1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "jvm.h" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "code/nmethod.hpp" 31 #include "code/pcDesc.hpp" 32 #include "code/scopeDesc.hpp" 33 #include "interpreter/bytecode.hpp" 34 #include "interpreter/interpreter.hpp" 35 #include "interpreter/oopMapCache.hpp" 36 #include "memory/allocation.inline.hpp" 37 #include "memory/oopFactory.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "oops/method.hpp" 40 #include "oops/objArrayOop.inline.hpp" 41 #include "oops/oop.inline.hpp" 42 #include "oops/fieldStreams.hpp" 43 #include "oops/typeArrayOop.inline.hpp" 44 #include "oops/valueArrayKlass.hpp" 45 #include "oops/valueArrayOop.hpp" 46 #include "oops/valueKlass.hpp" 47 #include "oops/verifyOopClosure.hpp" 48 #include "prims/jvmtiThreadState.hpp" 49 #include "runtime/biasedLocking.hpp" 50 #include "runtime/compilationPolicy.hpp" 51 #include "runtime/deoptimization.hpp" 52 #include "runtime/frame.inline.hpp" 53 #include "runtime/interfaceSupport.inline.hpp" 54 #include "runtime/safepointVerifiers.hpp" 55 #include "runtime/sharedRuntime.hpp" 56 #include "runtime/signature.hpp" 57 #include "runtime/stubRoutines.hpp" 58 #include "runtime/thread.hpp" 59 #include "runtime/threadSMR.hpp" 60 #include "runtime/vframe.hpp" 61 #include "runtime/vframeArray.hpp" 62 #include "runtime/vframe_hp.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/preserveException.hpp" 65 #include "utilities/xmlstream.hpp" 66 67 #if INCLUDE_JVMCI 68 #include "jvmci/jvmciRuntime.hpp" 69 #include "jvmci/jvmciJavaClasses.hpp" 70 #endif 71 72 73 bool DeoptimizationMarker::_is_active = false; 74 75 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 76 int caller_adjustment, 77 int caller_actual_parameters, 78 int number_of_frames, 79 intptr_t* frame_sizes, 80 address* frame_pcs, 81 BasicType return_type, 82 int exec_mode) { 83 _size_of_deoptimized_frame = size_of_deoptimized_frame; 84 _caller_adjustment = caller_adjustment; 85 _caller_actual_parameters = caller_actual_parameters; 86 _number_of_frames = number_of_frames; 87 _frame_sizes = frame_sizes; 88 _frame_pcs = frame_pcs; 89 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 90 _return_type = return_type; 91 _initial_info = 0; 92 // PD (x86 only) 93 _counter_temp = 0; 94 _unpack_kind = exec_mode; 95 _sender_sp_temp = 0; 96 97 _total_frame_sizes = size_of_frames(); 98 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 99 } 100 101 102 Deoptimization::UnrollBlock::~UnrollBlock() { 103 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 104 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 105 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 106 } 107 108 109 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 110 assert(register_number < RegisterMap::reg_count, "checking register number"); 111 return &_register_block[register_number * 2]; 112 } 113 114 115 116 int Deoptimization::UnrollBlock::size_of_frames() const { 117 // Acount first for the adjustment of the initial frame 118 int result = _caller_adjustment; 119 for (int index = 0; index < number_of_frames(); index++) { 120 result += frame_sizes()[index]; 121 } 122 return result; 123 } 124 125 126 void Deoptimization::UnrollBlock::print() { 127 ttyLocker ttyl; 128 tty->print_cr("UnrollBlock"); 129 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 130 tty->print( " frame_sizes: "); 131 for (int index = 0; index < number_of_frames(); index++) { 132 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 133 } 134 tty->cr(); 135 } 136 137 138 // In order to make fetch_unroll_info work properly with escape 139 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 140 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 141 // of previously eliminated objects occurs in realloc_objects, which is 142 // called from the method fetch_unroll_info_helper below. 143 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 144 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 145 // but makes the entry a little slower. There is however a little dance we have to 146 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 147 148 // fetch_unroll_info() is called at the beginning of the deoptimization 149 // handler. Note this fact before we start generating temporary frames 150 // that can confuse an asynchronous stack walker. This counter is 151 // decremented at the end of unpack_frames(). 152 if (TraceDeoptimization) { 153 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 154 } 155 thread->inc_in_deopt_handler(); 156 157 return fetch_unroll_info_helper(thread, exec_mode); 158 JRT_END 159 160 161 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 162 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 163 164 // Note: there is a safepoint safety issue here. No matter whether we enter 165 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 166 // the vframeArray is created. 167 // 168 169 // Allocate our special deoptimization ResourceMark 170 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 171 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 172 thread->set_deopt_mark(dmark); 173 174 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 175 RegisterMap map(thread, true); 176 RegisterMap dummy_map(thread, false); 177 // Now get the deoptee with a valid map 178 frame deoptee = stub_frame.sender(&map); 179 // Set the deoptee nmethod 180 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 181 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 182 thread->set_deopt_compiled_method(cm); 183 184 if (VerifyStack) { 185 thread->validate_frame_layout(); 186 } 187 188 // Create a growable array of VFrames where each VFrame represents an inlined 189 // Java frame. This storage is allocated with the usual system arena. 190 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 191 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 192 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 193 while (!vf->is_top()) { 194 assert(vf->is_compiled_frame(), "Wrong frame type"); 195 chunk->push(compiledVFrame::cast(vf)); 196 vf = vf->sender(); 197 } 198 assert(vf->is_compiled_frame(), "Wrong frame type"); 199 chunk->push(compiledVFrame::cast(vf)); 200 201 bool realloc_failures = false; 202 203 #if COMPILER2_OR_JVMCI 204 // Reallocate the non-escaping objects and restore their fields. Then 205 // relock objects if synchronization on them was eliminated. 206 #if !INCLUDE_JVMCI 207 if (DoEscapeAnalysis || EliminateNestedLocks) { 208 if (EliminateAllocations) { 209 #endif // INCLUDE_JVMCI 210 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 211 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 212 213 // The flag return_oop() indicates call sites which return oop 214 // in compiled code. Such sites include java method calls, 215 // runtime calls (for example, used to allocate new objects/arrays 216 // on slow code path) and any other calls generated in compiled code. 217 // It is not guaranteed that we can get such information here only 218 // by analyzing bytecode in deoptimized frames. This is why this flag 219 // is set during method compilation (see Compile::Process_OopMap_Node()). 220 // If the previous frame was popped or if we are dispatching an exception, 221 // we don't have an oop result. 222 ScopeDesc* scope = chunk->at(0)->scope(); 223 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt); 224 // In case of the return of multiple values, we must take care 225 // of all oop return values. 226 GrowableArray<Handle> return_oops; 227 ValueKlass* vk = NULL; 228 if (save_oop_result && scope->return_vt()) { 229 vk = ValueKlass::returned_value_klass(map); 230 if (vk != NULL) { 231 vk->save_oop_fields(map, return_oops); 232 save_oop_result = false; 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 #if !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 #if !INCLUDE_JVMCI 307 } 308 } 309 #endif // INCLUDE_JVMCI 310 #endif // COMPILER2_OR_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 COMPILER2_OR_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 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 368 CodeBlob* cb = stub_frame.cb(); 369 // Verify we have the right vframeArray 370 assert(cb->frame_size() >= 0, "Unexpected frame size"); 371 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 372 373 // If the deopt call site is a MethodHandle invoke call site we have 374 // to adjust the unpack_sp. 375 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 376 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 377 unpack_sp = deoptee.unextended_sp(); 378 379 #ifdef ASSERT 380 assert(cb->is_deoptimization_stub() || 381 cb->is_uncommon_trap_stub() || 382 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 383 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 384 "unexpected code blob: %s", cb->name()); 385 #endif 386 387 // This is a guarantee instead of an assert because if vframe doesn't match 388 // we will unpack the wrong deoptimized frame and wind up in strange places 389 // where it will be very difficult to figure out what went wrong. Better 390 // to die an early death here than some very obscure death later when the 391 // trail is cold. 392 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 393 // in that it will fail to detect a problem when there is one. This needs 394 // more work in tiger timeframe. 395 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 396 397 int number_of_frames = array->frames(); 398 399 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 400 // virtual activation, which is the reverse of the elements in the vframes array. 401 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 402 // +1 because we always have an interpreter return address for the final slot. 403 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 404 int popframe_extra_args = 0; 405 // Create an interpreter return address for the stub to use as its return 406 // address so the skeletal frames are perfectly walkable 407 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 408 409 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 410 // activation be put back on the expression stack of the caller for reexecution 411 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 412 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 413 } 414 415 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 416 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 417 // than simply use array->sender.pc(). This requires us to walk the current set of frames 418 // 419 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 420 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 421 422 // It's possible that the number of parameters at the call site is 423 // different than number of arguments in the callee when method 424 // handles are used. If the caller is interpreted get the real 425 // value so that the proper amount of space can be added to it's 426 // frame. 427 bool caller_was_method_handle = false; 428 if (deopt_sender.is_interpreted_frame()) { 429 methodHandle method = deopt_sender.interpreter_frame_method(); 430 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 431 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 432 // Method handle invokes may involve fairly arbitrary chains of 433 // calls so it's impossible to know how much actual space the 434 // caller has for locals. 435 caller_was_method_handle = true; 436 } 437 } 438 439 // 440 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 441 // frame_sizes/frame_pcs[1] next oldest frame (int) 442 // frame_sizes/frame_pcs[n] youngest frame (int) 443 // 444 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 445 // owns the space for the return address to it's caller). Confusing ain't it. 446 // 447 // The vframe array can address vframes with indices running from 448 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 449 // When we create the skeletal frames we need the oldest frame to be in the zero slot 450 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 451 // so things look a little strange in this loop. 452 // 453 int callee_parameters = 0; 454 int callee_locals = 0; 455 for (int index = 0; index < array->frames(); index++ ) { 456 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 457 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 458 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 459 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 460 callee_locals, 461 index == 0, 462 popframe_extra_args); 463 // This pc doesn't have to be perfect just good enough to identify the frame 464 // as interpreted so the skeleton frame will be walkable 465 // The correct pc will be set when the skeleton frame is completely filled out 466 // The final pc we store in the loop is wrong and will be overwritten below 467 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 468 469 callee_parameters = array->element(index)->method()->size_of_parameters(); 470 callee_locals = array->element(index)->method()->max_locals(); 471 popframe_extra_args = 0; 472 } 473 474 // Compute whether the root vframe returns a float or double value. 475 BasicType return_type; 476 { 477 methodHandle method(thread, array->element(0)->method()); 478 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 479 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 480 } 481 482 // Compute information for handling adapters and adjusting the frame size of the caller. 483 int caller_adjustment = 0; 484 485 // Compute the amount the oldest interpreter frame will have to adjust 486 // its caller's stack by. If the caller is a compiled frame then 487 // we pretend that the callee has no parameters so that the 488 // extension counts for the full amount of locals and not just 489 // locals-parms. This is because without a c2i adapter the parm 490 // area as created by the compiled frame will not be usable by 491 // the interpreter. (Depending on the calling convention there 492 // may not even be enough space). 493 494 // QQQ I'd rather see this pushed down into last_frame_adjust 495 // and have it take the sender (aka caller). 496 497 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 498 caller_adjustment = last_frame_adjust(0, callee_locals); 499 } else if (callee_locals > callee_parameters) { 500 // The caller frame may need extending to accommodate 501 // non-parameter locals of the first unpacked interpreted frame. 502 // Compute that adjustment. 503 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 504 } 505 506 // If the sender is deoptimized the we must retrieve the address of the handler 507 // since the frame will "magically" show the original pc before the deopt 508 // and we'd undo the deopt. 509 510 frame_pcs[0] = deopt_sender.raw_pc(); 511 512 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 513 514 #if INCLUDE_JVMCI 515 if (exceptionObject() != NULL) { 516 thread->set_exception_oop(exceptionObject()); 517 exec_mode = Unpack_exception; 518 } 519 #endif 520 521 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 522 assert(thread->has_pending_exception(), "should have thrown OOME"); 523 thread->set_exception_oop(thread->pending_exception()); 524 thread->clear_pending_exception(); 525 exec_mode = Unpack_exception; 526 } 527 528 #if INCLUDE_JVMCI 529 if (thread->frames_to_pop_failed_realloc() > 0) { 530 thread->set_pending_monitorenter(false); 531 } 532 #endif 533 534 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 535 caller_adjustment * BytesPerWord, 536 caller_was_method_handle ? 0 : callee_parameters, 537 number_of_frames, 538 frame_sizes, 539 frame_pcs, 540 return_type, 541 exec_mode); 542 // On some platforms, we need a way to pass some platform dependent 543 // information to the unpacking code so the skeletal frames come out 544 // correct (initial fp value, unextended sp, ...) 545 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 546 547 if (array->frames() > 1) { 548 if (VerifyStack && TraceDeoptimization) { 549 ttyLocker ttyl; 550 tty->print_cr("Deoptimizing method containing inlining"); 551 } 552 } 553 554 array->set_unroll_block(info); 555 return info; 556 } 557 558 // Called to cleanup deoptimization data structures in normal case 559 // after unpacking to stack and when stack overflow error occurs 560 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 561 vframeArray *array) { 562 563 // Get array if coming from exception 564 if (array == NULL) { 565 array = thread->vframe_array_head(); 566 } 567 thread->set_vframe_array_head(NULL); 568 569 // Free the previous UnrollBlock 570 vframeArray* old_array = thread->vframe_array_last(); 571 thread->set_vframe_array_last(array); 572 573 if (old_array != NULL) { 574 UnrollBlock* old_info = old_array->unroll_block(); 575 old_array->set_unroll_block(NULL); 576 delete old_info; 577 delete old_array; 578 } 579 580 // Deallocate any resource creating in this routine and any ResourceObjs allocated 581 // inside the vframeArray (StackValueCollections) 582 583 delete thread->deopt_mark(); 584 thread->set_deopt_mark(NULL); 585 thread->set_deopt_compiled_method(NULL); 586 587 588 if (JvmtiExport::can_pop_frame()) { 589 #ifndef CC_INTERP 590 // Regardless of whether we entered this routine with the pending 591 // popframe condition bit set, we should always clear it now 592 thread->clear_popframe_condition(); 593 #else 594 // C++ interpreter will clear has_pending_popframe when it enters 595 // with method_resume. For deopt_resume2 we clear it now. 596 if (thread->popframe_forcing_deopt_reexecution()) 597 thread->clear_popframe_condition(); 598 #endif /* CC_INTERP */ 599 } 600 601 // unpack_frames() is called at the end of the deoptimization handler 602 // and (in C2) at the end of the uncommon trap handler. Note this fact 603 // so that an asynchronous stack walker can work again. This counter is 604 // incremented at the beginning of fetch_unroll_info() and (in C2) at 605 // the beginning of uncommon_trap(). 606 thread->dec_in_deopt_handler(); 607 } 608 609 // Moved from cpu directories because none of the cpus has callee save values. 610 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 611 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 612 613 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 614 // the days we had adapter frames. When we deoptimize a situation where a 615 // compiled caller calls a compiled caller will have registers it expects 616 // to survive the call to the callee. If we deoptimize the callee the only 617 // way we can restore these registers is to have the oldest interpreter 618 // frame that we create restore these values. That is what this routine 619 // will accomplish. 620 621 // At the moment we have modified c2 to not have any callee save registers 622 // so this problem does not exist and this routine is just a place holder. 623 624 assert(f->is_interpreted_frame(), "must be interpreted"); 625 } 626 627 // Return BasicType of value being returned 628 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 629 630 // We are already active in the special DeoptResourceMark any ResourceObj's we 631 // allocate will be freed at the end of the routine. 632 633 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 634 // but makes the entry a little slower. There is however a little dance we have to 635 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 636 ResetNoHandleMark rnhm; // No-op in release/product versions 637 HandleMark hm; 638 639 frame stub_frame = thread->last_frame(); 640 641 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 642 // must point to the vframeArray for the unpack frame. 643 vframeArray* array = thread->vframe_array_head(); 644 645 #ifndef PRODUCT 646 if (TraceDeoptimization) { 647 ttyLocker ttyl; 648 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 649 p2i(thread), p2i(array), exec_mode); 650 } 651 #endif 652 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 653 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 654 655 UnrollBlock* info = array->unroll_block(); 656 657 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 658 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 659 660 BasicType bt = info->return_type(); 661 662 // If we have an exception pending, claim that the return type is an oop 663 // so the deopt_blob does not overwrite the exception_oop. 664 665 if (exec_mode == Unpack_exception) 666 bt = T_OBJECT; 667 668 // Cleanup thread deopt data 669 cleanup_deopt_info(thread, array); 670 671 #ifndef PRODUCT 672 if (VerifyStack) { 673 ResourceMark res_mark; 674 // Clear pending exception to not break verification code (restored afterwards) 675 PRESERVE_EXCEPTION_MARK; 676 677 thread->validate_frame_layout(); 678 679 // Verify that the just-unpacked frames match the interpreter's 680 // notions of expression stack and locals 681 vframeArray* cur_array = thread->vframe_array_last(); 682 RegisterMap rm(thread, false); 683 rm.set_include_argument_oops(false); 684 bool is_top_frame = true; 685 int callee_size_of_parameters = 0; 686 int callee_max_locals = 0; 687 for (int i = 0; i < cur_array->frames(); i++) { 688 vframeArrayElement* el = cur_array->element(i); 689 frame* iframe = el->iframe(); 690 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 691 692 // Get the oop map for this bci 693 InterpreterOopMap mask; 694 int cur_invoke_parameter_size = 0; 695 bool try_next_mask = false; 696 int next_mask_expression_stack_size = -1; 697 int top_frame_expression_stack_adjustment = 0; 698 methodHandle mh(thread, iframe->interpreter_frame_method()); 699 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 700 BytecodeStream str(mh); 701 str.set_start(iframe->interpreter_frame_bci()); 702 int max_bci = mh->code_size(); 703 // Get to the next bytecode if possible 704 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 705 // Check to see if we can grab the number of outgoing arguments 706 // at an uncommon trap for an invoke (where the compiler 707 // generates debug info before the invoke has executed) 708 Bytecodes::Code cur_code = str.next(); 709 if (Bytecodes::is_invoke(cur_code)) { 710 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 711 cur_invoke_parameter_size = invoke.size_of_parameters(); 712 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 713 callee_size_of_parameters++; 714 } 715 } 716 if (str.bci() < max_bci) { 717 Bytecodes::Code next_code = str.next(); 718 if (next_code >= 0) { 719 // The interpreter oop map generator reports results before 720 // the current bytecode has executed except in the case of 721 // calls. It seems to be hard to tell whether the compiler 722 // has emitted debug information matching the "state before" 723 // a given bytecode or the state after, so we try both 724 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 725 // Get expression stack size for the next bytecode 726 if (Bytecodes::is_invoke(next_code)) { 727 Bytecode_invoke invoke(mh, str.bci()); 728 next_mask_expression_stack_size = invoke.size_of_parameters(); 729 } else { 730 InterpreterOopMap next_mask; 731 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 732 next_mask_expression_stack_size = next_mask.expression_stack_size(); 733 } 734 // Need to subtract off the size of the result type of 735 // the bytecode because this is not described in the 736 // debug info but returned to the interpreter in the TOS 737 // caching register 738 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 739 if (bytecode_result_type != T_ILLEGAL) { 740 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 741 } 742 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 743 try_next_mask = true; 744 } 745 } 746 } 747 748 // Verify stack depth and oops in frame 749 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 750 if (!( 751 /* SPARC */ 752 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 753 /* x86 */ 754 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 755 (try_next_mask && 756 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 757 top_frame_expression_stack_adjustment))) || 758 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 759 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 760 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 761 )) { 762 ttyLocker ttyl; 763 764 // Print out some information that will help us debug the problem 765 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 766 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 767 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 768 iframe->interpreter_frame_expression_stack_size()); 769 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 770 tty->print_cr(" try_next_mask = %d", try_next_mask); 771 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 772 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 773 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 774 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 775 tty->print_cr(" exec_mode = %d", exec_mode); 776 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 777 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 778 tty->print_cr(" Interpreted frames:"); 779 for (int k = 0; k < cur_array->frames(); k++) { 780 vframeArrayElement* el = cur_array->element(k); 781 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 782 } 783 cur_array->print_on_2(tty); 784 guarantee(false, "wrong number of expression stack elements during deopt"); 785 } 786 VerifyOopClosure verify; 787 iframe->oops_interpreted_do(&verify, &rm, false); 788 callee_size_of_parameters = mh->size_of_parameters(); 789 callee_max_locals = mh->max_locals(); 790 is_top_frame = false; 791 } 792 } 793 #endif /* !PRODUCT */ 794 795 796 return bt; 797 JRT_END 798 799 800 int Deoptimization::deoptimize_dependents() { 801 Threads::deoptimized_wrt_marked_nmethods(); 802 return 0; 803 } 804 805 Deoptimization::DeoptAction Deoptimization::_unloaded_action 806 = Deoptimization::Action_reinterpret; 807 808 #if COMPILER2_OR_JVMCI 809 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 810 Handle pending_exception(THREAD, thread->pending_exception()); 811 const char* exception_file = thread->exception_file(); 812 int exception_line = thread->exception_line(); 813 thread->clear_pending_exception(); 814 815 bool failures = false; 816 817 for (int i = 0; i < objects->length(); i++) { 818 assert(objects->at(i)->is_object(), "invalid debug information"); 819 ObjectValue* sv = (ObjectValue*) objects->at(i); 820 821 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 822 oop obj = NULL; 823 824 if (k->is_instance_klass()) { 825 InstanceKlass* ik = InstanceKlass::cast(k); 826 obj = ik->allocate_instance(THREAD); 827 } else if (k->is_valueArray_klass()) { 828 ValueArrayKlass* ak = ValueArrayKlass::cast(k); 829 // Value type array must be zeroed because not all memory is reassigned 830 obj = ak->allocate(sv->field_size(), true, THREAD); 831 } else if (k->is_typeArray_klass()) { 832 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 833 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 834 int len = sv->field_size() / type2size[ak->element_type()]; 835 obj = ak->allocate(len, THREAD); 836 } else if (k->is_objArray_klass()) { 837 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 838 obj = ak->allocate(sv->field_size(), THREAD); 839 } 840 841 if (obj == NULL) { 842 failures = true; 843 } 844 845 assert(sv->value().is_null(), "redundant reallocation"); 846 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 847 CLEAR_PENDING_EXCEPTION; 848 sv->set_value(obj); 849 } 850 851 if (failures) { 852 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 853 } else if (pending_exception.not_null()) { 854 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 855 } 856 857 return failures; 858 } 859 860 // We're deoptimizing at the return of a call, value type fields are 861 // in registers. When we go back to the interpreter, it will expect a 862 // reference to a value type instance. Allocate and initialize it from 863 // the register values here. 864 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) { 865 oop new_vt = vk->realloc_result(map, return_oops, false, THREAD); 866 if (new_vt == NULL) { 867 CLEAR_PENDING_EXCEPTION; 868 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true); 869 } 870 return_oops.clear(); 871 return_oops.push(Handle(THREAD, new_vt)); 872 return false; 873 } 874 875 // restore elements of an eliminated type array 876 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 877 int index = 0; 878 intptr_t val; 879 880 for (int i = 0; i < sv->field_size(); i++) { 881 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 882 switch(type) { 883 case T_LONG: case T_DOUBLE: { 884 assert(value->type() == T_INT, "Agreement."); 885 StackValue* low = 886 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 887 #ifdef _LP64 888 jlong res = (jlong)low->get_int(); 889 #else 890 #ifdef SPARC 891 // For SPARC we have to swap high and low words. 892 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 893 #else 894 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 895 #endif //SPARC 896 #endif 897 obj->long_at_put(index, res); 898 break; 899 } 900 901 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 902 case T_INT: case T_FLOAT: { // 4 bytes. 903 assert(value->type() == T_INT, "Agreement."); 904 bool big_value = false; 905 if (i + 1 < sv->field_size() && type == T_INT) { 906 if (sv->field_at(i)->is_location()) { 907 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 908 if (type == Location::dbl || type == Location::lng) { 909 big_value = true; 910 } 911 } else if (sv->field_at(i)->is_constant_int()) { 912 ScopeValue* next_scope_field = sv->field_at(i + 1); 913 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 914 big_value = true; 915 } 916 } 917 } 918 919 if (big_value) { 920 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 921 #ifdef _LP64 922 jlong res = (jlong)low->get_int(); 923 #else 924 #ifdef SPARC 925 // For SPARC we have to swap high and low words. 926 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 927 #else 928 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 929 #endif //SPARC 930 #endif 931 obj->int_at_put(index, (jint)*((jint*)&res)); 932 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 933 } else { 934 val = value->get_int(); 935 obj->int_at_put(index, (jint)*((jint*)&val)); 936 } 937 break; 938 } 939 940 case T_SHORT: 941 assert(value->type() == T_INT, "Agreement."); 942 val = value->get_int(); 943 obj->short_at_put(index, (jshort)*((jint*)&val)); 944 break; 945 946 case T_CHAR: 947 assert(value->type() == T_INT, "Agreement."); 948 val = value->get_int(); 949 obj->char_at_put(index, (jchar)*((jint*)&val)); 950 break; 951 952 case T_BYTE: 953 assert(value->type() == T_INT, "Agreement."); 954 val = value->get_int(); 955 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 956 break; 957 958 case T_BOOLEAN: 959 assert(value->type() == T_INT, "Agreement."); 960 val = value->get_int(); 961 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 962 break; 963 964 default: 965 ShouldNotReachHere(); 966 } 967 index++; 968 } 969 } 970 971 972 // restore fields of an eliminated object array 973 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 974 for (int i = 0; i < sv->field_size(); i++) { 975 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 976 assert(value->type() == T_OBJECT, "object element expected"); 977 obj->obj_at_put(i, value->get_obj()()); 978 } 979 } 980 981 class ReassignedField { 982 public: 983 int _offset; 984 BasicType _type; 985 InstanceKlass* _klass; 986 public: 987 ReassignedField() { 988 _offset = 0; 989 _type = T_ILLEGAL; 990 _klass = NULL; 991 } 992 }; 993 994 int compare(ReassignedField* left, ReassignedField* right) { 995 return left->_offset - right->_offset; 996 } 997 998 // Restore fields of an eliminated instance object using the same field order 999 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1000 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) { 1001 if (klass->superklass() != NULL) { 1002 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0); 1003 } 1004 1005 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1006 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1007 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1008 ReassignedField field; 1009 field._offset = fs.offset(); 1010 field._type = fs.is_flattened() ? T_VALUETYPE : FieldType::basic_type(fs.signature()); 1011 if (field._type == T_VALUETYPE) { 1012 if (fs.is_flattened()) { 1013 // Resolve klass of flattened value type field 1014 SignatureStream ss(fs.signature(), false); 1015 Klass* vk = ss.as_klass(Handle(THREAD, klass->class_loader()), Handle(THREAD, klass->protection_domain()), SignatureStream::NCDFError, THREAD); 1016 guarantee(!HAS_PENDING_EXCEPTION, "Should not have any exceptions pending"); 1017 assert(vk->is_value(), "must be a ValueKlass"); 1018 field._klass = InstanceKlass::cast(vk); 1019 } else { 1020 // Non-flattened value type field 1021 field._type = T_VALUETYPEPTR; 1022 } 1023 } 1024 fields->append(field); 1025 } 1026 } 1027 fields->sort(compare); 1028 for (int i = 0; i < fields->length(); i++) { 1029 intptr_t val; 1030 ScopeValue* scope_field = sv->field_at(svIndex); 1031 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1032 int offset = base_offset + fields->at(i)._offset; 1033 BasicType type = fields->at(i)._type; 1034 switch (type) { 1035 case T_OBJECT: 1036 case T_VALUETYPEPTR: 1037 case T_ARRAY: 1038 assert(value->type() == T_OBJECT, "Agreement."); 1039 obj->obj_field_put(offset, value->get_obj()()); 1040 break; 1041 1042 case T_VALUETYPE: { 1043 // Recursively re-assign flattened value type fields 1044 InstanceKlass* vk = fields->at(i)._klass; 1045 assert(vk != NULL, "must be resolved"); 1046 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 1047 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 1048 continue; // Continue because we don't need to increment svIndex 1049 } 1050 1051 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1052 case T_INT: case T_FLOAT: { // 4 bytes. 1053 assert(value->type() == T_INT, "Agreement."); 1054 bool big_value = false; 1055 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1056 if (scope_field->is_location()) { 1057 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1058 if (type == Location::dbl || type == Location::lng) { 1059 big_value = true; 1060 } 1061 } 1062 if (scope_field->is_constant_int()) { 1063 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1064 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1065 big_value = true; 1066 } 1067 } 1068 } 1069 1070 if (big_value) { 1071 i++; 1072 assert(i < fields->length(), "second T_INT field needed"); 1073 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1074 } else { 1075 val = value->get_int(); 1076 obj->int_field_put(offset, (jint)*((jint*)&val)); 1077 break; 1078 } 1079 } 1080 /* no break */ 1081 1082 case T_LONG: case T_DOUBLE: { 1083 assert(value->type() == T_INT, "Agreement."); 1084 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1085 #ifdef _LP64 1086 jlong res = (jlong)low->get_int(); 1087 #else 1088 #ifdef SPARC 1089 // For SPARC we have to swap high and low words. 1090 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1091 #else 1092 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1093 #endif //SPARC 1094 #endif 1095 obj->long_field_put(offset, res); 1096 break; 1097 } 1098 1099 case T_SHORT: 1100 assert(value->type() == T_INT, "Agreement."); 1101 val = value->get_int(); 1102 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1103 break; 1104 1105 case T_CHAR: 1106 assert(value->type() == T_INT, "Agreement."); 1107 val = value->get_int(); 1108 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1109 break; 1110 1111 case T_BYTE: 1112 assert(value->type() == T_INT, "Agreement."); 1113 val = value->get_int(); 1114 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1115 break; 1116 1117 case T_BOOLEAN: 1118 assert(value->type() == T_INT, "Agreement."); 1119 val = value->get_int(); 1120 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1121 break; 1122 1123 default: 1124 ShouldNotReachHere(); 1125 } 1126 svIndex++; 1127 } 1128 return svIndex; 1129 } 1130 1131 // restore fields of an eliminated value type array 1132 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) { 1133 ValueKlass* vk = vak->element_klass(); 1134 assert(vk->flatten_array(), "should only be used for flattened value type arrays"); 1135 // Adjust offset to omit oop header 1136 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset(); 1137 // Initialize all elements of the flattened value type array 1138 for (int i = 0; i < sv->field_size(); i++) { 1139 ScopeValue* val = sv->field_at(i); 1140 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper())); 1141 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK); 1142 } 1143 } 1144 1145 // restore fields of all eliminated objects and arrays 1146 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1147 for (int i = 0; i < objects->length(); i++) { 1148 ObjectValue* sv = (ObjectValue*) objects->at(i); 1149 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1150 Handle obj = sv->value(); 1151 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1152 if (PrintDeoptimizationDetails) { 1153 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1154 } 1155 if (obj.is_null()) { 1156 continue; 1157 } 1158 1159 if (k->is_instance_klass()) { 1160 InstanceKlass* ik = InstanceKlass::cast(k); 1161 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1162 } else if (k->is_valueArray_klass()) { 1163 ValueArrayKlass* vak = ValueArrayKlass::cast(k); 1164 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK); 1165 } else if (k->is_typeArray_klass()) { 1166 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1167 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1168 } else if (k->is_objArray_klass()) { 1169 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1170 } 1171 } 1172 } 1173 1174 1175 // relock objects for which synchronization was eliminated 1176 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1177 for (int i = 0; i < monitors->length(); i++) { 1178 MonitorInfo* mon_info = monitors->at(i); 1179 if (mon_info->eliminated()) { 1180 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1181 if (!mon_info->owner_is_scalar_replaced()) { 1182 Handle obj(thread, mon_info->owner()); 1183 markOop mark = obj->mark(); 1184 if (UseBiasedLocking && mark->has_bias_pattern()) { 1185 // New allocated objects may have the mark set to anonymously biased. 1186 // Also the deoptimized method may called methods with synchronization 1187 // where the thread-local object is bias locked to the current thread. 1188 assert(mark->is_biased_anonymously() || 1189 mark->biased_locker() == thread, "should be locked to current thread"); 1190 // Reset mark word to unbiased prototype. 1191 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1192 obj->set_mark(unbiased_prototype); 1193 } 1194 BasicLock* lock = mon_info->lock(); 1195 ObjectSynchronizer::slow_enter(obj, lock, thread); 1196 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1197 } 1198 } 1199 } 1200 } 1201 1202 1203 #ifndef PRODUCT 1204 // print information about reallocated objects 1205 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1206 fieldDescriptor fd; 1207 1208 for (int i = 0; i < objects->length(); i++) { 1209 ObjectValue* sv = (ObjectValue*) objects->at(i); 1210 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1211 Handle obj = sv->value(); 1212 1213 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1214 k->print_value(); 1215 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1216 if (obj.is_null()) { 1217 tty->print(" allocation failed"); 1218 } else { 1219 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1220 } 1221 tty->cr(); 1222 1223 if (Verbose && !obj.is_null()) { 1224 k->oop_print_on(obj(), tty); 1225 } 1226 } 1227 } 1228 #endif 1229 #endif // COMPILER2_OR_JVMCI 1230 1231 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1232 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1233 1234 #ifndef PRODUCT 1235 if (PrintDeoptimizationDetails) { 1236 ttyLocker ttyl; 1237 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1238 fr.print_on(tty); 1239 tty->print_cr(" Virtual frames (innermost first):"); 1240 for (int index = 0; index < chunk->length(); index++) { 1241 compiledVFrame* vf = chunk->at(index); 1242 tty->print(" %2d - ", index); 1243 vf->print_value(); 1244 int bci = chunk->at(index)->raw_bci(); 1245 const char* code_name; 1246 if (bci == SynchronizationEntryBCI) { 1247 code_name = "sync entry"; 1248 } else { 1249 Bytecodes::Code code = vf->method()->code_at(bci); 1250 code_name = Bytecodes::name(code); 1251 } 1252 tty->print(" - %s", code_name); 1253 tty->print_cr(" @ bci %d ", bci); 1254 if (Verbose) { 1255 vf->print(); 1256 tty->cr(); 1257 } 1258 } 1259 } 1260 #endif 1261 1262 // Register map for next frame (used for stack crawl). We capture 1263 // the state of the deopt'ing frame's caller. Thus if we need to 1264 // stuff a C2I adapter we can properly fill in the callee-save 1265 // register locations. 1266 frame caller = fr.sender(reg_map); 1267 int frame_size = caller.sp() - fr.sp(); 1268 1269 frame sender = caller; 1270 1271 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1272 // the vframeArray containing the unpacking information is allocated in the C heap. 1273 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1274 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1275 1276 // Compare the vframeArray to the collected vframes 1277 assert(array->structural_compare(thread, chunk), "just checking"); 1278 1279 #ifndef PRODUCT 1280 if (PrintDeoptimizationDetails) { 1281 ttyLocker ttyl; 1282 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1283 } 1284 #endif // PRODUCT 1285 1286 return array; 1287 } 1288 1289 #if COMPILER2_OR_JVMCI 1290 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1291 // Reallocation of some scalar replaced objects failed. Record 1292 // that we need to pop all the interpreter frames for the 1293 // deoptimized compiled frame. 1294 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1295 thread->set_frames_to_pop_failed_realloc(array->frames()); 1296 // Unlock all monitors here otherwise the interpreter will see a 1297 // mix of locked and unlocked monitors (because of failed 1298 // reallocations of synchronized objects) and be confused. 1299 for (int i = 0; i < array->frames(); i++) { 1300 MonitorChunk* monitors = array->element(i)->monitors(); 1301 if (monitors != NULL) { 1302 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1303 BasicObjectLock* src = monitors->at(j); 1304 if (src->obj() != NULL) { 1305 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1306 } 1307 } 1308 array->element(i)->free_monitors(thread); 1309 #ifdef ASSERT 1310 array->element(i)->set_removed_monitors(); 1311 #endif 1312 } 1313 } 1314 } 1315 #endif 1316 1317 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1318 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1319 Thread* thread = Thread::current(); 1320 for (int i = 0; i < monitors->length(); i++) { 1321 MonitorInfo* mon_info = monitors->at(i); 1322 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1323 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1324 } 1325 } 1326 } 1327 1328 1329 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1330 if (!UseBiasedLocking) { 1331 return; 1332 } 1333 1334 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1335 1336 // Unfortunately we don't have a RegisterMap available in most of 1337 // the places we want to call this routine so we need to walk the 1338 // stack again to update the register map. 1339 if (map == NULL || !map->update_map()) { 1340 StackFrameStream sfs(thread, true); 1341 bool found = false; 1342 while (!found && !sfs.is_done()) { 1343 frame* cur = sfs.current(); 1344 sfs.next(); 1345 found = cur->id() == fr.id(); 1346 } 1347 assert(found, "frame to be deoptimized not found on target thread's stack"); 1348 map = sfs.register_map(); 1349 } 1350 1351 vframe* vf = vframe::new_vframe(&fr, map, thread); 1352 compiledVFrame* cvf = compiledVFrame::cast(vf); 1353 // Revoke monitors' biases in all scopes 1354 while (!cvf->is_top()) { 1355 collect_monitors(cvf, objects_to_revoke); 1356 cvf = compiledVFrame::cast(cvf->sender()); 1357 } 1358 collect_monitors(cvf, objects_to_revoke); 1359 1360 if (SafepointSynchronize::is_at_safepoint()) { 1361 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1362 } else { 1363 BiasedLocking::revoke(objects_to_revoke); 1364 } 1365 } 1366 1367 1368 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1369 if (!UseBiasedLocking) { 1370 return; 1371 } 1372 1373 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1374 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1375 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 1376 if (jt->has_last_Java_frame()) { 1377 StackFrameStream sfs(jt, true); 1378 while (!sfs.is_done()) { 1379 frame* cur = sfs.current(); 1380 if (cb->contains(cur->pc())) { 1381 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1382 compiledVFrame* cvf = compiledVFrame::cast(vf); 1383 // Revoke monitors' biases in all scopes 1384 while (!cvf->is_top()) { 1385 collect_monitors(cvf, objects_to_revoke); 1386 cvf = compiledVFrame::cast(cvf->sender()); 1387 } 1388 collect_monitors(cvf, objects_to_revoke); 1389 } 1390 sfs.next(); 1391 } 1392 } 1393 } 1394 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1395 } 1396 1397 1398 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1399 assert(fr.can_be_deoptimized(), "checking frame type"); 1400 1401 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1402 1403 if (LogCompilation && xtty != NULL) { 1404 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1405 assert(cm != NULL, "only compiled methods can deopt"); 1406 1407 ttyLocker ttyl; 1408 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1409 cm->log_identity(xtty); 1410 xtty->end_head(); 1411 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1412 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1413 xtty->method(sd->method()); 1414 xtty->end_elem(); 1415 if (sd->is_top()) break; 1416 } 1417 xtty->tail("deoptimized"); 1418 } 1419 1420 // Patch the compiled method so that when execution returns to it we will 1421 // deopt the execution state and return to the interpreter. 1422 fr.deoptimize(thread); 1423 } 1424 1425 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1426 deoptimize(thread, fr, map, Reason_constraint); 1427 } 1428 1429 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1430 // Deoptimize only if the frame comes from compile code. 1431 // Do not deoptimize the frame which is already patched 1432 // during the execution of the loops below. 1433 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1434 return; 1435 } 1436 ResourceMark rm; 1437 DeoptimizationMarker dm; 1438 if (UseBiasedLocking) { 1439 revoke_biases_of_monitors(thread, fr, map); 1440 } 1441 deoptimize_single_frame(thread, fr, reason); 1442 1443 } 1444 1445 #if INCLUDE_JVMCI 1446 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1447 // there is no exception handler for this pc => deoptimize 1448 cm->make_not_entrant(); 1449 1450 // Use Deoptimization::deoptimize for all of its side-effects: 1451 // revoking biases of monitors, gathering traps statistics, logging... 1452 // it also patches the return pc but we do not care about that 1453 // since we return a continuation to the deopt_blob below. 1454 JavaThread* thread = JavaThread::current(); 1455 RegisterMap reg_map(thread, UseBiasedLocking); 1456 frame runtime_frame = thread->last_frame(); 1457 frame caller_frame = runtime_frame.sender(®_map); 1458 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1459 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1460 1461 MethodData* trap_mdo = get_method_data(thread, cm->method(), true); 1462 if (trap_mdo != NULL) { 1463 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1464 } 1465 1466 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1467 } 1468 #endif 1469 1470 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1471 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1472 "can only deoptimize other thread at a safepoint"); 1473 // Compute frame and register map based on thread and sp. 1474 RegisterMap reg_map(thread, UseBiasedLocking); 1475 frame fr = thread->last_frame(); 1476 while (fr.id() != id) { 1477 fr = fr.sender(®_map); 1478 } 1479 deoptimize(thread, fr, ®_map, reason); 1480 } 1481 1482 1483 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1484 if (thread == Thread::current()) { 1485 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1486 } else { 1487 VM_DeoptimizeFrame deopt(thread, id, reason); 1488 VMThread::execute(&deopt); 1489 } 1490 } 1491 1492 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1493 deoptimize_frame(thread, id, Reason_constraint); 1494 } 1495 1496 // JVMTI PopFrame support 1497 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1498 { 1499 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1500 } 1501 JRT_END 1502 1503 MethodData* 1504 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1505 bool create_if_missing) { 1506 Thread* THREAD = thread; 1507 MethodData* mdo = m()->method_data(); 1508 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1509 // Build an MDO. Ignore errors like OutOfMemory; 1510 // that simply means we won't have an MDO to update. 1511 Method::build_interpreter_method_data(m, THREAD); 1512 if (HAS_PENDING_EXCEPTION) { 1513 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1514 CLEAR_PENDING_EXCEPTION; 1515 } 1516 mdo = m()->method_data(); 1517 } 1518 return mdo; 1519 } 1520 1521 #if COMPILER2_OR_JVMCI 1522 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1523 // in case of an unresolved klass entry, load the class. 1524 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1525 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1526 return; 1527 } 1528 1529 if (!constant_pool->tag_at(index).is_symbol()) return; 1530 1531 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1532 Symbol* symbol = constant_pool->symbol_at(index); 1533 1534 // class name? 1535 if (symbol->byte_at(0) != '(') { 1536 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1537 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1538 return; 1539 } 1540 1541 // then it must be a signature! 1542 ResourceMark rm(THREAD); 1543 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1544 if (ss.is_object()) { 1545 Symbol* class_name = ss.as_symbol(CHECK); 1546 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1547 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1548 } 1549 } 1550 } 1551 1552 1553 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1554 EXCEPTION_MARK; 1555 load_class_by_index(constant_pool, index, THREAD); 1556 if (HAS_PENDING_EXCEPTION) { 1557 // Exception happened during classloading. We ignore the exception here, since it 1558 // is going to be rethrown since the current activation is going to be deoptimized and 1559 // the interpreter will re-execute the bytecode. 1560 CLEAR_PENDING_EXCEPTION; 1561 // Class loading called java code which may have caused a stack 1562 // overflow. If the exception was thrown right before the return 1563 // to the runtime the stack is no longer guarded. Reguard the 1564 // stack otherwise if we return to the uncommon trap blob and the 1565 // stack bang causes a stack overflow we crash. 1566 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1567 JavaThread* thread = (JavaThread*)THREAD; 1568 bool guard_pages_enabled = thread->stack_guards_enabled(); 1569 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1570 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1571 } 1572 } 1573 1574 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1575 HandleMark hm; 1576 1577 // uncommon_trap() is called at the beginning of the uncommon trap 1578 // handler. Note this fact before we start generating temporary frames 1579 // that can confuse an asynchronous stack walker. This counter is 1580 // decremented at the end of unpack_frames(). 1581 thread->inc_in_deopt_handler(); 1582 1583 // We need to update the map if we have biased locking. 1584 #if INCLUDE_JVMCI 1585 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1586 RegisterMap reg_map(thread, true); 1587 #else 1588 RegisterMap reg_map(thread, UseBiasedLocking); 1589 #endif 1590 frame stub_frame = thread->last_frame(); 1591 frame fr = stub_frame.sender(®_map); 1592 // Make sure the calling nmethod is not getting deoptimized and removed 1593 // before we are done with it. 1594 nmethodLocker nl(fr.pc()); 1595 1596 // Log a message 1597 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1598 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1599 1600 { 1601 ResourceMark rm; 1602 1603 // Revoke biases of any monitors in the frame to ensure we can migrate them 1604 revoke_biases_of_monitors(thread, fr, ®_map); 1605 1606 DeoptReason reason = trap_request_reason(trap_request); 1607 DeoptAction action = trap_request_action(trap_request); 1608 #if INCLUDE_JVMCI 1609 int debug_id = trap_request_debug_id(trap_request); 1610 #endif 1611 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1612 1613 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1614 compiledVFrame* cvf = compiledVFrame::cast(vf); 1615 1616 CompiledMethod* nm = cvf->code(); 1617 1618 ScopeDesc* trap_scope = cvf->scope(); 1619 1620 if (TraceDeoptimization) { 1621 ttyLocker ttyl; 1622 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() 1623 #if INCLUDE_JVMCI 1624 , debug_id 1625 #endif 1626 ); 1627 } 1628 1629 methodHandle trap_method = trap_scope->method(); 1630 int trap_bci = trap_scope->bci(); 1631 #if INCLUDE_JVMCI 1632 oop speculation = thread->pending_failed_speculation(); 1633 if (nm->is_compiled_by_jvmci()) { 1634 if (speculation != NULL) { 1635 oop speculation_log = nm->as_nmethod()->speculation_log(); 1636 if (speculation_log != NULL) { 1637 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1638 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) { 1639 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1640 } 1641 } 1642 if (TraceDeoptimization) { 1643 tty->print_cr("Saving speculation to speculation log"); 1644 } 1645 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1646 } else { 1647 if (TraceDeoptimization) { 1648 tty->print_cr("Speculation present but no speculation log"); 1649 } 1650 } 1651 thread->set_pending_failed_speculation(NULL); 1652 } else { 1653 if (TraceDeoptimization) { 1654 tty->print_cr("No speculation"); 1655 } 1656 } 1657 } else { 1658 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1659 } 1660 1661 if (trap_bci == SynchronizationEntryBCI) { 1662 trap_bci = 0; 1663 thread->set_pending_monitorenter(true); 1664 } 1665 1666 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1667 thread->set_pending_transfer_to_interpreter(true); 1668 } 1669 #endif 1670 1671 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1672 // Record this event in the histogram. 1673 gather_statistics(reason, action, trap_bc); 1674 1675 // Ensure that we can record deopt. history: 1676 // Need MDO to record RTM code generation state. 1677 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1678 1679 methodHandle profiled_method; 1680 #if INCLUDE_JVMCI 1681 if (nm->is_compiled_by_jvmci()) { 1682 profiled_method = nm->method(); 1683 } else { 1684 profiled_method = trap_method; 1685 } 1686 #else 1687 profiled_method = trap_method; 1688 #endif 1689 1690 MethodData* trap_mdo = 1691 get_method_data(thread, profiled_method, create_if_missing); 1692 1693 // Log a message 1694 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1695 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1696 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1697 1698 // Print a bunch of diagnostics, if requested. 1699 if (TraceDeoptimization || LogCompilation) { 1700 ResourceMark rm; 1701 ttyLocker ttyl; 1702 char buf[100]; 1703 if (xtty != NULL) { 1704 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1705 os::current_thread_id(), 1706 format_trap_request(buf, sizeof(buf), trap_request)); 1707 nm->log_identity(xtty); 1708 } 1709 Symbol* class_name = NULL; 1710 bool unresolved = false; 1711 if (unloaded_class_index >= 0) { 1712 constantPoolHandle constants (THREAD, trap_method->constants()); 1713 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1714 class_name = constants->klass_name_at(unloaded_class_index); 1715 unresolved = true; 1716 if (xtty != NULL) 1717 xtty->print(" unresolved='1'"); 1718 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1719 class_name = constants->symbol_at(unloaded_class_index); 1720 } 1721 if (xtty != NULL) 1722 xtty->name(class_name); 1723 } 1724 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1725 // Dump the relevant MDO state. 1726 // This is the deopt count for the current reason, any previous 1727 // reasons or recompiles seen at this point. 1728 int dcnt = trap_mdo->trap_count(reason); 1729 if (dcnt != 0) 1730 xtty->print(" count='%d'", dcnt); 1731 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1732 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1733 if (dos != 0) { 1734 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1735 if (trap_state_is_recompiled(dos)) { 1736 int recnt2 = trap_mdo->overflow_recompile_count(); 1737 if (recnt2 != 0) 1738 xtty->print(" recompiles2='%d'", recnt2); 1739 } 1740 } 1741 } 1742 if (xtty != NULL) { 1743 xtty->stamp(); 1744 xtty->end_head(); 1745 } 1746 if (TraceDeoptimization) { // make noise on the tty 1747 tty->print("Uncommon trap occurred in"); 1748 nm->method()->print_short_name(tty); 1749 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1750 #if INCLUDE_JVMCI 1751 if (nm->is_nmethod()) { 1752 char* installed_code_name = nm->as_nmethod()->jvmci_installed_code_name(buf, sizeof(buf)); 1753 if (installed_code_name != NULL) { 1754 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1755 } 1756 } 1757 #endif 1758 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1759 p2i(fr.pc()), 1760 os::current_thread_id(), 1761 trap_reason_name(reason), 1762 trap_action_name(action), 1763 unloaded_class_index 1764 #if INCLUDE_JVMCI 1765 , debug_id 1766 #endif 1767 ); 1768 if (class_name != NULL) { 1769 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1770 class_name->print_symbol_on(tty); 1771 } 1772 tty->cr(); 1773 } 1774 if (xtty != NULL) { 1775 // Log the precise location of the trap. 1776 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1777 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1778 xtty->method(sd->method()); 1779 xtty->end_elem(); 1780 if (sd->is_top()) break; 1781 } 1782 xtty->tail("uncommon_trap"); 1783 } 1784 } 1785 // (End diagnostic printout.) 1786 1787 // Load class if necessary 1788 if (unloaded_class_index >= 0) { 1789 constantPoolHandle constants(THREAD, trap_method->constants()); 1790 load_class_by_index(constants, unloaded_class_index); 1791 } 1792 1793 // Flush the nmethod if necessary and desirable. 1794 // 1795 // We need to avoid situations where we are re-flushing the nmethod 1796 // because of a hot deoptimization site. Repeated flushes at the same 1797 // point need to be detected by the compiler and avoided. If the compiler 1798 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1799 // module must take measures to avoid an infinite cycle of recompilation 1800 // and deoptimization. There are several such measures: 1801 // 1802 // 1. If a recompilation is ordered a second time at some site X 1803 // and for the same reason R, the action is adjusted to 'reinterpret', 1804 // to give the interpreter time to exercise the method more thoroughly. 1805 // If this happens, the method's overflow_recompile_count is incremented. 1806 // 1807 // 2. If the compiler fails to reduce the deoptimization rate, then 1808 // the method's overflow_recompile_count will begin to exceed the set 1809 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1810 // is adjusted to 'make_not_compilable', and the method is abandoned 1811 // to the interpreter. This is a performance hit for hot methods, 1812 // but is better than a disastrous infinite cycle of recompilations. 1813 // (Actually, only the method containing the site X is abandoned.) 1814 // 1815 // 3. In parallel with the previous measures, if the total number of 1816 // recompilations of a method exceeds the much larger set limit 1817 // PerMethodRecompilationCutoff, the method is abandoned. 1818 // This should only happen if the method is very large and has 1819 // many "lukewarm" deoptimizations. The code which enforces this 1820 // limit is elsewhere (class nmethod, class Method). 1821 // 1822 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1823 // to recompile at each bytecode independently of the per-BCI cutoff. 1824 // 1825 // The decision to update code is up to the compiler, and is encoded 1826 // in the Action_xxx code. If the compiler requests Action_none 1827 // no trap state is changed, no compiled code is changed, and the 1828 // computation suffers along in the interpreter. 1829 // 1830 // The other action codes specify various tactics for decompilation 1831 // and recompilation. Action_maybe_recompile is the loosest, and 1832 // allows the compiled code to stay around until enough traps are seen, 1833 // and until the compiler gets around to recompiling the trapping method. 1834 // 1835 // The other actions cause immediate removal of the present code. 1836 1837 // Traps caused by injected profile shouldn't pollute trap counts. 1838 bool injected_profile_trap = trap_method->has_injected_profile() && 1839 (reason == Reason_intrinsic || reason == Reason_unreached); 1840 1841 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1842 bool make_not_entrant = false; 1843 bool make_not_compilable = false; 1844 bool reprofile = false; 1845 switch (action) { 1846 case Action_none: 1847 // Keep the old code. 1848 update_trap_state = false; 1849 break; 1850 case Action_maybe_recompile: 1851 // Do not need to invalidate the present code, but we can 1852 // initiate another 1853 // Start compiler without (necessarily) invalidating the nmethod. 1854 // The system will tolerate the old code, but new code should be 1855 // generated when possible. 1856 break; 1857 case Action_reinterpret: 1858 // Go back into the interpreter for a while, and then consider 1859 // recompiling form scratch. 1860 make_not_entrant = true; 1861 // Reset invocation counter for outer most method. 1862 // This will allow the interpreter to exercise the bytecodes 1863 // for a while before recompiling. 1864 // By contrast, Action_make_not_entrant is immediate. 1865 // 1866 // Note that the compiler will track null_check, null_assert, 1867 // range_check, and class_check events and log them as if they 1868 // had been traps taken from compiled code. This will update 1869 // the MDO trap history so that the next compilation will 1870 // properly detect hot trap sites. 1871 reprofile = true; 1872 break; 1873 case Action_make_not_entrant: 1874 // Request immediate recompilation, and get rid of the old code. 1875 // Make them not entrant, so next time they are called they get 1876 // recompiled. Unloaded classes are loaded now so recompile before next 1877 // time they are called. Same for uninitialized. The interpreter will 1878 // link the missing class, if any. 1879 make_not_entrant = true; 1880 break; 1881 case Action_make_not_compilable: 1882 // Give up on compiling this method at all. 1883 make_not_entrant = true; 1884 make_not_compilable = true; 1885 break; 1886 default: 1887 ShouldNotReachHere(); 1888 } 1889 1890 // Setting +ProfileTraps fixes the following, on all platforms: 1891 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1892 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1893 // recompile relies on a MethodData* to record heroic opt failures. 1894 1895 // Whether the interpreter is producing MDO data or not, we also need 1896 // to use the MDO to detect hot deoptimization points and control 1897 // aggressive optimization. 1898 bool inc_recompile_count = false; 1899 ProfileData* pdata = NULL; 1900 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1901 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1902 uint this_trap_count = 0; 1903 bool maybe_prior_trap = false; 1904 bool maybe_prior_recompile = false; 1905 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1906 #if INCLUDE_JVMCI 1907 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1908 #endif 1909 nm->method(), 1910 //outputs: 1911 this_trap_count, 1912 maybe_prior_trap, 1913 maybe_prior_recompile); 1914 // Because the interpreter also counts null, div0, range, and class 1915 // checks, these traps from compiled code are double-counted. 1916 // This is harmless; it just means that the PerXTrapLimit values 1917 // are in effect a little smaller than they look. 1918 1919 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1920 if (per_bc_reason != Reason_none) { 1921 // Now take action based on the partially known per-BCI history. 1922 if (maybe_prior_trap 1923 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1924 // If there are too many traps at this BCI, force a recompile. 1925 // This will allow the compiler to see the limit overflow, and 1926 // take corrective action, if possible. The compiler generally 1927 // does not use the exact PerBytecodeTrapLimit value, but instead 1928 // changes its tactics if it sees any traps at all. This provides 1929 // a little hysteresis, delaying a recompile until a trap happens 1930 // several times. 1931 // 1932 // Actually, since there is only one bit of counter per BCI, 1933 // the possible per-BCI counts are {0,1,(per-method count)}. 1934 // This produces accurate results if in fact there is only 1935 // one hot trap site, but begins to get fuzzy if there are 1936 // many sites. For example, if there are ten sites each 1937 // trapping two or more times, they each get the blame for 1938 // all of their traps. 1939 make_not_entrant = true; 1940 } 1941 1942 // Detect repeated recompilation at the same BCI, and enforce a limit. 1943 if (make_not_entrant && maybe_prior_recompile) { 1944 // More than one recompile at this point. 1945 inc_recompile_count = maybe_prior_trap; 1946 } 1947 } else { 1948 // For reasons which are not recorded per-bytecode, we simply 1949 // force recompiles unconditionally. 1950 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1951 make_not_entrant = true; 1952 } 1953 1954 // Go back to the compiler if there are too many traps in this method. 1955 if (this_trap_count >= per_method_trap_limit(reason)) { 1956 // If there are too many traps in this method, force a recompile. 1957 // This will allow the compiler to see the limit overflow, and 1958 // take corrective action, if possible. 1959 // (This condition is an unlikely backstop only, because the 1960 // PerBytecodeTrapLimit is more likely to take effect first, 1961 // if it is applicable.) 1962 make_not_entrant = true; 1963 } 1964 1965 // Here's more hysteresis: If there has been a recompile at 1966 // this trap point already, run the method in the interpreter 1967 // for a while to exercise it more thoroughly. 1968 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1969 reprofile = true; 1970 } 1971 } 1972 1973 // Take requested actions on the method: 1974 1975 // Recompile 1976 if (make_not_entrant) { 1977 if (!nm->make_not_entrant()) { 1978 return; // the call did not change nmethod's state 1979 } 1980 1981 if (pdata != NULL) { 1982 // Record the recompilation event, if any. 1983 int tstate0 = pdata->trap_state(); 1984 int tstate1 = trap_state_set_recompiled(tstate0, true); 1985 if (tstate1 != tstate0) 1986 pdata->set_trap_state(tstate1); 1987 } 1988 1989 #if INCLUDE_RTM_OPT 1990 // Restart collecting RTM locking abort statistic if the method 1991 // is recompiled for a reason other than RTM state change. 1992 // Assume that in new recompiled code the statistic could be different, 1993 // for example, due to different inlining. 1994 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1995 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 1996 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1997 } 1998 #endif 1999 // For code aging we count traps separately here, using make_not_entrant() 2000 // as a guard against simultaneous deopts in multiple threads. 2001 if (reason == Reason_tenured && trap_mdo != NULL) { 2002 trap_mdo->inc_tenure_traps(); 2003 } 2004 } 2005 2006 if (inc_recompile_count) { 2007 trap_mdo->inc_overflow_recompile_count(); 2008 if ((uint)trap_mdo->overflow_recompile_count() > 2009 (uint)PerBytecodeRecompilationCutoff) { 2010 // Give up on the method containing the bad BCI. 2011 if (trap_method() == nm->method()) { 2012 make_not_compilable = true; 2013 } else { 2014 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 2015 // But give grace to the enclosing nm->method(). 2016 } 2017 } 2018 } 2019 2020 // Reprofile 2021 if (reprofile) { 2022 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2023 } 2024 2025 // Give up compiling 2026 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2027 assert(make_not_entrant, "consistent"); 2028 nm->method()->set_not_compilable(CompLevel_full_optimization); 2029 } 2030 2031 } // Free marked resources 2032 2033 } 2034 JRT_END 2035 2036 ProfileData* 2037 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2038 int trap_bci, 2039 Deoptimization::DeoptReason reason, 2040 bool update_total_trap_count, 2041 #if INCLUDE_JVMCI 2042 bool is_osr, 2043 #endif 2044 Method* compiled_method, 2045 //outputs: 2046 uint& ret_this_trap_count, 2047 bool& ret_maybe_prior_trap, 2048 bool& ret_maybe_prior_recompile) { 2049 bool maybe_prior_trap = false; 2050 bool maybe_prior_recompile = false; 2051 uint this_trap_count = 0; 2052 if (update_total_trap_count) { 2053 uint idx = reason; 2054 #if INCLUDE_JVMCI 2055 if (is_osr) { 2056 idx += Reason_LIMIT; 2057 } 2058 #endif 2059 uint prior_trap_count = trap_mdo->trap_count(idx); 2060 this_trap_count = trap_mdo->inc_trap_count(idx); 2061 2062 // If the runtime cannot find a place to store trap history, 2063 // it is estimated based on the general condition of the method. 2064 // If the method has ever been recompiled, or has ever incurred 2065 // a trap with the present reason , then this BCI is assumed 2066 // (pessimistically) to be the culprit. 2067 maybe_prior_trap = (prior_trap_count != 0); 2068 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2069 } 2070 ProfileData* pdata = NULL; 2071 2072 2073 // For reasons which are recorded per bytecode, we check per-BCI data. 2074 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2075 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2076 if (per_bc_reason != Reason_none) { 2077 // Find the profile data for this BCI. If there isn't one, 2078 // try to allocate one from the MDO's set of spares. 2079 // This will let us detect a repeated trap at this point. 2080 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2081 2082 if (pdata != NULL) { 2083 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2084 if (LogCompilation && xtty != NULL) { 2085 ttyLocker ttyl; 2086 // no more room for speculative traps in this MDO 2087 xtty->elem("speculative_traps_oom"); 2088 } 2089 } 2090 // Query the trap state of this profile datum. 2091 int tstate0 = pdata->trap_state(); 2092 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2093 maybe_prior_trap = false; 2094 if (!trap_state_is_recompiled(tstate0)) 2095 maybe_prior_recompile = false; 2096 2097 // Update the trap state of this profile datum. 2098 int tstate1 = tstate0; 2099 // Record the reason. 2100 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2101 // Store the updated state on the MDO, for next time. 2102 if (tstate1 != tstate0) 2103 pdata->set_trap_state(tstate1); 2104 } else { 2105 if (LogCompilation && xtty != NULL) { 2106 ttyLocker ttyl; 2107 // Missing MDP? Leave a small complaint in the log. 2108 xtty->elem("missing_mdp bci='%d'", trap_bci); 2109 } 2110 } 2111 } 2112 2113 // Return results: 2114 ret_this_trap_count = this_trap_count; 2115 ret_maybe_prior_trap = maybe_prior_trap; 2116 ret_maybe_prior_recompile = maybe_prior_recompile; 2117 return pdata; 2118 } 2119 2120 void 2121 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2122 ResourceMark rm; 2123 // Ignored outputs: 2124 uint ignore_this_trap_count; 2125 bool ignore_maybe_prior_trap; 2126 bool ignore_maybe_prior_recompile; 2127 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2128 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2129 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true); 2130 query_update_method_data(trap_mdo, trap_bci, 2131 (DeoptReason)reason, 2132 update_total_counts, 2133 #if INCLUDE_JVMCI 2134 false, 2135 #endif 2136 NULL, 2137 ignore_this_trap_count, 2138 ignore_maybe_prior_trap, 2139 ignore_maybe_prior_recompile); 2140 } 2141 2142 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2143 if (TraceDeoptimization) { 2144 tty->print("Uncommon trap "); 2145 } 2146 // Still in Java no safepoints 2147 { 2148 // This enters VM and may safepoint 2149 uncommon_trap_inner(thread, trap_request); 2150 } 2151 return fetch_unroll_info_helper(thread, exec_mode); 2152 } 2153 2154 // Local derived constants. 2155 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2156 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 2157 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2158 2159 //---------------------------trap_state_reason--------------------------------- 2160 Deoptimization::DeoptReason 2161 Deoptimization::trap_state_reason(int trap_state) { 2162 // This assert provides the link between the width of DataLayout::trap_bits 2163 // and the encoding of "recorded" reasons. It ensures there are enough 2164 // bits to store all needed reasons in the per-BCI MDO profile. 2165 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2166 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2167 trap_state -= recompile_bit; 2168 if (trap_state == DS_REASON_MASK) { 2169 return Reason_many; 2170 } else { 2171 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2172 return (DeoptReason)trap_state; 2173 } 2174 } 2175 //-------------------------trap_state_has_reason------------------------------- 2176 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2177 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2178 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2179 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2180 trap_state -= recompile_bit; 2181 if (trap_state == DS_REASON_MASK) { 2182 return -1; // true, unspecifically (bottom of state lattice) 2183 } else if (trap_state == reason) { 2184 return 1; // true, definitely 2185 } else if (trap_state == 0) { 2186 return 0; // false, definitely (top of state lattice) 2187 } else { 2188 return 0; // false, definitely 2189 } 2190 } 2191 //-------------------------trap_state_add_reason------------------------------- 2192 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2193 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2194 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2195 trap_state -= recompile_bit; 2196 if (trap_state == DS_REASON_MASK) { 2197 return trap_state + recompile_bit; // already at state lattice bottom 2198 } else if (trap_state == reason) { 2199 return trap_state + recompile_bit; // the condition is already true 2200 } else if (trap_state == 0) { 2201 return reason + recompile_bit; // no condition has yet been true 2202 } else { 2203 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2204 } 2205 } 2206 //-----------------------trap_state_is_recompiled------------------------------ 2207 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2208 return (trap_state & DS_RECOMPILE_BIT) != 0; 2209 } 2210 //-----------------------trap_state_set_recompiled----------------------------- 2211 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2212 if (z) return trap_state | DS_RECOMPILE_BIT; 2213 else return trap_state & ~DS_RECOMPILE_BIT; 2214 } 2215 //---------------------------format_trap_state--------------------------------- 2216 // This is used for debugging and diagnostics, including LogFile output. 2217 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2218 int trap_state) { 2219 assert(buflen > 0, "sanity"); 2220 DeoptReason reason = trap_state_reason(trap_state); 2221 bool recomp_flag = trap_state_is_recompiled(trap_state); 2222 // Re-encode the state from its decoded components. 2223 int decoded_state = 0; 2224 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2225 decoded_state = trap_state_add_reason(decoded_state, reason); 2226 if (recomp_flag) 2227 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2228 // If the state re-encodes properly, format it symbolically. 2229 // Because this routine is used for debugging and diagnostics, 2230 // be robust even if the state is a strange value. 2231 size_t len; 2232 if (decoded_state != trap_state) { 2233 // Random buggy state that doesn't decode?? 2234 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2235 } else { 2236 len = jio_snprintf(buf, buflen, "%s%s", 2237 trap_reason_name(reason), 2238 recomp_flag ? " recompiled" : ""); 2239 } 2240 return buf; 2241 } 2242 2243 2244 //--------------------------------statics-------------------------------------- 2245 const char* Deoptimization::_trap_reason_name[] = { 2246 // Note: Keep this in sync. with enum DeoptReason. 2247 "none", 2248 "null_check", 2249 "null_assert" JVMCI_ONLY("_or_unreached0"), 2250 "range_check", 2251 "class_check", 2252 "array_check", 2253 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2254 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2255 "unloaded", 2256 "uninitialized", 2257 "unreached", 2258 "unhandled", 2259 "constraint", 2260 "div0_check", 2261 "age", 2262 "predicate", 2263 "loop_limit_check", 2264 "speculate_class_check", 2265 "speculate_null_check", 2266 "speculate_null_assert", 2267 "rtm_state_change", 2268 "unstable_if", 2269 "unstable_fused_if", 2270 #if INCLUDE_JVMCI 2271 "aliasing", 2272 "transfer_to_interpreter", 2273 "not_compiled_exception_handler", 2274 "unresolved", 2275 "jsr_mismatch", 2276 #endif 2277 "tenured" 2278 }; 2279 const char* Deoptimization::_trap_action_name[] = { 2280 // Note: Keep this in sync. with enum DeoptAction. 2281 "none", 2282 "maybe_recompile", 2283 "reinterpret", 2284 "make_not_entrant", 2285 "make_not_compilable" 2286 }; 2287 2288 const char* Deoptimization::trap_reason_name(int reason) { 2289 // Check that every reason has a name 2290 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2291 2292 if (reason == Reason_many) return "many"; 2293 if ((uint)reason < Reason_LIMIT) 2294 return _trap_reason_name[reason]; 2295 static char buf[20]; 2296 sprintf(buf, "reason%d", reason); 2297 return buf; 2298 } 2299 const char* Deoptimization::trap_action_name(int action) { 2300 // Check that every action has a name 2301 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2302 2303 if ((uint)action < Action_LIMIT) 2304 return _trap_action_name[action]; 2305 static char buf[20]; 2306 sprintf(buf, "action%d", action); 2307 return buf; 2308 } 2309 2310 // This is used for debugging and diagnostics, including LogFile output. 2311 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2312 int trap_request) { 2313 jint unloaded_class_index = trap_request_index(trap_request); 2314 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2315 const char* action = trap_action_name(trap_request_action(trap_request)); 2316 #if INCLUDE_JVMCI 2317 int debug_id = trap_request_debug_id(trap_request); 2318 #endif 2319 size_t len; 2320 if (unloaded_class_index < 0) { 2321 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2322 reason, action 2323 #if INCLUDE_JVMCI 2324 ,debug_id 2325 #endif 2326 ); 2327 } else { 2328 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2329 reason, action, unloaded_class_index 2330 #if INCLUDE_JVMCI 2331 ,debug_id 2332 #endif 2333 ); 2334 } 2335 return buf; 2336 } 2337 2338 juint Deoptimization::_deoptimization_hist 2339 [Deoptimization::Reason_LIMIT] 2340 [1 + Deoptimization::Action_LIMIT] 2341 [Deoptimization::BC_CASE_LIMIT] 2342 = {0}; 2343 2344 enum { 2345 LSB_BITS = 8, 2346 LSB_MASK = right_n_bits(LSB_BITS) 2347 }; 2348 2349 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2350 Bytecodes::Code bc) { 2351 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2352 assert(action >= 0 && action < Action_LIMIT, "oob"); 2353 _deoptimization_hist[Reason_none][0][0] += 1; // total 2354 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2355 juint* cases = _deoptimization_hist[reason][1+action]; 2356 juint* bc_counter_addr = NULL; 2357 juint bc_counter = 0; 2358 // Look for an unused counter, or an exact match to this BC. 2359 if (bc != Bytecodes::_illegal) { 2360 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2361 juint* counter_addr = &cases[bc_case]; 2362 juint counter = *counter_addr; 2363 if ((counter == 0 && bc_counter_addr == NULL) 2364 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2365 // this counter is either free or is already devoted to this BC 2366 bc_counter_addr = counter_addr; 2367 bc_counter = counter | bc; 2368 } 2369 } 2370 } 2371 if (bc_counter_addr == NULL) { 2372 // Overflow, or no given bytecode. 2373 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2374 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2375 } 2376 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2377 } 2378 2379 jint Deoptimization::total_deoptimization_count() { 2380 return _deoptimization_hist[Reason_none][0][0]; 2381 } 2382 2383 jint Deoptimization::deoptimization_count(DeoptReason reason) { 2384 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2385 return _deoptimization_hist[reason][0][0]; 2386 } 2387 2388 void Deoptimization::print_statistics() { 2389 juint total = total_deoptimization_count(); 2390 juint account = total; 2391 if (total != 0) { 2392 ttyLocker ttyl; 2393 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2394 tty->print_cr("Deoptimization traps recorded:"); 2395 #define PRINT_STAT_LINE(name, r) \ 2396 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2397 PRINT_STAT_LINE("total", total); 2398 // For each non-zero entry in the histogram, print the reason, 2399 // the action, and (if specifically known) the type of bytecode. 2400 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2401 for (int action = 0; action < Action_LIMIT; action++) { 2402 juint* cases = _deoptimization_hist[reason][1+action]; 2403 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2404 juint counter = cases[bc_case]; 2405 if (counter != 0) { 2406 char name[1*K]; 2407 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2408 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2409 bc = Bytecodes::_illegal; 2410 sprintf(name, "%s/%s/%s", 2411 trap_reason_name(reason), 2412 trap_action_name(action), 2413 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2414 juint r = counter >> LSB_BITS; 2415 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2416 account -= r; 2417 } 2418 } 2419 } 2420 } 2421 if (account != 0) { 2422 PRINT_STAT_LINE("unaccounted", account); 2423 } 2424 #undef PRINT_STAT_LINE 2425 if (xtty != NULL) xtty->tail("statistics"); 2426 } 2427 } 2428 #else // COMPILER2_OR_JVMCI 2429 2430 2431 // Stubs for C1 only system. 2432 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2433 return false; 2434 } 2435 2436 const char* Deoptimization::trap_reason_name(int reason) { 2437 return "unknown"; 2438 } 2439 2440 void Deoptimization::print_statistics() { 2441 // no output 2442 } 2443 2444 void 2445 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2446 // no udpate 2447 } 2448 2449 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2450 return 0; 2451 } 2452 2453 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2454 Bytecodes::Code bc) { 2455 // no update 2456 } 2457 2458 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2459 int trap_state) { 2460 jio_snprintf(buf, buflen, "#%d", trap_state); 2461 return buf; 2462 } 2463 2464 #endif // COMPILER2_OR_JVMCI