1 /* 2 * Copyright (c) 1997, 2019, 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 "jvm.h" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "code/nmethod.hpp" 32 #include "code/pcDesc.hpp" 33 #include "code/scopeDesc.hpp" 34 #include "interpreter/bytecode.hpp" 35 #include "interpreter/interpreter.hpp" 36 #include "interpreter/oopMapCache.hpp" 37 #include "memory/allocation.inline.hpp" 38 #include "memory/oopFactory.hpp" 39 #include "memory/resourceArea.hpp" 40 #include "memory/universe.hpp" 41 #include "oops/constantPool.hpp" 42 #include "oops/method.hpp" 43 #include "oops/objArrayKlass.hpp" 44 #include "oops/objArrayOop.inline.hpp" 45 #include "oops/oop.inline.hpp" 46 #include "oops/fieldStreams.hpp" 47 #include "oops/typeArrayOop.inline.hpp" 48 #include "oops/verifyOopClosure.hpp" 49 #include "prims/jvmtiThreadState.hpp" 50 #include "runtime/biasedLocking.hpp" 51 #include "runtime/compilationPolicy.hpp" 52 #include "runtime/deoptimization.hpp" 53 #include "runtime/fieldDescriptor.hpp" 54 #include "runtime/fieldDescriptor.inline.hpp" 55 #include "runtime/frame.inline.hpp" 56 #include "runtime/jniHandles.inline.hpp" 57 #include "runtime/handles.inline.hpp" 58 #include "runtime/interfaceSupport.inline.hpp" 59 #include "runtime/objectMonitor.inline.hpp" 60 #include "runtime/safepointVerifiers.hpp" 61 #include "runtime/sharedRuntime.hpp" 62 #include "runtime/signature.hpp" 63 #include "runtime/stubRoutines.hpp" 64 #include "runtime/thread.hpp" 65 #include "runtime/threadSMR.hpp" 66 #include "runtime/vframe.hpp" 67 #include "runtime/vframeArray.hpp" 68 #include "runtime/vframe_hp.hpp" 69 #include "utilities/events.hpp" 70 #include "utilities/preserveException.hpp" 71 #include "utilities/xmlstream.hpp" 72 73 74 bool DeoptimizationMarker::_is_active = false; 75 76 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 77 int caller_adjustment, 78 int caller_actual_parameters, 79 int number_of_frames, 80 intptr_t* frame_sizes, 81 address* frame_pcs, 82 BasicType return_type, 83 int exec_mode) { 84 _size_of_deoptimized_frame = size_of_deoptimized_frame; 85 _caller_adjustment = caller_adjustment; 86 _caller_actual_parameters = caller_actual_parameters; 87 _number_of_frames = number_of_frames; 88 _frame_sizes = frame_sizes; 89 _frame_pcs = frame_pcs; 90 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 91 _return_type = return_type; 92 _initial_info = 0; 93 // PD (x86 only) 94 _counter_temp = 0; 95 _unpack_kind = exec_mode; 96 _sender_sp_temp = 0; 97 98 _total_frame_sizes = size_of_frames(); 99 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 100 } 101 102 103 Deoptimization::UnrollBlock::~UnrollBlock() { 104 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 105 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 106 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 107 } 108 109 110 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 111 assert(register_number < RegisterMap::reg_count, "checking register number"); 112 return &_register_block[register_number * 2]; 113 } 114 115 116 117 int Deoptimization::UnrollBlock::size_of_frames() const { 118 // Acount first for the adjustment of the initial frame 119 int result = _caller_adjustment; 120 for (int index = 0; index < number_of_frames(); index++) { 121 result += frame_sizes()[index]; 122 } 123 return result; 124 } 125 126 127 void Deoptimization::UnrollBlock::print() { 128 ttyLocker ttyl; 129 tty->print_cr("UnrollBlock"); 130 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 131 tty->print( " frame_sizes: "); 132 for (int index = 0; index < number_of_frames(); index++) { 133 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 134 } 135 tty->cr(); 136 } 137 138 139 // In order to make fetch_unroll_info work properly with escape 140 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 141 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 142 // of previously eliminated objects occurs in realloc_objects, which is 143 // called from the method fetch_unroll_info_helper below. 144 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 145 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 146 // but makes the entry a little slower. There is however a little dance we have to 147 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 148 149 // fetch_unroll_info() is called at the beginning of the deoptimization 150 // handler. Note this fact before we start generating temporary frames 151 // that can confuse an asynchronous stack walker. This counter is 152 // decremented at the end of unpack_frames(). 153 if (TraceDeoptimization) { 154 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 155 } 156 thread->inc_in_deopt_handler(); 157 158 return fetch_unroll_info_helper(thread, exec_mode); 159 JRT_END 160 161 162 163 164 #if COMPILER2_OR_JVMCI 165 // Deoptimize objects, that is reallocate and relock them. Either because the holding 166 // compiled frame is being replaced by corresponding interpreter frames or because 167 // they are about to escape through JVMTI (exec_mode == Unpack_none). 168 bool Deoptimization::deoptimize_objects(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool& realloc_failures, int exec_mode) { 169 bool deoptimized_objects = false; 170 171 NOT_JVMCI(if (DoEscapeAnalysis || EliminateNestedLocks)) 172 { 173 frame deoptee = chunk->at(0)->fr(); 174 JavaThread* deoptee_thread = chunk->at(0)->thread(); 175 const RegisterMap* map = chunk->at(0)->register_map(); 176 177 assert(!JVMTIEscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once"); 178 assert(exec_mode == Unpack_none || (deoptee_thread == thread), "a frame can only be deoptimized by the owner thread"); 179 180 NOT_JVMCI(if (EliminateAllocations)) 181 { 182 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 183 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 184 185 if (objects != NULL) { 186 if (exec_mode == Unpack_none) { 187 assert(thread->thread_state() == _thread_in_vm, "assumption"); 188 Thread* THREAD = thread; 189 // Clear pending OOM if reallocation fails and return false, i.e. no objects deoptimized. 190 realloc_failures = realloc_objects(thread, &deoptee, map, objects, CHECK_AND_CLEAR_false); 191 } else { 192 // The flag return_oop() indicates call sites which return oop 193 // in compiled code. Such sites include java method calls, 194 // runtime calls (for example, used to allocate new objects/arrays 195 // on slow code path) and any other calls generated in compiled code. 196 // It is not guaranteed that we can get such information here only 197 // by analyzing bytecode in deoptimized frames. This is why this flag 198 // is set during method compilation (see Compile::Process_OopMap_Node()). 199 // If the previous frame was popped or if we are dispatching an exception, 200 // we don't have an oop result. 201 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt); 202 Handle return_value; 203 if (save_oop_result) { 204 // Reallocation may trigger GC. If deoptimization happened on return from 205 // call which returns oop we need to save it since it is not in oopmap. 206 oop result = deoptee.saved_oop_result(map); 207 assert(oopDesc::is_oop_or_null(result), "must be oop"); 208 return_value = Handle(thread, result); 209 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 210 if (TraceDeoptimization) { 211 ttyLocker ttyl; 212 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 213 } 214 } 215 JRT_BLOCK 216 realloc_failures = realloc_objects(thread, &deoptee, map, objects, THREAD); 217 JRT_END 218 if (save_oop_result) { 219 // Restore result. 220 deoptee.set_saved_oop_result(map, return_value()); 221 } 222 if (JVMTIEscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id())) { 223 // A concurrent JVMTI agent thread stop the current thread in the JRT_BLOCK above 224 // and deoptimized its objects 225 realloc_failures = false; // ignore realloc failures if any occurred 226 return false; // current thread did not deoptimize objects 227 } 228 } 229 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 230 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci(); 231 reassign_fields(&deoptee, map, objects, realloc_failures, skip_internal); 232 deoptimized_objects = true; 233 #ifndef PRODUCT 234 if (TraceDeoptimization) { 235 ttyLocker ttyl; 236 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread)); 237 print_objects(objects, realloc_failures); 238 } 239 #endif 240 } 241 } 242 NOT_JVMCI(if (EliminateLocks)) 243 { 244 #ifndef PRODUCT 245 bool first = true; 246 #endif 247 for (int i = 0; i < chunk->length(); i++) { 248 compiledVFrame* cvf = chunk->at(i); 249 assert (cvf->scope() != NULL,"expect only compiled java frames"); 250 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 251 if (monitors->is_nonempty()) { 252 bool relocked = relock_objects(thread, monitors, deoptee_thread, &deoptee, exec_mode, realloc_failures); 253 deoptimized_objects = deoptimized_objects || relocked; 254 #ifndef PRODUCT 255 if (PrintDeoptimizationDetails) { 256 ttyLocker ttyl; 257 for (int j = 0; j < monitors->length(); j++) { 258 MonitorInfo* mi = monitors->at(j); 259 if (mi->eliminated()) { 260 if (first) { 261 first = false; 262 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 263 } 264 if (exec_mode == Deoptimization::Unpack_none) { 265 ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor(); 266 if (monitor != NULL && (oop)monitor->object() == mi->owner()) { 267 tty->print_cr(" object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner())); 268 continue; 269 } 270 } 271 if (mi->owner_is_scalar_replaced()) { 272 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 273 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 274 } else { 275 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 276 } 277 } 278 } 279 } 280 #endif // !PRODUCT 281 } 282 } 283 } 284 } 285 return deoptimized_objects; 286 } 287 #endif // COMPILER2_OR_JVMCI 288 289 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 290 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 291 292 // Note: there is a safepoint safety issue here. No matter whether we enter 293 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 294 // the vframeArray is created. 295 // 296 297 // Allocate our special deoptimization ResourceMark 298 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 299 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 300 thread->set_deopt_mark(dmark); 301 302 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 303 RegisterMap map(thread, true); 304 RegisterMap dummy_map(thread, false); 305 // Now get the deoptee with a valid map 306 frame deoptee = stub_frame.sender(&map); 307 // Set the deoptee nmethod 308 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 309 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 310 thread->set_deopt_compiled_method(cm); 311 312 if (VerifyStack) { 313 thread->validate_frame_layout(); 314 } 315 316 // Create a growable array of VFrames where each VFrame represents an inlined 317 // Java frame. This storage is allocated with the usual system arena. 318 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 319 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 320 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 321 while (!vf->is_top()) { 322 assert(vf->is_compiled_frame(), "Wrong frame type"); 323 chunk->push(compiledVFrame::cast(vf)); 324 vf = vf->sender(); 325 } 326 assert(vf->is_compiled_frame(), "Wrong frame type"); 327 chunk->push(compiledVFrame::cast(vf)); 328 329 bool realloc_failures = false; 330 331 #if COMPILER2_OR_JVMCI 332 // Reallocate the non-escaping objects and restore their fields. Then 333 // relock objects if synchronization on them was eliminated. 334 if (!JVMTIEscapeBarrier::objs_are_deoptimized(thread, deoptee.id())) { 335 // objects are not yet deoptimized, do it now 336 deoptimize_objects(thread, chunk, realloc_failures, exec_mode); 337 } else { 338 // objects have been deoptimized already for JVMTI access 339 if (TraceDeoptimization) { 340 ttyLocker ttyl; 341 tty->print_cr("ALREADY DEOPTIMIZED OBJECTS for thread " INTPTR_FORMAT, p2i(thread)); 342 } 343 } 344 #endif // COMPILER2_OR_JVMCI 345 346 ScopeDesc* trap_scope = chunk->at(0)->scope(); 347 Handle exceptionObject; 348 if (trap_scope->rethrow_exception()) { 349 if (PrintDeoptimizationDetails) { 350 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()); 351 } 352 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 353 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 354 ScopeValue* topOfStack = expressions->top(); 355 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 356 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 357 } 358 359 // Ensure that no safepoint is taken after pointers have been stored 360 // in fields of rematerialized objects. If a safepoint occurs from here on 361 // out the java state residing in the vframeArray will be missed. 