1 /* 2 * Copyright (c) 1997, 2017, 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 "classfile/systemDictionary.hpp" 27 #include "code/codeCache.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/nmethod.hpp" 30 #include "code/pcDesc.hpp" 31 #include "code/scopeDesc.hpp" 32 #include "interpreter/bytecode.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "interpreter/oopMapCache.hpp" 35 #include "memory/allocation.inline.hpp" 36 #include "memory/oopFactory.hpp" 37 #include "memory/resourceArea.hpp" 38 #include "oops/method.hpp" 39 #include "oops/objArrayOop.inline.hpp" 40 #include "oops/oop.inline.hpp" 41 #include "oops/fieldStreams.hpp" 42 #include "oops/valueArrayKlass.hpp" 43 #include "oops/valueArrayOop.hpp" 44 #include "oops/valueKlass.hpp" 45 #include "oops/verifyOopClosure.hpp" 46 #include "prims/jvm.h" 47 #include "prims/jvmtiThreadState.hpp" 48 #include "runtime/biasedLocking.hpp" 49 #include "runtime/compilationPolicy.hpp" 50 #include "runtime/deoptimization.hpp" 51 #include "runtime/interfaceSupport.hpp" 52 #include "runtime/sharedRuntime.hpp" 53 #include "runtime/signature.hpp" 54 #include "runtime/stubRoutines.hpp" 55 #include "runtime/thread.hpp" 56 #include "runtime/vframe.hpp" 57 #include "runtime/vframeArray.hpp" 58 #include "runtime/vframe_hp.hpp" 59 #include "utilities/events.hpp" 60 #include "utilities/xmlstream.hpp" 61 62 #if INCLUDE_JVMCI 63 #include "jvmci/jvmciRuntime.hpp" 64 #include "jvmci/jvmciJavaClasses.hpp" 65 #endif 66 67 68 bool DeoptimizationMarker::_is_active = false; 69 70 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 71 int caller_adjustment, 72 int caller_actual_parameters, 73 int number_of_frames, 74 intptr_t* frame_sizes, 75 address* frame_pcs, 76 BasicType return_type, 77 int exec_mode) { 78 _size_of_deoptimized_frame = size_of_deoptimized_frame; 79 _caller_adjustment = caller_adjustment; 80 _caller_actual_parameters = caller_actual_parameters; 81 _number_of_frames = number_of_frames; 82 _frame_sizes = frame_sizes; 83 _frame_pcs = frame_pcs; 84 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 85 _return_type = return_type; 86 _initial_info = 0; 87 // PD (x86 only) 88 _counter_temp = 0; 89 _unpack_kind = exec_mode; 90 _sender_sp_temp = 0; 91 92 _total_frame_sizes = size_of_frames(); 93 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 94 } 95 96 97 Deoptimization::UnrollBlock::~UnrollBlock() { 98 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 99 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 100 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 101 } 102 103 104 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 105 assert(register_number < RegisterMap::reg_count, "checking register number"); 106 return &_register_block[register_number * 2]; 107 } 108 109 110 111 int Deoptimization::UnrollBlock::size_of_frames() const { 112 // Acount first for the adjustment of the initial frame 113 int result = _caller_adjustment; 114 for (int index = 0; index < number_of_frames(); index++) { 115 result += frame_sizes()[index]; 116 } 117 return result; 118 } 119 120 121 void Deoptimization::UnrollBlock::print() { 122 ttyLocker ttyl; 123 tty->print_cr("UnrollBlock"); 124 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 125 tty->print( " frame_sizes: "); 126 for (int index = 0; index < number_of_frames(); index++) { 127 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 128 } 129 tty->cr(); 130 } 131 132 133 // In order to make fetch_unroll_info work properly with escape 134 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 135 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 136 // of previously eliminated objects occurs in realloc_objects, which is 137 // called from the method fetch_unroll_info_helper below. 138 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 139 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 140 // but makes the entry a little slower. There is however a little dance we have to 141 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 142 143 // fetch_unroll_info() is called at the beginning of the deoptimization 144 // handler. Note this fact before we start generating temporary frames 145 // that can confuse an asynchronous stack walker. This counter is 146 // decremented at the end of unpack_frames(). 147 if (TraceDeoptimization) { 148 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 149 } 150 thread->inc_in_deopt_handler(); 151 152 return fetch_unroll_info_helper(thread, exec_mode); 153 JRT_END 154 155 156 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 157 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 158 159 // Note: there is a safepoint safety issue here. No matter whether we enter 160 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 161 // the vframeArray is created. 162 // 163 164 // Allocate our special deoptimization ResourceMark 165 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 166 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 167 thread->set_deopt_mark(dmark); 168 169 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 170 RegisterMap map(thread, true); 171 RegisterMap dummy_map(thread, false); 172 // Now get the deoptee with a valid map 173 frame deoptee = stub_frame.sender(&map); 174 // Set the deoptee nmethod 175 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 176 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 177 thread->set_deopt_compiled_method(cm); 178 179 if (VerifyStack) { 180 thread->validate_frame_layout(); 181 } 182 183 // Create a growable array of VFrames where each VFrame represents an inlined 184 // Java frame. This storage is allocated with the usual system arena. 185 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 186 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 187 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 188 while (!vf->is_top()) { 189 assert(vf->is_compiled_frame(), "Wrong frame type"); 190 chunk->push(compiledVFrame::cast(vf)); 191 vf = vf->sender(); 192 } 193 assert(vf->is_compiled_frame(), "Wrong frame type"); 194 chunk->push(compiledVFrame::cast(vf)); 195 196 bool realloc_failures = false; 197 198 #if defined(COMPILER2) || INCLUDE_JVMCI 199 // Reallocate the non-escaping objects and restore their fields. Then 200 // relock objects if synchronization on them was eliminated. 201 #ifndef INCLUDE_JVMCI 202 if (DoEscapeAnalysis || EliminateNestedLocks) { 203 if (EliminateAllocations) { 204 #endif // INCLUDE_JVMCI 205 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 206 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 207 208 // The flag return_oop() indicates call sites which return oop 209 // in compiled code. Such sites include java method calls, 210 // runtime calls (for example, used to allocate new objects/arrays 211 // on slow code path) and any other calls generated in compiled code. 212 // It is not guaranteed that we can get such information here only 213 // by analyzing bytecode in deoptimized frames. This is why this flag 214 // is set during method compilation (see Compile::Process_OopMap_Node()). 215 // If the previous frame was popped or if we are dispatching an exception, 216 // we don't have an oop result. 217 ScopeDesc* scope = chunk->at(0)->scope(); 218 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt); 219 // In case of the return of multiple values, we must take care 220 // of all oop return values. 221 GrowableArray<Handle> return_oops; 222 ValueKlass* vk = NULL; 223 if (save_oop_result) { 224 if (scope->return_vt()) { 225 vk = ValueKlass::returned_value_type(map); 226 if (vk != NULL) { 227 bool success = vk->save_oop_results(map, return_oops); 228 assert(success, "found klass ptr being returned: saving oops can't fail"); 229 save_oop_result = false; 230 } else { 231 vk = NULL; 232 } 233 } 234 } 235 if (save_oop_result) { 236 // Reallocation may trigger GC. If deoptimization happened on return from 237 // call which returns oop we need to save it since it is not in oopmap. 238 oop result = deoptee.saved_oop_result(&map); 239 assert(oopDesc::is_oop_or_null(result), "must be oop"); 240 return_oops.push(Handle(thread, result)); 241 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 242 if (TraceDeoptimization) { 243 ttyLocker ttyl; 244 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 245 } 246 } 247 if (objects != NULL || vk != NULL) { 248 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci(); 249 JRT_BLOCK 250 if (vk != NULL) { 251 realloc_failures = realloc_value_type_result(vk, map, return_oops, THREAD); 252 } 253 if (objects != NULL) { 254 realloc_failures = realloc_failures || realloc_objects(thread, &deoptee, objects, THREAD); 255 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 256 } 257 JRT_END 258 #ifndef PRODUCT 259 if (TraceDeoptimization) { 260 ttyLocker ttyl; 261 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 262 print_objects(objects, realloc_failures); 263 } 264 #endif 265 } 266 if (save_oop_result || vk != NULL) { 267 // Restore result. 268 assert(return_oops.length() == 1, "no value type"); 269 deoptee.set_saved_oop_result(&map, return_oops.pop()()); 270 } 271 #ifndef INCLUDE_JVMCI 272 } 273 if (EliminateLocks) { 274 #endif // INCLUDE_JVMCI 275 #ifndef PRODUCT 276 bool first = true; 277 #endif 278 for (int i = 0; i < chunk->length(); i++) { 279 compiledVFrame* cvf = chunk->at(i); 280 assert (cvf->scope() != NULL,"expect only compiled java frames"); 281 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 282 if (monitors->is_nonempty()) { 283 relock_objects(monitors, thread, realloc_failures); 284 #ifndef PRODUCT 285 if (PrintDeoptimizationDetails) { 286 ttyLocker ttyl; 287 for (int j = 0; j < monitors->length(); j++) { 288 MonitorInfo* mi = monitors->at(j); 289 if (mi->eliminated()) { 290 if (first) { 291 first = false; 292 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 293 } 294 if (mi->owner_is_scalar_replaced()) { 295 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 296 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 297 } else { 298 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 299 } 300 } 301 } 302 } 303 #endif // !PRODUCT 304 } 305 } 306 #ifndef INCLUDE_JVMCI 307 } 308 } 309 #endif // INCLUDE_JVMCI 310 #endif // COMPILER2 || INCLUDE_JVMCI 311 312 ScopeDesc* trap_scope = chunk->at(0)->scope(); 313 Handle exceptionObject; 314 if (trap_scope->rethrow_exception()) { 315 if (PrintDeoptimizationDetails) { 316 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()); 317 } 318 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 319 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 320 ScopeValue* topOfStack = expressions->top(); 321 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 322 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 323 } 324 325 // Ensure that no safepoint is taken after pointers have been stored 326 // in fields of rematerialized objects. If a safepoint occurs from here on 327 // out the java state residing in the vframeArray will be missed. 