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 // Resolve klass of flattened value type field 1031 SignatureStream ss(fs.signature(), false); 1032 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD); 1033 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending"); 1034 assert(vk->is_value(), "must be a ValueKlass"); 1035 field._klass = InstanceKlass::cast(vk); 1036 } 1037 fields->append(field); 1038 } 1039 } 1040 fields->sort(compare); 1041 for (int i = 0; i < fields->length(); i++) { 1042 intptr_t val; 1043 ScopeValue* scope_field = sv->field_at(svIndex); 1044 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1045 int offset = base_offset + fields->at(i)._offset; 1046 BasicType type = fields->at(i)._type; 1047 switch (type) { 1048 case T_OBJECT: case T_ARRAY: 1049 assert(value->type() == T_OBJECT, "Agreement."); 1050 obj->obj_field_put(offset, value->get_obj()()); 1051 break; 1052 1053 case T_VALUETYPE: { 1054 // Recursively re-assign flattened value type fields 1055 InstanceKlass* vk = fields->at(i)._klass; 1056 assert(vk != NULL, "must be resolved"); 1057 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1058 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1059 continue; // Continue because we don't need to increment svIndex 1060 } 1061 1062 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1063 case T_INT: case T_FLOAT: { // 4 bytes. 1064 assert(value->type() == T_INT, "Agreement."); 1065 bool big_value = false; 1066 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1067 if (scope_field->is_location()) { 1068 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1069 if (type == Location::dbl || type == Location::lng) { 1070 big_value = true; 1071 } 1072 } 1073 if (scope_field->is_constant_int()) { 1074 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1075 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1076 big_value = true; 1077 } 1078 } 1079 } 1080 1081 if (big_value) { 1082 i++; 1083 assert(i < fields->length(), "second T_INT field needed"); 1084 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1085 } else { 1086 val = value->get_int(); 1087 obj->int_field_put(offset, (jint)*((jint*)&val)); 1088 break; 1089 } 1090 } 1091 /* no break */ 1092 1093 case T_LONG: case T_DOUBLE: { 1094 assert(value->type() == T_INT, "Agreement."); 1095 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1096 #ifdef _LP64 1097 jlong res = (jlong)low->get_int(); 1098 #else 1099 #ifdef SPARC 1100 // For SPARC we have to swap high and low words. 1101 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1102 #else 1103 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1104 #endif //SPARC 1105 #endif 1106 obj->long_field_put(offset, res); 1107 break; 1108 } 1109 1110 case T_SHORT: 1111 assert(value->type() == T_INT, "Agreement."); 1112 val = value->get_int(); 1113 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1114 break; 1115 1116 case T_CHAR: 1117 assert(value->type() == T_INT, "Agreement."); 1118 val = value->get_int(); 1119 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1120 break; 1121 1122 case T_BYTE: 1123 assert(value->type() == T_INT, "Agreement."); 1124 val = value->get_int(); 1125 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1126 break; 1127 1128 case T_BOOLEAN: 1129 assert(value->type() == T_INT, "Agreement."); 1130 val = value->get_int(); 1131 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1132 break; 1133 1134 default: 1135 ShouldNotReachHere(); 1136 } 1137 svIndex++; 1138 } 1139 return svIndex; 1140 } 1141 1142 // restore fields of an eliminated value type array 1143 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) { 1144 ValueKlass* vk = vak->element_klass(); 1145 assert(vk->flatten_array(), "should only be used for flattened value type arrays"); 1146 // Adjust offset to omit oop header 1147 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset(); 1148 // Initialize all elements of the flattened value type array 1149 for (int i = 0; i < sv->field_size(); i++) { 1150 ScopeValue* val = sv->field_at(i); 1151 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1152 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK); 1153 } 1154 } 1155 1156 // restore fields of all eliminated objects and arrays 1157 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1158 for (int i = 0; i < objects->length(); i++) { 1159 ObjectValue* sv = (ObjectValue*) objects->at(i); 1160 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1161 Handle obj = sv->value(); 1162 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1163 if (PrintDeoptimizationDetails) { 1164 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1165 } 1166 if (obj.is_null()) { 1167 continue; 1168 } 1169 1170 if (k->is_instance_klass()) { 1171 InstanceKlass* ik = InstanceKlass::cast(k); 1172 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1173 } else if (k->is_valueArray_klass()) { 1174 ValueArrayKlass* vak = ValueArrayKlass::cast(k); 1175 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK); 1176 } else if (k->is_typeArray_klass()) { 1177 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1178 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1179 } else if (k->is_objArray_klass()) { 1180 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1181 } 1182 } 1183 } 1184 1185 1186 // relock objects for which synchronization was eliminated 1187 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1188 for (int i = 0; i < monitors->length(); i++) { 1189 MonitorInfo* mon_info = monitors->at(i); 1190 if (mon_info->eliminated()) { 1191 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1192 if (!mon_info->owner_is_scalar_replaced()) { 1193 Handle obj(thread, mon_info->owner()); 1194 markOop mark = obj->mark(); 1195 if (UseBiasedLocking && mark->has_bias_pattern()) { 1196 // New allocated objects may have the mark set to anonymously biased. 1197 // Also the deoptimized method may called methods with synchronization 1198 // where the thread-local object is bias locked to the current thread. 1199 assert(mark->is_biased_anonymously() || 1200 mark->biased_locker() == thread, "should be locked to current thread"); 1201 // Reset mark word to unbiased prototype. 