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 "compiler/compilationPolicy.hpp" 35 #include "interpreter/bytecode.hpp" 36 #include "interpreter/interpreter.hpp" 37 #include "interpreter/oopMapCache.hpp" 38 #include "memory/allocation.inline.hpp" 39 #include "memory/oopFactory.hpp" 40 #include "memory/resourceArea.hpp" 41 #include "memory/universe.hpp" 42 #include "oops/constantPool.hpp" 43 #include "oops/method.hpp" 44 #include "oops/objArrayKlass.hpp" 45 #include "oops/objArrayOop.inline.hpp" 46 #include "oops/oop.inline.hpp" 47 #include "oops/fieldStreams.inline.hpp" 48 #include "oops/typeArrayOop.inline.hpp" 49 #include "oops/verifyOopClosure.hpp" 50 #include "prims/jvmtiThreadState.hpp" 51 #include "runtime/biasedLocking.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 && EliminateAllocations)) { 299 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk); 300 } 301 #endif // COMPILER2_OR_JVMCI 302 303 // Revoke biases, done with in java state. 304 // No safepoints allowed after this 305 revoke_from_deopt_handler(thread, deoptee, &map); 306 307 // Ensure that no safepoint is taken after pointers have been stored 308 // in fields of rematerialized objects. If a safepoint occurs from here on 309 // out the java state residing in the vframeArray will be missed. 310 // Locks may be rebaised in a safepoint. 311 NoSafepointVerifier no_safepoint; 312 313 #if COMPILER2_OR_JVMCI 314 if (jvmci_enabled || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks)) { 315 eliminate_locks(thread, chunk, realloc_failures); 316 } 317 #endif // COMPILER2_OR_JVMCI 318 319 ScopeDesc* trap_scope = chunk->at(0)->scope(); 320 Handle exceptionObject; 321 if (trap_scope->rethrow_exception()) { 322 if (PrintDeoptimizationDetails) { 323 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()); 324 } 325 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 326 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 327 ScopeValue* topOfStack = expressions->top(); 328 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 329 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 330 } 331 332 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 333 #if COMPILER2_OR_JVMCI 334 if (realloc_failures) { 335 pop_frames_failed_reallocs(thread, array); 336 } 337 #endif 338 339 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 340 thread->set_vframe_array_head(array); 341 342 // Now that the vframeArray has been created if we have any deferred local writes 343 // added by jvmti then we can free up that structure as the data is now in the 344 // vframeArray 345 346 if (thread->deferred_locals() != NULL) { 347 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 348 int i = 0; 349 do { 350 // Because of inlining we could have multiple vframes for a single frame 351 // and several of the vframes could have deferred writes. Find them all. 352 if (list->at(i)->id() == array->original().id()) { 353 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 354 list->remove_at(i); 355 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 356 delete dlv; 357 } else { 358 i++; 359 } 360 } while ( i < list->length() ); 361 if (list->length() == 0) { 362 thread->set_deferred_locals(NULL); 363 // free the list and elements back to C heap. 364 delete list; 365 } 366 367 } 368 369 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 370 CodeBlob* cb = stub_frame.cb(); 371 // Verify we have the right vframeArray 372 assert(cb->frame_size() >= 0, "Unexpected frame size"); 373 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 374 375 // If the deopt call site is a MethodHandle invoke call site we have 376 // to adjust the unpack_sp. 377 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 378 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 379 unpack_sp = deoptee.unextended_sp(); 380 381 #ifdef ASSERT 382 assert(cb->is_deoptimization_stub() || 383 cb->is_uncommon_trap_stub() || 384 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 385 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 386 "unexpected code blob: %s", cb->name()); 387 #endif 388 389 // This is a guarantee instead of an assert because if vframe doesn't match 390 // we will unpack the wrong deoptimized frame and wind up in strange places 391 // where it will be very difficult to figure out what went wrong. Better 392 // to die an early death here than some very obscure death later when the 393 // trail is cold. 394 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 395 // in that it will fail to detect a problem when there is one. This needs 396 // more work in tiger timeframe. 397 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 398 399 int number_of_frames = array->frames(); 400 401 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 402 // virtual activation, which is the reverse of the elements in the vframes array. 403 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 404 // +1 because we always have an interpreter return address for the final slot. 405 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 406 int popframe_extra_args = 0; 407 // Create an interpreter return address for the stub to use as its return 408 // address so the skeletal frames are perfectly walkable 409 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 410 411 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 412 // activation be put back on the expression stack of the caller for reexecution 413 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 414 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 415 } 416 417 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 418 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 419 // than simply use array->sender.pc(). This requires us to walk the current set of frames 420 // 421 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 422 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 423 424 // It's possible that the number of parameters at the call site is 425 // different than number of arguments in the callee when method 426 // handles are used. If the caller is interpreted get the real 427 // value so that the proper amount of space can be added to it's 428 // frame. 429 bool caller_was_method_handle = false; 430 if (deopt_sender.is_interpreted_frame()) { 431 methodHandle method(thread, deopt_sender.interpreter_frame_method()); 432 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 433 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 434 // Method handle invokes may involve fairly arbitrary chains of 435 // calls so it's impossible to know how much actual space the 436 // caller has for locals. 437 caller_was_method_handle = true; 438 } 439 } 440 441 // 442 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 443 // frame_sizes/frame_pcs[1] next oldest frame (int) 444 // frame_sizes/frame_pcs[n] youngest frame (int) 445 // 446 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 447 // owns the space for the return address to it's caller). Confusing ain't it. 448 // 449 // The vframe array can address vframes with indices running from 450 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 451 // When we create the skeletal frames we need the oldest frame to be in the zero slot 452 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 453 // so things look a little strange in this loop. 454 // 455 int callee_parameters = 0; 456 int callee_locals = 0; 457 for (int index = 0; index < array->frames(); index++ ) { 458 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 459 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 460 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 461 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 462 callee_locals, 463 index == 0, 464 popframe_extra_args); 465 // This pc doesn't have to be perfect just good enough to identify the frame 466 // as interpreted so the skeleton frame will be walkable 467 // The correct pc will be set when the skeleton frame is completely filled out 468 // The final pc we store in the loop is wrong and will be overwritten below 469 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 470 471 callee_parameters = array->element(index)->method()->size_of_parameters(); 472 callee_locals = array->element(index)->method()->max_locals(); 473 popframe_extra_args = 0; 474 } 475 476 // Compute whether the root vframe returns a float or double value. 477 BasicType return_type; 478 { 479 methodHandle method(thread, array->element(0)->method()); 480 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 481 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 482 } 483 484 // Compute information for handling adapters and adjusting the frame size of the caller. 485 int caller_adjustment = 0; 486 487 // Compute the amount the oldest interpreter frame will have to adjust 488 // its caller's stack by. If the caller is a compiled frame then 489 // we pretend that the callee has no parameters so that the 490 // extension counts for the full amount of locals and not just 491 // locals-parms. This is because without a c2i adapter the parm 492 // area as created by the compiled frame will not be usable by 493 // the interpreter. (Depending on the calling convention there 494 // may not even be enough space). 495 496 // QQQ I'd rather see this pushed down into last_frame_adjust 497 // and have it take the sender (aka caller). 498 499 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 500 caller_adjustment = last_frame_adjust(0, callee_locals); 501 } else if (callee_locals > callee_parameters) { 502 // The caller frame may need extending to accommodate 503 // non-parameter locals of the first unpacked interpreted frame. 504 // Compute that adjustment. 505 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 506 } 507 508 // If the sender is deoptimized the we must retrieve the address of the handler 509 // since the frame will "magically" show the original pc before the deopt 510 // and we'd undo the deopt. 