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/valueArrayKlass.hpp" 46 #include "oops/valueArrayOop.hpp" 47 #include "oops/valueKlass.hpp" 48 #include "oops/verifyOopClosure.hpp" 49 #include "prims/jvmtiThreadState.hpp" 50 #include "runtime/biasedLocking.hpp" 51 #include "runtime/compilationPolicy.hpp" 52 #include "runtime/deoptimization.hpp" 53 #include "runtime/frame.inline.hpp" 54 #include "runtime/handles.inline.hpp" 55 #include "runtime/interfaceSupport.inline.hpp" 56 #include "runtime/safepointVerifiers.hpp" 57 #include "runtime/sharedRuntime.hpp" 58 #include "runtime/signature.hpp" 59 #include "runtime/stubRoutines.hpp" 60 #include "runtime/thread.hpp" 61 #include "runtime/threadSMR.hpp" 62 #include "runtime/vframe.hpp" 63 #include "runtime/vframeArray.hpp" 64 #include "runtime/vframe_hp.hpp" 65 #include "utilities/events.hpp" 66 #include "utilities/preserveException.hpp" 67 #include "utilities/xmlstream.hpp" 68 69 #if INCLUDE_JVMCI 70 #include "jvmci/jvmciRuntime.hpp" 71 #include "jvmci/jvmciJavaClasses.hpp" 72 #endif 73 74 75 bool DeoptimizationMarker::_is_active = false; 76 77 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 78 int caller_adjustment, 79 int caller_actual_parameters, 80 int number_of_frames, 81 intptr_t* frame_sizes, 82 address* frame_pcs, 83 BasicType return_type, 84 int exec_mode) { 85 _size_of_deoptimized_frame = size_of_deoptimized_frame; 86 _caller_adjustment = caller_adjustment; 87 _caller_actual_parameters = caller_actual_parameters; 88 _number_of_frames = number_of_frames; 89 _frame_sizes = frame_sizes; 90 _frame_pcs = frame_pcs; 91 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 92 _return_type = return_type; 93 _initial_info = 0; 94 // PD (x86 only) 95 _counter_temp = 0; 96 _unpack_kind = exec_mode; 97 _sender_sp_temp = 0; 98 99 _total_frame_sizes = size_of_frames(); 100 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 101 } 102 103 104 Deoptimization::UnrollBlock::~UnrollBlock() { 105 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 106 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 107 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 108 } 109 110 111 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 112 assert(register_number < RegisterMap::reg_count, "checking register number"); 113 return &_register_block[register_number * 2]; 114 } 115 116 117 118 int Deoptimization::UnrollBlock::size_of_frames() const { 119 // Acount first for the adjustment of the initial frame 120 int result = _caller_adjustment; 121 for (int index = 0; index < number_of_frames(); index++) { 122 result += frame_sizes()[index]; 123 } 124 return result; 125 } 126 127 128 void Deoptimization::UnrollBlock::print() { 129 ttyLocker ttyl; 130 tty->print_cr("UnrollBlock"); 131 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 132 tty->print( " frame_sizes: "); 133 for (int index = 0; index < number_of_frames(); index++) { 134 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 135 } 136 tty->cr(); 137 } 138 139 140 // In order to make fetch_unroll_info work properly with escape 141 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 142 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 143 // of previously eliminated objects occurs in realloc_objects, which is 144 // called from the method fetch_unroll_info_helper below. 145 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 146 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 147 // but makes the entry a little slower. There is however a little dance we have to 148 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 149 150 // fetch_unroll_info() is called at the beginning of the deoptimization 151 // handler. Note this fact before we start generating temporary frames 152 // that can confuse an asynchronous stack walker. This counter is 153 // decremented at the end of unpack_frames(). 154 if (TraceDeoptimization) { 155 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 156 } 157 thread->inc_in_deopt_handler(); 158 159 return fetch_unroll_info_helper(thread, exec_mode); 160 JRT_END 161 162 163 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 164 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 165 166 // Note: there is a safepoint safety issue here. No matter whether we enter 167 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 168 // the vframeArray is created. 169 // 170 171 // Allocate our special deoptimization ResourceMark 172 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 173 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 174 thread->set_deopt_mark(dmark); 175 176 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 177 RegisterMap map(thread, true); 178 RegisterMap dummy_map(thread, false); 179 // Now get the deoptee with a valid map 180 frame deoptee = stub_frame.sender(&map); 181 // Set the deoptee nmethod 182 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 183 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 184 thread->set_deopt_compiled_method(cm); 185 186 if (VerifyStack) { 187 thread->validate_frame_layout(); 188 } 189 190 // Create a growable array of VFrames where each VFrame represents an inlined 191 // Java frame. This storage is allocated with the usual system arena. 192 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 193 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 194 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 195 while (!vf->is_top()) { 196 assert(vf->is_compiled_frame(), "Wrong frame type"); 197 chunk->push(compiledVFrame::cast(vf)); 198 vf = vf->sender(); 199 } 200 assert(vf->is_compiled_frame(), "Wrong frame type"); 201 chunk->push(compiledVFrame::cast(vf)); 202 203 bool realloc_failures = false; 204 205 #if COMPILER2_OR_JVMCI 206 // Reallocate the non-escaping objects and restore their fields. Then 207 // relock objects if synchronization on them was eliminated. 208 #if !INCLUDE_JVMCI 209 if (DoEscapeAnalysis || EliminateNestedLocks) { 210 if (EliminateAllocations) { 211 #endif // INCLUDE_JVMCI 212 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 213 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 214 215 // The flag return_oop() indicates call sites which return oop 216 // in compiled code. Such sites include java method calls, 217 // runtime calls (for example, used to allocate new objects/arrays 218 // on slow code path) and any other calls generated in compiled code. 219 // It is not guaranteed that we can get such information here only 220 // by analyzing bytecode in deoptimized frames. This is why this flag 221 // is set during method compilation (see Compile::Process_OopMap_Node()). 222 // If the previous frame was popped or if we are dispatching an exception, 223 // we don't have an oop result. 224 ScopeDesc* scope = chunk->at(0)->scope(); 225 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt); 226 // In case of the return of multiple values, we must take care 227 // of all oop return values. 228 GrowableArray<Handle> return_oops; 229 ValueKlass* vk = NULL; 230 if (save_oop_result && scope->return_vt()) { 231 vk = ValueKlass::returned_value_klass(map); 232 if (vk != NULL) { 233 vk->save_oop_fields(map, return_oops); 234 save_oop_result = false; 235 } 236 } 237 if (save_oop_result) { 238 // Reallocation may trigger GC. If deoptimization happened on return from 239 // call which returns oop we need to save it since it is not in oopmap. 240 oop result = deoptee.saved_oop_result(&map); 241 assert(oopDesc::is_oop_or_null(result), "must be oop"); 242 return_oops.push(Handle(thread, result)); 243 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 244 if (TraceDeoptimization) { 245 ttyLocker ttyl; 246 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 247 } 248 } 249 if (objects != NULL || vk != NULL) { 250 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci(); 251 JRT_BLOCK 252 if (vk != NULL) { 253 realloc_failures = realloc_value_type_result(vk, map, return_oops, THREAD); 254 } 255 if (objects != NULL) { 256 realloc_failures = realloc_failures || realloc_objects(thread, &deoptee, objects, THREAD); 257 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 258 } 259 JRT_END 260 #ifndef PRODUCT 261 if (TraceDeoptimization) { 262 ttyLocker ttyl; 263 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 264 print_objects(objects, realloc_failures); 265 } 266 #endif 267 } 268 if (save_oop_result || vk != NULL) { 269 // Restore result. 270 assert(return_oops.length() == 1, "no value type"); 271 deoptee.set_saved_oop_result(&map, return_oops.pop()()); 272 } 273 #if !INCLUDE_JVMCI 274 } 275 if (EliminateLocks) { 276 #endif // INCLUDE_JVMCI 277 #ifndef PRODUCT 278 bool first = true; 279 #endif 280 for (int i = 0; i < chunk->length(); i++) { 281 compiledVFrame* cvf = chunk->at(i); 282 assert (cvf->scope() != NULL,"expect only compiled java frames"); 283 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 284 if (monitors->is_nonempty()) { 285 relock_objects(monitors, thread, realloc_failures); 286 #ifndef PRODUCT 287 if (PrintDeoptimizationDetails) { 288 ttyLocker ttyl; 289 for (int j = 0; j < monitors->length(); j++) { 290 MonitorInfo* mi = monitors->at(j); 291 if (mi->eliminated()) { 292 if (first) { 293 first = false; 294 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 295 } 296 if (mi->owner_is_scalar_replaced()) { 297 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 298 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 299 } else { 300 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 301 } 302 } 303 } 304 } 305 #endif // !PRODUCT 306 } 307 } 308 #if !INCLUDE_JVMCI 309 } 310 } 311 #endif // INCLUDE_JVMCI 312 #endif // COMPILER2_OR_JVMCI 313 314 ScopeDesc* trap_scope = chunk->at(0)->scope(); 315 Handle exceptionObject; 316 if (trap_scope->rethrow_exception()) { 317 if (PrintDeoptimizationDetails) { 318 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()); 319 } 320 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 321 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 322 ScopeValue* topOfStack = expressions->top(); 323 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 324 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 325 } 326 327 // Ensure that no safepoint is taken after pointers have been stored 328 // in fields of rematerialized objects. If a safepoint occurs from here on 329 // out the java state residing in the vframeArray will be missed. 