any fields 1543 if (fields->at(i)._is_flat) { 1544 // Recursively re-assign flat inline type fields 1545 InstanceKlass* vk = fields->at(i)._klass; 1546 assert(vk != nullptr, "must be resolved"); 1547 offset -= InlineKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1548 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1549 continue; // Continue because we don't need to increment svIndex 1550 } 1551 ScopeValue* scope_field = sv->field_at(svIndex); 1552 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1553 switch (type) { 1554 case T_OBJECT: 1555 case T_ARRAY: 1556 assert(value->type() == T_OBJECT, "Agreement."); 1557 obj->obj_field_put(offset, value->get_obj()()); 1558 break; 1559 1560 case T_INT: case T_FLOAT: { // 4 bytes. 1561 assert(value->type() == T_INT, "Agreement."); 1562 bool big_value = false; 1563 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1564 if (scope_field->is_location()) { 1565 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1566 if (type == Location::dbl || type == Location::lng) { 1567 big_value = true; 1568 } 1569 } 1570 if (scope_field->is_constant_int()) { 1571 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1572 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1573 big_value = true; 1574 } 1575 } 1576 } 1577 1578 if (big_value) { 1579 i++; 1580 assert(i < fields->length(), "second T_INT field needed"); 1581 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1582 } else { 1583 obj->int_field_put(offset, value->get_jint()); 1584 break; 1585 } 1586 } 1587 /* no break */ 1588 1589 case T_LONG: case T_DOUBLE: { 1590 assert(value->type() == T_INT, "Agreement."); 1591 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1592 #ifdef _LP64 1593 jlong res = (jlong)low->get_intptr(); 1594 #else 1595 jlong res = jlong_from(value->get_jint(), low->get_jint()); 1596 #endif 1597 obj->long_field_put(offset, res); 1598 break; 1599 } 1600 1601 case T_SHORT: 1602 assert(value->type() == T_INT, "Agreement."); 1603 obj->short_field_put(offset, (jshort)value->get_jint()); 1604 break; 1605 1606 case T_CHAR: 1607 assert(value->type() == T_INT, "Agreement."); 1608 obj->char_field_put(offset, (jchar)value->get_jint()); 1609 break; 1610 1611 case T_BYTE: 1612 assert(value->type() == T_INT, "Agreement."); 1613 obj->byte_field_put(offset, (jbyte)value->get_jint()); 1614 break; 1615 1616 case T_BOOLEAN: 1617 assert(value->type() == T_INT, "Agreement."); 1618 obj->bool_field_put(offset, (jboolean)value->get_jint()); 1619 break; 1620 1621 default: 1622 ShouldNotReachHere(); 1623 } 1624 svIndex++; 1625 } 1626 return svIndex; 1627 } 1628 1629 // restore fields of an eliminated inline type array 1630 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool skip_internal, TRAPS) { 1631 InlineKlass* vk = vak->element_klass(); 1632 assert(vk->flat_array(), "should only be used for flat inline type arrays"); 1633 // Adjust offset to omit oop header 1634 int base_offset = arrayOopDesc::base_offset_in_bytes(T_PRIMITIVE_OBJECT) - InlineKlass::cast(vk)->first_field_offset(); 1635 // Initialize all elements of the flat inline type array 1636 for (int i = 0; i < sv->field_size(); i++) { 1637 ScopeValue* val = sv->field_at(i); 1638 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1639 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, skip_internal, offset, CHECK); 1640 } 1641 } 1642 1643 // restore fields of all eliminated objects and arrays 1644 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1645 for (int i = 0; i < objects->length(); i++) { 1646 assert(objects->at(i)->is_object(), "invalid debug information"); 1647 ObjectValue* sv = (ObjectValue*) objects->at(i); 1648 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1649 Handle obj = sv->value(); 1650 assert(obj.not_null() || realloc_failures || sv->maybe_null(), "reallocation was missed"); 1651 #ifndef PRODUCT 1652 if (PrintDeoptimizationDetails) { 1653 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1654 } 1655 #endif // !PRODUCT 1656 1657 if (obj.is_null()) { 1658 continue; 1659 } 1660 1661 #if INCLUDE_JVMCI 1662 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1663 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1664 continue; 1665 } 1666 #endif // INCLUDE_JVMCI 1667 #if COMPILER2_OR_JVMCI 1668 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1669 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1670 ScopeValue* payload = sv->field_at(0); 1671 if (payload->is_location() && 1672 payload->as_LocationValue()->location().type() == Location::vector) { 1673 #ifndef PRODUCT 1674 if (PrintDeoptimizationDetails) { 1675 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1676 if (Verbose) { 1677 Handle obj = sv->value(); 1678 k->oop_print_on(obj(), tty); 1679 } 1680 } 1681 #endif // !PRODUCT 1682 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1683 } 1684 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1685 // which could be restored after vector object allocation. 1686 } 1687 #endif /* !COMPILER2_OR_JVMCI */ 1688 if (k->is_instance_klass()) { 1689 InstanceKlass* ik = InstanceKlass::cast(k); 1690 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1691 } else if (k->is_flatArray_klass()) { 1692 FlatArrayKlass* vak = FlatArrayKlass::cast(k); 1693 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, skip_internal, CHECK); 1694 } else if (k->is_typeArray_klass()) { 1695 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1696 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1697 } else if (k->is_objArray_klass()) { 1698 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1699 } 1700 } 1701 // These objects may escape when we return to Interpreter after deoptimization. 1702 // We need barrier so that stores that initialize these objects can't be reordered 1703 // with subsequent stores that make these objects accessible by other threads. 1704 OrderAccess::storestore(); 1705 } 1706 1707 1708 // relock objects for which synchronization was eliminated 1709 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1710 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1711 bool relocked_objects = false; 1712 for (int i = 0; i < monitors->length(); i++) { 1713 MonitorInfo* mon_info = monitors->at(i); 1714 if (mon_info->eliminated()) { 1715 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1716 relocked_objects = true; 1717 if (!mon_info->owner_is_scalar_replaced()) { 1718 Handle obj(thread, mon_info->owner()); 1719 markWord mark = obj->mark(); 1720 if (exec_mode == Unpack_none) { 1721 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) { 1722 // With exec_mode == Unpack_none obj may be thread local and locked in 1723 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header. 