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