330 NoSafepointVerifier no_safepoint; 331 332 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 333 #if COMPILER2_OR_JVMCI 334 if (realloc_failures) { 335 pop_frames_failed_reallocs(thread, array); 336 } 337 #endif 338 339 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 340 thread->set_vframe_array_head(array); 341 342 // Now that the vframeArray has been created if we have any deferred local writes 343 // added by jvmti then we can free up that structure as the data is now in the 344 // vframeArray 345 346 if (thread->deferred_locals() != NULL) { 347 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 348 int i = 0; 349 do { 350 // Because of inlining we could have multiple vframes for a single frame 351 // and several of the vframes could have deferred writes. Find them all. 352 if (list->at(i)->id() == array->original().id()) { 353 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 354 list->remove_at(i); 355 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 356 delete dlv; 357 } else { 358 i++; 359 } 360 } while ( i < list->length() ); 361 if (list->length() == 0) { 362 thread->set_deferred_locals(NULL); 363 // free the list and elements back to C heap. 364 delete list; 365 } 366 367 } 368 369 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 370 CodeBlob* cb = stub_frame.cb(); 371 // Verify we have the right vframeArray 372 assert(cb->frame_size() >= 0, "Unexpected frame size"); 373 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 374 375 // If the deopt call site is a MethodHandle invoke call site we have 376 // to adjust the unpack_sp. 377 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 378 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 379 unpack_sp = deoptee.unextended_sp(); 380 381 #ifdef ASSERT 382 assert(cb->is_deoptimization_stub() || 383 cb->is_uncommon_trap_stub() || 384 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 385 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 386 "unexpected code blob: %s", cb->name()); 387 #endif 388 389 // This is a guarantee instead of an assert because if vframe doesn't match 390 // we will unpack the wrong deoptimized frame and wind up in strange places 391 // where it will be very difficult to figure out what went wrong. Better 392 // to die an early death here than some very obscure death later when the 393 // trail is cold. 394 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 395 // in that it will fail to detect a problem when there is one. This needs 396 // more work in tiger timeframe. 397 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 398 399 int number_of_frames = array->frames(); 400 401 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 402 // virtual activation, which is the reverse of the elements in the vframes array. 403 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 404 // +1 because we always have an interpreter return address for the final slot. 405 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 406 int popframe_extra_args = 0; 407 // Create an interpreter return address for the stub to use as its return 408 // address so the skeletal frames are perfectly walkable 409 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 410 411 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 412 // activation be put back on the expression stack of the caller for reexecution 413 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 414 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 415 } 416 417 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 418 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 419 // than simply use array->sender.pc(). This requires us to walk the current set of frames 420 // 421 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 422 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 423 424 // It's possible that the number of parameters at the call site is 425 // different than number of arguments in the callee when method 426 // handles are used. If the caller is interpreted get the real 427 // value so that the proper amount of space can be added to it's 428 // frame. 429 bool caller_was_method_handle = false; 430 if (deopt_sender.is_interpreted_frame()) { 431 methodHandle method = deopt_sender.interpreter_frame_method(); 432 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 433 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 434 // Method handle invokes may involve fairly arbitrary chains of 435 // calls so it's impossible to know how much actual space the 436 // caller has for locals. 437 caller_was_method_handle = true; 438 } 439 } 440 441 // 442 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 443 // frame_sizes/frame_pcs[1] next oldest frame (int) 444 // frame_sizes/frame_pcs[n] youngest frame (int) 445 // 446 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 447 // owns the space for the return address to it's caller). Confusing ain't it. 448 // 449 // The vframe array can address vframes with indices running from 450 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 451 // When we create the skeletal frames we need the oldest frame to be in the zero slot 452 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 453 // so things look a little strange in this loop. 454 // 455 int callee_parameters = 0; 456 int callee_locals = 0; 457 for (int index = 0; index < array->frames(); index++ ) { 458 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 459 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 460 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 461 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 462 callee_locals, 463 index == 0, 464 popframe_extra_args); 465 // This pc doesn't have to be perfect just good enough to identify the frame 466 // as interpreted so the skeleton frame will be walkable 467 // The correct pc will be set when the skeleton frame is completely filled out 468 // The final pc we store in the loop is wrong and will be overwritten below 469 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 470 471 callee_parameters = array->element(index)->method()->size_of_parameters(); 472 callee_locals = array->element(index)->method()->max_locals(); 473 popframe_extra_args = 0; 474 } 475 476 // Compute whether the root vframe returns a float or double value. 477 BasicType return_type; 478 { 479 methodHandle method(thread, array->element(0)->method()); 480 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 481 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 482 } 483 484 // Compute information for handling adapters and adjusting the frame size of the caller. 485 int caller_adjustment = 0; 486 487 // Compute the amount the oldest interpreter frame will have to adjust 488 // its caller's stack by. If the caller is a compiled frame then 489 // we pretend that the callee has no parameters so that the 490 // extension counts for the full amount of locals and not just 491 // locals-parms. This is because without a c2i adapter the parm 492 // area as created by the compiled frame will not be usable by 493 // the interpreter. (Depending on the calling convention there 494 // may not even be enough space). 495 496 // QQQ I'd rather see this pushed down into last_frame_adjust 497 // and have it take the sender (aka caller). 498 499 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 500 caller_adjustment = last_frame_adjust(0, callee_locals); 501 } else if (callee_locals > callee_parameters) { 502 // The caller frame may need extending to accommodate 503 // non-parameter locals of the first unpacked interpreted frame. 504 // Compute that adjustment. 505 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 506 } 507 508 // If the sender is deoptimized we must retrieve the address of the handler 509 // since the frame will "magically" show the original pc before the deopt 510 // and we'd undo the deopt. 511 512 frame_pcs[0] = deopt_sender.raw_pc(); 513 514 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 515 516 #if INCLUDE_JVMCI 517 if (exceptionObject() != NULL) { 518 thread->set_exception_oop(exceptionObject()); 519 exec_mode = Unpack_exception; 520 } 521 #endif 522 523 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 524 assert(thread->has_pending_exception(), "should have thrown OOME"); 525 thread->set_exception_oop(thread->pending_exception()); 526 thread->clear_pending_exception(); 527 exec_mode = Unpack_exception; 528 } 529 530 #if INCLUDE_JVMCI 531 if (thread->frames_to_pop_failed_realloc() > 0) { 532 thread->set_pending_monitorenter(false); 533 } 534 #endif 535 536 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 537 caller_adjustment * BytesPerWord, 538 caller_was_method_handle ? 0 : callee_parameters, 539 number_of_frames, 540 frame_sizes, 541 frame_pcs, 542 return_type, 543 exec_mode); 544 // On some platforms, we need a way to pass some platform dependent 545 // information to the unpacking code so the skeletal frames come out 546 // correct (initial fp value, unextended sp, ...) 547 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 548 549 if (array->frames() > 1) { 550 if (VerifyStack && TraceDeoptimization) { 551 ttyLocker ttyl; 552 tty->print_cr("Deoptimizing method containing inlining"); 553 } 554 } 555 556 array->set_unroll_block(info); 557 return info; 558 } 559 560 // Called to cleanup deoptimization data structures in normal case 561 // after unpacking to stack and when stack overflow error occurs 562 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 563 vframeArray *array) { 564 565 // Get array if coming from exception 566 if (array == NULL) { 567 array = thread->vframe_array_head(); 568 } 569 thread->set_vframe_array_head(NULL); 570 571 // Free the previous UnrollBlock 572 vframeArray* old_array = thread->vframe_array_last(); 573 thread->set_vframe_array_last(array); 574 575 if (old_array != NULL) { 576 UnrollBlock* old_info = old_array->unroll_block(); 577 old_array->set_unroll_block(NULL); 578 delete old_info; 579 delete old_array; 580 } 581 582 // Deallocate any resource creating in this routine and any ResourceObjs allocated 583 // inside the vframeArray (StackValueCollections) 584 585 delete thread->deopt_mark(); 586 thread->set_deopt_mark(NULL); 587 thread->set_deopt_compiled_method(NULL); 588 589 590 if (JvmtiExport::can_pop_frame()) { 591 #ifndef CC_INTERP 592 // Regardless of whether we entered this routine with the pending 593 // popframe condition bit set, we should always clear it now 594 thread->clear_popframe_condition(); 595 #else 596 // C++ interpreter will clear has_pending_popframe when it enters 597 // with method_resume. For deopt_resume2 we clear it now. 598 if (thread->popframe_forcing_deopt_reexecution()) 599 thread->clear_popframe_condition(); 600 #endif /* CC_INTERP */ 601 } 602 603 // unpack_frames() is called at the end of the deoptimization handler 604 // and (in C2) at the end of the uncommon trap handler. Note this fact 605 // so that an asynchronous stack walker can work again. This counter is 606 // incremented at the beginning of fetch_unroll_info() and (in C2) at 607 // the beginning of uncommon_trap(). 608 thread->dec_in_deopt_handler(); 609 } 610 611 // Moved from cpu directories because none of the cpus has callee save values. 