1 /* 2 * Copyright (c) 1997, 2023, 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/threadSMR.hpp" 90 #include "runtime/threadWXSetters.inline.hpp" 91 #include "runtime/vframe.hpp" 92 #include "runtime/vframeArray.hpp" 93 #include "runtime/vframe_hp.hpp" 94 #include "runtime/vmOperations.hpp" 95 #include "utilities/events.hpp" 96 #include "utilities/growableArray.hpp" 97 #include "utilities/macros.hpp" 98 #include "utilities/preserveException.hpp" 99 #include "utilities/xmlstream.hpp" 100 #if INCLUDE_JFR 101 #include "jfr/jfrEvents.hpp" 102 #include "jfr/metadata/jfrSerializer.hpp" 103 #endif 104 105 uint64_t DeoptimizationScope::_committed_deopt_gen = 0; 106 uint64_t DeoptimizationScope::_active_deopt_gen = 1; 107 bool DeoptimizationScope::_committing_in_progress = false; 108 109 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) { 110 DEBUG_ONLY(_deopted = false;) 111 112 MutexLocker ml(CompiledMethod_lock, Mutex::_no_safepoint_check_flag); 113 // If there is nothing to deopt _required_gen is the same as comitted. 114 _required_gen = DeoptimizationScope::_committed_deopt_gen; 115 } 116 117 DeoptimizationScope::~DeoptimizationScope() { 118 assert(_deopted, "Deopt not executed"); 119 } 120 121 void DeoptimizationScope::mark(CompiledMethod* cm, bool inc_recompile_counts) { 122 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock, 123 Mutex::_no_safepoint_check_flag); 124 125 // If it's already marked but we still need it to be deopted. 126 if (cm->is_marked_for_deoptimization()) { 127 dependent(cm); 128 return; 129 } 130 131 CompiledMethod::DeoptimizationStatus status = 132 inc_recompile_counts ? CompiledMethod::deoptimize : CompiledMethod::deoptimize_noupdate; 133 Atomic::store(&cm->_deoptimization_status, status); 134 135 // Make sure active is not committed 136 assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be"); 137 assert(cm->_deoptimization_generation == 0, "Is already marked"); 138 139 cm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen; 140 _required_gen = DeoptimizationScope::_active_deopt_gen; 141 } 142 143 void DeoptimizationScope::dependent(CompiledMethod* cm) { 144 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock, 145 Mutex::_no_safepoint_check_flag); 146 // A method marked by someone else may have a _required_gen lower than what we marked with. 147 // Therefore only store it if it's higher than _required_gen. 148 if (_required_gen < cm->_deoptimization_generation) { 149 _required_gen = cm->_deoptimization_generation; 150 } 151 } 152 153 void DeoptimizationScope::deoptimize_marked() { 154 assert(!_deopted, "Already deopted"); 155 156 // We are not alive yet. 157 if (!Universe::is_fully_initialized()) { 158 DEBUG_ONLY(_deopted = true;) 159 return; 160 } 161 162 // Safepoints are a special case, handled here. 163 if (SafepointSynchronize::is_at_safepoint()) { 164 DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen; 165 DeoptimizationScope::_active_deopt_gen++; 166 Deoptimization::deoptimize_all_marked(); 167 DEBUG_ONLY(_deopted = true;) 168 return; 169 } 170 171 uint64_t comitting = 0; 172 bool wait = false; 173 while (true) { 174 { 175 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock, 176 Mutex::_no_safepoint_check_flag); 177 // First we check if we or someone else already deopted the gen we want. 178 if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) { 179 DEBUG_ONLY(_deopted = true;) 180 return; 181 } 182 if (!_committing_in_progress) { 183 // The version we are about to commit. 184 comitting = DeoptimizationScope::_active_deopt_gen; 185 // Make sure new marks use a higher gen. 186 DeoptimizationScope::_active_deopt_gen++; 187 _committing_in_progress = true; 188 wait = false; 189 } else { 190 // Another thread is handshaking and committing a gen. 191 wait = true; 192 } 193 } 194 if (wait) { 195 // Wait and let the concurrent handshake be performed. 196 ThreadBlockInVM tbivm(JavaThread::current()); 197 os::naked_yield(); 198 } else { 199 // Performs the handshake. 200 Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred. 201 DEBUG_ONLY(_deopted = true;) 202 { 203 MutexLocker ml(CompiledMethod_lock->owned_by_self() ? nullptr : CompiledMethod_lock, 204 Mutex::_no_safepoint_check_flag); 205 // Make sure that committed doesn't go backwards. 206 // Should only happen if we did a deopt during a safepoint above. 207 if (DeoptimizationScope::_committed_deopt_gen < comitting) { 208 DeoptimizationScope::_committed_deopt_gen = comitting; 209 } 210 _committing_in_progress = false; 211 212 assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be"); 213 214 return; 215 } 216 } 217 } 218 } 219 220 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 221 int caller_adjustment, 222 int caller_actual_parameters, 223 int number_of_frames, 224 intptr_t* frame_sizes, 225 address* frame_pcs, 226 BasicType return_type, 227 int exec_mode) { 228 _size_of_deoptimized_frame = size_of_deoptimized_frame; 229 _caller_adjustment = caller_adjustment; 230 _caller_actual_parameters = caller_actual_parameters; 231 _number_of_frames = number_of_frames; 232 _frame_sizes = frame_sizes; 233 _frame_pcs = frame_pcs; 234 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 235 _return_type = return_type; 236 _initial_info = 0; 237 // PD (x86 only) 238 _counter_temp = 0; 239 _unpack_kind = exec_mode; 240 _sender_sp_temp = 0; 241 242 _total_frame_sizes = size_of_frames(); 243 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 244 } 245 246 Deoptimization::UnrollBlock::~UnrollBlock() { 247 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 248 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 249 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 250 } 251 252 int Deoptimization::UnrollBlock::size_of_frames() const { 253 // Account first for the adjustment of the initial frame 254 int result = _caller_adjustment; 255 for (int index = 0; index < number_of_frames(); index++) { 256 result += frame_sizes()[index]; 257 } 258 return result; 259 } 260 261 void Deoptimization::UnrollBlock::print() { 262 ResourceMark rm; 263 stringStream st; 264 st.print_cr("UnrollBlock"); 265 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 266 st.print( " frame_sizes: "); 267 for (int index = 0; index < number_of_frames(); index++) { 268 st.print(INTX_FORMAT " ", frame_sizes()[index]); 269 } 270 st.cr(); 271 tty->print_raw(st.freeze()); 272 } 273 274 // In order to make fetch_unroll_info work properly with escape 275 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 276 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 277 // which is called from the method fetch_unroll_info_helper below. 278 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 279 // fetch_unroll_info() is called at the beginning of the deoptimization 280 // handler. Note this fact before we start generating temporary frames 281 // that can confuse an asynchronous stack walker. This counter is 282 // decremented at the end of unpack_frames(). 283 current->inc_in_deopt_handler(); 284 285 if (exec_mode == Unpack_exception) { 286 // When we get here, a callee has thrown an exception into a deoptimized 287 // frame. That throw might have deferred stack watermark checking until 288 // after unwinding. So we deal with such deferred requests here. 289 StackWatermarkSet::after_unwind(current); 290 } 291 292 return fetch_unroll_info_helper(current, exec_mode); 293 JRT_END 294 295 #if COMPILER2_OR_JVMCI 296 // print information about reallocated objects 297 static void print_objects(JavaThread* deoptee_thread, 298 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 299 ResourceMark rm; 300 stringStream st; // change to logStream with logging 301 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 302 fieldDescriptor fd; 303 304 for (int i = 0; i < objects->length(); i++) { 305 ObjectValue* sv = (ObjectValue*) objects->at(i); 306 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 307 Handle obj = sv->value(); 308 309 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 310 k->print_value_on(&st); 311 assert(obj.not_null() || k->is_inline_klass() || realloc_failures, "reallocation was missed"); 312 if (obj.is_null()) { 313 if (k->is_inline_klass()) { 314 st.print(" is null"); 315 } else { 316 st.print(" allocation failed"); 317 } 318 } else { 319 st.print(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize); 320 } 321 st.cr(); 322 323 if (Verbose && !obj.is_null()) { 324 k->oop_print_on(obj(), &st); 325 } 326 } 327 tty->print_raw(st.freeze()); 328 } 329 330 static bool rematerialize_objects(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 331 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 332 bool& deoptimized_objects) { 333 bool realloc_failures = false; 334 assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames"); 335 336 JavaThread* deoptee_thread = chunk->at(0)->thread(); 337 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 338 "a frame can only be deoptimized by the owner thread"); 339 340 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map); 341 342 // The flag return_oop() indicates call sites which return oop 343 // in compiled code. Such sites include java method calls, 344 // runtime calls (for example, used to allocate new objects/arrays 345 // on slow code path) and any other calls generated in compiled code. 346 // It is not guaranteed that we can get such information here only 347 // by analyzing bytecode in deoptimized frames. This is why this flag 348 // is set during method compilation (see Compile::Process_OopMap_Node()). 349 // If the previous frame was popped or if we are dispatching an exception, 350 // we don't have an oop result. 351 ScopeDesc* scope = chunk->at(0)->scope(); 352 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 353 // In case of the return of multiple values, we must take care 354 // of all oop return values. 355 GrowableArray<Handle> return_oops; 356 InlineKlass* vk = nullptr; 357 if (save_oop_result && scope->return_scalarized()) { 358 vk = InlineKlass::returned_inline_klass(map); 359 if (vk != nullptr) { 360 vk->save_oop_fields(map, return_oops); 361 save_oop_result = false; 362 } 363 } 364 if (save_oop_result) { 365 // Reallocation may trigger GC. If deoptimization happened on return from 366 // call which returns oop we need to save it since it is not in oopmap. 367 oop result = deoptee.saved_oop_result(&map); 368 assert(oopDesc::is_oop_or_null(result), "must be oop"); 369 return_oops.push(Handle(thread, result)); 370 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 371 if (TraceDeoptimization) { 372 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 373 tty->cr(); 374 } 375 } 376 if (objects != nullptr || vk != nullptr) { 377 if (exec_mode == Deoptimization::Unpack_none) { 378 assert(thread->thread_state() == _thread_in_vm, "assumption"); 379 JavaThread* THREAD = thread; // For exception macros. 380 // Clear pending OOM if reallocation fails and return true indicating allocation failure 381 if (vk != nullptr) { 382 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, CHECK_AND_CLEAR_(true)); 383 } 384 if (objects != nullptr) { 385 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 386 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 387 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, CHECK_AND_CLEAR_(true)); 388 } 389 deoptimized_objects = true; 390 } else { 391 JavaThread* current = thread; // For JRT_BLOCK 392 JRT_BLOCK 393 if (vk != nullptr) { 394 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, THREAD); 395 } 396 if (objects != nullptr) { 397 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 398 bool skip_internal = (compiled_method != nullptr) && !compiled_method->is_compiled_by_jvmci(); 399 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 400 } 401 JRT_END 402 } 403 if (TraceDeoptimization) { 404 print_objects(deoptee_thread, objects, realloc_failures); 405 } 406 } 407 if (save_oop_result || vk != nullptr) { 408 // Restore result. 409 assert(return_oops.length() == 1, "no inline type"); 410 deoptee.set_saved_oop_result(&map, return_oops.pop()()); 411 } 412 return realloc_failures; 413 } 414 415 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 416 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 417 JavaThread* deoptee_thread = chunk->at(0)->thread(); 418 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 419 assert(thread == Thread::current(), "should be"); 420 HandleMark hm(thread); 421 #ifndef PRODUCT 422 bool first = true; 423 #endif // !PRODUCT 424 for (int i = 0; i < chunk->length(); i++) { 425 compiledVFrame* cvf = chunk->at(i); 426 assert (cvf->scope() != nullptr,"expect only compiled java frames"); 427 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 428 if (monitors->is_nonempty()) { 429 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 430 exec_mode, realloc_failures); 431 deoptimized_objects = deoptimized_objects || relocked; 432 #ifndef PRODUCT 433 if (PrintDeoptimizationDetails) { 434 ResourceMark rm; 435 stringStream st; 436 for (int j = 0; j < monitors->length(); j++) { 437 MonitorInfo* mi = monitors->at(j); 438 if (mi->eliminated()) { 439 if (first) { 440 first = false; 441 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 442 } 443 if (exec_mode == Deoptimization::Unpack_none) { 444 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 445 if (monitor != nullptr && monitor->object() == mi->owner()) { 446 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 447 continue; 448 } 449 } 450 if (mi->owner_is_scalar_replaced()) { 451 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 452 st.print_cr(" failed reallocation for klass %s", k->external_name()); 453 } else { 454 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 455 } 456 } 457 } 458 tty->print_raw(st.freeze()); 459 } 460 #endif // !PRODUCT 461 } 462 } 463 } 464 465 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 466 // The given vframes cover one physical frame. 