1724 markWord dmw = mark.displaced_mark_helper(); 1725 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr)); 1726 obj->set_mark(dmw); 1727 } 1728 if (mark.has_monitor()) { 1729 // defer relocking if the deoptee thread is currently waiting for obj 1730 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1731 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) { 1732 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1733 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1734 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1735 continue; 1736 } 1737 } 1738 } 1739 if (LockingMode == LM_LIGHTWEIGHT && exec_mode == Unpack_none) { 1740 // We have lost information about the correct state of the lock stack. 1741 // Inflate the locks instead. Enter then inflate to avoid races with 1742 // deflation. 1743 ObjectSynchronizer::enter_for(obj, nullptr, deoptee_thread); 1744 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1745 ObjectMonitor* mon = ObjectSynchronizer::inflate_for(deoptee_thread, obj(), ObjectSynchronizer::inflate_cause_vm_internal); 1746 assert(mon->owner() == deoptee_thread, "must be"); 1747 } else { 1748 BasicLock* lock = mon_info->lock(); 1749 ObjectSynchronizer::enter_for(obj, lock, deoptee_thread); 1750 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1751 } 1752 } 1753 } 1754 } 1755 return relocked_objects; 1756 } 1757 #endif // COMPILER2_OR_JVMCI 1758 1759 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1760 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1761 1762 // Register map for next frame (used for stack crawl). We capture 1763 // the state of the deopt'ing frame's caller. Thus if we need to 1764 // stuff a C2I adapter we can properly fill in the callee-save 1765 // register locations. 1766 frame caller = fr.sender(reg_map); 1767 int frame_size = pointer_delta_as_int(caller.sp(), fr.sp()); 1768 1769 frame sender = caller; 1770 1771 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1772 // the vframeArray containing the unpacking information is allocated in the C heap. 1773 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1774 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1775 1776 // Compare the vframeArray to the collected vframes 1777 assert(array->structural_compare(thread, chunk), "just checking"); 1778 1779 if (TraceDeoptimization) { 1780 ResourceMark rm; 1781 stringStream st; 1782 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1783 st.print(" "); 1784 fr.print_on(&st); 1785 st.print_cr(" Virtual frames (innermost/newest first):"); 1786 for (int index = 0; index < chunk->length(); index++) { 1787 compiledVFrame* vf = chunk->at(index); 1788 int bci = vf->raw_bci(); 1789 const char* code_name; 1790 if (bci == SynchronizationEntryBCI) { 1791 code_name = "sync entry"; 1792 } else { 1793 Bytecodes::Code code = vf->method()->code_at(bci); 1794 code_name = Bytecodes::name(code); 1795 } 1796 1797 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1798 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1799 st.print(" - %s", code_name); 1800 st.print_cr(" @ bci=%d ", bci); 1801 } 1802 tty->print_raw(st.freeze()); 1803 tty->cr(); 1804 } 1805 1806 return array; 1807 } 1808 1809 #if COMPILER2_OR_JVMCI 1810 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1811 // Reallocation of some scalar replaced objects failed. Record 1812 // that we need to pop all the interpreter frames for the 1813 // deoptimized compiled frame. 1814 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1815 thread->set_frames_to_pop_failed_realloc(array->frames()); 1816 // Unlock all monitors here otherwise the interpreter will see a 1817 // mix of locked and unlocked monitors (because of failed 1818 // reallocations of synchronized objects) and be confused. 1819 for (int i = 0; i < array->frames(); i++) { 1820 MonitorChunk* monitors = array->element(i)->monitors(); 1821 if (monitors != nullptr) { 1822 // Unlock in reverse order starting from most nested monitor. 1823 for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) { 1824 BasicObjectLock* src = monitors->at(j); 1825 if (src->obj() != nullptr) { 1826 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1827 } 1828 } 1829 array->element(i)->free_monitors(thread); 1830 #ifdef ASSERT 1831 array->element(i)->set_removed_monitors(); 1832 #endif 1833 } 1834 } 1835 } 1836 #endif 1837 1838 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1839 assert(fr.can_be_deoptimized(), "checking frame type"); 1840 1841 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1842 1843 if (LogCompilation && xtty != nullptr) { 1844 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1845 assert(cm != nullptr, "only compiled methods can deopt"); 1846 1847 ttyLocker ttyl; 1848 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1849 cm->log_identity(xtty); 1850 xtty->end_head(); 1851 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1852 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1853 xtty->method(sd->method()); 1854 xtty->end_elem(); 1855 if (sd->is_top()) break; 1856 } 1857 xtty->tail("deoptimized"); 1858 } 1859 1860 Continuation::notify_deopt(thread, fr.sp()); 1861 1862 // Patch the compiled method so that when execution returns to it we will 1863 // deopt the execution state and return to the interpreter. 1864 fr.deoptimize(thread); 1865 } 1866 1867 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1868 // Deoptimize only if the frame comes from compiled code. 1869 // Do not deoptimize the frame which is already patched 1870 // during the execution of the loops below. 1871 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1872 return; 1873 } 1874 ResourceMark rm; 1875 deoptimize_single_frame(thread, fr, reason); 1876 } 1877 1878 #if INCLUDE_JVMCI 1879 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1880 // there is no exception handler for this pc => deoptimize 1881 cm->make_not_entrant(); 1882 1883 // Use Deoptimization::deoptimize for all of its side-effects: 1884 // gathering traps statistics, logging... 1885 // it also patches the return pc but we do not care about that 1886 // since we return a continuation to the deopt_blob below. 