612 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 613 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 614 615 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 616 // the days we had adapter frames. When we deoptimize a situation where a 617 // compiled caller calls a compiled caller will have registers it expects 618 // to survive the call to the callee. If we deoptimize the callee the only 619 // way we can restore these registers is to have the oldest interpreter 620 // frame that we create restore these values. That is what this routine 621 // will accomplish. 622 623 // At the moment we have modified c2 to not have any callee save registers 624 // so this problem does not exist and this routine is just a place holder. 625 626 assert(f->is_interpreted_frame(), "must be interpreted"); 627 } 628 629 // Return BasicType of value being returned 630 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 631 632 // We are already active in the special DeoptResourceMark any ResourceObj's we 633 // allocate will be freed at the end of the routine. 634 635 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 636 // but makes the entry a little slower. There is however a little dance we have to 637 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 638 ResetNoHandleMark rnhm; // No-op in release/product versions 639 HandleMark hm; 640 641 frame stub_frame = thread->last_frame(); 642 643 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 644 // must point to the vframeArray for the unpack frame. 645 vframeArray* array = thread->vframe_array_head(); 646 647 #ifndef PRODUCT 648 if (TraceDeoptimization) { 649 ttyLocker ttyl; 650 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 651 p2i(thread), p2i(array), exec_mode); 652 } 653 #endif 654 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 655 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 656 657 UnrollBlock* info = array->unroll_block(); 658 659 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 660 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 661 662 BasicType bt = info->return_type(); 663 664 // If we have an exception pending, claim that the return type is an oop 665 // so the deopt_blob does not overwrite the exception_oop. 666 667 if (exec_mode == Unpack_exception) 668 bt = T_OBJECT; 669 670 // Cleanup thread deopt data 671 cleanup_deopt_info(thread, array); 672 673 #ifndef PRODUCT 674 if (VerifyStack) { 675 ResourceMark res_mark; 676 // Clear pending exception to not break verification code (restored afterwards) 677 PRESERVE_EXCEPTION_MARK; 678 679 thread->validate_frame_layout(); 680 681 // Verify that the just-unpacked frames match the interpreter's 682 // notions of expression stack and locals 683 vframeArray* cur_array = thread->vframe_array_last(); 684 RegisterMap rm(thread, false); 685 rm.set_include_argument_oops(false); 686 bool is_top_frame = true; 687 int callee_size_of_parameters = 0; 688 int callee_max_locals = 0; 689 for (int i = 0; i < cur_array->frames(); i++) { 690 vframeArrayElement* el = cur_array->element(i); 691 frame* iframe = el->iframe(); 692 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 693 694 // Get the oop map for this bci 695 InterpreterOopMap mask; 696 int cur_invoke_parameter_size = 0; 697 bool try_next_mask = false; 698 int next_mask_expression_stack_size = -1; 699 int top_frame_expression_stack_adjustment = 0; 700 methodHandle mh(thread, iframe->interpreter_frame_method()); 701 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 702 BytecodeStream str(mh); 703 str.set_start(iframe->interpreter_frame_bci()); 704 int max_bci = mh->code_size(); 705 // Get to the next bytecode if possible 706 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 707 // Check to see if we can grab the number of outgoing arguments 708 // at an uncommon trap for an invoke (where the compiler 709 // generates debug info before the invoke has executed) 710 Bytecodes::Code cur_code = str.next(); 711 if (Bytecodes::is_invoke(cur_code)) { 712 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 713 cur_invoke_parameter_size = invoke.size_of_parameters(); 714 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 715 callee_size_of_parameters++; 716 } 717 } 718 if (str.bci() < max_bci) { 719 Bytecodes::Code next_code = str.next(); 720 if (next_code >= 0) { 721 // The interpreter oop map generator reports results before 722 // the current bytecode has executed except in the case of 723 // calls. It seems to be hard to tell whether the compiler 724 // has emitted debug information matching the "state before" 725 // a given bytecode or the state after, so we try both 726 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 727 // Get expression stack size for the next bytecode 728 InterpreterOopMap next_mask; 729 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 730 next_mask_expression_stack_size = next_mask.expression_stack_size(); 731 if (Bytecodes::is_invoke(next_code)) { 732 Bytecode_invoke invoke(mh, str.bci()); 733 next_mask_expression_stack_size += invoke.size_of_parameters(); 734 } 735 // Need to subtract off the size of the result type of 736 // the bytecode because this is not described in the 737 // debug info but returned to the interpreter in the TOS 738 // caching register 739 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 740 if (bytecode_result_type != T_ILLEGAL) { 741 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 742 } 743 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 744 try_next_mask = true; 745 } 746 } 747 } 748 749 // Verify stack depth and oops in frame 750 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 751 if (!( 752 /* SPARC */ 753 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 754 /* x86 */ 755 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 756 (try_next_mask && 757 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 758 top_frame_expression_stack_adjustment))) || 759 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 760 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 761 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 762 )) { 763 { 764 ttyLocker ttyl; 765 766 // Print out some information that will help us debug the problem 767 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 768 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 769 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 770 iframe->interpreter_frame_expression_stack_size()); 771 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 772 tty->print_cr(" try_next_mask = %d", try_next_mask); 773 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 774 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 775 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 776 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 777 tty->print_cr(" exec_mode = %d", exec_mode); 778 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 779 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 780 tty->print_cr(" Interpreted frames:"); 781 for (int k = 0; k < cur_array->frames(); k++) { 782 vframeArrayElement* el = cur_array->element(k); 783 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 784 } 785 cur_array->print_on_2(tty); 786 } // release tty lock before calling guarantee 787 guarantee(false, "wrong number of expression stack elements during deopt"); 788 } 789 VerifyOopClosure verify; 790 iframe->oops_interpreted_do(&verify, &rm, false); 791 callee_size_of_parameters = mh->size_of_parameters(); 792 callee_max_locals = mh->max_locals(); 793 is_top_frame = false; 794 } 795 } 796 #endif /* !PRODUCT */ 797 798 799 return bt; 800 JRT_END 801 802 803 int Deoptimization::deoptimize_dependents() { 804 Threads::deoptimized_wrt_marked_nmethods(); 805 return 0; 806 } 807 808 Deoptimization::DeoptAction Deoptimization::_unloaded_action 809 = Deoptimization::Action_reinterpret; 810 811 #if COMPILER2_OR_JVMCI 812 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 813 Handle pending_exception(THREAD, thread->pending_exception()); 814 const char* exception_file = thread->exception_file(); 815 int exception_line = thread->exception_line(); 816 thread->clear_pending_exception(); 817 818 bool failures = false; 819 820 for (int i = 0; i < objects->length(); i++) { 821 assert(objects->at(i)->is_object(), "invalid debug information"); 822 ObjectValue* sv = (ObjectValue*) objects->at(i); 823 824 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 825 oop obj = NULL; 826 827 if (k->is_instance_klass()) { 828 InstanceKlass* ik = InstanceKlass::cast(k); 829 obj = ik->allocate_instance(THREAD); 830 } else if (k->is_valueArray_klass()) { 831 ValueArrayKlass* ak = ValueArrayKlass::cast(k); 832 // Value type array must be zeroed because not all memory is reassigned 833 obj = ak->allocate(sv->field_size(), THREAD); 834 } else if (k->is_typeArray_klass()) { 835 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 836 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 837 int len = sv->field_size() / type2size[ak->element_type()]; 838 obj = ak->allocate(len, THREAD); 839 } else if (k->is_objArray_klass()) { 840 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 841 obj = ak->allocate(sv->field_size(), THREAD); 842 } 843 844 if (obj == NULL) { 845 failures = true; 846 } 847 848 assert(sv->value().is_null(), "redundant reallocation"); 849 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 850 CLEAR_PENDING_EXCEPTION; 851 sv->set_value(obj); 852 } 853 854 if (failures) { 855 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 856 } else if (pending_exception.not_null()) { 857 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 858 } 859 860 return failures; 861 } 862 863 // We're deoptimizing at the return of a call, value type fields are 864 // in registers. When we go back to the interpreter, it will expect a 865 // reference to a value type instance. Allocate and initialize it from 866 // the register values here. 867 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 868 oop new_vt = vk->realloc_result(map, return_oops, THREAD); 869 if (new_vt == NULL) { 870 CLEAR_PENDING_EXCEPTION; 871 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 872 } 873 return_oops.clear(); 874 return_oops.push(Handle(THREAD, new_vt)); 875 return false; 876 } 877 878 // restore elements of an eliminated type array 879 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 880 int index = 0; 881 intptr_t val; 882 883 for (int i = 0; i < sv->field_size(); i++) { 884 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 885 switch(type) { 886 case T_LONG: case T_DOUBLE: { 887 assert(value->type() == T_INT, "Agreement."); 888 StackValue* low = 889 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 890 #ifdef _LP64 891 jlong res = (jlong)low->get_int(); 892 #else 893 #ifdef SPARC 894 // For SPARC we have to swap high and low words. 