511 512 frame_pcs[0] = deopt_sender.raw_pc(); 513 514 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 515 516 #if INCLUDE_JVMCI 517 if (exceptionObject() != NULL) { 518 thread->set_exception_oop(exceptionObject()); 519 exec_mode = Unpack_exception; 520 } 521 #endif 522 523 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 524 assert(thread->has_pending_exception(), "should have thrown OOME"); 525 thread->set_exception_oop(thread->pending_exception()); 526 thread->clear_pending_exception(); 527 exec_mode = Unpack_exception; 528 } 529 530 #if INCLUDE_JVMCI 531 if (thread->frames_to_pop_failed_realloc() > 0) { 532 thread->set_pending_monitorenter(false); 533 } 534 #endif 535 536 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 537 caller_adjustment * BytesPerWord, 538 caller_was_method_handle ? 0 : callee_parameters, 539 number_of_frames, 540 frame_sizes, 541 frame_pcs, 542 return_type, 543 exec_mode); 544 // On some platforms, we need a way to pass some platform dependent 545 // information to the unpacking code so the skeletal frames come out 546 // correct (initial fp value, unextended sp, ...) 547 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 548 549 if (array->frames() > 1) { 550 if (VerifyStack && TraceDeoptimization) { 551 ttyLocker ttyl; 552 tty->print_cr("Deoptimizing method containing inlining"); 553 } 554 } 555 556 array->set_unroll_block(info); 557 return info; 558 } 559 560 // Called to cleanup deoptimization data structures in normal case 561 // after unpacking to stack and when stack overflow error occurs 562 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 563 vframeArray *array) { 564 565 // Get array if coming from exception 566 if (array == NULL) { 567 array = thread->vframe_array_head(); 568 } 569 thread->set_vframe_array_head(NULL); 570 571 // Free the previous UnrollBlock 572 vframeArray* old_array = thread->vframe_array_last(); 573 thread->set_vframe_array_last(array); 574 575 if (old_array != NULL) { 576 UnrollBlock* old_info = old_array->unroll_block(); 577 old_array->set_unroll_block(NULL); 578 delete old_info; 579 delete old_array; 580 } 581 582 // Deallocate any resource creating in this routine and any ResourceObjs allocated 583 // inside the vframeArray (StackValueCollections) 584 585 delete thread->deopt_mark(); 586 thread->set_deopt_mark(NULL); 587 thread->set_deopt_compiled_method(NULL); 588 589 590 if (JvmtiExport::can_pop_frame()) { 591 #ifndef CC_INTERP 592 // Regardless of whether we entered this routine with the pending 593 // popframe condition bit set, we should always clear it now 594 thread->clear_popframe_condition(); 595 #else 596 // C++ interpreter will clear has_pending_popframe when it enters 597 // with method_resume. For deopt_resume2 we clear it now. 598 if (thread->popframe_forcing_deopt_reexecution()) 599 thread->clear_popframe_condition(); 600 #endif /* CC_INTERP */ 601 } 602 603 // unpack_frames() is called at the end of the deoptimization handler 604 // and (in C2) at the end of the uncommon trap handler. Note this fact 605 // so that an asynchronous stack walker can work again. This counter is 606 // incremented at the beginning of fetch_unroll_info() and (in C2) at 607 // the beginning of uncommon_trap(). 608 thread->dec_in_deopt_handler(); 609 } 610 611 // Moved from cpu directories because none of the cpus has callee save values. 612 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 613 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 614 615 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 616 // the days we had adapter frames. When we deoptimize a situation where a 617 // compiled caller calls a compiled caller will have registers it expects 618 // to survive the call to the callee. If we deoptimize the callee the only 619 // way we can restore these registers is to have the oldest interpreter 620 // frame that we create restore these values. That is what this routine 621 // will accomplish. 622 623 // At the moment we have modified c2 to not have any callee save registers 624 // so this problem does not exist and this routine is just a place holder. 625 626 assert(f->is_interpreted_frame(), "must be interpreted"); 627 } 628 629 // Return BasicType of value being returned 630 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 631 632 // We are already active in the special DeoptResourceMark any ResourceObj's we 633 // allocate will be freed at the end of the routine. 634 635 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 636 // but makes the entry a little slower. There is however a little dance we have to 637 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 638 ResetNoHandleMark rnhm; // No-op in release/product versions 639 HandleMark hm; 640 641 frame stub_frame = thread->last_frame(); 642 643 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 644 // must point to the vframeArray for the unpack frame. 645 vframeArray* array = thread->vframe_array_head(); 646 647 #ifndef PRODUCT 648 if (TraceDeoptimization) { 649 ttyLocker ttyl; 650 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 651 p2i(thread), p2i(array), exec_mode); 652 } 653 #endif 654 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 655 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 656 657 UnrollBlock* info = array->unroll_block(); 658 659 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 660 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 661 662 BasicType bt = info->return_type(); 663 664 // If we have an exception pending, claim that the return type is an oop 665 // so the deopt_blob does not overwrite the exception_oop. 666 667 if (exec_mode == Unpack_exception) 668 bt = T_OBJECT; 669 670 // Cleanup thread deopt data 671 cleanup_deopt_info(thread, array); 672 673 #ifndef PRODUCT 674 if (VerifyStack) { 675 ResourceMark res_mark; 676 // Clear pending exception to not break verification code (restored afterwards) 677 PRESERVE_EXCEPTION_MARK; 678 679 thread->validate_frame_layout(); 680 681 // Verify that the just-unpacked frames match the interpreter's 682 // notions of expression stack and locals 683 vframeArray* cur_array = thread->vframe_array_last(); 684 RegisterMap rm(thread, false); 685 rm.set_include_argument_oops(false); 686 bool is_top_frame = true; 687 int callee_size_of_parameters = 0; 688 int callee_max_locals = 0; 689 for (int i = 0; i < cur_array->frames(); i++) { 690 vframeArrayElement* el = cur_array->element(i); 691 frame* iframe = el->iframe(); 692 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 693 694 // Get the oop map for this bci 695 InterpreterOopMap mask; 696 int cur_invoke_parameter_size = 0; 697 bool try_next_mask = false; 698 int next_mask_expression_stack_size = -1; 699 int top_frame_expression_stack_adjustment = 0; 700 methodHandle mh(thread, iframe->interpreter_frame_method()); 701 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 702 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 703 int max_bci = mh->code_size(); 704 // Get to the next bytecode if possible 705 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 706 // Check to see if we can grab the number of outgoing arguments 707 // at an uncommon trap for an invoke (where the compiler 708 // generates debug info before the invoke has executed) 709 Bytecodes::Code cur_code = str.next(); 710 if (Bytecodes::is_invoke(cur_code)) { 711 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 712 cur_invoke_parameter_size = invoke.size_of_parameters(); 713 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 714 callee_size_of_parameters++; 715 } 716 } 717 if (str.bci() < max_bci) { 718 Bytecodes::Code next_code = str.next(); 719 if (next_code >= 0) { 720 // The interpreter oop map generator reports results before 721 // the current bytecode has executed except in the case of 722 // calls. It seems to be hard to tell whether the compiler 723 // has emitted debug information matching the "state before" 724 // a given bytecode or the state after, so we try both 725 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 726 // Get expression stack size for the next bytecode 727 InterpreterOopMap next_mask; 728 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 729 next_mask_expression_stack_size = next_mask.expression_stack_size(); 730 if (Bytecodes::is_invoke(next_code)) { 731 Bytecode_invoke invoke(mh, str.bci()); 732 next_mask_expression_stack_size += invoke.size_of_parameters(); 733 } 734 // Need to subtract off the size of the result type of 735 // the bytecode because this is not described in the 736 // debug info but returned to the interpreter in the TOS 737 // caching register 738 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 739 if (bytecode_result_type != T_ILLEGAL) { 740 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 741 } 742 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 743 try_next_mask = true; 744 } 745 } 746 } 747 748 // Verify stack depth and oops in frame 749 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 750 if (!