467 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 468 bool& realloc_failures) { 469 frame deoptee = chunk->at(0)->fr(); 470 JavaThread* deoptee_thread = chunk->at(0)->thread(); 471 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 472 RegisterMap map(chunk->at(0)->register_map()); 473 bool deoptimized_objects = false; 474 475 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false); 476 477 // Reallocate the non-escaping objects and restore their fields. 478 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 479 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 480 realloc_failures = rematerialize_objects(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects); 481 } 482 483 // MonitorInfo structures used in eliminate_locks are not GC safe. 484 NoSafepointVerifier no_safepoint; 485 486 // Now relock objects if synchronization on them was eliminated. 487 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 488 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 489 } 490 return deoptimized_objects; 491 } 492 #endif // COMPILER2_OR_JVMCI 493 494 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 495 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 496 // When we get here we are about to unwind the deoptee frame. In order to 497 // catch not yet safe to use frames, the following stack watermark barrier 498 // poll will make such frames safe to use. 499 StackWatermarkSet::before_unwind(current); 500 501 // Note: there is a safepoint safety issue here. No matter whether we enter 502 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 503 // the vframeArray is created. 504 // 505 506 // Allocate our special deoptimization ResourceMark 507 DeoptResourceMark* dmark = new DeoptResourceMark(current); 508 assert(current->deopt_mark() == nullptr, "Pending deopt!"); 509 current->set_deopt_mark(dmark); 510 511 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 512 RegisterMap map(current, 513 RegisterMap::UpdateMap::include, 514 RegisterMap::ProcessFrames::include, 515 RegisterMap::WalkContinuation::skip); 516 RegisterMap dummy_map(current, 517 RegisterMap::UpdateMap::skip, 518 RegisterMap::ProcessFrames::include, 519 RegisterMap::WalkContinuation::skip); 520 // Now get the deoptee with a valid map 521 frame deoptee = stub_frame.sender(&map); 522 // Set the deoptee nmethod 523 assert(current->deopt_compiled_method() == nullptr, "Pending deopt!"); 524 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 525 current->set_deopt_compiled_method(cm); 526 527 if (VerifyStack) { 528 current->validate_frame_layout(); 529 } 530 531 // Create a growable array of VFrames where each VFrame represents an inlined 532 // Java frame. This storage is allocated with the usual system arena. 533 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 534 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 535 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 536 while (!vf->is_top()) { 537 assert(vf->is_compiled_frame(), "Wrong frame type"); 538 chunk->push(compiledVFrame::cast(vf)); 539 vf = vf->sender(); 540 } 541 assert(vf->is_compiled_frame(), "Wrong frame type"); 542 chunk->push(compiledVFrame::cast(vf)); 543 544 bool realloc_failures = false; 545 546 #if COMPILER2_OR_JVMCI 547 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 548 549 // Reallocate the non-escaping objects and restore their fields. Then 550 // relock objects if synchronization on them was eliminated. 551 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 552 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 553 bool unused; 554 realloc_failures = rematerialize_objects(current, exec_mode, cm, deoptee, map, chunk, unused); 555 } 556 #endif // COMPILER2_OR_JVMCI 557 558 // Ensure that no safepoint is taken after pointers have been stored 559 // in fields of rematerialized objects. If a safepoint occurs from here on 560 // out the java state residing in the vframeArray will be missed. 561 // Locks may be rebaised in a safepoint. 562 NoSafepointVerifier no_safepoint; 563 564 #if COMPILER2_OR_JVMCI 565 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 566 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 567 bool unused; 568 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 569 } 570 #endif // COMPILER2_OR_JVMCI 571 572 ScopeDesc* trap_scope = chunk->at(0)->scope(); 573 Handle exceptionObject; 574 if (trap_scope->rethrow_exception()) { 575 #ifndef PRODUCT 576 if (PrintDeoptimizationDetails) { 577 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()); 578 } 579 #endif // !PRODUCT 580 581 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 582 guarantee(expressions != nullptr && expressions->length() > 0, "must have exception to throw"); 583 ScopeValue* topOfStack = expressions->top(); 584 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 585 guarantee(exceptionObject() != nullptr, "exception oop can not be null"); 586 } 587 588 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 589 #if COMPILER2_OR_JVMCI 590 if (realloc_failures) { 591 // This destroys all ScopedValue bindings. 592 current->clear_scopedValueBindings(); 593 pop_frames_failed_reallocs(current, array); 594 } 595 #endif 596 597 assert(current->vframe_array_head() == nullptr, "Pending deopt!"); 598 current->set_vframe_array_head(array); 599 600 // Now that the vframeArray has been created if we have any deferred local writes 601 // added by jvmti then we can free up that structure as the data is now in the 602 // vframeArray 603 604 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 605 606 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 607 CodeBlob* cb = stub_frame.cb(); 608 // Verify we have the right vframeArray 609 assert(cb->frame_size() >= 0, "Unexpected frame size"); 610 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 611 612 // If the deopt call site is a MethodHandle invoke call site we have 613 // to adjust the unpack_sp. 614 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 615 if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc())) 616 unpack_sp = deoptee.unextended_sp(); 617 618 #ifdef ASSERT 619 assert(cb->is_deoptimization_stub() || 620 cb->is_uncommon_trap_stub() || 621 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 622 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 623 "unexpected code blob: %s", cb->name()); 624 #endif 625 626 // This is a guarantee instead of an assert because if vframe doesn't match 627 // we will unpack the wrong deoptimized frame and wind up in strange places 628 // where it will be very difficult to figure out what went wrong. Better 629 // to die an early death here than some very obscure death later when the 630 // trail is cold. 631 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 632 // in that it will fail to detect a problem when there is one. This needs 633 // more work in tiger timeframe. 634 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 635 636 int number_of_frames = array->frames(); 637 638 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 639 // virtual activation, which is the reverse of the elements in the vframes array. 640 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 641 // +1 because we always have an interpreter return address for the final slot. 642 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 643 int popframe_extra_args = 0; 644 // Create an interpreter return address for the stub to use as its return 645 // address so the skeletal frames are perfectly walkable 646 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 647 648 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 649 // activation be put back on the expression stack of the caller for reexecution 650 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 651 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 652 } 653 654 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 655 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 656 // than simply use array->sender.pc(). This requires us to walk the current set of frames 657 // 658 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 659 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 660 661 // It's possible that the number of parameters at the call site is 662 // different than number of arguments in the callee when method 663 // handles are used. If the caller is interpreted get the real 664 // value so that the proper amount of space can be added to it's 665 // frame. 666 bool caller_was_method_handle = false; 667 if (deopt_sender.is_interpreted_frame()) { 668 methodHandle method(current, deopt_sender.interpreter_frame_method()); 669 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 670 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 671 // Method handle invokes may involve fairly arbitrary chains of 672 // calls so it's impossible to know how much actual space the 673 // caller has for locals. 674 caller_was_method_handle = true; 675 } 676 } 677 678 // 679 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 680 // frame_sizes/frame_pcs[1] next oldest frame (int) 681 // frame_sizes/frame_pcs[n] youngest frame (int) 682 // 683 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 684 // owns the space for the return address to it's caller). Confusing ain't it. 685 // 686 // The vframe array can address vframes with indices running from 687 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 688 // When we create the skeletal frames we need the oldest frame to be in the zero slot 689 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 690 // so things look a little strange in this loop. 691 // 692 int callee_parameters = 0; 693 int callee_locals = 0; 694 for (int index = 0; index < array->frames(); index++ ) { 695 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 696 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 697 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 698 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 699 callee_locals, 700 index == 0, 701 popframe_extra_args); 702 // This pc doesn't have to be perfect just good enough to identify the frame 703 // as interpreted so the skeleton frame will be walkable 704 // The correct pc will be set when the skeleton frame is completely filled out 705 // The final pc we store in the loop is wrong and will be overwritten below 706 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 707 708 callee_parameters = array->element(index)->method()->size_of_parameters(); 709 callee_locals = array->element(index)->method()->max_locals(); 710 popframe_extra_args = 0; 711 } 712 713 // Compute whether the root vframe returns a float or double value. 714 BasicType return_type; 715 { 716 methodHandle method(current, array->element(0)->method()); 717 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 718 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 719 } 720 721 // Compute information for handling adapters and adjusting the frame size of the caller. 722 int caller_adjustment = 0; 723 724 // Compute the amount the oldest interpreter frame will have to adjust 725 // its caller's stack by. If the caller is a compiled frame then 726 // we pretend that the callee has no parameters so that the 727 // extension counts for the full amount of locals and not just 728 // locals-parms. This is because without a c2i adapter the parm 729 // area as created by the compiled frame will not be usable by 730 // the interpreter. (Depending on the calling convention there 731 // may not even be enough space). 732 733 // QQQ I'd rather see this pushed down into last_frame_adjust 734 // and have it take the sender (aka caller). 735 736 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 737 caller_adjustment = last_frame_adjust(0, callee_locals); 738 } else if (callee_locals > callee_parameters) { 739 // The caller frame may need extending to accommodate 740 // non-parameter locals of the first unpacked interpreted frame. 741 // Compute that adjustment. 742 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 743 } 744 745 // If the sender is deoptimized we must retrieve the address of the handler 746 // since the frame will "magically" show the original pc before the deopt 747 // and we'd undo the deopt. 748 749 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 750 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 751 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 752 } 753 754 assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc"); 755 756 #if INCLUDE_JVMCI 757 if (exceptionObject() != nullptr) { 758 current->set_exception_oop(exceptionObject()); 759 exec_mode = Unpack_exception; 760 } 761 #endif 762 763 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 764 assert(current->has_pending_exception(), "should have thrown OOME"); 765 current->set_exception_oop(current->pending_exception()); 766 current->clear_pending_exception(); 767 exec_mode = Unpack_exception; 768 } 769 770 #if INCLUDE_JVMCI 771 if (current->frames_to_pop_failed_realloc() > 0) { 772 current->set_pending_monitorenter(false); 773 } 774 #endif 775 776 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 777 caller_adjustment * BytesPerWord, 778 caller_was_method_handle ? 