1202 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1203 obj->set_mark(unbiased_prototype); 1204 } 1205 BasicLock* lock = mon_info->lock(); 1206 ObjectSynchronizer::slow_enter(obj, lock, thread); 1207 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1208 } 1209 } 1210 } 1211 } 1212 1213 1214 #ifndef PRODUCT 1215 // print information about reallocated objects 1216 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1217 fieldDescriptor fd; 1218 1219 for (int i = 0; i < objects->length(); i++) { 1220 ObjectValue* sv = (ObjectValue*) objects->at(i); 1221 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1222 Handle obj = sv->value(); 1223 1224 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1225 k->print_value(); 1226 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1227 if (obj.is_null()) { 1228 tty->print(" allocation failed"); 1229 } else { 1230 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1231 } 1232 tty->cr(); 1233 1234 if (Verbose && !obj.is_null()) { 1235 k->oop_print_on(obj(), tty); 1236 } 1237 } 1238 } 1239 #endif 1240 #endif // COMPILER2 || INCLUDE_JVMCI 1241 1242 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1243 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1244 1245 #ifndef PRODUCT 1246 if (PrintDeoptimizationDetails) { 1247 ttyLocker ttyl; 1248 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1249 fr.print_on(tty); 1250 tty->print_cr(" Virtual frames (innermost first):"); 1251 for (int index = 0; index < chunk->length(); index++) { 1252 compiledVFrame* vf = chunk->at(index); 1253 tty->print(" %2d - ", index); 1254 vf->print_value(); 1255 int bci = chunk->at(index)->raw_bci(); 1256 const char* code_name; 1257 if (bci == SynchronizationEntryBCI) { 1258 code_name = "sync entry"; 1259 } else { 1260 Bytecodes::Code code = vf->method()->code_at(bci); 1261 code_name = Bytecodes::name(code); 1262 } 1263 tty->print(" - %s", code_name); 1264 tty->print_cr(" @ bci %d ", bci); 1265 if (Verbose) { 1266 vf->print(); 1267 tty->cr(); 1268 } 1269 } 1270 } 1271 #endif 1272 1273 // Register map for next frame (used for stack crawl). We capture 1274 // the state of the deopt'ing frame's caller. Thus if we need to 1275 // stuff a C2I adapter we can properly fill in the callee-save 1276 // register locations. 1277 frame caller = fr.sender(reg_map); 1278 int frame_size = caller.sp() - fr.sp(); 1279 1280 frame sender = caller; 1281 1282 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1283 // the vframeArray containing the unpacking information is allocated in the C heap. 1284 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1285 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1286 1287 // Compare the vframeArray to the collected vframes 1288 assert(array->structural_compare(thread, chunk), "just checking"); 1289 1290 #ifndef PRODUCT 1291 if (PrintDeoptimizationDetails) { 1292 ttyLocker ttyl; 1293 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1294 } 1295 #endif // PRODUCT 1296 1297 return array; 1298 } 1299 1300 #if defined(COMPILER2) || INCLUDE_JVMCI 1301 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1302 // Reallocation of some scalar replaced objects failed. Record 1303 // that we need to pop all the interpreter frames for the 1304 // deoptimized compiled frame. 1305 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1306 thread->set_frames_to_pop_failed_realloc(array->frames()); 1307 // Unlock all monitors here otherwise the interpreter will see a 1308 // mix of locked and unlocked monitors (because of failed 1309 // reallocations of synchronized objects) and be confused. 1310 for (int i = 0; i < array->frames(); i++) { 1311 MonitorChunk* monitors = array->element(i)->monitors(); 1312 if (monitors != NULL) { 1313 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1314 BasicObjectLock* src = monitors->at(j); 1315 if (src->obj() != NULL) { 1316 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1317 } 1318 } 1319 array->element(i)->free_monitors(thread); 1320 #ifdef ASSERT 1321 array->element(i)->set_removed_monitors(); 1322 #endif 1323 } 1324 } 1325 } 1326 #endif 1327 1328 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1329 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1330 Thread* thread = Thread::current(); 1331 for (int i = 0; i < monitors->length(); i++) { 1332 MonitorInfo* mon_info = monitors->at(i); 1333 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1334 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1335 } 1336 } 1337 } 1338 1339 1340 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1341 if (!UseBiasedLocking) { 1342 return; 1343 } 1344 1345 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1346 1347 // Unfortunately we don't have a RegisterMap available in most of 1348 // the places we want to call this routine so we need to walk the 1349 // stack again to update the register map. 1350 if (map == NULL || !map->update_map()) { 1351 StackFrameStream sfs(thread, true); 1352 bool found = false; 1353 while (!found && !sfs.is_done()) { 1354 frame* cur = sfs.current(); 1355 sfs.next(); 1356 found = cur->id() == fr.id(); 1357 } 1358 assert(found, "frame to be deoptimized not found on target thread's stack"); 1359 map = sfs.register_map(); 1360 } 1361 1362 vframe* vf = vframe::new_vframe(&fr, map, thread); 1363 compiledVFrame* cvf = compiledVFrame::cast(vf); 1364 // Revoke monitors' biases in all scopes 1365 while (!cvf->is_top()) { 1366 collect_monitors(cvf, objects_to_revoke); 1367 cvf = compiledVFrame::cast(cvf->sender()); 1368 } 1369 collect_monitors(cvf, objects_to_revoke); 1370 1371 if (SafepointSynchronize::is_at_safepoint()) { 1372 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1373 } else { 1374 BiasedLocking::revoke(objects_to_revoke); 1375 } 1376 } 1377 1378 1379 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1380 if (!UseBiasedLocking) { 1381 return; 1382 } 1383 1384 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1385 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1386 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1387 if (jt->has_last_Java_frame()) { 1388 StackFrameStream sfs(jt, true); 1389 while (!sfs.is_done()) { 1390 frame* cur = sfs.current(); 1391 if (cb->contains(cur->pc())) { 1392 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1393 compiledVFrame* cvf = compiledVFrame::cast(vf); 1394 // Revoke monitors' biases in all scopes 1395 while (!cvf->is_top()) { 1396 collect_monitors(cvf, objects_to_revoke); 1397 cvf = compiledVFrame::cast(cvf->sender()); 1398 } 1399 collect_monitors(cvf, objects_to_revoke); 1400 } 1401 sfs.next(); 