1887 JavaThread* thread = JavaThread::current(); 1888 RegisterMap reg_map(thread, 1889 RegisterMap::UpdateMap::skip, 1890 RegisterMap::ProcessFrames::include, 1891 RegisterMap::WalkContinuation::skip); 1892 frame runtime_frame = thread->last_frame(); 1893 frame caller_frame = runtime_frame.sender(®_map); 1894 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1895 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1896 compiledVFrame* cvf = compiledVFrame::cast(vf); 1897 ScopeDesc* imm_scope = cvf->scope(); 1898 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1899 if (imm_mdo != nullptr) { 1900 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 1901 MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 1902 1903 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr); 1904 if (pdata != nullptr && pdata->is_BitData()) { 1905 BitData* bit_data = (BitData*) pdata; 1906 bit_data->set_exception_seen(); 1907 } 1908 } 1909 1910 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1911 1912 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1913 if (trap_mdo != nullptr) { 1914 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1915 } 1916 1917 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1918 } 1919 #endif 1920 1921 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1922 assert(thread == Thread::current() || 1923 thread->is_handshake_safe_for(Thread::current()) || 1924 SafepointSynchronize::is_at_safepoint(), 1925 "can only deoptimize other thread at a safepoint/handshake"); 1926 // Compute frame and register map based on thread and sp. 1927 RegisterMap reg_map(thread, 1928 RegisterMap::UpdateMap::skip, 1929 RegisterMap::ProcessFrames::include, 1930 RegisterMap::WalkContinuation::skip); 1931 frame fr = thread->last_frame(); 1932 while (fr.id() != id) { 1933 fr = fr.sender(®_map); 1934 } 1935 deoptimize(thread, fr, reason); 1936 } 1937 1938 1939 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1940 Thread* current = Thread::current(); 1941 if (thread == current || thread->is_handshake_safe_for(current)) { 1942 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1943 } else { 1944 VM_DeoptimizeFrame deopt(thread, id, reason); 1945 VMThread::execute(&deopt); 1946 } 1947 } 1948 1949 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1950 deoptimize_frame(thread, id, Reason_constraint); 1951 } 1952 1953 // JVMTI PopFrame support 1954 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1955 { 1956 assert(thread == JavaThread::current(), "pre-condition"); 1957 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1958 } 1959 JRT_END 1960 1961 MethodData* 1962 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1963 bool create_if_missing) { 1964 JavaThread* THREAD = thread; // For exception macros. 1965 MethodData* mdo = m()->method_data(); 1966 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) { 1967 // Build an MDO. Ignore errors like OutOfMemory; 1968 // that simply means we won't have an MDO to update. 1969 Method::build_profiling_method_data(m, THREAD); 1970 if (HAS_PENDING_EXCEPTION) { 1971 // Only metaspace OOM is expected. No Java code executed. 1972 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1973 CLEAR_PENDING_EXCEPTION; 1974 } 1975 mdo = m()->method_data(); 1976 } 1977 return mdo; 1978 } 1979 1980 #if COMPILER2_OR_JVMCI 1981 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1982 // In case of an unresolved klass entry, load the class. 1983 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1984 // and probably nowhere else. 1985 // Even that case would benefit from simply re-interpreting the 1986 // bytecode, without paying special attention to the class index. 1987 // So this whole "class index" feature should probably be removed. 1988 1989 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1990 Klass* tk = constant_pool->klass_at(index, THREAD); 1991 if (HAS_PENDING_EXCEPTION) { 1992 // Exception happened during classloading. We ignore the exception here, since it 1993 // is going to be rethrown since the current activation is going to be deoptimized and 1994 // the interpreter will re-execute the bytecode. 1995 // Do not clear probable Async Exceptions. 1996 CLEAR_PENDING_NONASYNC_EXCEPTION; 1997 // Class loading called java code which may have caused a stack 1998 // overflow. If the exception was thrown right before the return 1999 // to the runtime the stack is no longer guarded. Reguard the 2000 // stack otherwise if we return to the uncommon trap blob and the 2001 // stack bang causes a stack overflow we crash. 2002 JavaThread* jt = THREAD; 2003 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 2004 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 2005 } 2006 return; 2007 } 2008 2009 assert(!constant_pool->tag_at(index).is_symbol(), 2010 "no symbolic names here, please"); 2011 } 2012 2013 #if INCLUDE_JFR 2014 2015 class DeoptReasonSerializer : public JfrSerializer { 2016 public: 2017 void serialize(JfrCheckpointWriter& writer) { 2018 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 2019 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 2020 writer.write_key((u8)i); 2021 writer.write(Deoptimization::trap_reason_name(i)); 2022 } 2023 } 2024 }; 2025 2026 class DeoptActionSerializer : public JfrSerializer { 2027 public: 2028 void serialize(JfrCheckpointWriter& writer) { 2029 static const u4 nof_actions = Deoptimization::Action_LIMIT; 2030 writer.write_count(nof_actions); 2031 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 2032 writer.write_key(i); 2033 writer.write(Deoptimization::trap_action_name((int)i)); 2034 } 2035 } 2036 }; 2037 2038 static void register_serializers() { 2039 static int critical_section = 0; 2040 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 2041 return; 2042 } 2043 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 2044 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 2045 } 2046 2047 static void post_deoptimization_event(CompiledMethod* nm, 2048 const Method* method, 2049 int trap_bci, 2050 int instruction, 2051 Deoptimization::DeoptReason reason, 2052 Deoptimization::DeoptAction action) { 2053 assert(nm != nullptr, "invariant"); 2054 assert(method != nullptr, "invariant"); 2055 if (EventDeoptimization::is_enabled()) { 2056 static bool serializers_registered = false; 2057 if (!serializers_registered) { 2058 register_serializers(); 2059 serializers_registered = true; 2060 } 2061 EventDeoptimization event; 2062 event.set_compileId(nm->compile_id()); 2063 event.set_compiler(nm->compiler_type()); 2064 event.set_method(method); 2065 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 2066 event.set_bci(trap_bci); 2067 event.set_instruction(instruction); 2068 event.set_reason(reason); 2069 event.set_action(action); 2070 event.commit(); 2071 } 2072 } 2073 2074 #endif // INCLUDE_JFR 2075 2076 static void log_deopt(CompiledMethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 2077 const char* reason_name, const char* reason_action) { 2078 LogTarget(Debug, deoptimization) lt; 2079 if (lt.is_enabled()) { 2080 LogStream ls(lt); 2081 bool is_osr = nm->is_osr_method(); 2082 ls.print("cid=%4d %s level=%d", 2083 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level()); 2084 ls.print(" %s", tm->name_and_sig_as_C_string()); 2085 ls.print(" trap_bci=%d ", trap_bci); 2086 if (is_osr) { 2087 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 2088 } 2089 ls.print("%s ", reason_name); 2090 ls.print("%s ", reason_action); 2091 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 2092 pc, fr.pc() - nm->code_begin()); 2093 } 2094 } 2095 2096 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 2097 HandleMark hm(current); 2098 2099 // uncommon_trap() is called at the beginning of the uncommon trap 2100 // handler. Note this fact before we start generating temporary frames 2101 // that can confuse an asynchronous stack walker. This counter is 2102 // decremented at the end of unpack_frames(). 