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