( 751 /* SPARC */ 752 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 753 /* x86 */ 754 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 755 (try_next_mask && 756 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 757 top_frame_expression_stack_adjustment))) || 758 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 759 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 760 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 761 )) { 762 { 763 ttyLocker ttyl; 764 765 // Print out some information that will help us debug the problem 766 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 767 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 768 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 769 iframe->interpreter_frame_expression_stack_size()); 770 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 771 tty->print_cr(" try_next_mask = %d", try_next_mask); 772 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 773 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 774 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 775 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 776 tty->print_cr(" exec_mode = %d", exec_mode); 777 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 778 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 779 tty->print_cr(" Interpreted frames:"); 780 for (int k = 0; k < cur_array->frames(); k++) { 781 vframeArrayElement* el = cur_array->element(k); 782 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 783 } 784 cur_array->print_on_2(tty); 785 } // release tty lock before calling guarantee 786 guarantee(false, "wrong number of expression stack elements during deopt"); 787 } 788 VerifyOopClosure verify; 789 iframe->oops_interpreted_do(&verify, &rm, false); 790 callee_size_of_parameters = mh->size_of_parameters(); 791 callee_max_locals = mh->max_locals(); 792 is_top_frame = false; 793 } 794 } 795 #endif /* !PRODUCT */ 796 797 798 return bt; 799 JRT_END 800 801 class DeoptimizeMarkedTC : public ThreadClosure { 802 public: 803 virtual void do_thread(Thread* thread) { 804 assert(thread->is_Java_thread(), "must be"); 805 JavaThread* jt = (JavaThread*)thread; 806 jt->deoptimize_marked_methods(); 807 } 808 }; 809 810 void Deoptimization::deoptimize_all_marked() { 811 ResourceMark rm; 812 DeoptimizationMarker dm; 813 814 if (SafepointSynchronize::is_at_safepoint()) { 815 DeoptimizeMarkedTC deopt; 816 // Make the dependent methods not entrant 817 CodeCache::make_marked_nmethods_not_entrant(); 818 Threads::java_threads_do(&deopt); 819 } else { 820 // Make the dependent methods not entrant 821 { 822 MutexLocker mu(CodeCache_lock, Mutex::_no_safepoint_check_flag); 823 CodeCache::make_marked_nmethods_not_entrant(); 824 } 825 DeoptimizeMarkedTC deopt; 826 Handshake::execute(&deopt); 827 } 828 } 829 830 Deoptimization::DeoptAction Deoptimization::_unloaded_action 831 = Deoptimization::Action_reinterpret; 832 833 834 835 #if INCLUDE_JVMCI || INCLUDE_AOT 836 template<typename CacheType> 837 class BoxCacheBase : public CHeapObj<mtCompiler> { 838 protected: 839 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { 840 ResourceMark rm; 841 char* klass_name_str = klass_name->as_C_string(); 842 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); 843 guarantee(k != NULL, "%s must be loaded", klass_name_str); 844 InstanceKlass* ik = InstanceKlass::cast(k); 845 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 846 CacheType::compute_offsets(ik); 847 return ik; 848 } 849 }; 850 851 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 852 PrimitiveType _low; 853 PrimitiveType _high; 854 jobject _cache; 855 protected: 856 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 857 BoxCache(Thread* thread) { 858 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); 859 objArrayOop cache = CacheType::cache(ik); 860 assert(cache->length() > 0, "Empty cache"); 861 _low = BoxType::value(cache->obj_at(0)); 862 _high = _low + cache->length() - 1; 863 _cache = JNIHandles::make_global(Handle(thread, cache)); 864 } 865 ~BoxCache() { 866 JNIHandles::destroy_global(_cache); 867 } 868 public: 869 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 870 if (_singleton == NULL) { 871 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 872 if (!Atomic::replace_if_null(s, &_singleton)) { 873 delete s; 874 } 875 } 876 return _singleton; 877 } 878 oop lookup(PrimitiveType value) { 879 if (_low <= value && value <= _high) { 880 int offset = value - _low; 881 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 882 } 883 return NULL; 884 } 885 oop lookup_raw(intptr_t raw_value) { 886 // Have to cast to avoid little/big-endian problems. 887 if (sizeof(PrimitiveType) > sizeof(jint)) { 888 jlong value = (jlong)raw_value; 889 return lookup(value); 890 } 891 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 892 return lookup(value); 893 } 894 }; 895 896 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 897 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 898 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 899 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 900 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 901 902 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 903 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 904 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 905 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 906 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 907 908 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 909 jobject _true_cache; 910 jobject _false_cache; 911 protected: 912 static BooleanBoxCache *_singleton; 913 BooleanBoxCache(Thread *thread) { 914 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); 915 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 916 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 917 } 918 ~BooleanBoxCache() { 919 JNIHandles::destroy_global(_true_cache); 920 JNIHandles::destroy_global(_false_cache); 921 } 922 public: 923 static BooleanBoxCache* singleton(Thread* thread) { 924 if (_singleton == NULL) { 925 BooleanBoxCache* s = new BooleanBoxCache(thread); 926 if (!Atomic::replace_if_null(s, &_singleton)) { 927 delete s; 928 } 929 } 930 return _singleton; 931 } 932 oop lookup_raw(intptr_t raw_value) { 933 // Have to cast to avoid little/big-endian problems. 934 jboolean value = (jboolean)*((jint*)&raw_value); 935 return lookup(value); 936 } 937 oop lookup(jboolean value) { 938 if (value != 0) { 939 return JNIHandles::resolve_non_null(_true_cache); 940 } 941 return JNIHandles::resolve_non_null(_false_cache); 942 } 943 }; 944 945 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 946 947 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { 948 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 949 BasicType box_type = SystemDictionary::box_klass_type(k); 950 if (box_type != T_OBJECT) { 951 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 952 switch(box_type) { 953 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 954 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 955 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 956 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 957 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 958 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 959 default:; 960 } 961 } 962 return NULL; 963 } 964 #endif // INCLUDE_JVMCI || INCLUDE_AOT 965 966 #if COMPILER2_OR_JVMCI 967 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 968 Handle pending_exception(THREAD, thread->pending_exception()); 969 const char* exception_file = thread->exception_file(); 970 int exception_line = thread->exception_line(); 971 thread->clear_pending_exception(); 972 973 bool failures = false; 974 975 for (int i = 0; i < objects->length(); i++) { 976 assert(objects->at(i)->is_object(), "invalid debug information"); 977 ObjectValue* sv = (ObjectValue*) objects->at(i); 978 979 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 980 oop obj = NULL; 981 982 if (k->is_instance_klass()) { 983 #if INCLUDE_JVMCI || INCLUDE_AOT 984 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 985 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 986 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 987 obj = get_cached_box(abv, fr, reg_map, THREAD); 988 if (obj != NULL) { 989 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 990 abv->set_cached(true); 991 } 992 } 993 #endif // INCLUDE_JVMCI || INCLUDE_AOT 994 InstanceKlass* ik = InstanceKlass::cast(k); 995 if (obj == NULL) { 996 obj = ik->allocate_instance(THREAD); 997 } 998 } else if (k->is_typeArray_klass()) { 999 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1000 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1001 int len = sv->field_size() / type2size[ak->element_type()]; 1002 obj = ak->allocate(len, THREAD); 1003 } else if (k->is_objArray_klass()) { 1004 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1005 obj = ak->allocate(sv->field_size(), THREAD); 1006 } 1007 1008 if (obj == NULL) { 1009 failures = true; 1010 } 1011 1012 assert(sv->value().is_null(), "redundant reallocation"); 1013 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1014 CLEAR_PENDING_EXCEPTION; 1015 sv->set_value(obj); 1016 } 1017 1018 if (failures) { 1019 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1020 } else if (pending_exception.not_null()) { 1021 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1022 } 1023 1024 return failures; 1025 } 1026 1027 // restore elements of an eliminated type array 1028 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1029 int index = 0; 1030 intptr_t val; 1031 1032 for (int i = 0; i < sv->field_size(); i++) { 1033 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1034 switch(type) { 1035 case T_LONG: case T_DOUBLE: { 1036 assert(value->type() == T_INT, "Agreement."); 1037 StackValue* low = 1038 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1039 #ifdef _LP64 1040 jlong res = (jlong)low->get_int(); 1041 #else 1042 #ifdef SPARC 1043 // For SPARC we have to swap high and low words. 1044 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1045 #else 1046 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1047 #endif //SPARC 1048 #endif 1049 obj->long_at_put(index, res); 1050 break; 1051 } 1052 1053 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1054 case T_INT: case T_FLOAT: { // 4 bytes. 1055 assert(value->type() == T_INT, "Agreement."); 1056 bool big_value = false; 1057 if (i + 1 < sv->field_size() && type == T_INT) { 1058 if (sv->field_at(i)->is_location()) { 1059 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1060 if (type == Location::dbl || type == Location::lng) { 1061 big_value = true; 1062 } 1063 } else if (sv->field_at(i)->is_constant_int()) { 1064 ScopeValue* next_scope_field = sv->field_at(i + 1); 1065 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1066 big_value = true; 1067 } 1068 } 1069 } 1070 1071 if (big_value) { 1072 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1073 #ifdef _LP64 1074 jlong res = (jlong)low->get_int(); 1075 #else 1076 #ifdef SPARC 1077 // For SPARC we have to swap high and low words. 