328 NoSafepointVerifier no_safepoint; 329 330 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 331 #if defined(COMPILER2) || INCLUDE_JVMCI 332 if (realloc_failures) { 333 pop_frames_failed_reallocs(thread, array); 334 } 335 #endif 336 337 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 338 thread->set_vframe_array_head(array); 339 340 // Now that the vframeArray has been created if we have any deferred local writes 341 // added by jvmti then we can free up that structure as the data is now in the 342 // vframeArray 343 344 if (thread->deferred_locals() != NULL) { 345 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 346 int i = 0; 347 do { 348 // Because of inlining we could have multiple vframes for a single frame 349 // and several of the vframes could have deferred writes. Find them all. 350 if (list->at(i)->id() == array->original().id()) { 351 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 352 list->remove_at(i); 353 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 354 delete dlv; 355 } else { 356 i++; 357 } 358 } while ( i < list->length() ); 359 if (list->length() == 0) { 360 thread->set_deferred_locals(NULL); 361 // free the list and elements back to C heap. 362 delete list; 363 } 364 365 } 366 367 #ifndef SHARK 368 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 369 CodeBlob* cb = stub_frame.cb(); 370 // Verify we have the right vframeArray 371 assert(cb->frame_size() >= 0, "Unexpected frame size"); 372 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 373 374 // If the deopt call site is a MethodHandle invoke call site we have 375 // to adjust the unpack_sp. 376 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 377 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 378 unpack_sp = deoptee.unextended_sp(); 379 380 #ifdef ASSERT 381 assert(cb->is_deoptimization_stub() || 382 cb->is_uncommon_trap_stub() || 383 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 384 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 385 "unexpected code blob: %s", cb->name()); 386 #endif 387 #else 388 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 389 #endif // !SHARK 390 391 // This is a guarantee instead of an assert because if vframe doesn't match 392 // we will unpack the wrong deoptimized frame and wind up in strange places 393 // where it will be very difficult to figure out what went wrong. Better 394 // to die an early death here than some very obscure death later when the 395 // trail is cold. 396 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 397 // in that it will fail to detect a problem when there is one. This needs 398 // more work in tiger timeframe. 399 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 400 401 int number_of_frames = array->frames(); 402 403 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 404 // virtual activation, which is the reverse of the elements in the vframes array. 405 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 406 // +1 because we always have an interpreter return address for the final slot. 407 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 408 int popframe_extra_args = 0; 409 // Create an interpreter return address for the stub to use as its return 410 // address so the skeletal frames are perfectly walkable 411 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 412 413 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 414 // activation be put back on the expression stack of the caller for reexecution 415 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 416 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 417 } 418 419 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 420 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 421 // than simply use array->sender.pc(). This requires us to walk the current set of frames 422 // 423 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 424 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 425 426 // It's possible that the number of parameters at the call site is 427 // different than number of arguments in the callee when method 428 // handles are used. If the caller is interpreted get the real 429 // value so that the proper amount of space can be added to it's 430 // frame. 431 bool caller_was_method_handle = false; 432 if (deopt_sender.is_interpreted_frame()) { 433 methodHandle method = deopt_sender.interpreter_frame_method(); 434 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 435 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 436 // Method handle invokes may involve fairly arbitrary chains of 437 // calls so it's impossible to know how much actual space the 438 // caller has for locals. 439 caller_was_method_handle = true; 440 } 441 } 442 443 // 444 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 445 // frame_sizes/frame_pcs[1] next oldest frame (int) 446 // frame_sizes/frame_pcs[n] youngest frame (int) 447 // 448 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 449 // owns the space for the return address to it's caller). Confusing ain't it. 450 // 451 // The vframe array can address vframes with indices running from 452 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 453 // When we create the skeletal frames we need the oldest frame to be in the zero slot 454 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 455 // so things look a little strange in this loop. 456 // 457 int callee_parameters = 0; 458 int callee_locals = 0; 459 for (int index = 0; index < array->frames(); index++ ) { 460 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 461 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 462 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 463 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 464 callee_locals, 465 index == 0, 466 popframe_extra_args); 467 // This pc doesn't have to be perfect just good enough to identify the frame 468 // as interpreted so the skeleton frame will be walkable 469 // The correct pc will be set when the skeleton frame is completely filled out 470 // The final pc we store in the loop is wrong and will be overwritten below 471 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 472 473 callee_parameters = array->element(index)->method()->size_of_parameters(); 474 callee_locals = array->element(index)->method()->max_locals(); 475 popframe_extra_args = 0; 476 } 477 478 // Compute whether the root vframe returns a float or double value. 479 BasicType return_type; 480 { 481 methodHandle method(thread, array->element(0)->method()); 482 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 483 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 484 } 485 486 // Compute information for handling adapters and adjusting the frame size of the caller. 487 int caller_adjustment = 0; 488 489 // Compute the amount the oldest interpreter frame will have to adjust 490 // its caller's stack by. If the caller is a compiled frame then 491 // we pretend that the callee has no parameters so that the 492 // extension counts for the full amount of locals and not just 493 // locals-parms. This is because without a c2i adapter the parm 494 // area as created by the compiled frame will not be usable by 495 // the interpreter. (Depending on the calling convention there 496 // may not even be enough space). 497 498 // QQQ I'd rather see this pushed down into last_frame_adjust 499 // and have it take the sender (aka caller). 500 501 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 502 caller_adjustment = last_frame_adjust(0, callee_locals); 503 } else if (callee_locals > callee_parameters) { 504 // The caller frame may need extending to accommodate 505 // non-parameter locals of the first unpacked interpreted frame. 506 // Compute that adjustment. 507 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 508 } 509 510 // If the sender is deoptimized the we must retrieve the address of the handler 511 // since the frame will "magically" show the original pc before the deopt 512 // and we'd undo the deopt. 513 514 frame_pcs[0] = deopt_sender.raw_pc(); 515 516 #ifndef SHARK 517 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 518 #endif // SHARK 519 520 #ifdef INCLUDE_JVMCI 521 if (exceptionObject() != NULL) { 522 thread->set_exception_oop(exceptionObject()); 523 exec_mode = Unpack_exception; 524 } 525 #endif 526 527 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 528 assert(thread->has_pending_exception(), "should have thrown OOME"); 529 thread->set_exception_oop(thread->pending_exception()); 530 thread->clear_pending_exception(); 531 exec_mode = Unpack_exception; 532 } 533 534 #if INCLUDE_JVMCI 535 if (thread->frames_to_pop_failed_realloc() > 0) { 536 thread->set_pending_monitorenter(false); 537 } 538 #endif 539 540 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 541 caller_adjustment * BytesPerWord, 542 caller_was_method_handle ? 0 : callee_parameters, 543 number_of_frames, 544 frame_sizes, 545 frame_pcs, 546 return_type, 547 exec_mode); 548 // On some platforms, we need a way to pass some platform dependent 549 // information to the unpacking code so the skeletal frames come out 550 // correct (initial fp value, unextended sp, ...) 551 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 552 553 if (array->frames() > 1) { 554 if (VerifyStack && TraceDeoptimization) { 555 ttyLocker ttyl; 556 tty->print_cr("Deoptimizing method containing inlining"); 557 } 558 } 559 560 array->set_unroll_block(info); 561 return info; 562 } 563 564 // Called to cleanup deoptimization data structures in normal case 565 // after unpacking to stack and when stack overflow error occurs 566 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 567 vframeArray *array) { 568 569 // Get array if coming from exception 570 if (array == NULL) { 571 array = thread->vframe_array_head(); 572 } 573 thread->set_vframe_array_head(NULL); 574 575 // Free the previous UnrollBlock 576 vframeArray* old_array = thread->vframe_array_last(); 577 thread->set_vframe_array_last(array); 578 579 if (old_array != NULL) { 580 UnrollBlock* old_info = old_array->unroll_block(); 581 old_array->set_unroll_block(NULL); 582 delete old_info; 583 delete old_array; 584 } 585 586 // Deallocate any resource creating in this routine and any ResourceObjs allocated 587 // inside the vframeArray (StackValueCollections) 588 589 delete thread->deopt_mark(); 590 thread->set_deopt_mark(NULL); 591 thread->set_deopt_compiled_method(NULL); 592 593 594 if (JvmtiExport::can_pop_frame()) { 595 #ifndef CC_INTERP 596 // Regardless of whether we entered this routine with the pending 597 // popframe condition bit set, we should always clear it now 598 thread->clear_popframe_condition(); 599 #else 600 // C++ interpreter will clear has_pending_popframe when it enters 601 // with method_resume. For deopt_resume2 we clear it now. 602 if (thread->popframe_forcing_deopt_reexecution()) 603 thread->clear_popframe_condition(); 604 #endif /* CC_INTERP */ 605 } 606 607 // unpack_frames() is called at the end of the deoptimization handler 608 // and (in C2) at the end of the uncommon trap handler. Note this fact 609 // so that an asynchronous stack walker can work again. This counter is 610 // incremented at the beginning of fetch_unroll_info() and (in C2) at 611 // the beginning of uncommon_trap(). 