0 : callee_parameters, 779 number_of_frames, 780 frame_sizes, 781 frame_pcs, 782 return_type, 783 exec_mode); 784 // On some platforms, we need a way to pass some platform dependent 785 // information to the unpacking code so the skeletal frames come out 786 // correct (initial fp value, unextended sp, ...) 787 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 788 789 if (array->frames() > 1) { 790 if (VerifyStack && TraceDeoptimization) { 791 tty->print_cr("Deoptimizing method containing inlining"); 792 } 793 } 794 795 array->set_unroll_block(info); 796 return info; 797 } 798 799 // Called to cleanup deoptimization data structures in normal case 800 // after unpacking to stack and when stack overflow error occurs 801 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 802 vframeArray *array) { 803 804 // Get array if coming from exception 805 if (array == nullptr) { 806 array = thread->vframe_array_head(); 807 } 808 thread->set_vframe_array_head(nullptr); 809 810 // Free the previous UnrollBlock 811 vframeArray* old_array = thread->vframe_array_last(); 812 thread->set_vframe_array_last(array); 813 814 if (old_array != nullptr) { 815 UnrollBlock* old_info = old_array->unroll_block(); 816 old_array->set_unroll_block(nullptr); 817 delete old_info; 818 delete old_array; 819 } 820 821 // Deallocate any resource creating in this routine and any ResourceObjs allocated 822 // inside the vframeArray (StackValueCollections) 823 824 delete thread->deopt_mark(); 825 thread->set_deopt_mark(nullptr); 826 thread->set_deopt_compiled_method(nullptr); 827 828 829 if (JvmtiExport::can_pop_frame()) { 830 // Regardless of whether we entered this routine with the pending 831 // popframe condition bit set, we should always clear it now 832 thread->clear_popframe_condition(); 833 } 834 835 // unpack_frames() is called at the end of the deoptimization handler 836 // and (in C2) at the end of the uncommon trap handler. Note this fact 837 // so that an asynchronous stack walker can work again. This counter is 838 // incremented at the beginning of fetch_unroll_info() and (in C2) at 839 // the beginning of uncommon_trap(). 840 thread->dec_in_deopt_handler(); 841 } 842 843 // Moved from cpu directories because none of the cpus has callee save values. 844 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 845 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 846 847 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 848 // the days we had adapter frames. When we deoptimize a situation where a 849 // compiled caller calls a compiled caller will have registers it expects 850 // to survive the call to the callee. If we deoptimize the callee the only 851 // way we can restore these registers is to have the oldest interpreter 852 // frame that we create restore these values. That is what this routine 853 // will accomplish. 854 855 // At the moment we have modified c2 to not have any callee save registers 856 // so this problem does not exist and this routine is just a place holder. 857 858 assert(f->is_interpreted_frame(), "must be interpreted"); 859 } 860 861 #ifndef PRODUCT 862 static bool falls_through(Bytecodes::Code bc) { 863 switch (bc) { 864 // List may be incomplete. Here we really only care about bytecodes where compiled code 865 // can deoptimize. 866 case Bytecodes::_goto: 867 case Bytecodes::_goto_w: 868 case Bytecodes::_athrow: 869 return false; 870 default: 871 return true; 872 } 873 } 874 #endif 875 876 // Return BasicType of value being returned 877 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 878 assert(thread == JavaThread::current(), "pre-condition"); 879 880 // We are already active in the special DeoptResourceMark any ResourceObj's we 881 // allocate will be freed at the end of the routine. 882 883 // JRT_LEAF methods don't normally allocate handles and there is a 884 // NoHandleMark to enforce that. It is actually safe to use Handles 885 // in a JRT_LEAF method, and sometimes desirable, but to do so we 886 // must use ResetNoHandleMark to bypass the NoHandleMark, and 887 // then use a HandleMark to ensure any Handles we do create are 888 // cleaned up in this scope. 889 ResetNoHandleMark rnhm; 890 HandleMark hm(thread); 891 892 frame stub_frame = thread->last_frame(); 893 894 Continuation::notify_deopt(thread, stub_frame.sp()); 895 896 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 897 // must point to the vframeArray for the unpack frame. 898 vframeArray* array = thread->vframe_array_head(); 899 UnrollBlock* info = array->unroll_block(); 900 901 // We set the last_Java frame. But the stack isn't really parsable here. So we 902 // clear it to make sure JFR understands not to try and walk stacks from events 903 // in here. 904 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 905 thread->frame_anchor()->set_last_Java_sp(nullptr); 906 907 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 908 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 909 910 thread->frame_anchor()->set_last_Java_sp(sp); 911 912 BasicType bt = info->return_type(); 913 914 // If we have an exception pending, claim that the return type is an oop 915 // so the deopt_blob does not overwrite the exception_oop. 916 917 if (exec_mode == Unpack_exception) 918 bt = T_OBJECT; 919 920 // Cleanup thread deopt data 921 cleanup_deopt_info(thread, array); 922 923 #ifndef PRODUCT 924 if (VerifyStack) { 925 ResourceMark res_mark; 926 // Clear pending exception to not break verification code (restored afterwards) 927 PreserveExceptionMark pm(thread); 928 929 thread->validate_frame_layout(); 930 931 // Verify that the just-unpacked frames match the interpreter's 932 // notions of expression stack and locals 933 vframeArray* cur_array = thread->vframe_array_last(); 934 RegisterMap rm(thread, 935 RegisterMap::UpdateMap::skip, 936 RegisterMap::ProcessFrames::include, 937 RegisterMap::WalkContinuation::skip); 938 rm.set_include_argument_oops(false); 939 bool is_top_frame = true; 940 int callee_size_of_parameters = 0; 941 int callee_max_locals = 0; 942 for (int i = 0; i < cur_array->frames(); i++) { 943 vframeArrayElement* el = cur_array->element(i); 944 frame* iframe = el->iframe(); 945 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 946 947 // Get the oop map for this bci 948 InterpreterOopMap mask; 949 int cur_invoke_parameter_size = 0; 950 bool try_next_mask = false; 951 int next_mask_expression_stack_size = -1; 952 int top_frame_expression_stack_adjustment = 0; 953 methodHandle mh(thread, iframe->interpreter_frame_method()); 954 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 955 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 956 int max_bci = mh->code_size(); 957 // Get to the next bytecode if possible 958 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 959 // Check to see if we can grab the number of outgoing arguments 960 // at an uncommon trap for an invoke (where the compiler 961 // generates debug info before the invoke has executed) 962 Bytecodes::Code cur_code = str.next(); 963 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere; 964 if (Bytecodes::is_invoke(cur_code)) { 965 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 966 cur_invoke_parameter_size = invoke.size_of_parameters(); 967 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 968 callee_size_of_parameters++; 969 } 970 } 971 if (str.bci() < max_bci) { 972 next_code = str.next(); 973 if (next_code >= 0) { 974 // The interpreter oop map generator reports results before 975 // the current bytecode has executed except in the case of 976 // calls. It seems to be hard to tell whether the compiler 977 // has emitted debug information matching the "state before" 978 // a given bytecode or the state after, so we try both 979 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) { 980 // Get expression stack size for the next bytecode 981 InterpreterOopMap next_mask; 982 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 983 next_mask_expression_stack_size = next_mask.expression_stack_size(); 984 if (Bytecodes::is_invoke(next_code)) { 985 Bytecode_invoke invoke(mh, str.bci()); 986 next_mask_expression_stack_size += invoke.size_of_parameters(); 987 } 988 // Need to subtract off the size of the result type of 989 // the bytecode because this is not described in the 990 // debug info but returned to the interpreter in the TOS 991 // caching register 992 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 993 if (bytecode_result_type != T_ILLEGAL) { 994 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 995 } 996 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 997 try_next_mask = true; 998 } 999 } 1000 } 1001 1002 // Verify stack depth and oops in frame 1003 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 1004 if (!( 1005 /* SPARC */ 1006 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 1007 /* x86 */ 1008 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 1009 (try_next_mask && 1010 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 1011 top_frame_expression_stack_adjustment))) || 1012 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 1013 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 1014 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 1015 )) { 1016 { 1017 // Print out some information that will help us debug the problem 1018 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 1019 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 1020 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 1021 if (try_next_mask) { 1022 tty->print_cr(" Next code %s", Bytecodes::name(next_code)); 1023 } 1024 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 1025 iframe->interpreter_frame_expression_stack_size()); 1026 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 1027 tty->print_cr(" try_next_mask = %d", try_next_mask); 1028 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 1029 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 1030 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 1031 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 1032 tty->print_cr(" exec_mode = %d", exec_mode); 1033 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 1034 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 1035 tty->print_cr(" Interpreted frames:"); 1036 for (int k = 0; k < cur_array->frames(); k++) { 1037 vframeArrayElement* el = cur_array->element(k); 1038 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 1039 } 1040 cur_array->print_on_2(tty); 1041 } 1042 guarantee(false, "wrong number of expression stack elements during deopt"); 1043 } 1044 VerifyOopClosure verify; 1045 iframe->oops_interpreted_do(&verify, &rm, false); 1046 callee_size_of_parameters = mh->size_of_parameters(); 1047 callee_max_locals = mh->max_locals(); 1048 is_top_frame = false; 1049 } 1050 } 1051 #endif // !PRODUCT 1052 1053 return bt; 1054 JRT_END 1055 1056 class DeoptimizeMarkedClosure : public HandshakeClosure { 1057 public: 1058 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 1059 void do_thread(Thread* thread) { 1060 JavaThread* jt = JavaThread::cast(thread); 1061 jt->deoptimize_marked_methods(); 1062 } 1063 }; 1064 1065 void Deoptimization::deoptimize_all_marked() { 1066 ResourceMark rm; 1067 1068 // Make the dependent methods not entrant 1069 CodeCache::make_marked_nmethods_deoptimized(); 1070 1071 DeoptimizeMarkedClosure deopt; 1072 if (SafepointSynchronize::is_at_safepoint()) { 1073 Threads::java_threads_do(&deopt); 1074 } else { 1075 Handshake::execute(&deopt); 1076 } 1077 } 1078 1079 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1080 = Deoptimization::Action_reinterpret; 1081 1082 #if INCLUDE_JVMCI 1083 template<typename CacheType> 1084 class BoxCacheBase : public CHeapObj<mtCompiler> { 1085 protected: 1086 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 1087 ResourceMark rm(thread); 1088 char* klass_name_str = klass_name->as_C_string(); 1089 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle()); 1090 guarantee(ik != nullptr, "%s must be loaded", klass_name_str); 1091 if (!ik->is_in_error_state()) { 1092 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 1093 CacheType::compute_offsets(ik); 1094 } 1095 return ik; 1096 } 1097 }; 1098 1099 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 1100 PrimitiveType _low; 1101 PrimitiveType _high; 1102 jobject _cache; 1103 protected: 1104 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 1105 BoxCache(Thread* thread) { 1106 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 1107 if (ik->is_in_error_state()) { 1108 _low = 1; 1109 _high = 0; 1110 _cache = nullptr; 1111 } else { 1112 objArrayOop cache = CacheType::cache(ik); 1113 assert(cache->length() > 0, "Empty cache"); 1114 _low = BoxType::value(cache->obj_at(0)); 1115 _high = _low + cache->length() - 1; 1116 _cache = JNIHandles::make_global(Handle(thread, cache)); 1117 } 1118 } 1119 ~BoxCache() { 1120 JNIHandles::destroy_global(_cache); 1121 } 1122 public: 1123 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1124 if (_singleton == nullptr) { 1125 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1126 if (!Atomic::replace_if_null(&_singleton, s)) { 1127 delete s; 1128 } 1129 } 1130 return _singleton; 1131 } 1132 oop lookup(PrimitiveType value) { 1133 if (_low <= value && value <= _high) { 1134 int offset = value - _low; 1135 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1136 } 1137 return nullptr; 1138 } 1139 oop lookup_raw(intptr_t raw_value, bool& cache_init_error) { 1140 if (_cache == nullptr) { 1141 cache_init_error = true; 1142 return nullptr; 1143 } 1144 // Have to cast to avoid little/big-endian problems. 