1078 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1079 #else 1080 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1081 #endif //SPARC 1082 #endif 1083 obj->int_at_put(index, (jint)*((jint*)&res)); 1084 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1085 } else { 1086 val = value->get_int(); 1087 obj->int_at_put(index, (jint)*((jint*)&val)); 1088 } 1089 break; 1090 } 1091 1092 case T_SHORT: 1093 assert(value->type() == T_INT, "Agreement."); 1094 val = value->get_int(); 1095 obj->short_at_put(index, (jshort)*((jint*)&val)); 1096 break; 1097 1098 case T_CHAR: 1099 assert(value->type() == T_INT, "Agreement."); 1100 val = value->get_int(); 1101 obj->char_at_put(index, (jchar)*((jint*)&val)); 1102 break; 1103 1104 case T_BYTE: 1105 assert(value->type() == T_INT, "Agreement."); 1106 val = value->get_int(); 1107 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1108 break; 1109 1110 case T_BOOLEAN: 1111 assert(value->type() == T_INT, "Agreement."); 1112 val = value->get_int(); 1113 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1114 break; 1115 1116 default: 1117 ShouldNotReachHere(); 1118 } 1119 index++; 1120 } 1121 } 1122 1123 1124 // restore fields of an eliminated object array 1125 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1126 for (int i = 0; i < sv->field_size(); i++) { 1127 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1128 assert(value->type() == T_OBJECT, "object element expected"); 1129 obj->obj_at_put(i, value->get_obj()()); 1130 } 1131 } 1132 1133 class ReassignedField { 1134 public: 1135 int _offset; 1136 BasicType _type; 1137 public: 1138 ReassignedField() { 1139 _offset = 0; 1140 _type = T_ILLEGAL; 1141 } 1142 }; 1143 1144 int compare(ReassignedField* left, ReassignedField* right) { 1145 return left->_offset - right->_offset; 1146 } 1147 1148 // Restore fields of an eliminated instance object using the same field order 1149 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1150 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1151 if (klass->superklass() != NULL) { 1152 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); 1153 } 1154 1155 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1156 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1157 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1158 ReassignedField field; 1159 field._offset = fs.offset(); 1160 field._type = FieldType::basic_type(fs.signature()); 1161 fields->append(field); 1162 } 1163 } 1164 fields->sort(compare); 1165 for (int i = 0; i < fields->length(); i++) { 1166 intptr_t val; 1167 ScopeValue* scope_field = sv->field_at(svIndex); 1168 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1169 int offset = fields->at(i)._offset; 1170 BasicType type = fields->at(i)._type; 1171 switch (type) { 1172 case T_OBJECT: case T_ARRAY: 1173 assert(value->type() == T_OBJECT, "Agreement."); 1174 obj->obj_field_put(offset, value->get_obj()()); 1175 break; 1176 1177 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1178 case T_INT: case T_FLOAT: { // 4 bytes. 1179 assert(value->type() == T_INT, "Agreement."); 1180 bool big_value = false; 1181 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1182 if (scope_field->is_location()) { 1183 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1184 if (type == Location::dbl || type == Location::lng) { 1185 big_value = true; 1186 } 1187 } 1188 if (scope_field->is_constant_int()) { 1189 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1190 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1191 big_value = true; 1192 } 1193 } 1194 } 1195 1196 if (big_value) { 1197 i++; 1198 assert(i < fields->length(), "second T_INT field needed"); 1199 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1200 } else { 1201 val = value->get_int(); 1202 obj->int_field_put(offset, (jint)*((jint*)&val)); 1203 break; 1204 } 1205 } 1206 /* no break */ 1207 1208 case T_LONG: case T_DOUBLE: { 1209 assert(value->type() == T_INT, "Agreement."); 1210 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1211 #ifdef _LP64 1212 jlong res = (jlong)low->get_int(); 1213 #else 1214 #ifdef SPARC 1215 // For SPARC we have to swap high and low words. 1216 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1217 #else 1218 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1219 #endif //SPARC 1220 #endif 1221 obj->long_field_put(offset, res); 1222 break; 1223 } 1224 1225 case T_SHORT: 1226 assert(value->type() == T_INT, "Agreement."); 1227 val = value->get_int(); 1228 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1229 break; 1230 1231 case T_CHAR: 1232 assert(value->type() == T_INT, "Agreement."); 1233 val = value->get_int(); 1234 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1235 break; 1236 1237 case T_BYTE: 1238 assert(value->type() == T_INT, "Agreement."); 1239 val = value->get_int(); 1240 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1241 break; 1242 1243 case T_BOOLEAN: 1244 assert(value->type() == T_INT, "Agreement."); 1245 val = value->get_int(); 1246 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1247 break; 1248 1249 default: 1250 ShouldNotReachHere(); 1251 } 1252 svIndex++; 1253 } 1254 return svIndex; 1255 } 1256 1257 // restore fields of all eliminated objects and arrays 1258 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1259 for (int i = 0; i < objects->length(); i++) { 1260 ObjectValue* sv = (ObjectValue*) objects->at(i); 1261 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1262 Handle obj = sv->value(); 1263 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1264 if (PrintDeoptimizationDetails) { 1265 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1266 } 1267 if (obj.is_null()) { 1268 continue; 1269 } 1270 #if INCLUDE_JVMCI || INCLUDE_AOT 1271 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1272 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1273 continue; 1274 } 1275 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1276 if (k->is_instance_klass()) { 1277 InstanceKlass* ik = InstanceKlass::cast(k); 1278 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1279 } else if (k->is_typeArray_klass()) { 1280 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1281 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1282 } else if (k->is_objArray_klass()) { 1283 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1284 } 1285 } 1286 } 1287 1288 1289 // relock objects for which synchronization was eliminated 1290 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1291 for (int i = 0; i < monitors->length(); i++) { 1292 MonitorInfo* mon_info = monitors->at(i); 1293 if (mon_info->eliminated()) { 1294 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1295 if (!mon_info->owner_is_scalar_replaced()) { 1296 Handle obj(thread, mon_info->owner()); 1297 markWord mark = obj->mark(); 1298 if (UseBiasedLocking && mark.has_bias_pattern()) { 1299 // New allocated objects may have the mark set to anonymously biased. 1300 // Also the deoptimized method may called methods with synchronization 1301 // where the thread-local object is bias locked to the current thread. 1302 assert(mark.is_biased_anonymously() || 1303 mark.biased_locker() == thread, "should be locked to current thread"); 1304 // Reset mark word to unbiased prototype. 1305 markWord unbiased_prototype = markWord::prototype().set_age(mark.age()); 1306 obj->set_mark(unbiased_prototype); 1307 } 1308 BasicLock* lock = mon_info->lock(); 1309 ObjectSynchronizer::enter(obj, lock, thread); 1310 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1311 } 1312 } 1313 } 1314 } 1315 1316 1317 #ifndef PRODUCT 1318 // print information about reallocated objects 1319 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1320 fieldDescriptor fd; 1321 1322 for (int i = 0; i < objects->length(); i++) { 1323 ObjectValue* sv = (ObjectValue*) objects->at(i); 1324 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1325 Handle obj = sv->value(); 1326 1327 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1328 k->print_value(); 1329 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1330 if (obj.is_null()) { 1331 tty->print(" allocation failed"); 1332 } else { 1333 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1334 } 1335 tty->cr(); 1336 1337 if (Verbose && !obj.is_null()) { 1338 k->oop_print_on(obj(), tty); 1339 } 1340 } 1341 } 1342 #endif 1343 #endif // COMPILER2_OR_JVMCI 1344 1345 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1346 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1347 1348 #ifndef PRODUCT 1349 if (PrintDeoptimizationDetails) { 1350 ttyLocker ttyl; 1351 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1352 fr.print_on(tty); 1353 tty->print_cr(" Virtual frames (innermost first):"); 1354 for (int index = 0; index < chunk->length(); index++) { 1355 compiledVFrame* vf = chunk->at(index); 1356 tty->print(" %2d - ", index); 1357 vf->print_value(); 1358 int bci = chunk->at(index)->raw_bci(); 1359 const char* code_name; 1360 if (bci == SynchronizationEntryBCI) { 1361 code_name = "sync entry"; 1362 } else { 1363 Bytecodes::Code code = vf->method()->code_at(bci); 1364 code_name = Bytecodes::name(code); 1365 } 1366 tty->print(" - %s", code_name); 1367 tty->print_cr(" @ bci %d ", bci); 1368 if (Verbose) { 1369 vf->print(); 1370 tty->cr(); 1371 } 1372 } 1373 } 1374 #endif 1375 1376 // Register map for next frame (used for stack crawl). We capture 1377 // the state of the deopt'ing frame's caller. Thus if we need to 1378 // stuff a C2I adapter we can properly fill in the callee-save 1379 // register locations. 1380 frame caller = fr.sender(reg_map); 1381 int frame_size = caller.sp() - fr.sp(); 1382 1383 frame sender = caller; 1384 1385 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1386 // the vframeArray containing the unpacking information is allocated in the C heap. 1387 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1388 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1389 1390 // Compare the vframeArray to the collected vframes 1391 assert(array->structural_compare(thread, chunk), "just checking"); 1392 1393 #ifndef PRODUCT 1394 if (PrintDeoptimizationDetails) { 1395 ttyLocker ttyl; 1396 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1397 } 1398 #endif // PRODUCT 1399 1400 return array; 1401 } 1402 1403 #if COMPILER2_OR_JVMCI 1404 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1405 // Reallocation of some scalar replaced objects failed. Record 1406 // that we need to pop all the interpreter frames for the 1407 // deoptimized compiled frame. 1408 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1409 thread->set_frames_to_pop_failed_realloc(array->frames()); 1410 // Unlock all monitors here otherwise the interpreter will see a 1411 // mix of locked and unlocked monitors (because of failed 1412 // reallocations of synchronized objects) and be confused. 1413 for (int i = 0; i < array->frames(); i++) { 1414 MonitorChunk* monitors = array->element(i)->monitors(); 1415 if (monitors != NULL) { 1416 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1417 BasicObjectLock* src = monitors->at(j); 1418 if (src->obj() != NULL) { 1419 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1420 } 1421 } 1422 array->element(i)->free_monitors(thread); 1423 #ifdef ASSERT 1424 array->element(i)->set_removed_monitors(); 1425 #endif 1426 } 1427 } 1428 } 1429 #endif 1430 1431 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1432 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1433 Thread* thread = Thread::current(); 1434 for (int i = 0; i < monitors->length(); i++) { 1435 MonitorInfo* mon_info = monitors->at(i); 1436 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1437 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1438 } 1439 } 1440 } 1441 1442 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) { 1443 // Unfortunately we don't have a RegisterMap available in most of 1444 // the places we want to call this routine so we need to walk the 1445 // stack again to update the register map. 1446 if (map == NULL || !map->update_map()) { 1447 StackFrameStream sfs(thread, true); 1448 bool found = false; 1449 while (!found && !sfs.is_done()) { 1450 frame* cur = sfs.current(); 1451 sfs.next(); 1452 found = cur->id() == fr.id(); 1453 } 1454 assert(found, "frame to be deoptimized not found on target thread's stack"); 1455 map = sfs.register_map(); 1456 } 1457 1458 vframe* vf = vframe::new_vframe(&fr, map, thread); 1459 compiledVFrame* cvf = compiledVFrame::cast(vf); 1460 // Revoke monitors' biases in all scopes 1461 while (!cvf->is_top()) { 1462 collect_monitors(cvf, objects_to_revoke); 1463 cvf = compiledVFrame::cast(cvf->sender()); 1464 } 1465 collect_monitors(cvf, objects_to_revoke); 1466 } 1467 1468 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) { 1469 if (!UseBiasedLocking) { 1470 return; 1471 } 1472 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1473 get_monitors_from_stack(objects_to_revoke, thread, fr, map); 1474 1475 int len = objects_to_revoke->length(); 1476 for (int i = 0; i < len; i++) { 1477 oop obj = (objects_to_revoke->at(i))(); 1478 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread); 1479 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1480 } 1481 } 1482 1483 1484 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1485 assert(fr.can_be_deoptimized(), "checking frame type"); 1486 1487 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1488 1489 if (LogCompilation && xtty != NULL) { 1490 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1491 assert(cm != NULL, "only compiled methods can deopt"); 1492 1493 ttyLocker ttyl; 1494 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1495 cm->log_identity(xtty); 1496 xtty->end_head(); 1497 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1498 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1499 xtty->method(sd->method()); 1500 xtty->end_elem(); 1501 if (sd->is_top()) break; 1502 } 1503 xtty->tail("deoptimized"); 1504 } 1505 1506 // Patch the compiled method so that when execution returns to it we will 1507 // deopt the execution state and return to the interpreter. 1508 fr.deoptimize(thread); 1509 } 1510 1511 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1512 // Deoptimize only if the frame comes from compile code. 1513 // Do not deoptimize the frame which is already patched 1514 // during the execution of the loops below. 1515 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1516 return; 1517 } 1518 ResourceMark rm; 1519 DeoptimizationMarker dm; 1520 deoptimize_single_frame(thread, fr, reason); 1521 } 1522 1523 #if INCLUDE_JVMCI 1524 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1525 // there is no exception handler for this pc => deoptimize 1526 cm->make_not_entrant(); 1527 1528 // Use Deoptimization::deoptimize for all of its side-effects: 1529 // revoking biases of monitors, gathering traps statistics, logging... 1530 // it also patches the return pc but we do not care about that 1531 // since we return a continuation to the deopt_blob below. 1532 JavaThread* thread = JavaThread::current(); 1533 RegisterMap reg_map(thread, UseBiasedLocking); 1534 frame runtime_frame = thread->last_frame(); 1535 frame caller_frame = runtime_frame.sender(®_map); 1536 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1537 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1538 1539 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1540 if (trap_mdo != NULL) { 1541 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1542 } 1543 1544 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1545 } 1546 #endif 1547 1548 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1549 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1550 "can only deoptimize other thread at a safepoint"); 1551 // Compute frame and register map based on thread and sp. 1552 RegisterMap reg_map(thread, UseBiasedLocking); 1553 frame fr = thread->last_frame(); 1554 while (fr.id() != id) { 1555 fr = fr.sender(®_map); 1556 } 1557 deoptimize(thread, fr, ®_map, reason); 1558 } 1559 1560 1561 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1562 if (thread == Thread::current()) { 1563 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1564 } else { 1565 VM_DeoptimizeFrame deopt(thread, id, reason); 1566 VMThread::execute(&deopt); 1567 } 1568 } 1569 1570 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1571 deoptimize_frame(thread, id, Reason_constraint); 1572 } 1573 1574 // JVMTI PopFrame support 1575 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1576 { 1577 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1578 } 1579 JRT_END 1580 1581 MethodData* 1582 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1583 bool create_if_missing) { 1584 Thread* THREAD = thread; 1585 MethodData* mdo = m()->method_data(); 1586 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1587 // Build an MDO. Ignore errors like OutOfMemory; 1588 // that simply means we won't have an MDO to update. 1589 Method::build_interpreter_method_data(m, THREAD); 1590 if (HAS_PENDING_EXCEPTION) { 1591 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1592 CLEAR_PENDING_EXCEPTION; 1593 } 1594 mdo = m()->method_data(); 1595 } 1596 return mdo; 1597 } 1598 1599 #if COMPILER2_OR_JVMCI 1600 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1601 // in case of an unresolved klass entry, load the class. 1602 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1603 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1604 return; 1605 } 1606 1607 if (!constant_pool->tag_at(index).is_symbol()) return; 1608 1609 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1610 Symbol* symbol = constant_pool->symbol_at(index); 1611 1612 // class name? 1613 if (symbol->char_at(0) != '(') { 1614 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1615 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1616 return; 1617 } 1618 1619 // then it must be a signature! 1620 ResourceMark rm(THREAD); 1621 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1622 if (ss.is_object()) { 1623 Symbol* class_name = ss.as_symbol(); 1624 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1625 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1626 } 1627 } 1628 } 1629 1630 1631 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1632 EXCEPTION_MARK; 1633 load_class_by_index(constant_pool, index, THREAD); 1634 if (HAS_PENDING_EXCEPTION) { 1635 // Exception happened during classloading. We ignore the exception here, since it 1636 // is going to be rethrown since the current activation is going to be deoptimized and 1637 // the interpreter will re-execute the bytecode. 1638 CLEAR_PENDING_EXCEPTION; 1639 // Class loading called java code which may have caused a stack 1640 // overflow. If the exception was thrown right before the return 1641 // to the runtime the stack is no longer guarded. Reguard the 1642 // stack otherwise if we return to the uncommon trap blob and the 1643 // stack bang causes a stack overflow we crash. 1644 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1645 JavaThread* thread = (JavaThread*)THREAD; 1646 bool guard_pages_enabled = thread->stack_guards_enabled(); 1647 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1648 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1649 } 1650 } 1651 1652 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1653 HandleMark hm; 1654 1655 // uncommon_trap() is called at the beginning of the uncommon trap 1656 // handler. Note this fact before we start generating temporary frames 1657 // that can confuse an asynchronous stack walker. This counter is 1658 // decremented at the end of unpack_frames(). 1659 thread->inc_in_deopt_handler(); 1660 1661 // We need to update the map if we have biased locking. 1662 #if INCLUDE_JVMCI 1663 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1664 RegisterMap reg_map(thread, true); 1665 #else 1666 RegisterMap reg_map(thread, UseBiasedLocking); 1667 #endif 1668 frame stub_frame = thread->last_frame(); 1669 frame fr = stub_frame.sender(®_map); 1670 // Make sure the calling nmethod is not getting deoptimized and removed 1671 // before we are done with it. 1672 nmethodLocker nl(fr.pc()); 1673 1674 // Log a message 1675 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1676 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1677 1678 { 1679 ResourceMark rm; 1680 1681 DeoptReason reason = trap_request_reason(trap_request); 1682 DeoptAction action = trap_request_action(trap_request); 1683 #if INCLUDE_JVMCI 1684 int debug_id = trap_request_debug_id(trap_request); 1685 #endif 1686 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1687 1688 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1689 compiledVFrame* cvf = compiledVFrame::cast(vf); 1690 1691 CompiledMethod* nm = cvf->code(); 1692 1693 ScopeDesc* trap_scope = cvf->scope(); 1694 1695 if (TraceDeoptimization) { 1696 ttyLocker ttyl; 1697 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() 1698 #if INCLUDE_JVMCI 1699 , debug_id 1700 #endif 1701 ); 1702 } 1703 1704 methodHandle trap_method(THREAD, trap_scope->method()); 1705 int trap_bci = trap_scope->bci(); 1706 #if INCLUDE_JVMCI 1707 jlong speculation = thread->pending_failed_speculation(); 1708 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods 1709 nm->as_nmethod()->update_speculation(thread); 1710 } else { 1711 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 1712 } 1713 1714 if (trap_bci == SynchronizationEntryBCI) { 1715 trap_bci = 0; 1716 thread->set_pending_monitorenter(true); 1717 } 1718 1719 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1720 thread->set_pending_transfer_to_interpreter(true); 1721 } 1722 #endif 1723 1724 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1725 // Record this event in the histogram. 1726 gather_statistics(reason, action, trap_bc); 1727 1728 // Ensure that we can record deopt. history: 1729 // Need MDO to record RTM code generation state. 1730 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1731 1732 methodHandle profiled_method; 1733 #if INCLUDE_JVMCI 1734 if (nm->is_compiled_by_jvmci()) { 1735 profiled_method = methodHandle(THREAD, nm->method()); 1736 } else { 1737 profiled_method = trap_method; 1738 } 1739 #else 1740 profiled_method = trap_method; 1741 #endif 1742 1743 MethodData* trap_mdo = 1744 get_method_data(thread, profiled_method, create_if_missing); 1745 1746 // Log a message 1747 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1748 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1749 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1750 1751 // Print a bunch of diagnostics, if requested. 1752 if (TraceDeoptimization || LogCompilation) { 1753 ResourceMark rm; 1754 ttyLocker ttyl; 1755 char buf[100]; 1756 if (xtty != NULL) { 1757 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1758 os::current_thread_id(), 1759 format_trap_request(buf, sizeof(buf), trap_request)); 1760 #if INCLUDE_JVMCI 1761 if (speculation != 0) { 1762 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 1763 } 1764 #endif 1765 nm->log_identity(xtty); 1766 } 1767 Symbol* class_name = NULL; 1768 bool unresolved = false; 1769 if (unloaded_class_index >= 0) { 1770 constantPoolHandle constants (THREAD, trap_method->constants()); 1771 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1772 class_name = constants->klass_name_at(unloaded_class_index); 1773 unresolved = true; 1774 if (xtty != NULL) 1775 xtty->print(" unresolved='1'"); 1776 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1777 class_name = constants->symbol_at(unloaded_class_index); 1778 } 1779 if (xtty != NULL) 1780 xtty->name(class_name); 1781 } 1782 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1783 // Dump the relevant MDO state. 1784 // This is the deopt count for the current reason, any previous 1785 // reasons or recompiles seen at this point. 1786 int dcnt = trap_mdo->trap_count(reason); 1787 if (dcnt != 0) 1788 xtty->print(" count='%d'", dcnt); 1789 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1790 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1791 if (dos != 0) { 1792 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1793 if (trap_state_is_recompiled(dos)) { 1794 int recnt2 = trap_mdo->overflow_recompile_count(); 1795 if (recnt2 != 0) 1796 xtty->print(" recompiles2='%d'", recnt2); 1797 } 1798 } 1799 } 1800 if (xtty != NULL) { 1801 xtty->stamp(); 1802 xtty->end_head(); 1803 } 1804 if (TraceDeoptimization) { // make noise on the tty 1805 tty->print("Uncommon trap occurred in"); 1806 nm->method()->print_short_name(tty); 1807 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1808 #if INCLUDE_JVMCI 1809 if (nm->is_nmethod()) { 1810 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 1811 if (installed_code_name != NULL) { 1812 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1813 } 1814 } 1815 #endif 1816 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1817 p2i(fr.pc()), 1818 os::current_thread_id(), 1819 trap_reason_name(reason), 1820 trap_action_name(action), 1821 unloaded_class_index 1822 #if INCLUDE_JVMCI 1823 , debug_id 1824 #endif 1825 ); 1826 if (class_name != NULL) { 1827 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1828 class_name->print_symbol_on(tty); 1829 } 1830 tty->cr(); 1831 } 1832 if (xtty != NULL) { 1833 // Log the precise location of the trap. 1834 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1835 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1836 xtty->method(sd->method()); 1837 xtty->end_elem(); 1838 if (sd->is_top()) break; 1839 } 1840 xtty->tail("uncommon_trap"); 1841 } 1842 } 1843 // (End diagnostic printout.) 1844 1845 // Load class if necessary 1846 if (unloaded_class_index >= 0) { 1847 constantPoolHandle constants(THREAD, trap_method->constants()); 1848 load_class_by_index(constants, unloaded_class_index); 1849 } 1850 1851 // Flush the nmethod if necessary and desirable. 1852 // 1853 // We need to avoid situations where we are re-flushing the nmethod 1854 // because of a hot deoptimization site. Repeated flushes at the same 1855 // point need to be detected by the compiler and avoided. If the compiler 1856 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1857 // module must take measures to avoid an infinite cycle of recompilation 1858 // and deoptimization. There are several such measures: 1859 // 1860 // 1. If a recompilation is ordered a second time at some site X 1861 // and for the same reason R, the action is adjusted to 'reinterpret', 1862 // to give the interpreter time to exercise the method more thoroughly. 1863 // If this happens, the method's overflow_recompile_count is incremented. 1864 // 1865 // 2. If the compiler fails to reduce the deoptimization rate, then 1866 // the method's overflow_recompile_count will begin to exceed the set 1867 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1868 // is adjusted to 'make_not_compilable', and the method is abandoned 1869 // to the interpreter. This is a performance hit for hot methods, 1870 // but is better than a disastrous infinite cycle of recompilations. 1871 // (Actually, only the method containing the site X is abandoned.) 1872 // 1873 // 3. In parallel with the previous measures, if the total number of 1874 // recompilations of a method exceeds the much larger set limit 1875 // PerMethodRecompilationCutoff, the method is abandoned. 1876 // This should only happen if the method is very large and has 1877 // many "lukewarm" deoptimizations. The code which enforces this 1878 // limit is elsewhere (class nmethod, class Method). 1879 // 1880 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1881 // to recompile at each bytecode independently of the per-BCI cutoff. 1882 // 1883 // The decision to update code is up to the compiler, and is encoded 1884 // in the Action_xxx code. If the compiler requests Action_none 1885 // no trap state is changed, no compiled code is changed, and the 1886 // computation suffers along in the interpreter. 1887 // 1888 // The other action codes specify various tactics for decompilation 1889 // and recompilation. Action_maybe_recompile is the loosest, and 1890 // allows the compiled code to stay around until enough traps are seen, 1891 // and until the compiler gets around to recompiling the trapping method. 1892 // 1893 // The other actions cause immediate removal of the present code. 1894 1895 // Traps caused by injected profile shouldn't pollute trap counts. 1896 bool injected_profile_trap = trap_method->has_injected_profile() && 1897 (reason == Reason_intrinsic || reason == Reason_unreached); 1898 1899 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1900 bool make_not_entrant = false; 1901 bool make_not_compilable = false; 1902 bool reprofile = false; 1903 switch (action) { 1904 case Action_none: 1905 // Keep the old code. 1906 update_trap_state = false; 1907 break; 1908 case Action_maybe_recompile: 1909 // Do not need to invalidate the present code, but we can 1910 // initiate another 1911 // Start compiler without (necessarily) invalidating the nmethod. 1912 // The system will tolerate the old code, but new code should be 1913 // generated when possible. 1914 break; 1915 case Action_reinterpret: 1916 // Go back into the interpreter for a while, and then consider 1917 // recompiling form scratch. 1918 make_not_entrant = true; 1919 // Reset invocation counter for outer most method. 1920 // This will allow the interpreter to exercise the bytecodes 1921 // for a while before recompiling. 1922 // By contrast, Action_make_not_entrant is immediate. 1923 // 1924 // Note that the compiler will track null_check, null_assert, 1925 // range_check, and class_check events and log them as if they 1926 // had been traps taken from compiled code. This will update 1927 // the MDO trap history so that the next compilation will 1928 // properly detect hot trap sites. 1929 reprofile = true; 1930 break; 1931 case Action_make_not_entrant: 1932 // Request immediate recompilation, and get rid of the old code. 1933 // Make them not entrant, so next time they are called they get 1934 // recompiled. Unloaded classes are loaded now so recompile before next 1935 // time they are called. Same for uninitialized. The interpreter will 1936 // link the missing class, if any. 1937 make_not_entrant = true; 1938 break; 1939 case Action_make_not_compilable: 1940 // Give up on compiling this method at all. 1941 make_not_entrant = true; 1942 make_not_compilable = true; 1943 break; 1944 default: 1945 ShouldNotReachHere(); 1946 } 1947 1948 // Setting +ProfileTraps fixes the following, on all platforms: 1949 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1950 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1951 // recompile relies on a MethodData* to record heroic opt failures. 1952 1953 // Whether the interpreter is producing MDO data or not, we also need 1954 // to use the MDO to detect hot deoptimization points and control 1955 // aggressive optimization. 1956 bool inc_recompile_count = false; 1957 ProfileData* pdata = NULL; 1958 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1959 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1960 uint this_trap_count = 0; 1961 bool maybe_prior_trap = false; 1962 bool maybe_prior_recompile = false; 1963 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1964 #if INCLUDE_JVMCI 1965 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1966 #endif 1967 nm->method(), 1968 //outputs: 1969 this_trap_count, 1970 maybe_prior_trap, 1971 maybe_prior_recompile); 1972 // Because the interpreter also counts null, div0, range, and class 1973 // checks, these traps from compiled code are double-counted. 1974 // This is harmless; it just means that the PerXTrapLimit values 1975 // are in effect a little smaller than they look. 1976 1977 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1978 if (per_bc_reason != Reason_none) { 1979 // Now take action based on the partially known per-BCI history. 1980 if (maybe_prior_trap 1981 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1982 // If there are too many traps at this BCI, force a recompile. 1983 // This will allow the compiler to see the limit overflow, and 1984 // take corrective action, if possible. The compiler generally 1985 // does not use the exact PerBytecodeTrapLimit value, but instead 1986 // changes its tactics if it sees any traps at all. This provides 1987 // a little hysteresis, delaying a recompile until a trap happens 1988 // several times. 1989 // 1990 // Actually, since there is only one bit of counter per BCI, 1991 // the possible per-BCI counts are {0,1,(per-method count)}. 1992 // This produces accurate results if in fact there is only 1993 // one hot trap site, but begins to get fuzzy if there are 1994 // many sites. For example, if there are ten sites each 1995 // trapping two or more times, they each get the blame for 1996 // all of their traps. 1997 make_not_entrant = true; 1998 } 1999 2000 // Detect repeated recompilation at the same BCI, and enforce a limit. 2001 if (make_not_entrant && maybe_prior_recompile) { 2002 // More than one recompile at this point. 2003 inc_recompile_count = maybe_prior_trap; 2004 } 2005 } else { 2006 // For reasons which are not recorded per-bytecode, we simply 2007 // force recompiles unconditionally. 2008 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2009 make_not_entrant = true; 2010 } 2011 2012 // Go back to the compiler if there are too many traps in this method. 2013 if (this_trap_count >= per_method_trap_limit(reason)) { 2014 // If there are too many traps in this method, force a recompile. 2015 // This will allow the compiler to see the limit overflow, and 2016 // take corrective action, if possible. 2017 // (This condition is an unlikely backstop only, because the 2018 // PerBytecodeTrapLimit is more likely to take effect first, 2019 // if it is applicable.) 2020 make_not_entrant = true; 2021 } 2022 2023 // Here's more hysteresis: If there has been a recompile at 2024 // this trap point already, run the method in the interpreter 2025 // for a while to exercise it more thoroughly. 2026 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2027 reprofile = true; 2028 } 2029 } 2030 2031 // Take requested actions on the method: 2032 2033 // Recompile 2034 if (make_not_entrant) { 2035 if (!nm->make_not_entrant()) { 2036 return; // the call did not change nmethod's state 2037 } 2038 2039 if (pdata != NULL) { 2040 // Record the recompilation event, if any. 2041 int tstate0 = pdata->trap_state(); 2042 int tstate1 = trap_state_set_recompiled(tstate0, true); 2043 if (tstate1 != tstate0) 2044 pdata->set_trap_state(tstate1); 2045 } 2046 2047 #if INCLUDE_RTM_OPT 2048 // Restart collecting RTM locking abort statistic if the method 2049 // is recompiled for a reason other than RTM state change. 2050 // Assume that in new recompiled code the statistic could be different, 2051 // for example, due to different inlining. 2052 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2053 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2054 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2055 } 2056 #endif 2057 // For code aging we count traps separately here, using make_not_entrant() 2058 // as a guard against simultaneous deopts in multiple threads. 2059 if (reason == Reason_tenured && trap_mdo != NULL) { 2060 trap_mdo->inc_tenure_traps(); 2061 } 2062 } 2063 2064 if (inc_recompile_count) { 2065 trap_mdo->inc_overflow_recompile_count(); 2066 if ((uint)trap_mdo->overflow_recompile_count() > 2067 (uint)PerBytecodeRecompilationCutoff) { 2068 // Give up on the method containing the bad BCI. 2069 if (trap_method() == nm->method()) { 2070 make_not_compilable = true; 2071 } else { 2072 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2073 // But give grace to the enclosing nm->method(). 2074 } 2075 } 2076 } 2077 2078 // Reprofile 2079 if (reprofile) { 2080 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2081 } 2082 2083 // Give up compiling 2084 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2085 assert(make_not_entrant, "consistent"); 2086 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2087 } 2088 2089 } // Free marked resources 2090 2091 } 2092 JRT_END 2093 2094 ProfileData* 2095 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2096 int trap_bci, 2097 Deoptimization::DeoptReason reason, 2098 bool update_total_trap_count, 2099 #if INCLUDE_JVMCI 2100 bool is_osr, 2101 #endif 2102 Method* compiled_method, 2103 //outputs: 2104 uint& ret_this_trap_count, 2105 bool& ret_maybe_prior_trap, 2106 bool& ret_maybe_prior_recompile) { 2107 bool maybe_prior_trap = false; 2108 bool maybe_prior_recompile = false; 2109 uint this_trap_count = 0; 2110 if (update_total_trap_count) { 2111 uint idx = reason; 2112 #if INCLUDE_JVMCI 2113 if (is_osr) { 2114 idx += Reason_LIMIT; 2115 } 2116 #endif 2117 uint prior_trap_count = trap_mdo->trap_count(idx); 2118 this_trap_count = trap_mdo->inc_trap_count(idx); 2119 2120 // If the runtime cannot find a place to store trap history, 2121 // it is estimated based on the general condition of the method. 2122 // If the method has ever been recompiled, or has ever incurred 2123 // a trap with the present reason , then this BCI is assumed 2124 // (pessimistically) to be the culprit. 2125 maybe_prior_trap = (prior_trap_count != 0); 2126 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2127 } 2128 ProfileData* pdata = NULL; 2129 2130 2131 // For reasons which are recorded per bytecode, we check per-BCI data. 2132 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2133 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2134 if (per_bc_reason != Reason_none) { 2135 // Find the profile data for this BCI. If there isn't one, 2136 // try to allocate one from the MDO's set of spares. 2137 // This will let us detect a repeated trap at this point. 2138 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2139 2140 if (pdata != NULL) { 2141 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2142 if (LogCompilation && xtty != NULL) { 2143 ttyLocker ttyl; 2144 // no more room for speculative traps in this MDO 2145 xtty->elem("speculative_traps_oom"); 2146 } 2147 } 2148 // Query the trap state of this profile datum. 2149 int tstate0 = pdata->trap_state(); 2150 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2151 maybe_prior_trap = false; 2152 if (!trap_state_is_recompiled(tstate0)) 2153 maybe_prior_recompile = false; 2154 2155 // Update the trap state of this profile datum. 2156 int tstate1 = tstate0; 2157 // Record the reason. 2158 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2159 // Store the updated state on the MDO, for next time. 2160 if (tstate1 != tstate0) 2161 pdata->set_trap_state(tstate1); 2162 } else { 2163 if (LogCompilation && xtty != NULL) { 2164 ttyLocker ttyl; 2165 // Missing MDP? Leave a small complaint in the log. 2166 xtty->elem("missing_mdp bci='%d'", trap_bci); 2167 } 2168 } 2169 } 2170 2171 // Return results: 2172 ret_this_trap_count = this_trap_count; 2173 ret_maybe_prior_trap = maybe_prior_trap; 2174 ret_maybe_prior_recompile = maybe_prior_recompile; 2175 return pdata; 2176 } 2177 2178 void 2179 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2180 ResourceMark rm; 2181 // Ignored outputs: 2182 uint ignore_this_trap_count; 2183 bool ignore_maybe_prior_trap; 2184 bool ignore_maybe_prior_recompile; 2185 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2186 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2187 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2188 query_update_method_data(trap_mdo, trap_bci, 2189 (DeoptReason)reason, 2190 update_total_counts, 2191 #if INCLUDE_JVMCI 2192 false, 2193 #endif 2194 NULL, 2195 ignore_this_trap_count, 2196 ignore_maybe_prior_trap, 2197 ignore_maybe_prior_recompile); 2198 } 2199 2200 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2201 if (TraceDeoptimization) { 2202 tty->print("Uncommon trap "); 2203 } 2204 // Still in Java no safepoints 2205 { 2206 // This enters VM and may safepoint 2207 uncommon_trap_inner(thread, trap_request); 2208 } 2209 return fetch_unroll_info_helper(thread, exec_mode); 2210 } 2211 2212 // Local derived constants. 2213 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2214 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2215 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2216 2217 //---------------------------trap_state_reason--------------------------------- 2218 Deoptimization::DeoptReason 2219 Deoptimization::trap_state_reason(int trap_state) { 2220 // This assert provides the link between the width of DataLayout::trap_bits 2221 // and the encoding of "recorded" reasons. It ensures there are enough 2222 // bits to store all needed reasons in the per-BCI MDO profile. 2223 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2224 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2225 trap_state -= recompile_bit; 2226 if (trap_state == DS_REASON_MASK) { 2227 return Reason_many; 2228 } else { 2229 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2230 return (DeoptReason)trap_state; 2231 } 2232 } 2233 //-------------------------trap_state_has_reason------------------------------- 2234 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2235 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2236 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2237 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2238 trap_state -= recompile_bit; 2239 if (trap_state == DS_REASON_MASK) { 2240 return -1; // true, unspecifically (bottom of state lattice) 2241 } else if (trap_state == reason) { 2242 return 1; // true, definitely 2243 } else if (trap_state == 0) { 2244 return 0; // false, definitely (top of state lattice) 2245 } else { 2246 return 0; // false, definitely 2247 } 2248 } 2249 //-------------------------trap_state_add_reason------------------------------- 2250 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2251 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2252 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2253 trap_state -= recompile_bit; 2254 if (trap_state == DS_REASON_MASK) { 2255 return trap_state + recompile_bit; // already at state lattice bottom 2256 } else if (trap_state == reason) { 2257 return trap_state + recompile_bit; // the condition is already true 2258 } else if (trap_state == 0) { 2259 return reason + recompile_bit; // no condition has yet been true 2260 } else { 2261 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2262 } 2263 } 2264 //-----------------------trap_state_is_recompiled------------------------------ 2265 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2266 return (trap_state & DS_RECOMPILE_BIT) != 0; 2267 } 2268 //-----------------------trap_state_set_recompiled----------------------------- 2269 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2270 if (z) return trap_state | DS_RECOMPILE_BIT; 2271 else return trap_state & ~DS_RECOMPILE_BIT; 2272 } 2273 //---------------------------format_trap_state--------------------------------- 2274 // This is used for debugging and diagnostics, including LogFile output. 2275 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2276 int trap_state) { 2277 assert(buflen > 0, "sanity"); 2278 DeoptReason reason = trap_state_reason(trap_state); 2279 bool recomp_flag = trap_state_is_recompiled(trap_state); 2280 // Re-encode the state from its decoded components. 2281 int decoded_state = 0; 2282 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2283 decoded_state = trap_state_add_reason(decoded_state, reason); 2284 if (recomp_flag) 2285 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2286 // If the state re-encodes properly, format it symbolically. 2287 // Because this routine is used for debugging and diagnostics, 2288 // be robust even if the state is a strange value. 2289 size_t len; 2290 if (decoded_state != trap_state) { 2291 // Random buggy state that doesn't decode?? 2292 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2293 } else { 2294 len = jio_snprintf(buf, buflen, "%s%s", 2295 trap_reason_name(reason), 2296 recomp_flag ? " recompiled" : ""); 2297 } 2298 return buf; 2299 } 2300 2301 2302 //--------------------------------statics-------------------------------------- 2303 const char* Deoptimization::_trap_reason_name[] = { 2304 // Note: Keep this in sync. with enum DeoptReason. 2305 "none", 2306 "null_check", 2307 "null_assert" JVMCI_ONLY("_or_unreached0"), 2308 "range_check", 2309 "class_check", 2310 "array_check", 2311 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2312 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2313 "profile_predicate", 2314 "unloaded", 2315 "uninitialized", 2316 "initialized", 2317 "unreached", 2318 "unhandled", 2319 "constraint", 2320 "div0_check", 2321 "age", 2322 "predicate", 2323 "loop_limit_check", 2324 "speculate_class_check", 2325 "speculate_null_check", 2326 "speculate_null_assert", 2327 "rtm_state_change", 2328 "unstable_if", 2329 "unstable_fused_if", 2330 #if INCLUDE_JVMCI 2331 "aliasing", 2332 "transfer_to_interpreter", 2333 "not_compiled_exception_handler", 2334 "unresolved", 2335 "jsr_mismatch", 2336 #endif 2337 "tenured" 2338 }; 2339 const char* Deoptimization::_trap_action_name[] = { 2340 // Note: Keep this in sync. with enum DeoptAction. 2341 "none", 2342 "maybe_recompile", 2343 "reinterpret", 2344 "make_not_entrant", 2345 "make_not_compilable" 2346 }; 2347 2348 const char* Deoptimization::trap_reason_name(int reason) { 2349 // Check that every reason has a name 2350 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2351 2352 if (reason == Reason_many) return "many"; 2353 if ((uint)reason < Reason_LIMIT) 2354 return _trap_reason_name[reason]; 2355 static char buf[20]; 2356 sprintf(buf, "reason%d", reason); 2357 return buf; 2358 } 2359 const char* Deoptimization::trap_action_name(int action) { 2360 // Check that every action has a name 2361 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2362 2363 if ((uint)action < Action_LIMIT) 2364 return _trap_action_name[action]; 2365 static char buf[20]; 2366 sprintf(buf, "action%d", action); 2367 return buf; 2368 } 2369 2370 // This is used for debugging and diagnostics, including LogFile output. 2371 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2372 int trap_request) { 2373 jint unloaded_class_index = trap_request_index(trap_request); 2374 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2375 const char* action = trap_action_name(trap_request_action(trap_request)); 2376 #if INCLUDE_JVMCI 2377 int debug_id = trap_request_debug_id(trap_request); 2378 #endif 2379 size_t len; 2380 if (unloaded_class_index < 0) { 2381 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2382 reason, action 2383 #if INCLUDE_JVMCI 2384 ,debug_id 2385 #endif 2386 ); 2387 } else { 2388 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2389 reason, action, unloaded_class_index 2390 #if INCLUDE_JVMCI 2391 ,debug_id 2392 #endif 2393 ); 2394 } 2395 return buf; 2396 } 2397 2398 juint Deoptimization::_deoptimization_hist 2399 [Deoptimization::Reason_LIMIT] 2400 [1 + Deoptimization::Action_LIMIT] 2401 [Deoptimization::BC_CASE_LIMIT] 2402 = {0}; 2403 2404 enum { 2405 LSB_BITS = 8, 2406 LSB_MASK = right_n_bits(LSB_BITS) 2407 }; 2408 2409 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2410 Bytecodes::Code bc) { 2411 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2412 assert(action >= 0 && action < Action_LIMIT, "oob"); 2413 _deoptimization_hist[Reason_none][0][0] += 1; // total 2414 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2415 juint* cases = _deoptimization_hist[reason][1+action]; 2416 juint* bc_counter_addr = NULL; 2417 juint bc_counter = 0; 2418 // Look for an unused counter, or an exact match to this BC. 2419 if (bc != Bytecodes::_illegal) { 2420 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2421 juint* counter_addr = &cases[bc_case]; 2422 juint counter = *counter_addr; 2423 if ((counter == 0 && bc_counter_addr == NULL) 2424 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2425 // this counter is either free or is already devoted to this BC 2426 bc_counter_addr = counter_addr; 2427 bc_counter = counter | bc; 2428 } 2429 } 2430 } 2431 if (bc_counter_addr == NULL) { 2432 // Overflow, or no given bytecode. 2433 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2434 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2435 } 2436 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2437 } 2438 2439 jint Deoptimization::total_deoptimization_count() { 2440 return _deoptimization_hist[Reason_none][0][0]; 2441 } 2442 2443 void Deoptimization::print_statistics() { 2444 juint total = total_deoptimization_count(); 2445 juint account = total; 2446 if (total != 0) { 2447 ttyLocker ttyl; 2448 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2449 tty->print_cr("Deoptimization traps recorded:"); 2450 #define PRINT_STAT_LINE(name, r) \ 2451 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2452 PRINT_STAT_LINE("total", total); 2453 // For each non-zero entry in the histogram, print the reason, 2454 // the action, and (if specifically known) the type of bytecode. 2455 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2456 for (int action = 0; action < Action_LIMIT; action++) { 2457 juint* cases = _deoptimization_hist[reason][1+action]; 2458 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2459 juint counter = cases[bc_case]; 2460 if (counter != 0) { 2461 char name[1*K]; 2462 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2463 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2464 bc = Bytecodes::_illegal; 2465 sprintf(name, "%s/%s/%s", 2466 trap_reason_name(reason), 2467 trap_action_name(action), 2468 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2469 juint r = counter >> LSB_BITS; 2470 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2471 account -= r; 2472 } 2473 } 2474 } 2475 } 2476 if (account != 0) { 2477 PRINT_STAT_LINE("unaccounted", account); 2478 } 2479 #undef PRINT_STAT_LINE 2480 if (xtty != NULL) xtty->tail("statistics"); 2481 } 2482 } 2483 #else // COMPILER2_OR_JVMCI 2484 2485 2486 // Stubs for C1 only system. 2487 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2488 return false; 2489 } 2490 2491 const char* Deoptimization::trap_reason_name(int reason) { 2492 return "unknown"; 2493 } 2494 2495 void Deoptimization::print_statistics() { 2496 // no output 2497 } 2498 2499 void 2500 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2501 // no udpate 2502 } 2503 2504 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2505 return 0; 2506 } 2507 2508 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2509 Bytecodes::Code bc) { 2510 // no update 2511 } 2512 2513 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2514 int trap_state) { 2515 jio_snprintf(buf, buflen, "#%d", trap_state); 2516 return buf; 2517 } 2518 2519 #endif // COMPILER2_OR_JVMCI