895 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 896 #else 897 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 898 #endif //SPARC 899 #endif 900 obj->long_at_put(index, res); 901 break; 902 } 903 904 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 905 case T_INT: case T_FLOAT: { // 4 bytes. 906 assert(value->type() == T_INT, "Agreement."); 907 bool big_value = false; 908 if (i + 1 < sv->field_size() && type == T_INT) { 909 if (sv->field_at(i)->is_location()) { 910 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 911 if (type == Location::dbl || type == Location::lng) { 912 big_value = true; 913 } 914 } else if (sv->field_at(i)->is_constant_int()) { 915 ScopeValue* next_scope_field = sv->field_at(i + 1); 916 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 917 big_value = true; 918 } 919 } 920 } 921 922 if (big_value) { 923 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 924 #ifdef _LP64 925 jlong res = (jlong)low->get_int(); 926 #else 927 #ifdef SPARC 928 // For SPARC we have to swap high and low words. 929 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 930 #else 931 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 932 #endif //SPARC 933 #endif 934 obj->int_at_put(index, (jint)*((jint*)&res)); 935 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 936 } else { 937 val = value->get_int(); 938 obj->int_at_put(index, (jint)*((jint*)&val)); 939 } 940 break; 941 } 942 943 case T_SHORT: 944 assert(value->type() == T_INT, "Agreement."); 945 val = value->get_int(); 946 obj->short_at_put(index, (jshort)*((jint*)&val)); 947 break; 948 949 case T_CHAR: 950 assert(value->type() == T_INT, "Agreement."); 951 val = value->get_int(); 952 obj->char_at_put(index, (jchar)*((jint*)&val)); 953 break; 954 955 case T_BYTE: 956 assert(value->type() == T_INT, "Agreement."); 957 val = value->get_int(); 958 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 959 break; 960 961 case T_BOOLEAN: 962 assert(value->type() == T_INT, "Agreement."); 963 val = value->get_int(); 964 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 965 break; 966 967 default: 968 ShouldNotReachHere(); 969 } 970 index++; 971 } 972 } 973 974 975 // restore fields of an eliminated object array 976 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 977 for (int i = 0; i < sv->field_size(); i++) { 978 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 979 assert(value->type() == T_OBJECT, "object element expected"); 980 obj->obj_at_put(i, value->get_obj()()); 981 } 982 } 983 984 class ReassignedField { 985 public: 986 int _offset; 987 BasicType _type; 988 InstanceKlass* _klass; 989 public: 990 ReassignedField() { 991 _offset = 0; 992 _type = T_ILLEGAL; 993 _klass = NULL; 994 } 995 }; 996 997 int compare(ReassignedField* left, ReassignedField* right) { 998 return left->_offset - right->_offset; 999 } 1000 1001 // Restore fields of an eliminated instance object using the same field order 1002 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1003 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) { 1004 if (klass->superklass() != NULL) { 1005 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0); 1006 } 1007 1008 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1009 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1010 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1011 ReassignedField field; 1012 field._offset = fs.offset(); 1013 field._type = FieldType::basic_type(fs.signature()); 1014 if (field._type == T_VALUETYPE) { 1015 field._type = T_OBJECT; 1016 } 1017 if (fs.is_flattened()) { 1018 // Resolve klass of flattened value type field 1019 Klass* vk = klass->get_value_field_klass(fs.index()); 1020 field._klass = ValueKlass::cast(vk); 1021 field._type = T_VALUETYPE; 1022 } 1023 fields->append(field); 1024 } 1025 } 1026 fields->sort(compare); 1027 for (int i = 0; i < fields->length(); i++) { 1028 intptr_t val; 1029 ScopeValue* scope_field = sv->field_at(svIndex); 1030 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1031 int offset = base_offset + fields->at(i)._offset; 1032 BasicType type = fields->at(i)._type; 1033 switch (type) { 1034 case T_OBJECT: 1035 case T_ARRAY: 1036 assert(value->type() == T_OBJECT, "Agreement."); 1037 obj->obj_field_put(offset, value->get_obj()()); 1038 break; 1039 1040 case T_VALUETYPE: { 1041 // Recursively re-assign flattened value type fields 1042 InstanceKlass* vk = fields->at(i)._klass; 1043 assert(vk != NULL, "must be resolved"); 1044 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1045 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1046 continue; // Continue because we don't need to increment svIndex 1047 } 1048 1049 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1050 case T_INT: case T_FLOAT: { // 4 bytes. 1051 assert(value->type() == T_INT, "Agreement."); 1052 bool big_value = false; 1053 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1054 if (scope_field->is_location()) { 1055 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1056 if (type == Location::dbl || type == Location::lng) { 1057 big_value = true; 1058 } 1059 } 1060 if (scope_field->is_constant_int()) { 1061 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1062 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1063 big_value = true; 1064 } 1065 } 1066 } 1067 1068 if (big_value) { 1069 i++; 1070 assert(i < fields->length(), "second T_INT field needed"); 1071 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1072 } else { 1073 val = value->get_int(); 1074 obj->int_field_put(offset, (jint)*((jint*)&val)); 1075 break; 1076 } 1077 } 1078 /* no break */ 1079 1080 case T_LONG: case T_DOUBLE: { 1081 assert(value->type() == T_INT, "Agreement."); 1082 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1083 #ifdef _LP64 1084 jlong res = (jlong)low->get_int(); 1085 #else 1086 #ifdef SPARC 1087 // For SPARC we have to swap high and low words. 1088 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1089 #else 1090 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1091 #endif //SPARC 1092 #endif 1093 obj->long_field_put(offset, res); 1094 break; 1095 } 1096 1097 case T_SHORT: 1098 assert(value->type() == T_INT, "Agreement."); 1099 val = value->get_int(); 1100 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1101 break; 1102 1103 case T_CHAR: 1104 assert(value->type() == T_INT, "Agreement."); 1105 val = value->get_int(); 1106 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1107 break; 1108 1109 case T_BYTE: 1110 assert(value->type() == T_INT, "Agreement."); 1111 val = value->get_int(); 1112 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1113 break; 1114 1115 case T_BOOLEAN: 1116 assert(value->type() == T_INT, "Agreement."); 1117 val = value->get_int(); 1118 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1119 break; 1120 1121 default: 1122 ShouldNotReachHere(); 1123 } 1124 svIndex++; 1125 } 1126 return svIndex; 1127 } 1128 1129 // restore fields of an eliminated value type array 1130 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) { 1131 ValueKlass* vk = vak->element_klass(); 1132 assert(vk->flatten_array(), "should only be used for flattened value type arrays"); 1133 // Adjust offset to omit oop header 1134 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset(); 1135 // Initialize all elements of the flattened value type array 1136 for (int i = 0; i < sv->field_size(); i++) { 1137 ScopeValue* val = sv->field_at(i); 1138 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1139 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK); 1140 } 1141 } 1142 1143 // restore fields of all eliminated objects and arrays 1144 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1145 for (int i = 0; i < objects->length(); i++) { 1146 ObjectValue* sv = (ObjectValue*) objects->at(i); 1147 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1148 Handle obj = sv->value(); 1149 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1150 if (PrintDeoptimizationDetails) { 1151 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1152 } 1153 if (obj.is_null()) { 1154 continue; 1155 } 1156 1157 if (k->is_instance_klass()) { 1158 InstanceKlass* ik = InstanceKlass::cast(k); 1159 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1160 } else if (k->is_valueArray_klass()) { 1161 ValueArrayKlass* vak = ValueArrayKlass::cast(k); 1162 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK); 1163 } else if (k->is_typeArray_klass()) { 1164 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1165 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1166 } else if (k->is_objArray_klass()) { 1167 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1168 } 1169 } 1170 } 1171 1172 1173 // relock objects for which synchronization was eliminated 1174 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1175 for (int i = 0; i < monitors->length(); i++) { 1176 MonitorInfo* mon_info = monitors->at(i); 1177 if (mon_info->eliminated()) { 1178 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1179 if (!mon_info->owner_is_scalar_replaced()) { 1180 Handle obj(thread, mon_info->owner()); 1181 markOop mark = obj->mark(); 1182 if (UseBiasedLocking && mark->has_bias_pattern()) { 1183 // New allocated objects may have the mark set to anonymously biased. 1184 // Also the deoptimized method may called methods with synchronization 1185 // where the thread-local object is bias locked to the current thread. 1186 assert(mark->is_biased_anonymously() || 1187 mark->biased_locker() == thread, "should be locked to current thread"); 1188 // Reset mark word to unbiased prototype. 1189 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1190 obj->set_mark(unbiased_prototype); 1191 } 1192 BasicLock* lock = mon_info->lock(); 1193 ObjectSynchronizer::slow_enter(obj, lock, thread); 1194 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1195 } 1196 } 1197 } 1198 } 1199 1200 1201 #ifndef PRODUCT 1202 // print information about reallocated objects 1203 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1204 fieldDescriptor fd; 1205 1206 for (int i = 0; i < objects->length(); i++) { 1207 ObjectValue* sv = (ObjectValue*) objects->at(i); 1208 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1209 Handle obj = sv->value(); 1210 1211 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1212 k->print_value(); 1213 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1214 if (obj.is_null()) { 1215 tty->print(" allocation failed"); 1216 } else { 1217 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1218 } 1219 tty->cr(); 1220 1221 if (Verbose && !obj.is_null()) { 1222 k->oop_print_on(obj(), tty); 1223 } 1224 } 1225 } 1226 #endif 1227 #endif // COMPILER2_OR_JVMCI 1228 1229 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1230 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1231 1232 #ifndef PRODUCT 1233 if (PrintDeoptimizationDetails) { 1234 ttyLocker ttyl; 1235 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1236 fr.print_on(tty); 1237 tty->print_cr(" Virtual frames (innermost first):"); 1238 for (int index = 0; index < chunk->length(); index++) { 1239 compiledVFrame* vf = chunk->at(index); 1240 tty->print(" %2d - ", index); 1241 vf->print_value(); 1242 int bci = chunk->at(index)->raw_bci(); 1243 const char* code_name; 1244 if (bci == SynchronizationEntryBCI) { 1245 code_name = "sync entry"; 1246 } else { 1247 Bytecodes::Code code = vf->method()->code_at(bci); 1248 code_name = Bytecodes::name(code); 1249 } 1250 tty->print(" - %s", code_name); 1251 tty->print_cr(" @ bci %d ", bci); 1252 if (Verbose) { 1253 vf->print(); 1254 tty->cr(); 1255 } 1256 } 1257 } 1258 #endif 1259 1260 // Register map for next frame (used for stack crawl). We capture 1261 // the state of the deopt'ing frame's caller. Thus if we need to 1262 // stuff a C2I adapter we can properly fill in the callee-save 1263 // register locations. 1264 frame caller = fr.sender(reg_map); 1265 int frame_size = caller.sp() - fr.sp(); 1266 1267 frame sender = caller; 1268 1269 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1270 // the vframeArray containing the unpacking information is allocated in the C heap. 1271 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1272 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1273 1274 // Compare the vframeArray to the collected vframes 1275 assert(array->structural_compare(thread, chunk), "just checking"); 1276 1277 #ifndef PRODUCT 1278 if (PrintDeoptimizationDetails) { 1279 ttyLocker ttyl; 1280 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1281 } 1282 #endif // PRODUCT 1283 1284 return array; 1285 } 1286 1287 #if COMPILER2_OR_JVMCI 1288 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1289 // Reallocation of some scalar replaced objects failed. Record 1290 // that we need to pop all the interpreter frames for the 1291 // deoptimized compiled frame. 1292 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1293 thread->set_frames_to_pop_failed_realloc(array->frames()); 1294 // Unlock all monitors here otherwise the interpreter will see a 1295 // mix of locked and unlocked monitors (because of failed 1296 // reallocations of synchronized objects) and be confused. 1297 for (int i = 0; i < array->frames(); i++) { 1298 MonitorChunk* monitors = array->element(i)->monitors(); 1299 if (monitors != NULL) { 1300 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1301 BasicObjectLock* src = monitors->at(j); 1302 if (src->obj() != NULL) { 1303 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1304 } 1305 } 1306 array->element(i)->free_monitors(thread); 1307 #ifdef ASSERT 1308 array->element(i)->set_removed_monitors(); 1309 #endif 1310 } 1311 } 1312 } 1313 #endif 1314 1315 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1316 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1317 Thread* thread = Thread::current(); 1318 for (int i = 0; i < monitors->length(); i++) { 1319 MonitorInfo* mon_info = monitors->at(i); 1320 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1321 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1322 } 1323 } 1324 } 1325 1326 1327 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1328 if (!UseBiasedLocking) { 1329 return; 1330 } 1331 1332 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1333 1334 // Unfortunately we don't have a RegisterMap available in most of 1335 // the places we want to call this routine so we need to walk the 1336 // stack again to update the register map. 1337 if (map == NULL || !map->update_map()) { 1338 StackFrameStream sfs(thread, true); 1339 bool found = false; 1340 while (!found && !sfs.is_done()) { 1341 frame* cur = sfs.current(); 1342 sfs.next(); 1343 found = cur->id() == fr.id(); 1344 } 1345 assert(found, "frame to be deoptimized not found on target thread's stack"); 1346 map = sfs.register_map(); 1347 } 1348 1349 vframe* vf = vframe::new_vframe(&fr, map, thread); 1350 compiledVFrame* cvf = compiledVFrame::cast(vf); 1351 // Revoke monitors' biases in all scopes 1352 while (!cvf->is_top()) { 1353 collect_monitors(cvf, objects_to_revoke); 1354 cvf = compiledVFrame::cast(cvf->sender()); 1355 } 1356 collect_monitors(cvf, objects_to_revoke); 1357 1358 if (SafepointSynchronize::is_at_safepoint()) { 1359 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1360 } else { 1361 BiasedLocking::revoke(objects_to_revoke); 1362 } 1363 } 1364 1365 1366 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1367 assert(fr.can_be_deoptimized(), "checking frame type"); 1368 1369 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1370 1371 if (LogCompilation && xtty != NULL) { 1372 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1373 assert(cm != NULL, "only compiled methods can deopt"); 1374 1375 ttyLocker ttyl; 1376 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1377 cm->log_identity(xtty); 1378 xtty->end_head(); 1379 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1380 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1381 xtty->method(sd->method()); 1382 xtty->end_elem(); 1383 if (sd->is_top()) break; 1384 } 1385 xtty->tail("deoptimized"); 1386 } 1387 1388 // Patch the compiled method so that when execution returns to it we will 1389 // deopt the execution state and return to the interpreter. 1390 fr.deoptimize(thread); 1391 } 1392 1393 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1394 deoptimize(thread, fr, map, Reason_constraint); 1395 } 1396 1397 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1398 // Deoptimize only if the frame comes from compile code. 1399 // Do not deoptimize the frame which is already patched 1400 // during the execution of the loops below. 1401 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1402 return; 1403 } 1404 ResourceMark rm; 1405 DeoptimizationMarker dm; 1406 if (UseBiasedLocking) { 1407 revoke_biases_of_monitors(thread, fr, map); 1408 } 1409 deoptimize_single_frame(thread, fr, reason); 1410 1411 } 1412 1413 #if INCLUDE_JVMCI 1414 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1415 // there is no exception handler for this pc => deoptimize 1416 cm->make_not_entrant(); 1417 1418 // Use Deoptimization::deoptimize for all of its side-effects: 1419 // revoking biases of monitors, gathering traps statistics, logging... 1420 // it also patches the return pc but we do not care about that 1421 // since we return a continuation to the deopt_blob below. 1422 JavaThread* thread = JavaThread::current(); 1423 RegisterMap reg_map(thread, UseBiasedLocking); 1424 frame runtime_frame = thread->last_frame(); 1425 frame caller_frame = runtime_frame.sender(®_map); 1426 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1427 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1428 1429 MethodData* trap_mdo = get_method_data(thread, cm->method(), true); 1430 if (trap_mdo != NULL) { 1431 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1432 } 1433 1434 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1435 } 1436 #endif 1437 1438 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1439 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1440 "can only deoptimize other thread at a safepoint"); 1441 // Compute frame and register map based on thread and sp. 1442 RegisterMap reg_map(thread, UseBiasedLocking); 1443 frame fr = thread->last_frame(); 1444 while (fr.id() != id) { 1445 fr = fr.sender(®_map); 1446 } 1447 deoptimize(thread, fr, ®_map, reason); 1448 } 1449 1450 1451 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1452 if (thread == Thread::current()) { 1453 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1454 } else { 1455 VM_DeoptimizeFrame deopt(thread, id, reason); 1456 VMThread::execute(&deopt); 1457 } 1458 } 1459 1460 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1461 deoptimize_frame(thread, id, Reason_constraint); 1462 } 1463 1464 // JVMTI PopFrame support 1465 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1466 { 1467 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1468 } 1469 JRT_END 1470 1471 MethodData* 1472 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1473 bool create_if_missing) { 1474 Thread* THREAD = thread; 1475 MethodData* mdo = m()->method_data(); 1476 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1477 // Build an MDO. Ignore errors like OutOfMemory; 1478 // that simply means we won't have an MDO to update. 1479 Method::build_interpreter_method_data(m, THREAD); 1480 if (HAS_PENDING_EXCEPTION) { 1481 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1482 CLEAR_PENDING_EXCEPTION; 1483 } 1484 mdo = m()->method_data(); 1485 } 1486 return mdo; 1487 } 1488 1489 #if COMPILER2_OR_JVMCI 1490 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1491 // in case of an unresolved klass entry, load the class. 1492 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1493 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1494 return; 1495 } 1496 1497 if (!constant_pool->tag_at(index).is_symbol()) return; 1498 1499 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1500 Symbol* symbol = constant_pool->symbol_at(index); 1501 1502 // class name? 1503 if (symbol->char_at(0) != '(') { 1504 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1505 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1506 return; 1507 } 1508 1509 // then it must be a signature! 1510 ResourceMark rm(THREAD); 1511 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1512 if (ss.is_object()) { 1513 Symbol* class_name = ss.as_symbol(CHECK); 1514 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1515 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1516 } 1517 } 1518 } 1519 1520 1521 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1522 EXCEPTION_MARK; 1523 load_class_by_index(constant_pool, index, THREAD); 1524 if (HAS_PENDING_EXCEPTION) { 1525 // Exception happened during classloading. We ignore the exception here, since it 1526 // is going to be rethrown since the current activation is going to be deoptimized and 1527 // the interpreter will re-execute the bytecode. 1528 CLEAR_PENDING_EXCEPTION; 1529 // Class loading called java code which may have caused a stack 1530 // overflow. If the exception was thrown right before the return 1531 // to the runtime the stack is no longer guarded. Reguard the 1532 // stack otherwise if we return to the uncommon trap blob and the 1533 // stack bang causes a stack overflow we crash. 1534 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1535 JavaThread* thread = (JavaThread*)THREAD; 1536 bool guard_pages_enabled = thread->stack_guards_enabled(); 1537 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1538 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1539 } 1540 } 1541 1542 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1543 HandleMark hm; 1544 1545 // uncommon_trap() is called at the beginning of the uncommon trap 1546 // handler. Note this fact before we start generating temporary frames 1547 // that can confuse an asynchronous stack walker. This counter is 1548 // decremented at the end of unpack_frames(). 1549 thread->inc_in_deopt_handler(); 1550 1551 // We need to update the map if we have biased locking. 1552 #if INCLUDE_JVMCI 1553 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1554 RegisterMap reg_map(thread, true); 1555 #else 1556 RegisterMap reg_map(thread, UseBiasedLocking); 1557 #endif 1558 frame stub_frame = thread->last_frame(); 1559 frame fr = stub_frame.sender(®_map); 1560 // Make sure the calling nmethod is not getting deoptimized and removed 1561 // before we are done with it. 1562 nmethodLocker nl(fr.pc()); 1563 1564 // Log a message 1565 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1566 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1567 1568 { 1569 ResourceMark rm; 1570 1571 // Revoke biases of any monitors in the frame to ensure we can migrate them 1572 revoke_biases_of_monitors(thread, fr, ®_map); 1573 1574 DeoptReason reason = trap_request_reason(trap_request); 1575 DeoptAction action = trap_request_action(trap_request); 1576 #if INCLUDE_JVMCI 1577 int debug_id = trap_request_debug_id(trap_request); 1578 #endif 1579 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1580 1581 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1582 compiledVFrame* cvf = compiledVFrame::cast(vf); 1583 1584 CompiledMethod* nm = cvf->code(); 1585 1586 ScopeDesc* trap_scope = cvf->scope(); 1587 1588 if (TraceDeoptimization) { 1589 ttyLocker ttyl; 1590 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() 1591 #if INCLUDE_JVMCI 1592 , debug_id 1593 #endif 1594 ); 1595 } 1596 1597 methodHandle trap_method = trap_scope->method(); 1598 int trap_bci = trap_scope->bci(); 1599 #if INCLUDE_JVMCI 1600 long speculation = thread->pending_failed_speculation(); 1601 if (nm->is_compiled_by_jvmci()) { 1602 if (speculation != 0) { 1603 oop speculation_log = nm->as_nmethod()->speculation_log(); 1604 if (speculation_log != NULL) { 1605 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1606 if (HotSpotSpeculationLog::lastFailed(speculation_log) != 0) { 1607 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1608 } 1609 } 1610 if (TraceDeoptimization) { 1611 tty->print_cr("Saving speculation to speculation log"); 1612 } 1613 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1614 } else { 1615 if (TraceDeoptimization) { 1616 tty->print_cr("Speculation present but no speculation log"); 1617 } 1618 } 1619 thread->set_pending_failed_speculation(0); 1620 } else { 1621 if (TraceDeoptimization) { 1622 tty->print_cr("No speculation"); 1623 } 1624 } 1625 } else { 1626 assert(speculation == 0, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1627 } 1628 1629 if (trap_bci == SynchronizationEntryBCI) { 1630 trap_bci = 0; 1631 thread->set_pending_monitorenter(true); 1632 } 1633 1634 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1635 thread->set_pending_transfer_to_interpreter(true); 1636 } 1637 #endif 1638 1639 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1640 // Record this event in the histogram. 1641 gather_statistics(reason, action, trap_bc); 1642 1643 // Ensure that we can record deopt. history: 1644 // Need MDO to record RTM code generation state. 1645 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1646 1647 methodHandle profiled_method; 1648 #if INCLUDE_JVMCI 1649 if (nm->is_compiled_by_jvmci()) { 1650 profiled_method = nm->method(); 1651 } else { 1652 profiled_method = trap_method; 1653 } 1654 #else 1655 profiled_method = trap_method; 1656 #endif 1657 1658 MethodData* trap_mdo = 1659 get_method_data(thread, profiled_method, create_if_missing); 1660 1661 // Log a message 1662 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1663 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1664 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1665 1666 // Print a bunch of diagnostics, if requested. 1667 if (TraceDeoptimization || LogCompilation) { 1668 ResourceMark rm; 1669 ttyLocker ttyl; 1670 char buf[100]; 1671 if (xtty != NULL) { 1672 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1673 os::current_thread_id(), 1674 format_trap_request(buf, sizeof(buf), trap_request)); 1675 nm->log_identity(xtty); 1676 } 1677 Symbol* class_name = NULL; 1678 bool unresolved = false; 1679 if (unloaded_class_index >= 0) { 1680 constantPoolHandle constants (THREAD, trap_method->constants()); 1681 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1682 class_name = constants->klass_name_at(unloaded_class_index); 1683 unresolved = true; 1684 if (xtty != NULL) 1685 xtty->print(" unresolved='1'"); 1686 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1687 class_name = constants->symbol_at(unloaded_class_index); 1688 } 1689 if (xtty != NULL) 1690 xtty->name(class_name); 1691 } 1692 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1693 // Dump the relevant MDO state. 1694 // This is the deopt count for the current reason, any previous 1695 // reasons or recompiles seen at this point. 1696 int dcnt = trap_mdo->trap_count(reason); 1697 if (dcnt != 0) 1698 xtty->print(" count='%d'", dcnt); 1699 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1700 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1701 if (dos != 0) { 1702 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1703 if (trap_state_is_recompiled(dos)) { 1704 int recnt2 = trap_mdo->overflow_recompile_count(); 1705 if (recnt2 != 0) 1706 xtty->print(" recompiles2='%d'", recnt2); 1707 } 1708 } 1709 } 1710 if (xtty != NULL) { 1711 xtty->stamp(); 1712 xtty->end_head(); 1713 } 1714 if (TraceDeoptimization) { // make noise on the tty 1715 tty->print("Uncommon trap occurred in"); 1716 nm->method()->print_short_name(tty); 1717 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1718 #if INCLUDE_JVMCI 1719 if (nm->is_nmethod()) { 1720 char* installed_code_name = nm->as_nmethod()->jvmci_installed_code_name(buf, sizeof(buf)); 1721 if (installed_code_name != NULL) { 1722 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1723 } 1724 } 1725 #endif 1726 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1727 p2i(fr.pc()), 1728 os::current_thread_id(), 1729 trap_reason_name(reason), 1730 trap_action_name(action), 1731 unloaded_class_index 1732 #if INCLUDE_JVMCI 1733 , debug_id 1734 #endif 1735 ); 1736 if (class_name != NULL) { 1737 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1738 class_name->print_symbol_on(tty); 1739 } 1740 tty->cr(); 1741 } 1742 if (xtty != NULL) { 1743 // Log the precise location of the trap. 1744 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1745 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1746 xtty->method(sd->method()); 1747 xtty->end_elem(); 1748 if (sd->is_top()) break; 1749 } 1750 xtty->tail("uncommon_trap"); 1751 } 1752 } 1753 // (End diagnostic printout.) 1754 1755 // Load class if necessary 1756 if (unloaded_class_index >= 0) { 1757 constantPoolHandle constants(THREAD, trap_method->constants()); 1758 load_class_by_index(constants, unloaded_class_index); 1759 } 1760 1761 // Flush the nmethod if necessary and desirable. 1762 // 1763 // We need to avoid situations where we are re-flushing the nmethod 1764 // because of a hot deoptimization site. Repeated flushes at the same 1765 // point need to be detected by the compiler and avoided. If the compiler 1766 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1767 // module must take measures to avoid an infinite cycle of recompilation 1768 // and deoptimization. There are several such measures: 1769 // 1770 // 1. If a recompilation is ordered a second time at some site X 1771 // and for the same reason R, the action is adjusted to 'reinterpret', 1772 // to give the interpreter time to exercise the method more thoroughly. 1773 // If this happens, the method's overflow_recompile_count is incremented. 1774 // 1775 // 2. If the compiler fails to reduce the deoptimization rate, then 1776 // the method's overflow_recompile_count will begin to exceed the set 1777 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1778 // is adjusted to 'make_not_compilable', and the method is abandoned 1779 // to the interpreter. This is a performance hit for hot methods, 1780 // but is better than a disastrous infinite cycle of recompilations. 1781 // (Actually, only the method containing the site X is abandoned.) 1782 // 1783 // 3. In parallel with the previous measures, if the total number of 1784 // recompilations of a method exceeds the much larger set limit 1785 // PerMethodRecompilationCutoff, the method is abandoned. 1786 // This should only happen if the method is very large and has 1787 // many "lukewarm" deoptimizations. The code which enforces this 1788 // limit is elsewhere (class nmethod, class Method). 1789 // 1790 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1791 // to recompile at each bytecode independently of the per-BCI cutoff. 1792 // 1793 // The decision to update code is up to the compiler, and is encoded 1794 // in the Action_xxx code. If the compiler requests Action_none 1795 // no trap state is changed, no compiled code is changed, and the 1796 // computation suffers along in the interpreter. 1797 // 1798 // The other action codes specify various tactics for decompilation 1799 // and recompilation. Action_maybe_recompile is the loosest, and 1800 // allows the compiled code to stay around until enough traps are seen, 1801 // and until the compiler gets around to recompiling the trapping method. 1802 // 1803 // The other actions cause immediate removal of the present code. 1804 1805 // Traps caused by injected profile shouldn't pollute trap counts. 1806 bool injected_profile_trap = trap_method->has_injected_profile() && 1807 (reason == Reason_intrinsic || reason == Reason_unreached); 1808 1809 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1810 bool make_not_entrant = false; 1811 bool make_not_compilable = false; 1812 bool reprofile = false; 1813 switch (action) { 1814 case Action_none: 1815 // Keep the old code. 1816 update_trap_state = false; 1817 break; 1818 case Action_maybe_recompile: 1819 // Do not need to invalidate the present code, but we can 1820 // initiate another 1821 // Start compiler without (necessarily) invalidating the nmethod. 1822 // The system will tolerate the old code, but new code should be 1823 // generated when possible. 1824 break; 1825 case Action_reinterpret: 1826 // Go back into the interpreter for a while, and then consider 1827 // recompiling form scratch. 1828 make_not_entrant = true; 1829 // Reset invocation counter for outer most method. 1830 // This will allow the interpreter to exercise the bytecodes 1831 // for a while before recompiling. 1832 // By contrast, Action_make_not_entrant is immediate. 1833 // 1834 // Note that the compiler will track null_check, null_assert, 1835 // range_check, and class_check events and log them as if they 1836 // had been traps taken from compiled code. This will update 1837 // the MDO trap history so that the next compilation will 1838 // properly detect hot trap sites. 1839 reprofile = true; 1840 break; 1841 case Action_make_not_entrant: 1842 // Request immediate recompilation, and get rid of the old code. 1843 // Make them not entrant, so next time they are called they get 1844 // recompiled. Unloaded classes are loaded now so recompile before next 1845 // time they are called. Same for uninitialized. The interpreter will 1846 // link the missing class, if any. 1847 make_not_entrant = true; 1848 break; 1849 case Action_make_not_compilable: 1850 // Give up on compiling this method at all. 1851 make_not_entrant = true; 1852 make_not_compilable = true; 1853 break; 1854 default: 1855 ShouldNotReachHere(); 1856 } 1857 1858 // Setting +ProfileTraps fixes the following, on all platforms: 1859 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1860 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1861 // recompile relies on a MethodData* to record heroic opt failures. 1862 1863 // Whether the interpreter is producing MDO data or not, we also need 1864 // to use the MDO to detect hot deoptimization points and control 1865 // aggressive optimization. 1866 bool inc_recompile_count = false; 1867 ProfileData* pdata = NULL; 1868 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1869 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1870 uint this_trap_count = 0; 1871 bool maybe_prior_trap = false; 1872 bool maybe_prior_recompile = false; 1873 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1874 #if INCLUDE_JVMCI 1875 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1876 #endif 1877 nm->method(), 1878 //outputs: 1879 this_trap_count, 1880 maybe_prior_trap, 1881 maybe_prior_recompile); 1882 // Because the interpreter also counts null, div0, range, and class 1883 // checks, these traps from compiled code are double-counted. 1884 // This is harmless; it just means that the PerXTrapLimit values 1885 // are in effect a little smaller than they look. 1886 1887 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1888 if (per_bc_reason != Reason_none) { 1889 // Now take action based on the partially known per-BCI history. 1890 if (maybe_prior_trap 1891 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1892 // If there are too many traps at this BCI, force a recompile. 1893 // This will allow the compiler to see the limit overflow, and 1894 // take corrective action, if possible. The compiler generally 1895 // does not use the exact PerBytecodeTrapLimit value, but instead 1896 // changes its tactics if it sees any traps at all. This provides 1897 // a little hysteresis, delaying a recompile until a trap happens 1898 // several times. 1899 // 1900 // Actually, since there is only one bit of counter per BCI, 1901 // the possible per-BCI counts are {0,1,(per-method count)}. 1902 // This produces accurate results if in fact there is only 1903 // one hot trap site, but begins to get fuzzy if there are 1904 // many sites. For example, if there are ten sites each 1905 // trapping two or more times, they each get the blame for 1906 // all of their traps. 1907 make_not_entrant = true; 1908 } 1909 1910 // Detect repeated recompilation at the same BCI, and enforce a limit. 1911 if (make_not_entrant && maybe_prior_recompile) { 1912 // More than one recompile at this point. 1913 inc_recompile_count = maybe_prior_trap; 1914 } 1915 } else { 1916 // For reasons which are not recorded per-bytecode, we simply 1917 // force recompiles unconditionally. 1918 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1919 make_not_entrant = true; 1920 } 1921 1922 // Go back to the compiler if there are too many traps in this method. 1923 if (this_trap_count >= per_method_trap_limit(reason)) { 1924 // If there are too many traps in this method, force a recompile. 1925 // This will allow the compiler to see the limit overflow, and 1926 // take corrective action, if possible. 1927 // (This condition is an unlikely backstop only, because the 1928 // PerBytecodeTrapLimit is more likely to take effect first, 1929 // if it is applicable.) 1930 make_not_entrant = true; 1931 } 1932 1933 // Here's more hysteresis: If there has been a recompile at 1934 // this trap point already, run the method in the interpreter 1935 // for a while to exercise it more thoroughly. 1936 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1937 reprofile = true; 1938 } 1939 } 1940 1941 // Take requested actions on the method: 1942 1943 // Recompile 1944 if (make_not_entrant) { 1945 if (!nm->make_not_entrant()) { 1946 return; // the call did not change nmethod's state 1947 } 1948 1949 if (pdata != NULL) { 1950 // Record the recompilation event, if any. 1951 int tstate0 = pdata->trap_state(); 1952 int tstate1 = trap_state_set_recompiled(tstate0, true); 1953 if (tstate1 != tstate0) 1954 pdata->set_trap_state(tstate1); 1955 } 1956 1957 #if INCLUDE_RTM_OPT 1958 // Restart collecting RTM locking abort statistic if the method 1959 // is recompiled for a reason other than RTM state change. 1960 // Assume that in new recompiled code the statistic could be different, 1961 // for example, due to different inlining. 1962 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1963 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 1964 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1965 } 1966 #endif 1967 // For code aging we count traps separately here, using make_not_entrant() 1968 // as a guard against simultaneous deopts in multiple threads. 1969 if (reason == Reason_tenured && trap_mdo != NULL) { 1970 trap_mdo->inc_tenure_traps(); 1971 } 1972 } 1973 1974 if (inc_recompile_count) { 1975 trap_mdo->inc_overflow_recompile_count(); 1976 if ((uint)trap_mdo->overflow_recompile_count() > 1977 (uint)PerBytecodeRecompilationCutoff) { 1978 // Give up on the method containing the bad BCI. 1979 if (trap_method() == nm->method()) { 1980 make_not_compilable = true; 1981 } else { 1982 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1983 // But give grace to the enclosing nm->method(). 1984 } 1985 } 1986 } 1987 1988 // Reprofile 1989 if (reprofile) { 1990 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1991 } 1992 1993 // Give up compiling 1994 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1995 assert(make_not_entrant, "consistent"); 1996 nm->method()->set_not_compilable(CompLevel_full_optimization); 1997 } 1998 1999 } // Free marked resources 2000 2001 } 2002 JRT_END 2003 2004 ProfileData* 2005 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2006 int trap_bci, 2007 Deoptimization::DeoptReason reason, 2008 bool update_total_trap_count, 2009 #if INCLUDE_JVMCI 2010 bool is_osr, 2011 #endif 2012 Method* compiled_method, 2013 //outputs: 2014 uint& ret_this_trap_count, 2015 bool& ret_maybe_prior_trap, 2016 bool& ret_maybe_prior_recompile) { 2017 bool maybe_prior_trap = false; 2018 bool maybe_prior_recompile = false; 2019 uint this_trap_count = 0; 2020 if (update_total_trap_count) { 2021 uint idx = reason; 2022 #if INCLUDE_JVMCI 2023 if (is_osr) { 2024 idx += Reason_LIMIT; 2025 } 2026 #endif 2027 uint prior_trap_count = trap_mdo->trap_count(idx); 2028 this_trap_count = trap_mdo->inc_trap_count(idx); 2029 2030 // If the runtime cannot find a place to store trap history, 2031 // it is estimated based on the general condition of the method. 2032 // If the method has ever been recompiled, or has ever incurred 2033 // a trap with the present reason , then this BCI is assumed 2034 // (pessimistically) to be the culprit. 2035 maybe_prior_trap = (prior_trap_count != 0); 2036 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2037 } 2038 ProfileData* pdata = NULL; 2039 2040 2041 // For reasons which are recorded per bytecode, we check per-BCI data. 2042 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2043 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2044 if (per_bc_reason != Reason_none) { 2045 // Find the profile data for this BCI. If there isn't one, 2046 // try to allocate one from the MDO's set of spares. 2047 // This will let us detect a repeated trap at this point. 2048 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2049 2050 if (pdata != NULL) { 2051 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2052 if (LogCompilation && xtty != NULL) { 2053 ttyLocker ttyl; 2054 // no more room for speculative traps in this MDO 2055 xtty->elem("speculative_traps_oom"); 2056 } 2057 } 2058 // Query the trap state of this profile datum. 2059 int tstate0 = pdata->trap_state(); 2060 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2061 maybe_prior_trap = false; 2062 if (!trap_state_is_recompiled(tstate0)) 2063 maybe_prior_recompile = false; 2064 2065 // Update the trap state of this profile datum. 2066 int tstate1 = tstate0; 2067 // Record the reason. 2068 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2069 // Store the updated state on the MDO, for next time. 2070 if (tstate1 != tstate0) 2071 pdata->set_trap_state(tstate1); 2072 } else { 2073 if (LogCompilation && xtty != NULL) { 2074 ttyLocker ttyl; 2075 // Missing MDP? Leave a small complaint in the log. 2076 xtty->elem("missing_mdp bci='%d'", trap_bci); 2077 } 2078 } 2079 } 2080 2081 // Return results: 2082 ret_this_trap_count = this_trap_count; 2083 ret_maybe_prior_trap = maybe_prior_trap; 2084 ret_maybe_prior_recompile = maybe_prior_recompile; 2085 return pdata; 2086 } 2087 2088 void 2089 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2090 ResourceMark rm; 2091 // Ignored outputs: 2092 uint ignore_this_trap_count; 2093 bool ignore_maybe_prior_trap; 2094 bool ignore_maybe_prior_recompile; 2095 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2096 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2097 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2098 query_update_method_data(trap_mdo, trap_bci, 2099 (DeoptReason)reason, 2100 update_total_counts, 2101 #if INCLUDE_JVMCI 2102 false, 2103 #endif 2104 NULL, 2105 ignore_this_trap_count, 2106 ignore_maybe_prior_trap, 2107 ignore_maybe_prior_recompile); 2108 } 2109 2110 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2111 if (TraceDeoptimization) { 2112 tty->print("Uncommon trap "); 2113 } 2114 // Still in Java no safepoints 2115 { 2116 // This enters VM and may safepoint 2117 uncommon_trap_inner(thread, trap_request); 2118 } 2119 return fetch_unroll_info_helper(thread, exec_mode); 2120 } 2121 2122 // Local derived constants. 