1 /* 2 * Copyright (c) 1997, 2022, 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/interpreter.hpp" 40 #include "interpreter/oopMapCache.hpp" 41 #include "jvm.h" 42 #include "logging/log.hpp" 43 #include "logging/logLevel.hpp" 44 #include "logging/logMessage.hpp" 45 #include "logging/logStream.hpp" 46 #include "memory/allocation.inline.hpp" 47 #include "memory/oopFactory.hpp" 48 #include "memory/resourceArea.hpp" 49 #include "memory/universe.hpp" 50 #include "oops/constantPool.hpp" 51 #include "oops/flatArrayKlass.hpp" 52 #include "oops/flatArrayOop.hpp" 53 #include "oops/fieldStreams.inline.hpp" 54 #include "oops/method.hpp" 55 #include "oops/objArrayKlass.hpp" 56 #include "oops/objArrayOop.inline.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "oops/inlineKlass.inline.hpp" 59 #include "oops/typeArrayOop.inline.hpp" 60 #include "oops/verifyOopClosure.hpp" 61 #include "prims/jvmtiDeferredUpdates.hpp" 62 #include "prims/jvmtiExport.hpp" 63 #include "prims/jvmtiThreadState.hpp" 64 #include "prims/methodHandles.hpp" 65 #include "prims/vectorSupport.hpp" 66 #include "runtime/atomic.hpp" 67 #include "runtime/continuation.hpp" 68 #include "runtime/continuationEntry.inline.hpp" 69 #include "runtime/deoptimization.hpp" 70 #include "runtime/escapeBarrier.hpp" 71 #include "runtime/fieldDescriptor.hpp" 72 #include "runtime/fieldDescriptor.inline.hpp" 73 #include "runtime/frame.inline.hpp" 74 #include "runtime/handles.inline.hpp" 75 #include "runtime/interfaceSupport.inline.hpp" 76 #include "runtime/javaThread.hpp" 77 #include "runtime/jniHandles.inline.hpp" 78 #include "runtime/keepStackGCProcessed.hpp" 79 #include "runtime/objectMonitor.inline.hpp" 80 #include "runtime/osThread.hpp" 81 #include "runtime/safepointVerifiers.hpp" 82 #include "runtime/sharedRuntime.hpp" 83 #include "runtime/signature.hpp" 84 #include "runtime/stackFrameStream.inline.hpp" 85 #include "runtime/stackValue.hpp" 86 #include "runtime/stackWatermarkSet.hpp" 87 #include "runtime/stubRoutines.hpp" 88 #include "runtime/threadSMR.hpp" 89 #include "runtime/threadWXSetters.inline.hpp" 90 #include "runtime/vframe.hpp" 91 #include "runtime/vframeArray.hpp" 92 #include "runtime/vframe_hp.hpp" 93 #include "runtime/vmOperations.hpp" 94 #include "utilities/events.hpp" 95 #include "utilities/growableArray.hpp" 96 #include "utilities/macros.hpp" 97 #include "utilities/preserveException.hpp" 98 #include "utilities/xmlstream.hpp" 99 #if INCLUDE_JFR 100 #include "jfr/jfrEvents.hpp" 101 #include "jfr/metadata/jfrSerializer.hpp" 102 #endif 103 104 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 105 int caller_adjustment, 106 int caller_actual_parameters, 107 int number_of_frames, 108 intptr_t* frame_sizes, 109 address* frame_pcs, 110 BasicType return_type, 111 int exec_mode) { 112 _size_of_deoptimized_frame = size_of_deoptimized_frame; 113 _caller_adjustment = caller_adjustment; 114 _caller_actual_parameters = caller_actual_parameters; 115 _number_of_frames = number_of_frames; 116 _frame_sizes = frame_sizes; 117 _frame_pcs = frame_pcs; 118 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 119 _return_type = return_type; 120 _initial_info = 0; 121 // PD (x86 only) 122 _counter_temp = 0; 123 _unpack_kind = exec_mode; 124 _sender_sp_temp = 0; 125 126 _total_frame_sizes = size_of_frames(); 127 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 128 } 129 130 Deoptimization::UnrollBlock::~UnrollBlock() { 131 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 132 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 133 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 134 } 135 136 int Deoptimization::UnrollBlock::size_of_frames() const { 137 // Account first for the adjustment of the initial frame 138 int result = _caller_adjustment; 139 for (int index = 0; index < number_of_frames(); index++) { 140 result += frame_sizes()[index]; 141 } 142 return result; 143 } 144 145 void Deoptimization::UnrollBlock::print() { 146 ResourceMark rm; 147 stringStream st; 148 st.print_cr("UnrollBlock"); 149 st.print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 150 st.print( " frame_sizes: "); 151 for (int index = 0; index < number_of_frames(); index++) { 152 st.print(INTX_FORMAT " ", frame_sizes()[index]); 153 } 154 st.cr(); 155 tty->print_raw(st.freeze()); 156 } 157 158 // In order to make fetch_unroll_info work properly with escape 159 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY. 160 // The actual reallocation of previously eliminated objects occurs in realloc_objects, 161 // which is called from the method fetch_unroll_info_helper below. 162 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode)) 163 // fetch_unroll_info() is called at the beginning of the deoptimization 164 // handler. Note this fact before we start generating temporary frames 165 // that can confuse an asynchronous stack walker. This counter is 166 // decremented at the end of unpack_frames(). 167 current->inc_in_deopt_handler(); 168 169 if (exec_mode == Unpack_exception) { 170 // When we get here, a callee has thrown an exception into a deoptimized 171 // frame. That throw might have deferred stack watermark checking until 172 // after unwinding. So we deal with such deferred requests here. 173 StackWatermarkSet::after_unwind(current); 174 } 175 176 return fetch_unroll_info_helper(current, exec_mode); 177 JRT_END 178 179 #if COMPILER2_OR_JVMCI 180 // print information about reallocated objects 181 static void print_objects(JavaThread* deoptee_thread, 182 GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 183 ResourceMark rm; 184 stringStream st; // change to logStream with logging 185 st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 186 fieldDescriptor fd; 187 188 for (int i = 0; i < objects->length(); i++) { 189 ObjectValue* sv = (ObjectValue*) objects->at(i); 190 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 191 Handle obj = sv->value(); 192 193 st.print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 194 k->print_value_on(&st); 195 assert(obj.not_null() || k->is_inline_klass() || realloc_failures, "reallocation was missed"); 196 if (obj.is_null()) { 197 if (k->is_inline_klass()) { 198 st.print(" is null"); 199 } else { 200 st.print(" allocation failed"); 201 } 202 } else { 203 st.print(" allocated (" SIZE_FORMAT " bytes)", obj->size() * HeapWordSize); 204 } 205 st.cr(); 206 207 if (Verbose && !obj.is_null()) { 208 k->oop_print_on(obj(), &st); 209 } 210 } 211 tty->print_raw(st.freeze()); 212 } 213 214 static bool rematerialize_objects(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 215 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk, 216 bool& deoptimized_objects) { 217 bool realloc_failures = false; 218 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 219 220 JavaThread* deoptee_thread = chunk->at(0)->thread(); 221 assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), 222 "a frame can only be deoptimized by the owner thread"); 223 224 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 225 226 // The flag return_oop() indicates call sites which return oop 227 // in compiled code. Such sites include java method calls, 228 // runtime calls (for example, used to allocate new objects/arrays 229 // on slow code path) and any other calls generated in compiled code. 230 // It is not guaranteed that we can get such information here only 231 // by analyzing bytecode in deoptimized frames. This is why this flag 232 // is set during method compilation (see Compile::Process_OopMap_Node()). 233 // If the previous frame was popped or if we are dispatching an exception, 234 // we don't have an oop result. 235 ScopeDesc* scope = chunk->at(0)->scope(); 236 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 237 // In case of the return of multiple values, we must take care 238 // of all oop return values. 239 GrowableArray<Handle> return_oops; 240 InlineKlass* vk = NULL; 241 if (save_oop_result && scope->return_scalarized()) { 242 vk = InlineKlass::returned_inline_klass(map); 243 if (vk != NULL) { 244 vk->save_oop_fields(map, return_oops); 245 save_oop_result = false; 246 } 247 } 248 if (save_oop_result) { 249 // Reallocation may trigger GC. If deoptimization happened on return from 250 // call which returns oop we need to save it since it is not in oopmap. 251 oop result = deoptee.saved_oop_result(&map); 252 assert(oopDesc::is_oop_or_null(result), "must be oop"); 253 return_oops.push(Handle(thread, result)); 254 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 255 if (TraceDeoptimization) { 256 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 257 tty->cr(); 258 } 259 } 260 if (objects != NULL || vk != NULL) { 261 if (exec_mode == Deoptimization::Unpack_none) { 262 assert(thread->thread_state() == _thread_in_vm, "assumption"); 263 JavaThread* THREAD = thread; // For exception macros. 264 // Clear pending OOM if reallocation fails and return true indicating allocation failure 265 if (vk != NULL) { 266 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, CHECK_AND_CLEAR_(true)); 267 } 268 if (objects != NULL) { 269 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true)); 270 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci(); 271 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, CHECK_AND_CLEAR_(true)); 272 } 273 deoptimized_objects = true; 274 } else { 275 JavaThread* current = thread; // For JRT_BLOCK 276 JRT_BLOCK 277 if (vk != NULL) { 278 realloc_failures = Deoptimization::realloc_inline_type_result(vk, map, return_oops, THREAD); 279 } 280 if (objects != NULL) { 281 realloc_failures = realloc_failures || Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 282 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci(); 283 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 284 } 285 JRT_END 286 } 287 if (TraceDeoptimization) { 288 print_objects(deoptee_thread, objects, realloc_failures); 289 } 290 } 291 if (save_oop_result || vk != NULL) { 292 // Restore result. 293 assert(return_oops.length() == 1, "no inline type"); 294 deoptee.set_saved_oop_result(&map, return_oops.pop()()); 295 } 296 return realloc_failures; 297 } 298 299 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures, 300 frame& deoptee, int exec_mode, bool& deoptimized_objects) { 301 JavaThread* deoptee_thread = chunk->at(0)->thread(); 302 assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 303 assert(thread == Thread::current(), "should be"); 304 HandleMark hm(thread); 305 #ifndef PRODUCT 306 bool first = true; 307 #endif // !PRODUCT 308 for (int i = 0; i < chunk->length(); i++) { 309 compiledVFrame* cvf = chunk->at(i); 310 assert (cvf->scope() != NULL,"expect only compiled java frames"); 311 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 312 if (monitors->is_nonempty()) { 313 bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, 314 exec_mode, realloc_failures); 315 deoptimized_objects = deoptimized_objects || relocked; 316 #ifndef PRODUCT 317 if (PrintDeoptimizationDetails) { 318 ResourceMark rm; 319 stringStream st; 320 for (int j = 0; j < monitors->length(); j++) { 321 MonitorInfo* mi = monitors->at(j); 322 if (mi->eliminated()) { 323 if (first) { 324 first = false; 325 st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 326 } 327 if (exec_mode == Deoptimization::Unpack_none) { 328 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 329 if (monitor != NULL && monitor->object() == mi->owner()) { 330 st.print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 331 continue; 332 } 333 } 334 if (mi->owner_is_scalar_replaced()) { 335 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 336 st.print_cr(" failed reallocation for klass %s", k->external_name()); 337 } else { 338 st.print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 339 } 340 } 341 } 342 tty->print_raw(st.freeze()); 343 } 344 #endif // !PRODUCT 345 } 346 } 347 } 348 349 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI. 350 // The given vframes cover one physical frame. 351 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, 352 bool& realloc_failures) { 353 frame deoptee = chunk->at(0)->fr(); 354 JavaThread* deoptee_thread = chunk->at(0)->thread(); 355 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 356 RegisterMap map(chunk->at(0)->register_map()); 357 bool deoptimized_objects = false; 358 359 bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false); 360 361 // Reallocate the non-escaping objects and restore their fields. 362 if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations) 363 || EliminateAutoBox || EnableVectorAggressiveReboxing)) { 364 realloc_failures = rematerialize_objects(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects); 365 } 366 367 // MonitorInfo structures used in eliminate_locks are not GC safe. 368 NoSafepointVerifier no_safepoint; 369 370 // Now relock objects if synchronization on them was eliminated. 371 if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) { 372 restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects); 373 } 374 return deoptimized_objects; 375 } 376 #endif // COMPILER2_OR_JVMCI 377 378 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 379 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) { 380 // When we get here we are about to unwind the deoptee frame. In order to 381 // catch not yet safe to use frames, the following stack watermark barrier 382 // poll will make such frames safe to use. 383 StackWatermarkSet::before_unwind(current); 384 385 // Note: there is a safepoint safety issue here. No matter whether we enter 386 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 387 // the vframeArray is created. 