612 thread->dec_in_deopt_handler(); 613 } 614 615 // Moved from cpu directories because none of the cpus has callee save values. 616 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 617 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 618 619 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 620 // the days we had adapter frames. When we deoptimize a situation where a 621 // compiled caller calls a compiled caller will have registers it expects 622 // to survive the call to the callee. If we deoptimize the callee the only 623 // way we can restore these registers is to have the oldest interpreter 624 // frame that we create restore these values. That is what this routine 625 // will accomplish. 626 627 // At the moment we have modified c2 to not have any callee save registers 628 // so this problem does not exist and this routine is just a place holder. 629 630 assert(f->is_interpreted_frame(), "must be interpreted"); 631 } 632 633 // Return BasicType of value being returned 634 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 635 636 // We are already active int he special DeoptResourceMark any ResourceObj's we 637 // allocate will be freed at the end of the routine. 638 639 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 640 // but makes the entry a little slower. There is however a little dance we have to 641 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 642 ResetNoHandleMark rnhm; // No-op in release/product versions 643 HandleMark hm; 644 645 frame stub_frame = thread->last_frame(); 646 647 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 648 // must point to the vframeArray for the unpack frame. 649 vframeArray* array = thread->vframe_array_head(); 650 651 #ifndef PRODUCT 652 if (TraceDeoptimization) { 653 ttyLocker ttyl; 654 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 655 p2i(thread), p2i(array), exec_mode); 656 } 657 #endif 658 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 659 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 660 661 UnrollBlock* info = array->unroll_block(); 662 663 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 664 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 665 666 BasicType bt = info->return_type(); 667 668 // If we have an exception pending, claim that the return type is an oop 669 // so the deopt_blob does not overwrite the exception_oop. 670 671 if (exec_mode == Unpack_exception) 672 bt = T_OBJECT; 673 674 // Cleanup thread deopt data 675 cleanup_deopt_info(thread, array); 676 677 #ifndef PRODUCT 678 if (VerifyStack) { 679 ResourceMark res_mark; 680 681 thread->validate_frame_layout(); 682 683 // Verify that the just-unpacked frames match the interpreter's 684 // notions of expression stack and locals 685 vframeArray* cur_array = thread->vframe_array_last(); 686 RegisterMap rm(thread, false); 687 rm.set_include_argument_oops(false); 688 bool is_top_frame = true; 689 int callee_size_of_parameters = 0; 690 int callee_max_locals = 0; 691 for (int i = 0; i < cur_array->frames(); i++) { 692 vframeArrayElement* el = cur_array->element(i); 693 frame* iframe = el->iframe(); 694 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 695 696 // Get the oop map for this bci 697 InterpreterOopMap mask; 698 int cur_invoke_parameter_size = 0; 699 bool try_next_mask = false; 700 int next_mask_expression_stack_size = -1; 701 int top_frame_expression_stack_adjustment = 0; 702 methodHandle mh(thread, iframe->interpreter_frame_method()); 703 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 704 BytecodeStream str(mh); 705 str.set_start(iframe->interpreter_frame_bci()); 706 int max_bci = mh->code_size(); 707 // Get to the next bytecode if possible 708 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 709 // Check to see if we can grab the number of outgoing arguments 710 // at an uncommon trap for an invoke (where the compiler 711 // generates debug info before the invoke has executed) 712 Bytecodes::Code cur_code = str.next(); 713 if (cur_code == Bytecodes::_invokevirtual || 714 cur_code == Bytecodes::_invokespecial || 715 cur_code == Bytecodes::_invokestatic || 716 cur_code == Bytecodes::_invokeinterface || 717 cur_code == Bytecodes::_invokedynamic) { 718 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 719 Symbol* signature = invoke.signature(); 720 ArgumentSizeComputer asc(signature); 721 cur_invoke_parameter_size = asc.size(); 722 if (invoke.has_receiver()) { 723 // Add in receiver 724 ++cur_invoke_parameter_size; 725 } 726 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 727 callee_size_of_parameters++; 728 } 729 } 730 if (str.bci() < max_bci) { 731 Bytecodes::Code bc = str.next(); 732 if (bc >= 0) { 733 // The interpreter oop map generator reports results before 734 // the current bytecode has executed except in the case of 735 // calls. It seems to be hard to tell whether the compiler 736 // has emitted debug information matching the "state before" 737 // a given bytecode or the state after, so we try both 738 switch (cur_code) { 739 case Bytecodes::_invokevirtual: 740 case Bytecodes::_invokespecial: 741 case Bytecodes::_invokestatic: 742 case Bytecodes::_invokeinterface: 743 case Bytecodes::_invokedynamic: 744 case Bytecodes::_athrow: 745 break; 746 default: { 747 InterpreterOopMap next_mask; 748 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 749 next_mask_expression_stack_size = next_mask.expression_stack_size(); 750 // Need to subtract off the size of the result type of 751 // the bytecode because this is not described in the 752 // debug info but returned to the interpreter in the TOS 753 // caching register 754 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 755 if (bytecode_result_type != T_ILLEGAL) { 756 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 757 } 758 assert(top_frame_expression_stack_adjustment >= 0, ""); 759 try_next_mask = true; 760 break; 761 } 762 } 763 } 764 } 765 766 // Verify stack depth and oops in frame 767 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 768 if (!( 769 /* SPARC */ 770 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 771 /* x86 */ 772 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 773 (try_next_mask && 774 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 775 top_frame_expression_stack_adjustment))) || 776 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 777 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 778 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 779 )) { 780 ttyLocker ttyl; 781 782 // Print out some information that will help us debug the problem 783 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 784 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 785 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 786 iframe->interpreter_frame_expression_stack_size()); 787 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 788 tty->print_cr(" try_next_mask = %d", try_next_mask); 789 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 790 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 791 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 792 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 793 tty->print_cr(" exec_mode = %d", exec_mode); 794 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 795 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 796 tty->print_cr(" Interpreted frames:"); 797 for (int k = 0; k < cur_array->frames(); k++) { 798 vframeArrayElement* el = cur_array->element(k); 799 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 800 } 801 cur_array->print_on_2(tty); 802 guarantee(false, "wrong number of expression stack elements during deopt"); 803 } 804 VerifyOopClosure verify; 805 iframe->oops_interpreted_do(&verify, &rm, false); 806 callee_size_of_parameters = mh->size_of_parameters(); 807 callee_max_locals = mh->max_locals(); 808 is_top_frame = false; 809 } 810 } 811 #endif /* !PRODUCT */ 812 813 814 return bt; 815 JRT_END 816 817 818 int Deoptimization::deoptimize_dependents() { 819 Threads::deoptimized_wrt_marked_nmethods(); 820 return 0; 821 } 822 823 Deoptimization::DeoptAction Deoptimization::_unloaded_action 824 = Deoptimization::Action_reinterpret; 825 826 #if defined(COMPILER2) || INCLUDE_JVMCI 827 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 828 Handle pending_exception(THREAD, thread->pending_exception()); 829 const char* exception_file = thread->exception_file(); 830 int exception_line = thread->exception_line(); 831 thread->clear_pending_exception(); 832 833 bool failures = false; 834 835 for (int i = 0; i < objects->length(); i++) { 836 assert(objects->at(i)->is_object(), "invalid debug information"); 837 ObjectValue* sv = (ObjectValue*) objects->at(i); 838 839 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 840 oop obj = NULL; 841 842 if (k->is_instance_klass()) { 843 InstanceKlass* ik = InstanceKlass::cast(k); 844 obj = ik->allocate_instance(THREAD); 845 } else if (k->is_valueArray_klass()) { 846 ValueArrayKlass* ak = ValueArrayKlass::cast(k); 847 // Value type array must be zeroed because not all memory is reassigned 848 obj = ak->allocate(sv->field_size(), true, THREAD); 849 } else if (k->is_typeArray_klass()) { 850 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 851 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 852 int len = sv->field_size() / type2size[ak->element_type()]; 853 obj = ak->allocate(len, THREAD); 854 } else if (k->is_objArray_klass()) { 855 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 856 obj = ak->allocate(sv->field_size(), THREAD); 857 } 858 859 if (obj == NULL) { 860 failures = true; 861 } 862 863 assert(sv->value().is_null(), "redundant reallocation"); 864 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 865 CLEAR_PENDING_EXCEPTION; 866 sv->set_value(obj); 867 } 868 869 if (failures) { 870 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 871 } else if (pending_exception.not_null()) { 872 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 873 } 874 875 return failures; 876 } 877 878 // We're deoptimizing at the return of a call, value type fields are 879 // in registers. When we go back to the interpreter, it will expect a 880 // reference to a value type instance. Allocate and initialize it from 881 // the register values here. 882 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 883 oop new_vt = vk->realloc_result(map, return_oops, false, THREAD); 884 if (new_vt == NULL) { 885 CLEAR_PENDING_EXCEPTION; 886 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 887 } 888 return_oops.clear(); 889 return_oops.push(Handle(THREAD, new_vt)); 890 return false; 891 } 892 893 // restore elements of an eliminated type array 894 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 895 int index = 0; 896 intptr_t val; 897 898 for (int i = 0; i < sv->field_size(); i++) { 899 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 900 switch(type) { 901 case T_LONG: case T_DOUBLE: { 902 assert(value->type() == T_INT, "Agreement."); 903 StackValue* low = 904 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 905 #ifdef _LP64 906 jlong res = (jlong)low->get_int(); 907 #else 908 #ifdef SPARC 909 // For SPARC we have to swap high and low words. 