1145 if (sizeof(PrimitiveType) > sizeof(jint)) { 1146 jlong value = (jlong)raw_value; 1147 return lookup(value); 1148 } 1149 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1150 return lookup(value); 1151 } 1152 }; 1153 1154 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1155 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1156 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1157 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1158 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1159 1160 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr; 1161 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr; 1162 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr; 1163 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr; 1164 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr; 1165 1166 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1167 jobject _true_cache; 1168 jobject _false_cache; 1169 protected: 1170 static BooleanBoxCache *_singleton; 1171 BooleanBoxCache(Thread *thread) { 1172 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1173 if (ik->is_in_error_state()) { 1174 _true_cache = nullptr; 1175 _false_cache = nullptr; 1176 } else { 1177 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1178 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1179 } 1180 } 1181 ~BooleanBoxCache() { 1182 JNIHandles::destroy_global(_true_cache); 1183 JNIHandles::destroy_global(_false_cache); 1184 } 1185 public: 1186 static BooleanBoxCache* singleton(Thread* thread) { 1187 if (_singleton == nullptr) { 1188 BooleanBoxCache* s = new BooleanBoxCache(thread); 1189 if (!Atomic::replace_if_null(&_singleton, s)) { 1190 delete s; 1191 } 1192 } 1193 return _singleton; 1194 } 1195 oop lookup_raw(intptr_t raw_value, bool& cache_in_error) { 1196 if (_true_cache == nullptr) { 1197 cache_in_error = true; 1198 return nullptr; 1199 } 1200 // Have to cast to avoid little/big-endian problems. 1201 jboolean value = (jboolean)*((jint*)&raw_value); 1202 return lookup(value); 1203 } 1204 oop lookup(jboolean value) { 1205 if (value != 0) { 1206 return JNIHandles::resolve_non_null(_true_cache); 1207 } 1208 return JNIHandles::resolve_non_null(_false_cache); 1209 } 1210 }; 1211 1212 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr; 1213 1214 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) { 1215 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1216 BasicType box_type = vmClasses::box_klass_type(k); 1217 if (box_type != T_OBJECT) { 1218 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1219 switch(box_type) { 1220 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1221 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1222 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1223 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1224 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1225 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int(), cache_init_error); 1226 default:; 1227 } 1228 } 1229 return nullptr; 1230 } 1231 #endif // INCLUDE_JVMCI 1232 1233 #if COMPILER2_OR_JVMCI 1234 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1235 Handle pending_exception(THREAD, thread->pending_exception()); 1236 const char* exception_file = thread->exception_file(); 1237 int exception_line = thread->exception_line(); 1238 thread->clear_pending_exception(); 1239 1240 bool failures = false; 1241 1242 for (int i = 0; i < objects->length(); i++) { 1243 assert(objects->at(i)->is_object(), "invalid debug information"); 1244 ObjectValue* sv = (ObjectValue*) objects->at(i); 1245 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1246 1247 // Check if the object may be null and has an additional is_init input that needs 1248 // to be checked before using the field values. Skip re-allocation if it is null. 1249 if (sv->maybe_null()) { 1250 assert(k->is_inline_klass(), "must be an inline klass"); 1251 intptr_t init_value = StackValue::create_stack_value(fr, reg_map, sv->is_init())->get_int(); 1252 jint is_init = (jint)*((jint*)&init_value); 1253 if (is_init == 0) { 1254 continue; 1255 } 1256 } 1257 1258 oop obj = nullptr; 1259 bool cache_init_error = false; 1260 if (k->is_instance_klass()) { 1261 #if INCLUDE_JVMCI 1262 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 1263 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1264 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1265 obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD); 1266 if (obj != nullptr) { 1267 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1268 abv->set_cached(true); 1269 } else if (cache_init_error) { 1270 // Results in an OOME which is valid (as opposed to a class initialization error) 1271 // and is fine for the rare case a cache initialization failing. 1272 failures = true; 1273 } 1274 } 1275 #endif // INCLUDE_JVMCI 1276 1277 InstanceKlass* ik = InstanceKlass::cast(k); 1278 if (obj == nullptr && !cache_init_error) { 1279 #ifdef COMPILER2 1280 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1281 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1282 } else { 1283 obj = ik->allocate_instance(THREAD); 1284 } 1285 #else 1286 obj = ik->allocate_instance(THREAD); 1287 #endif // COMPILER2 1288 } 1289 } else if (k->is_flatArray_klass()) { 1290 FlatArrayKlass* ak = FlatArrayKlass::cast(k); 1291 // Inline type array must be zeroed because not all memory is reassigned 1292 obj = ak->allocate(sv->field_size(), THREAD); 1293 } else if (k->is_typeArray_klass()) { 1294 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1295 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1296 int len = sv->field_size() / type2size[ak->element_type()]; 1297 obj = ak->allocate(len, THREAD); 1298 } else if (k->is_objArray_klass()) { 1299 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1300 obj = ak->allocate(sv->field_size(), THREAD); 1301 } 1302 1303 if (obj == nullptr) { 1304 failures = true; 1305 } 1306 1307 assert(sv->value().is_null(), "redundant reallocation"); 1308 assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception"); 1309 CLEAR_PENDING_EXCEPTION; 1310 sv->set_value(obj); 1311 } 1312 1313 if (failures) { 1314 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1315 } else if (pending_exception.not_null()) { 1316 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1317 } 1318 1319 return failures; 1320 } 1321 1322 // We're deoptimizing at the return of a call, inline type fields are 1323 // in registers. When we go back to the interpreter, it will expect a 1324 // reference to an inline type instance. Allocate and initialize it from 1325 // the register values here. 1326 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 1327 oop new_vt = vk->realloc_result(map, return_oops, THREAD); 1328 if (new_vt == nullptr) { 1329 CLEAR_PENDING_EXCEPTION; 1330 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 1331 } 1332 return_oops.clear(); 1333 return_oops.push(Handle(THREAD, new_vt)); 1334 return false; 1335 } 1336 1337 #if INCLUDE_JVMCI 1338 /** 1339 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1340 * we need to somehow be able to recover the actual kind to be able to write the correct 1341 * amount of bytes. 1342 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1343 * the entries at index n + 1 to n + i are 'markers'. 1344 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1345 * expected form of the array would be: 1346 * 1347 * {b0, b1, b2, b3, INT, marker, b6, b7} 1348 * 1349 * Thus, in order to get back the size of the entry, we simply need to count the number 1350 * of marked entries 1351 * 1352 * @param virtualArray the virtualized byte array 1353 * @param i index of the virtual entry we are recovering 1354 * @return The number of bytes the entry spans 1355 */ 1356 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1357 int index = i; 1358 while (++index < virtualArray->field_size() && 1359 virtualArray->field_at(index)->is_marker()) {} 1360 return index - i; 1361 } 1362 1363 /** 1364 * If there was a guarantee for byte array to always start aligned to a long, we could 1365 * do a simple check on the parity of the index. Unfortunately, that is not always the 1366 * case. Thus, we check alignment of the actual address we are writing to. 1367 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1368 * write a long to index 3. 1369 */ 1370 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1371 jbyte* res = obj->byte_at_addr(index); 1372 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1373 return res; 1374 } 1375 1376 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) { 1377 switch (byte_count) { 1378 case 1: 1379 obj->byte_at_put(index, (jbyte) *((jint *) &val)); 1380 break; 1381 case 2: 1382 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val); 1383 break; 1384 case 4: 1385 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val); 1386 break; 1387 case 8: 1388 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val); 1389 break; 1390 default: 1391 ShouldNotReachHere(); 1392 } 1393 } 1394 #endif // INCLUDE_JVMCI 1395 1396 1397 // restore elements of an eliminated type array 1398 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1399 int index = 0; 1400 intptr_t val; 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_int(); 1411 #else 1412 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1413 #endif 1414 obj->long_at_put(index, res); 1415 break; 1416 } 1417 1418 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1419 case T_INT: case T_FLOAT: { // 4 bytes. 1420 assert(value->type() == T_INT, "Agreement."); 1421 bool big_value = false; 1422 if (i + 1 < sv->field_size() && type == T_INT) { 1423 if (sv->field_at(i)->is_location()) { 1424 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1425 if (type == Location::dbl || type == Location::lng) { 1426 big_value = true; 1427 } 1428 } else if (sv->field_at(i)->is_constant_int()) { 1429 ScopeValue* next_scope_field = sv->field_at(i + 1); 1430 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1431 big_value = true; 1432 } 1433 } 1434 } 1435 1436 if (big_value) { 1437 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1438 #ifdef _LP64 1439 jlong res = (jlong)low->get_int(); 1440 #else 1441 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1442 #endif 1443 obj->int_at_put(index, (jint)*((jint*)&res)); 1444 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1445 } else { 1446 val = value->get_int(); 1447 obj->int_at_put(index, (jint)*((jint*)&val)); 1448 } 1449 break; 1450 } 1451 1452 case T_SHORT: 1453 assert(value->type() == T_INT, "Agreement."); 1454 val = value->get_int(); 1455 obj->short_at_put(index, (jshort)*((jint*)&val)); 1456 break; 1457 1458 case T_CHAR: 1459 assert(value->type() == T_INT, "Agreement."); 1460 val = value->get_int(); 1461 obj->char_at_put(index, (jchar)*((jint*)&val)); 1462 break; 1463 1464 case T_BYTE: { 1465 assert(value->type() == T_INT, "Agreement."); 1466 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1467 val = value->get_int(); 1468 #if INCLUDE_JVMCI 1469 int byte_count = count_number_of_bytes_for_entry(sv, i); 1470 byte_array_put(obj, val, index, byte_count); 1471 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1472 i += byte_count - 1; // Balance the loop counter. 1473 index += byte_count; 1474 // index has been updated so continue at top of loop 1475 continue; 1476 #else 1477 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1478 break; 1479 #endif // INCLUDE_JVMCI 1480 } 1481 1482 case T_BOOLEAN: { 1483 assert(value->type() == T_INT, "Agreement."); 1484 val = value->get_int(); 1485 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1486 break; 1487 } 1488 1489 default: 1490 ShouldNotReachHere(); 1491 } 1492 index++; 1493 } 1494 } 1495 1496 // restore fields of an eliminated object array 1497 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1498 for (int i = 0; i < sv->field_size(); i++) { 1499 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1500 assert(value->type() == T_OBJECT, "object element expected"); 1501 obj->obj_at_put(i, value->get_obj()()); 1502 } 1503 } 1504 1505 class ReassignedField { 1506 public: 1507 int _offset; 1508 BasicType _type; 1509 InstanceKlass* _klass; 1510 bool _is_flat; 1511 public: 1512 ReassignedField() : _offset(0), _type(T_ILLEGAL), _klass(nullptr), _is_flat(false) { } 1513 }; 1514 1515 int compare(ReassignedField* left, ReassignedField* right) { 1516 return left->_offset - right->_offset; 1517 } 1518 1519 // Restore fields of an eliminated instance object using the same field order 1520 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1521 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) { 1522 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1523 InstanceKlass* ik = klass; 1524 while (ik != nullptr) { 1525 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1526 if (!fs.access_flags().is_static() && (!skip_internal || !fs.field_flags().is_injected())) { 1527 ReassignedField field; 1528 field._offset = fs.offset(); 1529 field._type = Signature::basic_type(fs.signature()); 1530 if (fs.is_null_free_inline_type()) { 1531 if (fs.is_flat()) { 1532 field._is_flat = true; 1533 // Resolve klass of flat inline type field 1534 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index())); 1535 } else { 1536 field._type = T_OBJECT; // Can be removed once Q-descriptors have been removed. 