388 // 389 390 // Allocate our special deoptimization ResourceMark 391 DeoptResourceMark* dmark = new DeoptResourceMark(current); 392 assert(current->deopt_mark() == NULL, "Pending deopt!"); 393 current->set_deopt_mark(dmark); 394 395 frame stub_frame = current->last_frame(); // Makes stack walkable as side effect 396 RegisterMap map(current, 397 RegisterMap::UpdateMap::include, 398 RegisterMap::ProcessFrames::include, 399 RegisterMap::WalkContinuation::skip); 400 RegisterMap dummy_map(current, 401 RegisterMap::UpdateMap::skip, 402 RegisterMap::ProcessFrames::include, 403 RegisterMap::WalkContinuation::skip); 404 // Now get the deoptee with a valid map 405 frame deoptee = stub_frame.sender(&map); 406 // Set the deoptee nmethod 407 assert(current->deopt_compiled_method() == NULL, "Pending deopt!"); 408 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 409 current->set_deopt_compiled_method(cm); 410 411 if (VerifyStack) { 412 current->validate_frame_layout(); 413 } 414 415 // Create a growable array of VFrames where each VFrame represents an inlined 416 // Java frame. This storage is allocated with the usual system arena. 417 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 418 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 419 vframe* vf = vframe::new_vframe(&deoptee, &map, current); 420 while (!vf->is_top()) { 421 assert(vf->is_compiled_frame(), "Wrong frame type"); 422 chunk->push(compiledVFrame::cast(vf)); 423 vf = vf->sender(); 424 } 425 assert(vf->is_compiled_frame(), "Wrong frame type"); 426 chunk->push(compiledVFrame::cast(vf)); 427 428 bool realloc_failures = false; 429 430 #if COMPILER2_OR_JVMCI 431 bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false); 432 433 // Reallocate the non-escaping objects and restore their fields. Then 434 // relock objects if synchronization on them was eliminated. 435 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) 436 || EliminateAutoBox || EnableVectorAggressiveReboxing )) { 437 bool unused; 438 realloc_failures = rematerialize_objects(current, exec_mode, cm, deoptee, map, chunk, unused); 439 } 440 #endif // COMPILER2_OR_JVMCI 441 442 // Ensure that no safepoint is taken after pointers have been stored 443 // in fields of rematerialized objects. If a safepoint occurs from here on 444 // out the java state residing in the vframeArray will be missed. 445 // Locks may be rebaised in a safepoint. 446 NoSafepointVerifier no_safepoint; 447 448 #if COMPILER2_OR_JVMCI 449 if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) 450 && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) { 451 bool unused; 452 restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused); 453 } 454 #endif // COMPILER2_OR_JVMCI 455 456 ScopeDesc* trap_scope = chunk->at(0)->scope(); 457 Handle exceptionObject; 458 if (trap_scope->rethrow_exception()) { 459 #ifndef PRODUCT 460 if (PrintDeoptimizationDetails) { 461 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()); 462 } 463 #endif // !PRODUCT 464 465 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 466 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 467 ScopeValue* topOfStack = expressions->top(); 468 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 469 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 470 } 471 472 vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures); 473 #if COMPILER2_OR_JVMCI 474 if (realloc_failures) { 475 // This destroys all ScopedValue bindings. 476 current->clear_scopedValueBindings(); 477 pop_frames_failed_reallocs(current, array); 478 } 479 #endif 480 481 assert(current->vframe_array_head() == NULL, "Pending deopt!"); 482 current->set_vframe_array_head(array); 483 484 // Now that the vframeArray has been created if we have any deferred local writes 485 // added by jvmti then we can free up that structure as the data is now in the 486 // vframeArray 487 488 JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id()); 489 490 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 491 CodeBlob* cb = stub_frame.cb(); 492 // Verify we have the right vframeArray 493 assert(cb->frame_size() >= 0, "Unexpected frame size"); 494 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 495 496 // If the deopt call site is a MethodHandle invoke call site we have 497 // to adjust the unpack_sp. 498 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 499 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 500 unpack_sp = deoptee.unextended_sp(); 501 502 #ifdef ASSERT 503 assert(cb->is_deoptimization_stub() || 504 cb->is_uncommon_trap_stub() || 505 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 506 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 507 "unexpected code blob: %s", cb->name()); 508 #endif 509 510 // This is a guarantee instead of an assert because if vframe doesn't match 511 // we will unpack the wrong deoptimized frame and wind up in strange places 512 // where it will be very difficult to figure out what went wrong. Better 513 // to die an early death here than some very obscure death later when the 514 // trail is cold. 515 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 516 // in that it will fail to detect a problem when there is one. This needs 517 // more work in tiger timeframe. 518 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 519 520 int number_of_frames = array->frames(); 521 522 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 523 // virtual activation, which is the reverse of the elements in the vframes array. 524 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 525 // +1 because we always have an interpreter return address for the final slot. 526 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 527 int popframe_extra_args = 0; 528 // Create an interpreter return address for the stub to use as its return 529 // address so the skeletal frames are perfectly walkable 530 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 531 532 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 533 // activation be put back on the expression stack of the caller for reexecution 534 if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) { 535 popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words()); 536 } 537 538 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 539 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 540 // than simply use array->sender.pc(). This requires us to walk the current set of frames 541 // 542 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 543 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 544 545 // It's possible that the number of parameters at the call site is 546 // different than number of arguments in the callee when method 547 // handles are used. If the caller is interpreted get the real 548 // value so that the proper amount of space can be added to it's 549 // frame. 550 bool caller_was_method_handle = false; 551 if (deopt_sender.is_interpreted_frame()) { 552 methodHandle method(current, deopt_sender.interpreter_frame_method()); 553 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 554 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 555 // Method handle invokes may involve fairly arbitrary chains of 556 // calls so it's impossible to know how much actual space the 557 // caller has for locals. 558 caller_was_method_handle = true; 559 } 560 } 561 562 // 563 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 564 // frame_sizes/frame_pcs[1] next oldest frame (int) 565 // frame_sizes/frame_pcs[n] youngest frame (int) 566 // 567 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 568 // owns the space for the return address to it's caller). Confusing ain't it. 569 // 570 // The vframe array can address vframes with indices running from 571 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 572 // When we create the skeletal frames we need the oldest frame to be in the zero slot 573 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 574 // so things look a little strange in this loop. 575 // 576 int callee_parameters = 0; 577 int callee_locals = 0; 578 for (int index = 0; index < array->frames(); index++ ) { 579 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 580 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 581 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 582 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 583 callee_locals, 584 index == 0, 585 popframe_extra_args); 586 // This pc doesn't have to be perfect just good enough to identify the frame 587 // as interpreted so the skeleton frame will be walkable 588 // The correct pc will be set when the skeleton frame is completely filled out 589 // The final pc we store in the loop is wrong and will be overwritten below 590 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 591 592 callee_parameters = array->element(index)->method()->size_of_parameters(); 593 callee_locals = array->element(index)->method()->max_locals(); 594 popframe_extra_args = 0; 595 } 596 597 // Compute whether the root vframe returns a float or double value. 598 BasicType return_type; 599 { 600 methodHandle method(current, array->element(0)->method()); 601 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 602 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 603 } 604 605 // Compute information for handling adapters and adjusting the frame size of the caller. 606 int caller_adjustment = 0; 607 608 // Compute the amount the oldest interpreter frame will have to adjust 609 // its caller's stack by. If the caller is a compiled frame then 610 // we pretend that the callee has no parameters so that the 611 // extension counts for the full amount of locals and not just 612 // locals-parms. This is because without a c2i adapter the parm 613 // area as created by the compiled frame will not be usable by 614 // the interpreter. (Depending on the calling convention there 615 // may not even be enough space). 616 617 // QQQ I'd rather see this pushed down into last_frame_adjust 618 // and have it take the sender (aka caller). 619 620 if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) { 621 caller_adjustment = last_frame_adjust(0, callee_locals); 622 } else if (callee_locals > callee_parameters) { 623 // The caller frame may need extending to accommodate 624 // non-parameter locals of the first unpacked interpreted frame. 625 // Compute that adjustment. 626 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 627 } 628 629 // If the sender is deoptimized we must retrieve the address of the handler 630 // since the frame will "magically" show the original pc before the deopt 631 // and we'd undo the deopt. 632 633 frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc(); 634 if (Continuation::is_continuation_enterSpecial(deopt_sender)) { 635 ContinuationEntry::from_frame(deopt_sender)->set_argsize(0); 636 } 637 638 assert(CodeCache::find_blob(frame_pcs[0]) != NULL, "bad pc"); 639 640 #if INCLUDE_JVMCI 641 if (exceptionObject() != NULL) { 642 current->set_exception_oop(exceptionObject()); 643 exec_mode = Unpack_exception; 644 } 645 #endif 646 647 if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 648 assert(current->has_pending_exception(), "should have thrown OOME"); 649 current->set_exception_oop(current->pending_exception()); 650 current->clear_pending_exception(); 651 exec_mode = Unpack_exception; 652 } 653 654 #if INCLUDE_JVMCI 655 if (current->frames_to_pop_failed_realloc() > 0) { 656 current->set_pending_monitorenter(false); 657 } 658 #endif 659 660 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 661 caller_adjustment * BytesPerWord, 662 caller_was_method_handle ? 0 : callee_parameters, 663 number_of_frames, 664 frame_sizes, 665 frame_pcs, 666 return_type, 667 exec_mode); 668 // On some platforms, we need a way to pass some platform dependent 669 // information to the unpacking code so the skeletal frames come out 670 // correct (initial fp value, unextended sp, ...) 671 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 672 673 if (array->frames() > 1) { 674 if (VerifyStack && TraceDeoptimization) { 675 tty->print_cr("Deoptimizing method containing inlining"); 676 } 677 } 678 679 array->set_unroll_block(info); 680 return info; 681 } 682 683 // Called to cleanup deoptimization data structures in normal case 684 // after unpacking to stack and when stack overflow error occurs 685 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 686 vframeArray *array) { 687 688 // Get array if coming from exception 689 if (array == NULL) { 690 array = thread->vframe_array_head(); 691 } 692 thread->set_vframe_array_head(NULL); 693 694 // Free the previous UnrollBlock 695 vframeArray* old_array = thread->vframe_array_last(); 696 thread->set_vframe_array_last(array); 697 698 if (old_array != NULL) { 699 UnrollBlock* old_info = old_array->unroll_block(); 700 old_array->set_unroll_block(NULL); 701 delete old_info; 702 delete old_array; 703 } 704 705 // Deallocate any resource creating in this routine and any ResourceObjs allocated 706 // inside the vframeArray (StackValueCollections) 707 708 delete thread->deopt_mark(); 709 thread->set_deopt_mark(NULL); 710 thread->set_deopt_compiled_method(NULL); 711 712 713 if (JvmtiExport::can_pop_frame()) { 714 // Regardless of whether we entered this routine with the pending 715 // popframe condition bit set, we should always clear it now 716 thread->clear_popframe_condition(); 717 } 718 719 // unpack_frames() is called at the end of the deoptimization handler 720 // and (in C2) at the end of the uncommon trap handler. Note this fact 721 // so that an asynchronous stack walker can work again. This counter is 722 // incremented at the beginning of fetch_unroll_info() and (in C2) at 723 // the beginning of uncommon_trap(). 