2103 2104 current->inc_in_deopt_handler(); 2105 2106 #if INCLUDE_JVMCI 2107 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 2108 RegisterMap reg_map(current, 2109 RegisterMap::UpdateMap::include, 2110 RegisterMap::ProcessFrames::include, 2111 RegisterMap::WalkContinuation::skip); 2112 #else 2113 RegisterMap reg_map(current, 2114 RegisterMap::UpdateMap::skip, 2115 RegisterMap::ProcessFrames::include, 2116 RegisterMap::WalkContinuation::skip); 2117 #endif 2118 frame stub_frame = current->last_frame(); 2119 frame fr = stub_frame.sender(®_map); 2120 2121 // Log a message 2122 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 2123 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 2124 2125 { 2126 ResourceMark rm; 2127 2128 DeoptReason reason = trap_request_reason(trap_request); 2129 DeoptAction action = trap_request_action(trap_request); 2130 #if INCLUDE_JVMCI 2131 int debug_id = trap_request_debug_id(trap_request); 2132 #endif 2133 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 2134 2135 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 2136 compiledVFrame* cvf = compiledVFrame::cast(vf); 2137 2138 CompiledMethod* nm = cvf->code(); 2139 2140 ScopeDesc* trap_scope = cvf->scope(); 2141 2142 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2143 2144 if (is_receiver_constraint_failure) { 2145 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2146 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2147 JVMCI_ONLY(COMMA debug_id)); 2148 } 2149 2150 methodHandle trap_method(current, trap_scope->method()); 2151 int trap_bci = trap_scope->bci(); 2152 #if INCLUDE_JVMCI 2153 jlong speculation = current->pending_failed_speculation(); 2154 if (nm->is_compiled_by_jvmci()) { 2155 nm->as_nmethod()->update_speculation(current); 2156 } else { 2157 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2158 } 2159 2160 if (trap_bci == SynchronizationEntryBCI) { 2161 trap_bci = 0; 2162 current->set_pending_monitorenter(true); 2163 } 2164 2165 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2166 current->set_pending_transfer_to_interpreter(true); 2167 } 2168 #endif 2169 2170 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2171 // Record this event in the histogram. 2172 gather_statistics(reason, action, trap_bc); 2173 2174 // Ensure that we can record deopt. history: 2175 // Need MDO to record RTM code generation state. 2176 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 2177 2178 methodHandle profiled_method; 2179 #if INCLUDE_JVMCI 2180 if (nm->is_compiled_by_jvmci()) { 2181 profiled_method = methodHandle(current, nm->method()); 2182 } else { 2183 profiled_method = trap_method; 2184 } 2185 #else 2186 profiled_method = trap_method; 2187 #endif 2188 2189 MethodData* trap_mdo = 2190 get_method_data(current, profiled_method, create_if_missing); 2191 2192 { // Log Deoptimization event for JFR, UL and event system 2193 Method* tm = trap_method(); 2194 const char* reason_name = trap_reason_name(reason); 2195 const char* reason_action = trap_action_name(action); 2196 intptr_t pc = p2i(fr.pc()); 2197 2198 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2199 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2200 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 2201 reason_name, reason_action, pc, 2202 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 2203 } 2204 2205 // Print a bunch of diagnostics, if requested. 2206 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2207 ResourceMark rm; 2208 2209 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2210 // We must do this already now, since we cannot acquire this lock while 2211 // holding the tty lock (lock ordering by rank). 2212 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2213 2214 ttyLocker ttyl; 2215 2216 char buf[100]; 2217 if (xtty != nullptr) { 2218 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 2219 os::current_thread_id(), 2220 format_trap_request(buf, sizeof(buf), trap_request)); 2221 #if INCLUDE_JVMCI 2222 if (speculation != 0) { 2223 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2224 } 2225 #endif 2226 nm->log_identity(xtty); 2227 } 2228 Symbol* class_name = nullptr; 2229 bool unresolved = false; 2230 if (unloaded_class_index >= 0) { 2231 constantPoolHandle constants (current, trap_method->constants()); 2232 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2233 class_name = constants->klass_name_at(unloaded_class_index); 2234 unresolved = true; 2235 if (xtty != nullptr) 2236 xtty->print(" unresolved='1'"); 2237 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2238 class_name = constants->symbol_at(unloaded_class_index); 2239 } 2240 if (xtty != nullptr) 2241 xtty->name(class_name); 2242 } 2243 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) { 2244 // Dump the relevant MDO state. 2245 // This is the deopt count for the current reason, any previous 2246 // reasons or recompiles seen at this point. 2247 int dcnt = trap_mdo->trap_count(reason); 2248 if (dcnt != 0) 2249 xtty->print(" count='%d'", dcnt); 2250 2251 // We need to lock to read the ProfileData. But to keep the locks ordered, we need to 2252 // lock extra_data_lock before the tty lock. 2253 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2254 int dos = (pdata == nullptr)? 0: pdata->trap_state(); 2255 if (dos != 0) { 2256 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2257 if (trap_state_is_recompiled(dos)) { 2258 int recnt2 = trap_mdo->overflow_recompile_count(); 2259 if (recnt2 != 0) 2260 xtty->print(" recompiles2='%d'", recnt2); 2261 } 2262 } 2263 } 2264 if (xtty != nullptr) { 2265 xtty->stamp(); 2266 xtty->end_head(); 2267 } 2268 if (TraceDeoptimization) { // make noise on the tty 2269 stringStream st; 2270 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2271 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2272 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2273 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 2274 #if INCLUDE_JVMCI 2275 if (nm->is_nmethod()) { 2276 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 2277 if (installed_code_name != nullptr) { 2278 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2279 } 2280 } 2281 #endif 2282 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2283 p2i(fr.pc()), 2284 os::current_thread_id(), 2285 trap_reason_name(reason), 2286 trap_action_name(action), 2287 unloaded_class_index 2288 #if INCLUDE_JVMCI 2289 , debug_id 2290 #endif 2291 ); 2292 if (class_name != nullptr) { 2293 st.print(unresolved ? " unresolved class: " : " symbol: "); 2294 class_name->print_symbol_on(&st); 2295 } 2296 st.cr(); 2297 tty->print_raw(st.freeze()); 2298 } 2299 if (xtty != nullptr) { 2300 // Log the precise location of the trap. 2301 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2302 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2303 xtty->method(sd->method()); 2304 xtty->end_elem(); 2305 if (sd->is_top()) break; 2306 } 2307 xtty->tail("uncommon_trap"); 2308 } 2309 } 2310 // (End diagnostic printout.) 