1537 } 1538 } 1539 fields->append(field); 1540 } 1541 } 1542 ik = ik->superklass(); 1543 } 1544 fields->sort(compare); 1545 for (int i = 0; i < fields->length(); i++) { 1546 BasicType type = fields->at(i)._type; 1547 int offset = base_offset + fields->at(i)._offset; 1548 // Check for flat inline type field before accessing the ScopeValue because it might not have any fields 1549 if (fields->at(i)._is_flat) { 1550 // Recursively re-assign flat inline type fields 1551 InstanceKlass* vk = fields->at(i)._klass; 1552 assert(vk != nullptr, "must be resolved"); 1553 offset -= InlineKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1554 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1555 continue; // Continue because we don't need to increment svIndex 1556 } 1557 intptr_t val; 1558 ScopeValue* scope_field = sv->field_at(svIndex); 1559 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1560 switch (type) { 1561 case T_OBJECT: 1562 case T_ARRAY: 1563 assert(value->type() == T_OBJECT, "Agreement."); 1564 obj->obj_field_put(offset, value->get_obj()()); 1565 break; 1566 1567 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1568 case T_INT: case T_FLOAT: { // 4 bytes. 1569 assert(value->type() == T_INT, "Agreement."); 1570 bool big_value = false; 1571 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1572 if (scope_field->is_location()) { 1573 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1574 if (type == Location::dbl || type == Location::lng) { 1575 big_value = true; 1576 } 1577 } 1578 if (scope_field->is_constant_int()) { 1579 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1580 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1581 big_value = true; 1582 } 1583 } 1584 } 1585 1586 if (big_value) { 1587 i++; 1588 assert(i < fields->length(), "second T_INT field needed"); 1589 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1590 } else { 1591 val = value->get_int(); 1592 obj->int_field_put(offset, (jint)*((jint*)&val)); 1593 break; 1594 } 1595 } 1596 /* no break */ 1597 1598 case T_LONG: case T_DOUBLE: { 1599 assert(value->type() == T_INT, "Agreement."); 1600 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1601 #ifdef _LP64 1602 jlong res = (jlong)low->get_int(); 1603 #else 1604 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1605 #endif 1606 obj->long_field_put(offset, res); 1607 break; 1608 } 1609 1610 case T_SHORT: 1611 assert(value->type() == T_INT, "Agreement."); 1612 val = value->get_int(); 1613 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1614 break; 1615 1616 case T_CHAR: 1617 assert(value->type() == T_INT, "Agreement."); 1618 val = value->get_int(); 1619 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1620 break; 1621 1622 case T_BYTE: 1623 assert(value->type() == T_INT, "Agreement."); 1624 val = value->get_int(); 1625 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1626 break; 1627 1628 case T_BOOLEAN: 1629 assert(value->type() == T_INT, "Agreement."); 1630 val = value->get_int(); 1631 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1632 break; 1633 1634 default: 1635 ShouldNotReachHere(); 1636 } 1637 svIndex++; 1638 } 1639 return svIndex; 1640 } 1641 1642 // restore fields of an eliminated inline type array 1643 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool skip_internal, TRAPS) { 1644 InlineKlass* vk = vak->element_klass(); 1645 assert(vk->flat_array(), "should only be used for flat inline type arrays"); 1646 // Adjust offset to omit oop header 1647 int base_offset = arrayOopDesc::base_offset_in_bytes(T_PRIMITIVE_OBJECT) - InlineKlass::cast(vk)->first_field_offset(); 1648 // Initialize all elements of the flat inline type array 1649 for (int i = 0; i < sv->field_size(); i++) { 1650 ScopeValue* val = sv->field_at(i); 1651 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1652 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, skip_internal, offset, CHECK); 1653 } 1654 } 1655 1656 // restore fields of all eliminated objects and arrays 1657 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1658 for (int i = 0; i < objects->length(); i++) { 1659 assert(objects->at(i)->is_object(), "invalid debug information"); 1660 ObjectValue* sv = (ObjectValue*) objects->at(i); 1661 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1662 Handle obj = sv->value(); 1663 assert(obj.not_null() || realloc_failures || sv->maybe_null(), "reallocation was missed"); 1664 #ifndef PRODUCT 1665 if (PrintDeoptimizationDetails) { 1666 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1667 } 1668 #endif // !PRODUCT 1669 1670 if (obj.is_null()) { 1671 continue; 1672 } 1673 1674 #if INCLUDE_JVMCI 1675 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1676 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1677 continue; 1678 } 1679 #endif // INCLUDE_JVMCI 1680 #ifdef COMPILER2 1681 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1682 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1683 ScopeValue* payload = sv->field_at(0); 1684 if (payload->is_location() && 1685 payload->as_LocationValue()->location().type() == Location::vector) { 1686 #ifndef PRODUCT 1687 if (PrintDeoptimizationDetails) { 1688 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1689 if (Verbose) { 1690 Handle obj = sv->value(); 1691 k->oop_print_on(obj(), tty); 1692 } 1693 } 1694 #endif // !PRODUCT 1695 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1696 } 1697 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1698 // which could be restored after vector object allocation. 1699 } 1700 #endif /* !COMPILER2 */ 1701 if (k->is_instance_klass()) { 1702 InstanceKlass* ik = InstanceKlass::cast(k); 1703 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1704 } else if (k->is_flatArray_klass()) { 1705 FlatArrayKlass* vak = FlatArrayKlass::cast(k); 1706 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, skip_internal, CHECK); 1707 } else if (k->is_typeArray_klass()) { 1708 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1709 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1710 } else if (k->is_objArray_klass()) { 1711 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1712 } 1713 } 1714 } 1715 1716 1717 // relock objects for which synchronization was eliminated 1718 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1719 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1720 bool relocked_objects = false; 1721 for (int i = 0; i < monitors->length(); i++) { 1722 MonitorInfo* mon_info = monitors->at(i); 1723 if (mon_info->eliminated()) { 1724 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1725 relocked_objects = true; 1726 if (!mon_info->owner_is_scalar_replaced()) { 1727 Handle obj(thread, mon_info->owner()); 1728 markWord mark = obj->mark(); 1729 if (exec_mode == Unpack_none) { 1730 if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) { 1731 // With exec_mode == Unpack_none obj may be thread local and locked in 1732 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header. 1733 markWord dmw = mark.displaced_mark_helper(); 1734 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr)); 1735 obj->set_mark(dmw); 1736 } 1737 if (mark.has_monitor()) { 1738 // defer relocking if the deoptee thread is currently waiting for obj 1739 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1740 if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) { 1741 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1742 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1743 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1744 continue; 1745 } 1746 } 1747 } 1748 BasicLock* lock = mon_info->lock(); 1749 ObjectSynchronizer::enter(obj, lock, deoptee_thread); 1750 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1751 } 1752 } 1753 } 1754 return relocked_objects; 1755 } 1756 #endif // COMPILER2_OR_JVMCI 1757 1758 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1759 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1760 1761 // Register map for next frame (used for stack crawl). We capture 1762 // the state of the deopt'ing frame's caller. Thus if we need to 1763 // stuff a C2I adapter we can properly fill in the callee-save 1764 // register locations. 1765 frame caller = fr.sender(reg_map); 1766 int frame_size = caller.sp() - fr.sp(); 1767 1768 frame sender = caller; 1769 1770 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1771 // the vframeArray containing the unpacking information is allocated in the C heap. 1772 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1773 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1774 1775 // Compare the vframeArray to the collected vframes 1776 assert(array->structural_compare(thread, chunk), "just checking"); 1777 1778 if (TraceDeoptimization) { 1779 ResourceMark rm; 1780 stringStream st; 1781 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1782 st.print(" "); 1783 fr.print_on(&st); 1784 st.print_cr(" Virtual frames (innermost/newest first):"); 1785 for (int index = 0; index < chunk->length(); index++) { 1786 compiledVFrame* vf = chunk->at(index); 1787 int bci = vf->raw_bci(); 1788 const char* code_name; 1789 if (bci == SynchronizationEntryBCI) { 1790 code_name = "sync entry"; 1791 } else { 1792 Bytecodes::Code code = vf->method()->code_at(bci); 1793 code_name = Bytecodes::name(code); 1794 } 1795 1796 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1797 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1798 st.print(" - %s", code_name); 1799 st.print_cr(" @ bci=%d ", bci); 1800 } 1801 tty->print_raw(st.freeze()); 1802 tty->cr(); 1803 } 1804 1805 return array; 1806 } 1807 1808 #if COMPILER2_OR_JVMCI 1809 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1810 // Reallocation of some scalar replaced objects failed. Record 1811 // that we need to pop all the interpreter frames for the 1812 // deoptimized compiled frame. 1813 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1814 thread->set_frames_to_pop_failed_realloc(array->frames()); 1815 // Unlock all monitors here otherwise the interpreter will see a 1816 // mix of locked and unlocked monitors (because of failed 1817 // reallocations of synchronized objects) and be confused. 1818 for (int i = 0; i < array->frames(); i++) { 1819 MonitorChunk* monitors = array->element(i)->monitors(); 1820 if (monitors != nullptr) { 1821 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1822 BasicObjectLock* src = monitors->at(j); 1823 if (src->obj() != nullptr) { 1824 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1825 } 1826 } 1827 array->element(i)->free_monitors(thread); 1828 #ifdef ASSERT 1829 array->element(i)->set_removed_monitors(); 1830 #endif 1831 } 1832 } 1833 } 1834 #endif 1835 1836 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1837 assert(fr.can_be_deoptimized(), "checking frame type"); 1838 1839 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1840 1841 if (LogCompilation && xtty != nullptr) { 1842 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1843 assert(cm != nullptr, "only compiled methods can deopt"); 1844 1845 ttyLocker ttyl; 1846 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1847 cm->log_identity(xtty); 1848 xtty->end_head(); 1849 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1850 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1851 xtty->method(sd->method()); 1852 xtty->end_elem(); 1853 if (sd->is_top()) break; 1854 } 1855 xtty->tail("deoptimized"); 1856 } 1857 1858 Continuation::notify_deopt(thread, fr.sp()); 1859 1860 // Patch the compiled method so that when execution returns to it we will 1861 // deopt the execution state and return to the interpreter. 1862 fr.deoptimize(thread); 1863 } 1864 1865 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1866 // Deoptimize only if the frame comes from compiled code. 1867 // Do not deoptimize the frame which is already patched 1868 // during the execution of the loops below. 1869 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1870 return; 1871 } 1872 ResourceMark rm; 1873 deoptimize_single_frame(thread, fr, reason); 1874 } 1875 1876 #if INCLUDE_JVMCI 1877 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1878 // there is no exception handler for this pc => deoptimize 1879 cm->make_not_entrant(); 1880 1881 // Use Deoptimization::deoptimize for all of its side-effects: 1882 // gathering traps statistics, logging... 1883 // it also patches the return pc but we do not care about that 1884 // since we return a continuation to the deopt_blob below. 