724 thread->dec_in_deopt_handler(); 725 } 726 727 // Moved from cpu directories because none of the cpus has callee save values. 728 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 729 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 730 731 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 732 // the days we had adapter frames. When we deoptimize a situation where a 733 // compiled caller calls a compiled caller will have registers it expects 734 // to survive the call to the callee. If we deoptimize the callee the only 735 // way we can restore these registers is to have the oldest interpreter 736 // frame that we create restore these values. That is what this routine 737 // will accomplish. 738 739 // At the moment we have modified c2 to not have any callee save registers 740 // so this problem does not exist and this routine is just a place holder. 741 742 assert(f->is_interpreted_frame(), "must be interpreted"); 743 } 744 745 #ifndef PRODUCT 746 static bool falls_through(Bytecodes::Code bc) { 747 switch (bc) { 748 // List may be incomplete. Here we really only care about bytecodes where compiled code 749 // can deoptimize. 750 case Bytecodes::_goto: 751 case Bytecodes::_goto_w: 752 case Bytecodes::_athrow: 753 return false; 754 default: 755 return true; 756 } 757 } 758 #endif 759 760 // Return BasicType of value being returned 761 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 762 assert(thread == JavaThread::current(), "pre-condition"); 763 764 // We are already active in the special DeoptResourceMark any ResourceObj's we 765 // allocate will be freed at the end of the routine. 766 767 // JRT_LEAF methods don't normally allocate handles and there is a 768 // NoHandleMark to enforce that. It is actually safe to use Handles 769 // in a JRT_LEAF method, and sometimes desirable, but to do so we 770 // must use ResetNoHandleMark to bypass the NoHandleMark, and 771 // then use a HandleMark to ensure any Handles we do create are 772 // cleaned up in this scope. 773 ResetNoHandleMark rnhm; 774 HandleMark hm(thread); 775 776 frame stub_frame = thread->last_frame(); 777 778 Continuation::notify_deopt(thread, stub_frame.sp()); 779 780 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 781 // must point to the vframeArray for the unpack frame. 782 vframeArray* array = thread->vframe_array_head(); 783 UnrollBlock* info = array->unroll_block(); 784 785 // We set the last_Java frame. But the stack isn't really parsable here. So we 786 // clear it to make sure JFR understands not to try and walk stacks from events 787 // in here. 788 intptr_t* sp = thread->frame_anchor()->last_Java_sp(); 789 thread->frame_anchor()->set_last_Java_sp(NULL); 790 791 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 792 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 793 794 thread->frame_anchor()->set_last_Java_sp(sp); 795 796 BasicType bt = info->return_type(); 797 798 // If we have an exception pending, claim that the return type is an oop 799 // so the deopt_blob does not overwrite the exception_oop. 800 801 if (exec_mode == Unpack_exception) 802 bt = T_OBJECT; 803 804 // Cleanup thread deopt data 805 cleanup_deopt_info(thread, array); 806 807 #ifndef PRODUCT 808 if (VerifyStack) { 809 ResourceMark res_mark; 810 // Clear pending exception to not break verification code (restored afterwards) 811 PreserveExceptionMark pm(thread); 812 813 thread->validate_frame_layout(); 814 815 // Verify that the just-unpacked frames match the interpreter's 816 // notions of expression stack and locals 817 vframeArray* cur_array = thread->vframe_array_last(); 818 RegisterMap rm(thread, 819 RegisterMap::UpdateMap::skip, 820 RegisterMap::ProcessFrames::include, 821 RegisterMap::WalkContinuation::skip); 822 rm.set_include_argument_oops(false); 823 bool is_top_frame = true; 824 int callee_size_of_parameters = 0; 825 int callee_max_locals = 0; 826 for (int i = 0; i < cur_array->frames(); i++) { 827 vframeArrayElement* el = cur_array->element(i); 828 frame* iframe = el->iframe(); 829 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 830 831 // Get the oop map for this bci 832 InterpreterOopMap mask; 833 int cur_invoke_parameter_size = 0; 834 bool try_next_mask = false; 835 int next_mask_expression_stack_size = -1; 836 int top_frame_expression_stack_adjustment = 0; 837 methodHandle mh(thread, iframe->interpreter_frame_method()); 838 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 839 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 840 int max_bci = mh->code_size(); 841 // Get to the next bytecode if possible 842 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 843 // Check to see if we can grab the number of outgoing arguments 844 // at an uncommon trap for an invoke (where the compiler 845 // generates debug info before the invoke has executed) 846 Bytecodes::Code cur_code = str.next(); 847 Bytecodes::Code next_code = Bytecodes::_shouldnotreachhere; 848 if (Bytecodes::is_invoke(cur_code)) { 849 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 850 cur_invoke_parameter_size = invoke.size_of_parameters(); 851 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 852 callee_size_of_parameters++; 853 } 854 } 855 if (str.bci() < max_bci) { 856 next_code = str.next(); 857 if (next_code >= 0) { 858 // The interpreter oop map generator reports results before 859 // the current bytecode has executed except in the case of 860 // calls. It seems to be hard to tell whether the compiler 861 // has emitted debug information matching the "state before" 862 // a given bytecode or the state after, so we try both 863 if (!Bytecodes::is_invoke(cur_code) && falls_through(cur_code)) { 864 // Get expression stack size for the next bytecode 865 InterpreterOopMap next_mask; 866 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 867 next_mask_expression_stack_size = next_mask.expression_stack_size(); 868 if (Bytecodes::is_invoke(next_code)) { 869 Bytecode_invoke invoke(mh, str.bci()); 870 next_mask_expression_stack_size += invoke.size_of_parameters(); 871 } 872 // Need to subtract off the size of the result type of 873 // the bytecode because this is not described in the 874 // debug info but returned to the interpreter in the TOS 875 // caching register 876 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 877 if (bytecode_result_type != T_ILLEGAL) { 878 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 879 } 880 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 881 try_next_mask = true; 882 } 883 } 884 } 885 886 // Verify stack depth and oops in frame 887 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 888 if (!( 889 /* SPARC */ 890 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 891 /* x86 */ 892 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 893 (try_next_mask && 894 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 895 top_frame_expression_stack_adjustment))) || 896 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 897 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 898 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 899 )) { 900 { 901 // Print out some information that will help us debug the problem 902 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 903 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 904 tty->print_cr(" Current code %s", Bytecodes::name(cur_code)); 905 if (try_next_mask) { 906 tty->print_cr(" Next code %s", Bytecodes::name(next_code)); 907 } 908 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 909 iframe->interpreter_frame_expression_stack_size()); 910 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 911 tty->print_cr(" try_next_mask = %d", try_next_mask); 912 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 913 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 914 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 915 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 916 tty->print_cr(" exec_mode = %d", exec_mode); 917 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 918 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 919 tty->print_cr(" Interpreted frames:"); 920 for (int k = 0; k < cur_array->frames(); k++) { 921 vframeArrayElement* el = cur_array->element(k); 922 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 923 } 924 cur_array->print_on_2(tty); 925 } 926 guarantee(false, "wrong number of expression stack elements during deopt"); 927 } 928 VerifyOopClosure verify; 929 iframe->oops_interpreted_do(&verify, &rm, false); 930 callee_size_of_parameters = mh->size_of_parameters(); 931 callee_max_locals = mh->max_locals(); 932 is_top_frame = false; 933 } 934 } 935 #endif // !PRODUCT 936 937 return bt; 938 JRT_END 939 940 class DeoptimizeMarkedClosure : public HandshakeClosure { 941 public: 942 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 943 void do_thread(Thread* thread) { 944 JavaThread* jt = JavaThread::cast(thread); 945 jt->deoptimize_marked_methods(); 946 } 947 }; 948 949 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) { 950 ResourceMark rm; 951 952 // Make the dependent methods not entrant 953 if (nmethod_only != NULL) { 954 nmethod_only->mark_for_deoptimization(); 955 nmethod_only->make_not_entrant(); 956 CodeCache::make_nmethod_deoptimized(nmethod_only); 957 } else { 958 CodeCache::make_marked_nmethods_deoptimized(); 959 } 960 961 DeoptimizeMarkedClosure deopt; 962 if (SafepointSynchronize::is_at_safepoint()) { 963 Threads::java_threads_do(&deopt); 964 } else { 965 Handshake::execute(&deopt); 966 } 967 } 968 969 Deoptimization::DeoptAction Deoptimization::_unloaded_action 970 = Deoptimization::Action_reinterpret; 971 972 #if INCLUDE_JVMCI 973 template<typename CacheType> 974 class BoxCacheBase : public CHeapObj<mtCompiler> { 975 protected: 976 static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) { 977 ResourceMark rm(thread); 978 char* klass_name_str = klass_name->as_C_string(); 979 InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle(), Handle()); 980 guarantee(ik != NULL, "%s must be loaded", klass_name_str); 981 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 982 CacheType::compute_offsets(ik); 983 return ik; 984 } 985 }; 986 987 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 988 PrimitiveType _low; 989 PrimitiveType _high; 990 jobject _cache; 991 protected: 992 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 993 BoxCache(Thread* thread) { 994 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol()); 995 objArrayOop cache = CacheType::cache(ik); 996 assert(cache->length() > 0, "Empty cache"); 997 _low = BoxType::value(cache->obj_at(0)); 998 _high = _low + cache->length() - 1; 999 _cache = JNIHandles::make_global(Handle(thread, cache)); 1000 } 1001 ~BoxCache() { 1002 JNIHandles::destroy_global(_cache); 1003 } 1004 public: 1005 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 1006 if (_singleton == NULL) { 1007 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 1008 if (!Atomic::replace_if_null(&_singleton, s)) { 1009 delete s; 1010 } 1011 } 1012 return _singleton; 1013 } 1014 oop lookup(PrimitiveType value) { 1015 if (_low <= value && value <= _high) { 1016 int offset = value - _low; 1017 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 1018 } 1019 return NULL; 1020 } 1021 oop lookup_raw(intptr_t raw_value) { 1022 // Have to cast to avoid little/big-endian problems. 1023 if (sizeof(PrimitiveType) > sizeof(jint)) { 1024 jlong value = (jlong)raw_value; 1025 return lookup(value); 1026 } 1027 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 1028 return lookup(value); 1029 } 1030 }; 1031 1032 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 1033 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 1034 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 1035 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 1036 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 1037 1038 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 1039 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 1040 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 1041 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 1042 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 1043 1044 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 1045 jobject _true_cache; 1046 jobject _false_cache; 1047 protected: 1048 static BooleanBoxCache *_singleton; 1049 BooleanBoxCache(Thread *thread) { 1050 InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol()); 1051 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 1052 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 1053 } 1054 ~BooleanBoxCache() { 1055 JNIHandles::destroy_global(_true_cache); 1056 JNIHandles::destroy_global(_false_cache); 1057 } 1058 public: 1059 static BooleanBoxCache* singleton(Thread* thread) { 1060 if (_singleton == NULL) { 1061 BooleanBoxCache* s = new BooleanBoxCache(thread); 1062 if (!Atomic::replace_if_null(&_singleton, s)) { 1063 delete s; 1064 } 1065 } 1066 return _singleton; 1067 } 1068 oop lookup_raw(intptr_t raw_value) { 1069 // Have to cast to avoid little/big-endian problems. 