362 NoSafepointVerifier no_safepoint; 363 364 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 365 #if COMPILER2_OR_JVMCI 366 if (realloc_failures) { 367 pop_frames_failed_reallocs(thread, array); 368 } 369 #endif 370 371 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 372 thread->set_vframe_array_head(array); 373 374 // Now that the vframeArray has been created if we have any deferred local writes 375 // added by jvmti then we can free up that structure as the data is now in the 376 // vframeArray 377 378 if (thread->deferred_locals() != NULL) { 379 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 380 int i = 0; 381 do { 382 // Because of inlining we could have multiple vframes for a single frame 383 // and several of the vframes could have deferred writes. Find them all. 384 if (list->at(i)->id() == array->original().id()) { 385 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 386 list->remove_at(i); 387 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 388 delete dlv; 389 } else { 390 i++; 391 } 392 } while ( i < list->length() ); 393 if (list->length() == 0) { 394 JvmtiDeferredUpdates* updates = thread->deferred_updates(); 395 thread->reset_deferred_updates(); 396 // free deferred updates. 397 delete updates; 398 } 399 400 } 401 402 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 403 CodeBlob* cb = stub_frame.cb(); 404 // Verify we have the right vframeArray 405 assert(cb->frame_size() >= 0, "Unexpected frame size"); 406 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 407 408 // If the deopt call site is a MethodHandle invoke call site we have 409 // to adjust the unpack_sp. 410 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 411 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 412 unpack_sp = deoptee.unextended_sp(); 413 414 #ifdef ASSERT 415 assert(cb->is_deoptimization_stub() || 416 cb->is_uncommon_trap_stub() || 417 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 418 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 419 "unexpected code blob: %s", cb->name()); 420 #endif 421 422 // This is a guarantee instead of an assert because if vframe doesn't match 423 // we will unpack the wrong deoptimized frame and wind up in strange places 424 // where it will be very difficult to figure out what went wrong. Better 425 // to die an early death here than some very obscure death later when the 426 // trail is cold. 427 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 428 // in that it will fail to detect a problem when there is one. This needs 429 // more work in tiger timeframe. 430 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 431 432 int number_of_frames = array->frames(); 433 434 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 435 // virtual activation, which is the reverse of the elements in the vframes array. 436 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 437 // +1 because we always have an interpreter return address for the final slot. 438 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 439 int popframe_extra_args = 0; 440 // Create an interpreter return address for the stub to use as its return 441 // address so the skeletal frames are perfectly walkable 442 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 443 444 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 445 // activation be put back on the expression stack of the caller for reexecution 446 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 447 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 448 } 449 450 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 451 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 452 // than simply use array->sender.pc(). This requires us to walk the current set of frames 453 // 454 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 455 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 456 457 // It's possible that the number of parameters at the call site is 458 // different than number of arguments in the callee when method 459 // handles are used. If the caller is interpreted get the real 460 // value so that the proper amount of space can be added to it's 461 // frame. 462 bool caller_was_method_handle = false; 463 if (deopt_sender.is_interpreted_frame()) { 464 methodHandle method = deopt_sender.interpreter_frame_method(); 465 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 466 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 467 // Method handle invokes may involve fairly arbitrary chains of 468 // calls so it's impossible to know how much actual space the 469 // caller has for locals. 470 caller_was_method_handle = true; 471 } 472 } 473 474 // 475 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 476 // frame_sizes/frame_pcs[1] next oldest frame (int) 477 // frame_sizes/frame_pcs[n] youngest frame (int) 478 // 479 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 480 // owns the space for the return address to it's caller). Confusing ain't it. 481 // 482 // The vframe array can address vframes with indices running from 483 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 484 // When we create the skeletal frames we need the oldest frame to be in the zero slot 485 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 486 // so things look a little strange in this loop. 487 // 488 int callee_parameters = 0; 489 int callee_locals = 0; 490 for (int index = 0; index < array->frames(); index++ ) { 491 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 492 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 493 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 494 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 495 callee_locals, 496 index == 0, 497 popframe_extra_args); 498 // This pc doesn't have to be perfect just good enough to identify the frame 499 // as interpreted so the skeleton frame will be walkable 500 // The correct pc will be set when the skeleton frame is completely filled out 501 // The final pc we store in the loop is wrong and will be overwritten below 502 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 503 504 callee_parameters = array->element(index)->method()->size_of_parameters(); 505 callee_locals = array->element(index)->method()->max_locals(); 506 popframe_extra_args = 0; 507 } 508 509 // Compute whether the root vframe returns a float or double value. 510 BasicType return_type; 511 { 512 methodHandle method(thread, array->element(0)->method()); 513 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 514 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 515 } 516 517 // Compute information for handling adapters and adjusting the frame size of the caller. 518 int caller_adjustment = 0; 519 520 // Compute the amount the oldest interpreter frame will have to adjust 521 // its caller's stack by. If the caller is a compiled frame then 522 // we pretend that the callee has no parameters so that the 523 // extension counts for the full amount of locals and not just 524 // locals-parms. This is because without a c2i adapter the parm 525 // area as created by the compiled frame will not be usable by 526 // the interpreter. (Depending on the calling convention there 527 // may not even be enough space). 528 529 // QQQ I'd rather see this pushed down into last_frame_adjust 530 // and have it take the sender (aka caller). 531 532 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 533 caller_adjustment = last_frame_adjust(0, callee_locals); 534 } else if (callee_locals > callee_parameters) { 535 // The caller frame may need extending to accommodate 536 // non-parameter locals of the first unpacked interpreted frame. 537 // Compute that adjustment. 538 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 539 } 540 541 // If the sender is deoptimized the we must retrieve the address of the handler 542 // since the frame will "magically" show the original pc before the deopt 543 // and we'd undo the deopt. 544 545 frame_pcs[0] = deopt_sender.raw_pc(); 546 547 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 548 549 #if INCLUDE_JVMCI 550 if (exceptionObject() != NULL) { 551 thread->set_exception_oop(exceptionObject()); 552 exec_mode = Unpack_exception; 553 } 554 #endif 555 556 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 557 assert(thread->has_pending_exception(), "should have thrown OOME"); 558 thread->set_exception_oop(thread->pending_exception()); 559 thread->clear_pending_exception(); 560 exec_mode = Unpack_exception; 561 } 562 563 #if INCLUDE_JVMCI 564 if (thread->frames_to_pop_failed_realloc() > 0) { 565 thread->set_pending_monitorenter(false); 566 } 567 #endif 568 569 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 570 caller_adjustment * BytesPerWord, 571 caller_was_method_handle ? 0 : callee_parameters, 572 number_of_frames, 573 frame_sizes, 574 frame_pcs, 575 return_type, 576 exec_mode); 577 // On some platforms, we need a way to pass some platform dependent 578 // information to the unpacking code so the skeletal frames come out 579 // correct (initial fp value, unextended sp, ...) 580 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 581 582 if (array->frames() > 1) { 583 if (VerifyStack && TraceDeoptimization) { 584 ttyLocker ttyl; 585 tty->print_cr("Deoptimizing method containing inlining"); 586 } 587 } 588 589 array->set_unroll_block(info); 590 return info; 591 } 592 593 // Called to cleanup deoptimization data structures in normal case 594 // after unpacking to stack and when stack overflow error occurs 595 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 596 vframeArray *array) { 597 598 // Get array if coming from exception 599 if (array == NULL) { 600 array = thread->vframe_array_head(); 601 } 602 thread->set_vframe_array_head(NULL); 603 604 // Free the previous UnrollBlock 605 vframeArray* old_array = thread->vframe_array_last(); 606 thread->set_vframe_array_last(array); 607 608 if (old_array != NULL) { 609 UnrollBlock* old_info = old_array->unroll_block(); 610 old_array->set_unroll_block(NULL); 611 delete old_info; 612 delete old_array; 613 } 614 615 // Deallocate any resource creating in this routine and any ResourceObjs allocated 616 // inside the vframeArray (StackValueCollections) 617 618 delete thread->deopt_mark(); 619 thread->set_deopt_mark(NULL); 620 thread->set_deopt_compiled_method(NULL); 621 622 623 if (JvmtiExport::can_pop_frame()) { 624 #ifndef CC_INTERP 625 // Regardless of whether we entered this routine with the pending 626 // popframe condition bit set, we should always clear it now 627 thread->clear_popframe_condition(); 628 #else 629 // C++ interpreter will clear has_pending_popframe when it enters 630 // with method_resume. For deopt_resume2 we clear it now. 631 if (thread->popframe_forcing_deopt_reexecution()) 632 thread->clear_popframe_condition(); 633 #endif /* CC_INTERP */ 634 } 635 636 // unpack_frames() is called at the end of the deoptimization handler 637 // and (in C2) at the end of the uncommon trap handler. Note this fact 638 // so that an asynchronous stack walker can work again. This counter is 639 // incremented at the beginning of fetch_unroll_info() and (in C2) at 640 // the beginning of uncommon_trap(). 641 thread->dec_in_deopt_handler(); 642 } 643 644 // Moved from cpu directories because none of the cpus has callee save values. 645 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 646 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 647 648 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 649 // the days we had adapter frames. When we deoptimize a situation where a 650 // compiled caller calls a compiled caller will have registers it expects 651 // to survive the call to the callee. If we deoptimize the callee the only 652 // way we can restore these registers is to have the oldest interpreter 653 // frame that we create restore these values. That is what this routine 654 // will accomplish. 655 656 // At the moment we have modified c2 to not have any callee save registers 657 // so this problem does not exist and this routine is just a place holder. 658 659 assert(f->is_interpreted_frame(), "must be interpreted"); 660 } 661 662 // Return BasicType of value being returned 663 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 664 665 // We are already active in the special DeoptResourceMark any ResourceObj's we 666 // allocate will be freed at the end of the routine. 667 668 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 669 // but makes the entry a little slower. There is however a little dance we have to 670 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 671 ResetNoHandleMark rnhm; // No-op in release/product versions 672 HandleMark hm; 673 674 frame stub_frame = thread->last_frame(); 675 676 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 677 // must point to the vframeArray for the unpack frame. 678 vframeArray* array = thread->vframe_array_head(); 679 680 #ifndef PRODUCT 681 if (TraceDeoptimization) { 682 ttyLocker ttyl; 683 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 684 p2i(thread), p2i(array), exec_mode); 685 } 686 #endif 687 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 688 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 689 690 UnrollBlock* info = array->unroll_block(); 691 692 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 693 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 694 695 BasicType bt = info->return_type(); 696 697 // If we have an exception pending, claim that the return type is an oop 698 // so the deopt_blob does not overwrite the exception_oop. 699 700 if (exec_mode == Unpack_exception) 701 bt = T_OBJECT; 702 703 // Cleanup thread deopt data 704 cleanup_deopt_info(thread, array); 705 706 #ifndef PRODUCT 707 if (VerifyStack) { 708 ResourceMark res_mark; 709 // Clear pending exception to not break verification code (restored afterwards) 710 PRESERVE_EXCEPTION_MARK; 711 712 thread->validate_frame_layout(); 713 714 // Verify that the just-unpacked frames match the interpreter's 715 // notions of expression stack and locals 716 vframeArray* cur_array = thread->vframe_array_last(); 717 RegisterMap rm(thread, false); 718 rm.set_include_argument_oops(false); 719 bool is_top_frame = true; 720 int callee_size_of_parameters = 0; 721 int callee_max_locals = 0; 722 for (int i = 0; i < cur_array->frames(); i++) { 723 vframeArrayElement* el = cur_array->element(i); 724 frame* iframe = el->iframe(); 725 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 726 727 // Get the oop map for this bci 728 InterpreterOopMap mask; 729 int cur_invoke_parameter_size = 0; 730 bool try_next_mask = false; 731 int next_mask_expression_stack_size = -1; 732 int top_frame_expression_stack_adjustment = 0; 733 methodHandle mh(thread, iframe->interpreter_frame_method()); 734 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 735 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 736 int max_bci = mh->code_size(); 737 // Get to the next bytecode if possible 738 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 739 // Check to see if we can grab the number of outgoing arguments 740 // at an uncommon trap for an invoke (where the compiler 741 // generates debug info before the invoke has executed) 742 Bytecodes::Code cur_code = str.next(); 743 if (Bytecodes::is_invoke(cur_code)) { 744 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 745 cur_invoke_parameter_size = invoke.size_of_parameters(); 746 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 747 callee_size_of_parameters++; 748 } 749 } 750 if (str.bci() < max_bci) { 751 Bytecodes::Code next_code = str.next(); 752 if (next_code >= 0) { 753 // The interpreter oop map generator reports results before 754 // the current bytecode has executed except in the case of 755 // calls. It seems to be hard to tell whether the compiler 756 // has emitted debug information matching the "state before" 757 // a given bytecode or the state after, so we try both 758 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 759 // Get expression stack size for the next bytecode 760 InterpreterOopMap next_mask; 761 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 762 next_mask_expression_stack_size = next_mask.expression_stack_size(); 763 if (Bytecodes::is_invoke(next_code)) { 764 Bytecode_invoke invoke(mh, str.bci()); 765 next_mask_expression_stack_size += invoke.size_of_parameters(); 766 } 767 // Need to subtract off the size of the result type of 768 // the bytecode because this is not described in the 769 // debug info but returned to the interpreter in the TOS 770 // caching register 771 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 772 if (bytecode_result_type != T_ILLEGAL) { 773 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 774 } 775 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 776 try_next_mask = true; 777 } 778 } 779 } 780 781 // Verify stack depth and oops in frame 782 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 783 if (!( 784 /* SPARC */ 785 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 786 /* x86 */ 787 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 788 (try_next_mask && 789 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 790 top_frame_expression_stack_adjustment))) || 791 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 792 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 793 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 794 )) { 795 { 796 ttyLocker ttyl; 797 798 // Print out some information that will help us debug the problem 799 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 800 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 801 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 802 iframe->interpreter_frame_expression_stack_size()); 803 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 804 tty->print_cr(" try_next_mask = %d", try_next_mask); 805 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 806 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 807 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 808 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 809 tty->print_cr(" exec_mode = %d", exec_mode); 810 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 811 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 812 tty->print_cr(" Interpreted frames:"); 813 for (int k = 0; k < cur_array->frames(); k++) { 814 vframeArrayElement* el = cur_array->element(k); 815 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 816 } 817 cur_array->print_on_2(tty); 818 } // release tty lock before calling guarantee 819 guarantee(false, "wrong number of expression stack elements during deopt"); 820 } 821 VerifyOopClosure verify; 822 iframe->oops_interpreted_do(&verify, &rm, false); 823 callee_size_of_parameters = mh->size_of_parameters(); 824 callee_max_locals = mh->max_locals(); 825 is_top_frame = false; 826 } 827 } 828 #endif /* !PRODUCT */ 829 830 831 return bt; 832 JRT_END 833 834 835 int Deoptimization::deoptimize_dependents() { 836 Threads::deoptimized_wrt_marked_nmethods(); 837 return 0; 838 } 839 840 Deoptimization::DeoptAction Deoptimization::_unloaded_action 841 = Deoptimization::Action_reinterpret; 842 843 844 845 #if INCLUDE_JVMCI || INCLUDE_AOT 846 template<typename CacheType> 847 class BoxCacheBase : public CHeapObj<mtCompiler> { 848 protected: 849 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { 850 ResourceMark rm; 851 char* klass_name_str = klass_name->as_C_string(); 852 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); 853 guarantee(k != NULL, "%s must be loaded", klass_name_str); 854 InstanceKlass* ik = InstanceKlass::cast(k); 855 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 856 CacheType::compute_offsets(ik); 857 return ik; 858 } 859 }; 860 861 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 862 PrimitiveType _low; 863 PrimitiveType _high; 864 jobject _cache; 865 protected: 866 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 867 BoxCache(Thread* thread) { 868 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); 869 objArrayOop cache = CacheType::cache(ik); 870 assert(cache->length() > 0, "Empty cache"); 871 _low = BoxType::value(cache->obj_at(0)); 872 _high = _low + cache->length() - 1; 873 _cache = JNIHandles::make_global(Handle(thread, cache)); 874 } 875 ~BoxCache() { 876 JNIHandles::destroy_global(_cache); 877 } 878 public: 879 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 880 if (_singleton == NULL) { 881 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 882 if (!Atomic::replace_if_null(s, &_singleton)) { 883 delete s; 884 } 885 } 886 return _singleton; 887 } 888 oop lookup(PrimitiveType value) { 889 if (_low <= value && value <= _high) { 890 int offset = value - _low; 891 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 892 } 893 return NULL; 894 } 895 oop lookup_raw(intptr_t raw_value) { 896 // Have to cast to avoid little/big-endian problems. 897 if (sizeof(PrimitiveType) > sizeof(jint)) { 898 jlong value = (jlong)raw_value; 899 return lookup(value); 900 } 901 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 902 return lookup(value); 903 } 904 }; 905 906 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 907 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 908 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 909 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 910 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 911 912 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 913 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 914 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 915 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 916 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 917 918 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 919 jobject _true_cache; 920 jobject _false_cache; 921 protected: 922 static BooleanBoxCache *_singleton; 923 BooleanBoxCache(Thread *thread) { 924 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); 925 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 926 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 927 } 928 ~BooleanBoxCache() { 929 JNIHandles::destroy_global(_true_cache); 930 JNIHandles::destroy_global(_false_cache); 931 } 932 public: 933 static BooleanBoxCache* singleton(Thread* thread) { 934 if (_singleton == NULL) { 935 BooleanBoxCache* s = new BooleanBoxCache(thread); 936 if (!Atomic::replace_if_null(s, &_singleton)) { 937 delete s; 938 } 939 } 940 return _singleton; 941 } 942 oop lookup_raw(intptr_t raw_value) { 943 // Have to cast to avoid little/big-endian problems. 944 jboolean value = (jboolean)*((jint*)&raw_value); 945 return lookup(value); 946 } 947 oop lookup(jboolean value) { 948 if (value != 0) { 949 return JNIHandles::resolve_non_null(_true_cache); 950 } 951 return JNIHandles::resolve_non_null(_false_cache); 952 } 953 }; 954 955 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 956 957 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, const RegisterMap* reg_map, TRAPS) { 958 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 959 BasicType box_type = SystemDictionary::box_klass_type(k); 960 if (box_type != T_OBJECT) { 961 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 962 switch(box_type) { 963 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 964 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 965 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 966 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 967 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 968 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 969 default:; 970 } 971 } 972 return NULL; 973 } 974 #endif // INCLUDE_JVMCI || INCLUDE_AOT 975 976 #if COMPILER2_OR_JVMCI 977 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, const RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 978 Handle pending_exception(THREAD, thread->pending_exception()); 979 const char* exception_file = thread->exception_file(); 980 int exception_line = thread->exception_line(); 981 thread->clear_pending_exception(); 982 983 bool failures = false; 984 985 for (int i = 0; i < objects->length(); i++) { 986 assert(objects->at(i)->is_object(), "invalid debug information"); 987 ObjectValue* sv = (ObjectValue*) objects->at(i); 988 989 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 990 oop obj = NULL; 991 992 if (k->is_instance_klass()) { 993 #if INCLUDE_JVMCI || INCLUDE_AOT 994 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 995 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 996 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 997 obj = get_cached_box(abv, fr, reg_map, THREAD); 998 if (obj != NULL) { 999 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 1000 abv->set_cached(true); 1001 } 1002 } 1003 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1004 InstanceKlass* ik = InstanceKlass::cast(k); 1005 if (obj == NULL) { 1006 obj = ik->allocate_instance(THREAD); 1007 } 1008 } else if (k->is_typeArray_klass()) { 1009 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1010 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1011 int len = sv->field_size() / type2size[ak->element_type()]; 1012 obj = ak->allocate(len, THREAD); 1013 } else if (k->is_objArray_klass()) { 1014 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1015 obj = ak->allocate(sv->field_size(), THREAD); 1016 } 1017 1018 if (obj == NULL) { 1019 failures = true; 1020 } 1021 1022 assert(sv->value().is_null(), "redundant reallocation"); 1023 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1024 CLEAR_PENDING_EXCEPTION; 1025 sv->set_value(obj); 1026 } 1027 1028 if (failures) { 1029 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1030 } else if (pending_exception.not_null()) { 1031 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1032 } 1033 1034 return failures; 1035 } 1036 1037 // restore elements of an eliminated type array 1038 void Deoptimization::reassign_type_array_elements(frame* fr, const RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1039 int index = 0; 1040 intptr_t val; 1041 1042 for (int i = 0; i < sv->field_size(); i++) { 1043 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1044 switch(type) { 1045 case T_LONG: case T_DOUBLE: { 1046 assert(value->type() == T_INT, "Agreement."); 1047 StackValue* low = 1048 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1049 #ifdef _LP64 1050 jlong res = (jlong)low->get_int(); 1051 #else 1052 #ifdef SPARC 1053 // For SPARC we have to swap high and low words. 1054 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1055 #else 1056 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1057 #endif //SPARC 1058 #endif 1059 obj->long_at_put(index, res); 1060 break; 1061 } 1062 1063 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1064 case T_INT: case T_FLOAT: { // 4 bytes. 1065 assert(value->type() == T_INT, "Agreement."); 1066 bool big_value = false; 1067 if (i + 1 < sv->field_size() && type == T_INT) { 1068 if (sv->field_at(i)->is_location()) { 1069 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1070 if (type == Location::dbl || type == Location::lng) { 1071 big_value = true; 1072 } 1073 } else if (sv->field_at(i)->is_constant_int()) { 1074 ScopeValue* next_scope_field = sv->field_at(i + 1); 1075 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1076 big_value = true; 1077 } 1078 } 1079 } 1080 1081 if (big_value) { 1082 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1083 #ifdef _LP64 1084 jlong res = (jlong)low->get_int(); 1085 #else 1086 #ifdef SPARC 1087 // For SPARC we have to swap high and low words. 