910 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 911 #else 912 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 913 #endif //SPARC 914 #endif 915 obj->long_at_put(index, res); 916 break; 917 } 918 919 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 920 case T_INT: case T_FLOAT: { // 4 bytes. 921 assert(value->type() == T_INT, "Agreement."); 922 bool big_value = false; 923 if (i + 1 < sv->field_size() && type == T_INT) { 924 if (sv->field_at(i)->is_location()) { 925 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 926 if (type == Location::dbl || type == Location::lng) { 927 big_value = true; 928 } 929 } else if (sv->field_at(i)->is_constant_int()) { 930 ScopeValue* next_scope_field = sv->field_at(i + 1); 931 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 932 big_value = true; 933 } 934 } 935 } 936 937 if (big_value) { 938 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 939 #ifdef _LP64 940 jlong res = (jlong)low->get_int(); 941 #else 942 #ifdef SPARC 943 // For SPARC we have to swap high and low words. 944 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 945 #else 946 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 947 #endif //SPARC 948 #endif 949 obj->int_at_put(index, (jint)*((jint*)&res)); 950 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 951 } else { 952 val = value->get_int(); 953 obj->int_at_put(index, (jint)*((jint*)&val)); 954 } 955 break; 956 } 957 958 case T_SHORT: 959 assert(value->type() == T_INT, "Agreement."); 960 val = value->get_int(); 961 obj->short_at_put(index, (jshort)*((jint*)&val)); 962 break; 963 964 case T_CHAR: 965 assert(value->type() == T_INT, "Agreement."); 966 val = value->get_int(); 967 obj->char_at_put(index, (jchar)*((jint*)&val)); 968 break; 969 970 case T_BYTE: 971 assert(value->type() == T_INT, "Agreement."); 972 val = value->get_int(); 973 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 974 break; 975 976 case T_BOOLEAN: 977 assert(value->type() == T_INT, "Agreement."); 978 val = value->get_int(); 979 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 980 break; 981 982 default: 983 ShouldNotReachHere(); 984 } 985 index++; 986 } 987 } 988 989 990 // restore fields of an eliminated object array 991 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 992 for (int i = 0; i < sv->field_size(); i++) { 993 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 994 assert(value->type() == T_OBJECT, "object element expected"); 995 obj->obj_at_put(i, value->get_obj()()); 996 } 997 } 998 999 class ReassignedField { 1000 public: 1001 int _offset; 1002 BasicType _type; 1003 InstanceKlass* _klass; 1004 public: 1005 ReassignedField() { 1006 _offset = 0; 1007 _type = T_ILLEGAL; 1008 _klass = NULL; 1009 } 1010 }; 1011 1012 int compare(ReassignedField* left, ReassignedField* right) { 1013 return left->_offset - right->_offset; 1014 } 1015 1016 // Restore fields of an eliminated instance object using the same field order 1017 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1018 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) { 1019 if (klass->superklass() != NULL) { 1020 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0); 1021 } 1022 1023 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1024 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1025 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1026 ReassignedField field; 1027 field._offset = fs.offset(); 1028 field._type = FieldType::basic_type(fs.signature()); 1029 if (field._type == T_VALUETYPE) { 1030 if (fs.is_flatten()) { 1031 // Resolve klass of flattened value type field 1032 SignatureStream ss(fs.signature(), false); 1033 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD); 1034 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending"); 1035 assert(vk->is_value(), "must be a ValueKlass"); 1036 field._klass = InstanceKlass::cast(vk); 1037 } else { 1038 // Non-flattened value type field 1039 field._type = T_VALUETYPEPTR; 1040 } 1041 } 1042 fields->append(field); 1043 } 1044 } 1045 fields->sort(compare); 1046 for (int i = 0; i < fields->length(); i++) { 1047 intptr_t val; 1048 ScopeValue* scope_field = sv->field_at(svIndex); 1049 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1050 int offset = base_offset + fields->at(i)._offset; 1051 BasicType type = fields->at(i)._type; 1052 switch (type) { 1053 case T_OBJECT: 1054 case T_VALUETYPEPTR: 1055 case T_ARRAY: 1056 assert(value->type() == T_OBJECT, "Agreement."); 1057 obj->obj_field_put(offset, value->get_obj()()); 1058 break; 1059 1060 case T_VALUETYPE: { 1061 // Recursively re-assign flattened value type fields 1062 InstanceKlass* vk = fields->at(i)._klass; 1063 assert(vk != NULL, "must be resolved"); 1064 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1065 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1066 continue; // Continue because we don't need to increment svIndex 1067 } 1068 1069 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1070 case T_INT: case T_FLOAT: { // 4 bytes. 1071 assert(value->type() == T_INT, "Agreement."); 1072 bool big_value = false; 1073 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1074 if (scope_field->is_location()) { 1075 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1076 if (type == Location::dbl || type == Location::lng) { 1077 big_value = true; 1078 } 1079 } 1080 if (scope_field->is_constant_int()) { 1081 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1082 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1083 big_value = true; 1084 } 1085 } 1086 } 1087 1088 if (big_value) { 1089 i++; 1090 assert(i < fields->length(), "second T_INT field needed"); 1091 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1092 } else { 1093 val = value->get_int(); 1094 obj->int_field_put(offset, (jint)*((jint*)&val)); 1095 break; 1096 } 1097 } 1098 /* no break */ 1099 1100 case T_LONG: case T_DOUBLE: { 1101 assert(value->type() == T_INT, "Agreement."); 1102 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1103 #ifdef _LP64 1104 jlong res = (jlong)low->get_int(); 1105 #else 1106 #ifdef SPARC 1107 // For SPARC we have to swap high and low words. 1108 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1109 #else 1110 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1111 #endif //SPARC 1112 #endif 1113 obj->long_field_put(offset, res); 1114 break; 1115 } 1116 1117 case T_SHORT: 1118 assert(value->type() == T_INT, "Agreement."); 1119 val = value->get_int(); 1120 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1121 break; 1122 1123 case T_CHAR: 1124 assert(value->type() == T_INT, "Agreement."); 1125 val = value->get_int(); 1126 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1127 break; 1128 1129 case T_BYTE: 1130 assert(value->type() == T_INT, "Agreement."); 1131 val = value->get_int(); 1132 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1133 break; 1134 1135 case T_BOOLEAN: 1136 assert(value->type() == T_INT, "Agreement."); 1137 val = value->get_int(); 1138 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1139 break; 1140 1141 default: 1142 ShouldNotReachHere(); 1143 } 1144 svIndex++; 1145 } 1146 return svIndex; 1147 } 1148 1149 // restore fields of an eliminated value type array 1150 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) { 1151 ValueKlass* vk = vak->element_klass(); 1152 assert(vk->flatten_array(), "should only be used for flattened value type arrays"); 1153 // Adjust offset to omit oop header 1154 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset(); 1155 // Initialize all elements of the flattened value type array 1156 for (int i = 0; i < sv->field_size(); i++) { 1157 ScopeValue* val = sv->field_at(i); 1158 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1159 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK); 1160 } 1161 } 1162 1163 // restore fields of all eliminated objects and arrays 1164 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1165 for (int i = 0; i < objects->length(); i++) { 1166 ObjectValue* sv = (ObjectValue*) objects->at(i); 1167 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1168 Handle obj = sv->value(); 1169 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1170 if (PrintDeoptimizationDetails) { 1171 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1172 } 1173 if (obj.is_null()) { 1174 continue; 1175 } 1176 1177 if (k->is_instance_klass()) { 1178 InstanceKlass* ik = InstanceKlass::cast(k); 1179 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1180 } else if (k->is_valueArray_klass()) { 1181 ValueArrayKlass* vak = ValueArrayKlass::cast(k); 1182 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK); 1183 } else if (k->is_typeArray_klass()) { 1184 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1185 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1186 } else if (k->is_objArray_klass()) { 1187 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1188 } 1189 } 1190 } 1191 1192 1193 // relock objects for which synchronization was eliminated 1194 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1195 for (int i = 0; i < monitors->length(); i++) { 1196 MonitorInfo* mon_info = monitors->at(i); 1197 if (mon_info->eliminated()) { 1198 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1199 if (!mon_info->owner_is_scalar_replaced()) { 1200 Handle obj(thread, mon_info->owner()); 1201 markOop mark = obj->mark(); 1202 if (UseBiasedLocking && mark->has_bias_pattern()) { 1203 // New allocated objects may have the mark set to anonymously biased. 1204 // Also the deoptimized method may called methods with synchronization 1205 // where the thread-local object is bias locked to the current thread. 1206 assert(mark->is_biased_anonymously() || 1207 mark->biased_locker() == thread, "should be locked to current thread"); 1208 // Reset mark word to unbiased prototype. 1209 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1210 obj->set_mark(unbiased_prototype); 1211 } 1212 BasicLock* lock = mon_info->lock(); 1213 ObjectSynchronizer::slow_enter(obj, lock, thread); 1214 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1215 } 1216 } 1217 } 1218 } 1219 1220 1221 #ifndef PRODUCT 1222 // print information about reallocated objects 1223 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1224 fieldDescriptor fd; 1225 1226 for (int i = 0; i < objects->length(); i++) { 1227 ObjectValue* sv = (ObjectValue*) objects->at(i); 1228 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1229 Handle obj = sv->value(); 1230 1231 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1232 k->print_value(); 1233 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1234 if (obj.is_null()) { 1235 tty->print(" allocation failed"); 1236 } else { 1237 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1238 } 1239 tty->cr(); 1240 1241 if (Verbose && !obj.is_null()) { 1242 k->oop_print_on(obj(), tty); 1243 } 1244 } 1245 } 1246 #endif 1247 #endif // COMPILER2 || INCLUDE_JVMCI 1248 1249 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1250 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1251 1252 #ifndef PRODUCT 1253 if (PrintDeoptimizationDetails) { 1254 ttyLocker ttyl; 1255 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1256 fr.print_on(tty); 1257 tty->print_cr(" Virtual frames (innermost first):"); 1258 for (int index = 0; index < chunk->length(); index++) { 1259 compiledVFrame* vf = chunk->at(index); 1260 tty->print(" %2d - ", index); 1261 vf->print_value(); 1262 int bci = chunk->at(index)->raw_bci(); 1263 const char* code_name; 1264 if (bci == SynchronizationEntryBCI) { 1265 code_name = "sync entry"; 1266 } else { 1267 Bytecodes::Code code = vf->method()->code_at(bci); 1268 code_name = Bytecodes::name(code); 1269 } 1270 tty->print(" - %s", code_name); 1271 tty->print_cr(" @ bci %d ", bci); 1272 if (Verbose) { 1273 vf->print(); 1274 tty->cr(); 1275 } 1276 } 1277 } 1278 #endif 1279 1280 // Register map for next frame (used for stack crawl). We capture 1281 // the state of the deopt'ing frame's caller. Thus if we need to 1282 // stuff a C2I adapter we can properly fill in the callee-save 1283 // register locations. 