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