1088 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1089 #else 1090 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1091 #endif //SPARC 1092 #endif 1093 obj->int_at_put(index, (jint)*((jint*)&res)); 1094 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1095 } else { 1096 val = value->get_int(); 1097 obj->int_at_put(index, (jint)*((jint*)&val)); 1098 } 1099 break; 1100 } 1101 1102 case T_SHORT: 1103 assert(value->type() == T_INT, "Agreement."); 1104 val = value->get_int(); 1105 obj->short_at_put(index, (jshort)*((jint*)&val)); 1106 break; 1107 1108 case T_CHAR: 1109 assert(value->type() == T_INT, "Agreement."); 1110 val = value->get_int(); 1111 obj->char_at_put(index, (jchar)*((jint*)&val)); 1112 break; 1113 1114 case T_BYTE: 1115 assert(value->type() == T_INT, "Agreement."); 1116 val = value->get_int(); 1117 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1118 break; 1119 1120 case T_BOOLEAN: 1121 assert(value->type() == T_INT, "Agreement."); 1122 val = value->get_int(); 1123 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1124 break; 1125 1126 default: 1127 ShouldNotReachHere(); 1128 } 1129 index++; 1130 } 1131 } 1132 1133 1134 // restore fields of an eliminated object array 1135 void Deoptimization::reassign_object_array_elements(frame* fr, const RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1136 for (int i = 0; i < sv->field_size(); i++) { 1137 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1138 assert(value->type() == T_OBJECT, "object element expected"); 1139 obj->obj_at_put(i, value->get_obj()()); 1140 } 1141 } 1142 1143 class ReassignedField { 1144 public: 1145 int _offset; 1146 BasicType _type; 1147 public: 1148 ReassignedField() { 1149 _offset = 0; 1150 _type = T_ILLEGAL; 1151 } 1152 }; 1153 1154 int compare(ReassignedField* left, ReassignedField* right) { 1155 return left->_offset - right->_offset; 1156 } 1157 1158 // Restore fields of an eliminated instance object using the same field order 1159 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1160 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, const RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1161 if (klass->superklass() != NULL) { 1162 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); 1163 } 1164 1165 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1166 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1167 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1168 ReassignedField field; 1169 field._offset = fs.offset(); 1170 field._type = FieldType::basic_type(fs.signature()); 1171 fields->append(field); 1172 } 1173 } 1174 fields->sort(compare); 1175 for (int i = 0; i < fields->length(); i++) { 1176 intptr_t val; 1177 ScopeValue* scope_field = sv->field_at(svIndex); 1178 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1179 int offset = fields->at(i)._offset; 1180 BasicType type = fields->at(i)._type; 1181 switch (type) { 1182 case T_OBJECT: case T_ARRAY: 1183 assert(value->type() == T_OBJECT, "Agreement."); 1184 obj->obj_field_put(offset, value->get_obj()()); 1185 break; 1186 1187 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1188 case T_INT: case T_FLOAT: { // 4 bytes. 1189 assert(value->type() == T_INT, "Agreement."); 1190 bool big_value = false; 1191 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1192 if (scope_field->is_location()) { 1193 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1194 if (type == Location::dbl || type == Location::lng) { 1195 big_value = true; 1196 } 1197 } 1198 if (scope_field->is_constant_int()) { 1199 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1200 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1201 big_value = true; 1202 } 1203 } 1204 } 1205 1206 if (big_value) { 1207 i++; 1208 assert(i < fields->length(), "second T_INT field needed"); 1209 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1210 } else { 1211 val = value->get_int(); 1212 obj->int_field_put(offset, (jint)*((jint*)&val)); 1213 break; 1214 } 1215 } 1216 /* no break */ 1217 1218 case T_LONG: case T_DOUBLE: { 1219 assert(value->type() == T_INT, "Agreement."); 1220 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1221 #ifdef _LP64 1222 jlong res = (jlong)low->get_int(); 1223 #else 1224 #ifdef SPARC 1225 // For SPARC we have to swap high and low words. 1226 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1227 #else 1228 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1229 #endif //SPARC 1230 #endif 1231 obj->long_field_put(offset, res); 1232 break; 1233 } 1234 1235 case T_SHORT: 1236 assert(value->type() == T_INT, "Agreement."); 1237 val = value->get_int(); 1238 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1239 break; 1240 1241 case T_CHAR: 1242 assert(value->type() == T_INT, "Agreement."); 1243 val = value->get_int(); 1244 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1245 break; 1246 1247 case T_BYTE: 1248 assert(value->type() == T_INT, "Agreement."); 1249 val = value->get_int(); 1250 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1251 break; 1252 1253 case T_BOOLEAN: 1254 assert(value->type() == T_INT, "Agreement."); 1255 val = value->get_int(); 1256 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1257 break; 1258 1259 default: 1260 ShouldNotReachHere(); 1261 } 1262 svIndex++; 1263 } 1264 return svIndex; 1265 } 1266 1267 // restore fields of all eliminated objects and arrays 1268 void Deoptimization::reassign_fields(frame* fr, const RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1269 for (int i = 0; i < objects->length(); i++) { 1270 ObjectValue* sv = (ObjectValue*) objects->at(i); 1271 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1272 Handle obj = sv->value(); 1273 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1274 if (PrintDeoptimizationDetails) { 1275 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1276 } 1277 if (obj.is_null()) { 1278 continue; 1279 } 1280 #if INCLUDE_JVMCI || INCLUDE_AOT 1281 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1282 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1283 continue; 1284 } 1285 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1286 if (k->is_instance_klass()) { 1287 InstanceKlass* ik = InstanceKlass::cast(k); 1288 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1289 } else if (k->is_typeArray_klass()) { 1290 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1291 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1292 } else if (k->is_objArray_klass()) { 1293 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1294 } 1295 } 1296 } 1297 1298 1299 // relock objects for which synchronization was eliminated 1300 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors, 1301 JavaThread* deoptee_thread, frame* fr, int exec_mode, bool realloc_failures) { 1302 bool relocked_objects = false; 1303 for (int i = 0; i < monitors->length(); i++) { 1304 MonitorInfo* mon_info = monitors->at(i); 1305 if (mon_info->eliminated()) { 1306 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1307 relocked_objects = true; 1308 if (!mon_info->owner_is_scalar_replaced()) { 1309 Handle obj(thread, mon_info->owner()); 1310 markWord mark = obj->mark(); 1311 if (UseBiasedLocking && mark.has_bias_pattern()) { 1312 // New allocated objects may have the mark set to anonymously biased. 1313 // Also the deoptimized method may called methods with synchronization 1314 // where the thread-local object is bias locked to the current thread. 1315 assert(mark.is_biased_anonymously() || 1316 mark.biased_locker() == deoptee_thread, "should be locked to current thread"); 1317 // Reset mark word to unbiased prototype. 1318 markWord unbiased_prototype = markWord::prototype().set_age(mark.age()); 1319 obj->set_mark(unbiased_prototype); 1320 } else if (exec_mode == Unpack_none) { 1321 if (mark.has_locker() && fr->sp() > (intptr_t*)mark.locker()) { 1322 // With exec_mode == Unpack_none obj may be thread local and locked in 1323 // a callee frame. In this case the bias was revoked before. 1324 // Make the lock in the callee a recursive lock and restore the displaced header. 1325 markWord dmw = mark.displaced_mark_helper(); 1326 mark.locker()->set_displaced_header(markWord::encode((BasicLock*) NULL)); 1327 obj->set_mark(dmw); 1328 } 1329 if (mark.has_monitor()) { 1330 // defer relocking if the deoptee thread is currently waiting for obj 1331 ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor(); 1332 if (waiting_monitor != NULL && (oop)waiting_monitor->object() == obj()) { 1333 mon_info->lock()->set_displaced_header(markWord::unused_mark()); 1334 deoptee_thread->inc_relock_count_after_wait(); 1335 continue; 1336 } 1337 } 1338 } 1339 BasicLock* lock = mon_info->lock(); 1340 ObjectSynchronizer::enter(obj, lock, deoptee_thread); 1341 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1342 } 1343 } 1344 } 1345 return relocked_objects; 1346 } 1347 1348 1349 #ifndef PRODUCT 1350 // print information about reallocated objects 1351 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1352 fieldDescriptor fd; 1353 1354 for (int i = 0; i < objects->length(); i++) { 1355 ObjectValue* sv = (ObjectValue*) objects->at(i); 1356 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1357 Handle obj = sv->value(); 1358 1359 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1360 k->print_value(); 1361 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1362 if (obj.is_null()) { 1363 tty->print(" allocation failed"); 1364 } else { 1365 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1366 } 1367 tty->cr(); 1368 1369 if (Verbose && !obj.is_null()) { 1370 k->oop_print_on(obj(), tty); 1371 } 1372 } 1373 } 1374 #endif 1375 #endif // COMPILER2_OR_JVMCI 1376 1377 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1378 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1379 1380 #ifndef PRODUCT 1381 if (PrintDeoptimizationDetails) { 1382 ttyLocker ttyl; 1383 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1384 fr.print_on(tty); 1385 tty->print_cr(" Virtual frames (innermost first):"); 1386 for (int index = 0; index < chunk->length(); index++) { 1387 compiledVFrame* vf = chunk->at(index); 1388 tty->print(" %2d - ", index); 1389 vf->print_value(); 1390 int bci = chunk->at(index)->raw_bci(); 1391 const char* code_name; 1392 if (bci == SynchronizationEntryBCI) { 1393 code_name = "sync entry"; 1394 } else { 1395 Bytecodes::Code code = vf->method()->code_at(bci); 1396 code_name = Bytecodes::name(code); 1397 } 1398 tty->print(" - %s", code_name); 1399 tty->print_cr(" @ bci %d ", bci); 1400 if (Verbose) { 1401 vf->print(); 1402 tty->cr(); 1403 } 1404 } 1405 } 1406 #endif 1407 1408 // Register map for next frame (used for stack crawl). We capture 1409 // the state of the deopt'ing frame's caller. Thus if we need to 1410 // stuff a C2I adapter we can properly fill in the callee-save 1411 // register locations. 1412 frame caller = fr.sender(reg_map); 1413 int frame_size = caller.sp() - fr.sp(); 1414 1415 frame sender = caller; 1416 1417 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1418 // the vframeArray containing the unpacking information is allocated in the C heap. 1419 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1420 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1421 1422 // Compare the vframeArray to the collected vframes 1423 assert(array->structural_compare(thread, chunk), "just checking"); 1424 1425 #ifndef PRODUCT 1426 if (PrintDeoptimizationDetails) { 1427 ttyLocker ttyl; 1428 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1429 } 1430 #endif // PRODUCT 1431 1432 return array; 1433 } 1434 1435 #if COMPILER2_OR_JVMCI 1436 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1437 // Reallocation of some scalar replaced objects failed. Record 1438 // that we need to pop all the interpreter frames for the 1439 // deoptimized compiled frame. 1440 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1441 thread->set_frames_to_pop_failed_realloc(array->frames()); 1442 // Unlock all monitors here otherwise the interpreter will see a 1443 // mix of locked and unlocked monitors (because of failed 1444 // reallocations of synchronized objects) and be confused. 