1284 frame caller = fr.sender(reg_map); 1285 int frame_size = caller.sp() - fr.sp(); 1286 1287 frame sender = caller; 1288 1289 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1290 // the vframeArray containing the unpacking information is allocated in the C heap. 1291 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1292 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1293 1294 // Compare the vframeArray to the collected vframes 1295 assert(array->structural_compare(thread, chunk), "just checking"); 1296 1297 #ifndef PRODUCT 1298 if (PrintDeoptimizationDetails) { 1299 ttyLocker ttyl; 1300 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1301 } 1302 #endif // PRODUCT 1303 1304 return array; 1305 } 1306 1307 #if defined(COMPILER2) || INCLUDE_JVMCI 1308 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1309 // Reallocation of some scalar replaced objects failed. Record 1310 // that we need to pop all the interpreter frames for the 1311 // deoptimized compiled frame. 1312 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1313 thread->set_frames_to_pop_failed_realloc(array->frames()); 1314 // Unlock all monitors here otherwise the interpreter will see a 1315 // mix of locked and unlocked monitors (because of failed 1316 // reallocations of synchronized objects) and be confused. 1317 for (int i = 0; i < array->frames(); i++) { 1318 MonitorChunk* monitors = array->element(i)->monitors(); 1319 if (monitors != NULL) { 1320 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1321 BasicObjectLock* src = monitors->at(j); 1322 if (src->obj() != NULL) { 1323 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1324 } 1325 } 1326 array->element(i)->free_monitors(thread); 1327 #ifdef ASSERT 1328 array->element(i)->set_removed_monitors(); 1329 #endif 1330 } 1331 } 1332 } 1333 #endif 1334 1335 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1336 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1337 Thread* thread = Thread::current(); 1338 for (int i = 0; i < monitors->length(); i++) { 1339 MonitorInfo* mon_info = monitors->at(i); 1340 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1341 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1342 } 1343 } 1344 } 1345 1346 1347 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1348 if (!UseBiasedLocking) { 1349 return; 1350 } 1351 1352 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1353 1354 // Unfortunately we don't have a RegisterMap available in most of 1355 // the places we want to call this routine so we need to walk the 1356 // stack again to update the register map. 1357 if (map == NULL || !map->update_map()) { 1358 StackFrameStream sfs(thread, true); 1359 bool found = false; 1360 while (!found && !sfs.is_done()) { 1361 frame* cur = sfs.current(); 1362 sfs.next(); 1363 found = cur->id() == fr.id(); 1364 } 1365 assert(found, "frame to be deoptimized not found on target thread's stack"); 1366 map = sfs.register_map(); 1367 } 1368 1369 vframe* vf = vframe::new_vframe(&fr, map, thread); 1370 compiledVFrame* cvf = compiledVFrame::cast(vf); 1371 // Revoke monitors' biases in all scopes 1372 while (!cvf->is_top()) { 1373 collect_monitors(cvf, objects_to_revoke); 1374 cvf = compiledVFrame::cast(cvf->sender()); 1375 } 1376 collect_monitors(cvf, objects_to_revoke); 1377 1378 if (SafepointSynchronize::is_at_safepoint()) { 1379 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1380 } else { 1381 BiasedLocking::revoke(objects_to_revoke); 1382 } 1383 } 1384 1385 1386 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1387 if (!UseBiasedLocking) { 1388 return; 1389 } 1390 1391 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1392 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1393 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1394 if (jt->has_last_Java_frame()) { 1395 StackFrameStream sfs(jt, true); 1396 while (!sfs.is_done()) { 1397 frame* cur = sfs.current(); 1398 if (cb->contains(cur->pc())) { 1399 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1400 compiledVFrame* cvf = compiledVFrame::cast(vf); 1401 // Revoke monitors' biases in all scopes 1402 while (!cvf->is_top()) { 1403 collect_monitors(cvf, objects_to_revoke); 1404 cvf = compiledVFrame::cast(cvf->sender()); 1405 } 1406 collect_monitors(cvf, objects_to_revoke); 1407 } 1408 sfs.next(); 1409 } 1410 } 1411 } 1412 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1413 } 1414 1415 1416 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1417 assert(fr.can_be_deoptimized(), "checking frame type"); 1418 1419 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1420 1421 if (LogCompilation && xtty != NULL) { 1422 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1423 assert(cm != NULL, "only compiled methods can deopt"); 1424 1425 ttyLocker ttyl; 1426 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1427 cm->log_identity(xtty); 1428 xtty->end_head(); 1429 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1430 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1431 xtty->method(sd->method()); 1432 xtty->end_elem(); 1433 if (sd->is_top()) break; 1434 } 1435 xtty->tail("deoptimized"); 1436 } 1437 1438 // Patch the compiled method so that when execution returns to it we will 1439 // deopt the execution state and return to the interpreter. 1440 fr.deoptimize(thread); 1441 } 1442 1443 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1444 deoptimize(thread, fr, map, Reason_constraint); 1445 } 1446 1447 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1448 // Deoptimize only if the frame comes from compile code. 1449 // Do not deoptimize the frame which is already patched 1450 // during the execution of the loops below. 1451 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1452 return; 1453 } 1454 ResourceMark rm; 1455 DeoptimizationMarker dm; 1456 if (UseBiasedLocking) { 1457 revoke_biases_of_monitors(thread, fr, map); 1458 } 1459 deoptimize_single_frame(thread, fr, reason); 1460 1461 } 1462 1463 #if INCLUDE_JVMCI 1464 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1465 // there is no exception handler for this pc => deoptimize 1466 cm->make_not_entrant(); 1467 1468 // Use Deoptimization::deoptimize for all of its side-effects: 1469 // revoking biases of monitors, gathering traps statistics, logging... 1470 // it also patches the return pc but we do not care about that 1471 // since we return a continuation to the deopt_blob below. 1472 JavaThread* thread = JavaThread::current(); 1473 RegisterMap reg_map(thread, UseBiasedLocking); 1474 frame runtime_frame = thread->last_frame(); 1475 frame caller_frame = runtime_frame.sender(®_map); 1476 assert(caller_frame.cb()->as_nmethod_or_null() == cm, "expect top frame nmethod"); 1477 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1478 1479 MethodData* trap_mdo = get_method_data(thread, cm->method(), true); 1480 if (trap_mdo != NULL) { 1481 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1482 } 1483 1484 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1485 } 1486 #endif 1487 1488 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1489 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1490 "can only deoptimize other thread at a safepoint"); 1491 // Compute frame and register map based on thread and sp. 1492 RegisterMap reg_map(thread, UseBiasedLocking); 1493 frame fr = thread->last_frame(); 1494 while (fr.id() != id) { 1495 fr = fr.sender(®_map); 1496 } 1497 deoptimize(thread, fr, ®_map, reason); 1498 } 1499 1500 1501 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1502 if (thread == Thread::current()) { 1503 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1504 } else { 1505 VM_DeoptimizeFrame deopt(thread, id, reason); 1506 VMThread::execute(&deopt); 1507 } 1508 } 1509 1510 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1511 deoptimize_frame(thread, id, Reason_constraint); 1512 } 1513 1514 // JVMTI PopFrame support 1515 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1516 { 1517 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1518 } 1519 JRT_END 1520 1521 MethodData* 1522 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1523 bool create_if_missing) { 1524 Thread* THREAD = thread; 1525 MethodData* mdo = m()->method_data(); 1526 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1527 // Build an MDO. Ignore errors like OutOfMemory; 1528 // that simply means we won't have an MDO to update. 1529 Method::build_interpreter_method_data(m, THREAD); 1530 if (HAS_PENDING_EXCEPTION) { 1531 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1532 CLEAR_PENDING_EXCEPTION; 1533 } 1534 mdo = m()->method_data(); 1535 } 1536 return mdo; 1537 } 1538 1539 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI 1540 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1541 // in case of an unresolved klass entry, load the class. 1542 if (constant_pool->tag_at(index).is_unresolved_klass() || 1543 constant_pool->tag_at(index).is_unresolved_value_type()) { 1544 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1545 return; 1546 } 1547 1548 if (!constant_pool->tag_at(index).is_symbol()) return; 1549 1550 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1551 Symbol* symbol = constant_pool->symbol_at(index); 1552 1553 // class name? 1554 if (symbol->byte_at(0) != '(') { 1555 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1556 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1557 return; 1558 } 1559 1560 // then it must be a signature! 1561 ResourceMark rm(THREAD); 1562 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1563 if (ss.is_object()) { 1564 Symbol* class_name = ss.as_symbol(CHECK); 1565 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1566 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1567 } 1568 } 1569 } 1570 1571 1572 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1573 EXCEPTION_MARK; 1574 load_class_by_index(constant_pool, index, THREAD); 1575 if (HAS_PENDING_EXCEPTION) { 1576 // Exception happened during classloading. We ignore the exception here, since it 1577 // is going to be rethrown since the current activation is going to be deoptimized and 1578 // the interpreter will re-execute the bytecode. 1579 CLEAR_PENDING_EXCEPTION; 1580 // Class loading called java code which may have caused a stack 1581 // overflow. If the exception was thrown right before the return 1582 // to the runtime the stack is no longer guarded. Reguard the 1583 // stack otherwise if we return to the uncommon trap blob and the 1584 // stack bang causes a stack overflow we crash. 1585 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1586 JavaThread* thread = (JavaThread*)THREAD; 1587 bool guard_pages_enabled = thread->stack_guards_enabled(); 1588 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1589 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1590 } 1591 } 1592 1593 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1594 HandleMark hm; 1595 1596 // uncommon_trap() is called at the beginning of the uncommon trap 1597 // handler. Note this fact before we start generating temporary frames 1598 // that can confuse an asynchronous stack walker. This counter is 1599 // decremented at the end of unpack_frames(). 1600 thread->inc_in_deopt_handler(); 1601 1602 // We need to update the map if we have biased locking. 1603 #if INCLUDE_JVMCI 1604 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1605 RegisterMap reg_map(thread, true); 1606 #else 1607 RegisterMap reg_map(thread, UseBiasedLocking); 1608 #endif 1609 frame stub_frame = thread->last_frame(); 1610 frame fr = stub_frame.sender(®_map); 1611 // Make sure the calling nmethod is not getting deoptimized and removed 1612 // before we are done with it. 1613 nmethodLocker nl(fr.pc()); 1614 1615 // Log a message 1616 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1617 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1618 1619 { 1620 ResourceMark rm; 1621 1622 // Revoke biases of any monitors in the frame to ensure we can migrate them 1623 revoke_biases_of_monitors(thread, fr, ®_map); 1624 1625 DeoptReason reason = trap_request_reason(trap_request); 1626 DeoptAction action = trap_request_action(trap_request); 1627 #if INCLUDE_JVMCI 1628 int debug_id = trap_request_debug_id(trap_request); 1629 #endif 1630 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1631 1632 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1633 compiledVFrame* cvf = compiledVFrame::cast(vf); 1634 1635 CompiledMethod* nm = cvf->code(); 1636 1637 ScopeDesc* trap_scope = cvf->scope(); 1638 1639 if (TraceDeoptimization) { 1640 ttyLocker ttyl; 1641 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() 1642 #if INCLUDE_JVMCI 1643 , debug_id 1644 #endif 1645 ); 1646 } 1647 1648 methodHandle trap_method = trap_scope->method(); 1649 int trap_bci = trap_scope->bci(); 1650 #if INCLUDE_JVMCI 1651 oop speculation = thread->pending_failed_speculation(); 1652 if (nm->is_compiled_by_jvmci()) { 1653 if (speculation != NULL) { 1654 oop speculation_log = nm->as_nmethod()->speculation_log(); 1655 if (speculation_log != NULL) { 1656 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1657 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) { 1658 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1659 } 1660 } 1661 if (TraceDeoptimization) { 1662 tty->print_cr("Saving speculation to speculation log"); 1663 } 1664 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1665 } else { 1666 if (TraceDeoptimization) { 1667 tty->print_cr("Speculation present but no speculation log"); 1668 } 1669 } 1670 thread->set_pending_failed_speculation(NULL); 1671 } else { 1672 if (TraceDeoptimization) { 1673 tty->print_cr("No speculation"); 1674 } 1675 } 1676 } else { 1677 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1678 } 1679 1680 if (trap_bci == SynchronizationEntryBCI) { 1681 trap_bci = 0; 1682 thread->set_pending_monitorenter(true); 1683 } 1684 1685 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1686 thread->set_pending_transfer_to_interpreter(true); 1687 } 1688 #endif 1689 1690 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1691 // Record this event in the histogram. 1692 gather_statistics(reason, action, trap_bc); 1693 1694 // Ensure that we can record deopt. history: 1695 // Need MDO to record RTM code generation state. 1696 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1697 1698 methodHandle profiled_method; 1699 #if INCLUDE_JVMCI 1700 if (nm->is_compiled_by_jvmci()) { 1701 profiled_method = nm->method(); 1702 } else { 1703 profiled_method = trap_method; 1704 } 1705 #else 1706 profiled_method = trap_method; 1707 #endif 1708 1709 MethodData* trap_mdo = 1710 get_method_data(thread, profiled_method, create_if_missing); 1711 1712 // Log a message 1713 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1714 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1715 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1716 1717 // Print a bunch of diagnostics, if requested. 1718 if (TraceDeoptimization || LogCompilation) { 1719 ResourceMark rm; 1720 ttyLocker ttyl; 1721 char buf[100]; 1722 if (xtty != NULL) { 1723 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1724 os::current_thread_id(), 1725 format_trap_request(buf, sizeof(buf), trap_request)); 1726 nm->log_identity(xtty); 1727 } 1728 Symbol* class_name = NULL; 1729 bool unresolved = false; 1730 if (unloaded_class_index >= 0) { 1731 constantPoolHandle constants (THREAD, trap_method->constants()); 1732 if (constants->tag_at(unloaded_class_index).is_unresolved_klass() || 1733 constants->tag_at(unloaded_class_index).is_unresolved_value_type()) { 1734 class_name = constants->klass_name_at(unloaded_class_index); 1735 unresolved = true; 1736 if (xtty != NULL) 1737 xtty->print(" unresolved='1'"); 1738 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1739 class_name = constants->symbol_at(unloaded_class_index); 1740 } 1741 if (xtty != NULL) 1742 xtty->name(class_name); 1743 } 1744 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1745 // Dump the relevant MDO state. 1746 // This is the deopt count for the current reason, any previous 1747 // reasons or recompiles seen at this point. 1748 int dcnt = trap_mdo->trap_count(reason); 1749 if (dcnt != 0) 1750 xtty->print(" count='%d'", dcnt); 1751 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1752 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1753 if (dos != 0) { 1754 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1755 if (trap_state_is_recompiled(dos)) { 1756 int recnt2 = trap_mdo->overflow_recompile_count(); 1757 if (recnt2 != 0) 1758 xtty->print(" recompiles2='%d'", recnt2); 1759 } 1760 } 1761 } 1762 if (xtty != NULL) { 1763 xtty->stamp(); 1764 xtty->end_head(); 1765 } 1766 if (TraceDeoptimization) { // make noise on the tty 1767 tty->print("Uncommon trap occurred in"); 1768 nm->method()->print_short_name(tty); 1769 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1770 #if INCLUDE_JVMCI 1771 if (nm->is_nmethod()) { 1772 oop installedCode = nm->as_nmethod()->jvmci_installed_code(); 1773 if (installedCode != NULL) { 1774 oop installedCodeName = NULL; 1775 if (installedCode->is_a(InstalledCode::klass())) { 1776 installedCodeName = InstalledCode::name(installedCode); 1777 } 1778 if (installedCodeName != NULL) { 1779 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName)); 1780 } else { 1781 tty->print(" (JVMCI: installed code has no name) "); 1782 } 1783 } else if (nm->is_compiled_by_jvmci()) { 1784 tty->print(" (JVMCI: no installed code) "); 1785 } 1786 } 1787 #endif 1788 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1789 p2i(fr.pc()), 1790 os::current_thread_id(), 1791 trap_reason_name(reason), 1792 trap_action_name(action), 1793 unloaded_class_index 1794 #if INCLUDE_JVMCI 1795 , debug_id 1796 #endif 1797 ); 1798 if (class_name != NULL) { 1799 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1800 class_name->print_symbol_on(tty); 1801 } 1802 tty->cr(); 1803 } 1804 if (xtty != NULL) { 1805 // Log the precise location of the trap. 1806 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1807 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1808 xtty->method(sd->method()); 1809 xtty->end_elem(); 1810 if (sd->is_top()) break; 1811 } 1812 xtty->tail("uncommon_trap"); 1813 } 1814 } 1815 // (End diagnostic printout.) 1816 1817 // Load class if necessary 1818 if (unloaded_class_index >= 0) { 1819 constantPoolHandle constants(THREAD, trap_method->constants()); 1820 load_class_by_index(constants, unloaded_class_index); 1821 } 1822 1823 // Flush the nmethod if necessary and desirable. 1824 // 1825 // We need to avoid situations where we are re-flushing the nmethod 1826 // because of a hot deoptimization site. Repeated flushes at the same 1827 // point need to be detected by the compiler and avoided. If the compiler 1828 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1829 // module must take measures to avoid an infinite cycle of recompilation 1830 // and deoptimization. There are several such measures: 1831 // 1832 // 1. If a recompilation is ordered a second time at some site X 1833 // and for the same reason R, the action is adjusted to 'reinterpret', 1834 // to give the interpreter time to exercise the method more thoroughly. 1835 // If this happens, the method's overflow_recompile_count is incremented. 1836 // 1837 // 2. If the compiler fails to reduce the deoptimization rate, then 1838 // the method's overflow_recompile_count will begin to exceed the set 1839 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1840 // is adjusted to 'make_not_compilable', and the method is abandoned 1841 // to the interpreter. This is a performance hit for hot methods, 1842 // but is better than a disastrous infinite cycle of recompilations. 1843 // (Actually, only the method containing the site X is abandoned.) 1844 // 1845 // 3. In parallel with the previous measures, if the total number of 1846 // recompilations of a method exceeds the much larger set limit 1847 // PerMethodRecompilationCutoff, the method is abandoned. 1848 // This should only happen if the method is very large and has 1849 // many "lukewarm" deoptimizations. The code which enforces this 1850 // limit is elsewhere (class nmethod, class Method). 1851 // 1852 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1853 // to recompile at each bytecode independently of the per-BCI cutoff. 1854 // 1855 // The decision to update code is up to the compiler, and is encoded 1856 // in the Action_xxx code. If the compiler requests Action_none 1857 // no trap state is changed, no compiled code is changed, and the 1858 // computation suffers along in the interpreter. 1859 // 1860 // The other action codes specify various tactics for decompilation 1861 // and recompilation. Action_maybe_recompile is the loosest, and 1862 // allows the compiled code to stay around until enough traps are seen, 1863 // and until the compiler gets around to recompiling the trapping method. 1864 // 1865 // The other actions cause immediate removal of the present code. 1866 1867 // Traps caused by injected profile shouldn't pollute trap counts. 1868 bool injected_profile_trap = trap_method->has_injected_profile() && 1869 (reason == Reason_intrinsic || reason == Reason_unreached); 1870 1871 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1872 bool make_not_entrant = false; 1873 bool make_not_compilable = false; 1874 bool reprofile = false; 1875 switch (action) { 1876 case Action_none: 1877 // Keep the old code. 1878 update_trap_state = false; 1879 break; 1880 case Action_maybe_recompile: 1881 // Do not need to invalidate the present code, but we can 1882 // initiate another 1883 // Start compiler without (necessarily) invalidating the nmethod. 1884 // The system will tolerate the old code, but new code should be 1885 // generated when possible. 1886 break; 1887 case Action_reinterpret: 1888 // Go back into the interpreter for a while, and then consider 1889 // recompiling form scratch. 1890 make_not_entrant = true; 1891 // Reset invocation counter for outer most method. 1892 // This will allow the interpreter to exercise the bytecodes 1893 // for a while before recompiling. 1894 // By contrast, Action_make_not_entrant is immediate. 1895 // 1896 // Note that the compiler will track null_check, null_assert, 1897 // range_check, and class_check events and log them as if they 1898 // had been traps taken from compiled code. This will update 1899 // the MDO trap history so that the next compilation will 1900 // properly detect hot trap sites. 1901 reprofile = true; 1902 break; 1903 case Action_make_not_entrant: 1904 // Request immediate recompilation, and get rid of the old code. 1905 // Make them not entrant, so next time they are called they get 1906 // recompiled. Unloaded classes are loaded now so recompile before next 1907 // time they are called. Same for uninitialized. The interpreter will 1908 // link the missing class, if any. 1909 make_not_entrant = true; 1910 break; 1911 case Action_make_not_compilable: 1912 // Give up on compiling this method at all. 1913 make_not_entrant = true; 1914 make_not_compilable = true; 1915 break; 1916 default: 1917 ShouldNotReachHere(); 1918 } 1919 1920 // Setting +ProfileTraps fixes the following, on all platforms: 1921 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1922 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1923 // recompile relies on a MethodData* to record heroic opt failures. 1924 1925 // Whether the interpreter is producing MDO data or not, we also need 1926 // to use the MDO to detect hot deoptimization points and control 1927 // aggressive optimization. 1928 bool inc_recompile_count = false; 1929 ProfileData* pdata = NULL; 1930 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1931 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1932 uint this_trap_count = 0; 1933 bool maybe_prior_trap = false; 1934 bool maybe_prior_recompile = false; 1935 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1936 #if INCLUDE_JVMCI 1937 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1938 #endif 1939 nm->method(), 1940 //outputs: 1941 this_trap_count, 1942 maybe_prior_trap, 1943 maybe_prior_recompile); 1944 // Because the interpreter also counts null, div0, range, and class 1945 // checks, these traps from compiled code are double-counted. 1946 // This is harmless; it just means that the PerXTrapLimit values 1947 // are in effect a little smaller than they look. 1948 1949 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1950 if (per_bc_reason != Reason_none) { 1951 // Now take action based on the partially known per-BCI history. 1952 if (maybe_prior_trap 1953 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1954 // If there are too many traps at this BCI, force a recompile. 1955 // This will allow the compiler to see the limit overflow, and 1956 // take corrective action, if possible. The compiler generally 1957 // does not use the exact PerBytecodeTrapLimit value, but instead 1958 // changes its tactics if it sees any traps at all. This provides 1959 // a little hysteresis, delaying a recompile until a trap happens 1960 // several times. 1961 // 1962 // Actually, since there is only one bit of counter per BCI, 1963 // the possible per-BCI counts are {0,1,(per-method count)}. 1964 // This produces accurate results if in fact there is only 1965 // one hot trap site, but begins to get fuzzy if there are 1966 // many sites. For example, if there are ten sites each 1967 // trapping two or more times, they each get the blame for 1968 // all of their traps. 1969 make_not_entrant = true; 1970 } 1971 1972 // Detect repeated recompilation at the same BCI, and enforce a limit. 1973 if (make_not_entrant && maybe_prior_recompile) { 1974 // More than one recompile at this point. 1975 inc_recompile_count = maybe_prior_trap; 1976 } 1977 } else { 1978 // For reasons which are not recorded per-bytecode, we simply 1979 // force recompiles unconditionally. 1980 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1981 make_not_entrant = true; 1982 } 1983 1984 // Go back to the compiler if there are too many traps in this method. 1985 if (this_trap_count >= per_method_trap_limit(reason)) { 1986 // If there are too many traps in this method, force a recompile. 1987 // This will allow the compiler to see the limit overflow, and 1988 // take corrective action, if possible. 1989 // (This condition is an unlikely backstop only, because the 1990 // PerBytecodeTrapLimit is more likely to take effect first, 1991 // if it is applicable.) 1992 make_not_entrant = true; 1993 } 1994 1995 // Here's more hysteresis: If there has been a recompile at 1996 // this trap point already, run the method in the interpreter 1997 // for a while to exercise it more thoroughly. 1998 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1999 reprofile = true; 2000 } 2001 } 2002 2003 // Take requested actions on the method: 2004 2005 // Recompile 2006 if (make_not_entrant) { 2007 if (!nm->make_not_entrant()) { 2008 return; // the call did not change nmethod's state 2009 } 2010 2011 if (pdata != NULL) { 2012 // Record the recompilation event, if any. 2013 int tstate0 = pdata->trap_state(); 2014 int tstate1 = trap_state_set_recompiled(tstate0, true); 2015 if (tstate1 != tstate0) 2016 pdata->set_trap_state(tstate1); 2017 } 2018 2019 #if INCLUDE_RTM_OPT 2020 // Restart collecting RTM locking abort statistic if the method 2021 // is recompiled for a reason other than RTM state change. 2022 // Assume that in new recompiled code the statistic could be different, 2023 // for example, due to different inlining. 2024 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2025 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2026 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2027 } 2028 #endif 2029 // For code aging we count traps separately here, using make_not_entrant() 2030 // as a guard against simultaneous deopts in multiple threads. 2031 if (reason == Reason_tenured && trap_mdo != NULL) { 2032 trap_mdo->inc_tenure_traps(); 2033 } 2034 } 2035 2036 if (inc_recompile_count) { 2037 trap_mdo->inc_overflow_recompile_count(); 2038 if ((uint)trap_mdo->overflow_recompile_count() > 2039 (uint)PerBytecodeRecompilationCutoff) { 2040 // Give up on the method containing the bad BCI. 2041 if (trap_method() == nm->method()) { 2042 make_not_compilable = true; 2043 } else { 2044 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 2045 // But give grace to the enclosing nm->method(). 2046 } 2047 } 2048 } 2049 2050 // Reprofile 2051 if (reprofile) { 2052 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2053 } 2054 2055 // Give up compiling 2056 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2057 assert(make_not_entrant, "consistent"); 2058 nm->method()->set_not_compilable(CompLevel_full_optimization); 2059 } 2060 2061 } // Free marked resources 2062 2063 } 2064 JRT_END 2065 2066 ProfileData* 2067 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2068 int trap_bci, 2069 Deoptimization::DeoptReason reason, 2070 bool update_total_trap_count, 2071 #if INCLUDE_JVMCI 2072 bool is_osr, 2073 #endif 2074 Method* compiled_method, 2075 //outputs: 2076 uint& ret_this_trap_count, 2077 bool& ret_maybe_prior_trap, 2078 bool& ret_maybe_prior_recompile) { 2079 bool maybe_prior_trap = false; 2080 bool maybe_prior_recompile = false; 2081 uint this_trap_count = 0; 2082 if (update_total_trap_count) { 2083 uint idx = reason; 2084 #if INCLUDE_JVMCI 2085 if (is_osr) { 2086 idx += Reason_LIMIT; 2087 } 2088 #endif 2089 uint prior_trap_count = trap_mdo->trap_count(idx); 2090 this_trap_count = trap_mdo->inc_trap_count(idx); 2091 2092 // If the runtime cannot find a place to store trap history, 2093 // it is estimated based on the general condition of the method. 2094 // If the method has ever been recompiled, or has ever incurred 2095 // a trap with the present reason , then this BCI is assumed 2096 // (pessimistically) to be the culprit. 2097 maybe_prior_trap = (prior_trap_count != 0); 2098 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2099 } 2100 ProfileData* pdata = NULL; 2101 2102 2103 // For reasons which are recorded per bytecode, we check per-BCI data. 2104 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2105 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2106 if (per_bc_reason != Reason_none) { 2107 // Find the profile data for this BCI. If there isn't one, 2108 // try to allocate one from the MDO's set of spares. 2109 // This will let us detect a repeated trap at this point. 2110 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2111 2112 if (pdata != NULL) { 2113 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2114 if (LogCompilation && xtty != NULL) { 2115 ttyLocker ttyl; 2116 // no more room for speculative traps in this MDO 2117 xtty->elem("speculative_traps_oom"); 2118 } 2119 } 2120 // Query the trap state of this profile datum. 2121 int tstate0 = pdata->trap_state(); 2122 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2123 maybe_prior_trap = false; 2124 if (!trap_state_is_recompiled(tstate0)) 2125 maybe_prior_recompile = false; 2126 2127 // Update the trap state of this profile datum. 2128 int tstate1 = tstate0; 2129 // Record the reason. 2130 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2131 // Store the updated state on the MDO, for next time. 2132 if (tstate1 != tstate0) 2133 pdata->set_trap_state(tstate1); 2134 } else { 2135 if (LogCompilation && xtty != NULL) { 2136 ttyLocker ttyl; 2137 // Missing MDP? Leave a small complaint in the log. 2138 xtty->elem("missing_mdp bci='%d'", trap_bci); 2139 } 2140 } 2141 } 2142 2143 // Return results: 2144 ret_this_trap_count = this_trap_count; 2145 ret_maybe_prior_trap = maybe_prior_trap; 2146 ret_maybe_prior_recompile = maybe_prior_recompile; 2147 return pdata; 2148 } 2149 2150 void 2151 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2152 ResourceMark rm; 2153 // Ignored outputs: 2154 uint ignore_this_trap_count; 2155 bool ignore_maybe_prior_trap; 2156 bool ignore_maybe_prior_recompile; 2157 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2158 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2159 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true); 2160 query_update_method_data(trap_mdo, trap_bci, 2161 (DeoptReason)reason, 2162 update_total_counts, 2163 #if INCLUDE_JVMCI 2164 false, 2165 #endif 2166 NULL, 2167 ignore_this_trap_count, 2168 ignore_maybe_prior_trap, 2169 ignore_maybe_prior_recompile); 2170 } 2171 2172 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2173 if (TraceDeoptimization) { 2174 tty->print("Uncommon trap "); 2175 } 2176 // Still in Java no safepoints 2177 { 2178 // This enters VM and may safepoint 2179 uncommon_trap_inner(thread, trap_request); 2180 } 2181 return fetch_unroll_info_helper(thread, exec_mode); 2182 } 2183 2184 // Local derived constants. 