1402 } 1403 } 1404 } 1405 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1406 } 1407 1408 1409 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1410 assert(fr.can_be_deoptimized(), "checking frame type"); 1411 1412 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1413 1414 if (LogCompilation && xtty != NULL) { 1415 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1416 assert(cm != NULL, "only compiled methods can deopt"); 1417 1418 ttyLocker ttyl; 1419 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1420 cm->log_identity(xtty); 1421 xtty->end_head(); 1422 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1423 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1424 xtty->method(sd->method()); 1425 xtty->end_elem(); 1426 if (sd->is_top()) break; 1427 } 1428 xtty->tail("deoptimized"); 1429 } 1430 1431 // Patch the compiled method so that when execution returns to it we will 1432 // deopt the execution state and return to the interpreter. 1433 fr.deoptimize(thread); 1434 } 1435 1436 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1437 deoptimize(thread, fr, map, Reason_constraint); 1438 } 1439 1440 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1441 // Deoptimize only if the frame comes from compile code. 1442 // Do not deoptimize the frame which is already patched 1443 // during the execution of the loops below. 1444 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1445 return; 1446 } 1447 ResourceMark rm; 1448 DeoptimizationMarker dm; 1449 if (UseBiasedLocking) { 1450 revoke_biases_of_monitors(thread, fr, map); 1451 } 1452 deoptimize_single_frame(thread, fr, reason); 1453 1454 } 1455 1456 1457 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1458 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1459 "can only deoptimize other thread at a safepoint"); 1460 // Compute frame and register map based on thread and sp. 1461 RegisterMap reg_map(thread, UseBiasedLocking); 1462 frame fr = thread->last_frame(); 1463 while (fr.id() != id) { 1464 fr = fr.sender(®_map); 1465 } 1466 deoptimize(thread, fr, ®_map, reason); 1467 } 1468 1469 1470 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1471 if (thread == Thread::current()) { 1472 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1473 } else { 1474 VM_DeoptimizeFrame deopt(thread, id, reason); 1475 VMThread::execute(&deopt); 1476 } 1477 } 1478 1479 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1480 deoptimize_frame(thread, id, Reason_constraint); 1481 } 1482 1483 // JVMTI PopFrame support 1484 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1485 { 1486 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1487 } 1488 JRT_END 1489 1490 MethodData* 1491 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1492 bool create_if_missing) { 1493 Thread* THREAD = thread; 1494 MethodData* mdo = m()->method_data(); 1495 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1496 // Build an MDO. Ignore errors like OutOfMemory; 1497 // that simply means we won't have an MDO to update. 1498 Method::build_interpreter_method_data(m, THREAD); 1499 if (HAS_PENDING_EXCEPTION) { 1500 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1501 CLEAR_PENDING_EXCEPTION; 1502 } 1503 mdo = m()->method_data(); 1504 } 1505 return mdo; 1506 } 1507 1508 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI 1509 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1510 // in case of an unresolved klass entry, load the class. 1511 if (constant_pool->tag_at(index).is_unresolved_klass() || 1512 constant_pool->tag_at(index).is_unresolved_value_type()) { 1513 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1514 return; 1515 } 1516 1517 if (!constant_pool->tag_at(index).is_symbol()) return; 1518 1519 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1520 Symbol* symbol = constant_pool->symbol_at(index); 1521 1522 // class name? 1523 if (symbol->byte_at(0) != '(') { 1524 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1525 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1526 return; 1527 } 1528 1529 // then it must be a signature! 1530 ResourceMark rm(THREAD); 1531 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1532 if (ss.is_object()) { 1533 Symbol* class_name = ss.as_symbol(CHECK); 1534 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1535 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1536 } 1537 } 1538 } 1539 1540 1541 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1542 EXCEPTION_MARK; 1543 load_class_by_index(constant_pool, index, THREAD); 1544 if (HAS_PENDING_EXCEPTION) { 1545 // Exception happened during classloading. We ignore the exception here, since it 1546 // is going to be rethrown since the current activation is going to be deoptimized and 1547 // the interpreter will re-execute the bytecode. 1548 CLEAR_PENDING_EXCEPTION; 1549 // Class loading called java code which may have caused a stack 1550 // overflow. If the exception was thrown right before the return 1551 // to the runtime the stack is no longer guarded. Reguard the 1552 // stack otherwise if we return to the uncommon trap blob and the 1553 // stack bang causes a stack overflow we crash. 1554 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1555 JavaThread* thread = (JavaThread*)THREAD; 1556 bool guard_pages_enabled = thread->stack_guards_enabled(); 1557 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1558 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1559 } 1560 } 1561 1562 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1563 HandleMark hm; 1564 1565 // uncommon_trap() is called at the beginning of the uncommon trap 1566 // handler. Note this fact before we start generating temporary frames 1567 // that can confuse an asynchronous stack walker. This counter is 1568 // decremented at the end of unpack_frames(). 1569 thread->inc_in_deopt_handler(); 1570 1571 // We need to update the map if we have biased locking. 