1445 for (int i = 0; i < array->frames(); i++) { 1446 MonitorChunk* monitors = array->element(i)->monitors(); 1447 if (monitors != NULL) { 1448 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1449 BasicObjectLock* src = monitors->at(j); 1450 if (src->obj() != NULL) { 1451 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1452 } 1453 } 1454 array->element(i)->free_monitors(thread); 1455 #ifdef ASSERT 1456 array->element(i)->set_removed_monitors(); 1457 #endif 1458 } 1459 } 1460 } 1461 #endif 1462 1463 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1464 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1465 Thread* thread = Thread::current(); 1466 for (int i = 0; i < monitors->length(); i++) { 1467 MonitorInfo* mon_info = monitors->at(i); 1468 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1469 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1470 } 1471 } 1472 } 1473 1474 1475 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1476 if (!UseBiasedLocking) { 1477 return; 1478 } 1479 1480 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1481 1482 // Unfortunately we don't have a RegisterMap available in most of 1483 // the places we want to call this routine so we need to walk the 1484 // stack again to update the register map. 1485 if (map == NULL || !map->update_map()) { 1486 StackFrameStream sfs(thread, true); 1487 bool found = false; 1488 while (!found && !sfs.is_done()) { 1489 frame* cur = sfs.current(); 1490 sfs.next(); 1491 found = cur->id() == fr.id(); 1492 } 1493 assert(found, "frame to be deoptimized not found on target thread's stack"); 1494 map = sfs.register_map(); 1495 } 1496 1497 vframe* vf = vframe::new_vframe(&fr, map, thread); 1498 compiledVFrame* cvf = compiledVFrame::cast(vf); 1499 // Revoke monitors' biases in all scopes 1500 while (!cvf->is_top()) { 1501 collect_monitors(cvf, objects_to_revoke); 1502 cvf = compiledVFrame::cast(cvf->sender()); 1503 } 1504 collect_monitors(cvf, objects_to_revoke); 1505 1506 if (SafepointSynchronize::is_at_safepoint()) { 1507 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1508 } else { 1509 BiasedLocking::revoke(objects_to_revoke, thread); 1510 } 1511 } 1512 1513 1514 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1515 assert(fr.can_be_deoptimized(), "checking frame type"); 1516 1517 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1518 1519 if (LogCompilation && xtty != NULL) { 1520 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1521 assert(cm != NULL, "only compiled methods can deopt"); 1522 1523 ttyLocker ttyl; 1524 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1525 cm->log_identity(xtty); 1526 xtty->end_head(); 1527 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1528 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1529 xtty->method(sd->method()); 1530 xtty->end_elem(); 1531 if (sd->is_top()) break; 1532 } 1533 xtty->tail("deoptimized"); 1534 } 1535 1536 // Patch the compiled method so that when execution returns to it we will 1537 // deopt the execution state and return to the interpreter. 1538 fr.deoptimize(thread); 1539 } 1540 1541 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1542 deoptimize(thread, fr, map, Reason_constraint); 1543 } 1544 1545 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1546 // Deoptimize only if the frame comes from compile code. 1547 // Do not deoptimize the frame which is already patched 1548 // during the execution of the loops below. 1549 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1550 return; 1551 } 1552 ResourceMark rm; 1553 DeoptimizationMarker dm; 1554 if (UseBiasedLocking) { 1555 revoke_biases_of_monitors(thread, fr, map); 1556 } 1557 deoptimize_single_frame(thread, fr, reason); 1558 1559 } 1560 1561 #if INCLUDE_JVMCI 1562 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1563 // there is no exception handler for this pc => deoptimize 1564 cm->make_not_entrant(); 1565 1566 // Use Deoptimization::deoptimize for all of its side-effects: 1567 // revoking biases of monitors, gathering traps statistics, logging... 1568 // it also patches the return pc but we do not care about that 1569 // since we return a continuation to the deopt_blob below. 1570 JavaThread* thread = JavaThread::current(); 1571 RegisterMap reg_map(thread, UseBiasedLocking); 1572 frame runtime_frame = thread->last_frame(); 1573 frame caller_frame = runtime_frame.sender(®_map); 1574 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1575 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1576 1577 MethodData* trap_mdo = get_method_data(thread, cm->method(), true); 1578 if (trap_mdo != NULL) { 1579 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1580 } 1581 1582 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1583 } 1584 #endif 1585 1586 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1587 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1588 "can only deoptimize other thread at a safepoint"); 1589 // Compute frame and register map based on thread and sp. 1590 RegisterMap reg_map(thread, UseBiasedLocking); 1591 frame fr = thread->last_frame(); 1592 while (fr.id() != id) { 1593 fr = fr.sender(®_map); 1594 } 1595 deoptimize(thread, fr, ®_map, reason); 1596 } 1597 1598 1599 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1600 if (thread == Thread::current()) { 1601 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1602 } else { 1603 VM_DeoptimizeFrame deopt(thread, id, reason); 1604 VMThread::execute(&deopt); 1605 } 1606 } 1607 1608 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1609 deoptimize_frame(thread, id, Reason_constraint); 1610 } 1611 1612 // JVMTI PopFrame support 1613 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1614 { 1615 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1616 } 1617 JRT_END 1618 1619 MethodData* 1620 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1621 bool create_if_missing) { 1622 Thread* THREAD = thread; 1623 MethodData* mdo = m()->method_data(); 1624 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1625 // Build an MDO. Ignore errors like OutOfMemory; 1626 // that simply means we won't have an MDO to update. 1627 Method::build_interpreter_method_data(m, THREAD); 1628 if (HAS_PENDING_EXCEPTION) { 1629 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1630 CLEAR_PENDING_EXCEPTION; 1631 } 1632 mdo = m()->method_data(); 1633 } 1634 return mdo; 1635 } 1636 1637 #if COMPILER2_OR_JVMCI 1638 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1639 // in case of an unresolved klass entry, load the class. 1640 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1641 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1642 return; 1643 } 1644 1645 if (!constant_pool->tag_at(index).is_symbol()) return; 1646 1647 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1648 Symbol* symbol = constant_pool->symbol_at(index); 1649 1650 // class name? 1651 if (symbol->char_at(0) != '(') { 1652 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1653 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1654 return; 1655 } 1656 1657 // then it must be a signature! 1658 ResourceMark rm(THREAD); 1659 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1660 if (ss.is_object()) { 1661 Symbol* class_name = ss.as_symbol(); 1662 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1663 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1664 } 1665 } 1666 } 1667 1668 1669 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1670 EXCEPTION_MARK; 1671 load_class_by_index(constant_pool, index, THREAD); 1672 if (HAS_PENDING_EXCEPTION) { 1673 // Exception happened during classloading. We ignore the exception here, since it 1674 // is going to be rethrown since the current activation is going to be deoptimized and 1675 // the interpreter will re-execute the bytecode. 1676 CLEAR_PENDING_EXCEPTION; 1677 // Class loading called java code which may have caused a stack 1678 // overflow. If the exception was thrown right before the return 1679 // to the runtime the stack is no longer guarded. Reguard the 1680 // stack otherwise if we return to the uncommon trap blob and the 1681 // stack bang causes a stack overflow we crash. 1682 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1683 JavaThread* thread = (JavaThread*)THREAD; 1684 bool guard_pages_enabled = thread->stack_guards_enabled(); 1685 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1686 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1687 } 1688 } 1689 1690 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1691 HandleMark hm; 1692 1693 // uncommon_trap() is called at the beginning of the uncommon trap 1694 // handler. Note this fact before we start generating temporary frames 1695 // that can confuse an asynchronous stack walker. This counter is 1696 // decremented at the end of unpack_frames(). 1697 thread->inc_in_deopt_handler(); 1698 1699 // We need to update the map if we have biased locking. 1700 #if INCLUDE_JVMCI 1701 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1702 RegisterMap reg_map(thread, true); 1703 #else 1704 RegisterMap reg_map(thread, UseBiasedLocking); 1705 #endif 1706 frame stub_frame = thread->last_frame(); 1707 frame fr = stub_frame.sender(®_map); 1708 // Make sure the calling nmethod is not getting deoptimized and removed 1709 // before we are done with it. 1710 nmethodLocker nl(fr.pc()); 1711 1712 // Log a message 1713 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1714 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1715 1716 { 1717 ResourceMark rm; 1718 1719 // Revoke biases of any monitors in the frame to ensure we can migrate them 1720 revoke_biases_of_monitors(thread, fr, ®_map); 1721 1722 DeoptReason reason = trap_request_reason(trap_request); 1723 DeoptAction action = trap_request_action(trap_request); 1724 #if INCLUDE_JVMCI 1725 int debug_id = trap_request_debug_id(trap_request); 1726 #endif 1727 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1728 1729 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1730 compiledVFrame* cvf = compiledVFrame::cast(vf); 1731 1732 CompiledMethod* nm = cvf->code(); 1733 1734 ScopeDesc* trap_scope = cvf->scope(); 1735 1736 if (TraceDeoptimization) { 1737 ttyLocker ttyl; 1738 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 1739 #if INCLUDE_JVMCI 1740 , debug_id 1741 #endif 1742 ); 1743 } 1744 1745 methodHandle trap_method = trap_scope->method(); 1746 int trap_bci = trap_scope->bci(); 1747 #if INCLUDE_JVMCI 1748 jlong speculation = thread->pending_failed_speculation(); 1749 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods 1750 nm->as_nmethod()->update_speculation(thread); 1751 } else { 1752 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 1753 } 1754 1755 if (trap_bci == SynchronizationEntryBCI) { 1756 trap_bci = 0; 1757 thread->set_pending_monitorenter(true); 1758 } 1759 1760 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1761 thread->set_pending_transfer_to_interpreter(true); 1762 } 1763 #endif 1764 1765 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1766 // Record this event in the histogram. 1767 gather_statistics(reason, action, trap_bc); 1768 1769 // Ensure that we can record deopt. history: 1770 // Need MDO to record RTM code generation state. 1771 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1772 1773 methodHandle profiled_method; 1774 #if INCLUDE_JVMCI 1775 if (nm->is_compiled_by_jvmci()) { 1776 profiled_method = nm->method(); 1777 } else { 1778 profiled_method = trap_method; 1779 } 1780 #else 1781 profiled_method = trap_method; 1782 #endif 1783 1784 MethodData* trap_mdo = 1785 get_method_data(thread, profiled_method, create_if_missing); 1786 1787 // Log a message 1788 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1789 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1790 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1791 1792 // Print a bunch of diagnostics, if requested. 1793 if (TraceDeoptimization || LogCompilation) { 1794 ResourceMark rm; 1795 ttyLocker ttyl; 1796 char buf[100]; 1797 if (xtty != NULL) { 1798 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1799 os::current_thread_id(), 1800 format_trap_request(buf, sizeof(buf), trap_request)); 1801 #if INCLUDE_JVMCI 1802 if (speculation != 0) { 1803 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 1804 } 1805 #endif 1806 nm->log_identity(xtty); 1807 } 1808 Symbol* class_name = NULL; 1809 bool unresolved = false; 1810 if (unloaded_class_index >= 0) { 1811 constantPoolHandle constants (THREAD, trap_method->constants()); 1812 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1813 class_name = constants->klass_name_at(unloaded_class_index); 1814 unresolved = true; 1815 if (xtty != NULL) 1816 xtty->print(" unresolved='1'"); 1817 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1818 class_name = constants->symbol_at(unloaded_class_index); 1819 } 1820 if (xtty != NULL) 1821 xtty->name(class_name); 1822 } 1823 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1824 // Dump the relevant MDO state. 1825 // This is the deopt count for the current reason, any previous 1826 // reasons or recompiles seen at this point. 1827 int dcnt = trap_mdo->trap_count(reason); 1828 if (dcnt != 0) 1829 xtty->print(" count='%d'", dcnt); 1830 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1831 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1832 if (dos != 0) { 1833 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1834 if (trap_state_is_recompiled(dos)) { 1835 int recnt2 = trap_mdo->overflow_recompile_count(); 1836 if (recnt2 != 0) 1837 xtty->print(" recompiles2='%d'", recnt2); 1838 } 1839 } 1840 } 1841 if (xtty != NULL) { 1842 xtty->stamp(); 1843 xtty->end_head(); 1844 } 1845 if (TraceDeoptimization) { // make noise on the tty 1846 tty->print("Uncommon trap occurred in"); 1847 nm->method()->print_short_name(tty); 1848 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1849 #if INCLUDE_JVMCI 1850 if (nm->is_nmethod()) { 1851 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 1852 if (installed_code_name != NULL) { 1853 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1854 } 1855 } 1856 #endif 1857 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1858 p2i(fr.pc()), 1859 os::current_thread_id(), 1860 trap_reason_name(reason), 1861 trap_action_name(action), 1862 unloaded_class_index 1863 #if INCLUDE_JVMCI 1864 , debug_id 1865 #endif 1866 ); 1867 if (class_name != NULL) { 1868 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1869 class_name->print_symbol_on(tty); 1870 } 1871 tty->cr(); 1872 } 1873 if (xtty != NULL) { 1874 // Log the precise location of the trap. 1875 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1876 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1877 xtty->method(sd->method()); 1878 xtty->end_elem(); 1879 if (sd->is_top()) break; 1880 } 1881 xtty->tail("uncommon_trap"); 1882 } 1883 } 1884 // (End diagnostic printout.) 1885 1886 // Load class if necessary 1887 if (unloaded_class_index >= 0) { 1888 constantPoolHandle constants(THREAD, trap_method->constants()); 1889 load_class_by_index(constants, unloaded_class_index); 1890 } 1891 1892 // Flush the nmethod if necessary and desirable. 1893 // 1894 // We need to avoid situations where we are re-flushing the nmethod 1895 // because of a hot deoptimization site. Repeated flushes at the same 1896 // point need to be detected by the compiler and avoided. If the compiler 1897 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1898 // module must take measures to avoid an infinite cycle of recompilation 1899 // and deoptimization. There are several such measures: 1900 // 1901 // 1. If a recompilation is ordered a second time at some site X 1902 // and for the same reason R, the action is adjusted to 'reinterpret', 1903 // to give the interpreter time to exercise the method more thoroughly. 1904 // If this happens, the method's overflow_recompile_count is incremented. 1905 // 1906 // 2. If the compiler fails to reduce the deoptimization rate, then 1907 // the method's overflow_recompile_count will begin to exceed the set 1908 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1909 // is adjusted to 'make_not_compilable', and the method is abandoned 1910 // to the interpreter. This is a performance hit for hot methods, 1911 // but is better than a disastrous infinite cycle of recompilations. 1912 // (Actually, only the method containing the site X is abandoned.) 1913 // 1914 // 3. In parallel with the previous measures, if the total number of 1915 // recompilations of a method exceeds the much larger set limit 1916 // PerMethodRecompilationCutoff, the method is abandoned. 1917 // This should only happen if the method is very large and has 1918 // many "lukewarm" deoptimizations. The code which enforces this 1919 // limit is elsewhere (class nmethod, class Method). 1920 // 1921 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1922 // to recompile at each bytecode independently of the per-BCI cutoff. 1923 // 1924 // The decision to update code is up to the compiler, and is encoded 1925 // in the Action_xxx code. If the compiler requests Action_none 1926 // no trap state is changed, no compiled code is changed, and the 1927 // computation suffers along in the interpreter. 1928 // 1929 // The other action codes specify various tactics for decompilation 1930 // and recompilation. Action_maybe_recompile is the loosest, and 1931 // allows the compiled code to stay around until enough traps are seen, 1932 // and until the compiler gets around to recompiling the trapping method. 1933 // 1934 // The other actions cause immediate removal of the present code. 1935 1936 // Traps caused by injected profile shouldn't pollute trap counts. 1937 bool injected_profile_trap = trap_method->has_injected_profile() && 1938 (reason == Reason_intrinsic || reason == Reason_unreached); 1939 1940 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1941 bool make_not_entrant = false; 1942 bool make_not_compilable = false; 1943 bool reprofile = false; 1944 switch (action) { 1945 case Action_none: 1946 // Keep the old code. 1947 update_trap_state = false; 1948 break; 1949 case Action_maybe_recompile: 1950 // Do not need to invalidate the present code, but we can 1951 // initiate another 1952 // Start compiler without (necessarily) invalidating the nmethod. 1953 // The system will tolerate the old code, but new code should be 1954 // generated when possible. 1955 break; 1956 case Action_reinterpret: 1957 // Go back into the interpreter for a while, and then consider 1958 // recompiling form scratch. 1959 make_not_entrant = true; 1960 // Reset invocation counter for outer most method. 1961 // This will allow the interpreter to exercise the bytecodes 1962 // for a while before recompiling. 1963 // By contrast, Action_make_not_entrant is immediate. 1964 // 1965 // Note that the compiler will track null_check, null_assert, 1966 // range_check, and class_check events and log them as if they 1967 // had been traps taken from compiled code. This will update 1968 // the MDO trap history so that the next compilation will 1969 // properly detect hot trap sites. 1970 reprofile = true; 1971 break; 1972 case Action_make_not_entrant: 1973 // Request immediate recompilation, and get rid of the old code. 1974 // Make them not entrant, so next time they are called they get 1975 // recompiled. Unloaded classes are loaded now so recompile before next 1976 // time they are called. Same for uninitialized. The interpreter will 1977 // link the missing class, if any. 1978 make_not_entrant = true; 1979 break; 1980 case Action_make_not_compilable: 1981 // Give up on compiling this method at all. 1982 make_not_entrant = true; 1983 make_not_compilable = true; 1984 break; 1985 default: 1986 ShouldNotReachHere(); 1987 } 1988 1989 // Setting +ProfileTraps fixes the following, on all platforms: 1990 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1991 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1992 // recompile relies on a MethodData* to record heroic opt failures. 1993 1994 // Whether the interpreter is producing MDO data or not, we also need 1995 // to use the MDO to detect hot deoptimization points and control 1996 // aggressive optimization. 1997 bool inc_recompile_count = false; 1998 ProfileData* pdata = NULL; 1999 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 2000 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 2001 uint this_trap_count = 0; 2002 bool maybe_prior_trap = false; 2003 bool maybe_prior_recompile = false; 2004 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 2005 #if INCLUDE_JVMCI 2006 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 2007 #endif 2008 nm->method(), 2009 //outputs: 2010 this_trap_count, 2011 maybe_prior_trap, 2012 maybe_prior_recompile); 2013 // Because the interpreter also counts null, div0, range, and class 2014 // checks, these traps from compiled code are double-counted. 2015 // This is harmless; it just means that the PerXTrapLimit values 2016 // are in effect a little smaller than they look. 2017 2018 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2019 if (per_bc_reason != Reason_none) { 2020 // Now take action based on the partially known per-BCI history. 2021 if (maybe_prior_trap 2022 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 2023 // If there are too many traps at this BCI, force a recompile. 2024 // This will allow the compiler to see the limit overflow, and 2025 // take corrective action, if possible. The compiler generally 2026 // does not use the exact PerBytecodeTrapLimit value, but instead 2027 // changes its tactics if it sees any traps at all. This provides 2028 // a little hysteresis, delaying a recompile until a trap happens 2029 // several times. 2030 // 2031 // Actually, since there is only one bit of counter per BCI, 2032 // the possible per-BCI counts are {0,1,(per-method count)}. 2033 // This produces accurate results if in fact there is only 2034 // one hot trap site, but begins to get fuzzy if there are 2035 // many sites. For example, if there are ten sites each 2036 // trapping two or more times, they each get the blame for 2037 // all of their traps. 2038 make_not_entrant = true; 2039 } 2040 2041 // Detect repeated recompilation at the same BCI, and enforce a limit. 2042 if (make_not_entrant && maybe_prior_recompile) { 2043 // More than one recompile at this point. 2044 inc_recompile_count = maybe_prior_trap; 2045 } 2046 } else { 2047 // For reasons which are not recorded per-bytecode, we simply 2048 // force recompiles unconditionally. 2049 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2050 make_not_entrant = true; 2051 } 2052 2053 // Go back to the compiler if there are too many traps in this method. 2054 if (this_trap_count >= per_method_trap_limit(reason)) { 2055 // If there are too many traps in this method, force a recompile. 2056 // This will allow the compiler to see the limit overflow, and 2057 // take corrective action, if possible. 2058 // (This condition is an unlikely backstop only, because the 2059 // PerBytecodeTrapLimit is more likely to take effect first, 2060 // if it is applicable.) 2061 make_not_entrant = true; 2062 } 2063 2064 // Here's more hysteresis: If there has been a recompile at 2065 // this trap point already, run the method in the interpreter 2066 // for a while to exercise it more thoroughly. 2067 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2068 reprofile = true; 2069 } 2070 } 2071 2072 // Take requested actions on the method: 2073 2074 // Recompile 2075 if (make_not_entrant) { 2076 if (!nm->make_not_entrant()) { 2077 return; // the call did not change nmethod's state 2078 } 2079 2080 if (pdata != NULL) { 2081 // Record the recompilation event, if any. 2082 int tstate0 = pdata->trap_state(); 2083 int tstate1 = trap_state_set_recompiled(tstate0, true); 2084 if (tstate1 != tstate0) 2085 pdata->set_trap_state(tstate1); 2086 } 2087 2088 #if INCLUDE_RTM_OPT 2089 // Restart collecting RTM locking abort statistic if the method 2090 // is recompiled for a reason other than RTM state change. 2091 // Assume that in new recompiled code the statistic could be different, 2092 // for example, due to different inlining. 2093 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2094 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2095 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2096 } 2097 #endif 2098 // For code aging we count traps separately here, using make_not_entrant() 2099 // as a guard against simultaneous deopts in multiple threads. 2100 if (reason == Reason_tenured && trap_mdo != NULL) { 2101 trap_mdo->inc_tenure_traps(); 2102 } 2103 } 2104 2105 if (inc_recompile_count) { 2106 trap_mdo->inc_overflow_recompile_count(); 2107 if ((uint)trap_mdo->overflow_recompile_count() > 2108 (uint)PerBytecodeRecompilationCutoff) { 2109 // Give up on the method containing the bad BCI. 2110 if (trap_method() == nm->method()) { 2111 make_not_compilable = true; 2112 } else { 2113 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2114 // But give grace to the enclosing nm->method(). 2115 } 2116 } 2117 } 2118 2119 // Reprofile 2120 if (reprofile) { 2121 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2122 } 2123 2124 // Give up compiling 2125 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2126 assert(make_not_entrant, "consistent"); 2127 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2128 } 2129 2130 } // Free marked resources 2131 2132 } 2133 JRT_END 2134 2135 ProfileData* 2136 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2137 int trap_bci, 2138 Deoptimization::DeoptReason reason, 2139 bool update_total_trap_count, 2140 #if INCLUDE_JVMCI 2141 bool is_osr, 2142 #endif 2143 Method* compiled_method, 2144 //outputs: 2145 uint& ret_this_trap_count, 2146 bool& ret_maybe_prior_trap, 2147 bool& ret_maybe_prior_recompile) { 2148 bool maybe_prior_trap = false; 2149 bool maybe_prior_recompile = false; 2150 uint this_trap_count = 0; 2151 if (update_total_trap_count) { 2152 uint idx = reason; 2153 #if INCLUDE_JVMCI 2154 if (is_osr) { 2155 idx += Reason_LIMIT; 2156 } 2157 #endif 2158 uint prior_trap_count = trap_mdo->trap_count(idx); 2159 this_trap_count = trap_mdo->inc_trap_count(idx); 2160 2161 // If the runtime cannot find a place to store trap history, 2162 // it is estimated based on the general condition of the method. 2163 // If the method has ever been recompiled, or has ever incurred 2164 // a trap with the present reason , then this BCI is assumed 2165 // (pessimistically) to be the culprit. 2166 maybe_prior_trap = (prior_trap_count != 0); 2167 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2168 } 2169 ProfileData* pdata = NULL; 2170 2171 2172 // For reasons which are recorded per bytecode, we check per-BCI data. 2173 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2174 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2175 if (per_bc_reason != Reason_none) { 2176 // Find the profile data for this BCI. If there isn't one, 2177 // try to allocate one from the MDO's set of spares. 2178 // This will let us detect a repeated trap at this point. 2179 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2180 2181 if (pdata != NULL) { 2182 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2183 if (LogCompilation && xtty != NULL) { 2184 ttyLocker ttyl; 2185 // no more room for speculative traps in this MDO 2186 xtty->elem("speculative_traps_oom"); 2187 } 2188 } 2189 // Query the trap state of this profile datum. 2190 int tstate0 = pdata->trap_state(); 2191 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2192 maybe_prior_trap = false; 2193 if (!trap_state_is_recompiled(tstate0)) 2194 maybe_prior_recompile = false; 2195 2196 // Update the trap state of this profile datum. 2197 int tstate1 = tstate0; 2198 // Record the reason. 2199 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2200 // Store the updated state on the MDO, for next time. 2201 if (tstate1 != tstate0) 2202 pdata->set_trap_state(tstate1); 2203 } else { 2204 if (LogCompilation && xtty != NULL) { 2205 ttyLocker ttyl; 2206 // Missing MDP? Leave a small complaint in the log. 2207 xtty->elem("missing_mdp bci='%d'", trap_bci); 2208 } 2209 } 2210 } 2211 2212 // Return results: 2213 ret_this_trap_count = this_trap_count; 2214 ret_maybe_prior_trap = maybe_prior_trap; 2215 ret_maybe_prior_recompile = maybe_prior_recompile; 2216 return pdata; 2217 } 2218 2219 void 2220 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2221 ResourceMark rm; 2222 // Ignored outputs: 2223 uint ignore_this_trap_count; 2224 bool ignore_maybe_prior_trap; 2225 bool ignore_maybe_prior_recompile; 2226 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2227 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2228 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2229 query_update_method_data(trap_mdo, trap_bci, 2230 (DeoptReason)reason, 2231 update_total_counts, 2232 #if INCLUDE_JVMCI 2233 false, 2234 #endif 2235 NULL, 2236 ignore_this_trap_count, 2237 ignore_maybe_prior_trap, 2238 ignore_maybe_prior_recompile); 2239 } 2240 2241 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2242 if (TraceDeoptimization) { 2243 tty->print("Uncommon trap "); 2244 } 2245 // Still in Java no safepoints 2246 { 2247 // This enters VM and may safepoint 2248 uncommon_trap_inner(thread, trap_request); 2249 } 2250 return fetch_unroll_info_helper(thread, exec_mode); 2251 } 2252 2253 // Local derived constants. 2254 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2255 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2256 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2257 2258 //---------------------------trap_state_reason--------------------------------- 2259 Deoptimization::DeoptReason 2260 Deoptimization::trap_state_reason(int trap_state) { 2261 // This assert provides the link between the width of DataLayout::trap_bits 2262 // and the encoding of "recorded" reasons. It ensures there are enough 2263 // bits to store all needed reasons in the per-BCI MDO profile. 2264 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2265 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2266 trap_state -= recompile_bit; 2267 if (trap_state == DS_REASON_MASK) { 2268 return Reason_many; 2269 } else { 2270 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2271 return (DeoptReason)trap_state; 2272 } 2273 } 2274 //-------------------------trap_state_has_reason------------------------------- 2275 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2276 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2277 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2278 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2279 trap_state -= recompile_bit; 2280 if (trap_state == DS_REASON_MASK) { 2281 return -1; // true, unspecifically (bottom of state lattice) 2282 } else if (trap_state == reason) { 2283 return 1; // true, definitely 2284 } else if (trap_state == 0) { 2285 return 0; // false, definitely (top of state lattice) 2286 } else { 2287 return 0; // false, definitely 2288 } 2289 } 2290 //-------------------------trap_state_add_reason------------------------------- 2291 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2292 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2293 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2294 trap_state -= recompile_bit; 2295 if (trap_state == DS_REASON_MASK) { 2296 return trap_state + recompile_bit; // already at state lattice bottom 2297 } else if (trap_state == reason) { 2298 return trap_state + recompile_bit; // the condition is already true 2299 } else if (trap_state == 0) { 2300 return reason + recompile_bit; // no condition has yet been true 2301 } else { 2302 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2303 } 2304 } 2305 //-----------------------trap_state_is_recompiled------------------------------ 2306 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2307 return (trap_state & DS_RECOMPILE_BIT) != 0; 2308 } 2309 //-----------------------trap_state_set_recompiled----------------------------- 2310 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2311 if (z) return trap_state | DS_RECOMPILE_BIT; 2312 else return trap_state & ~DS_RECOMPILE_BIT; 2313 } 2314 //---------------------------format_trap_state--------------------------------- 2315 // This is used for debugging and diagnostics, including LogFile output. 2316 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2317 int trap_state) { 2318 assert(buflen > 0, "sanity"); 2319 DeoptReason reason = trap_state_reason(trap_state); 2320 bool recomp_flag = trap_state_is_recompiled(trap_state); 2321 // Re-encode the state from its decoded components. 2322 int decoded_state = 0; 2323 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2324 decoded_state = trap_state_add_reason(decoded_state, reason); 2325 if (recomp_flag) 2326 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2327 // If the state re-encodes properly, format it symbolically. 2328 // Because this routine is used for debugging and diagnostics, 2329 // be robust even if the state is a strange value. 2330 size_t len; 2331 if (decoded_state != trap_state) { 2332 // Random buggy state that doesn't decode?? 2333 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2334 } else { 2335 len = jio_snprintf(buf, buflen, "%s%s", 2336 trap_reason_name(reason), 2337 recomp_flag ? " recompiled" : ""); 2338 } 2339 return buf; 2340 } 2341 2342 2343 //--------------------------------statics-------------------------------------- 2344 const char* Deoptimization::_trap_reason_name[] = { 2345 // Note: Keep this in sync. with enum DeoptReason. 2346 "none", 2347 "null_check", 2348 "null_assert" JVMCI_ONLY("_or_unreached0"), 2349 "range_check", 2350 "class_check", 2351 "array_check", 2352 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2353 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2354 "profile_predicate", 2355 "unloaded", 2356 "uninitialized", 2357 "initialized", 2358 "unreached", 2359 "unhandled", 2360 "constraint", 2361 "div0_check", 2362 "age", 2363 "predicate", 2364 "loop_limit_check", 2365 "speculate_class_check", 2366 "speculate_null_check", 2367 "speculate_null_assert", 2368 "rtm_state_change", 2369 "unstable_if", 2370 "unstable_fused_if", 2371 #if INCLUDE_JVMCI 2372 "aliasing", 2373 "transfer_to_interpreter", 2374 "not_compiled_exception_handler", 2375 "unresolved", 2376 "jsr_mismatch", 2377 #endif 2378 "tenured" 2379 }; 2380 const char* Deoptimization::_trap_action_name[] = { 2381 // Note: Keep this in sync. with enum DeoptAction. 2382 "none", 2383 "maybe_recompile", 2384 "reinterpret", 2385 "make_not_entrant", 2386 "make_not_compilable" 2387 }; 2388 2389 const char* Deoptimization::trap_reason_name(int reason) { 2390 // Check that every reason has a name 2391 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2392 2393 if (reason == Reason_many) return "many"; 2394 if ((uint)reason < Reason_LIMIT) 2395 return _trap_reason_name[reason]; 2396 static char buf[20]; 2397 sprintf(buf, "reason%d", reason); 2398 return buf; 2399 } 2400 const char* Deoptimization::trap_action_name(int action) { 2401 // Check that every action has a name 2402 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2403 2404 if ((uint)action < Action_LIMIT) 2405 return _trap_action_name[action]; 2406 static char buf[20]; 2407 sprintf(buf, "action%d", action); 2408 return buf; 2409 } 2410 2411 // This is used for debugging and diagnostics, including LogFile output. 2412 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2413 int trap_request) { 2414 jint unloaded_class_index = trap_request_index(trap_request); 2415 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2416 const char* action = trap_action_name(trap_request_action(trap_request)); 2417 #if INCLUDE_JVMCI 2418 int debug_id = trap_request_debug_id(trap_request); 2419 #endif 2420 size_t len; 2421 if (unloaded_class_index < 0) { 2422 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2423 reason, action 2424 #if INCLUDE_JVMCI 2425 ,debug_id 2426 #endif 2427 ); 2428 } else { 2429 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2430 reason, action, unloaded_class_index 2431 #if INCLUDE_JVMCI 2432 ,debug_id 2433 #endif 2434 ); 2435 } 2436 return buf; 2437 } 2438 2439 juint Deoptimization::_deoptimization_hist 2440 [Deoptimization::Reason_LIMIT] 2441 [1 + Deoptimization::Action_LIMIT] 2442 [Deoptimization::BC_CASE_LIMIT] 2443 = {0}; 2444 2445 enum { 2446 LSB_BITS = 8, 2447 LSB_MASK = right_n_bits(LSB_BITS) 2448 }; 2449 2450 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2451 Bytecodes::Code bc) { 2452 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2453 assert(action >= 0 && action < Action_LIMIT, "oob"); 2454 _deoptimization_hist[Reason_none][0][0] += 1; // total 2455 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2456 juint* cases = _deoptimization_hist[reason][1+action]; 2457 juint* bc_counter_addr = NULL; 2458 juint bc_counter = 0; 2459 // Look for an unused counter, or an exact match to this BC. 2460 if (bc != Bytecodes::_illegal) { 2461 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2462 juint* counter_addr = &cases[bc_case]; 2463 juint counter = *counter_addr; 2464 if ((counter == 0 && bc_counter_addr == NULL) 2465 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2466 // this counter is either free or is already devoted to this BC 2467 bc_counter_addr = counter_addr; 2468 bc_counter = counter | bc; 2469 } 2470 } 2471 } 2472 if (bc_counter_addr == NULL) { 2473 // Overflow, or no given bytecode. 2474 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2475 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2476 } 2477 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2478 } 2479 2480 jint Deoptimization::total_deoptimization_count() { 2481 return _deoptimization_hist[Reason_none][0][0]; 2482 } 2483 2484 void Deoptimization::print_statistics() { 2485 juint total = total_deoptimization_count(); 2486 juint account = total; 2487 if (total != 0) { 2488 ttyLocker ttyl; 2489 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2490 tty->print_cr("Deoptimization traps recorded:"); 2491 #define PRINT_STAT_LINE(name, r) \ 2492 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2493 PRINT_STAT_LINE("total", total); 2494 // For each non-zero entry in the histogram, print the reason, 2495 // the action, and (if specifically known) the type of bytecode. 2496 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2497 for (int action = 0; action < Action_LIMIT; action++) { 2498 juint* cases = _deoptimization_hist[reason][1+action]; 2499 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2500 juint counter = cases[bc_case]; 2501 if (counter != 0) { 2502 char name[1*K]; 2503 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2504 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2505 bc = Bytecodes::_illegal; 2506 sprintf(name, "%s/%s/%s", 2507 trap_reason_name(reason), 2508 trap_action_name(action), 2509 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2510 juint r = counter >> LSB_BITS; 2511 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2512 account -= r; 2513 } 2514 } 2515 } 2516 } 2517 if (account != 0) { 2518 PRINT_STAT_LINE("unaccounted", account); 2519 } 2520 #undef PRINT_STAT_LINE 2521 if (xtty != NULL) xtty->tail("statistics"); 2522 } 2523 } 2524 2525 #ifdef ASSERT 2526 // Revert optimizations based on escape analysis for all compiled frames of all Java threads as if 2527 // objects local to a frame or a thread were escaping. Do it every DeoptimizeObjectsALotInterval 2528 // milliseconds. 2529 void Deoptimization::deoptimize_objects_alot_loop() { 2530 JavaThread* ct = JavaThread::current(); 2531 HandleMark hm(ct); 2532 while (!ct->is_terminated()) { 2533 { // Begin new scope for escape barrier 2534 HandleMarkCleaner hmc(ct); 2535 ResourceMark rm(ct); 2536 JVMTIEscapeBarrier eb(ct, true); 2537 eb.deoptimize_objects_all_threads(); 2538 } 2539 // Now sleep after the escape barriers destructor resumed the java threads. 2540 os::sleep(ct, DeoptimizeObjectsALotInterval, true); 2541 } 2542 } 2543 #endif // !ASSERT 2544 2545 // Returns true iff objects were reallocated and relocked because of access through JVMTI 2546 bool JVMTIEscapeBarrier::objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) { 2547 // first/oldest update holds the flag 2548 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 2549 bool result = false; 2550 if (list != NULL ) { 2551 for (int i = 0; i < list->length(); i++) { 2552 if (list->at(i)->matches(fr_id)) { 2553 result = list->at(i)->objects_are_deoptimized(); 2554 break; 2555 } 2556 } 2557 } 2558 return result; 2559 } 2560 2561 // Remember that objects were reallocated and relocked for the compiled frame with the given id 2562 void JVMTIEscapeBarrier::set_objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) { 2563 // set in first/oldest update 2564 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 2565 DEBUG_ONLY(bool found = false); 2566 if (list != NULL ) { 2567 for (int i = 0; i < list->length(); i++) { 2568 if (list->at(i)->matches(fr_id)) { 2569 DEBUG_ONLY(found = true); 2570 list->at(i)->set_objs_are_deoptimized(); 2571 break; 2572 } 2573 } 2574 } 2575 assert(found, "variable set should exist at least for one vframe"); 2576 } 2577 2578 bool JVMTIEscapeBarrier::deoptimize_objects(compiledVFrame* cvf) { 2579 return !barrier_active() || deoptimize_objects(deoptee_thread(), cvf->fr(), cvf->register_map()); 2580 } 2581 2582 // Deoptimize frames with non escaping objects. Deoptimize objects with optimizations based on 2583 // escape analysis. Do it for all frames within the given depth and continue from there until the 2584 // entry frame is reached, because thread local objects passed as arguments might escape from callee 2585 // frames within the given depth. 2586 bool JVMTIEscapeBarrier::deoptimize_objects(int depth) { 2587 if (barrier_active() && deoptee_thread()->has_last_Java_frame()) { 2588 ResourceMark rm; 2589 HandleMark hm; 2590 RegisterMap reg_map(deoptee_thread()); 2591 vframe* vf = deoptee_thread()->last_java_vframe(®_map); 2592 int cur_depth = 0; 2593 while (vf != NULL && ((cur_depth <= depth) || !vf->is_entry_frame())) { 2594 if (vf->is_compiled_frame()) { 2595 compiledVFrame* cvf = compiledVFrame::cast(vf); 2596 // Deoptimize frame and local objects if any exist. 2597 // If cvf is deeper than depth, then we must only deoptimize if local objects are passed as args. 2598 bool should_deopt = cur_depth <= depth ? cvf->not_global_escape_in_scope() : cvf->arg_escape(); 2599 if (should_deopt && !deoptimize_objects(cvf)) { 2600 // reallocation of scalar replaced objects failed, because heap is exhausted 2601 return false; 2602 } 2603 } 2604 2605 // move to next physical frame 2606 while(!vf->is_top()) { 2607 cur_depth++; 2608 vf = vf->sender(); 2609 } 2610 cur_depth++; 2611 vf = vf->sender(); 2612 } 2613 } 2614 return true; 2615 } 2616 2617 bool JVMTIEscapeBarrier::deoptimize_objects(intptr_t* fr_id) { 2618 if (!barrier_active()) return true; 2619 // Compute frame and register map based on thread and sp. 2620 RegisterMap reg_map(deoptee_thread()); 2621 frame fr = deoptee_thread()->last_frame(); 2622 while (fr.id() != fr_id) { 2623 fr = fr.sender(®_map); 2624 } 2625 return deoptimize_objects(deoptee_thread(), fr, ®_map); 2626 } 2627 2628 2629 bool JVMTIEscapeBarrier::deoptimize_objects_all_threads() { 2630 if (!barrier_active()) return true; 2631 ResourceMark rm(calling_thread()); 2632 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2633 if (jt->has_last_Java_frame()) { 2634 RegisterMap reg_map(jt); 2635 vframe* vf = jt->last_java_vframe(®_map); 2636 assert(jt->frame_anchor()->walkable(), 2637 "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d", 2638 p2i(jt), jt->thread_state()); 2639 while (vf != NULL) { 2640 if (vf->is_compiled_frame()) { 2641 compiledVFrame* cvf = compiledVFrame::cast(vf); 2642 if ((cvf->not_global_escape_in_scope() || cvf->arg_escape()) && 2643 !deoptimize_objects(jt, cvf->fr(), cvf->register_map())) { 2644 return false; // reallocation failure 2645 } 2646 } 2647 // move to next physical frame 2648 while(!vf->is_top()) { 2649 vf = vf->sender(); 2650 } 2651 vf = vf->sender(); 2652 } 2653 } 2654 } 2655 return true; // success 2656 } 2657 2658 bool JVMTIEscapeBarrier::_deoptimizing_objects_for_all_threads = false; 2659 bool JVMTIEscapeBarrier::_self_deoptimization_in_progress = false; 2660 2661 bool JVMTIEscapeBarrier::deoptimizing_objects_for_all_threads() { 2662 assert(Threads_lock->owned_by_self(), "Threads_lock required"); 2663 return _deoptimizing_objects_for_all_threads; 2664 } 2665 2666 void JVMTIEscapeBarrier::set_deoptimizing_objects_for_all_threads(bool v) { 2667 assert(Threads_lock->owned_by_self(), "Threads_lock required"); 2668 _deoptimizing_objects_for_all_threads = v; 2669 if (!_deoptimizing_objects_for_all_threads) { 2670 Threads_lock->notify_all(); // notify waiting threads 2671 } 2672 } 2673 2674 void JVMTIEscapeBarrier::sync_and_suspend_one() { 2675 assert(_calling_thread != NULL, "calling thread must not be NULL"); 2676 assert(_deoptee_thread != NULL, "deoptee thread must not be NULL"); 2677 assert(barrier_active(), "should not call"); 2678 2679 // Sync with other threads that might be doing deoptimizations 2680 { 2681 MutexLocker ml(JvmtiObjReallocRelock_lock); 2682 while (_self_deoptimization_in_progress) { 2683 JvmtiObjReallocRelock_lock->wait(); 2684 } 2685 2686 if (self_deopt()) { 2687 _self_deoptimization_in_progress = true; 2688 } 2689 2690 while (_deoptee_thread->is_ea_obj_deopt_suspend()) { 2691 JvmtiObjReallocRelock_lock->wait(); 2692 } 2693 2694 if (self_deopt()) { 2695 return; 2696 } 2697 2698 // set suspend flag for target thread 2699 _deoptee_thread->set_ea_obj_deopt_flag(); 2700 } 2701 2702 // suspend target thread 2703 uint32_t debug_bits = 0; 2704 if (!_deoptee_thread->is_thread_fully_suspended(false, &debug_bits)) { 2705 class NopClosure : public ThreadClosure { 2706 void do_thread(Thread* th) { } 2707 } nop; 2708 Handshake::execute(&nop, _deoptee_thread); 2709 } 2710 assert(!_deoptee_thread->has_last_Java_frame() || _deoptee_thread->frame_anchor()->walkable(), 2711 "stack should be walkable now"); 2712 } 2713 2714 class VM_ThreadSuspendAllForObjDeopt : public VM_Operation { 2715 public: 2716 VMOp_Type type() const { return VMOp_ThreadSuspendAllForObjDeopt; } 2717 virtual void doit() { 2718 Thread* ct = calling_thread(); 2719 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2720 if (jt->is_hidden_from_external_view()) continue; 2721 assert(!jt->is_ea_obj_deopt_suspend(), "bad synchronization"); 2722 if (ct != jt) { 2723 jt->set_ea_obj_deopt_flag(); 2724 } 2725 } 2726 JVMTIEscapeBarrier::set_deoptimizing_objects_for_all_threads(true); 2727 } 2728 }; 2729 2730 void JVMTIEscapeBarrier::sync_and_suspend_all() { 2731 assert(barrier_active(), "should not call"); 2732 assert(_calling_thread != NULL, "calling thread must not be NULL"); 2733 assert(all_threads(), "sanity"); 2734 2735 // Sync with other threads that might be doing deoptimizations 2736 { 2737 MutexLocker ml(JvmtiObjReallocRelock_lock); 2738 while (_self_deoptimization_in_progress) { 2739 JvmtiObjReallocRelock_lock->wait(); 2740 } 2741 2742 _self_deoptimization_in_progress = true; 2743 2744 bool deopt_in_progress; 2745 do { 2746 deopt_in_progress = false; 2747 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2748 if (jt->is_ea_obj_deopt_suspend()) { 2749 deopt_in_progress = true; 2750 JvmtiObjReallocRelock_lock->wait(); 2751 break; // check all threads again 2752 } 2753 } 2754 } while(deopt_in_progress); 2755 } 2756 2757 VM_ThreadSuspendAllForObjDeopt vm_suspend_all; 2758 VMThread::execute(&vm_suspend_all); 2759 #ifdef ASSERT 2760 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2761 if (jt->is_hidden_from_external_view()) continue; 2762 assert(!jt->has_last_Java_frame() || jt->frame_anchor()->walkable(), 2763 "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d", 2764 p2i(jt), jt->thread_state()); 2765 } 2766 #endif // ASSERT 2767 } 2768 2769 void JVMTIEscapeBarrier::resume_one() { 2770 assert(barrier_active(), "should not call"); 2771 assert(!all_threads(), "use resume_all()"); 2772 MutexLocker ml(JvmtiObjReallocRelock_lock); 2773 if (self_deopt()) { 2774 assert(_self_deoptimization_in_progress, "incorrect synchronization"); 2775 _self_deoptimization_in_progress = false; 2776 } else { 2777 _deoptee_thread->clear_ea_obj_deopt_flag(); 2778 } 2779 JvmtiObjReallocRelock_lock->notify_all(); 2780 } 2781 2782 void JVMTIEscapeBarrier::resume_all() { 2783 assert(barrier_active(), "should not call"); 2784 assert(all_threads(), "use resume_one()"); 2785 { 2786 MutexLocker l1(Threads_lock); 2787 set_deoptimizing_objects_for_all_threads(false); 2788 } 2789 MutexLocker l2(JvmtiObjReallocRelock_lock); 2790 assert(_self_deoptimization_in_progress, "incorrect synchronization"); 2791 _self_deoptimization_in_progress = false; 2792 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2793 jt->clear_ea_obj_deopt_flag(); 2794 } 2795 JvmtiObjReallocRelock_lock->notify_all(); 2796 } 2797 2798 // Deoptimize the given frame and deoptimize objects with optimizations based on escape analysis, 2799 // i.e. reallocate scalar replaced objects on the heap and relock objects if locking has been 2800 // eliminated. 2801 // Deoptimized objects are kept as JVMTI deferred updates until the compiled frame is replaced with interpreter frames. 2802 // Returns false iff at least one reallocation failed. 2803 bool JVMTIEscapeBarrier::deoptimize_objects(JavaThread* deoptee, frame fr, const RegisterMap *reg_map) { 2804 if (!barrier_active()) return true; 2805 2806 JavaThread* ct = calling_thread(); 2807 bool realloc_failures = false; 2808 2809 assert(fr.is_compiled_frame(), "only compiled frames can contain stack allocated objects"); 2810 assert(reg_map->update_map(), "e.g. for values in callee saved registers"); 2811 2812 if (!objs_are_deoptimized(deoptee, fr.id())) { 2813 // Execution must not continue in the compiled method, so we deoptimize the frame. 2814 // As a side effect all locking biases will be removed which makes relocking 2815 // of eliminated nested locks easier. 2816 compiledVFrame* last_cvf = compiledVFrame::cast(vframe::new_vframe(&fr, reg_map, deoptee)); 2817 if (!fr.is_deoptimized_frame()) { 2818 Deoptimization::deoptimize_frame(deoptee, fr.id()); 2819 2820 // the frame fr is stale after the deoptimization, we have to fetch it again 2821 StackFrameStream fst(deoptee); 2822 while (fst.current()->id() != fr.id() && !fst.is_done()) { 2823 fst.next(); 2824 } 2825 assert(fst.current()->id() == fr.id(), "frame not found after deoptimization"); 2826 last_cvf = compiledVFrame::cast(vframe::new_vframe(fst.current(), fst.register_map(), deoptee)); 2827 } 2828 2829 // collect inlined frames 2830 compiledVFrame* cvf = last_cvf; 2831 GrowableArray<compiledVFrame*>* vfs = new GrowableArray<compiledVFrame*>(10); 2832 while (!cvf->is_top()) { 2833 vfs->push(cvf); 2834 cvf = compiledVFrame::cast(cvf->sender()); 2835 } 2836 vfs->push(cvf); 2837 2838 // With the exception of not escaping owners biases where revoked when the deoptimization of fr 2839 // was requested. Among the non escaping owners of eliminated locks might be some that are still 2840 // locked in callee frames. We need their biases revoked and do it here, because we cannot 2841 // safepoint in relock_objects(). Note that the markword of such an owner will then point to a 2842 // callee frame. This will be fixed in relock_objects(). 2843 if (UseBiasedLocking && last_cvf->arg_escape()) { 2844 GrowableArray<Handle>* arg_esc_owners = new GrowableArray<Handle>(); 2845 for (int i = 0; i < vfs->length(); i++) { 2846 GrowableArray<MonitorInfo*>* monitors = vfs->at(i)->monitors(); 2847 for (int j = 0; j < monitors->length(); j++) { 2848 MonitorInfo* mon_info = monitors->at(j); 2849 oop owner = mon_info->owner_is_scalar_replaced() ? oop(NULL) : mon_info->owner(); 2850 if (mon_info->eliminated() && owner != NULL) { 2851 markWord mark = owner->mark(); 2852 if (mark.has_bias_pattern() && !mark.is_biased_anonymously()) { 2853 assert(mark.biased_locker() == deoptee, "not escaping object can only be biased to current thread"); 2854 arg_esc_owners->push(Handle(ct, owner)); 2855 } 2856 } 2857 } 2858 } 2859 if (arg_esc_owners->length() > 0) { 2860 BiasedLocking::revoke(arg_esc_owners, deoptee); 2861 } 2862 } 2863 2864 // reallocate and relock optimized objects 2865 bool deoptimized_objects = Deoptimization::deoptimize_objects(ct, vfs, realloc_failures, Deoptimization::Unpack_none); 2866 if (!realloc_failures && deoptimized_objects) { 2867 // now do the updates 2868 for (int frame_index = 0; frame_index < vfs->length(); frame_index++) { 2869 cvf = vfs->at(frame_index); 2870 2871 // locals 2872 GrowableArray<ScopeValue*>* scopeLocals = cvf->scope()->locals(); 2873 StackValueCollection* locals = cvf->locals(); 2874 if (locals != NULL) { 2875 for (int i2 = 0; i2 < locals->size(); i2++) { 2876 StackValue* var = locals->at(i2); 2877 if (var->type() == T_OBJECT && scopeLocals->at(i2)->is_object()) { 2878 jvalue val; 2879 val.l = (jobject) locals->at(i2)->get_obj()(); 2880 cvf->update_local(T_OBJECT, i2, val); 2881 } 2882 } 2883 } 2884 2885 // expressions 2886 GrowableArray<ScopeValue*>* scopeExpressions = cvf->scope()->expressions(); 2887 StackValueCollection* expressions = cvf->expressions(); 2888 if (expressions != NULL) { 2889 for (int i2 = 0; i2 < expressions->size(); i2++) { 2890 StackValue* var = expressions->at(i2); 2891 if (var->type() == T_OBJECT && scopeExpressions->at(i2)->is_object()) { 2892 jvalue val; 2893 val.l = (jobject) expressions->at(i2)->get_obj()(); 2894 cvf->update_stack(T_OBJECT, i2, val); 2895 } 2896 } 2897 } 2898 2899 // monitors 2900 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 2901 if (monitors != NULL) { 2902 for (int i2 = 0; i2 < monitors->length(); i2++) { 2903 if (monitors->at(i2)->eliminated()) { 2904 assert(!monitors->at(i2)->owner_is_scalar_replaced(), "reallocation failure, should not update"); 2905 cvf->update_monitor(i2, monitors->at(i2)); 2906 } 2907 } 2908 } 2909 } 2910 set_objs_are_deoptimized(deoptee, fr.id()); 2911 } 2912 } 2913 return !realloc_failures; 2914 } 2915 2916 #else // COMPILER2_OR_JVMCI 2917 2918 2919 // Stubs for C1 only system. 2920 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2921 return false; 2922 } 2923 2924 const char* Deoptimization::trap_reason_name(int reason) { 2925 return "unknown"; 2926 } 2927 2928 void Deoptimization::print_statistics() { 2929 // no output 2930 } 2931 2932 void 2933 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2934 // no udpate 2935 } 2936 2937 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2938 return 0; 2939 } 2940 2941 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2942 Bytecodes::Code bc) { 2943 // no update 2944 } 2945 2946 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2947 int trap_state) { 2948 jio_snprintf(buf, buflen, "#%d", trap_state); 2949 return buf; 2950 } 2951 2952 #endif // COMPILER2_OR_JVMCI