2311 2312 if (is_receiver_constraint_failure) { 2313 fatal("missing receiver type check"); 2314 } 2315 2316 // Load class if necessary 2317 if (unloaded_class_index >= 0) { 2318 constantPoolHandle constants(current, trap_method->constants()); 2319 load_class_by_index(constants, unloaded_class_index, THREAD); 2320 } 2321 2322 // Flush the nmethod if necessary and desirable. 2323 // 2324 // We need to avoid situations where we are re-flushing the nmethod 2325 // because of a hot deoptimization site. Repeated flushes at the same 2326 // point need to be detected by the compiler and avoided. If the compiler 2327 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2328 // module must take measures to avoid an infinite cycle of recompilation 2329 // and deoptimization. There are several such measures: 2330 // 2331 // 1. If a recompilation is ordered a second time at some site X 2332 // and for the same reason R, the action is adjusted to 'reinterpret', 2333 // to give the interpreter time to exercise the method more thoroughly. 2334 // If this happens, the method's overflow_recompile_count is incremented. 2335 // 2336 // 2. If the compiler fails to reduce the deoptimization rate, then 2337 // the method's overflow_recompile_count will begin to exceed the set 2338 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2339 // is adjusted to 'make_not_compilable', and the method is abandoned 2340 // to the interpreter. This is a performance hit for hot methods, 2341 // but is better than a disastrous infinite cycle of recompilations. 2342 // (Actually, only the method containing the site X is abandoned.) 2343 // 2344 // 3. In parallel with the previous measures, if the total number of 2345 // recompilations of a method exceeds the much larger set limit 2346 // PerMethodRecompilationCutoff, the method is abandoned. 2347 // This should only happen if the method is very large and has 2348 // many "lukewarm" deoptimizations. The code which enforces this 2349 // limit is elsewhere (class nmethod, class Method). 2350 // 2351 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2352 // to recompile at each bytecode independently of the per-BCI cutoff. 2353 // 2354 // The decision to update code is up to the compiler, and is encoded 2355 // in the Action_xxx code. If the compiler requests Action_none 2356 // no trap state is changed, no compiled code is changed, and the 2357 // computation suffers along in the interpreter. 2358 // 2359 // The other action codes specify various tactics for decompilation 2360 // and recompilation. Action_maybe_recompile is the loosest, and 2361 // allows the compiled code to stay around until enough traps are seen, 2362 // and until the compiler gets around to recompiling the trapping method. 2363 // 2364 // The other actions cause immediate removal of the present code. 2365 2366 // Traps caused by injected profile shouldn't pollute trap counts. 2367 bool injected_profile_trap = trap_method->has_injected_profile() && 2368 (reason == Reason_intrinsic || reason == Reason_unreached); 2369 2370 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2371 bool make_not_entrant = false; 2372 bool make_not_compilable = false; 2373 bool reprofile = false; 2374 switch (action) { 2375 case Action_none: 2376 // Keep the old code. 2377 update_trap_state = false; 2378 break; 2379 case Action_maybe_recompile: 2380 // Do not need to invalidate the present code, but we can 2381 // initiate another 2382 // Start compiler without (necessarily) invalidating the nmethod. 2383 // The system will tolerate the old code, but new code should be 2384 // generated when possible. 2385 break; 2386 case Action_reinterpret: 2387 // Go back into the interpreter for a while, and then consider 2388 // recompiling form scratch. 2389 make_not_entrant = true; 2390 // Reset invocation counter for outer most method. 2391 // This will allow the interpreter to exercise the bytecodes 2392 // for a while before recompiling. 2393 // By contrast, Action_make_not_entrant is immediate. 2394 // 2395 // Note that the compiler will track null_check, null_assert, 2396 // range_check, and class_check events and log them as if they 2397 // had been traps taken from compiled code. This will update 2398 // the MDO trap history so that the next compilation will 2399 // properly detect hot trap sites. 2400 reprofile = true; 2401 break; 2402 case Action_make_not_entrant: 2403 // Request immediate recompilation, and get rid of the old code. 2404 // Make them not entrant, so next time they are called they get 2405 // recompiled. Unloaded classes are loaded now so recompile before next 2406 // time they are called. Same for uninitialized. The interpreter will 2407 // link the missing class, if any. 2408 make_not_entrant = true; 2409 break; 2410 case Action_make_not_compilable: 2411 // Give up on compiling this method at all. 2412 make_not_entrant = true; 2413 make_not_compilable = true; 2414 break; 2415 default: 2416 ShouldNotReachHere(); 2417 } 2418 2419 // Setting +ProfileTraps fixes the following, on all platforms: 2420 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2421 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2422 // recompile relies on a MethodData* to record heroic opt failures. 2423 2424 // Whether the interpreter is producing MDO data or not, we also need 2425 // to use the MDO to detect hot deoptimization points and control 2426 // aggressive optimization. 2427 bool inc_recompile_count = false; 2428 2429 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2430 ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr, 2431 (trap_mdo != nullptr), 2432 Mutex::_no_safepoint_check_flag); 2433 ProfileData* pdata = nullptr; 2434 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) { 2435 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2436 uint this_trap_count = 0; 2437 bool maybe_prior_trap = false; 2438 bool maybe_prior_recompile = false; 2439 2440 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2441 #if INCLUDE_JVMCI 2442 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2443 #endif 2444 nm->method(), 2445 //outputs: 2446 this_trap_count, 2447 maybe_prior_trap, 2448 maybe_prior_recompile); 2449 // Because the interpreter also counts null, div0, range, and class 2450 // checks, these traps from compiled code are double-counted. 2451 // This is harmless; it just means that the PerXTrapLimit values 2452 // are in effect a little smaller than they look. 2453 2454 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2455 if (per_bc_reason != Reason_none) { 2456 // Now take action based on the partially known per-BCI history. 2457 if (maybe_prior_trap 2458 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2459 // If there are too many traps at this BCI, force a recompile. 