1572 #if INCLUDE_JVMCI 1573 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1574 RegisterMap reg_map(thread, true); 1575 #else 1576 RegisterMap reg_map(thread, UseBiasedLocking); 1577 #endif 1578 frame stub_frame = thread->last_frame(); 1579 frame fr = stub_frame.sender(®_map); 1580 // Make sure the calling nmethod is not getting deoptimized and removed 1581 // before we are done with it. 1582 nmethodLocker nl(fr.pc()); 1583 1584 // Log a message 1585 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1586 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1587 1588 { 1589 ResourceMark rm; 1590 1591 // Revoke biases of any monitors in the frame to ensure we can migrate them 1592 revoke_biases_of_monitors(thread, fr, ®_map); 1593 1594 DeoptReason reason = trap_request_reason(trap_request); 1595 DeoptAction action = trap_request_action(trap_request); 1596 #if INCLUDE_JVMCI 1597 int debug_id = trap_request_debug_id(trap_request); 1598 #endif 1599 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1600 1601 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1602 compiledVFrame* cvf = compiledVFrame::cast(vf); 1603 1604 CompiledMethod* nm = cvf->code(); 1605 1606 ScopeDesc* trap_scope = cvf->scope(); 1607 1608 if (TraceDeoptimization) { 1609 ttyLocker ttyl; 1610 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() 1611 #if INCLUDE_JVMCI 1612 , debug_id 1613 #endif 1614 ); 1615 } 1616 1617 methodHandle trap_method = trap_scope->method(); 1618 int trap_bci = trap_scope->bci(); 1619 #if INCLUDE_JVMCI 1620 oop speculation = thread->pending_failed_speculation(); 1621 if (nm->is_compiled_by_jvmci()) { 1622 if (speculation != NULL) { 1623 oop speculation_log = nm->as_nmethod()->speculation_log(); 1624 if (speculation_log != NULL) { 1625 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1626 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) { 1627 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1628 } 1629 } 1630 if (TraceDeoptimization) { 1631 tty->print_cr("Saving speculation to speculation log"); 1632 } 1633 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1634 } else { 1635 if (TraceDeoptimization) { 1636 tty->print_cr("Speculation present but no speculation log"); 1637 } 1638 } 1639 thread->set_pending_failed_speculation(NULL); 1640 } else { 1641 if (TraceDeoptimization) { 1642 tty->print_cr("No speculation"); 1643 } 1644 } 1645 } else { 1646 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1647 } 1648 1649 if (trap_bci == SynchronizationEntryBCI) { 1650 trap_bci = 0; 1651 thread->set_pending_monitorenter(true); 1652 } 1653 1654 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1655 thread->set_pending_transfer_to_interpreter(true); 1656 } 1657 #endif 1658 1659 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1660 // Record this event in the histogram. 1661 gather_statistics(reason, action, trap_bc); 1662 1663 // Ensure that we can record deopt. history: 1664 // Need MDO to record RTM code generation state. 1665 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1666 1667 methodHandle profiled_method; 1668 #if INCLUDE_JVMCI 1669 if (nm->is_compiled_by_jvmci()) { 1670 profiled_method = nm->method(); 1671 } else { 1672 profiled_method = trap_method; 1673 } 1674 #else 1675 profiled_method = trap_method; 1676 #endif 1677 1678 MethodData* trap_mdo = 1679 get_method_data(thread, profiled_method, create_if_missing); 1680 1681 // Log a message 1682 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1683 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1684 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1685 1686 // Print a bunch of diagnostics, if requested. 1687 if (TraceDeoptimization || LogCompilation) { 1688 ResourceMark rm; 1689 ttyLocker ttyl; 1690 char buf[100]; 1691 if (xtty != NULL) { 1692 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1693 os::current_thread_id(), 1694 format_trap_request(buf, sizeof(buf), trap_request)); 1695 nm->log_identity(xtty); 1696 } 1697 Symbol* class_name = NULL; 1698 bool unresolved = false; 1699 if (unloaded_class_index >= 0) { 1700 constantPoolHandle constants (THREAD, trap_method->constants()); 1701 if (constants->tag_at(unloaded_class_index).is_unresolved_klass() || 1702 constants->tag_at(unloaded_class_index).is_unresolved_value_type()) { 1703 class_name = constants->klass_name_at(unloaded_class_index); 1704 unresolved = true; 1705 if (xtty != NULL) 1706 xtty->print(" unresolved='1'"); 1707 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1708 class_name = constants->symbol_at(unloaded_class_index); 1709 } 1710 if (xtty != NULL) 1711 xtty->name(class_name); 1712 } 1713 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1714 // Dump the relevant MDO state. 1715 // This is the deopt count for the current reason, any previous 1716 // reasons or recompiles seen at this point. 1717 int dcnt = trap_mdo->trap_count(reason); 1718 if (dcnt != 0) 1719 xtty->print(" count='%d'", dcnt); 1720 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1721 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1722 if (dos != 0) { 1723 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1724 if (trap_state_is_recompiled(dos)) { 1725 int recnt2 = trap_mdo->overflow_recompile_count(); 1726 if (recnt2 != 0) 1727 xtty->print(" recompiles2='%d'", recnt2); 1728 } 1729 } 1730 } 1731 if (xtty != NULL) { 1732 xtty->stamp(); 1733 xtty->end_head(); 1734 } 1735 if (TraceDeoptimization) { // make noise on the tty 1736 tty->print("Uncommon trap occurred in"); 1737 nm->method()->print_short_name(tty); 1738 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1739 #if INCLUDE_JVMCI 1740 if (nm->is_nmethod()) { 1741 oop installedCode = nm->as_nmethod()->jvmci_installed_code(); 1742 if (installedCode != NULL) { 1743 oop installedCodeName = NULL; 1744 if (installedCode->is_a(InstalledCode::klass())) { 1745 installedCodeName = InstalledCode::name(installedCode); 1746 } 1747 if (installedCodeName != NULL) { 1748 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName)); 1749 } else { 1750 tty->print(" (JVMCI: installed code has no name) "); 1751 } 1752 } else if (nm->is_compiled_by_jvmci()) { 1753 tty->print(" (JVMCI: no installed code) "); 1754 } 1755 } 1756 #endif 1757 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1758 p2i(fr.pc()), 1759 os::current_thread_id(), 1760 trap_reason_name(reason), 1761 trap_action_name(action), 1762 unloaded_class_index 1763 #if INCLUDE_JVMCI 1764 , debug_id 1765 #endif 1766 ); 1767 if (class_name != NULL) { 1768 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1769 class_name->print_symbol_on(tty); 1770 } 1771 tty->cr(); 1772 } 1773 if (xtty != NULL) { 1774 // Log the precise location of the trap. 1775 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1776 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1777 xtty->method(sd->method()); 1778 xtty->end_elem(); 1779 if (sd->is_top()) break; 1780 } 1781 xtty->tail("uncommon_trap"); 1782 } 1783 } 1784 // (End diagnostic printout.) 1785 1786 // Load class if necessary 1787 if (unloaded_class_index >= 0) { 1788 constantPoolHandle constants(THREAD, trap_method->constants()); 1789 load_class_by_index(constants, unloaded_class_index); 1790 } 1791 1792 // Flush the nmethod if necessary and desirable. 1793 // 1794 // We need to avoid situations where we are re-flushing the nmethod 1795 // because of a hot deoptimization site. Repeated flushes at the same 1796 // point need to be detected by the compiler and avoided. If the compiler 1797 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1798 // module must take measures to avoid an infinite cycle of recompilation 1799 // and deoptimization. There are several such measures: 1800 // 1801 // 1. If a recompilation is ordered a second time at some site X 1802 // and for the same reason R, the action is adjusted to 'reinterpret', 1803 // to give the interpreter time to exercise the method more thoroughly. 1804 // If this happens, the method's overflow_recompile_count is incremented. 1805 // 1806 // 2. If the compiler fails to reduce the deoptimization rate, then 1807 // the method's overflow_recompile_count will begin to exceed the set 1808 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1809 // is adjusted to 'make_not_compilable', and the method is abandoned 1810 // to the interpreter. This is a performance hit for hot methods, 1811 // but is better than a disastrous infinite cycle of recompilations. 1812 // (Actually, only the method containing the site X is abandoned.) 1813 // 1814 // 3. In parallel with the previous measures, if the total number of 1815 // recompilations of a method exceeds the much larger set limit 1816 // PerMethodRecompilationCutoff, the method is abandoned. 1817 // This should only happen if the method is very large and has 1818 // many "lukewarm" deoptimizations. The code which enforces this 1819 // limit is elsewhere (class nmethod, class Method). 1820 // 1821 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1822 // to recompile at each bytecode independently of the per-BCI cutoff. 1823 // 1824 // The decision to update code is up to the compiler, and is encoded 1825 // in the Action_xxx code. If the compiler requests Action_none 1826 // no trap state is changed, no compiled code is changed, and the 1827 // computation suffers along in the interpreter. 1828 // 1829 // The other action codes specify various tactics for decompilation 1830 // and recompilation. Action_maybe_recompile is the loosest, and 1831 // allows the compiled code to stay around until enough traps are seen, 1832 // and until the compiler gets around to recompiling the trapping method. 1833 // 1834 // The other actions cause immediate removal of the present code. 1835 1836 // Traps caused by injected profile shouldn't pollute trap counts. 1837 bool injected_profile_trap = trap_method->has_injected_profile() && 1838 (reason == Reason_intrinsic || reason == Reason_unreached); 1839 1840 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1841 bool make_not_entrant = false; 1842 bool make_not_compilable = false; 1843 bool reprofile = false; 1844 switch (action) { 1845 case Action_none: 1846 // Keep the old code. 1847 update_trap_state = false; 1848 break; 1849 case Action_maybe_recompile: 1850 // Do not need to invalidate the present code, but we can 1851 // initiate another 1852 // Start compiler without (necessarily) invalidating the nmethod. 1853 // The system will tolerate the old code, but new code should be 1854 // generated when possible. 1855 break; 1856 case Action_reinterpret: 1857 // Go back into the interpreter for a while, and then consider 1858 // recompiling form scratch. 1859 make_not_entrant = true; 1860 // Reset invocation counter for outer most method. 1861 // This will allow the interpreter to exercise the bytecodes 1862 // for a while before recompiling. 1863 // By contrast, Action_make_not_entrant is immediate. 1864 // 1865 // Note that the compiler will track null_check, null_assert, 1866 // range_check, and class_check events and log them as if they 1867 // had been traps taken from compiled code. This will update 1868 // the MDO trap history so that the next compilation will 1869 // properly detect hot trap sites. 1870 reprofile = true; 1871 break; 1872 case Action_make_not_entrant: 1873 // Request immediate recompilation, and get rid of the old code. 1874 // Make them not entrant, so next time they are called they get 1875 // recompiled. Unloaded classes are loaded now so recompile before next 1876 // time they are called. Same for uninitialized. The interpreter will 1877 // link the missing class, if any. 1878 make_not_entrant = true; 1879 break; 1880 case Action_make_not_compilable: 1881 // Give up on compiling this method at all. 1882 make_not_entrant = true; 1883 make_not_compilable = true; 1884 break; 1885 default: 1886 ShouldNotReachHere(); 1887 } 1888 1889 // Setting +ProfileTraps fixes the following, on all platforms: 1890 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1891 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1892 // recompile relies on a MethodData* to record heroic opt failures. 1893 1894 // Whether the interpreter is producing MDO data or not, we also need 1895 // to use the MDO to detect hot deoptimization points and control 1896 // aggressive optimization. 1897 bool inc_recompile_count = false; 1898 ProfileData* pdata = NULL; 1899 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1900 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1901 uint this_trap_count = 0; 1902 bool maybe_prior_trap = false; 1903 bool maybe_prior_recompile = false; 1904 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1905 #if INCLUDE_JVMCI 1906 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1907 #endif 1908 nm->method(), 1909 //outputs: 1910 this_trap_count, 1911 maybe_prior_trap, 1912 maybe_prior_recompile); 1913 // Because the interpreter also counts null, div0, range, and class 1914 // checks, these traps from compiled code are double-counted. 1915 // This is harmless; it just means that the PerXTrapLimit values 1916 // are in effect a little smaller than they look. 1917 1918 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1919 if (per_bc_reason != Reason_none) { 1920 // Now take action based on the partially known per-BCI history. 