1885 JavaThread* thread = JavaThread::current(); 1886 RegisterMap reg_map(thread, 1887 RegisterMap::UpdateMap::skip, 1888 RegisterMap::ProcessFrames::include, 1889 RegisterMap::WalkContinuation::skip); 1890 frame runtime_frame = thread->last_frame(); 1891 frame caller_frame = runtime_frame.sender(®_map); 1892 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1893 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1894 compiledVFrame* cvf = compiledVFrame::cast(vf); 1895 ScopeDesc* imm_scope = cvf->scope(); 1896 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1897 if (imm_mdo != nullptr) { 1898 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr); 1899 if (pdata != nullptr && pdata->is_BitData()) { 1900 BitData* bit_data = (BitData*) pdata; 1901 bit_data->set_exception_seen(); 1902 } 1903 } 1904 1905 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1906 1907 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1908 if (trap_mdo != nullptr) { 1909 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1910 } 1911 1912 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1913 } 1914 #endif 1915 1916 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1917 assert(thread == Thread::current() || 1918 thread->is_handshake_safe_for(Thread::current()) || 1919 SafepointSynchronize::is_at_safepoint(), 1920 "can only deoptimize other thread at a safepoint/handshake"); 1921 // Compute frame and register map based on thread and sp. 1922 RegisterMap reg_map(thread, 1923 RegisterMap::UpdateMap::skip, 1924 RegisterMap::ProcessFrames::include, 1925 RegisterMap::WalkContinuation::skip); 1926 frame fr = thread->last_frame(); 1927 while (fr.id() != id) { 1928 fr = fr.sender(®_map); 1929 } 1930 deoptimize(thread, fr, reason); 1931 } 1932 1933 1934 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1935 Thread* current = Thread::current(); 1936 if (thread == current || thread->is_handshake_safe_for(current)) { 1937 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1938 } else { 1939 VM_DeoptimizeFrame deopt(thread, id, reason); 1940 VMThread::execute(&deopt); 1941 } 1942 } 1943 1944 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1945 deoptimize_frame(thread, id, Reason_constraint); 1946 } 1947 1948 // JVMTI PopFrame support 1949 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1950 { 1951 assert(thread == JavaThread::current(), "pre-condition"); 1952 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1953 } 1954 JRT_END 1955 1956 MethodData* 1957 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1958 bool create_if_missing) { 1959 JavaThread* THREAD = thread; // For exception macros. 1960 MethodData* mdo = m()->method_data(); 1961 if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) { 1962 // Build an MDO. Ignore errors like OutOfMemory; 1963 // that simply means we won't have an MDO to update. 1964 Method::build_profiling_method_data(m, THREAD); 1965 if (HAS_PENDING_EXCEPTION) { 1966 // Only metaspace OOM is expected. No Java code executed. 1967 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1968 CLEAR_PENDING_EXCEPTION; 1969 } 1970 mdo = m()->method_data(); 1971 } 1972 return mdo; 1973 } 1974 1975 #if COMPILER2_OR_JVMCI 1976 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1977 // In case of an unresolved klass entry, load the class. 1978 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1979 // and probably nowhere else. 1980 // Even that case would benefit from simply re-interpreting the 1981 // bytecode, without paying special attention to the class index. 1982 // So this whole "class index" feature should probably be removed. 1983 1984 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1985 Klass* tk = constant_pool->klass_at(index, THREAD); 1986 if (HAS_PENDING_EXCEPTION) { 1987 // Exception happened during classloading. We ignore the exception here, since it 1988 // is going to be rethrown since the current activation is going to be deoptimized and 1989 // the interpreter will re-execute the bytecode. 1990 // Do not clear probable Async Exceptions. 1991 CLEAR_PENDING_NONASYNC_EXCEPTION; 1992 // Class loading called java code which may have caused a stack 1993 // overflow. If the exception was thrown right before the return 1994 // to the runtime the stack is no longer guarded. Reguard the 1995 // stack otherwise if we return to the uncommon trap blob and the 1996 // stack bang causes a stack overflow we crash. 1997 JavaThread* jt = THREAD; 1998 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 1999 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 2000 } 2001 return; 2002 } 2003 2004 assert(!constant_pool->tag_at(index).is_symbol(), 2005 "no symbolic names here, please"); 2006 } 2007 2008 #if INCLUDE_JFR 2009 2010 class DeoptReasonSerializer : public JfrSerializer { 2011 public: 2012 void serialize(JfrCheckpointWriter& writer) { 2013 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 2014 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 2015 writer.write_key((u8)i); 2016 writer.write(Deoptimization::trap_reason_name(i)); 2017 } 2018 } 2019 }; 2020 2021 class DeoptActionSerializer : public JfrSerializer { 2022 public: 2023 void serialize(JfrCheckpointWriter& writer) { 2024 static const u4 nof_actions = Deoptimization::Action_LIMIT; 2025 writer.write_count(nof_actions); 2026 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 2027 writer.write_key(i); 2028 writer.write(Deoptimization::trap_action_name((int)i)); 2029 } 2030 } 2031 }; 2032 2033 static void register_serializers() { 2034 static int critical_section = 0; 2035 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 2036 return; 2037 } 2038 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 2039 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 2040 } 2041 2042 static void post_deoptimization_event(CompiledMethod* nm, 2043 const Method* method, 2044 int trap_bci, 2045 int instruction, 2046 Deoptimization::DeoptReason reason, 2047 Deoptimization::DeoptAction action) { 2048 assert(nm != nullptr, "invariant"); 2049 assert(method != nullptr, "invariant"); 2050 if (EventDeoptimization::is_enabled()) { 2051 static bool serializers_registered = false; 2052 if (!serializers_registered) { 2053 register_serializers(); 2054 serializers_registered = true; 2055 } 2056 EventDeoptimization event; 2057 event.set_compileId(nm->compile_id()); 2058 event.set_compiler(nm->compiler_type()); 2059 event.set_method(method); 2060 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 2061 event.set_bci(trap_bci); 2062 event.set_instruction(instruction); 2063 event.set_reason(reason); 2064 event.set_action(action); 2065 event.commit(); 2066 } 2067 } 2068 2069 #endif // INCLUDE_JFR 2070 2071 static void log_deopt(CompiledMethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 2072 const char* reason_name, const char* reason_action) { 2073 LogTarget(Debug, deoptimization) lt; 2074 if (lt.is_enabled()) { 2075 LogStream ls(lt); 2076 bool is_osr = nm->is_osr_method(); 2077 ls.print("cid=%4d %s level=%d", 2078 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level()); 2079 ls.print(" %s", tm->name_and_sig_as_C_string()); 2080 ls.print(" trap_bci=%d ", trap_bci); 2081 if (is_osr) { 2082 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 2083 } 2084 ls.print("%s ", reason_name); 2085 ls.print("%s ", reason_action); 2086 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 2087 pc, fr.pc() - nm->code_begin()); 2088 } 2089 } 2090 2091 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 2092 HandleMark hm(current); 2093 2094 // uncommon_trap() is called at the beginning of the uncommon trap 2095 // handler. Note this fact before we start generating temporary frames 2096 // that can confuse an asynchronous stack walker. This counter is 2097 // decremented at the end of unpack_frames(). 2098 2099 current->inc_in_deopt_handler(); 2100 2101 #if INCLUDE_JVMCI 2102 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 2103 RegisterMap reg_map(current, 2104 RegisterMap::UpdateMap::include, 2105 RegisterMap::ProcessFrames::include, 2106 RegisterMap::WalkContinuation::skip); 2107 #else 2108 RegisterMap reg_map(current, 2109 RegisterMap::UpdateMap::skip, 2110 RegisterMap::ProcessFrames::include, 2111 RegisterMap::WalkContinuation::skip); 2112 #endif 2113 frame stub_frame = current->last_frame(); 2114 frame fr = stub_frame.sender(®_map); 2115 2116 // Log a message 2117 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 2118 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 2119 2120 { 2121 ResourceMark rm; 2122 2123 DeoptReason reason = trap_request_reason(trap_request); 2124 DeoptAction action = trap_request_action(trap_request); 2125 #if INCLUDE_JVMCI 2126 int debug_id = trap_request_debug_id(trap_request); 2127 #endif 2128 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 2129 2130 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 2131 compiledVFrame* cvf = compiledVFrame::cast(vf); 2132 2133 CompiledMethod* nm = cvf->code(); 2134 2135 ScopeDesc* trap_scope = cvf->scope(); 2136 2137 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2138 2139 if (is_receiver_constraint_failure) { 2140 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2141 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2142 JVMCI_ONLY(COMMA debug_id)); 2143 } 2144 2145 methodHandle trap_method(current, trap_scope->method()); 2146 int trap_bci = trap_scope->bci(); 2147 #if INCLUDE_JVMCI 2148 jlong speculation = current->pending_failed_speculation(); 2149 if (nm->is_compiled_by_jvmci()) { 2150 nm->as_nmethod()->update_speculation(current); 2151 } else { 2152 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2153 } 2154 2155 if (trap_bci == SynchronizationEntryBCI) { 2156 trap_bci = 0; 2157 current->set_pending_monitorenter(true); 2158 } 2159 2160 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2161 current->set_pending_transfer_to_interpreter(true); 2162 } 2163 #endif 2164 2165 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2166 // Record this event in the histogram. 2167 gather_statistics(reason, action, trap_bc); 2168 2169 // Ensure that we can record deopt. history: 2170 // Need MDO to record RTM code generation state. 2171 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 2172 2173 methodHandle profiled_method; 2174 #if INCLUDE_JVMCI 2175 if (nm->is_compiled_by_jvmci()) { 2176 profiled_method = methodHandle(current, nm->method()); 2177 } else { 2178 profiled_method = trap_method; 2179 } 2180 #else 2181 profiled_method = trap_method; 2182 #endif 2183 2184 MethodData* trap_mdo = 2185 get_method_data(current, profiled_method, create_if_missing); 2186 2187 { // Log Deoptimization event for JFR, UL and event system 2188 Method* tm = trap_method(); 2189 const char* reason_name = trap_reason_name(reason); 2190 const char* reason_action = trap_action_name(action); 2191 intptr_t pc = p2i(fr.pc()); 2192 2193 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2194 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2195 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 2196 reason_name, reason_action, pc, 2197 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 2198 } 2199 2200 // Print a bunch of diagnostics, if requested. 2201 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2202 ResourceMark rm; 2203 ttyLocker ttyl; 2204 char buf[100]; 2205 if (xtty != nullptr) { 2206 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 2207 os::current_thread_id(), 2208 format_trap_request(buf, sizeof(buf), trap_request)); 2209 #if INCLUDE_JVMCI 2210 if (speculation != 0) { 2211 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2212 } 2213 #endif 2214 nm->log_identity(xtty); 2215 } 2216 Symbol* class_name = nullptr; 2217 bool unresolved = false; 2218 if (unloaded_class_index >= 0) { 2219 constantPoolHandle constants (current, trap_method->constants()); 2220 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2221 class_name = constants->klass_name_at(unloaded_class_index); 2222 unresolved = true; 2223 if (xtty != nullptr) 2224 xtty->print(" unresolved='1'"); 2225 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2226 class_name = constants->symbol_at(unloaded_class_index); 2227 } 2228 if (xtty != nullptr) 2229 xtty->name(class_name); 2230 } 2231 if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) { 2232 // Dump the relevant MDO state. 2233 // This is the deopt count for the current reason, any previous 2234 // reasons or recompiles seen at this point. 2235 int dcnt = trap_mdo->trap_count(reason); 2236 if (dcnt != 0) 2237 xtty->print(" count='%d'", dcnt); 2238 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2239 int dos = (pdata == nullptr)? 0: pdata->trap_state(); 2240 if (dos != 0) { 2241 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2242 if (trap_state_is_recompiled(dos)) { 2243 int recnt2 = trap_mdo->overflow_recompile_count(); 2244 if (recnt2 != 0) 2245 xtty->print(" recompiles2='%d'", recnt2); 2246 } 2247 } 2248 } 2249 if (xtty != nullptr) { 2250 xtty->stamp(); 2251 xtty->end_head(); 2252 } 2253 if (TraceDeoptimization) { // make noise on the tty 2254 stringStream st; 2255 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2256 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2257 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2258 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 2259 #if INCLUDE_JVMCI 2260 if (nm->is_nmethod()) { 2261 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 2262 if (installed_code_name != nullptr) { 2263 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2264 } 2265 } 2266 #endif 2267 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2268 p2i(fr.pc()), 2269 os::current_thread_id(), 2270 trap_reason_name(reason), 2271 trap_action_name(action), 2272 unloaded_class_index 2273 #if INCLUDE_JVMCI 2274 , debug_id 2275 #endif 2276 ); 2277 if (class_name != nullptr) { 2278 st.print(unresolved ? " unresolved class: " : " symbol: "); 2279 class_name->print_symbol_on(&st); 2280 } 2281 st.cr(); 2282 tty->print_raw(st.freeze()); 2283 } 2284 if (xtty != nullptr) { 2285 // Log the precise location of the trap. 2286 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2287 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2288 xtty->method(sd->method()); 2289 xtty->end_elem(); 2290 if (sd->is_top()) break; 2291 } 2292 xtty->tail("uncommon_trap"); 2293 } 2294 } 2295 // (End diagnostic printout.) 2296 2297 if (is_receiver_constraint_failure) { 2298 fatal("missing receiver type check"); 2299 } 2300 2301 // Load class if necessary 2302 if (unloaded_class_index >= 0) { 2303 constantPoolHandle constants(current, trap_method->constants()); 2304 load_class_by_index(constants, unloaded_class_index, THREAD); 2305 } 2306 2307 // Flush the nmethod if necessary and desirable. 2308 // 2309 // We need to avoid situations where we are re-flushing the nmethod 2310 // because of a hot deoptimization site. Repeated flushes at the same 2311 // point need to be detected by the compiler and avoided. If the compiler 2312 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2313 // module must take measures to avoid an infinite cycle of recompilation 2314 // and deoptimization. There are several such measures: 2315 // 2316 // 1. If a recompilation is ordered a second time at some site X 2317 // and for the same reason R, the action is adjusted to 'reinterpret', 2318 // to give the interpreter time to exercise the method more thoroughly. 2319 // If this happens, the method's overflow_recompile_count is incremented. 2320 // 2321 // 2. If the compiler fails to reduce the deoptimization rate, then 2322 // the method's overflow_recompile_count will begin to exceed the set 2323 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2324 // is adjusted to 'make_not_compilable', and the method is abandoned 2325 // to the interpreter. This is a performance hit for hot methods, 2326 // but is better than a disastrous infinite cycle of recompilations. 2327 // (Actually, only the method containing the site X is abandoned.) 2328 // 2329 // 3. In parallel with the previous measures, if the total number of 2330 // recompilations of a method exceeds the much larger set limit 2331 // PerMethodRecompilationCutoff, the method is abandoned. 2332 // This should only happen if the method is very large and has 2333 // many "lukewarm" deoptimizations. The code which enforces this 2334 // limit is elsewhere (class nmethod, class Method). 2335 // 2336 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2337 // to recompile at each bytecode independently of the per-BCI cutoff. 2338 // 2339 // The decision to update code is up to the compiler, and is encoded 2340 // in the Action_xxx code. If the compiler requests Action_none 2341 // no trap state is changed, no compiled code is changed, and the 2342 // computation suffers along in the interpreter. 2343 // 2344 // The other action codes specify various tactics for decompilation 2345 // and recompilation. Action_maybe_recompile is the loosest, and 2346 // allows the compiled code to stay around until enough traps are seen, 2347 // and until the compiler gets around to recompiling the trapping method. 2348 // 2349 // The other actions cause immediate removal of the present code. 2350 2351 // Traps caused by injected profile shouldn't pollute trap counts. 2352 bool injected_profile_trap = trap_method->has_injected_profile() && 2353 (reason == Reason_intrinsic || reason == Reason_unreached); 2354 2355 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2356 bool make_not_entrant = false; 2357 bool make_not_compilable = false; 2358 bool reprofile = false; 2359 switch (action) { 2360 case Action_none: 2361 // Keep the old code. 2362 update_trap_state = false; 2363 break; 2364 case Action_maybe_recompile: 2365 // Do not need to invalidate the present code, but we can 2366 // initiate another 2367 // Start compiler without (necessarily) invalidating the nmethod. 2368 // The system will tolerate the old code, but new code should be 2369 // generated when possible. 2370 break; 2371 case Action_reinterpret: 2372 // Go back into the interpreter for a while, and then consider 2373 // recompiling form scratch. 2374 make_not_entrant = true; 2375 // Reset invocation counter for outer most method. 2376 // This will allow the interpreter to exercise the bytecodes 2377 // for a while before recompiling. 2378 // By contrast, Action_make_not_entrant is immediate. 2379 // 2380 // Note that the compiler will track null_check, null_assert, 2381 // range_check, and class_check events and log them as if they 2382 // had been traps taken from compiled code. This will update 2383 // the MDO trap history so that the next compilation will 2384 // properly detect hot trap sites. 2385 reprofile = true; 2386 break; 2387 case Action_make_not_entrant: 2388 // Request immediate recompilation, and get rid of the old code. 2389 // Make them not entrant, so next time they are called they get 2390 // recompiled. Unloaded classes are loaded now so recompile before next 2391 // time they are called. Same for uninitialized. The interpreter will 2392 // link the missing class, if any. 2393 make_not_entrant = true; 2394 break; 2395 case Action_make_not_compilable: 2396 // Give up on compiling this method at all. 2397 make_not_entrant = true; 2398 make_not_compilable = true; 2399 break; 2400 default: 2401 ShouldNotReachHere(); 2402 } 2403 2404 // Setting +ProfileTraps fixes the following, on all platforms: 2405 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2406 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2407 // recompile relies on a MethodData* to record heroic opt failures. 2408 2409 // Whether the interpreter is producing MDO data or not, we also need 2410 // to use the MDO to detect hot deoptimization points and control 2411 // aggressive optimization. 2412 bool inc_recompile_count = false; 2413 ProfileData* pdata = nullptr; 2414 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) { 2415 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2416 uint this_trap_count = 0; 2417 bool maybe_prior_trap = false; 2418 bool maybe_prior_recompile = false; 2419 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2420 #if INCLUDE_JVMCI 2421 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2422 #endif 2423 nm->method(), 2424 //outputs: 2425 this_trap_count, 2426 maybe_prior_trap, 2427 maybe_prior_recompile); 2428 // Because the interpreter also counts null, div0, range, and class 2429 // checks, these traps from compiled code are double-counted. 2430 // This is harmless; it just means that the PerXTrapLimit values 2431 // are in effect a little smaller than they look. 2432 2433 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2434 if (per_bc_reason != Reason_none) { 2435 // Now take action based on the partially known per-BCI history. 2436 if (maybe_prior_trap 2437 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2438 // If there are too many traps at this BCI, force a recompile. 2439 // This will allow the compiler to see the limit overflow, and 2440 // take corrective action, if possible. The compiler generally 2441 // does not use the exact PerBytecodeTrapLimit value, but instead 2442 // changes its tactics if it sees any traps at all. This provides 2443 // a little hysteresis, delaying a recompile until a trap happens 2444 // several times. 2445 // 2446 // Actually, since there is only one bit of counter per BCI, 2447 // the possible per-BCI counts are {0,1,(per-method count)}. 2448 // This produces accurate results if in fact there is only 2449 // one hot trap site, but begins to get fuzzy if there are 2450 // many sites. For example, if there are ten sites each 2451 // trapping two or more times, they each get the blame for 2452 // all of their traps. 2453 make_not_entrant = true; 2454 } 2455 2456 // Detect repeated recompilation at the same BCI, and enforce a limit. 2457 if (make_not_entrant && maybe_prior_recompile) { 2458 // More than one recompile at this point. 2459 inc_recompile_count = maybe_prior_trap; 2460 } 2461 } else { 2462 // For reasons which are not recorded per-bytecode, we simply 2463 // force recompiles unconditionally. 2464 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2465 make_not_entrant = true; 2466 } 2467 2468 // Go back to the compiler if there are too many traps in this method. 2469 if (this_trap_count >= per_method_trap_limit(reason)) { 2470 // If there are too many traps in this method, force a recompile. 2471 // This will allow the compiler to see the limit overflow, and 2472 // take corrective action, if possible. 2473 // (This condition is an unlikely backstop only, because the 2474 // PerBytecodeTrapLimit is more likely to take effect first, 2475 // if it is applicable.) 2476 make_not_entrant = true; 2477 } 2478 2479 // Here's more hysteresis: If there has been a recompile at 2480 // this trap point already, run the method in the interpreter 2481 // for a while to exercise it more thoroughly. 2482 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2483 reprofile = true; 2484 } 2485 } 2486 2487 // Take requested actions on the method: 2488 2489 // Recompile 2490 if (make_not_entrant) { 2491 if (!nm->make_not_entrant()) { 2492 return; // the call did not change nmethod's state 2493 } 2494 2495 if (pdata != nullptr) { 2496 // Record the recompilation event, if any. 2497 int tstate0 = pdata->trap_state(); 2498 int tstate1 = trap_state_set_recompiled(tstate0, true); 2499 if (tstate1 != tstate0) 2500 pdata->set_trap_state(tstate1); 2501 } 2502 2503 #if INCLUDE_RTM_OPT 2504 // Restart collecting RTM locking abort statistic if the method 2505 // is recompiled for a reason other than RTM state change. 2506 // Assume that in new recompiled code the statistic could be different, 2507 // for example, due to different inlining. 2508 if ((reason != Reason_rtm_state_change) && (trap_mdo != nullptr) && 2509 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2510 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2511 } 2512 #endif 2513 // For code aging we count traps separately here, using make_not_entrant() 2514 // as a guard against simultaneous deopts in multiple threads. 2515 if (reason == Reason_tenured && trap_mdo != nullptr) { 2516 trap_mdo->inc_tenure_traps(); 2517 } 2518 } 2519 2520 if (inc_recompile_count) { 2521 trap_mdo->inc_overflow_recompile_count(); 2522 if ((uint)trap_mdo->overflow_recompile_count() > 2523 (uint)PerBytecodeRecompilationCutoff) { 2524 // Give up on the method containing the bad BCI. 2525 if (trap_method() == nm->method()) { 2526 make_not_compilable = true; 2527 } else { 2528 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2529 // But give grace to the enclosing nm->method(). 2530 } 2531 } 2532 } 2533 2534 // Reprofile 2535 if (reprofile) { 2536 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2537 } 2538 2539 // Give up compiling 2540 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2541 assert(make_not_entrant, "consistent"); 2542 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2543 } 2544 2545 } // Free marked resources 2546 2547 } 2548 JRT_END 2549 2550 ProfileData* 2551 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2552 int trap_bci, 2553 Deoptimization::DeoptReason reason, 2554 bool update_total_trap_count, 2555 #if INCLUDE_JVMCI 2556 bool is_osr, 2557 #endif 2558 Method* compiled_method, 2559 //outputs: 2560 uint& ret_this_trap_count, 2561 bool& ret_maybe_prior_trap, 2562 bool& ret_maybe_prior_recompile) { 2563 bool maybe_prior_trap = false; 2564 bool maybe_prior_recompile = false; 2565 uint this_trap_count = 0; 2566 if (update_total_trap_count) { 2567 uint idx = reason; 2568 #if INCLUDE_JVMCI 2569 if (is_osr) { 2570 // Upper half of history array used for traps in OSR compilations 2571 idx += Reason_TRAP_HISTORY_LENGTH; 2572 } 2573 #endif 2574 uint prior_trap_count = trap_mdo->trap_count(idx); 2575 this_trap_count = trap_mdo->inc_trap_count(idx); 2576 2577 // If the runtime cannot find a place to store trap history, 2578 // it is estimated based on the general condition of the method. 2579 // If the method has ever been recompiled, or has ever incurred 2580 // a trap with the present reason , then this BCI is assumed 2581 // (pessimistically) to be the culprit. 2582 maybe_prior_trap = (prior_trap_count != 0); 2583 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2584 } 2585 ProfileData* pdata = nullptr; 2586 2587 2588 // For reasons which are recorded per bytecode, we check per-BCI data. 2589 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2590 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2591 if (per_bc_reason != Reason_none) { 2592 // Find the profile data for this BCI. If there isn't one, 2593 // try to allocate one from the MDO's set of spares. 2594 // This will let us detect a repeated trap at this point. 2595 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr); 2596 2597 if (pdata != nullptr) { 2598 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2599 if (LogCompilation && xtty != nullptr) { 2600 ttyLocker ttyl; 2601 // no more room for speculative traps in this MDO 2602 xtty->elem("speculative_traps_oom"); 2603 } 2604 } 2605 // Query the trap state of this profile datum. 2606 int tstate0 = pdata->trap_state(); 2607 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2608 maybe_prior_trap = false; 2609 if (!trap_state_is_recompiled(tstate0)) 2610 maybe_prior_recompile = false; 2611 2612 // Update the trap state of this profile datum. 2613 int tstate1 = tstate0; 2614 // Record the reason. 