2460 // This will allow the compiler to see the limit overflow, and 2461 // take corrective action, if possible. The compiler generally 2462 // does not use the exact PerBytecodeTrapLimit value, but instead 2463 // changes its tactics if it sees any traps at all. This provides 2464 // a little hysteresis, delaying a recompile until a trap happens 2465 // several times. 2466 // 2467 // Actually, since there is only one bit of counter per BCI, 2468 // the possible per-BCI counts are {0,1,(per-method count)}. 2469 // This produces accurate results if in fact there is only 2470 // one hot trap site, but begins to get fuzzy if there are 2471 // many sites. For example, if there are ten sites each 2472 // trapping two or more times, they each get the blame for 2473 // all of their traps. 2474 make_not_entrant = true; 2475 } 2476 2477 // Detect repeated recompilation at the same BCI, and enforce a limit. 2478 if (make_not_entrant && maybe_prior_recompile) { 2479 // More than one recompile at this point. 2480 inc_recompile_count = maybe_prior_trap; 2481 } 2482 } else { 2483 // For reasons which are not recorded per-bytecode, we simply 2484 // force recompiles unconditionally. 2485 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2486 make_not_entrant = true; 2487 } 2488 2489 // Go back to the compiler if there are too many traps in this method. 2490 if (this_trap_count >= per_method_trap_limit(reason)) { 2491 // If there are too many traps in this method, force a recompile. 2492 // This will allow the compiler to see the limit overflow, and 2493 // take corrective action, if possible. 2494 // (This condition is an unlikely backstop only, because the 2495 // PerBytecodeTrapLimit is more likely to take effect first, 2496 // if it is applicable.) 2497 make_not_entrant = true; 2498 } 2499 2500 // Here's more hysteresis: If there has been a recompile at 2501 // this trap point already, run the method in the interpreter 2502 // for a while to exercise it more thoroughly. 2503 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2504 reprofile = true; 2505 } 2506 } 2507 2508 // Take requested actions on the method: 2509 2510 // Recompile 2511 if (make_not_entrant) { 2512 if (!nm->make_not_entrant()) { 2513 return; // the call did not change nmethod's state 2514 } 2515 2516 if (pdata != nullptr) { 2517 // Record the recompilation event, if any. 2518 int tstate0 = pdata->trap_state(); 2519 int tstate1 = trap_state_set_recompiled(tstate0, true); 2520 if (tstate1 != tstate0) 2521 pdata->set_trap_state(tstate1); 2522 } 2523 2524 #if INCLUDE_RTM_OPT 2525 // Restart collecting RTM locking abort statistic if the method 2526 // is recompiled for a reason other than RTM state change. 2527 // Assume that in new recompiled code the statistic could be different, 2528 // for example, due to different inlining. 2529 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) && 2530 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2531 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2532 } 2533 #endif 2534 // For code aging we count traps separately here, using make_not_entrant() 2535 // as a guard against simultaneous deopts in multiple threads. 2536 if (reason == Reason_tenured && trap_mdo != nullptr) { 2537 trap_mdo->inc_tenure_traps(); 2538 } 2539 } 2540 2541 if (inc_recompile_count) { 2542 trap_mdo->inc_overflow_recompile_count(); 2543 if ((uint)trap_mdo->overflow_recompile_count() > 2544 (uint)PerBytecodeRecompilationCutoff) { 2545 // Give up on the method containing the bad BCI. 2546 if (trap_method() == nm->method()) { 2547 make_not_compilable = true; 2548 } else { 2549 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2550 // But give grace to the enclosing nm->method(). 2551 } 2552 } 2553 } 2554 2555 // Reprofile 2556 if (reprofile) { 2557 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2558 } 2559 2560 // Give up compiling 2561 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2562 assert(make_not_entrant, "consistent"); 2563 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2564 } 2565 2566 if (ProfileExceptionHandlers && trap_mdo != nullptr) { 2567 BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci); 2568 if (exception_handler_data != nullptr) { 2569 // uncommon trap at the start of an exception handler. 2570 // C2 generates these for un-entered exception handlers. 2571 // mark the handler as entered to avoid generating 2572 // another uncommon trap the next time the handler is compiled 2573 exception_handler_data->set_exception_handler_entered(); 2574 } 2575 } 2576 2577 } // Free marked resources 2578 2579 } 2580 JRT_END 2581 2582 ProfileData* 2583 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2584 int trap_bci, 2585 Deoptimization::DeoptReason reason, 2586 bool update_total_trap_count, 2587 #if INCLUDE_JVMCI 2588 bool is_osr, 2589 #endif 2590 Method* compiled_method, 2591 //outputs: 2592 uint& ret_this_trap_count, 2593 bool& ret_maybe_prior_trap, 2594 bool& ret_maybe_prior_recompile) { 2595 trap_mdo->check_extra_data_locked(); 2596 2597 bool maybe_prior_trap = false; 2598 bool maybe_prior_recompile = false; 2599 uint this_trap_count = 0; 2600 if (update_total_trap_count) { 2601 uint idx = reason; 2602 #if INCLUDE_JVMCI 2603 if (is_osr) { 2604 // Upper half of history array used for traps in OSR compilations 2605 idx += Reason_TRAP_HISTORY_LENGTH; 2606 } 2607 #endif 2608 uint prior_trap_count = trap_mdo->trap_count(idx); 2609 this_trap_count = trap_mdo->inc_trap_count(idx); 2610 2611 // If the runtime cannot find a place to store trap history, 2612 // it is estimated based on the general condition of the method. 2613 // If the method has ever been recompiled, or has ever incurred 2614 // a trap with the present reason , then this BCI is assumed 2615 // (pessimistically) to be the culprit. 2616 maybe_prior_trap = (prior_trap_count != 0); 2617 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2618 } 2619 ProfileData* pdata = nullptr; 2620 2621 2622 // For reasons which are recorded per bytecode, we check per-BCI data. 2623 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2624 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2625 if (per_bc_reason != Reason_none) { 2626 // Find the profile data for this BCI. If there isn't one, 2627 // try to allocate one from the MDO's set of spares. 2628 // This will let us detect a repeated trap at this point. 2629 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr); 2630 2631 if (pdata != nullptr) { 2632 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2633 if (LogCompilation && xtty != nullptr) { 2634 ttyLocker ttyl; 2635 // no more room for speculative traps in this MDO 2636 xtty->elem("speculative_traps_oom"); 2637 } 2638 } 2639 // Query the trap state of this profile datum. 2640 int tstate0 = pdata->trap_state(); 2641 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2642 maybe_prior_trap = false; 2643 if (!trap_state_is_recompiled(tstate0)) 2644 maybe_prior_recompile = false; 2645 2646 // Update the trap state of this profile datum. 2647 int tstate1 = tstate0; 2648 // Record the reason. 2649 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2650 // Store the updated state on the MDO, for next time. 2651 if (tstate1 != tstate0) 2652 pdata->set_trap_state(tstate1); 2653 } else { 2654 if (LogCompilation && xtty != nullptr) { 2655 ttyLocker ttyl; 2656 // Missing MDP? Leave a small complaint in the log. 2657 xtty->elem("missing_mdp bci='%d'", trap_bci); 2658 } 2659 } 2660 } 2661 2662 // Return results: 2663 ret_this_trap_count = this_trap_count; 2664 ret_maybe_prior_trap = maybe_prior_trap; 2665 ret_maybe_prior_recompile = maybe_prior_recompile; 2666 return pdata; 2667 } 2668 2669 void 2670 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2671 ResourceMark rm; 2672 // Ignored outputs: 2673 uint ignore_this_trap_count; 2674 bool ignore_maybe_prior_trap; 2675 bool ignore_maybe_prior_recompile; 2676 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2677 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2678 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2679 2680 // Lock to read ProfileData, and ensure lock is not broken by a safepoint 2681 MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag); 2682 2683 query_update_method_data(trap_mdo, trap_bci, 2684 (DeoptReason)reason, 2685 update_total_counts, 2686 #if INCLUDE_JVMCI 2687 false, 2688 #endif 2689 nullptr, 2690 ignore_this_trap_count, 2691 ignore_maybe_prior_trap, 2692 ignore_maybe_prior_recompile); 2693 } 2694 2695 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) { 2696 // Enable WXWrite: current function is called from methods compiled by C2 directly 2697 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2698 2699 // Still in Java no safepoints 2700 { 2701 // This enters VM and may safepoint 2702 uncommon_trap_inner(current, trap_request); 2703 } 2704 HandleMark hm(current); 2705 return fetch_unroll_info_helper(current, exec_mode); 2706 } 2707 2708 // Local derived constants. 2709 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2710 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2711 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2712 2713 //---------------------------trap_state_reason--------------------------------- 2714 Deoptimization::DeoptReason 2715 Deoptimization::trap_state_reason(int trap_state) { 2716 // This assert provides the link between the width of DataLayout::trap_bits 2717 // and the encoding of "recorded" reasons. It ensures there are enough 2718 // bits to store all needed reasons in the per-BCI MDO profile. 2719 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2720 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2721 trap_state -= recompile_bit; 2722 if (trap_state == DS_REASON_MASK) { 2723 return Reason_many; 2724 } else { 2725 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2726 return (DeoptReason)trap_state; 2727 } 2728 } 2729 //-------------------------trap_state_has_reason------------------------------- 2730 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2731 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2732 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2733 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2734 trap_state -= recompile_bit; 2735 if (trap_state == DS_REASON_MASK) { 2736 return -1; // true, unspecifically (bottom of state lattice) 2737 } else if (trap_state == reason) { 2738 return 1; // true, definitely 2739 } else if (trap_state == 0) { 2740 return 0; // false, definitely (top of state lattice) 2741 } else { 2742 return 0; // false, definitely 2743 } 2744 } 2745 //-------------------------trap_state_add_reason------------------------------- 2746 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2747 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2748 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2749 trap_state -= recompile_bit; 2750 if (trap_state == DS_REASON_MASK) { 2751 return trap_state + recompile_bit; // already at state lattice bottom 2752 } else if (trap_state == reason) { 2753 return trap_state + recompile_bit; // the condition is already true 2754 } else if (trap_state == 0) { 2755 return reason + recompile_bit; // no condition has yet been true 2756 } else { 2757 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2758 } 2759 } 2760 //-----------------------trap_state_is_recompiled------------------------------ 2761 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2762 return (trap_state & DS_RECOMPILE_BIT) != 0; 2763 } 2764 //-----------------------trap_state_set_recompiled----------------------------- 2765 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2766 if (z) return trap_state | DS_RECOMPILE_BIT; 2767 else return trap_state & ~DS_RECOMPILE_BIT; 2768 } 2769 //---------------------------format_trap_state--------------------------------- 2770 // This is used for debugging and diagnostics, including LogFile output. 2771 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2772 int trap_state) { 2773 assert(buflen > 0, "sanity"); 2774 DeoptReason reason = trap_state_reason(trap_state); 2775 bool recomp_flag = trap_state_is_recompiled(trap_state); 2776 // Re-encode the state from its decoded components. 2777 int decoded_state = 0; 2778 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2779 decoded_state = trap_state_add_reason(decoded_state, reason); 2780 if (recomp_flag) 2781 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2782 // If the state re-encodes properly, format it symbolically. 2783 // Because this routine is used for debugging and diagnostics, 2784 // be robust even if the state is a strange value. 2785 size_t len; 2786 if (decoded_state != trap_state) { 2787 // Random buggy state that doesn't decode?? 2788 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2789 } else { 2790 len = jio_snprintf(buf, buflen, "%s%s", 2791 trap_reason_name(reason), 2792 recomp_flag ? " recompiled" : ""); 2793 } 2794 return buf; 2795 } 2796 2797 2798 //--------------------------------statics-------------------------------------- 2799 const char* Deoptimization::_trap_reason_name[] = { 2800 // Note: Keep this in sync. with enum DeoptReason. 2801 "none", 2802 "null_check", 2803 "null_assert" JVMCI_ONLY("_or_unreached0"), 2804 "range_check", 2805 "class_check", 2806 "array_check", 2807 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2808 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2809 "profile_predicate", 2810 "unloaded", 2811 "uninitialized", 2812 "initialized", 2813 "unreached", 2814 "unhandled", 2815 "constraint", 2816 "div0_check", 2817 "age", 2818 "predicate", 2819 "loop_limit_check", 2820 "speculate_class_check", 2821 "speculate_null_check", 2822 "speculate_null_assert", 2823 "rtm_state_change", 2824 "unstable_if", 2825 "unstable_fused_if", 2826 "receiver_constraint", 2827 #if INCLUDE_JVMCI 2828 "aliasing", 2829 "transfer_to_interpreter", 2830 "not_compiled_exception_handler", 2831 "unresolved", 2832 "jsr_mismatch", 2833 #endif 2834 "tenured" 2835 }; 2836 const char* Deoptimization::_trap_action_name[] = { 2837 // Note: Keep this in sync. with enum DeoptAction. 2838 "none", 2839 "maybe_recompile", 2840 "reinterpret", 2841 "make_not_entrant", 2842 "make_not_compilable" 2843 }; 2844 2845 const char* Deoptimization::trap_reason_name(int reason) { 2846 // Check that every reason has a name 2847 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2848 2849 if (reason == Reason_many) return "many"; 2850 if ((uint)reason < Reason_LIMIT) 2851 return _trap_reason_name[reason]; 2852 static char buf[20]; 2853 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2854 return buf; 2855 } 2856 const char* Deoptimization::trap_action_name(int action) { 2857 // Check that every action has a name 2858 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2859 2860 if ((uint)action < Action_LIMIT) 2861 return _trap_action_name[action]; 2862 static char buf[20]; 2863 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2864 return buf; 2865 } 2866 2867 // This is used for debugging and diagnostics, including LogFile output. 2868 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2869 int trap_request) { 2870 jint unloaded_class_index = trap_request_index(trap_request); 2871 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2872 const char* action = trap_action_name(trap_request_action(trap_request)); 2873 #if INCLUDE_JVMCI 2874 int debug_id = trap_request_debug_id(trap_request); 2875 #endif 2876 size_t len; 2877 if (unloaded_class_index < 0) { 2878 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2879 reason, action 2880 #if INCLUDE_JVMCI 2881 ,debug_id 2882 #endif 2883 ); 2884 } else { 2885 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2886 reason, action, unloaded_class_index 2887 #if INCLUDE_JVMCI 2888 ,debug_id 2889 #endif 2890 ); 2891 } 2892 return buf; 2893 } 2894 2895 juint Deoptimization::_deoptimization_hist 2896 [Deoptimization::Reason_LIMIT] 2897 [1 + Deoptimization::Action_LIMIT] 2898 [Deoptimization::BC_CASE_LIMIT] 2899 = {0}; 2900 2901 enum { 2902 LSB_BITS = 8, 2903 LSB_MASK = right_n_bits(LSB_BITS) 2904 }; 2905 2906 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2907 Bytecodes::Code bc) { 2908 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2909 assert(action >= 0 && action < Action_LIMIT, "oob"); 2910 _deoptimization_hist[Reason_none][0][0] += 1; // total 2911 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2912 juint* cases = _deoptimization_hist[reason][1+action]; 2913 juint* bc_counter_addr = nullptr; 2914 juint bc_counter = 0; 2915 // Look for an unused counter, or an exact match to this BC. 2916 if (bc != Bytecodes::_illegal) { 2917 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2918 juint* counter_addr = &cases[bc_case]; 2919 juint counter = *counter_addr; 2920 if ((counter == 0 && bc_counter_addr == nullptr) 2921 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2922 // this counter is either free or is already devoted to this BC 2923 bc_counter_addr = counter_addr; 2924 bc_counter = counter | bc; 2925 } 2926 } 2927 } 2928 if (bc_counter_addr == nullptr) { 2929 // Overflow, or no given bytecode. 2930 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2931 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2932 } 2933 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2934 } 2935 2936 jint Deoptimization::total_deoptimization_count() { 2937 return _deoptimization_hist[Reason_none][0][0]; 2938 } 2939 2940 // Get the deopt count for a specific reason and a specific action. If either 2941 // one of 'reason' or 'action' is null, the method returns the sum of all 2942 // deoptimizations with the specific 'action' or 'reason' respectively. 2943 // If both arguments are null, the method returns the total deopt count. 2944 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2945 if (reason_str == nullptr && action_str == nullptr) { 2946 return total_deoptimization_count(); 2947 } 2948 juint counter = 0; 2949 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2950 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) { 2951 for (int action = 0; action < Action_LIMIT; action++) { 2952 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) { 2953 juint* cases = _deoptimization_hist[reason][1+action]; 2954 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2955 counter += cases[bc_case] >> LSB_BITS; 2956 } 2957 } 2958 } 2959 } 2960 } 2961 return counter; 2962 } 2963 2964 void Deoptimization::print_statistics() { 2965 juint total = total_deoptimization_count(); 2966 juint account = total; 2967 if (total != 0) { 2968 ttyLocker ttyl; 2969 if (xtty != nullptr) xtty->head("statistics type='deoptimization'"); 2970 tty->print_cr("Deoptimization traps recorded:"); 2971 #define PRINT_STAT_LINE(name, r) \ 2972 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2973 PRINT_STAT_LINE("total", total); 2974 // For each non-zero entry in the histogram, print the reason, 2975 // the action, and (if specifically known) the type of bytecode. 2976 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2977 for (int action = 0; action < Action_LIMIT; action++) { 2978 juint* cases = _deoptimization_hist[reason][1+action]; 2979 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2980 juint counter = cases[bc_case]; 2981 if (counter != 0) { 2982 char name[1*K]; 2983 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2984 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2985 bc = Bytecodes::_illegal; 2986 os::snprintf_checked(name, sizeof(name), "%s/%s/%s", 2987 trap_reason_name(reason), 2988 trap_action_name(action), 2989 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2990 juint r = counter >> LSB_BITS; 2991 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2992 account -= r; 2993 } 2994 } 2995 } 2996 } 2997 if (account != 0) { 2998 PRINT_STAT_LINE("unaccounted", account); 2999 } 3000 #undef PRINT_STAT_LINE 3001 if (xtty != nullptr) xtty->tail("statistics"); 3002 } 3003 } 3004 3005 #else // COMPILER2_OR_JVMCI 3006 3007 3008 // Stubs for C1 only system. 3009 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 3010 return false; 3011 } 3012 3013 const char* Deoptimization::trap_reason_name(int reason) { 3014 return "unknown"; 3015 } 3016 3017 jint Deoptimization::total_deoptimization_count() { 3018 return 0; 3019 } 3020 3021 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 3022 return 0; 3023 } 3024 3025 void Deoptimization::print_statistics() { 3026 // no output 3027 } 3028 3029 void 3030 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 3031 // no update 3032 } 3033 3034 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 3035 return 0; 3036 } 3037 3038 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 3039 Bytecodes::Code bc) { 3040 // no update 3041 } 3042 3043 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 3044 int trap_state) { 3045 jio_snprintf(buf, buflen, "#%d", trap_state); 3046 return buf; 3047 } 3048 3049 #endif // COMPILER2_OR_JVMCI