1070 jboolean value = (jboolean)*((jint*)&raw_value); 1071 return lookup(value); 1072 } 1073 oop lookup(jboolean value) { 1074 if (value != 0) { 1075 return JNIHandles::resolve_non_null(_true_cache); 1076 } 1077 return JNIHandles::resolve_non_null(_false_cache); 1078 } 1079 }; 1080 1081 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 1082 1083 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { 1084 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 1085 BasicType box_type = vmClasses::box_klass_type(k); 1086 if (box_type != T_OBJECT) { 1087 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 1088 switch(box_type) { 1089 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1090 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1091 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1092 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1093 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1094 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 1095 default:; 1096 } 1097 } 1098 return NULL; 1099 } 1100 #endif // INCLUDE_JVMCI 1101 1102 #if COMPILER2_OR_JVMCI 1103 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 1104 Handle pending_exception(THREAD, thread->pending_exception()); 1105 const char* exception_file = thread->exception_file(); 1106 int exception_line = thread->exception_line(); 1107 thread->clear_pending_exception(); 1108 1109 bool failures = false; 1110 1111 for (int i = 0; i < objects->length(); i++) { 1112 assert(objects->at(i)->is_object(), "invalid debug information"); 1113 ObjectValue* sv = (ObjectValue*) objects->at(i); 1114 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1115 1116 // Check if the object may be null and has an additional is_init input that needs 1117 // to be checked before using the field values. Skip re-allocation if it is null. 1118 if (sv->maybe_null()) { 1119 assert(k->is_inline_klass(), "must be an inline klass"); 1120 intptr_t init_value = StackValue::create_stack_value(fr, reg_map, sv->is_init())->get_int(); 1121 jint is_init = (jint)*((jint*)&init_value); 1122 if (is_init == 0) { 1123 continue; 1124 } 1125 } 1126 1127 oop obj = NULL; 1128 if (k->is_instance_klass()) { 1129 #if INCLUDE_JVMCI 1130 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 1131 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 1132 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 1133 obj = get_cached_box(abv, fr, reg_map, THREAD); 1134 if (obj != NULL) { 1135 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1136 abv->set_cached(true); 1137 } 1138 } 1139 #endif // INCLUDE_JVMCI 1140 1141 InstanceKlass* ik = InstanceKlass::cast(k); 1142 if (obj == NULL) { 1143 #ifdef COMPILER2 1144 if (EnableVectorSupport && VectorSupport::is_vector(ik)) { 1145 obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD); 1146 } else { 1147 obj = ik->allocate_instance(THREAD); 1148 } 1149 #else 1150 obj = ik->allocate_instance(THREAD); 1151 #endif // COMPILER2 1152 } 1153 } else if (k->is_flatArray_klass()) { 1154 FlatArrayKlass* ak = FlatArrayKlass::cast(k); 1155 // Inline type array must be zeroed because not all memory is reassigned 1156 obj = ak->allocate(sv->field_size(), THREAD); 1157 } else if (k->is_typeArray_klass()) { 1158 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1159 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1160 int len = sv->field_size() / type2size[ak->element_type()]; 1161 obj = ak->allocate(len, THREAD); 1162 } else if (k->is_objArray_klass()) { 1163 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1164 obj = ak->allocate(sv->field_size(), THREAD); 1165 } 1166 1167 if (obj == NULL) { 1168 failures = true; 1169 } 1170 1171 assert(sv->value().is_null(), "redundant reallocation"); 1172 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1173 CLEAR_PENDING_EXCEPTION; 1174 sv->set_value(obj); 1175 } 1176 1177 if (failures) { 1178 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1179 } else if (pending_exception.not_null()) { 1180 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1181 } 1182 1183 return failures; 1184 } 1185 1186 // We're deoptimizing at the return of a call, inline type fields are 1187 // in registers. When we go back to the interpreter, it will expect a 1188 // reference to an inline type instance. Allocate and initialize it from 1189 // the register values here. 1190 bool Deoptimization::realloc_inline_type_result(InlineKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 1191 oop new_vt = vk->realloc_result(map, return_oops, THREAD); 1192 if (new_vt == NULL) { 1193 CLEAR_PENDING_EXCEPTION; 1194 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 1195 } 1196 return_oops.clear(); 1197 return_oops.push(Handle(THREAD, new_vt)); 1198 return false; 1199 } 1200 1201 #if INCLUDE_JVMCI 1202 /** 1203 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints), 1204 * we need to somehow be able to recover the actual kind to be able to write the correct 1205 * amount of bytes. 1206 * For that purpose, this method assumes that, for an entry spanning n bytes at index i, 1207 * the entries at index n + 1 to n + i are 'markers'. 1208 * For example, if we were writing a short at index 4 of a byte array of size 8, the 1209 * expected form of the array would be: 1210 * 1211 * {b0, b1, b2, b3, INT, marker, b6, b7} 1212 * 1213 * Thus, in order to get back the size of the entry, we simply need to count the number 1214 * of marked entries 1215 * 1216 * @param virtualArray the virtualized byte array 1217 * @param i index of the virtual entry we are recovering 1218 * @return The number of bytes the entry spans 1219 */ 1220 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) { 1221 int index = i; 1222 while (++index < virtualArray->field_size() && 1223 virtualArray->field_at(index)->is_marker()) {} 1224 return index - i; 1225 } 1226 1227 /** 1228 * If there was a guarantee for byte array to always start aligned to a long, we could 1229 * do a simple check on the parity of the index. Unfortunately, that is not always the 1230 * case. Thus, we check alignment of the actual address we are writing to. 1231 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to 1232 * write a long to index 3. 1233 */ 1234 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) { 1235 jbyte* res = obj->byte_at_addr(index); 1236 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write"); 1237 return res; 1238 } 1239 1240 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) { 1241 switch (byte_count) { 1242 case 1: 1243 obj->byte_at_put(index, (jbyte) *((jint *) &val)); 1244 break; 1245 case 2: 1246 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val); 1247 break; 1248 case 4: 1249 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val); 1250 break; 1251 case 8: 1252 *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val); 1253 break; 1254 default: 1255 ShouldNotReachHere(); 1256 } 1257 } 1258 #endif // INCLUDE_JVMCI 1259 1260 1261 // restore elements of an eliminated type array 1262 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1263 int index = 0; 1264 intptr_t val; 1265 1266 for (int i = 0; i < sv->field_size(); i++) { 1267 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1268 switch(type) { 1269 case T_LONG: case T_DOUBLE: { 1270 assert(value->type() == T_INT, "Agreement."); 1271 StackValue* low = 1272 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1273 #ifdef _LP64 1274 jlong res = (jlong)low->get_int(); 1275 #else 1276 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1277 #endif 1278 obj->long_at_put(index, res); 1279 break; 1280 } 1281 1282 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1283 case T_INT: case T_FLOAT: { // 4 bytes. 1284 assert(value->type() == T_INT, "Agreement."); 1285 bool big_value = false; 1286 if (i + 1 < sv->field_size() && type == T_INT) { 1287 if (sv->field_at(i)->is_location()) { 1288 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1289 if (type == Location::dbl || type == Location::lng) { 1290 big_value = true; 1291 } 1292 } else if (sv->field_at(i)->is_constant_int()) { 1293 ScopeValue* next_scope_field = sv->field_at(i + 1); 1294 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1295 big_value = true; 1296 } 1297 } 1298 } 1299 1300 if (big_value) { 1301 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1302 #ifdef _LP64 1303 jlong res = (jlong)low->get_int(); 1304 #else 1305 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1306 #endif 1307 obj->int_at_put(index, (jint)*((jint*)&res)); 1308 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1309 } else { 1310 val = value->get_int(); 1311 obj->int_at_put(index, (jint)*((jint*)&val)); 1312 } 1313 break; 1314 } 1315 1316 case T_SHORT: 1317 assert(value->type() == T_INT, "Agreement."); 1318 val = value->get_int(); 1319 obj->short_at_put(index, (jshort)*((jint*)&val)); 1320 break; 1321 1322 case T_CHAR: 1323 assert(value->type() == T_INT, "Agreement."); 1324 val = value->get_int(); 1325 obj->char_at_put(index, (jchar)*((jint*)&val)); 1326 break; 1327 1328 case T_BYTE: { 1329 assert(value->type() == T_INT, "Agreement."); 1330 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'. 1331 val = value->get_int(); 1332 #if INCLUDE_JVMCI 1333 int byte_count = count_number_of_bytes_for_entry(sv, i); 1334 byte_array_put(obj, val, index, byte_count); 1335 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip. 1336 i += byte_count - 1; // Balance the loop counter. 1337 index += byte_count; 1338 // index has been updated so continue at top of loop 1339 continue; 1340 #else 1341 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1342 break; 1343 #endif // INCLUDE_JVMCI 1344 } 1345 1346 case T_BOOLEAN: { 1347 assert(value->type() == T_INT, "Agreement."); 1348 val = value->get_int(); 1349 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1350 break; 1351 } 1352 1353 default: 1354 ShouldNotReachHere(); 1355 } 1356 index++; 1357 } 1358 } 1359 1360 // restore fields of an eliminated object array 1361 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1362 for (int i = 0; i < sv->field_size(); i++) { 1363 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1364 assert(value->type() == T_OBJECT, "object element expected"); 1365 obj->obj_at_put(i, value->get_obj()()); 1366 } 1367 } 1368 1369 class ReassignedField { 1370 public: 1371 int _offset; 1372 BasicType _type; 1373 InstanceKlass* _klass; 1374 public: 1375 ReassignedField() { 1376 _offset = 0; 1377 _type = T_ILLEGAL; 1378 _klass = NULL; 1379 } 1380 }; 1381 1382 int compare(ReassignedField* left, ReassignedField* right) { 1383 return left->_offset - right->_offset; 1384 } 1385 1386 // Restore fields of an eliminated instance object using the same field order 1387 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1388 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) { 1389 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1390 InstanceKlass* ik = klass; 1391 while (ik != NULL) { 1392 for (AllFieldStream fs(ik); !fs.done(); fs.next()) { 1393 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1394 ReassignedField field; 1395 field._offset = fs.offset(); 1396 field._type = Signature::basic_type(fs.signature()); 1397 if (fs.signature()->is_Q_signature()) { 1398 if (fs.is_inlined()) { 1399 // Resolve klass of flattened inline type field 1400 field._klass = InlineKlass::cast(klass->get_inline_type_field_klass(fs.index())); 1401 } else { 1402 field._type = T_OBJECT; 1403 } 1404 } 1405 fields->append(field); 1406 } 1407 } 1408 ik = ik->superklass(); 1409 } 1410 fields->sort(compare); 1411 for (int i = 0; i < fields->length(); i++) { 1412 BasicType type = fields->at(i)._type; 1413 int offset = base_offset + fields->at(i)._offset; 1414 // Check for flattened inline type field before accessing the ScopeValue because it might not have any fields 1415 if (type == T_PRIMITIVE_OBJECT) { 1416 // Recursively re-assign flattened inline type fields 1417 InstanceKlass* vk = fields->at(i)._klass; 1418 assert(vk != NULL, "must be resolved"); 1419 offset -= InlineKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1420 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1421 continue; // Continue because we don't need to increment svIndex 1422 } 1423 intptr_t val; 1424 ScopeValue* scope_field = sv->field_at(svIndex); 1425 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1426 switch (type) { 1427 case T_OBJECT: 1428 case T_ARRAY: 1429 assert(value->type() == T_OBJECT, "Agreement."); 1430 obj->obj_field_put(offset, value->get_obj()()); 1431 break; 1432 1433 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1434 case T_INT: case T_FLOAT: { // 4 bytes. 1435 assert(value->type() == T_INT, "Agreement."); 1436 bool big_value = false; 1437 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1438 if (scope_field->is_location()) { 1439 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1440 if (type == Location::dbl || type == Location::lng) { 1441 big_value = true; 1442 } 1443 } 1444 if (scope_field->is_constant_int()) { 1445 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1446 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1447 big_value = true; 1448 } 1449 } 1450 } 1451 1452 if (big_value) { 1453 i++; 1454 assert(i < fields->length(), "second T_INT field needed"); 1455 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1456 } else { 1457 val = value->get_int(); 1458 obj->int_field_put(offset, (jint)*((jint*)&val)); 1459 break; 1460 } 1461 } 1462 /* no break */ 1463 1464 case T_LONG: case T_DOUBLE: { 1465 assert(value->type() == T_INT, "Agreement."); 1466 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1467 #ifdef _LP64 1468 jlong res = (jlong)low->get_int(); 1469 #else 1470 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1471 #endif 1472 obj->long_field_put(offset, res); 1473 break; 1474 } 1475 1476 case T_SHORT: 1477 assert(value->type() == T_INT, "Agreement."); 1478 val = value->get_int(); 1479 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1480 break; 1481 1482 case T_CHAR: 1483 assert(value->type() == T_INT, "Agreement."); 1484 val = value->get_int(); 1485 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1486 break; 1487 1488 case T_BYTE: 1489 assert(value->type() == T_INT, "Agreement."); 1490 val = value->get_int(); 1491 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1492 break; 1493 1494 case T_BOOLEAN: 1495 assert(value->type() == T_INT, "Agreement."); 1496 val = value->get_int(); 1497 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1498 break; 1499 1500 default: 1501 ShouldNotReachHere(); 1502 } 1503 svIndex++; 1504 } 1505 return svIndex; 1506 } 1507 1508 // restore fields of an eliminated inline type array 1509 void Deoptimization::reassign_flat_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, flatArrayOop obj, FlatArrayKlass* vak, bool skip_internal, TRAPS) { 1510 InlineKlass* vk = vak->element_klass(); 1511 assert(vk->flatten_array(), "should only be used for flattened inline type arrays"); 1512 // Adjust offset to omit oop header 1513 int base_offset = arrayOopDesc::base_offset_in_bytes(T_PRIMITIVE_OBJECT) - InlineKlass::cast(vk)->first_field_offset(); 1514 // Initialize all elements of the flattened inline type array 1515 for (int i = 0; i < sv->field_size(); i++) { 1516 ScopeValue* val = sv->field_at(i); 1517 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1518 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, skip_internal, offset, CHECK); 1519 } 1520 } 1521 1522 // restore fields of all eliminated objects and arrays 1523 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1524 for (int i = 0; i < objects->length(); i++) { 1525 ObjectValue* sv = (ObjectValue*) objects->at(i); 1526 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1527 Handle obj = sv->value(); 1528 assert(obj.not_null() || realloc_failures || sv->maybe_null(), "reallocation was missed"); 1529 #ifndef PRODUCT 1530 if (PrintDeoptimizationDetails) { 1531 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1532 } 1533 #endif // !PRODUCT 1534 1535 if (obj.is_null()) { 1536 continue; 1537 } 1538 1539 #if INCLUDE_JVMCI 1540 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1541 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1542 continue; 1543 } 1544 #endif // INCLUDE_JVMCI 1545 #ifdef COMPILER2 1546 if (EnableVectorSupport && VectorSupport::is_vector(k)) { 1547 assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string()); 1548 ScopeValue* payload = sv->field_at(0); 1549 if (payload->is_location() && 1550 payload->as_LocationValue()->location().type() == Location::vector) { 1551 #ifndef PRODUCT 1552 if (PrintDeoptimizationDetails) { 1553 tty->print_cr("skip field reassignment for this vector - it should be assigned already"); 1554 if (Verbose) { 1555 Handle obj = sv->value(); 1556 k->oop_print_on(obj(), tty); 1557 } 1558 } 1559 #endif // !PRODUCT 1560 continue; // Such vector's value was already restored in VectorSupport::allocate_vector(). 1561 } 1562 // Else fall-through to do assignment for scalar-replaced boxed vector representation 1563 // which could be restored after vector object allocation. 1564 } 1565 #endif /* !COMPILER2 */ 1566 if (k->is_instance_klass()) { 1567 InstanceKlass* ik = InstanceKlass::cast(k); 1568 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1569 } else if (k->is_flatArray_klass()) { 1570 FlatArrayKlass* vak = FlatArrayKlass::cast(k); 1571 reassign_flat_array_elements(fr, reg_map, sv, (flatArrayOop) obj(), vak, skip_internal, CHECK); 1572 } else if (k->is_typeArray_klass()) { 1573 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1574 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1575 } else if (k->is_objArray_klass()) { 1576 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1577 } 1578 } 1579 } 1580 1581 1582 // relock objects for which synchronization was eliminated 1583 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1584 JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) { 1585 bool relocked_objects = false; 1586 for (int i = 0; i < monitors->length(); i++) { 1587 MonitorInfo* mon_info = monitors->at(i); 1588 if (mon_info->eliminated()) { 1589 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1590 relocked_objects = true; 1591 if (!mon_info->owner_is_scalar_replaced()) { 1592 Handle obj(thread, mon_info->owner()); 1593 markWord mark = obj->mark(); 1594 if (exec_mode == Unpack_none) { 1595 if (mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) { 1596 // With exec_mode == Unpack_none obj may be thread local and locked in 1597 // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header. 1598 markWord dmw = mark.displaced_mark_helper(); 1599 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) NULL)); 1600 obj->set_mark(dmw); 1601 } 1602 if (mark.has_monitor()) { 1603 // defer relocking if the deoptee thread is currently waiting for obj 1604 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1605 if (waiting_monitor != NULL && waiting_monitor->object() == obj()) { 1606 assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization"); 1607 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1608 JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread); 1609 continue; 1610 } 1611 } 1612 } 1613 BasicLock* lock = mon_info->lock(); 1614 ObjectSynchronizer::enter(obj, lock, deoptee_thread); 1615 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1616 } 1617 } 1618 } 1619 return relocked_objects; 1620 } 1621 #endif // COMPILER2_OR_JVMCI 1622 1623 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1624 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1625 1626 // Register map for next frame (used for stack crawl). We capture 1627 // the state of the deopt'ing frame's caller. Thus if we need to 1628 // stuff a C2I adapter we can properly fill in the callee-save 1629 // register locations. 1630 frame caller = fr.sender(reg_map); 1631 int frame_size = caller.sp() - fr.sp(); 1632 1633 frame sender = caller; 1634 1635 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1636 // the vframeArray containing the unpacking information is allocated in the C heap. 1637 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1638 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1639 1640 // Compare the vframeArray to the collected vframes 1641 assert(array->structural_compare(thread, chunk), "just checking"); 1642 1643 if (TraceDeoptimization) { 1644 ResourceMark rm; 1645 stringStream st; 1646 st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array)); 1647 st.print(" "); 1648 fr.print_on(&st); 1649 st.print_cr(" Virtual frames (innermost/newest first):"); 1650 for (int index = 0; index < chunk->length(); index++) { 1651 compiledVFrame* vf = chunk->at(index); 1652 int bci = vf->raw_bci(); 1653 const char* code_name; 1654 if (bci == SynchronizationEntryBCI) { 1655 code_name = "sync entry"; 1656 } else { 1657 Bytecodes::Code code = vf->method()->code_at(bci); 1658 code_name = Bytecodes::name(code); 1659 } 1660 1661 st.print(" VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf)); 1662 st.print(" - %s", vf->method()->name_and_sig_as_C_string()); 1663 st.print(" - %s", code_name); 1664 st.print_cr(" @ bci=%d ", bci); 1665 } 1666 tty->print_raw(st.freeze()); 1667 tty->cr(); 1668 } 1669 1670 return array; 1671 } 1672 1673 #if COMPILER2_OR_JVMCI 1674 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1675 // Reallocation of some scalar replaced objects failed. Record 1676 // that we need to pop all the interpreter frames for the 1677 // deoptimized compiled frame. 1678 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1679 thread->set_frames_to_pop_failed_realloc(array->frames()); 1680 // Unlock all monitors here otherwise the interpreter will see a 1681 // mix of locked and unlocked monitors (because of failed 1682 // reallocations of synchronized objects) and be confused. 1683 for (int i = 0; i < array->frames(); i++) { 1684 MonitorChunk* monitors = array->element(i)->monitors(); 1685 if (monitors != NULL) { 1686 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1687 BasicObjectLock* src = monitors->at(j); 1688 if (src->obj() != NULL) { 1689 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1690 } 1691 } 1692 array->element(i)->free_monitors(thread); 1693 #ifdef ASSERT 1694 array->element(i)->set_removed_monitors(); 1695 #endif 1696 } 1697 } 1698 } 1699 #endif 1700 1701 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1702 assert(fr.can_be_deoptimized(), "checking frame type"); 1703 1704 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1705 1706 if (LogCompilation && xtty != NULL) { 1707 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1708 assert(cm != NULL, "only compiled methods can deopt"); 1709 1710 ttyLocker ttyl; 1711 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1712 cm->log_identity(xtty); 1713 xtty->end_head(); 1714 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1715 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1716 xtty->method(sd->method()); 1717 xtty->end_elem(); 1718 if (sd->is_top()) break; 1719 } 1720 xtty->tail("deoptimized"); 1721 } 1722 1723 Continuation::notify_deopt(thread, fr.sp()); 1724 1725 // Patch the compiled method so that when execution returns to it we will 1726 // deopt the execution state and return to the interpreter. 1727 fr.deoptimize(thread); 1728 } 1729 1730 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) { 1731 // Deoptimize only if the frame comes from compiled code. 1732 // Do not deoptimize the frame which is already patched 1733 // during the execution of the loops below. 1734 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1735 return; 1736 } 1737 ResourceMark rm; 1738 deoptimize_single_frame(thread, fr, reason); 1739 } 1740 1741 #if INCLUDE_JVMCI 1742 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1743 // there is no exception handler for this pc => deoptimize 1744 cm->make_not_entrant(); 1745 1746 // Use Deoptimization::deoptimize for all of its side-effects: 1747 // gathering traps statistics, logging... 1748 // it also patches the return pc but we do not care about that 1749 // since we return a continuation to the deopt_blob below. 1750 JavaThread* thread = JavaThread::current(); 1751 RegisterMap reg_map(thread, 1752 RegisterMap::UpdateMap::skip, 1753 RegisterMap::ProcessFrames::include, 1754 RegisterMap::WalkContinuation::skip); 1755 frame runtime_frame = thread->last_frame(); 1756 frame caller_frame = runtime_frame.sender(®_map); 1757 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1758 vframe* vf = vframe::new_vframe(&caller_frame, ®_map, thread); 1759 compiledVFrame* cvf = compiledVFrame::cast(vf); 1760 ScopeDesc* imm_scope = cvf->scope(); 1761 MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true); 1762 if (imm_mdo != NULL) { 1763 ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), NULL); 1764 if (pdata != NULL && pdata->is_BitData()) { 1765 BitData* bit_data = (BitData*) pdata; 1766 bit_data->set_exception_seen(); 1767 } 1768 } 1769 1770 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler); 1771 1772 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1773 if (trap_mdo != NULL) { 1774 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1775 } 1776 1777 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1778 } 1779 #endif 1780 1781 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1782 assert(thread == Thread::current() || 1783 thread->is_handshake_safe_for(Thread::current()) || 1784 SafepointSynchronize::is_at_safepoint(), 1785 "can only deoptimize other thread at a safepoint/handshake"); 1786 // Compute frame and register map based on thread and sp. 1787 RegisterMap reg_map(thread, 1788 RegisterMap::UpdateMap::skip, 1789 RegisterMap::ProcessFrames::include, 1790 RegisterMap::WalkContinuation::skip); 1791 frame fr = thread->last_frame(); 1792 while (fr.id() != id) { 1793 fr = fr.sender(®_map); 1794 } 1795 deoptimize(thread, fr, reason); 1796 } 1797 1798 1799 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1800 Thread* current = Thread::current(); 1801 if (thread == current || thread->is_handshake_safe_for(current)) { 1802 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1803 } else { 1804 VM_DeoptimizeFrame deopt(thread, id, reason); 1805 VMThread::execute(&deopt); 1806 } 1807 } 1808 1809 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1810 deoptimize_frame(thread, id, Reason_constraint); 1811 } 1812 1813 // JVMTI PopFrame support 1814 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1815 { 1816 assert(thread == JavaThread::current(), "pre-condition"); 1817 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1818 } 1819 JRT_END 1820 1821 MethodData* 1822 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1823 bool create_if_missing) { 1824 JavaThread* THREAD = thread; // For exception macros. 1825 MethodData* mdo = m()->method_data(); 1826 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1827 // Build an MDO. Ignore errors like OutOfMemory; 1828 // that simply means we won't have an MDO to update. 1829 Method::build_profiling_method_data(m, THREAD); 1830 if (HAS_PENDING_EXCEPTION) { 1831 // Only metaspace OOM is expected. No Java code executed. 1832 assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1833 CLEAR_PENDING_EXCEPTION; 1834 } 1835 mdo = m()->method_data(); 1836 } 1837 return mdo; 1838 } 1839 1840 #if COMPILER2_OR_JVMCI 1841 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1842 // In case of an unresolved klass entry, load the class. 1843 // This path is exercised from case _ldc in Parse::do_one_bytecode, 1844 // and probably nowhere else. 1845 // Even that case would benefit from simply re-interpreting the 1846 // bytecode, without paying special attention to the class index. 1847 // So this whole "class index" feature should probably be removed. 1848 1849 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1850 Klass* tk = constant_pool->klass_at(index, THREAD); 1851 if (HAS_PENDING_EXCEPTION) { 1852 // Exception happened during classloading. We ignore the exception here, since it 1853 // is going to be rethrown since the current activation is going to be deoptimized and 1854 // the interpreter will re-execute the bytecode. 1855 // Do not clear probable Async Exceptions. 1856 CLEAR_PENDING_NONASYNC_EXCEPTION; 1857 // Class loading called java code which may have caused a stack 1858 // overflow. If the exception was thrown right before the return 1859 // to the runtime the stack is no longer guarded. Reguard the 1860 // stack otherwise if we return to the uncommon trap blob and the 1861 // stack bang causes a stack overflow we crash. 1862 JavaThread* jt = THREAD; 1863 bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed(); 1864 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1865 } 1866 return; 1867 } 1868 1869 assert(!constant_pool->tag_at(index).is_symbol(), 1870 "no symbolic names here, please"); 1871 } 1872 1873 #if INCLUDE_JFR 1874 1875 class DeoptReasonSerializer : public JfrSerializer { 1876 public: 1877 void serialize(JfrCheckpointWriter& writer) { 1878 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1) 1879 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) { 1880 writer.write_key((u8)i); 1881 writer.write(Deoptimization::trap_reason_name(i)); 1882 } 1883 } 1884 }; 1885 1886 class DeoptActionSerializer : public JfrSerializer { 1887 public: 1888 void serialize(JfrCheckpointWriter& writer) { 1889 static const u4 nof_actions = Deoptimization::Action_LIMIT; 1890 writer.write_count(nof_actions); 1891 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) { 1892 writer.write_key(i); 1893 writer.write(Deoptimization::trap_action_name((int)i)); 1894 } 1895 } 1896 }; 1897 1898 static void register_serializers() { 1899 static int critical_section = 0; 1900 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) { 1901 return; 1902 } 1903 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer()); 1904 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer()); 1905 } 1906 1907 static void post_deoptimization_event(CompiledMethod* nm, 1908 const Method* method, 1909 int trap_bci, 1910 int instruction, 1911 Deoptimization::DeoptReason reason, 1912 Deoptimization::DeoptAction action) { 1913 assert(nm != NULL, "invariant"); 1914 assert(method != NULL, "invariant"); 1915 if (EventDeoptimization::is_enabled()) { 1916 static bool serializers_registered = false; 1917 if (!serializers_registered) { 1918 register_serializers(); 1919 serializers_registered = true; 1920 } 1921 EventDeoptimization event; 1922 event.set_compileId(nm->compile_id()); 1923 event.set_compiler(nm->compiler_type()); 1924 event.set_method(method); 1925 event.set_lineNumber(method->line_number_from_bci(trap_bci)); 1926 event.set_bci(trap_bci); 1927 event.set_instruction(instruction); 1928 event.set_reason(reason); 1929 event.set_action(action); 1930 event.commit(); 1931 } 1932 } 1933 1934 #endif // INCLUDE_JFR 1935 1936 static void log_deopt(CompiledMethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci, 1937 const char* reason_name, const char* reason_action) { 1938 LogTarget(Debug, deoptimization) lt; 1939 if (lt.is_enabled()) { 1940 LogStream ls(lt); 1941 bool is_osr = nm->is_osr_method(); 1942 ls.print("cid=%4d %s level=%d", 1943 nm->compile_id(), (is_osr ? "osr" : " "), nm->comp_level()); 1944 ls.print(" %s", tm->name_and_sig_as_C_string()); 1945 ls.print(" trap_bci=%d ", trap_bci); 1946 if (is_osr) { 1947 ls.print("osr_bci=%d ", nm->osr_entry_bci()); 1948 } 1949 ls.print("%s ", reason_name); 1950 ls.print("%s ", reason_action); 1951 ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT, 1952 pc, fr.pc() - nm->code_begin()); 1953 } 1954 } 1955 1956 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) { 1957 HandleMark hm(current); 1958 1959 // uncommon_trap() is called at the beginning of the uncommon trap 1960 // handler. Note this fact before we start generating temporary frames 1961 // that can confuse an asynchronous stack walker. This counter is 1962 // decremented at the end of unpack_frames(). 1963 1964 current->inc_in_deopt_handler(); 1965 1966 #if INCLUDE_JVMCI 1967 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1968 RegisterMap reg_map(current, 1969 RegisterMap::UpdateMap::include, 1970 RegisterMap::ProcessFrames::include, 1971 RegisterMap::WalkContinuation::skip); 1972 #else 1973 RegisterMap reg_map(current, 1974 RegisterMap::UpdateMap::skip, 1975 RegisterMap::ProcessFrames::include, 1976 RegisterMap::WalkContinuation::skip); 1977 #endif 1978 frame stub_frame = current->last_frame(); 1979 frame fr = stub_frame.sender(®_map); 1980 1981 // Log a message 1982 Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1983 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1984 1985 { 1986 ResourceMark rm; 1987 1988 DeoptReason reason = trap_request_reason(trap_request); 1989 DeoptAction action = trap_request_action(trap_request); 1990 #if INCLUDE_JVMCI 1991 int debug_id = trap_request_debug_id(trap_request); 1992 #endif 1993 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1994 1995 vframe* vf = vframe::new_vframe(&fr, ®_map, current); 1996 compiledVFrame* cvf = compiledVFrame::cast(vf); 1997 1998 CompiledMethod* nm = cvf->code(); 1999 2000 ScopeDesc* trap_scope = cvf->scope(); 2001 2002 bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint); 2003 2004 if (is_receiver_constraint_failure) { 2005 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), 2006 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 2007 JVMCI_ONLY(COMMA debug_id)); 2008 } 2009 2010 methodHandle trap_method(current, trap_scope->method()); 2011 int trap_bci = trap_scope->bci(); 2012 #if INCLUDE_JVMCI 2013 jlong speculation = current->pending_failed_speculation(); 2014 if (nm->is_compiled_by_jvmci()) { 2015 nm->as_nmethod()->update_speculation(current); 2016 } else { 2017 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 2018 } 2019 2020 if (trap_bci == SynchronizationEntryBCI) { 2021 trap_bci = 0; 2022 current->set_pending_monitorenter(true); 2023 } 2024 2025 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 2026 current->set_pending_transfer_to_interpreter(true); 2027 } 2028 #endif 2029 2030 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 2031 // Record this event in the histogram. 2032 gather_statistics(reason, action, trap_bc); 2033 2034 // Ensure that we can record deopt. history: 2035 // Need MDO to record RTM code generation state. 2036 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 2037 2038 methodHandle profiled_method; 2039 #if INCLUDE_JVMCI 2040 if (nm->is_compiled_by_jvmci()) { 2041 profiled_method = methodHandle(current, nm->method()); 2042 } else { 2043 profiled_method = trap_method; 2044 } 2045 #else 2046 profiled_method = trap_method; 2047 #endif 2048 2049 MethodData* trap_mdo = 2050 get_method_data(current, profiled_method, create_if_missing); 2051 2052 { // Log Deoptimization event for JFR, UL and event system 2053 Method* tm = trap_method(); 2054 const char* reason_name = trap_reason_name(reason); 2055 const char* reason_action = trap_action_name(action); 2056 intptr_t pc = p2i(fr.pc()); 2057 2058 JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);) 2059 log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action); 2060 Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 2061 reason_name, reason_action, pc, 2062 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 2063 } 2064 2065 // Print a bunch of diagnostics, if requested. 2066 if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) { 2067 ResourceMark rm; 2068 ttyLocker ttyl; 2069 char buf[100]; 2070 if (xtty != NULL) { 2071 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 2072 os::current_thread_id(), 2073 format_trap_request(buf, sizeof(buf), trap_request)); 2074 #if INCLUDE_JVMCI 2075 if (speculation != 0) { 2076 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 2077 } 2078 #endif 2079 nm->log_identity(xtty); 2080 } 2081 Symbol* class_name = NULL; 2082 bool unresolved = false; 2083 if (unloaded_class_index >= 0) { 2084 constantPoolHandle constants (current, trap_method->constants()); 2085 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 2086 class_name = constants->klass_name_at(unloaded_class_index); 2087 unresolved = true; 2088 if (xtty != NULL) 2089 xtty->print(" unresolved='1'"); 2090 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 2091 class_name = constants->symbol_at(unloaded_class_index); 2092 } 2093 if (xtty != NULL) 2094 xtty->name(class_name); 2095 } 2096 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 2097 // Dump the relevant MDO state. 2098 // This is the deopt count for the current reason, any previous 2099 // reasons or recompiles seen at this point. 2100 int dcnt = trap_mdo->trap_count(reason); 2101 if (dcnt != 0) 2102 xtty->print(" count='%d'", dcnt); 2103 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 2104 int dos = (pdata == NULL)? 0: pdata->trap_state(); 2105 if (dos != 0) { 2106 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 2107 if (trap_state_is_recompiled(dos)) { 2108 int recnt2 = trap_mdo->overflow_recompile_count(); 2109 if (recnt2 != 0) 2110 xtty->print(" recompiles2='%d'", recnt2); 2111 } 2112 } 2113 } 2114 if (xtty != NULL) { 2115 xtty->stamp(); 2116 xtty->end_head(); 2117 } 2118 if (TraceDeoptimization) { // make noise on the tty 2119 stringStream st; 2120 st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string()); 2121 st.print(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"), 2122 trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id)); 2123 st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 2124 #if INCLUDE_JVMCI 2125 if (nm->is_nmethod()) { 2126 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 2127 if (installed_code_name != NULL) { 2128 st.print(" (JVMCI: installed code name=%s) ", installed_code_name); 2129 } 2130 } 2131 #endif 2132 st.print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 2133 p2i(fr.pc()), 2134 os::current_thread_id(), 2135 trap_reason_name(reason), 2136 trap_action_name(action), 2137 unloaded_class_index 2138 #if INCLUDE_JVMCI 2139 , debug_id 2140 #endif 2141 ); 2142 if (class_name != NULL) { 2143 st.print(unresolved ? " unresolved class: " : " symbol: "); 2144 class_name->print_symbol_on(&st); 2145 } 2146 st.cr(); 2147 tty->print_raw(st.freeze()); 2148 } 2149 if (xtty != NULL) { 2150 // Log the precise location of the trap. 2151 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 2152 xtty->begin_elem("jvms bci='%d'", sd->bci()); 2153 xtty->method(sd->method()); 2154 xtty->end_elem(); 2155 if (sd->is_top()) break; 2156 } 2157 xtty->tail("uncommon_trap"); 2158 } 2159 } 2160 // (End diagnostic printout.) 2161 2162 if (is_receiver_constraint_failure) { 2163 fatal("missing receiver type check"); 2164 } 2165 2166 // Load class if necessary 2167 if (unloaded_class_index >= 0) { 2168 constantPoolHandle constants(current, trap_method->constants()); 2169 load_class_by_index(constants, unloaded_class_index, THREAD); 2170 } 2171 2172 // Flush the nmethod if necessary and desirable. 