2185 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2186 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 2187 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2188 2189 //---------------------------trap_state_reason--------------------------------- 2190 Deoptimization::DeoptReason 2191 Deoptimization::trap_state_reason(int trap_state) { 2192 // This assert provides the link between the width of DataLayout::trap_bits 2193 // and the encoding of "recorded" reasons. It ensures there are enough 2194 // bits to store all needed reasons in the per-BCI MDO profile. 2195 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2196 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2197 trap_state -= recompile_bit; 2198 if (trap_state == DS_REASON_MASK) { 2199 return Reason_many; 2200 } else { 2201 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2202 return (DeoptReason)trap_state; 2203 } 2204 } 2205 //-------------------------trap_state_has_reason------------------------------- 2206 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2207 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2208 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2209 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2210 trap_state -= recompile_bit; 2211 if (trap_state == DS_REASON_MASK) { 2212 return -1; // true, unspecifically (bottom of state lattice) 2213 } else if (trap_state == reason) { 2214 return 1; // true, definitely 2215 } else if (trap_state == 0) { 2216 return 0; // false, definitely (top of state lattice) 2217 } else { 2218 return 0; // false, definitely 2219 } 2220 } 2221 //-------------------------trap_state_add_reason------------------------------- 2222 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2223 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2224 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2225 trap_state -= recompile_bit; 2226 if (trap_state == DS_REASON_MASK) { 2227 return trap_state + recompile_bit; // already at state lattice bottom 2228 } else if (trap_state == reason) { 2229 return trap_state + recompile_bit; // the condition is already true 2230 } else if (trap_state == 0) { 2231 return reason + recompile_bit; // no condition has yet been true 2232 } else { 2233 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2234 } 2235 } 2236 //-----------------------trap_state_is_recompiled------------------------------ 2237 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2238 return (trap_state & DS_RECOMPILE_BIT) != 0; 2239 } 2240 //-----------------------trap_state_set_recompiled----------------------------- 2241 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2242 if (z) return trap_state | DS_RECOMPILE_BIT; 2243 else return trap_state & ~DS_RECOMPILE_BIT; 2244 } 2245 //---------------------------format_trap_state--------------------------------- 2246 // This is used for debugging and diagnostics, including LogFile output. 2247 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2248 int trap_state) { 2249 assert(buflen > 0, "sanity"); 2250 DeoptReason reason = trap_state_reason(trap_state); 2251 bool recomp_flag = trap_state_is_recompiled(trap_state); 2252 // Re-encode the state from its decoded components. 2253 int decoded_state = 0; 2254 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2255 decoded_state = trap_state_add_reason(decoded_state, reason); 2256 if (recomp_flag) 2257 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2258 // If the state re-encodes properly, format it symbolically. 2259 // Because this routine is used for debugging and diagnostics, 2260 // be robust even if the state is a strange value. 2261 size_t len; 2262 if (decoded_state != trap_state) { 2263 // Random buggy state that doesn't decode?? 2264 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2265 } else { 2266 len = jio_snprintf(buf, buflen, "%s%s", 2267 trap_reason_name(reason), 2268 recomp_flag ? " recompiled" : ""); 2269 } 2270 return buf; 2271 } 2272 2273 2274 //--------------------------------statics-------------------------------------- 2275 const char* Deoptimization::_trap_reason_name[] = { 2276 // Note: Keep this in sync. with enum DeoptReason. 2277 "none", 2278 "null_check", 2279 "null_assert" JVMCI_ONLY("_or_unreached0"), 2280 "range_check", 2281 "class_check", 2282 "array_check", 2283 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2284 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2285 "unloaded", 2286 "uninitialized", 2287 "unreached", 2288 "unhandled", 2289 "constraint", 2290 "div0_check", 2291 "age", 2292 "predicate", 2293 "loop_limit_check", 2294 "speculate_class_check", 2295 "speculate_null_check", 2296 "speculate_null_assert", 2297 "rtm_state_change", 2298 "unstable_if", 2299 "unstable_fused_if", 2300 #if INCLUDE_JVMCI 2301 "aliasing", 2302 "transfer_to_interpreter", 2303 "not_compiled_exception_handler", 2304 "unresolved", 2305 "jsr_mismatch", 2306 #endif 2307 "tenured" 2308 }; 2309 const char* Deoptimization::_trap_action_name[] = { 2310 // Note: Keep this in sync. with enum DeoptAction. 2311 "none", 2312 "maybe_recompile", 2313 "reinterpret", 2314 "make_not_entrant", 2315 "make_not_compilable" 2316 }; 2317 2318 const char* Deoptimization::trap_reason_name(int reason) { 2319 // Check that every reason has a name 2320 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2321 2322 if (reason == Reason_many) return "many"; 2323 if ((uint)reason < Reason_LIMIT) 2324 return _trap_reason_name[reason]; 2325 static char buf[20]; 2326 sprintf(buf, "reason%d", reason); 2327 return buf; 2328 } 2329 const char* Deoptimization::trap_action_name(int action) { 2330 // Check that every action has a name 2331 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2332 2333 if ((uint)action < Action_LIMIT) 2334 return _trap_action_name[action]; 2335 static char buf[20]; 2336 sprintf(buf, "action%d", action); 2337 return buf; 2338 } 2339 2340 // This is used for debugging and diagnostics, including LogFile output. 2341 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2342 int trap_request) { 2343 jint unloaded_class_index = trap_request_index(trap_request); 2344 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2345 const char* action = trap_action_name(trap_request_action(trap_request)); 2346 #if INCLUDE_JVMCI 2347 int debug_id = trap_request_debug_id(trap_request); 2348 #endif 2349 size_t len; 2350 if (unloaded_class_index < 0) { 2351 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2352 reason, action 2353 #if INCLUDE_JVMCI 2354 ,debug_id 2355 #endif 2356 ); 2357 } else { 2358 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2359 reason, action, unloaded_class_index 2360 #if INCLUDE_JVMCI 2361 ,debug_id 2362 #endif 2363 ); 2364 } 2365 return buf; 2366 } 2367 2368 juint Deoptimization::_deoptimization_hist 2369 [Deoptimization::Reason_LIMIT] 2370 [1 + Deoptimization::Action_LIMIT] 2371 [Deoptimization::BC_CASE_LIMIT] 2372 = {0}; 2373 2374 enum { 2375 LSB_BITS = 8, 2376 LSB_MASK = right_n_bits(LSB_BITS) 2377 }; 2378 2379 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2380 Bytecodes::Code bc) { 2381 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2382 assert(action >= 0 && action < Action_LIMIT, "oob"); 2383 _deoptimization_hist[Reason_none][0][0] += 1; // total 2384 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2385 juint* cases = _deoptimization_hist[reason][1+action]; 2386 juint* bc_counter_addr = NULL; 2387 juint bc_counter = 0; 2388 // Look for an unused counter, or an exact match to this BC. 2389 if (bc != Bytecodes::_illegal) { 2390 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2391 juint* counter_addr = &cases[bc_case]; 2392 juint counter = *counter_addr; 2393 if ((counter == 0 && bc_counter_addr == NULL) 2394 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2395 // this counter is either free or is already devoted to this BC 2396 bc_counter_addr = counter_addr; 2397 bc_counter = counter | bc; 2398 } 2399 } 2400 } 2401 if (bc_counter_addr == NULL) { 2402 // Overflow, or no given bytecode. 2403 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2404 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2405 } 2406 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2407 } 2408 2409 jint Deoptimization::total_deoptimization_count() { 2410 return _deoptimization_hist[Reason_none][0][0]; 2411 } 2412 2413 jint Deoptimization::deoptimization_count(DeoptReason reason) { 2414 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2415 return _deoptimization_hist[reason][0][0]; 2416 } 2417 2418 void Deoptimization::print_statistics() { 2419 juint total = total_deoptimization_count(); 2420 juint account = total; 2421 if (total != 0) { 2422 ttyLocker ttyl; 2423 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2424 tty->print_cr("Deoptimization traps recorded:"); 2425 #define PRINT_STAT_LINE(name, r) \ 2426 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2427 PRINT_STAT_LINE("total", total); 2428 // For each non-zero entry in the histogram, print the reason, 2429 // the action, and (if specifically known) the type of bytecode. 2430 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2431 for (int action = 0; action < Action_LIMIT; action++) { 2432 juint* cases = _deoptimization_hist[reason][1+action]; 2433 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2434 juint counter = cases[bc_case]; 2435 if (counter != 0) { 2436 char name[1*K]; 2437 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2438 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2439 bc = Bytecodes::_illegal; 2440 sprintf(name, "%s/%s/%s", 2441 trap_reason_name(reason), 2442 trap_action_name(action), 2443 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2444 juint r = counter >> LSB_BITS; 2445 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2446 account -= r; 2447 } 2448 } 2449 } 2450 } 2451 if (account != 0) { 2452 PRINT_STAT_LINE("unaccounted", account); 2453 } 2454 #undef PRINT_STAT_LINE 2455 if (xtty != NULL) xtty->tail("statistics"); 2456 } 2457 } 2458 #else // COMPILER2 || SHARK || INCLUDE_JVMCI 2459 2460 2461 // Stubs for C1 only system. 2462 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2463 return false; 2464 } 2465 2466 const char* Deoptimization::trap_reason_name(int reason) { 2467 return "unknown"; 2468 } 2469 2470 void Deoptimization::print_statistics() { 2471 // no output 2472 } 2473 2474 void 2475 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2476 // no udpate 2477 } 2478 2479 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2480 return 0; 2481 } 2482 2483 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2484 Bytecodes::Code bc) { 2485 // no update 2486 } 2487 2488 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2489 int trap_state) { 2490 jio_snprintf(buf, buflen, "#%d", trap_state); 2491 return buf; 2492 } 2493 2494 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI