1 /* 2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "jvm.h" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "code/nmethod.hpp" 32 #include "code/pcDesc.hpp" 33 #include "code/scopeDesc.hpp" 34 #include "interpreter/bytecode.hpp" 35 #include "interpreter/interpreter.hpp" 36 #include "interpreter/oopMapCache.hpp" 37 #include "memory/allocation.inline.hpp" 38 #include "memory/oopFactory.hpp" 39 #include "memory/resourceArea.hpp" 40 #include "memory/universe.hpp" 41 #include "oops/constantPool.hpp" 42 #include "oops/method.hpp" 43 #include "oops/objArrayKlass.hpp" 44 #include "oops/objArrayOop.inline.hpp" 45 #include "oops/oop.inline.hpp" 46 #include "oops/fieldStreams.hpp" 47 #include "oops/typeArrayOop.inline.hpp" 48 #include "oops/verifyOopClosure.hpp" 49 #include "prims/jvmtiThreadState.hpp" 50 #include "runtime/biasedLocking.hpp" 51 #include "runtime/compilationPolicy.hpp" 52 #include "runtime/deoptimization.hpp" 53 #include "runtime/fieldDescriptor.hpp" 54 #include "runtime/fieldDescriptor.inline.hpp" 55 #include "runtime/frame.inline.hpp" 56 #include "runtime/jniHandles.inline.hpp" 57 #include "runtime/handles.inline.hpp" 58 #include "runtime/interfaceSupport.inline.hpp" 59 #include "runtime/safepointVerifiers.hpp" 60 #include "runtime/sharedRuntime.hpp" 61 #include "runtime/signature.hpp" 62 #include "runtime/stubRoutines.hpp" 63 #include "runtime/thread.hpp" 64 #include "runtime/threadSMR.hpp" 65 #include "runtime/vframe.hpp" 66 #include "runtime/vframeArray.hpp" 67 #include "runtime/vframe_hp.hpp" 68 #include "utilities/events.hpp" 69 #include "utilities/preserveException.hpp" 70 #include "utilities/xmlstream.hpp" 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 #if COMPILER2_OR_JVMCI 161 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 162 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk) { 163 bool realloc_failures = false; 164 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 165 166 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 167 168 // The flag return_oop() indicates call sites which return oop 169 // in compiled code. Such sites include java method calls, 170 // runtime calls (for example, used to allocate new objects/arrays 171 // on slow code path) and any other calls generated in compiled code. 172 // It is not guaranteed that we can get such information here only 173 // by analyzing bytecode in deoptimized frames. This is why this flag 174 // is set during method compilation (see Compile::Process_OopMap_Node()). 175 // If the previous frame was popped or if we are dispatching an exception, 176 // we don't have an oop result. 177 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 178 Handle return_value; 179 if (save_oop_result) { 180 // Reallocation may trigger GC. If deoptimization happened on return from 181 // call which returns oop we need to save it since it is not in oopmap. 182 oop result = deoptee.saved_oop_result(&map); 183 assert(oopDesc::is_oop_or_null(result), "must be oop"); 184 return_value = Handle(thread, result); 185 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 186 if (TraceDeoptimization) { 187 ttyLocker ttyl; 188 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 189 } 190 } 191 if (objects != NULL) { 192 JRT_BLOCK 193 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 194 JRT_END 195 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci(); 196 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 197 #ifndef PRODUCT 198 if (TraceDeoptimization) { 199 ttyLocker ttyl; 200 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 201 Deoptimization::print_objects(objects, realloc_failures); 202 } 203 #endif 204 } 205 if (save_oop_result) { 206 // Restore result. 207 deoptee.set_saved_oop_result(&map, return_value()); 208 } 209 return realloc_failures; 210 } 211 212 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 213 #ifndef PRODUCT 214 bool first = true; 215 #endif 216 for (int i = 0; i < chunk->length(); i++) { 217 compiledVFrame* cvf = chunk->at(i); 218 assert (cvf->scope() != NULL,"expect only compiled java frames"); 219 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 220 if (monitors->is_nonempty()) { 221 Deoptimization::relock_objects(monitors, thread, realloc_failures); 222 #ifndef PRODUCT 223 if (PrintDeoptimizationDetails) { 224 ttyLocker ttyl; 225 for (int j = 0; j < monitors->length(); j++) { 226 MonitorInfo* mi = monitors->at(j); 227 if (mi->eliminated()) { 228 if (first) { 229 first = false; 230 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 231 } 232 if (mi->owner_is_scalar_replaced()) { 233 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 234 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 235 } else { 236 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 237 } 238 } 239 } 240 } 241 #endif // !PRODUCT 242 } 243 } 244 } 245 #endif // COMPILER2_OR_JVMCI 246 247 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 248 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 249 250 // Note: there is a safepoint safety issue here. No matter whether we enter 251 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 252 // the vframeArray is created. 253 // 254 255 // Allocate our special deoptimization ResourceMark 256 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 257 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 258 thread->set_deopt_mark(dmark); 259 260 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 261 RegisterMap map(thread, true); 262 RegisterMap dummy_map(thread, false); 263 // Now get the deoptee with a valid map 264 frame deoptee = stub_frame.sender(&map); 265 // Set the deoptee nmethod 266 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 267 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 268 thread->set_deopt_compiled_method(cm); 269 270 if (VerifyStack) { 271 thread->validate_frame_layout(); 272 } 273 274 // Create a growable array of VFrames where each VFrame represents an inlined 275 // Java frame. This storage is allocated with the usual system arena. 276 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 277 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 278 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 279 while (!vf->is_top()) { 280 assert(vf->is_compiled_frame(), "Wrong frame type"); 281 chunk->push(compiledVFrame::cast(vf)); 282 vf = vf->sender(); 283 } 284 assert(vf->is_compiled_frame(), "Wrong frame type"); 285 chunk->push(compiledVFrame::cast(vf)); 286 287 bool realloc_failures = false; 288 289 #if COMPILER2_OR_JVMCI 290 #if INCLUDE_JVMCI 291 bool jvmci_enabled = true; 292 #else 293 bool jvmci_enabled = false; 294 #endif 295 296 // Reallocate the non-escaping objects and restore their fields. Then 297 // relock objects if synchronization on them was eliminated. 298 if (jvmci_enabled || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateAllocations)) { 299 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk); 300 } 301 302 // Revoke biases, done with in java state. 303 // No safepoints allowed after this 304 revoke_from_deopt_handler(thread, deoptee, &map); 305 306 // Ensure that no safepoint is taken after pointers have been stored 307 // in fields of rematerialized objects. If a safepoint occurs from here on 308 // out the java state residing in the vframeArray will be missed. 309 // Locks may be rebaised in a safepoint. 310 NoSafepointVerifier no_safepoint; 311 312 if (jvmci_enabled || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks)) { 313 eliminate_locks(thread, chunk, realloc_failures); 314 } 315 #endif // COMPILER2_OR_JVMCI 316 317 ScopeDesc* trap_scope = chunk->at(0)->scope(); 318 Handle exceptionObject; 319 if (trap_scope->rethrow_exception()) { 320 if (PrintDeoptimizationDetails) { 321 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()); 322 } 323 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 324 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 325 ScopeValue* topOfStack = expressions->top(); 326 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 327 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 328 } 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_deopt_message(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, iframe->interpreter_frame_bci()); 701 int max_bci = mh->code_size(); 702 // Get to the next bytecode if possible 703 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 704 // Check to see if we can grab the number of outgoing arguments 705 // at an uncommon trap for an invoke (where the compiler 706 // generates debug info before the invoke has executed) 707 Bytecodes::Code cur_code = str.next(); 708 if (Bytecodes::is_invoke(cur_code)) { 709 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 710 cur_invoke_parameter_size = invoke.size_of_parameters(); 711 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 712 callee_size_of_parameters++; 713 } 714 } 715 if (str.bci() < max_bci) { 716 Bytecodes::Code next_code = str.next(); 717 if (next_code >= 0) { 718 // The interpreter oop map generator reports results before 719 // the current bytecode has executed except in the case of 720 // calls. It seems to be hard to tell whether the compiler 721 // has emitted debug information matching the "state before" 722 // a given bytecode or the state after, so we try both 723 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 724 // Get expression stack size for the next bytecode 725 InterpreterOopMap next_mask; 726 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 727 next_mask_expression_stack_size = next_mask.expression_stack_size(); 728 if (Bytecodes::is_invoke(next_code)) { 729 Bytecode_invoke invoke(mh, str.bci()); 730 next_mask_expression_stack_size += invoke.size_of_parameters(); 731 } 732 // Need to subtract off the size of the result type of 733 // the bytecode because this is not described in the 734 // debug info but returned to the interpreter in the TOS 735 // caching register 736 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 737 if (bytecode_result_type != T_ILLEGAL) { 738 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 739 } 740 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 741 try_next_mask = true; 742 } 743 } 744 } 745 746 // Verify stack depth and oops in frame 747 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 748 if (!( 749 /* SPARC */ 750 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 751 /* x86 */ 752 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 753 (try_next_mask && 754 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 755 top_frame_expression_stack_adjustment))) || 756 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 757 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 758 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 759 )) { 760 { 761 ttyLocker ttyl; 762 763 // Print out some information that will help us debug the problem 764 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 765 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 766 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 767 iframe->interpreter_frame_expression_stack_size()); 768 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 769 tty->print_cr(" try_next_mask = %d", try_next_mask); 770 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 771 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 772 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 773 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 774 tty->print_cr(" exec_mode = %d", exec_mode); 775 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 776 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 777 tty->print_cr(" Interpreted frames:"); 778 for (int k = 0; k < cur_array->frames(); k++) { 779 vframeArrayElement* el = cur_array->element(k); 780 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 781 } 782 cur_array->print_on_2(tty); 783 } // release tty lock before calling guarantee 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 class DeoptimizeMarkedTC : public ThreadClosure { 800 public: 801 virtual void do_thread(Thread* thread) { 802 assert(thread->is_Java_thread(), "must be"); 803 JavaThread* jt = (JavaThread*)thread; 804 jt->deoptimize_marked_methods(); 805 } 806 }; 807 808 void Deoptimization::deoptimize_all_marked() { 809 ResourceMark rm; 810 DeoptimizationMarker dm; 811 812 if (SafepointSynchronize::is_at_safepoint()) { 813 DeoptimizeMarkedTC deopt; 814 // Make the dependent methods not entrant 815 CodeCache::make_marked_nmethods_not_entrant(); 816 Threads::java_threads_do(&deopt); 817 } else { 818 // Make the dependent methods not entrant 819 { 820 MutexLocker mu(CodeCache_lock, Mutex::_no_safepoint_check_flag); 821 CodeCache::make_marked_nmethods_not_entrant(); 822 } 823 DeoptimizeMarkedTC deopt; 824 Handshake::execute(&deopt); 825 } 826 } 827 828 Deoptimization::DeoptAction Deoptimization::_unloaded_action 829 = Deoptimization::Action_reinterpret; 830 831 832 833 #if INCLUDE_JVMCI || INCLUDE_AOT 834 template<typename CacheType> 835 class BoxCacheBase : public CHeapObj<mtCompiler> { 836 protected: 837 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { 838 ResourceMark rm; 839 char* klass_name_str = klass_name->as_C_string(); 840 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); 841 guarantee(k != NULL, "%s must be loaded", klass_name_str); 842 InstanceKlass* ik = InstanceKlass::cast(k); 843 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 844 CacheType::compute_offsets(ik); 845 return ik; 846 } 847 }; 848 849 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 850 PrimitiveType _low; 851 PrimitiveType _high; 852 jobject _cache; 853 protected: 854 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 855 BoxCache(Thread* thread) { 856 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); 857 objArrayOop cache = CacheType::cache(ik); 858 assert(cache->length() > 0, "Empty cache"); 859 _low = BoxType::value(cache->obj_at(0)); 860 _high = _low + cache->length() - 1; 861 _cache = JNIHandles::make_global(Handle(thread, cache)); 862 } 863 ~BoxCache() { 864 JNIHandles::destroy_global(_cache); 865 } 866 public: 867 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 868 if (_singleton == NULL) { 869 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 870 if (!Atomic::replace_if_null(s, &_singleton)) { 871 delete s; 872 } 873 } 874 return _singleton; 875 } 876 oop lookup(PrimitiveType value) { 877 if (_low <= value && value <= _high) { 878 int offset = value - _low; 879 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 880 } 881 return NULL; 882 } 883 oop lookup_raw(intptr_t raw_value) { 884 // Have to cast to avoid little/big-endian problems. 885 if (sizeof(PrimitiveType) > sizeof(jint)) { 886 jlong value = (jlong)raw_value; 887 return lookup(value); 888 } 889 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 890 return lookup(value); 891 } 892 }; 893 894 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 895 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 896 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 897 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 898 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 899 900 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 901 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 902 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 903 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 904 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 905 906 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 907 jobject _true_cache; 908 jobject _false_cache; 909 protected: 910 static BooleanBoxCache *_singleton; 911 BooleanBoxCache(Thread *thread) { 912 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); 913 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 914 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 915 } 916 ~BooleanBoxCache() { 917 JNIHandles::destroy_global(_true_cache); 918 JNIHandles::destroy_global(_false_cache); 919 } 920 public: 921 static BooleanBoxCache* singleton(Thread* thread) { 922 if (_singleton == NULL) { 923 BooleanBoxCache* s = new BooleanBoxCache(thread); 924 if (!Atomic::replace_if_null(s, &_singleton)) { 925 delete s; 926 } 927 } 928 return _singleton; 929 } 930 oop lookup_raw(intptr_t raw_value) { 931 // Have to cast to avoid little/big-endian problems. 932 jboolean value = (jboolean)*((jint*)&raw_value); 933 return lookup(value); 934 } 935 oop lookup(jboolean value) { 936 if (value != 0) { 937 return JNIHandles::resolve_non_null(_true_cache); 938 } 939 return JNIHandles::resolve_non_null(_false_cache); 940 } 941 }; 942 943 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 944 945 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { 946 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 947 BasicType box_type = SystemDictionary::box_klass_type(k); 948 if (box_type != T_OBJECT) { 949 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 950 switch(box_type) { 951 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 952 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 953 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 954 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 955 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 956 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 957 default:; 958 } 959 } 960 return NULL; 961 } 962 #endif // INCLUDE_JVMCI || INCLUDE_AOT 963 964 #if COMPILER2_OR_JVMCI 965 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 966 Handle pending_exception(THREAD, thread->pending_exception()); 967 const char* exception_file = thread->exception_file(); 968 int exception_line = thread->exception_line(); 969 thread->clear_pending_exception(); 970 971 bool failures = false; 972 973 for (int i = 0; i < objects->length(); i++) { 974 assert(objects->at(i)->is_object(), "invalid debug information"); 975 ObjectValue* sv = (ObjectValue*) objects->at(i); 976 977 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 978 oop obj = NULL; 979 980 if (k->is_instance_klass()) { 981 #if INCLUDE_JVMCI || INCLUDE_AOT 982 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 983 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 984 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 985 obj = get_cached_box(abv, fr, reg_map, THREAD); 986 if (obj != NULL) { 987 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 988 abv->set_cached(true); 989 } 990 } 991 #endif // INCLUDE_JVMCI || INCLUDE_AOT 992 InstanceKlass* ik = InstanceKlass::cast(k); 993 if (obj == NULL) { 994 obj = ik->allocate_instance(THREAD); 995 } 996 } else if (k->is_typeArray_klass()) { 997 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 998 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 999 int len = sv->field_size() / type2size[ak->element_type()]; 1000 obj = ak->allocate(len, THREAD); 1001 } else if (k->is_objArray_klass()) { 1002 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1003 obj = ak->allocate(sv->field_size(), THREAD); 1004 } 1005 1006 if (obj == NULL) { 1007 failures = true; 1008 } 1009 1010 assert(sv->value().is_null(), "redundant reallocation"); 1011 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1012 CLEAR_PENDING_EXCEPTION; 1013 sv->set_value(obj); 1014 } 1015 1016 if (failures) { 1017 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1018 } else if (pending_exception.not_null()) { 1019 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1020 } 1021 1022 return failures; 1023 } 1024 1025 // restore elements of an eliminated type array 1026 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1027 int index = 0; 1028 intptr_t val; 1029 1030 for (int i = 0; i < sv->field_size(); i++) { 1031 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1032 switch(type) { 1033 case T_LONG: case T_DOUBLE: { 1034 assert(value->type() == T_INT, "Agreement."); 1035 StackValue* low = 1036 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1037 #ifdef _LP64 1038 jlong res = (jlong)low->get_int(); 1039 #else 1040 #ifdef SPARC 1041 // For SPARC we have to swap high and low words. 1042 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1043 #else 1044 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1045 #endif //SPARC 1046 #endif 1047 obj->long_at_put(index, res); 1048 break; 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 < sv->field_size() && type == T_INT) { 1056 if (sv->field_at(i)->is_location()) { 1057 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1058 if (type == Location::dbl || type == Location::lng) { 1059 big_value = true; 1060 } 1061 } else if (sv->field_at(i)->is_constant_int()) { 1062 ScopeValue* next_scope_field = sv->field_at(i + 1); 1063 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1064 big_value = true; 1065 } 1066 } 1067 } 1068 1069 if (big_value) { 1070 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1071 #ifdef _LP64 1072 jlong res = (jlong)low->get_int(); 1073 #else 1074 #ifdef SPARC 1075 // For SPARC we have to swap high and low words. 1076 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1077 #else 1078 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1079 #endif //SPARC 1080 #endif 1081 obj->int_at_put(index, (jint)*((jint*)&res)); 1082 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1083 } else { 1084 val = value->get_int(); 1085 obj->int_at_put(index, (jint)*((jint*)&val)); 1086 } 1087 break; 1088 } 1089 1090 case T_SHORT: 1091 assert(value->type() == T_INT, "Agreement."); 1092 val = value->get_int(); 1093 obj->short_at_put(index, (jshort)*((jint*)&val)); 1094 break; 1095 1096 case T_CHAR: 1097 assert(value->type() == T_INT, "Agreement."); 1098 val = value->get_int(); 1099 obj->char_at_put(index, (jchar)*((jint*)&val)); 1100 break; 1101 1102 case T_BYTE: 1103 assert(value->type() == T_INT, "Agreement."); 1104 val = value->get_int(); 1105 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1106 break; 1107 1108 case T_BOOLEAN: 1109 assert(value->type() == T_INT, "Agreement."); 1110 val = value->get_int(); 1111 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1112 break; 1113 1114 default: 1115 ShouldNotReachHere(); 1116 } 1117 index++; 1118 } 1119 } 1120 1121 1122 // restore fields of an eliminated object array 1123 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1124 for (int i = 0; i < sv->field_size(); i++) { 1125 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1126 assert(value->type() == T_OBJECT, "object element expected"); 1127 obj->obj_at_put(i, value->get_obj()()); 1128 } 1129 } 1130 1131 class ReassignedField { 1132 public: 1133 int _offset; 1134 BasicType _type; 1135 public: 1136 ReassignedField() { 1137 _offset = 0; 1138 _type = T_ILLEGAL; 1139 } 1140 }; 1141 1142 int compare(ReassignedField* left, ReassignedField* right) { 1143 return left->_offset - right->_offset; 1144 } 1145 1146 // Restore fields of an eliminated instance object using the same field order 1147 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1148 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1149 if (klass->superklass() != NULL) { 1150 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); 1151 } 1152 1153 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1154 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1155 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1156 ReassignedField field; 1157 field._offset = fs.offset(); 1158 field._type = FieldType::basic_type(fs.signature()); 1159 fields->append(field); 1160 } 1161 } 1162 fields->sort(compare); 1163 for (int i = 0; i < fields->length(); i++) { 1164 intptr_t val; 1165 ScopeValue* scope_field = sv->field_at(svIndex); 1166 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1167 int offset = fields->at(i)._offset; 1168 BasicType type = fields->at(i)._type; 1169 switch (type) { 1170 case T_OBJECT: case T_ARRAY: 1171 assert(value->type() == T_OBJECT, "Agreement."); 1172 obj->obj_field_put(offset, value->get_obj()()); 1173 break; 1174 1175 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1176 case T_INT: case T_FLOAT: { // 4 bytes. 1177 assert(value->type() == T_INT, "Agreement."); 1178 bool big_value = false; 1179 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1180 if (scope_field->is_location()) { 1181 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1182 if (type == Location::dbl || type == Location::lng) { 1183 big_value = true; 1184 } 1185 } 1186 if (scope_field->is_constant_int()) { 1187 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1188 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1189 big_value = true; 1190 } 1191 } 1192 } 1193 1194 if (big_value) { 1195 i++; 1196 assert(i < fields->length(), "second T_INT field needed"); 1197 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1198 } else { 1199 val = value->get_int(); 1200 obj->int_field_put(offset, (jint)*((jint*)&val)); 1201 break; 1202 } 1203 } 1204 /* no break */ 1205 1206 case T_LONG: case T_DOUBLE: { 1207 assert(value->type() == T_INT, "Agreement."); 1208 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1209 #ifdef _LP64 1210 jlong res = (jlong)low->get_int(); 1211 #else 1212 #ifdef SPARC 1213 // For SPARC we have to swap high and low words. 1214 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1215 #else 1216 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1217 #endif //SPARC 1218 #endif 1219 obj->long_field_put(offset, res); 1220 break; 1221 } 1222 1223 case T_SHORT: 1224 assert(value->type() == T_INT, "Agreement."); 1225 val = value->get_int(); 1226 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1227 break; 1228 1229 case T_CHAR: 1230 assert(value->type() == T_INT, "Agreement."); 1231 val = value->get_int(); 1232 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1233 break; 1234 1235 case T_BYTE: 1236 assert(value->type() == T_INT, "Agreement."); 1237 val = value->get_int(); 1238 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1239 break; 1240 1241 case T_BOOLEAN: 1242 assert(value->type() == T_INT, "Agreement."); 1243 val = value->get_int(); 1244 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1245 break; 1246 1247 default: 1248 ShouldNotReachHere(); 1249 } 1250 svIndex++; 1251 } 1252 return svIndex; 1253 } 1254 1255 // restore fields of all eliminated objects and arrays 1256 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1257 for (int i = 0; i < objects->length(); i++) { 1258 ObjectValue* sv = (ObjectValue*) objects->at(i); 1259 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1260 Handle obj = sv->value(); 1261 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1262 if (PrintDeoptimizationDetails) { 1263 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1264 } 1265 if (obj.is_null()) { 1266 continue; 1267 } 1268 #if INCLUDE_JVMCI || INCLUDE_AOT 1269 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1270 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1271 continue; 1272 } 1273 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1274 if (k->is_instance_klass()) { 1275 InstanceKlass* ik = InstanceKlass::cast(k); 1276 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1277 } else if (k->is_typeArray_klass()) { 1278 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1279 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1280 } else if (k->is_objArray_klass()) { 1281 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1282 } 1283 } 1284 } 1285 1286 1287 // relock objects for which synchronization was eliminated 1288 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1289 for (int i = 0; i < monitors->length(); i++) { 1290 MonitorInfo* mon_info = monitors->at(i); 1291 if (mon_info->eliminated()) { 1292 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1293 if (!mon_info->owner_is_scalar_replaced()) { 1294 Handle obj(thread, mon_info->owner()); 1295 markWord mark = obj->mark(); 1296 if (UseBiasedLocking && mark.has_bias_pattern()) { 1297 // New allocated objects may have the mark set to anonymously biased. 1298 // Also the deoptimized method may called methods with synchronization 1299 // where the thread-local object is bias locked to the current thread. 1300 assert(mark.is_biased_anonymously() || 1301 mark.biased_locker() == thread, "should be locked to current thread"); 1302 // Reset mark word to unbiased prototype. 1303 markWord unbiased_prototype = markWord::prototype().set_age(mark.age()); 1304 obj->set_mark(unbiased_prototype); 1305 } 1306 BasicLock* lock = mon_info->lock(); 1307 ObjectSynchronizer::enter(obj, lock, thread); 1308 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1309 } 1310 } 1311 } 1312 } 1313 1314 1315 #ifndef PRODUCT 1316 // print information about reallocated objects 1317 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1318 fieldDescriptor fd; 1319 1320 for (int i = 0; i < objects->length(); i++) { 1321 ObjectValue* sv = (ObjectValue*) objects->at(i); 1322 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1323 Handle obj = sv->value(); 1324 1325 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1326 k->print_value(); 1327 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1328 if (obj.is_null()) { 1329 tty->print(" allocation failed"); 1330 } else { 1331 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1332 } 1333 tty->cr(); 1334 1335 if (Verbose && !obj.is_null()) { 1336 k->oop_print_on(obj(), tty); 1337 } 1338 } 1339 } 1340 #endif 1341 #endif // COMPILER2_OR_JVMCI 1342 1343 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1344 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1345 1346 #ifndef PRODUCT 1347 if (PrintDeoptimizationDetails) { 1348 ttyLocker ttyl; 1349 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1350 fr.print_on(tty); 1351 tty->print_cr(" Virtual frames (innermost first):"); 1352 for (int index = 0; index < chunk->length(); index++) { 1353 compiledVFrame* vf = chunk->at(index); 1354 tty->print(" %2d - ", index); 1355 vf->print_value(); 1356 int bci = chunk->at(index)->raw_bci(); 1357 const char* code_name; 1358 if (bci == SynchronizationEntryBCI) { 1359 code_name = "sync entry"; 1360 } else { 1361 Bytecodes::Code code = vf->method()->code_at(bci); 1362 code_name = Bytecodes::name(code); 1363 } 1364 tty->print(" - %s", code_name); 1365 tty->print_cr(" @ bci %d ", bci); 1366 if (Verbose) { 1367 vf->print(); 1368 tty->cr(); 1369 } 1370 } 1371 } 1372 #endif 1373 1374 // Register map for next frame (used for stack crawl). We capture 1375 // the state of the deopt'ing frame's caller. Thus if we need to 1376 // stuff a C2I adapter we can properly fill in the callee-save 1377 // register locations. 1378 frame caller = fr.sender(reg_map); 1379 int frame_size = caller.sp() - fr.sp(); 1380 1381 frame sender = caller; 1382 1383 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1384 // the vframeArray containing the unpacking information is allocated in the C heap. 1385 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1386 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1387 1388 // Compare the vframeArray to the collected vframes 1389 assert(array->structural_compare(thread, chunk), "just checking"); 1390 1391 #ifndef PRODUCT 1392 if (PrintDeoptimizationDetails) { 1393 ttyLocker ttyl; 1394 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1395 } 1396 #endif // PRODUCT 1397 1398 return array; 1399 } 1400 1401 #if COMPILER2_OR_JVMCI 1402 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1403 // Reallocation of some scalar replaced objects failed. Record 1404 // that we need to pop all the interpreter frames for the 1405 // deoptimized compiled frame. 1406 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1407 thread->set_frames_to_pop_failed_realloc(array->frames()); 1408 // Unlock all monitors here otherwise the interpreter will see a 1409 // mix of locked and unlocked monitors (because of failed 1410 // reallocations of synchronized objects) and be confused. 1411 for (int i = 0; i < array->frames(); i++) { 1412 MonitorChunk* monitors = array->element(i)->monitors(); 1413 if (monitors != NULL) { 1414 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1415 BasicObjectLock* src = monitors->at(j); 1416 if (src->obj() != NULL) { 1417 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1418 } 1419 } 1420 array->element(i)->free_monitors(thread); 1421 #ifdef ASSERT 1422 array->element(i)->set_removed_monitors(); 1423 #endif 1424 } 1425 } 1426 } 1427 #endif 1428 1429 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1430 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1431 Thread* thread = Thread::current(); 1432 for (int i = 0; i < monitors->length(); i++) { 1433 MonitorInfo* mon_info = monitors->at(i); 1434 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1435 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1436 } 1437 } 1438 } 1439 1440 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) { 1441 // Unfortunately we don't have a RegisterMap available in most of 1442 // the places we want to call this routine so we need to walk the 1443 // stack again to update the register map. 1444 if (map == NULL || !map->update_map()) { 1445 StackFrameStream sfs(thread, true); 1446 bool found = false; 1447 while (!found && !sfs.is_done()) { 1448 frame* cur = sfs.current(); 1449 sfs.next(); 1450 found = cur->id() == fr.id(); 1451 } 1452 assert(found, "frame to be deoptimized not found on target thread's stack"); 1453 map = sfs.register_map(); 1454 } 1455 1456 vframe* vf = vframe::new_vframe(&fr, map, thread); 1457 compiledVFrame* cvf = compiledVFrame::cast(vf); 1458 // Revoke monitors' biases in all scopes 1459 while (!cvf->is_top()) { 1460 collect_monitors(cvf, objects_to_revoke); 1461 cvf = compiledVFrame::cast(cvf->sender()); 1462 } 1463 collect_monitors(cvf, objects_to_revoke); 1464 } 1465 1466 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) { 1467 if (!UseBiasedLocking) { 1468 return; 1469 } 1470 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1471 get_monitors_from_stack(objects_to_revoke, thread, fr, map); 1472 1473 int len = objects_to_revoke->length(); 1474 for (int i = 0; i < len; i++) { 1475 oop obj = (objects_to_revoke->at(i))(); 1476 BiasedLocking::revoke_own_locks(objects_to_revoke->at(i), thread); 1477 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1478 } 1479 } 1480 1481 1482 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1483 assert(fr.can_be_deoptimized(), "checking frame type"); 1484 1485 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1486 1487 if (LogCompilation && xtty != NULL) { 1488 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1489 assert(cm != NULL, "only compiled methods can deopt"); 1490 1491 ttyLocker ttyl; 1492 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1493 cm->log_identity(xtty); 1494 xtty->end_head(); 1495 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1496 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1497 xtty->method(sd->method()); 1498 xtty->end_elem(); 1499 if (sd->is_top()) break; 1500 } 1501 xtty->tail("deoptimized"); 1502 } 1503 1504 // Patch the compiled method so that when execution returns to it we will 1505 // deopt the execution state and return to the interpreter. 1506 fr.deoptimize(thread); 1507 } 1508 1509 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1510 // Deoptimize only if the frame comes from compile code. 1511 // Do not deoptimize the frame which is already patched 1512 // during the execution of the loops below. 1513 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1514 return; 1515 } 1516 ResourceMark rm; 1517 DeoptimizationMarker dm; 1518 deoptimize_single_frame(thread, fr, reason); 1519 } 1520 1521 #if INCLUDE_JVMCI 1522 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1523 // there is no exception handler for this pc => deoptimize 1524 cm->make_not_entrant(); 1525 1526 // Use Deoptimization::deoptimize for all of its side-effects: 1527 // revoking biases of monitors, gathering traps statistics, logging... 1528 // it also patches the return pc but we do not care about that 1529 // since we return a continuation to the deopt_blob below. 1530 JavaThread* thread = JavaThread::current(); 1531 RegisterMap reg_map(thread, UseBiasedLocking); 1532 frame runtime_frame = thread->last_frame(); 1533 frame caller_frame = runtime_frame.sender(®_map); 1534 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1535 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1536 1537 MethodData* trap_mdo = get_method_data(thread, cm->method(), true); 1538 if (trap_mdo != NULL) { 1539 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1540 } 1541 1542 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1543 } 1544 #endif 1545 1546 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1547 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1548 "can only deoptimize other thread at a safepoint"); 1549 // Compute frame and register map based on thread and sp. 1550 RegisterMap reg_map(thread, UseBiasedLocking); 1551 frame fr = thread->last_frame(); 1552 while (fr.id() != id) { 1553 fr = fr.sender(®_map); 1554 } 1555 deoptimize(thread, fr, ®_map, reason); 1556 } 1557 1558 1559 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1560 if (thread == Thread::current()) { 1561 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1562 } else { 1563 VM_DeoptimizeFrame deopt(thread, id, reason); 1564 VMThread::execute(&deopt); 1565 } 1566 } 1567 1568 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1569 deoptimize_frame(thread, id, Reason_constraint); 1570 } 1571 1572 // JVMTI PopFrame support 1573 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1574 { 1575 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1576 } 1577 JRT_END 1578 1579 MethodData* 1580 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1581 bool create_if_missing) { 1582 Thread* THREAD = thread; 1583 MethodData* mdo = m()->method_data(); 1584 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1585 // Build an MDO. Ignore errors like OutOfMemory; 1586 // that simply means we won't have an MDO to update. 1587 Method::build_interpreter_method_data(m, THREAD); 1588 if (HAS_PENDING_EXCEPTION) { 1589 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1590 CLEAR_PENDING_EXCEPTION; 1591 } 1592 mdo = m()->method_data(); 1593 } 1594 return mdo; 1595 } 1596 1597 #if COMPILER2_OR_JVMCI 1598 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1599 // in case of an unresolved klass entry, load the class. 1600 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1601 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1602 return; 1603 } 1604 1605 if (!constant_pool->tag_at(index).is_symbol()) return; 1606 1607 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1608 Symbol* symbol = constant_pool->symbol_at(index); 1609 1610 // class name? 1611 if (symbol->char_at(0) != '(') { 1612 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1613 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1614 return; 1615 } 1616 1617 // then it must be a signature! 1618 ResourceMark rm(THREAD); 1619 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1620 if (ss.is_object()) { 1621 Symbol* class_name = ss.as_symbol(); 1622 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1623 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1624 } 1625 } 1626 } 1627 1628 1629 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1630 EXCEPTION_MARK; 1631 load_class_by_index(constant_pool, index, THREAD); 1632 if (HAS_PENDING_EXCEPTION) { 1633 // Exception happened during classloading. We ignore the exception here, since it 1634 // is going to be rethrown since the current activation is going to be deoptimized and 1635 // the interpreter will re-execute the bytecode. 1636 CLEAR_PENDING_EXCEPTION; 1637 // Class loading called java code which may have caused a stack 1638 // overflow. If the exception was thrown right before the return 1639 // to the runtime the stack is no longer guarded. Reguard the 1640 // stack otherwise if we return to the uncommon trap blob and the 1641 // stack bang causes a stack overflow we crash. 1642 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1643 JavaThread* thread = (JavaThread*)THREAD; 1644 bool guard_pages_enabled = thread->stack_guards_enabled(); 1645 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1646 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1647 } 1648 } 1649 1650 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1651 HandleMark hm; 1652 1653 // uncommon_trap() is called at the beginning of the uncommon trap 1654 // handler. Note this fact before we start generating temporary frames 1655 // that can confuse an asynchronous stack walker. This counter is 1656 // decremented at the end of unpack_frames(). 1657 thread->inc_in_deopt_handler(); 1658 1659 // We need to update the map if we have biased locking. 1660 #if INCLUDE_JVMCI 1661 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1662 RegisterMap reg_map(thread, true); 1663 #else 1664 RegisterMap reg_map(thread, UseBiasedLocking); 1665 #endif 1666 frame stub_frame = thread->last_frame(); 1667 frame fr = stub_frame.sender(®_map); 1668 // Make sure the calling nmethod is not getting deoptimized and removed 1669 // before we are done with it. 1670 nmethodLocker nl(fr.pc()); 1671 1672 // Log a message 1673 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1674 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1675 1676 { 1677 ResourceMark rm; 1678 1679 DeoptReason reason = trap_request_reason(trap_request); 1680 DeoptAction action = trap_request_action(trap_request); 1681 #if INCLUDE_JVMCI 1682 int debug_id = trap_request_debug_id(trap_request); 1683 #endif 1684 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1685 1686 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1687 compiledVFrame* cvf = compiledVFrame::cast(vf); 1688 1689 CompiledMethod* nm = cvf->code(); 1690 1691 ScopeDesc* trap_scope = cvf->scope(); 1692 1693 if (TraceDeoptimization) { 1694 ttyLocker ttyl; 1695 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() 1696 #if INCLUDE_JVMCI 1697 , debug_id 1698 #endif 1699 ); 1700 } 1701 1702 methodHandle trap_method = trap_scope->method(); 1703 int trap_bci = trap_scope->bci(); 1704 #if INCLUDE_JVMCI 1705 jlong speculation = thread->pending_failed_speculation(); 1706 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods 1707 nm->as_nmethod()->update_speculation(thread); 1708 } else { 1709 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 1710 } 1711 1712 if (trap_bci == SynchronizationEntryBCI) { 1713 trap_bci = 0; 1714 thread->set_pending_monitorenter(true); 1715 } 1716 1717 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1718 thread->set_pending_transfer_to_interpreter(true); 1719 } 1720 #endif 1721 1722 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1723 // Record this event in the histogram. 1724 gather_statistics(reason, action, trap_bc); 1725 1726 // Ensure that we can record deopt. history: 1727 // Need MDO to record RTM code generation state. 1728 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1729 1730 methodHandle profiled_method; 1731 #if INCLUDE_JVMCI 1732 if (nm->is_compiled_by_jvmci()) { 1733 profiled_method = nm->method(); 1734 } else { 1735 profiled_method = trap_method; 1736 } 1737 #else 1738 profiled_method = trap_method; 1739 #endif 1740 1741 MethodData* trap_mdo = 1742 get_method_data(thread, profiled_method, create_if_missing); 1743 1744 // Log a message 1745 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1746 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1747 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1748 1749 // Print a bunch of diagnostics, if requested. 1750 if (TraceDeoptimization || LogCompilation) { 1751 ResourceMark rm; 1752 ttyLocker ttyl; 1753 char buf[100]; 1754 if (xtty != NULL) { 1755 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1756 os::current_thread_id(), 1757 format_trap_request(buf, sizeof(buf), trap_request)); 1758 #if INCLUDE_JVMCI 1759 if (speculation != 0) { 1760 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 1761 } 1762 #endif 1763 nm->log_identity(xtty); 1764 } 1765 Symbol* class_name = NULL; 1766 bool unresolved = false; 1767 if (unloaded_class_index >= 0) { 1768 constantPoolHandle constants (THREAD, trap_method->constants()); 1769 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1770 class_name = constants->klass_name_at(unloaded_class_index); 1771 unresolved = true; 1772 if (xtty != NULL) 1773 xtty->print(" unresolved='1'"); 1774 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1775 class_name = constants->symbol_at(unloaded_class_index); 1776 } 1777 if (xtty != NULL) 1778 xtty->name(class_name); 1779 } 1780 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1781 // Dump the relevant MDO state. 1782 // This is the deopt count for the current reason, any previous 1783 // reasons or recompiles seen at this point. 1784 int dcnt = trap_mdo->trap_count(reason); 1785 if (dcnt != 0) 1786 xtty->print(" count='%d'", dcnt); 1787 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1788 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1789 if (dos != 0) { 1790 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1791 if (trap_state_is_recompiled(dos)) { 1792 int recnt2 = trap_mdo->overflow_recompile_count(); 1793 if (recnt2 != 0) 1794 xtty->print(" recompiles2='%d'", recnt2); 1795 } 1796 } 1797 } 1798 if (xtty != NULL) { 1799 xtty->stamp(); 1800 xtty->end_head(); 1801 } 1802 if (TraceDeoptimization) { // make noise on the tty 1803 tty->print("Uncommon trap occurred in"); 1804 nm->method()->print_short_name(tty); 1805 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1806 #if INCLUDE_JVMCI 1807 if (nm->is_nmethod()) { 1808 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 1809 if (installed_code_name != NULL) { 1810 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1811 } 1812 } 1813 #endif 1814 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1815 p2i(fr.pc()), 1816 os::current_thread_id(), 1817 trap_reason_name(reason), 1818 trap_action_name(action), 1819 unloaded_class_index 1820 #if INCLUDE_JVMCI 1821 , debug_id 1822 #endif 1823 ); 1824 if (class_name != NULL) { 1825 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1826 class_name->print_symbol_on(tty); 1827 } 1828 tty->cr(); 1829 } 1830 if (xtty != NULL) { 1831 // Log the precise location of the trap. 1832 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1833 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1834 xtty->method(sd->method()); 1835 xtty->end_elem(); 1836 if (sd->is_top()) break; 1837 } 1838 xtty->tail("uncommon_trap"); 1839 } 1840 } 1841 // (End diagnostic printout.) 1842 1843 // Load class if necessary 1844 if (unloaded_class_index >= 0) { 1845 constantPoolHandle constants(THREAD, trap_method->constants()); 1846 load_class_by_index(constants, unloaded_class_index); 1847 } 1848 1849 // Flush the nmethod if necessary and desirable. 1850 // 1851 // We need to avoid situations where we are re-flushing the nmethod 1852 // because of a hot deoptimization site. Repeated flushes at the same 1853 // point need to be detected by the compiler and avoided. If the compiler 1854 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1855 // module must take measures to avoid an infinite cycle of recompilation 1856 // and deoptimization. There are several such measures: 1857 // 1858 // 1. If a recompilation is ordered a second time at some site X 1859 // and for the same reason R, the action is adjusted to 'reinterpret', 1860 // to give the interpreter time to exercise the method more thoroughly. 1861 // If this happens, the method's overflow_recompile_count is incremented. 1862 // 1863 // 2. If the compiler fails to reduce the deoptimization rate, then 1864 // the method's overflow_recompile_count will begin to exceed the set 1865 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1866 // is adjusted to 'make_not_compilable', and the method is abandoned 1867 // to the interpreter. This is a performance hit for hot methods, 1868 // but is better than a disastrous infinite cycle of recompilations. 1869 // (Actually, only the method containing the site X is abandoned.) 1870 // 1871 // 3. In parallel with the previous measures, if the total number of 1872 // recompilations of a method exceeds the much larger set limit 1873 // PerMethodRecompilationCutoff, the method is abandoned. 1874 // This should only happen if the method is very large and has 1875 // many "lukewarm" deoptimizations. The code which enforces this 1876 // limit is elsewhere (class nmethod, class Method). 1877 // 1878 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1879 // to recompile at each bytecode independently of the per-BCI cutoff. 1880 // 1881 // The decision to update code is up to the compiler, and is encoded 1882 // in the Action_xxx code. If the compiler requests Action_none 1883 // no trap state is changed, no compiled code is changed, and the 1884 // computation suffers along in the interpreter. 1885 // 1886 // The other action codes specify various tactics for decompilation 1887 // and recompilation. Action_maybe_recompile is the loosest, and 1888 // allows the compiled code to stay around until enough traps are seen, 1889 // and until the compiler gets around to recompiling the trapping method. 1890 // 1891 // The other actions cause immediate removal of the present code. 1892 1893 // Traps caused by injected profile shouldn't pollute trap counts. 1894 bool injected_profile_trap = trap_method->has_injected_profile() && 1895 (reason == Reason_intrinsic || reason == Reason_unreached); 1896 1897 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1898 bool make_not_entrant = false; 1899 bool make_not_compilable = false; 1900 bool reprofile = false; 1901 switch (action) { 1902 case Action_none: 1903 // Keep the old code. 1904 update_trap_state = false; 1905 break; 1906 case Action_maybe_recompile: 1907 // Do not need to invalidate the present code, but we can 1908 // initiate another 1909 // Start compiler without (necessarily) invalidating the nmethod. 1910 // The system will tolerate the old code, but new code should be 1911 // generated when possible. 1912 break; 1913 case Action_reinterpret: 1914 // Go back into the interpreter for a while, and then consider 1915 // recompiling form scratch. 1916 make_not_entrant = true; 1917 // Reset invocation counter for outer most method. 1918 // This will allow the interpreter to exercise the bytecodes 1919 // for a while before recompiling. 1920 // By contrast, Action_make_not_entrant is immediate. 1921 // 1922 // Note that the compiler will track null_check, null_assert, 1923 // range_check, and class_check events and log them as if they 1924 // had been traps taken from compiled code. This will update 1925 // the MDO trap history so that the next compilation will 1926 // properly detect hot trap sites. 1927 reprofile = true; 1928 break; 1929 case Action_make_not_entrant: 1930 // Request immediate recompilation, and get rid of the old code. 1931 // Make them not entrant, so next time they are called they get 1932 // recompiled. Unloaded classes are loaded now so recompile before next 1933 // time they are called. Same for uninitialized. The interpreter will 1934 // link the missing class, if any. 1935 make_not_entrant = true; 1936 break; 1937 case Action_make_not_compilable: 1938 // Give up on compiling this method at all. 1939 make_not_entrant = true; 1940 make_not_compilable = true; 1941 break; 1942 default: 1943 ShouldNotReachHere(); 1944 } 1945 1946 // Setting +ProfileTraps fixes the following, on all platforms: 1947 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1948 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1949 // recompile relies on a MethodData* to record heroic opt failures. 1950 1951 // Whether the interpreter is producing MDO data or not, we also need 1952 // to use the MDO to detect hot deoptimization points and control 1953 // aggressive optimization. 1954 bool inc_recompile_count = false; 1955 ProfileData* pdata = NULL; 1956 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1957 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1958 uint this_trap_count = 0; 1959 bool maybe_prior_trap = false; 1960 bool maybe_prior_recompile = false; 1961 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1962 #if INCLUDE_JVMCI 1963 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1964 #endif 1965 nm->method(), 1966 //outputs: 1967 this_trap_count, 1968 maybe_prior_trap, 1969 maybe_prior_recompile); 1970 // Because the interpreter also counts null, div0, range, and class 1971 // checks, these traps from compiled code are double-counted. 1972 // This is harmless; it just means that the PerXTrapLimit values 1973 // are in effect a little smaller than they look. 1974 1975 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1976 if (per_bc_reason != Reason_none) { 1977 // Now take action based on the partially known per-BCI history. 1978 if (maybe_prior_trap 1979 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1980 // If there are too many traps at this BCI, force a recompile. 1981 // This will allow the compiler to see the limit overflow, and 1982 // take corrective action, if possible. The compiler generally 1983 // does not use the exact PerBytecodeTrapLimit value, but instead 1984 // changes its tactics if it sees any traps at all. This provides 1985 // a little hysteresis, delaying a recompile until a trap happens 1986 // several times. 1987 // 1988 // Actually, since there is only one bit of counter per BCI, 1989 // the possible per-BCI counts are {0,1,(per-method count)}. 1990 // This produces accurate results if in fact there is only 1991 // one hot trap site, but begins to get fuzzy if there are 1992 // many sites. For example, if there are ten sites each 1993 // trapping two or more times, they each get the blame for 1994 // all of their traps. 1995 make_not_entrant = true; 1996 } 1997 1998 // Detect repeated recompilation at the same BCI, and enforce a limit. 1999 if (make_not_entrant && maybe_prior_recompile) { 2000 // More than one recompile at this point. 2001 inc_recompile_count = maybe_prior_trap; 2002 } 2003 } else { 2004 // For reasons which are not recorded per-bytecode, we simply 2005 // force recompiles unconditionally. 2006 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2007 make_not_entrant = true; 2008 } 2009 2010 // Go back to the compiler if there are too many traps in this method. 2011 if (this_trap_count >= per_method_trap_limit(reason)) { 2012 // If there are too many traps in this method, force a recompile. 2013 // This will allow the compiler to see the limit overflow, and 2014 // take corrective action, if possible. 2015 // (This condition is an unlikely backstop only, because the 2016 // PerBytecodeTrapLimit is more likely to take effect first, 2017 // if it is applicable.) 2018 make_not_entrant = true; 2019 } 2020 2021 // Here's more hysteresis: If there has been a recompile at 2022 // this trap point already, run the method in the interpreter 2023 // for a while to exercise it more thoroughly. 2024 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2025 reprofile = true; 2026 } 2027 } 2028 2029 // Take requested actions on the method: 2030 2031 // Recompile 2032 if (make_not_entrant) { 2033 if (!nm->make_not_entrant()) { 2034 return; // the call did not change nmethod's state 2035 } 2036 2037 if (pdata != NULL) { 2038 // Record the recompilation event, if any. 2039 int tstate0 = pdata->trap_state(); 2040 int tstate1 = trap_state_set_recompiled(tstate0, true); 2041 if (tstate1 != tstate0) 2042 pdata->set_trap_state(tstate1); 2043 } 2044 2045 #if INCLUDE_RTM_OPT 2046 // Restart collecting RTM locking abort statistic if the method 2047 // is recompiled for a reason other than RTM state change. 2048 // Assume that in new recompiled code the statistic could be different, 2049 // for example, due to different inlining. 2050 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2051 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2052 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2053 } 2054 #endif 2055 // For code aging we count traps separately here, using make_not_entrant() 2056 // as a guard against simultaneous deopts in multiple threads. 2057 if (reason == Reason_tenured && trap_mdo != NULL) { 2058 trap_mdo->inc_tenure_traps(); 2059 } 2060 } 2061 2062 if (inc_recompile_count) { 2063 trap_mdo->inc_overflow_recompile_count(); 2064 if ((uint)trap_mdo->overflow_recompile_count() > 2065 (uint)PerBytecodeRecompilationCutoff) { 2066 // Give up on the method containing the bad BCI. 2067 if (trap_method() == nm->method()) { 2068 make_not_compilable = true; 2069 } else { 2070 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2071 // But give grace to the enclosing nm->method(). 2072 } 2073 } 2074 } 2075 2076 // Reprofile 2077 if (reprofile) { 2078 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2079 } 2080 2081 // Give up compiling 2082 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2083 assert(make_not_entrant, "consistent"); 2084 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2085 } 2086 2087 } // Free marked resources 2088 2089 } 2090 JRT_END 2091 2092 ProfileData* 2093 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2094 int trap_bci, 2095 Deoptimization::DeoptReason reason, 2096 bool update_total_trap_count, 2097 #if INCLUDE_JVMCI 2098 bool is_osr, 2099 #endif 2100 Method* compiled_method, 2101 //outputs: 2102 uint& ret_this_trap_count, 2103 bool& ret_maybe_prior_trap, 2104 bool& ret_maybe_prior_recompile) { 2105 bool maybe_prior_trap = false; 2106 bool maybe_prior_recompile = false; 2107 uint this_trap_count = 0; 2108 if (update_total_trap_count) { 2109 uint idx = reason; 2110 #if INCLUDE_JVMCI 2111 if (is_osr) { 2112 idx += Reason_LIMIT; 2113 } 2114 #endif 2115 uint prior_trap_count = trap_mdo->trap_count(idx); 2116 this_trap_count = trap_mdo->inc_trap_count(idx); 2117 2118 // If the runtime cannot find a place to store trap history, 2119 // it is estimated based on the general condition of the method. 2120 // If the method has ever been recompiled, or has ever incurred 2121 // a trap with the present reason , then this BCI is assumed 2122 // (pessimistically) to be the culprit. 2123 maybe_prior_trap = (prior_trap_count != 0); 2124 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2125 } 2126 ProfileData* pdata = NULL; 2127 2128 2129 // For reasons which are recorded per bytecode, we check per-BCI data. 2130 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2131 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2132 if (per_bc_reason != Reason_none) { 2133 // Find the profile data for this BCI. If there isn't one, 2134 // try to allocate one from the MDO's set of spares. 2135 // This will let us detect a repeated trap at this point. 2136 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2137 2138 if (pdata != NULL) { 2139 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2140 if (LogCompilation && xtty != NULL) { 2141 ttyLocker ttyl; 2142 // no more room for speculative traps in this MDO 2143 xtty->elem("speculative_traps_oom"); 2144 } 2145 } 2146 // Query the trap state of this profile datum. 2147 int tstate0 = pdata->trap_state(); 2148 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2149 maybe_prior_trap = false; 2150 if (!trap_state_is_recompiled(tstate0)) 2151 maybe_prior_recompile = false; 2152 2153 // Update the trap state of this profile datum. 2154 int tstate1 = tstate0; 2155 // Record the reason. 2156 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2157 // Store the updated state on the MDO, for next time. 2158 if (tstate1 != tstate0) 2159 pdata->set_trap_state(tstate1); 2160 } else { 2161 if (LogCompilation && xtty != NULL) { 2162 ttyLocker ttyl; 2163 // Missing MDP? Leave a small complaint in the log. 2164 xtty->elem("missing_mdp bci='%d'", trap_bci); 2165 } 2166 } 2167 } 2168 2169 // Return results: 2170 ret_this_trap_count = this_trap_count; 2171 ret_maybe_prior_trap = maybe_prior_trap; 2172 ret_maybe_prior_recompile = maybe_prior_recompile; 2173 return pdata; 2174 } 2175 2176 void 2177 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2178 ResourceMark rm; 2179 // Ignored outputs: 2180 uint ignore_this_trap_count; 2181 bool ignore_maybe_prior_trap; 2182 bool ignore_maybe_prior_recompile; 2183 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2184 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2185 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2186 query_update_method_data(trap_mdo, trap_bci, 2187 (DeoptReason)reason, 2188 update_total_counts, 2189 #if INCLUDE_JVMCI 2190 false, 2191 #endif 2192 NULL, 2193 ignore_this_trap_count, 2194 ignore_maybe_prior_trap, 2195 ignore_maybe_prior_recompile); 2196 } 2197 2198 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2199 if (TraceDeoptimization) { 2200 tty->print("Uncommon trap "); 2201 } 2202 // Still in Java no safepoints 2203 { 2204 // This enters VM and may safepoint 2205 uncommon_trap_inner(thread, trap_request); 2206 } 2207 return fetch_unroll_info_helper(thread, exec_mode); 2208 } 2209 2210 // Local derived constants. 2211 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2212 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2213 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2214 2215 //---------------------------trap_state_reason--------------------------------- 2216 Deoptimization::DeoptReason 2217 Deoptimization::trap_state_reason(int trap_state) { 2218 // This assert provides the link between the width of DataLayout::trap_bits 2219 // and the encoding of "recorded" reasons. It ensures there are enough 2220 // bits to store all needed reasons in the per-BCI MDO profile. 2221 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2222 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2223 trap_state -= recompile_bit; 2224 if (trap_state == DS_REASON_MASK) { 2225 return Reason_many; 2226 } else { 2227 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2228 return (DeoptReason)trap_state; 2229 } 2230 } 2231 //-------------------------trap_state_has_reason------------------------------- 2232 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2233 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2234 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2235 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2236 trap_state -= recompile_bit; 2237 if (trap_state == DS_REASON_MASK) { 2238 return -1; // true, unspecifically (bottom of state lattice) 2239 } else if (trap_state == reason) { 2240 return 1; // true, definitely 2241 } else if (trap_state == 0) { 2242 return 0; // false, definitely (top of state lattice) 2243 } else { 2244 return 0; // false, definitely 2245 } 2246 } 2247 //-------------------------trap_state_add_reason------------------------------- 2248 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2249 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2250 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2251 trap_state -= recompile_bit; 2252 if (trap_state == DS_REASON_MASK) { 2253 return trap_state + recompile_bit; // already at state lattice bottom 2254 } else if (trap_state == reason) { 2255 return trap_state + recompile_bit; // the condition is already true 2256 } else if (trap_state == 0) { 2257 return reason + recompile_bit; // no condition has yet been true 2258 } else { 2259 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2260 } 2261 } 2262 //-----------------------trap_state_is_recompiled------------------------------ 2263 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2264 return (trap_state & DS_RECOMPILE_BIT) != 0; 2265 } 2266 //-----------------------trap_state_set_recompiled----------------------------- 2267 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2268 if (z) return trap_state | DS_RECOMPILE_BIT; 2269 else return trap_state & ~DS_RECOMPILE_BIT; 2270 } 2271 //---------------------------format_trap_state--------------------------------- 2272 // This is used for debugging and diagnostics, including LogFile output. 2273 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2274 int trap_state) { 2275 assert(buflen > 0, "sanity"); 2276 DeoptReason reason = trap_state_reason(trap_state); 2277 bool recomp_flag = trap_state_is_recompiled(trap_state); 2278 // Re-encode the state from its decoded components. 2279 int decoded_state = 0; 2280 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2281 decoded_state = trap_state_add_reason(decoded_state, reason); 2282 if (recomp_flag) 2283 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2284 // If the state re-encodes properly, format it symbolically. 2285 // Because this routine is used for debugging and diagnostics, 2286 // be robust even if the state is a strange value. 2287 size_t len; 2288 if (decoded_state != trap_state) { 2289 // Random buggy state that doesn't decode?? 2290 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2291 } else { 2292 len = jio_snprintf(buf, buflen, "%s%s", 2293 trap_reason_name(reason), 2294 recomp_flag ? " recompiled" : ""); 2295 } 2296 return buf; 2297 } 2298 2299 2300 //--------------------------------statics-------------------------------------- 2301 const char* Deoptimization::_trap_reason_name[] = { 2302 // Note: Keep this in sync. with enum DeoptReason. 2303 "none", 2304 "null_check", 2305 "null_assert" JVMCI_ONLY("_or_unreached0"), 2306 "range_check", 2307 "class_check", 2308 "array_check", 2309 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2310 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2311 "profile_predicate", 2312 "unloaded", 2313 "uninitialized", 2314 "initialized", 2315 "unreached", 2316 "unhandled", 2317 "constraint", 2318 "div0_check", 2319 "age", 2320 "predicate", 2321 "loop_limit_check", 2322 "speculate_class_check", 2323 "speculate_null_check", 2324 "speculate_null_assert", 2325 "rtm_state_change", 2326 "unstable_if", 2327 "unstable_fused_if", 2328 #if INCLUDE_JVMCI 2329 "aliasing", 2330 "transfer_to_interpreter", 2331 "not_compiled_exception_handler", 2332 "unresolved", 2333 "jsr_mismatch", 2334 #endif 2335 "tenured" 2336 }; 2337 const char* Deoptimization::_trap_action_name[] = { 2338 // Note: Keep this in sync. with enum DeoptAction. 2339 "none", 2340 "maybe_recompile", 2341 "reinterpret", 2342 "make_not_entrant", 2343 "make_not_compilable" 2344 }; 2345 2346 const char* Deoptimization::trap_reason_name(int reason) { 2347 // Check that every reason has a name 2348 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2349 2350 if (reason == Reason_many) return "many"; 2351 if ((uint)reason < Reason_LIMIT) 2352 return _trap_reason_name[reason]; 2353 static char buf[20]; 2354 sprintf(buf, "reason%d", reason); 2355 return buf; 2356 } 2357 const char* Deoptimization::trap_action_name(int action) { 2358 // Check that every action has a name 2359 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2360 2361 if ((uint)action < Action_LIMIT) 2362 return _trap_action_name[action]; 2363 static char buf[20]; 2364 sprintf(buf, "action%d", action); 2365 return buf; 2366 } 2367 2368 // This is used for debugging and diagnostics, including LogFile output. 2369 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2370 int trap_request) { 2371 jint unloaded_class_index = trap_request_index(trap_request); 2372 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2373 const char* action = trap_action_name(trap_request_action(trap_request)); 2374 #if INCLUDE_JVMCI 2375 int debug_id = trap_request_debug_id(trap_request); 2376 #endif 2377 size_t len; 2378 if (unloaded_class_index < 0) { 2379 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2380 reason, action 2381 #if INCLUDE_JVMCI 2382 ,debug_id 2383 #endif 2384 ); 2385 } else { 2386 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2387 reason, action, unloaded_class_index 2388 #if INCLUDE_JVMCI 2389 ,debug_id 2390 #endif 2391 ); 2392 } 2393 return buf; 2394 } 2395 2396 juint Deoptimization::_deoptimization_hist 2397 [Deoptimization::Reason_LIMIT] 2398 [1 + Deoptimization::Action_LIMIT] 2399 [Deoptimization::BC_CASE_LIMIT] 2400 = {0}; 2401 2402 enum { 2403 LSB_BITS = 8, 2404 LSB_MASK = right_n_bits(LSB_BITS) 2405 }; 2406 2407 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2408 Bytecodes::Code bc) { 2409 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2410 assert(action >= 0 && action < Action_LIMIT, "oob"); 2411 _deoptimization_hist[Reason_none][0][0] += 1; // total 2412 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2413 juint* cases = _deoptimization_hist[reason][1+action]; 2414 juint* bc_counter_addr = NULL; 2415 juint bc_counter = 0; 2416 // Look for an unused counter, or an exact match to this BC. 2417 if (bc != Bytecodes::_illegal) { 2418 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2419 juint* counter_addr = &cases[bc_case]; 2420 juint counter = *counter_addr; 2421 if ((counter == 0 && bc_counter_addr == NULL) 2422 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2423 // this counter is either free or is already devoted to this BC 2424 bc_counter_addr = counter_addr; 2425 bc_counter = counter | bc; 2426 } 2427 } 2428 } 2429 if (bc_counter_addr == NULL) { 2430 // Overflow, or no given bytecode. 2431 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2432 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2433 } 2434 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2435 } 2436 2437 jint Deoptimization::total_deoptimization_count() { 2438 return _deoptimization_hist[Reason_none][0][0]; 2439 } 2440 2441 void Deoptimization::print_statistics() { 2442 juint total = total_deoptimization_count(); 2443 juint account = total; 2444 if (total != 0) { 2445 ttyLocker ttyl; 2446 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2447 tty->print_cr("Deoptimization traps recorded:"); 2448 #define PRINT_STAT_LINE(name, r) \ 2449 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2450 PRINT_STAT_LINE("total", total); 2451 // For each non-zero entry in the histogram, print the reason, 2452 // the action, and (if specifically known) the type of bytecode. 2453 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2454 for (int action = 0; action < Action_LIMIT; action++) { 2455 juint* cases = _deoptimization_hist[reason][1+action]; 2456 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2457 juint counter = cases[bc_case]; 2458 if (counter != 0) { 2459 char name[1*K]; 2460 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2461 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2462 bc = Bytecodes::_illegal; 2463 sprintf(name, "%s/%s/%s", 2464 trap_reason_name(reason), 2465 trap_action_name(action), 2466 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2467 juint r = counter >> LSB_BITS; 2468 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2469 account -= r; 2470 } 2471 } 2472 } 2473 } 2474 if (account != 0) { 2475 PRINT_STAT_LINE("unaccounted", account); 2476 } 2477 #undef PRINT_STAT_LINE 2478 if (xtty != NULL) xtty->tail("statistics"); 2479 } 2480 } 2481 #else // COMPILER2_OR_JVMCI 2482 2483 2484 // Stubs for C1 only system. 2485 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2486 return false; 2487 } 2488 2489 const char* Deoptimization::trap_reason_name(int reason) { 2490 return "unknown"; 2491 } 2492 2493 void Deoptimization::print_statistics() { 2494 // no output 2495 } 2496 2497 void 2498 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2499 // no udpate 2500 } 2501 2502 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2503 return 0; 2504 } 2505 2506 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2507 Bytecodes::Code bc) { 2508 // no update 2509 } 2510 2511 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2512 int trap_state) { 2513 jio_snprintf(buf, buflen, "#%d", trap_state); 2514 return buf; 2515 } 2516 2517 #endif // COMPILER2_OR_JVMCI