2173 // 2174 // We need to avoid situations where we are re-flushing the nmethod 2175 // because of a hot deoptimization site. Repeated flushes at the same 2176 // point need to be detected by the compiler and avoided. If the compiler 2177 // cannot avoid them (or has a bug and "refuses" to avoid them), this 2178 // module must take measures to avoid an infinite cycle of recompilation 2179 // and deoptimization. There are several such measures: 2180 // 2181 // 1. If a recompilation is ordered a second time at some site X 2182 // and for the same reason R, the action is adjusted to 'reinterpret', 2183 // to give the interpreter time to exercise the method more thoroughly. 2184 // If this happens, the method's overflow_recompile_count is incremented. 2185 // 2186 // 2. If the compiler fails to reduce the deoptimization rate, then 2187 // the method's overflow_recompile_count will begin to exceed the set 2188 // limit PerBytecodeRecompilationCutoff. If this happens, the action 2189 // is adjusted to 'make_not_compilable', and the method is abandoned 2190 // to the interpreter. This is a performance hit for hot methods, 2191 // but is better than a disastrous infinite cycle of recompilations. 2192 // (Actually, only the method containing the site X is abandoned.) 2193 // 2194 // 3. In parallel with the previous measures, if the total number of 2195 // recompilations of a method exceeds the much larger set limit 2196 // PerMethodRecompilationCutoff, the method is abandoned. 2197 // This should only happen if the method is very large and has 2198 // many "lukewarm" deoptimizations. The code which enforces this 2199 // limit is elsewhere (class nmethod, class Method). 2200 // 2201 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 2202 // to recompile at each bytecode independently of the per-BCI cutoff. 2203 // 2204 // The decision to update code is up to the compiler, and is encoded 2205 // in the Action_xxx code. If the compiler requests Action_none 2206 // no trap state is changed, no compiled code is changed, and the 2207 // computation suffers along in the interpreter. 2208 // 2209 // The other action codes specify various tactics for decompilation 2210 // and recompilation. Action_maybe_recompile is the loosest, and 2211 // allows the compiled code to stay around until enough traps are seen, 2212 // and until the compiler gets around to recompiling the trapping method. 2213 // 2214 // The other actions cause immediate removal of the present code. 2215 2216 // Traps caused by injected profile shouldn't pollute trap counts. 2217 bool injected_profile_trap = trap_method->has_injected_profile() && 2218 (reason == Reason_intrinsic || reason == Reason_unreached); 2219 2220 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 2221 bool make_not_entrant = false; 2222 bool make_not_compilable = false; 2223 bool reprofile = false; 2224 switch (action) { 2225 case Action_none: 2226 // Keep the old code. 2227 update_trap_state = false; 2228 break; 2229 case Action_maybe_recompile: 2230 // Do not need to invalidate the present code, but we can 2231 // initiate another 2232 // Start compiler without (necessarily) invalidating the nmethod. 2233 // The system will tolerate the old code, but new code should be 2234 // generated when possible. 2235 break; 2236 case Action_reinterpret: 2237 // Go back into the interpreter for a while, and then consider 2238 // recompiling form scratch. 2239 make_not_entrant = true; 2240 // Reset invocation counter for outer most method. 2241 // This will allow the interpreter to exercise the bytecodes 2242 // for a while before recompiling. 2243 // By contrast, Action_make_not_entrant is immediate. 2244 // 2245 // Note that the compiler will track null_check, null_assert, 2246 // range_check, and class_check events and log them as if they 2247 // had been traps taken from compiled code. This will update 2248 // the MDO trap history so that the next compilation will 2249 // properly detect hot trap sites. 2250 reprofile = true; 2251 break; 2252 case Action_make_not_entrant: 2253 // Request immediate recompilation, and get rid of the old code. 2254 // Make them not entrant, so next time they are called they get 2255 // recompiled. Unloaded classes are loaded now so recompile before next 2256 // time they are called. Same for uninitialized. The interpreter will 2257 // link the missing class, if any. 2258 make_not_entrant = true; 2259 break; 2260 case Action_make_not_compilable: 2261 // Give up on compiling this method at all. 2262 make_not_entrant = true; 2263 make_not_compilable = true; 2264 break; 2265 default: 2266 ShouldNotReachHere(); 2267 } 2268 2269 // Setting +ProfileTraps fixes the following, on all platforms: 2270 // 4852688: ProfileInterpreter is off by default for ia64. The result is 2271 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 2272 // recompile relies on a MethodData* to record heroic opt failures. 2273 2274 // Whether the interpreter is producing MDO data or not, we also need 2275 // to use the MDO to detect hot deoptimization points and control 2276 // aggressive optimization. 2277 bool inc_recompile_count = false; 2278 ProfileData* pdata = NULL; 2279 if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != NULL) { 2280 assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity"); 2281 uint this_trap_count = 0; 2282 bool maybe_prior_trap = false; 2283 bool maybe_prior_recompile = false; 2284 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2285 #if INCLUDE_JVMCI 2286 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2287 #endif 2288 nm->method(), 2289 //outputs: 2290 this_trap_count, 2291 maybe_prior_trap, 2292 maybe_prior_recompile); 2293 // Because the interpreter also counts null, div0, range, and class 2294 // checks, these traps from compiled code are double-counted. 2295 // This is harmless; it just means that the PerXTrapLimit values 2296 // are in effect a little smaller than they look. 2297 2298 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2299 if (per_bc_reason != Reason_none) { 2300 // Now take action based on the partially known per-BCI history. 2301 if (maybe_prior_trap 2302 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2303 // If there are too many traps at this BCI, force a recompile. 2304 // This will allow the compiler to see the limit overflow, and 2305 // take corrective action, if possible. The compiler generally 2306 // does not use the exact PerBytecodeTrapLimit value, but instead 2307 // changes its tactics if it sees any traps at all. This provides 2308 // a little hysteresis, delaying a recompile until a trap happens 2309 // several times. 2310 // 2311 // Actually, since there is only one bit of counter per BCI, 2312 // the possible per-BCI counts are {0,1,(per-method count)}. 2313 // This produces accurate results if in fact there is only 2314 // one hot trap site, but begins to get fuzzy if there are 2315 // many sites. For example, if there are ten sites each 2316 // trapping two or more times, they each get the blame for 2317 // all of their traps. 2318 make_not_entrant = true; 2319 } 2320 2321 // Detect repeated recompilation at the same BCI, and enforce a limit. 2322 if (make_not_entrant && maybe_prior_recompile) { 2323 // More than one recompile at this point. 2324 inc_recompile_count = maybe_prior_trap; 2325 } 2326 } else { 2327 // For reasons which are not recorded per-bytecode, we simply 2328 // force recompiles unconditionally. 2329 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2330 make_not_entrant = true; 2331 } 2332 2333 // Go back to the compiler if there are too many traps in this method. 2334 if (this_trap_count >= per_method_trap_limit(reason)) { 2335 // If there are too many traps in this method, force a recompile. 2336 // This will allow the compiler to see the limit overflow, and 2337 // take corrective action, if possible. 2338 // (This condition is an unlikely backstop only, because the 2339 // PerBytecodeTrapLimit is more likely to take effect first, 2340 // if it is applicable.) 2341 make_not_entrant = true; 2342 } 2343 2344 // Here's more hysteresis: If there has been a recompile at 2345 // this trap point already, run the method in the interpreter 2346 // for a while to exercise it more thoroughly. 2347 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2348 reprofile = true; 2349 } 2350 } 2351 2352 // Take requested actions on the method: 2353 2354 // Recompile 2355 if (make_not_entrant) { 2356 if (!nm->make_not_entrant()) { 2357 return; // the call did not change nmethod's state 2358 } 2359 2360 if (pdata != NULL) { 2361 // Record the recompilation event, if any. 2362 int tstate0 = pdata->trap_state(); 2363 int tstate1 = trap_state_set_recompiled(tstate0, true); 2364 if (tstate1 != tstate0) 2365 pdata->set_trap_state(tstate1); 2366 } 2367 2368 #if INCLUDE_RTM_OPT 2369 // Restart collecting RTM locking abort statistic if the method 2370 // is recompiled for a reason other than RTM state change. 2371 // Assume that in new recompiled code the statistic could be different, 2372 // for example, due to different inlining. 2373 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2374 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2375 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2376 } 2377 #endif 2378 // For code aging we count traps separately here, using make_not_entrant() 2379 // as a guard against simultaneous deopts in multiple threads. 2380 if (reason == Reason_tenured && trap_mdo != NULL) { 2381 trap_mdo->inc_tenure_traps(); 2382 } 2383 } 2384 2385 if (inc_recompile_count) { 2386 trap_mdo->inc_overflow_recompile_count(); 2387 if ((uint)trap_mdo->overflow_recompile_count() > 2388 (uint)PerBytecodeRecompilationCutoff) { 2389 // Give up on the method containing the bad BCI. 2390 if (trap_method() == nm->method()) { 2391 make_not_compilable = true; 2392 } else { 2393 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2394 // But give grace to the enclosing nm->method(). 2395 } 2396 } 2397 } 2398 2399 // Reprofile 2400 if (reprofile) { 2401 CompilationPolicy::reprofile(trap_scope, nm->is_osr_method()); 2402 } 2403 2404 // Give up compiling 2405 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2406 assert(make_not_entrant, "consistent"); 2407 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2408 } 2409 2410 } // Free marked resources 2411 2412 } 2413 JRT_END 2414 2415 ProfileData* 2416 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2417 int trap_bci, 2418 Deoptimization::DeoptReason reason, 2419 bool update_total_trap_count, 2420 #if INCLUDE_JVMCI 2421 bool is_osr, 2422 #endif 2423 Method* compiled_method, 2424 //outputs: 2425 uint& ret_this_trap_count, 2426 bool& ret_maybe_prior_trap, 2427 bool& ret_maybe_prior_recompile) { 2428 bool maybe_prior_trap = false; 2429 bool maybe_prior_recompile = false; 2430 uint this_trap_count = 0; 2431 if (update_total_trap_count) { 2432 uint idx = reason; 2433 #if INCLUDE_JVMCI 2434 if (is_osr) { 2435 // Upper half of history array used for traps in OSR compilations 2436 idx += Reason_TRAP_HISTORY_LENGTH; 2437 } 2438 #endif 2439 uint prior_trap_count = trap_mdo->trap_count(idx); 2440 this_trap_count = trap_mdo->inc_trap_count(idx); 2441 2442 // If the runtime cannot find a place to store trap history, 2443 // it is estimated based on the general condition of the method. 2444 // If the method has ever been recompiled, or has ever incurred 2445 // a trap with the present reason , then this BCI is assumed 2446 // (pessimistically) to be the culprit. 2447 maybe_prior_trap = (prior_trap_count != 0); 2448 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2449 } 2450 ProfileData* pdata = NULL; 2451 2452 2453 // For reasons which are recorded per bytecode, we check per-BCI data. 2454 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2455 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2456 if (per_bc_reason != Reason_none) { 2457 // Find the profile data for this BCI. If there isn't one, 2458 // try to allocate one from the MDO's set of spares. 2459 // This will let us detect a repeated trap at this point. 2460 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2461 2462 if (pdata != NULL) { 2463 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2464 if (LogCompilation && xtty != NULL) { 2465 ttyLocker ttyl; 2466 // no more room for speculative traps in this MDO 2467 xtty->elem("speculative_traps_oom"); 2468 } 2469 } 2470 // Query the trap state of this profile datum. 2471 int tstate0 = pdata->trap_state(); 2472 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2473 maybe_prior_trap = false; 2474 if (!trap_state_is_recompiled(tstate0)) 2475 maybe_prior_recompile = false; 2476 2477 // Update the trap state of this profile datum. 2478 int tstate1 = tstate0; 2479 // Record the reason. 2480 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2481 // Store the updated state on the MDO, for next time. 2482 if (tstate1 != tstate0) 2483 pdata->set_trap_state(tstate1); 2484 } else { 2485 if (LogCompilation && xtty != NULL) { 2486 ttyLocker ttyl; 2487 // Missing MDP? Leave a small complaint in the log. 2488 xtty->elem("missing_mdp bci='%d'", trap_bci); 2489 } 2490 } 2491 } 2492 2493 // Return results: 2494 ret_this_trap_count = this_trap_count; 2495 ret_maybe_prior_trap = maybe_prior_trap; 2496 ret_maybe_prior_recompile = maybe_prior_recompile; 2497 return pdata; 2498 } 2499 2500 void 2501 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2502 ResourceMark rm; 2503 // Ignored outputs: 2504 uint ignore_this_trap_count; 2505 bool ignore_maybe_prior_trap; 2506 bool ignore_maybe_prior_recompile; 2507 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2508 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2509 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2510 query_update_method_data(trap_mdo, trap_bci, 2511 (DeoptReason)reason, 2512 update_total_counts, 2513 #if INCLUDE_JVMCI 2514 false, 2515 #endif 2516 NULL, 2517 ignore_this_trap_count, 2518 ignore_maybe_prior_trap, 2519 ignore_maybe_prior_recompile); 2520 } 2521 2522 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) { 2523 // Enable WXWrite: current function is called from methods compiled by C2 directly 2524 MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current)); 2525 2526 // Still in Java no safepoints 2527 { 2528 // This enters VM and may safepoint 2529 uncommon_trap_inner(current, trap_request); 2530 } 2531 HandleMark hm(current); 2532 return fetch_unroll_info_helper(current, exec_mode); 2533 } 2534 2535 // Local derived constants. 2536 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2537 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2538 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2539 2540 //---------------------------trap_state_reason--------------------------------- 2541 Deoptimization::DeoptReason 2542 Deoptimization::trap_state_reason(int trap_state) { 2543 // This assert provides the link between the width of DataLayout::trap_bits 2544 // and the encoding of "recorded" reasons. It ensures there are enough 2545 // bits to store all needed reasons in the per-BCI MDO profile. 2546 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2547 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2548 trap_state -= recompile_bit; 2549 if (trap_state == DS_REASON_MASK) { 2550 return Reason_many; 2551 } else { 2552 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2553 return (DeoptReason)trap_state; 2554 } 2555 } 2556 //-------------------------trap_state_has_reason------------------------------- 2557 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2558 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2559 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2560 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2561 trap_state -= recompile_bit; 2562 if (trap_state == DS_REASON_MASK) { 2563 return -1; // true, unspecifically (bottom of state lattice) 2564 } else if (trap_state == reason) { 2565 return 1; // true, definitely 2566 } else if (trap_state == 0) { 2567 return 0; // false, definitely (top of state lattice) 2568 } else { 2569 return 0; // false, definitely 2570 } 2571 } 2572 //-------------------------trap_state_add_reason------------------------------- 2573 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2574 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2575 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2576 trap_state -= recompile_bit; 2577 if (trap_state == DS_REASON_MASK) { 2578 return trap_state + recompile_bit; // already at state lattice bottom 2579 } else if (trap_state == reason) { 2580 return trap_state + recompile_bit; // the condition is already true 2581 } else if (trap_state == 0) { 2582 return reason + recompile_bit; // no condition has yet been true 2583 } else { 2584 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2585 } 2586 } 2587 //-----------------------trap_state_is_recompiled------------------------------ 2588 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2589 return (trap_state & DS_RECOMPILE_BIT) != 0; 2590 } 2591 //-----------------------trap_state_set_recompiled----------------------------- 2592 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2593 if (z) return trap_state | DS_RECOMPILE_BIT; 2594 else return trap_state & ~DS_RECOMPILE_BIT; 2595 } 2596 //---------------------------format_trap_state--------------------------------- 2597 // This is used for debugging and diagnostics, including LogFile output. 2598 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2599 int trap_state) { 2600 assert(buflen > 0, "sanity"); 2601 DeoptReason reason = trap_state_reason(trap_state); 2602 bool recomp_flag = trap_state_is_recompiled(trap_state); 2603 // Re-encode the state from its decoded components. 2604 int decoded_state = 0; 2605 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2606 decoded_state = trap_state_add_reason(decoded_state, reason); 2607 if (recomp_flag) 2608 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2609 // If the state re-encodes properly, format it symbolically. 2610 // Because this routine is used for debugging and diagnostics, 2611 // be robust even if the state is a strange value. 2612 size_t len; 2613 if (decoded_state != trap_state) { 2614 // Random buggy state that doesn't decode?? 2615 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2616 } else { 2617 len = jio_snprintf(buf, buflen, "%s%s", 2618 trap_reason_name(reason), 2619 recomp_flag ? " recompiled" : ""); 2620 } 2621 return buf; 2622 } 2623 2624 2625 //--------------------------------statics-------------------------------------- 2626 const char* Deoptimization::_trap_reason_name[] = { 2627 // Note: Keep this in sync. with enum DeoptReason. 2628 "none", 2629 "null_check", 2630 "null_assert" JVMCI_ONLY("_or_unreached0"), 2631 "range_check", 2632 "class_check", 2633 "array_check", 2634 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2635 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2636 "profile_predicate", 2637 "unloaded", 2638 "uninitialized", 2639 "initialized", 2640 "unreached", 2641 "unhandled", 2642 "constraint", 2643 "div0_check", 2644 "age", 2645 "predicate", 2646 "loop_limit_check", 2647 "speculate_class_check", 2648 "speculate_null_check", 2649 "speculate_null_assert", 2650 "rtm_state_change", 2651 "unstable_if", 2652 "unstable_fused_if", 2653 "receiver_constraint", 2654 #if INCLUDE_JVMCI 2655 "aliasing", 2656 "transfer_to_interpreter", 2657 "not_compiled_exception_handler", 2658 "unresolved", 2659 "jsr_mismatch", 2660 #endif 2661 "tenured" 2662 }; 2663 const char* Deoptimization::_trap_action_name[] = { 2664 // Note: Keep this in sync. with enum DeoptAction. 2665 "none", 2666 "maybe_recompile", 2667 "reinterpret", 2668 "make_not_entrant", 2669 "make_not_compilable" 2670 }; 2671 2672 const char* Deoptimization::trap_reason_name(int reason) { 2673 // Check that every reason has a name 2674 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2675 2676 if (reason == Reason_many) return "many"; 2677 if ((uint)reason < Reason_LIMIT) 2678 return _trap_reason_name[reason]; 2679 static char buf[20]; 2680 os::snprintf_checked(buf, sizeof(buf), "reason%d", reason); 2681 return buf; 2682 } 2683 const char* Deoptimization::trap_action_name(int action) { 2684 // Check that every action has a name 2685 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2686 2687 if ((uint)action < Action_LIMIT) 2688 return _trap_action_name[action]; 2689 static char buf[20]; 2690 os::snprintf_checked(buf, sizeof(buf), "action%d", action); 2691 return buf; 2692 } 2693 2694 // This is used for debugging and diagnostics, including LogFile output. 2695 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2696 int trap_request) { 2697 jint unloaded_class_index = trap_request_index(trap_request); 2698 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2699 const char* action = trap_action_name(trap_request_action(trap_request)); 2700 #if INCLUDE_JVMCI 2701 int debug_id = trap_request_debug_id(trap_request); 2702 #endif 2703 size_t len; 2704 if (unloaded_class_index < 0) { 2705 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2706 reason, action 2707 #if INCLUDE_JVMCI 2708 ,debug_id 2709 #endif 2710 ); 2711 } else { 2712 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2713 reason, action, unloaded_class_index 2714 #if INCLUDE_JVMCI 2715 ,debug_id 2716 #endif 2717 ); 2718 } 2719 return buf; 2720 } 2721 2722 juint Deoptimization::_deoptimization_hist 2723 [Deoptimization::Reason_LIMIT] 2724 [1 + Deoptimization::Action_LIMIT] 2725 [Deoptimization::BC_CASE_LIMIT] 2726 = {0}; 2727 2728 enum { 2729 LSB_BITS = 8, 2730 LSB_MASK = right_n_bits(LSB_BITS) 2731 }; 2732 2733 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2734 Bytecodes::Code bc) { 2735 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2736 assert(action >= 0 && action < Action_LIMIT, "oob"); 2737 _deoptimization_hist[Reason_none][0][0] += 1; // total 2738 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2739 juint* cases = _deoptimization_hist[reason][1+action]; 2740 juint* bc_counter_addr = NULL; 2741 juint bc_counter = 0; 2742 // Look for an unused counter, or an exact match to this BC. 2743 if (bc != Bytecodes::_illegal) { 2744 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2745 juint* counter_addr = &cases[bc_case]; 2746 juint counter = *counter_addr; 2747 if ((counter == 0 && bc_counter_addr == NULL) 2748 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2749 // this counter is either free or is already devoted to this BC 2750 bc_counter_addr = counter_addr; 2751 bc_counter = counter | bc; 2752 } 2753 } 2754 } 2755 if (bc_counter_addr == NULL) { 2756 // Overflow, or no given bytecode. 2757 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2758 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2759 } 2760 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2761 } 2762 2763 jint Deoptimization::total_deoptimization_count() { 2764 return _deoptimization_hist[Reason_none][0][0]; 2765 } 2766 2767 // Get the deopt count for a specific reason and a specific action. If either 2768 // one of 'reason' or 'action' is null, the method returns the sum of all 2769 // deoptimizations with the specific 'action' or 'reason' respectively. 2770 // If both arguments are null, the method returns the total deopt count. 2771 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2772 if (reason_str == NULL && action_str == NULL) { 2773 return total_deoptimization_count(); 2774 } 2775 juint counter = 0; 2776 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2777 if (reason_str == NULL || !strcmp(reason_str, trap_reason_name(reason))) { 2778 for (int action = 0; action < Action_LIMIT; action++) { 2779 if (action_str == NULL || !strcmp(action_str, trap_action_name(action))) { 2780 juint* cases = _deoptimization_hist[reason][1+action]; 2781 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2782 counter += cases[bc_case] >> LSB_BITS; 2783 } 2784 } 2785 } 2786 } 2787 } 2788 return counter; 2789 } 2790 2791 void Deoptimization::print_statistics() { 2792 juint total = total_deoptimization_count(); 2793 juint account = total; 2794 if (total != 0) { 2795 ttyLocker ttyl; 2796 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2797 tty->print_cr("Deoptimization traps recorded:"); 2798 #define PRINT_STAT_LINE(name, r) \ 2799 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2800 PRINT_STAT_LINE("total", total); 2801 // For each non-zero entry in the histogram, print the reason, 2802 // the action, and (if specifically known) the type of bytecode. 2803 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2804 for (int action = 0; action < Action_LIMIT; action++) { 2805 juint* cases = _deoptimization_hist[reason][1+action]; 2806 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2807 juint counter = cases[bc_case]; 2808 if (counter != 0) { 2809 char name[1*K]; 2810 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2811 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2812 bc = Bytecodes::_illegal; 2813 os::snprintf_checked(name, sizeof(name), "%s/%s/%s", 2814 trap_reason_name(reason), 2815 trap_action_name(action), 2816 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2817 juint r = counter >> LSB_BITS; 2818 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2819 account -= r; 2820 } 2821 } 2822 } 2823 } 2824 if (account != 0) { 2825 PRINT_STAT_LINE("unaccounted", account); 2826 } 2827 #undef PRINT_STAT_LINE 2828 if (xtty != NULL) xtty->tail("statistics"); 2829 } 2830 } 2831 2832 #else // COMPILER2_OR_JVMCI 2833 2834 2835 // Stubs for C1 only system. 2836 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2837 return false; 2838 } 2839 2840 const char* Deoptimization::trap_reason_name(int reason) { 2841 return "unknown"; 2842 } 2843 2844 jint Deoptimization::total_deoptimization_count() { 2845 return 0; 2846 } 2847 2848 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) { 2849 return 0; 2850 } 2851 2852 void Deoptimization::print_statistics() { 2853 // no output 2854 } 2855 2856 void 2857 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2858 // no update 2859 } 2860 2861 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2862 return 0; 2863 } 2864 2865 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2866 Bytecodes::Code bc) { 2867 // no update 2868 } 2869 2870 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2871 int trap_state) { 2872 jio_snprintf(buf, buflen, "#%d", trap_state); 2873 return buf; 2874 } 2875 2876 #endif // COMPILER2_OR_JVMCI