2123 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2124 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2125 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2126 2127 //---------------------------trap_state_reason--------------------------------- 2128 Deoptimization::DeoptReason 2129 Deoptimization::trap_state_reason(int trap_state) { 2130 // This assert provides the link between the width of DataLayout::trap_bits 2131 // and the encoding of "recorded" reasons. It ensures there are enough 2132 // bits to store all needed reasons in the per-BCI MDO profile. 2133 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2134 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2135 trap_state -= recompile_bit; 2136 if (trap_state == DS_REASON_MASK) { 2137 return Reason_many; 2138 } else { 2139 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2140 return (DeoptReason)trap_state; 2141 } 2142 } 2143 //-------------------------trap_state_has_reason------------------------------- 2144 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2145 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2146 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2147 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2148 trap_state -= recompile_bit; 2149 if (trap_state == DS_REASON_MASK) { 2150 return -1; // true, unspecifically (bottom of state lattice) 2151 } else if (trap_state == reason) { 2152 return 1; // true, definitely 2153 } else if (trap_state == 0) { 2154 return 0; // false, definitely (top of state lattice) 2155 } else { 2156 return 0; // false, definitely 2157 } 2158 } 2159 //-------------------------trap_state_add_reason------------------------------- 2160 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2161 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2162 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2163 trap_state -= recompile_bit; 2164 if (trap_state == DS_REASON_MASK) { 2165 return trap_state + recompile_bit; // already at state lattice bottom 2166 } else if (trap_state == reason) { 2167 return trap_state + recompile_bit; // the condition is already true 2168 } else if (trap_state == 0) { 2169 return reason + recompile_bit; // no condition has yet been true 2170 } else { 2171 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2172 } 2173 } 2174 //-----------------------trap_state_is_recompiled------------------------------ 2175 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2176 return (trap_state & DS_RECOMPILE_BIT) != 0; 2177 } 2178 //-----------------------trap_state_set_recompiled----------------------------- 2179 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2180 if (z) return trap_state | DS_RECOMPILE_BIT; 2181 else return trap_state & ~DS_RECOMPILE_BIT; 2182 } 2183 //---------------------------format_trap_state--------------------------------- 2184 // This is used for debugging and diagnostics, including LogFile output. 2185 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2186 int trap_state) { 2187 assert(buflen > 0, "sanity"); 2188 DeoptReason reason = trap_state_reason(trap_state); 2189 bool recomp_flag = trap_state_is_recompiled(trap_state); 2190 // Re-encode the state from its decoded components. 2191 int decoded_state = 0; 2192 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2193 decoded_state = trap_state_add_reason(decoded_state, reason); 2194 if (recomp_flag) 2195 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2196 // If the state re-encodes properly, format it symbolically. 2197 // Because this routine is used for debugging and diagnostics, 2198 // be robust even if the state is a strange value. 2199 size_t len; 2200 if (decoded_state != trap_state) { 2201 // Random buggy state that doesn't decode?? 2202 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2203 } else { 2204 len = jio_snprintf(buf, buflen, "%s%s", 2205 trap_reason_name(reason), 2206 recomp_flag ? " recompiled" : ""); 2207 } 2208 return buf; 2209 } 2210 2211 2212 //--------------------------------statics-------------------------------------- 2213 const char* Deoptimization::_trap_reason_name[] = { 2214 // Note: Keep this in sync. with enum DeoptReason. 2215 "none", 2216 "null_check", 2217 "null_assert" JVMCI_ONLY("_or_unreached0"), 2218 "range_check", 2219 "class_check", 2220 "array_check", 2221 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2222 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2223 "profile_predicate", 2224 "unloaded", 2225 "uninitialized", 2226 "unreached", 2227 "unhandled", 2228 "constraint", 2229 "div0_check", 2230 "age", 2231 "predicate", 2232 "loop_limit_check", 2233 "speculate_class_check", 2234 "speculate_null_check", 2235 "speculate_null_assert", 2236 "rtm_state_change", 2237 "unstable_if", 2238 "unstable_fused_if", 2239 #if INCLUDE_JVMCI 2240 "aliasing", 2241 "transfer_to_interpreter", 2242 "not_compiled_exception_handler", 2243 "unresolved", 2244 "jsr_mismatch", 2245 #endif 2246 "tenured" 2247 }; 2248 const char* Deoptimization::_trap_action_name[] = { 2249 // Note: Keep this in sync. with enum DeoptAction. 2250 "none", 2251 "maybe_recompile", 2252 "reinterpret", 2253 "make_not_entrant", 2254 "make_not_compilable" 2255 }; 2256 2257 const char* Deoptimization::trap_reason_name(int reason) { 2258 // Check that every reason has a name 2259 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2260 2261 if (reason == Reason_many) return "many"; 2262 if ((uint)reason < Reason_LIMIT) 2263 return _trap_reason_name[reason]; 2264 static char buf[20]; 2265 sprintf(buf, "reason%d", reason); 2266 return buf; 2267 } 2268 const char* Deoptimization::trap_action_name(int action) { 2269 // Check that every action has a name 2270 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2271 2272 if ((uint)action < Action_LIMIT) 2273 return _trap_action_name[action]; 2274 static char buf[20]; 2275 sprintf(buf, "action%d", action); 2276 return buf; 2277 } 2278 2279 // This is used for debugging and diagnostics, including LogFile output. 2280 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2281 int trap_request) { 2282 jint unloaded_class_index = trap_request_index(trap_request); 2283 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2284 const char* action = trap_action_name(trap_request_action(trap_request)); 2285 #if INCLUDE_JVMCI 2286 int debug_id = trap_request_debug_id(trap_request); 2287 #endif 2288 size_t len; 2289 if (unloaded_class_index < 0) { 2290 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2291 reason, action 2292 #if INCLUDE_JVMCI 2293 ,debug_id 2294 #endif 2295 ); 2296 } else { 2297 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2298 reason, action, unloaded_class_index 2299 #if INCLUDE_JVMCI 2300 ,debug_id 2301 #endif 2302 ); 2303 } 2304 return buf; 2305 } 2306 2307 juint Deoptimization::_deoptimization_hist 2308 [Deoptimization::Reason_LIMIT] 2309 [1 + Deoptimization::Action_LIMIT] 2310 [Deoptimization::BC_CASE_LIMIT] 2311 = {0}; 2312 2313 enum { 2314 LSB_BITS = 8, 2315 LSB_MASK = right_n_bits(LSB_BITS) 2316 }; 2317 2318 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2319 Bytecodes::Code bc) { 2320 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2321 assert(action >= 0 && action < Action_LIMIT, "oob"); 2322 _deoptimization_hist[Reason_none][0][0] += 1; // total 2323 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2324 juint* cases = _deoptimization_hist[reason][1+action]; 2325 juint* bc_counter_addr = NULL; 2326 juint bc_counter = 0; 2327 // Look for an unused counter, or an exact match to this BC. 2328 if (bc != Bytecodes::_illegal) { 2329 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2330 juint* counter_addr = &cases[bc_case]; 2331 juint counter = *counter_addr; 2332 if ((counter == 0 && bc_counter_addr == NULL) 2333 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2334 // this counter is either free or is already devoted to this BC 2335 bc_counter_addr = counter_addr; 2336 bc_counter = counter | bc; 2337 } 2338 } 2339 } 2340 if (bc_counter_addr == NULL) { 2341 // Overflow, or no given bytecode. 2342 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2343 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2344 } 2345 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2346 } 2347 2348 jint Deoptimization::total_deoptimization_count() { 2349 return _deoptimization_hist[Reason_none][0][0]; 2350 } 2351 2352 void Deoptimization::print_statistics() { 2353 juint total = total_deoptimization_count(); 2354 juint account = total; 2355 if (total != 0) { 2356 ttyLocker ttyl; 2357 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2358 tty->print_cr("Deoptimization traps recorded:"); 2359 #define PRINT_STAT_LINE(name, r) \ 2360 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2361 PRINT_STAT_LINE("total", total); 2362 // For each non-zero entry in the histogram, print the reason, 2363 // the action, and (if specifically known) the type of bytecode. 2364 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2365 for (int action = 0; action < Action_LIMIT; action++) { 2366 juint* cases = _deoptimization_hist[reason][1+action]; 2367 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2368 juint counter = cases[bc_case]; 2369 if (counter != 0) { 2370 char name[1*K]; 2371 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2372 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2373 bc = Bytecodes::_illegal; 2374 sprintf(name, "%s/%s/%s", 2375 trap_reason_name(reason), 2376 trap_action_name(action), 2377 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2378 juint r = counter >> LSB_BITS; 2379 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2380 account -= r; 2381 } 2382 } 2383 } 2384 } 2385 if (account != 0) { 2386 PRINT_STAT_LINE("unaccounted", account); 2387 } 2388 #undef PRINT_STAT_LINE 2389 if (xtty != NULL) xtty->tail("statistics"); 2390 } 2391 } 2392 #else // COMPILER2_OR_JVMCI 2393 2394 2395 // Stubs for C1 only system. 2396 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2397 return false; 2398 } 2399 2400 const char* Deoptimization::trap_reason_name(int reason) { 2401 return "unknown"; 2402 } 2403 2404 void Deoptimization::print_statistics() { 2405 // no output 2406 } 2407 2408 void 2409 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2410 // no udpate 2411 } 2412 2413 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2414 return 0; 2415 } 2416 2417 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2418 Bytecodes::Code bc) { 2419 // no update 2420 } 2421 2422 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2423 int trap_state) { 2424 jio_snprintf(buf, buflen, "#%d", trap_state); 2425 return buf; 2426 } 2427 2428 #endif // COMPILER2_OR_JVMCI