2615 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2616 // Store the updated state on the MDO, for next time. 2617 if (tstate1 != tstate0) 2618 pdata->set_trap_state(tstate1); 2619 } else { 2620 if (LogCompilation && xtty != nullptr) { 2621 ttyLocker ttyl; 2622 // Missing MDP? Leave a small complaint in the log. 2623 xtty->elem("missing_mdp bci='%d'", trap_bci); 2624 } 2625 } 2626 } 2627 2628 // Return results: 2629 ret_this_trap_count = this_trap_count; 2630 ret_maybe_prior_trap = maybe_prior_trap; 2631 ret_maybe_prior_recompile = maybe_prior_recompile; 2632 return pdata; 2633 } 2634 2635 void 2636 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2637 ResourceMark rm; 2638 // Ignored outputs: 2639 uint ignore_this_trap_count; 2640 bool ignore_maybe_prior_trap; 2641 bool ignore_maybe_prior_recompile; 2642 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2643 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2644 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2645 query_update_method_data(trap_mdo, trap_bci, 2646 (DeoptReason)reason, 2647 update_total_counts, 2648 #if INCLUDE_JVMCI 2649 false, 2650 #endif 2651 nullptr, 2652 ignore_this_trap_count, 2653 ignore_maybe_prior_trap, 2654 ignore_maybe_prior_recompile); 2655 } 2656 2657 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) { 2658 // Enable WXWrite: current function is called from methods compiled by C2 directly 2659 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2660 2661 // Still in Java no safepoints 2662 { 2663 // This enters VM and may safepoint 2664 uncommon_trap_inner(current, trap_request); 2665 } 2666 HandleMark hm(current); 2667 return fetch_unroll_info_helper(current, exec_mode); 2668 } 2669 2670 // Local derived constants. 2671 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2672 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2673 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2674 2675 //---------------------------trap_state_reason--------------------------------- 2676 Deoptimization::DeoptReason 2677 Deoptimization::trap_state_reason(int trap_state) { 2678 // This assert provides the link between the width of DataLayout::trap_bits 2679 // and the encoding of "recorded" reasons. It ensures there are enough 2680 // bits to store all needed reasons in the per-BCI MDO profile. 2681 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2682 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2683 trap_state -= recompile_bit; 2684 if (trap_state == DS_REASON_MASK) { 2685 return Reason_many; 2686 } else { 2687 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2688 return (DeoptReason)trap_state; 2689 } 2690 } 2691 //-------------------------trap_state_has_reason------------------------------- 2692 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2693 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2694 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2695 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2696 trap_state -= recompile_bit; 2697 if (trap_state == DS_REASON_MASK) { 2698 return -1; // true, unspecifically (bottom of state lattice) 2699 } else if (trap_state == reason) { 2700 return 1; // true, definitely 2701 } else if (trap_state == 0) { 2702 return 0; // false, definitely (top of state lattice) 2703 } else { 2704 return 0; // false, definitely 2705 } 2706 } 2707 //-------------------------trap_state_add_reason------------------------------- 2708 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2709 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2710 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2711 trap_state -= recompile_bit; 2712 if (trap_state == DS_REASON_MASK) { 2713 return trap_state + recompile_bit; // already at state lattice bottom 2714 } else if (trap_state == reason) { 2715 return trap_state + recompile_bit; // the condition is already true 2716 } else if (trap_state == 0) { 2717 return reason + recompile_bit; // no condition has yet been true 2718 } else { 2719 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2720 } 2721 } 2722 //-----------------------trap_state_is_recompiled------------------------------ 2723 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2724 return (trap_state & DS_RECOMPILE_BIT) != 0; 2725 } 2726 //-----------------------trap_state_set_recompiled----------------------------- 2727 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2728 if (z) return trap_state | DS_RECOMPILE_BIT; 2729 else return trap_state & ~DS_RECOMPILE_BIT; 2730 } 2731 //---------------------------format_trap_state--------------------------------- 2732 // This is used for debugging and diagnostics, including LogFile output. 2733 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2734 int trap_state) { 2735 assert(buflen > 0, "sanity"); 2736 DeoptReason reason = trap_state_reason(trap_state); 2737 bool recomp_flag = trap_state_is_recompiled(trap_state); 2738 // Re-encode the state from its decoded components. 2739 int decoded_state = 0; 2740 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2741 decoded_state = trap_state_add_reason(decoded_state, reason); 2742 if (recomp_flag) 2743 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2744 // If the state re-encodes properly, format it symbolically. 2745 // Because this routine is used for debugging and diagnostics, 2746 // be robust even if the state is a strange value. 2747 size_t len; 2748 if (decoded_state != trap_state) { 2749 // Random buggy state that doesn't decode?? 2750 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2751 } else { 2752 len = jio_snprintf(buf, buflen, "%s%s", 2753 trap_reason_name(reason), 2754 recomp_flag ? " recompiled" : ""); 2755 } 2756 return buf; 2757 } 2758 2759 2760 //--------------------------------statics-------------------------------------- 2761 const char* Deoptimization::_trap_reason_name[] = { 2762 // Note: Keep this in sync. with enum DeoptReason. 2763 "none", 2764 "null_check", 2765 "null_assert" JVMCI_ONLY("_or_unreached0"), 2766 "range_check", 2767 "class_check", 2768 "array_check", 2769 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2770 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2771 "profile_predicate", 2772 "unloaded", 2773 "uninitialized", 2774 "initialized", 2775 "unreached", 2776 "unhandled", 2777 "constraint", 2778 "div0_check", 2779 "age", 2780 "predicate", 2781 "loop_limit_check", 2782 "speculate_class_check", 2783 "speculate_null_check", 2784 "speculate_null_assert", 2785 "rtm_state_change", 2786 "unstable_if", 2787 "unstable_fused_if", 2788 "receiver_constraint", 2789 #if INCLUDE_JVMCI 2790 "aliasing", 2791 "transfer_to_interpreter", 2792 "not_compiled_exception_handler", 2793 "unresolved", 2794 "jsr_mismatch", 2795 #endif 2796 "tenured" 2797 }; 2798 const char* Deoptimization::_trap_action_name[] = { 2799 // Note: Keep this in sync. with enum DeoptAction. 2800 "none", 2801 "maybe_recompile", 2802 "reinterpret", 2803 "make_not_entrant", 2804 "make_not_compilable" 2805 }; 2806 2807 const char* Deoptimization::trap_reason_name(int reason) { 2808 // Check that every reason has a name 2809 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2810 2811 if (reason == Reason_many) return "many"; 2812 if ((uint)reason < Reason_LIMIT) 2813 return _trap_reason_name[reason]; 2814 static char buf[20]; 2815 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2816 return buf; 2817 } 2818 const char* Deoptimization::trap_action_name(int action) { 2819 // Check that every action has a name 2820 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2821 2822 if ((uint)action < Action_LIMIT) 2823 return _trap_action_name[action]; 2824 static char buf[20]; 2825 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2826 return buf; 2827 } 2828 2829 // This is used for debugging and diagnostics, including LogFile output. 2830 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2831 int trap_request) { 2832 jint unloaded_class_index = trap_request_index(trap_request); 2833 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2834 const char* action = trap_action_name(trap_request_action(trap_request)); 2835 #if INCLUDE_JVMCI 2836 int debug_id = trap_request_debug_id(trap_request); 2837 #endif 2838 size_t len; 2839 if (unloaded_class_index < 0) { 2840 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2841 reason, action 2842 #if INCLUDE_JVMCI 2843 ,debug_id 2844 #endif 2845 ); 2846 } else { 2847 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2848 reason, action, unloaded_class_index 2849 #if INCLUDE_JVMCI 2850 ,debug_id 2851 #endif 2852 ); 2853 } 2854 return buf; 2855 } 2856 2857 juint Deoptimization::_deoptimization_hist 2858 [Deoptimization::Reason_LIMIT] 2859 [1 + Deoptimization::Action_LIMIT] 2860 [Deoptimization::BC_CASE_LIMIT] 2861 = {0}; 2862 2863 enum { 2864 LSB_BITS = 8, 2865 LSB_MASK = right_n_bits(LSB_BITS) 2866 }; 2867 2868 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2869 Bytecodes::Code bc) { 2870 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2871 assert(action >= 0 && action < Action_LIMIT, "oob"); 2872 _deoptimization_hist[Reason_none][0][0] += 1; // total 2873 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2874 juint* cases = _deoptimization_hist[reason][1+action]; 2875 juint* bc_counter_addr = nullptr; 2876 juint bc_counter = 0; 2877 // Look for an unused counter, or an exact match to this BC. 2878 if (bc != Bytecodes::_illegal) { 2879 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2880 juint* counter_addr = &cases[bc_case]; 2881 juint counter = *counter_addr; 2882 if ((counter == 0 && bc_counter_addr == nullptr) 2883 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2884 // this counter is either free or is already devoted to this BC 2885 bc_counter_addr = counter_addr; 2886 bc_counter = counter | bc; 2887 } 2888 } 2889 } 2890 if (bc_counter_addr == nullptr) { 2891 // Overflow, or no given bytecode. 2892 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2893 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2894 } 2895 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2896 } 2897 2898 jint Deoptimization::total_deoptimization_count() { 2899 return _deoptimization_hist[Reason_none][0][0]; 2900 } 2901 2902 // Get the deopt count for a specific reason and a specific action. If either 2903 // one of 'reason' or 'action' is null, the method returns the sum of all 2904 // deoptimizations with the specific 'action' or 'reason' respectively. 2905 // If both arguments are null, the method returns the total deopt count. 2906 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2907 if (reason_str == nullptr && action_str == nullptr) { 2908 return total_deoptimization_count(); 2909 } 2910 juint counter = 0; 2911 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2912 if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) { 2913 for (int action = 0; action < Action_LIMIT; action++) { 2914 if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) { 2915 juint* cases = _deoptimization_hist[reason][1+action]; 2916 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2917 counter += cases[bc_case] >> LSB_BITS; 2918 } 2919 } 2920 } 2921 } 2922 } 2923 return counter; 2924 } 2925 2926 void Deoptimization::print_statistics() { 2927 juint total = total_deoptimization_count(); 2928 juint account = total; 2929 if (total != 0) { 2930 ttyLocker ttyl; 2931 if (xtty != nullptr) xtty->head("statistics type='deoptimization'"); 2932 tty->print_cr("Deoptimization traps recorded:"); 2933 #define PRINT_STAT_LINE(name, r) \ 2934 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2935 PRINT_STAT_LINE("total", total); 2936 // For each non-zero entry in the histogram, print the reason, 2937 // the action, and (if specifically known) the type of bytecode. 2938 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2939 for (int action = 0; action < Action_LIMIT; action++) { 2940 juint* cases = _deoptimization_hist[reason][1+action]; 2941 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2942 juint counter = cases[bc_case]; 2943 if (counter != 0) { 2944 char name[1*K]; 2945 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2946 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2947 bc = Bytecodes::_illegal; 2948 os::snprintf_checked(name, sizeof(name), "%s/%s/%s", 2949 trap_reason_name(reason), 2950 trap_action_name(action), 2951 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2952 juint r = counter >> LSB_BITS; 2953 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2954 account -= r; 2955 } 2956 } 2957 } 2958 } 2959 if (account != 0) { 2960 PRINT_STAT_LINE("unaccounted", account); 2961 } 2962 #undef PRINT_STAT_LINE 2963 if (xtty != nullptr) xtty->tail("statistics"); 2964 } 2965 } 2966 2967 #else // COMPILER2_OR_JVMCI 2968 2969 2970 // Stubs for C1 only system. 2971 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2972 return false; 2973 } 2974 2975 const char* Deoptimization::trap_reason_name(int reason) { 2976 return "unknown"; 2977 } 2978 2979 jint Deoptimization::total_deoptimization_count() { 2980 return 0; 2981 } 2982 2983 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2984 return 0; 2985 } 2986 2987 void Deoptimization::print_statistics() { 2988 // no output 2989 } 2990 2991 void 2992 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2993 // no update 2994 } 2995 2996 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2997 return 0; 2998 } 2999 3000 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 3001 Bytecodes::Code bc) { 3002 // no update 3003 } 3004 3005 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 3006 int trap_state) { 3007 jio_snprintf(buf, buflen, "#%d", trap_state); 3008 return buf; 3009 } 3010 3011 #endif // COMPILER2_OR_JVMCI