1921 if (maybe_prior_trap 1922 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1923 // If there are too many traps at this BCI, force a recompile. 1924 // This will allow the compiler to see the limit overflow, and 1925 // take corrective action, if possible. The compiler generally 1926 // does not use the exact PerBytecodeTrapLimit value, but instead 1927 // changes its tactics if it sees any traps at all. This provides 1928 // a little hysteresis, delaying a recompile until a trap happens 1929 // several times. 1930 // 1931 // Actually, since there is only one bit of counter per BCI, 1932 // the possible per-BCI counts are {0,1,(per-method count)}. 1933 // This produces accurate results if in fact there is only 1934 // one hot trap site, but begins to get fuzzy if there are 1935 // many sites. For example, if there are ten sites each 1936 // trapping two or more times, they each get the blame for 1937 // all of their traps. 1938 make_not_entrant = true; 1939 } 1940 1941 // Detect repeated recompilation at the same BCI, and enforce a limit. 1942 if (make_not_entrant && maybe_prior_recompile) { 1943 // More than one recompile at this point. 1944 inc_recompile_count = maybe_prior_trap; 1945 } 1946 } else { 1947 // For reasons which are not recorded per-bytecode, we simply 1948 // force recompiles unconditionally. 1949 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1950 make_not_entrant = true; 1951 } 1952 1953 // Go back to the compiler if there are too many traps in this method. 1954 if (this_trap_count >= per_method_trap_limit(reason)) { 1955 // If there are too many traps in this method, force a recompile. 1956 // This will allow the compiler to see the limit overflow, and 1957 // take corrective action, if possible. 1958 // (This condition is an unlikely backstop only, because the 1959 // PerBytecodeTrapLimit is more likely to take effect first, 1960 // if it is applicable.) 1961 make_not_entrant = true; 1962 } 1963 1964 // Here's more hysteresis: If there has been a recompile at 1965 // this trap point already, run the method in the interpreter 1966 // for a while to exercise it more thoroughly. 1967 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1968 reprofile = true; 1969 } 1970 } 1971 1972 // Take requested actions on the method: 1973 1974 // Recompile 1975 if (make_not_entrant) { 1976 if (!nm->make_not_entrant()) { 1977 return; // the call did not change nmethod's state 1978 } 1979 1980 if (pdata != NULL) { 1981 // Record the recompilation event, if any. 1982 int tstate0 = pdata->trap_state(); 1983 int tstate1 = trap_state_set_recompiled(tstate0, true); 1984 if (tstate1 != tstate0) 1985 pdata->set_trap_state(tstate1); 1986 } 1987 1988 #if INCLUDE_RTM_OPT 1989 // Restart collecting RTM locking abort statistic if the method 1990 // is recompiled for a reason other than RTM state change. 1991 // Assume that in new recompiled code the statistic could be different, 1992 // for example, due to different inlining. 1993 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1994 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 1995 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1996 } 1997 #endif 1998 // For code aging we count traps separately here, using make_not_entrant() 1999 // as a guard against simultaneous deopts in multiple threads. 2000 if (reason == Reason_tenured && trap_mdo != NULL) { 2001 trap_mdo->inc_tenure_traps(); 2002 } 2003 } 2004 2005 if (inc_recompile_count) { 2006 trap_mdo->inc_overflow_recompile_count(); 2007 if ((uint)trap_mdo->overflow_recompile_count() > 2008 (uint)PerBytecodeRecompilationCutoff) { 2009 // Give up on the method containing the bad BCI. 2010 if (trap_method() == nm->method()) { 2011 make_not_compilable = true; 2012 } else { 2013 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 2014 // But give grace to the enclosing nm->method(). 2015 } 2016 } 2017 } 2018 2019 // Reprofile 2020 if (reprofile) { 2021 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2022 } 2023 2024 // Give up compiling 2025 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2026 assert(make_not_entrant, "consistent"); 2027 nm->method()->set_not_compilable(CompLevel_full_optimization); 2028 } 2029 2030 } // Free marked resources 2031 2032 } 2033 JRT_END 2034 2035 ProfileData* 2036 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2037 int trap_bci, 2038 Deoptimization::DeoptReason reason, 2039 bool update_total_trap_count, 2040 #if INCLUDE_JVMCI 2041 bool is_osr, 2042 #endif 2043 Method* compiled_method, 2044 //outputs: 2045 uint& ret_this_trap_count, 2046 bool& ret_maybe_prior_trap, 2047 bool& ret_maybe_prior_recompile) { 2048 bool maybe_prior_trap = false; 2049 bool maybe_prior_recompile = false; 2050 uint this_trap_count = 0; 2051 if (update_total_trap_count) { 2052 uint idx = reason; 2053 #if INCLUDE_JVMCI 2054 if (is_osr) { 2055 idx += Reason_LIMIT; 2056 } 2057 #endif 2058 uint prior_trap_count = trap_mdo->trap_count(idx); 2059 this_trap_count = trap_mdo->inc_trap_count(idx); 2060 2061 // If the runtime cannot find a place to store trap history, 2062 // it is estimated based on the general condition of the method. 2063 // If the method has ever been recompiled, or has ever incurred 2064 // a trap with the present reason , then this BCI is assumed 2065 // (pessimistically) to be the culprit. 2066 maybe_prior_trap = (prior_trap_count != 0); 2067 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2068 } 2069 ProfileData* pdata = NULL; 2070 2071 2072 // For reasons which are recorded per bytecode, we check per-BCI data. 2073 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2074 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2075 if (per_bc_reason != Reason_none) { 2076 // Find the profile data for this BCI. If there isn't one, 2077 // try to allocate one from the MDO's set of spares. 2078 // This will let us detect a repeated trap at this point. 2079 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2080 2081 if (pdata != NULL) { 2082 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2083 if (LogCompilation && xtty != NULL) { 2084 ttyLocker ttyl; 2085 // no more room for speculative traps in this MDO 2086 xtty->elem("speculative_traps_oom"); 2087 } 2088 } 2089 // Query the trap state of this profile datum. 2090 int tstate0 = pdata->trap_state(); 2091 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2092 maybe_prior_trap = false; 2093 if (!trap_state_is_recompiled(tstate0)) 2094 maybe_prior_recompile = false; 2095 2096 // Update the trap state of this profile datum. 2097 int tstate1 = tstate0; 2098 // Record the reason. 2099 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2100 // Store the updated state on the MDO, for next time. 2101 if (tstate1 != tstate0) 2102 pdata->set_trap_state(tstate1); 2103 } else { 2104 if (LogCompilation && xtty != NULL) { 2105 ttyLocker ttyl; 2106 // Missing MDP? Leave a small complaint in the log. 2107 xtty->elem("missing_mdp bci='%d'", trap_bci); 2108 } 2109 } 2110 } 2111 2112 // Return results: 2113 ret_this_trap_count = this_trap_count; 2114 ret_maybe_prior_trap = maybe_prior_trap; 2115 ret_maybe_prior_recompile = maybe_prior_recompile; 2116 return pdata; 2117 } 2118 2119 void 2120 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2121 ResourceMark rm; 2122 // Ignored outputs: 2123 uint ignore_this_trap_count; 2124 bool ignore_maybe_prior_trap; 2125 bool ignore_maybe_prior_recompile; 2126 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2127 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2128 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true); 2129 query_update_method_data(trap_mdo, trap_bci, 2130 (DeoptReason)reason, 2131 update_total_counts, 2132 #if INCLUDE_JVMCI 2133 false, 2134 #endif 2135 NULL, 2136 ignore_this_trap_count, 2137 ignore_maybe_prior_trap, 2138 ignore_maybe_prior_recompile); 2139 } 2140 2141 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2142 if (TraceDeoptimization) { 2143 tty->print("Uncommon trap "); 2144 } 2145 // Still in Java no safepoints 2146 { 2147 // This enters VM and may safepoint 2148 uncommon_trap_inner(thread, trap_request); 2149 } 2150 return fetch_unroll_info_helper(thread, exec_mode); 2151 } 2152 2153 // Local derived constants. 2154 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2155 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 2156 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2157 2158 //---------------------------trap_state_reason--------------------------------- 2159 Deoptimization::DeoptReason 2160 Deoptimization::trap_state_reason(int trap_state) { 2161 // This assert provides the link between the width of DataLayout::trap_bits 2162 // and the encoding of "recorded" reasons. It ensures there are enough 2163 // bits to store all needed reasons in the per-BCI MDO profile. 2164 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2165 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2166 trap_state -= recompile_bit; 2167 if (trap_state == DS_REASON_MASK) { 2168 return Reason_many; 2169 } else { 2170 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2171 return (DeoptReason)trap_state; 2172 } 2173 } 2174 //-------------------------trap_state_has_reason------------------------------- 2175 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2176 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2177 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2178 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2179 trap_state -= recompile_bit; 2180 if (trap_state == DS_REASON_MASK) { 2181 return -1; // true, unspecifically (bottom of state lattice) 2182 } else if (trap_state == reason) { 2183 return 1; // true, definitely 2184 } else if (trap_state == 0) { 2185 return 0; // false, definitely (top of state lattice) 2186 } else { 2187 return 0; // false, definitely 2188 } 2189 } 2190 //-------------------------trap_state_add_reason------------------------------- 2191 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2192 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2193 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2194 trap_state -= recompile_bit; 2195 if (trap_state == DS_REASON_MASK) { 2196 return trap_state + recompile_bit; // already at state lattice bottom 2197 } else if (trap_state == reason) { 2198 return trap_state + recompile_bit; // the condition is already true 2199 } else if (trap_state == 0) { 2200 return reason + recompile_bit; // no condition has yet been true 2201 } else { 2202 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2203 } 2204 } 2205 //-----------------------trap_state_is_recompiled------------------------------ 2206 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2207 return (trap_state & DS_RECOMPILE_BIT) != 0; 2208 } 2209 //-----------------------trap_state_set_recompiled----------------------------- 2210 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2211 if (z) return trap_state | DS_RECOMPILE_BIT; 2212 else return trap_state & ~DS_RECOMPILE_BIT; 2213 } 2214 //---------------------------format_trap_state--------------------------------- 2215 // This is used for debugging and diagnostics, including LogFile output. 2216 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2217 int trap_state) { 2218 assert(buflen > 0, "sanity"); 2219 DeoptReason reason = trap_state_reason(trap_state); 2220 bool recomp_flag = trap_state_is_recompiled(trap_state); 2221 // Re-encode the state from its decoded components. 2222 int decoded_state = 0; 2223 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2224 decoded_state = trap_state_add_reason(decoded_state, reason); 2225 if (recomp_flag) 2226 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2227 // If the state re-encodes properly, format it symbolically. 2228 // Because this routine is used for debugging and diagnostics, 2229 // be robust even if the state is a strange value. 2230 size_t len; 2231 if (decoded_state != trap_state) { 2232 // Random buggy state that doesn't decode?? 2233 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2234 } else { 2235 len = jio_snprintf(buf, buflen, "%s%s", 2236 trap_reason_name(reason), 2237 recomp_flag ? " recompiled" : ""); 2238 } 2239 return buf; 2240 } 2241 2242 2243 //--------------------------------statics-------------------------------------- 2244 const char* Deoptimization::_trap_reason_name[] = { 2245 // Note: Keep this in sync. with enum DeoptReason. 2246 "none", 2247 "null_check", 2248 "null_assert" JVMCI_ONLY("_or_unreached0"), 2249 "range_check", 2250 "class_check", 2251 "array_check", 2252 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2253 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2254 "unloaded", 2255 "uninitialized", 2256 "unreached", 2257 "unhandled", 2258 "constraint", 2259 "div0_check", 2260 "age", 2261 "predicate", 2262 "loop_limit_check", 2263 "speculate_class_check", 2264 "speculate_null_check", 2265 "speculate_null_assert", 2266 "rtm_state_change", 2267 "unstable_if", 2268 "unstable_fused_if", 2269 #if INCLUDE_JVMCI 2270 "aliasing", 2271 "transfer_to_interpreter", 2272 "not_compiled_exception_handler", 2273 "unresolved", 2274 "jsr_mismatch", 2275 #endif 2276 "tenured" 2277 }; 2278 const char* Deoptimization::_trap_action_name[] = { 2279 // Note: Keep this in sync. with enum DeoptAction. 2280 "none", 2281 "maybe_recompile", 2282 "reinterpret", 2283 "make_not_entrant", 2284 "make_not_compilable" 2285 }; 2286 2287 const char* Deoptimization::trap_reason_name(int reason) { 2288 // Check that every reason has a name 2289 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2290 2291 if (reason == Reason_many) return "many"; 2292 if ((uint)reason < Reason_LIMIT) 2293 return _trap_reason_name[reason]; 2294 static char buf[20]; 2295 sprintf(buf, "reason%d", reason); 2296 return buf; 2297 } 2298 const char* Deoptimization::trap_action_name(int action) { 2299 // Check that every action has a name 2300 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2301 2302 if ((uint)action < Action_LIMIT) 2303 return _trap_action_name[action]; 2304 static char buf[20]; 2305 sprintf(buf, "action%d", action); 2306 return buf; 2307 } 2308 2309 // This is used for debugging and diagnostics, including LogFile output. 2310 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2311 int trap_request) { 2312 jint unloaded_class_index = trap_request_index(trap_request); 2313 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2314 const char* action = trap_action_name(trap_request_action(trap_request)); 2315 #if INCLUDE_JVMCI 2316 int debug_id = trap_request_debug_id(trap_request); 2317 #endif 2318 size_t len; 2319 if (unloaded_class_index < 0) { 2320 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2321 reason, action 2322 #if INCLUDE_JVMCI 2323 ,debug_id 2324 #endif 2325 ); 2326 } else { 2327 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2328 reason, action, unloaded_class_index 2329 #if INCLUDE_JVMCI 2330 ,debug_id 2331 #endif 2332 ); 2333 } 2334 return buf; 2335 } 2336 2337 juint Deoptimization::_deoptimization_hist 2338 [Deoptimization::Reason_LIMIT] 2339 [1 + Deoptimization::Action_LIMIT] 2340 [Deoptimization::BC_CASE_LIMIT] 2341 = {0}; 2342 2343 enum { 2344 LSB_BITS = 8, 2345 LSB_MASK = right_n_bits(LSB_BITS) 2346 }; 2347 2348 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2349 Bytecodes::Code bc) { 2350 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2351 assert(action >= 0 && action < Action_LIMIT, "oob"); 2352 _deoptimization_hist[Reason_none][0][0] += 1; // total 2353 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2354 juint* cases = _deoptimization_hist[reason][1+action]; 2355 juint* bc_counter_addr = NULL; 2356 juint bc_counter = 0; 2357 // Look for an unused counter, or an exact match to this BC. 2358 if (bc != Bytecodes::_illegal) { 2359 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2360 juint* counter_addr = &cases[bc_case]; 2361 juint counter = *counter_addr; 2362 if ((counter == 0 && bc_counter_addr == NULL) 2363 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2364 // this counter is either free or is already devoted to this BC 2365 bc_counter_addr = counter_addr; 2366 bc_counter = counter | bc; 2367 } 2368 } 2369 } 2370 if (bc_counter_addr == NULL) { 2371 // Overflow, or no given bytecode. 2372 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2373 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2374 } 2375 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2376 } 2377 2378 jint Deoptimization::total_deoptimization_count() { 2379 return _deoptimization_hist[Reason_none][0][0]; 2380 } 2381 2382 jint Deoptimization::deoptimization_count(DeoptReason reason) { 2383 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2384 return _deoptimization_hist[reason][0][0]; 2385 } 2386 2387 void Deoptimization::print_statistics() { 2388 juint total = total_deoptimization_count(); 2389 juint account = total; 2390 if (total != 0) { 2391 ttyLocker ttyl; 2392 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2393 tty->print_cr("Deoptimization traps recorded:"); 2394 #define PRINT_STAT_LINE(name, r) \ 2395 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2396 PRINT_STAT_LINE("total", total); 2397 // For each non-zero entry in the histogram, print the reason, 2398 // the action, and (if specifically known) the type of bytecode. 2399 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2400 for (int action = 0; action < Action_LIMIT; action++) { 2401 juint* cases = _deoptimization_hist[reason][1+action]; 2402 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2403 juint counter = cases[bc_case]; 2404 if (counter != 0) { 2405 char name[1*K]; 2406 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2407 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2408 bc = Bytecodes::_illegal; 2409 sprintf(name, "%s/%s/%s", 2410 trap_reason_name(reason), 2411 trap_action_name(action), 2412 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2413 juint r = counter >> LSB_BITS; 2414 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2415 account -= r; 2416 } 2417 } 2418 } 2419 } 2420 if (account != 0) { 2421 PRINT_STAT_LINE("unaccounted", account); 2422 } 2423 #undef PRINT_STAT_LINE 2424 if (xtty != NULL) xtty->tail("statistics"); 2425 } 2426 } 2427 #else // COMPILER2 || SHARK || INCLUDE_JVMCI 2428 2429 2430 // Stubs for C1 only system. 2431 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2432 return false; 2433 } 2434 2435 const char* Deoptimization::trap_reason_name(int reason) { 2436 return "unknown"; 2437 } 2438 2439 void Deoptimization::print_statistics() { 2440 // no output 2441 } 2442 2443 void 2444 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2445 // no udpate 2446 } 2447 2448 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2449 return 0; 2450 } 2451 2452 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2453 Bytecodes::Code bc) { 2454 // no update 2455 } 2456 2457 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2458 int trap_state) { 2459 jio_snprintf(buf, buflen, "#%d", trap_state); 2460 return buf; 2461 } 2462 2463 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI