1 /* 2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "code/codeCache.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/nmethod.hpp" 30 #include "code/pcDesc.hpp" 31 #include "code/scopeDesc.hpp" 32 #include "interpreter/bytecode.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "interpreter/oopMapCache.hpp" 35 #include "memory/allocation.inline.hpp" 36 #include "memory/oopFactory.hpp" 37 #include "memory/resourceArea.hpp" 38 #include "oops/method.hpp" 39 #include "oops/objArrayOop.inline.hpp" 40 #include "oops/oop.inline.hpp" 41 #include "oops/fieldStreams.hpp" 42 #include "oops/verifyOopClosure.hpp" 43 #include "prims/jvmtiThreadState.hpp" 44 #include "runtime/biasedLocking.hpp" 45 #include "runtime/compilationPolicy.hpp" 46 #include "runtime/deoptimization.hpp" 47 #include "runtime/interfaceSupport.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 #include "runtime/signature.hpp" 50 #include "runtime/stubRoutines.hpp" 51 #include "runtime/thread.hpp" 52 #include "runtime/vframe.hpp" 53 #include "runtime/vframeArray.hpp" 54 #include "runtime/vframe_hp.hpp" 55 #include "utilities/events.hpp" 56 #include "utilities/xmlstream.hpp" 57 58 #if INCLUDE_JVMCI 59 #include "jvmci/jvmciRuntime.hpp" 60 #include "jvmci/jvmciJavaClasses.hpp" 61 #endif 62 63 64 bool DeoptimizationMarker::_is_active = false; 65 66 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 67 int caller_adjustment, 68 int caller_actual_parameters, 69 int number_of_frames, 70 intptr_t* frame_sizes, 71 address* frame_pcs, 72 BasicType return_type, 73 int exec_mode) { 74 _size_of_deoptimized_frame = size_of_deoptimized_frame; 75 _caller_adjustment = caller_adjustment; 76 _caller_actual_parameters = caller_actual_parameters; 77 _number_of_frames = number_of_frames; 78 _frame_sizes = frame_sizes; 79 _frame_pcs = frame_pcs; 80 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 81 _return_type = return_type; 82 _initial_info = 0; 83 // PD (x86 only) 84 _counter_temp = 0; 85 _unpack_kind = exec_mode; 86 _sender_sp_temp = 0; 87 88 _total_frame_sizes = size_of_frames(); 89 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 90 } 91 92 93 Deoptimization::UnrollBlock::~UnrollBlock() { 94 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 95 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 96 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 97 } 98 99 100 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 101 assert(register_number < RegisterMap::reg_count, "checking register number"); 102 return &_register_block[register_number * 2]; 103 } 104 105 106 107 int Deoptimization::UnrollBlock::size_of_frames() const { 108 // Acount first for the adjustment of the initial frame 109 int result = _caller_adjustment; 110 for (int index = 0; index < number_of_frames(); index++) { 111 result += frame_sizes()[index]; 112 } 113 return result; 114 } 115 116 117 void Deoptimization::UnrollBlock::print() { 118 ttyLocker ttyl; 119 tty->print_cr("UnrollBlock"); 120 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 121 tty->print( " frame_sizes: "); 122 for (int index = 0; index < number_of_frames(); index++) { 123 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 124 } 125 tty->cr(); 126 } 127 128 129 // In order to make fetch_unroll_info work properly with escape 130 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 131 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 132 // of previously eliminated objects occurs in realloc_objects, which is 133 // called from the method fetch_unroll_info_helper below. 134 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 135 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 136 // but makes the entry a little slower. There is however a little dance we have to 137 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 138 139 // fetch_unroll_info() is called at the beginning of the deoptimization 140 // handler. Note this fact before we start generating temporary frames 141 // that can confuse an asynchronous stack walker. This counter is 142 // decremented at the end of unpack_frames(). 143 if (TraceDeoptimization) { 144 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 145 } 146 thread->inc_in_deopt_handler(); 147 148 return fetch_unroll_info_helper(thread, exec_mode); 149 JRT_END 150 151 152 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 153 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 154 155 // Note: there is a safepoint safety issue here. No matter whether we enter 156 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 157 // the vframeArray is created. 158 // 159 160 // Allocate our special deoptimization ResourceMark 161 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 162 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 163 thread->set_deopt_mark(dmark); 164 165 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 166 RegisterMap map(thread, true); 167 RegisterMap dummy_map(thread, false); 168 // Now get the deoptee with a valid map 169 frame deoptee = stub_frame.sender(&map); 170 // Set the deoptee nmethod 171 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 172 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 173 bool skip_internal = thread->deopt_nmethod() != NULL && !thread->deopt_nmethod()->compiler()->is_jvmci(); 174 175 if (VerifyStack) { 176 thread->validate_frame_layout(); 177 } 178 179 // Create a growable array of VFrames where each VFrame represents an inlined 180 // Java frame. This storage is allocated with the usual system arena. 181 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 182 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 183 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 184 while (!vf->is_top()) { 185 assert(vf->is_compiled_frame(), "Wrong frame type"); 186 chunk->push(compiledVFrame::cast(vf)); 187 vf = vf->sender(); 188 } 189 assert(vf->is_compiled_frame(), "Wrong frame type"); 190 chunk->push(compiledVFrame::cast(vf)); 191 192 ScopeDesc* trap_scope = chunk->at(0)->scope(); 193 Handle exceptionObject; 194 if (trap_scope->rethrow_exception()) { 195 if (PrintDeoptimizationDetails) { 196 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()); 197 } 198 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 199 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 200 ScopeValue* topOfStack = expressions->top(); 201 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 202 assert(exceptionObject() != NULL, "exception oop can not be null"); 203 } 204 205 bool realloc_failures = false; 206 207 #if defined(COMPILER2) || INCLUDE_JVMCI 208 // Reallocate the non-escaping objects and restore their fields. Then 209 // relock objects if synchronization on them was eliminated. 210 #ifndef INCLUDE_JVMCI 211 if (DoEscapeAnalysis || EliminateNestedLocks) { 212 if (EliminateAllocations) { 213 #endif // INCLUDE_JVMCI 214 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 215 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 216 217 // The flag return_oop() indicates call sites which return oop 218 // in compiled code. Such sites include java method calls, 219 // runtime calls (for example, used to allocate new objects/arrays 220 // on slow code path) and any other calls generated in compiled code. 221 // It is not guaranteed that we can get such information here only 222 // by analyzing bytecode in deoptimized frames. This is why this flag 223 // is set during method compilation (see Compile::Process_OopMap_Node()). 224 // If the previous frame was popped, we don't have a result. 225 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution(); 226 Handle return_value; 227 if (save_oop_result) { 228 // Reallocation may trigger GC. If deoptimization happened on return from 229 // call which returns oop we need to save it since it is not in oopmap. 230 oop result = deoptee.saved_oop_result(&map); 231 assert(result == NULL || result->is_oop(), "must be oop"); 232 return_value = Handle(thread, result); 233 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 234 if (TraceDeoptimization) { 235 ttyLocker ttyl; 236 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 237 } 238 } 239 if (objects != NULL) { 240 JRT_BLOCK 241 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD); 242 JRT_END 243 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 244 #ifndef PRODUCT 245 if (TraceDeoptimization) { 246 ttyLocker ttyl; 247 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 248 print_objects(objects, realloc_failures); 249 } 250 #endif 251 } 252 if (save_oop_result) { 253 // Restore result. 254 deoptee.set_saved_oop_result(&map, return_value()); 255 } 256 #ifndef INCLUDE_JVMCI 257 } 258 if (EliminateLocks) { 259 #endif // INCLUDE_JVMCI 260 #ifndef PRODUCT 261 bool first = true; 262 #endif 263 for (int i = 0; i < chunk->length(); i++) { 264 compiledVFrame* cvf = chunk->at(i); 265 assert (cvf->scope() != NULL,"expect only compiled java frames"); 266 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 267 if (monitors->is_nonempty()) { 268 relock_objects(monitors, thread, realloc_failures); 269 #ifndef PRODUCT 270 if (PrintDeoptimizationDetails) { 271 ttyLocker ttyl; 272 for (int j = 0; j < monitors->length(); j++) { 273 MonitorInfo* mi = monitors->at(j); 274 if (mi->eliminated()) { 275 if (first) { 276 first = false; 277 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 278 } 279 if (mi->owner_is_scalar_replaced()) { 280 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 281 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 282 } else { 283 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 284 } 285 } 286 } 287 } 288 #endif // !PRODUCT 289 } 290 } 291 #ifndef INCLUDE_JVMCI 292 } 293 } 294 #endif // INCLUDE_JVMCI 295 #endif // COMPILER2 || INCLUDE_JVMCI 296 297 // Ensure that no safepoint is taken after pointers have been stored 298 // in fields of rematerialized objects. If a safepoint occurs from here on 299 // out the java state residing in the vframeArray will be missed. 300 No_Safepoint_Verifier no_safepoint; 301 302 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 303 #if defined(COMPILER2) || INCLUDE_JVMCI 304 if (realloc_failures) { 305 pop_frames_failed_reallocs(thread, array); 306 } 307 #endif 308 309 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 310 thread->set_vframe_array_head(array); 311 312 // Now that the vframeArray has been created if we have any deferred local writes 313 // added by jvmti then we can free up that structure as the data is now in the 314 // vframeArray 315 316 if (thread->deferred_locals() != NULL) { 317 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 318 int i = 0; 319 do { 320 // Because of inlining we could have multiple vframes for a single frame 321 // and several of the vframes could have deferred writes. Find them all. 322 if (list->at(i)->id() == array->original().id()) { 323 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 324 list->remove_at(i); 325 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 326 delete dlv; 327 } else { 328 i++; 329 } 330 } while ( i < list->length() ); 331 if (list->length() == 0) { 332 thread->set_deferred_locals(NULL); 333 // free the list and elements back to C heap. 334 delete list; 335 } 336 337 } 338 339 #ifndef SHARK 340 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 341 CodeBlob* cb = stub_frame.cb(); 342 // Verify we have the right vframeArray 343 assert(cb->frame_size() >= 0, "Unexpected frame size"); 344 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 345 346 // If the deopt call site is a MethodHandle invoke call site we have 347 // to adjust the unpack_sp. 348 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 349 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 350 unpack_sp = deoptee.unextended_sp(); 351 352 #ifdef ASSERT 353 assert(cb->is_deoptimization_stub() || 354 cb->is_uncommon_trap_stub() || 355 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 356 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 357 "unexpected code blob: %s", cb->name()); 358 #endif 359 #else 360 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 361 #endif // !SHARK 362 363 // This is a guarantee instead of an assert because if vframe doesn't match 364 // we will unpack the wrong deoptimized frame and wind up in strange places 365 // where it will be very difficult to figure out what went wrong. Better 366 // to die an early death here than some very obscure death later when the 367 // trail is cold. 368 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 369 // in that it will fail to detect a problem when there is one. This needs 370 // more work in tiger timeframe. 371 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 372 373 int number_of_frames = array->frames(); 374 375 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 376 // virtual activation, which is the reverse of the elements in the vframes array. 377 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 378 // +1 because we always have an interpreter return address for the final slot. 379 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 380 int popframe_extra_args = 0; 381 // Create an interpreter return address for the stub to use as its return 382 // address so the skeletal frames are perfectly walkable 383 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 384 385 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 386 // activation be put back on the expression stack of the caller for reexecution 387 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 388 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 389 } 390 391 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 392 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 393 // than simply use array->sender.pc(). This requires us to walk the current set of frames 394 // 395 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 396 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 397 398 // It's possible that the number of parameters at the call site is 399 // different than number of arguments in the callee when method 400 // handles are used. If the caller is interpreted get the real 401 // value so that the proper amount of space can be added to it's 402 // frame. 403 bool caller_was_method_handle = false; 404 if (deopt_sender.is_interpreted_frame()) { 405 methodHandle method = deopt_sender.interpreter_frame_method(); 406 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 407 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 408 // Method handle invokes may involve fairly arbitrary chains of 409 // calls so it's impossible to know how much actual space the 410 // caller has for locals. 411 caller_was_method_handle = true; 412 } 413 } 414 415 // 416 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 417 // frame_sizes/frame_pcs[1] next oldest frame (int) 418 // frame_sizes/frame_pcs[n] youngest frame (int) 419 // 420 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 421 // owns the space for the return address to it's caller). Confusing ain't it. 422 // 423 // The vframe array can address vframes with indices running from 424 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 425 // When we create the skeletal frames we need the oldest frame to be in the zero slot 426 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 427 // so things look a little strange in this loop. 428 // 429 int callee_parameters = 0; 430 int callee_locals = 0; 431 for (int index = 0; index < array->frames(); index++ ) { 432 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 433 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 434 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 435 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 436 callee_locals, 437 index == 0, 438 popframe_extra_args); 439 // This pc doesn't have to be perfect just good enough to identify the frame 440 // as interpreted so the skeleton frame will be walkable 441 // The correct pc will be set when the skeleton frame is completely filled out 442 // The final pc we store in the loop is wrong and will be overwritten below 443 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 444 445 callee_parameters = array->element(index)->method()->size_of_parameters(); 446 callee_locals = array->element(index)->method()->max_locals(); 447 popframe_extra_args = 0; 448 } 449 450 // Compute whether the root vframe returns a float or double value. 451 BasicType return_type; 452 { 453 HandleMark hm; 454 methodHandle method(thread, array->element(0)->method()); 455 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 456 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 457 } 458 459 // Compute information for handling adapters and adjusting the frame size of the caller. 460 int caller_adjustment = 0; 461 462 // Compute the amount the oldest interpreter frame will have to adjust 463 // its caller's stack by. If the caller is a compiled frame then 464 // we pretend that the callee has no parameters so that the 465 // extension counts for the full amount of locals and not just 466 // locals-parms. This is because without a c2i adapter the parm 467 // area as created by the compiled frame will not be usable by 468 // the interpreter. (Depending on the calling convention there 469 // may not even be enough space). 470 471 // QQQ I'd rather see this pushed down into last_frame_adjust 472 // and have it take the sender (aka caller). 473 474 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 475 caller_adjustment = last_frame_adjust(0, callee_locals); 476 } else if (callee_locals > callee_parameters) { 477 // The caller frame may need extending to accommodate 478 // non-parameter locals of the first unpacked interpreted frame. 479 // Compute that adjustment. 480 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 481 } 482 483 // If the sender is deoptimized the we must retrieve the address of the handler 484 // since the frame will "magically" show the original pc before the deopt 485 // and we'd undo the deopt. 486 487 frame_pcs[0] = deopt_sender.raw_pc(); 488 489 #ifndef SHARK 490 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 491 #endif // SHARK 492 493 #ifdef INCLUDE_JVMCI 494 if (exceptionObject() != NULL) { 495 thread->set_exception_oop(exceptionObject()); 496 exec_mode = Unpack_exception; 497 } 498 #endif 499 500 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 501 caller_adjustment * BytesPerWord, 502 caller_was_method_handle ? 0 : callee_parameters, 503 number_of_frames, 504 frame_sizes, 505 frame_pcs, 506 return_type, 507 exec_mode); 508 // On some platforms, we need a way to pass some platform dependent 509 // information to the unpacking code so the skeletal frames come out 510 // correct (initial fp value, unextended sp, ...) 511 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 512 513 if (array->frames() > 1) { 514 if (VerifyStack && TraceDeoptimization) { 515 ttyLocker ttyl; 516 tty->print_cr("Deoptimizing method containing inlining"); 517 } 518 } 519 520 array->set_unroll_block(info); 521 return info; 522 } 523 524 // Called to cleanup deoptimization data structures in normal case 525 // after unpacking to stack and when stack overflow error occurs 526 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 527 vframeArray *array) { 528 529 // Get array if coming from exception 530 if (array == NULL) { 531 array = thread->vframe_array_head(); 532 } 533 thread->set_vframe_array_head(NULL); 534 535 // Free the previous UnrollBlock 536 vframeArray* old_array = thread->vframe_array_last(); 537 thread->set_vframe_array_last(array); 538 539 if (old_array != NULL) { 540 UnrollBlock* old_info = old_array->unroll_block(); 541 old_array->set_unroll_block(NULL); 542 delete old_info; 543 delete old_array; 544 } 545 546 // Deallocate any resource creating in this routine and any ResourceObjs allocated 547 // inside the vframeArray (StackValueCollections) 548 549 delete thread->deopt_mark(); 550 thread->set_deopt_mark(NULL); 551 thread->set_deopt_nmethod(NULL); 552 553 554 if (JvmtiExport::can_pop_frame()) { 555 #ifndef CC_INTERP 556 // Regardless of whether we entered this routine with the pending 557 // popframe condition bit set, we should always clear it now 558 thread->clear_popframe_condition(); 559 #else 560 // C++ interpreter will clear has_pending_popframe when it enters 561 // with method_resume. For deopt_resume2 we clear it now. 562 if (thread->popframe_forcing_deopt_reexecution()) 563 thread->clear_popframe_condition(); 564 #endif /* CC_INTERP */ 565 } 566 567 // unpack_frames() is called at the end of the deoptimization handler 568 // and (in C2) at the end of the uncommon trap handler. Note this fact 569 // so that an asynchronous stack walker can work again. This counter is 570 // incremented at the beginning of fetch_unroll_info() and (in C2) at 571 // the beginning of uncommon_trap(). 572 thread->dec_in_deopt_handler(); 573 } 574 575 // Moved from cpu directories because none of the cpus has callee save values. 576 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 577 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 578 579 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 580 // the days we had adapter frames. When we deoptimize a situation where a 581 // compiled caller calls a compiled caller will have registers it expects 582 // to survive the call to the callee. If we deoptimize the callee the only 583 // way we can restore these registers is to have the oldest interpreter 584 // frame that we create restore these values. That is what this routine 585 // will accomplish. 586 587 // At the moment we have modified c2 to not have any callee save registers 588 // so this problem does not exist and this routine is just a place holder. 589 590 assert(f->is_interpreted_frame(), "must be interpreted"); 591 } 592 593 // Return BasicType of value being returned 594 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 595 596 // We are already active int he special DeoptResourceMark any ResourceObj's we 597 // allocate will be freed at the end of the routine. 598 599 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 600 // but makes the entry a little slower. There is however a little dance we have to 601 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 602 ResetNoHandleMark rnhm; // No-op in release/product versions 603 HandleMark hm; 604 605 frame stub_frame = thread->last_frame(); 606 607 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 608 // must point to the vframeArray for the unpack frame. 609 vframeArray* array = thread->vframe_array_head(); 610 611 #ifndef PRODUCT 612 if (TraceDeoptimization) { 613 ttyLocker ttyl; 614 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 615 p2i(thread), p2i(array), exec_mode); 616 } 617 #endif 618 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 619 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 620 621 UnrollBlock* info = array->unroll_block(); 622 623 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 624 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 625 626 BasicType bt = info->return_type(); 627 628 // If we have an exception pending, claim that the return type is an oop 629 // so the deopt_blob does not overwrite the exception_oop. 630 631 if (exec_mode == Unpack_exception) 632 bt = T_OBJECT; 633 634 // Cleanup thread deopt data 635 cleanup_deopt_info(thread, array); 636 637 #ifndef PRODUCT 638 if (VerifyStack) { 639 ResourceMark res_mark; 640 641 thread->validate_frame_layout(); 642 643 // Verify that the just-unpacked frames match the interpreter's 644 // notions of expression stack and locals 645 vframeArray* cur_array = thread->vframe_array_last(); 646 RegisterMap rm(thread, false); 647 rm.set_include_argument_oops(false); 648 bool is_top_frame = true; 649 int callee_size_of_parameters = 0; 650 int callee_max_locals = 0; 651 for (int i = 0; i < cur_array->frames(); i++) { 652 vframeArrayElement* el = cur_array->element(i); 653 frame* iframe = el->iframe(); 654 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 655 656 // Get the oop map for this bci 657 InterpreterOopMap mask; 658 int cur_invoke_parameter_size = 0; 659 bool try_next_mask = false; 660 int next_mask_expression_stack_size = -1; 661 int top_frame_expression_stack_adjustment = 0; 662 methodHandle mh(thread, iframe->interpreter_frame_method()); 663 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 664 BytecodeStream str(mh); 665 str.set_start(iframe->interpreter_frame_bci()); 666 int max_bci = mh->code_size(); 667 // Get to the next bytecode if possible 668 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 669 // Check to see if we can grab the number of outgoing arguments 670 // at an uncommon trap for an invoke (where the compiler 671 // generates debug info before the invoke has executed) 672 Bytecodes::Code cur_code = str.next(); 673 if (cur_code == Bytecodes::_invokevirtual || 674 cur_code == Bytecodes::_invokespecial || 675 cur_code == Bytecodes::_invokestatic || 676 cur_code == Bytecodes::_invokeinterface || 677 cur_code == Bytecodes::_invokedynamic) { 678 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 679 Symbol* signature = invoke.signature(); 680 ArgumentSizeComputer asc(signature); 681 cur_invoke_parameter_size = asc.size(); 682 if (invoke.has_receiver()) { 683 // Add in receiver 684 ++cur_invoke_parameter_size; 685 } 686 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 687 callee_size_of_parameters++; 688 } 689 } 690 if (str.bci() < max_bci) { 691 Bytecodes::Code bc = str.next(); 692 if (bc >= 0) { 693 // The interpreter oop map generator reports results before 694 // the current bytecode has executed except in the case of 695 // calls. It seems to be hard to tell whether the compiler 696 // has emitted debug information matching the "state before" 697 // a given bytecode or the state after, so we try both 698 switch (cur_code) { 699 case Bytecodes::_invokevirtual: 700 case Bytecodes::_invokespecial: 701 case Bytecodes::_invokestatic: 702 case Bytecodes::_invokeinterface: 703 case Bytecodes::_invokedynamic: 704 case Bytecodes::_athrow: 705 break; 706 default: { 707 InterpreterOopMap next_mask; 708 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 709 next_mask_expression_stack_size = next_mask.expression_stack_size(); 710 // Need to subtract off the size of the result type of 711 // the bytecode because this is not described in the 712 // debug info but returned to the interpreter in the TOS 713 // caching register 714 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 715 if (bytecode_result_type != T_ILLEGAL) { 716 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 717 } 718 assert(top_frame_expression_stack_adjustment >= 0, ""); 719 try_next_mask = true; 720 break; 721 } 722 } 723 } 724 } 725 726 // Verify stack depth and oops in frame 727 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 728 if (!( 729 /* SPARC */ 730 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 731 /* x86 */ 732 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 733 (try_next_mask && 734 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 735 top_frame_expression_stack_adjustment))) || 736 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 737 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 738 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 739 )) { 740 ttyLocker ttyl; 741 742 // Print out some information that will help us debug the problem 743 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 744 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 745 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 746 iframe->interpreter_frame_expression_stack_size()); 747 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 748 tty->print_cr(" try_next_mask = %d", try_next_mask); 749 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 750 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 751 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 752 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 753 tty->print_cr(" exec_mode = %d", exec_mode); 754 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 755 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 756 tty->print_cr(" Interpreted frames:"); 757 for (int k = 0; k < cur_array->frames(); k++) { 758 vframeArrayElement* el = cur_array->element(k); 759 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 760 } 761 cur_array->print_on_2(tty); 762 guarantee(false, "wrong number of expression stack elements during deopt"); 763 } 764 VerifyOopClosure verify; 765 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 766 callee_size_of_parameters = mh->size_of_parameters(); 767 callee_max_locals = mh->max_locals(); 768 is_top_frame = false; 769 } 770 } 771 #endif /* !PRODUCT */ 772 773 774 return bt; 775 JRT_END 776 777 778 int Deoptimization::deoptimize_dependents() { 779 Threads::deoptimized_wrt_marked_nmethods(); 780 return 0; 781 } 782 783 Deoptimization::DeoptAction Deoptimization::_unloaded_action 784 = Deoptimization::Action_reinterpret; 785 786 #if defined(COMPILER2) || INCLUDE_JVMCI 787 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 788 Handle pending_exception(thread->pending_exception()); 789 const char* exception_file = thread->exception_file(); 790 int exception_line = thread->exception_line(); 791 thread->clear_pending_exception(); 792 793 bool failures = false; 794 795 for (int i = 0; i < objects->length(); i++) { 796 assert(objects->at(i)->is_object(), "invalid debug information"); 797 ObjectValue* sv = (ObjectValue*) objects->at(i); 798 799 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 800 oop obj = NULL; 801 802 if (k->is_instance_klass()) { 803 InstanceKlass* ik = InstanceKlass::cast(k()); 804 obj = ik->allocate_instance(THREAD); 805 } else if (k->is_typeArray_klass()) { 806 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 807 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 808 int len = sv->field_size() / type2size[ak->element_type()]; 809 obj = ak->allocate(len, THREAD); 810 } else if (k->is_objArray_klass()) { 811 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 812 obj = ak->allocate(sv->field_size(), THREAD); 813 } 814 815 if (obj == NULL) { 816 failures = true; 817 } 818 819 assert(sv->value().is_null(), "redundant reallocation"); 820 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 821 CLEAR_PENDING_EXCEPTION; 822 sv->set_value(obj); 823 } 824 825 if (failures) { 826 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 827 } else if (pending_exception.not_null()) { 828 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 829 } 830 831 return failures; 832 } 833 834 // restore elements of an eliminated type array 835 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 836 int index = 0; 837 intptr_t val; 838 839 for (int i = 0; i < sv->field_size(); i++) { 840 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 841 switch(type) { 842 case T_LONG: case T_DOUBLE: { 843 assert(value->type() == T_INT, "Agreement."); 844 StackValue* low = 845 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 846 #ifdef _LP64 847 jlong res = (jlong)low->get_int(); 848 #else 849 #ifdef SPARC 850 // For SPARC we have to swap high and low words. 851 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 852 #else 853 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 854 #endif //SPARC 855 #endif 856 obj->long_at_put(index, res); 857 break; 858 } 859 860 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 861 case T_INT: case T_FLOAT: { // 4 bytes. 862 assert(value->type() == T_INT, "Agreement."); 863 bool big_value = false; 864 if (i + 1 < sv->field_size() && type == T_INT) { 865 if (sv->field_at(i)->is_location()) { 866 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 867 if (type == Location::dbl || type == Location::lng) { 868 big_value = true; 869 } 870 } else if (sv->field_at(i)->is_constant_int()) { 871 ScopeValue* next_scope_field = sv->field_at(i + 1); 872 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 873 big_value = true; 874 } 875 } 876 } 877 878 if (big_value) { 879 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 880 #ifdef _LP64 881 jlong res = (jlong)low->get_int(); 882 #else 883 #ifdef SPARC 884 // For SPARC we have to swap high and low words. 885 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 886 #else 887 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 888 #endif //SPARC 889 #endif 890 obj->int_at_put(index, (jint)*((jint*)&res)); 891 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 892 } else { 893 val = value->get_int(); 894 obj->int_at_put(index, (jint)*((jint*)&val)); 895 } 896 break; 897 } 898 899 case T_SHORT: case T_CHAR: // 2 bytes 900 assert(value->type() == T_INT, "Agreement."); 901 val = value->get_int(); 902 obj->short_at_put(index, (jshort)*((jint*)&val)); 903 break; 904 905 case T_BOOLEAN: case T_BYTE: // 1 byte 906 assert(value->type() == T_INT, "Agreement."); 907 val = value->get_int(); 908 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 909 break; 910 911 default: 912 ShouldNotReachHere(); 913 } 914 index++; 915 } 916 } 917 918 919 // restore fields of an eliminated object array 920 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 921 for (int i = 0; i < sv->field_size(); i++) { 922 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 923 assert(value->type() == T_OBJECT, "object element expected"); 924 obj->obj_at_put(i, value->get_obj()()); 925 } 926 } 927 928 class ReassignedField { 929 public: 930 int _offset; 931 BasicType _type; 932 public: 933 ReassignedField() { 934 _offset = 0; 935 _type = T_ILLEGAL; 936 } 937 }; 938 939 int compare(ReassignedField* left, ReassignedField* right) { 940 return left->_offset - right->_offset; 941 } 942 943 // Restore fields of an eliminated instance object using the same field order 944 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 945 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 946 if (klass->superklass() != NULL) { 947 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); 948 } 949 950 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 951 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 952 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 953 ReassignedField field; 954 field._offset = fs.offset(); 955 field._type = FieldType::basic_type(fs.signature()); 956 fields->append(field); 957 } 958 } 959 fields->sort(compare); 960 for (int i = 0; i < fields->length(); i++) { 961 intptr_t val; 962 ScopeValue* scope_field = sv->field_at(svIndex); 963 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 964 int offset = fields->at(i)._offset; 965 BasicType type = fields->at(i)._type; 966 switch (type) { 967 case T_OBJECT: case T_ARRAY: 968 assert(value->type() == T_OBJECT, "Agreement."); 969 obj->obj_field_put(offset, value->get_obj()()); 970 break; 971 972 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 973 case T_INT: case T_FLOAT: { // 4 bytes. 974 assert(value->type() == T_INT, "Agreement."); 975 bool big_value = false; 976 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 977 if (scope_field->is_location()) { 978 Location::Type type = ((LocationValue*) scope_field)->location().type(); 979 if (type == Location::dbl || type == Location::lng) { 980 big_value = true; 981 } 982 } 983 if (scope_field->is_constant_int()) { 984 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 985 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 986 big_value = true; 987 } 988 } 989 } 990 991 if (big_value) { 992 i++; 993 assert(i < fields->length(), "second T_INT field needed"); 994 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 995 } else { 996 val = value->get_int(); 997 obj->int_field_put(offset, (jint)*((jint*)&val)); 998 break; 999 } 1000 } 1001 /* no break */ 1002 1003 case T_LONG: case T_DOUBLE: { 1004 assert(value->type() == T_INT, "Agreement."); 1005 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1006 #ifdef _LP64 1007 jlong res = (jlong)low->get_int(); 1008 #else 1009 #ifdef SPARC 1010 // For SPARC we have to swap high and low words. 1011 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1012 #else 1013 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1014 #endif //SPARC 1015 #endif 1016 obj->long_field_put(offset, res); 1017 break; 1018 } 1019 1020 case T_SHORT: case T_CHAR: // 2 bytes 1021 assert(value->type() == T_INT, "Agreement."); 1022 val = value->get_int(); 1023 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1024 break; 1025 1026 case T_BOOLEAN: case T_BYTE: // 1 byte 1027 assert(value->type() == T_INT, "Agreement."); 1028 val = value->get_int(); 1029 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1030 break; 1031 1032 default: 1033 ShouldNotReachHere(); 1034 } 1035 svIndex++; 1036 } 1037 return svIndex; 1038 } 1039 1040 // restore fields of all eliminated objects and arrays 1041 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1042 for (int i = 0; i < objects->length(); i++) { 1043 ObjectValue* sv = (ObjectValue*) objects->at(i); 1044 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1045 Handle obj = sv->value(); 1046 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1047 if (PrintDeoptimizationDetails) { 1048 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1049 } 1050 if (obj.is_null()) { 1051 continue; 1052 } 1053 1054 if (k->is_instance_klass()) { 1055 InstanceKlass* ik = InstanceKlass::cast(k()); 1056 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1057 } else if (k->is_typeArray_klass()) { 1058 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 1059 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1060 } else if (k->is_objArray_klass()) { 1061 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1062 } 1063 } 1064 } 1065 1066 1067 // relock objects for which synchronization was eliminated 1068 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1069 for (int i = 0; i < monitors->length(); i++) { 1070 MonitorInfo* mon_info = monitors->at(i); 1071 if (mon_info->eliminated()) { 1072 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1073 if (!mon_info->owner_is_scalar_replaced()) { 1074 Handle obj = Handle(mon_info->owner()); 1075 markOop mark = obj->mark(); 1076 if (UseBiasedLocking && mark->has_bias_pattern()) { 1077 // New allocated objects may have the mark set to anonymously biased. 1078 // Also the deoptimized method may called methods with synchronization 1079 // where the thread-local object is bias locked to the current thread. 1080 assert(mark->is_biased_anonymously() || 1081 mark->biased_locker() == thread, "should be locked to current thread"); 1082 // Reset mark word to unbiased prototype. 1083 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1084 obj->set_mark(unbiased_prototype); 1085 } 1086 BasicLock* lock = mon_info->lock(); 1087 ObjectSynchronizer::slow_enter(obj, lock, thread); 1088 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1089 } 1090 } 1091 } 1092 } 1093 1094 1095 #ifndef PRODUCT 1096 // print information about reallocated objects 1097 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1098 fieldDescriptor fd; 1099 1100 for (int i = 0; i < objects->length(); i++) { 1101 ObjectValue* sv = (ObjectValue*) objects->at(i); 1102 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1103 Handle obj = sv->value(); 1104 1105 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1106 k->print_value(); 1107 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1108 if (obj.is_null()) { 1109 tty->print(" allocation failed"); 1110 } else { 1111 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1112 } 1113 tty->cr(); 1114 1115 if (Verbose && !obj.is_null()) { 1116 k->oop_print_on(obj(), tty); 1117 } 1118 } 1119 } 1120 #endif 1121 #endif // COMPILER2 || INCLUDE_JVMCI 1122 1123 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1124 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1125 1126 #ifndef PRODUCT 1127 if (PrintDeoptimizationDetails) { 1128 ttyLocker ttyl; 1129 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1130 fr.print_on(tty); 1131 tty->print_cr(" Virtual frames (innermost first):"); 1132 for (int index = 0; index < chunk->length(); index++) { 1133 compiledVFrame* vf = chunk->at(index); 1134 tty->print(" %2d - ", index); 1135 vf->print_value(); 1136 int bci = chunk->at(index)->raw_bci(); 1137 const char* code_name; 1138 if (bci == SynchronizationEntryBCI) { 1139 code_name = "sync entry"; 1140 } else { 1141 Bytecodes::Code code = vf->method()->code_at(bci); 1142 code_name = Bytecodes::name(code); 1143 } 1144 tty->print(" - %s", code_name); 1145 tty->print_cr(" @ bci %d ", bci); 1146 if (Verbose) { 1147 vf->print(); 1148 tty->cr(); 1149 } 1150 } 1151 } 1152 #endif 1153 1154 // Register map for next frame (used for stack crawl). We capture 1155 // the state of the deopt'ing frame's caller. Thus if we need to 1156 // stuff a C2I adapter we can properly fill in the callee-save 1157 // register locations. 1158 frame caller = fr.sender(reg_map); 1159 int frame_size = caller.sp() - fr.sp(); 1160 1161 frame sender = caller; 1162 1163 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1164 // the vframeArray containing the unpacking information is allocated in the C heap. 1165 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1166 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1167 1168 // Compare the vframeArray to the collected vframes 1169 assert(array->structural_compare(thread, chunk), "just checking"); 1170 1171 #ifndef PRODUCT 1172 if (PrintDeoptimizationDetails) { 1173 ttyLocker ttyl; 1174 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1175 } 1176 #endif // PRODUCT 1177 1178 return array; 1179 } 1180 1181 #if defined(COMPILER2) || INCLUDE_JVMCI 1182 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1183 // Reallocation of some scalar replaced objects failed. Record 1184 // that we need to pop all the interpreter frames for the 1185 // deoptimized compiled frame. 1186 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1187 thread->set_frames_to_pop_failed_realloc(array->frames()); 1188 // Unlock all monitors here otherwise the interpreter will see a 1189 // mix of locked and unlocked monitors (because of failed 1190 // reallocations of synchronized objects) and be confused. 1191 for (int i = 0; i < array->frames(); i++) { 1192 MonitorChunk* monitors = array->element(i)->monitors(); 1193 if (monitors != NULL) { 1194 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1195 BasicObjectLock* src = monitors->at(j); 1196 if (src->obj() != NULL) { 1197 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1198 } 1199 } 1200 array->element(i)->free_monitors(thread); 1201 #ifdef ASSERT 1202 array->element(i)->set_removed_monitors(); 1203 #endif 1204 } 1205 } 1206 } 1207 #endif 1208 1209 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1210 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1211 for (int i = 0; i < monitors->length(); i++) { 1212 MonitorInfo* mon_info = monitors->at(i); 1213 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1214 objects_to_revoke->append(Handle(mon_info->owner())); 1215 } 1216 } 1217 } 1218 1219 1220 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1221 if (!UseBiasedLocking) { 1222 return; 1223 } 1224 1225 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1226 1227 // Unfortunately we don't have a RegisterMap available in most of 1228 // the places we want to call this routine so we need to walk the 1229 // stack again to update the register map. 1230 if (map == NULL || !map->update_map()) { 1231 StackFrameStream sfs(thread, true); 1232 bool found = false; 1233 while (!found && !sfs.is_done()) { 1234 frame* cur = sfs.current(); 1235 sfs.next(); 1236 found = cur->id() == fr.id(); 1237 } 1238 assert(found, "frame to be deoptimized not found on target thread's stack"); 1239 map = sfs.register_map(); 1240 } 1241 1242 vframe* vf = vframe::new_vframe(&fr, map, thread); 1243 compiledVFrame* cvf = compiledVFrame::cast(vf); 1244 // Revoke monitors' biases in all scopes 1245 while (!cvf->is_top()) { 1246 collect_monitors(cvf, objects_to_revoke); 1247 cvf = compiledVFrame::cast(cvf->sender()); 1248 } 1249 collect_monitors(cvf, objects_to_revoke); 1250 1251 if (SafepointSynchronize::is_at_safepoint()) { 1252 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1253 } else { 1254 BiasedLocking::revoke(objects_to_revoke); 1255 } 1256 } 1257 1258 1259 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1260 if (!UseBiasedLocking) { 1261 return; 1262 } 1263 1264 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1265 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1266 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1267 if (jt->has_last_Java_frame()) { 1268 StackFrameStream sfs(jt, true); 1269 while (!sfs.is_done()) { 1270 frame* cur = sfs.current(); 1271 if (cb->contains(cur->pc())) { 1272 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1273 compiledVFrame* cvf = compiledVFrame::cast(vf); 1274 // Revoke monitors' biases in all scopes 1275 while (!cvf->is_top()) { 1276 collect_monitors(cvf, objects_to_revoke); 1277 cvf = compiledVFrame::cast(cvf->sender()); 1278 } 1279 collect_monitors(cvf, objects_to_revoke); 1280 } 1281 sfs.next(); 1282 } 1283 } 1284 } 1285 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1286 } 1287 1288 1289 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1290 assert(fr.can_be_deoptimized(), "checking frame type"); 1291 1292 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1293 1294 if (LogCompilation && xtty != NULL) { 1295 nmethod* nm = fr.cb()->as_nmethod_or_null(); 1296 assert(nm != NULL, "only compiled methods can deopt"); 1297 1298 ttyLocker ttyl; 1299 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "'", (uintx)thread->osthread()->thread_id()); 1300 nm->log_identity(xtty); 1301 xtty->end_head(); 1302 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1303 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1304 xtty->method(sd->method()); 1305 xtty->end_elem(); 1306 if (sd->is_top()) break; 1307 } 1308 xtty->tail("deoptimized"); 1309 } 1310 1311 // Patch the compiled method so that when execution returns to it we will 1312 // deopt the execution state and return to the interpreter. 1313 fr.deoptimize(thread); 1314 } 1315 1316 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1317 deoptimize(thread, fr, map, Reason_constraint); 1318 } 1319 1320 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1321 // Deoptimize only if the frame comes from compile code. 1322 // Do not deoptimize the frame which is already patched 1323 // during the execution of the loops below. 1324 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1325 return; 1326 } 1327 ResourceMark rm; 1328 DeoptimizationMarker dm; 1329 if (UseBiasedLocking) { 1330 revoke_biases_of_monitors(thread, fr, map); 1331 } 1332 deoptimize_single_frame(thread, fr, reason); 1333 1334 } 1335 1336 1337 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1338 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1339 "can only deoptimize other thread at a safepoint"); 1340 // Compute frame and register map based on thread and sp. 1341 RegisterMap reg_map(thread, UseBiasedLocking); 1342 frame fr = thread->last_frame(); 1343 while (fr.id() != id) { 1344 fr = fr.sender(®_map); 1345 } 1346 deoptimize(thread, fr, ®_map, reason); 1347 } 1348 1349 1350 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1351 if (thread == Thread::current()) { 1352 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1353 } else { 1354 VM_DeoptimizeFrame deopt(thread, id, reason); 1355 VMThread::execute(&deopt); 1356 } 1357 } 1358 1359 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1360 deoptimize_frame(thread, id, Reason_constraint); 1361 } 1362 1363 // JVMTI PopFrame support 1364 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1365 { 1366 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1367 } 1368 JRT_END 1369 1370 MethodData* 1371 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1372 bool create_if_missing) { 1373 Thread* THREAD = thread; 1374 MethodData* mdo = m()->method_data(); 1375 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1376 // Build an MDO. Ignore errors like OutOfMemory; 1377 // that simply means we won't have an MDO to update. 1378 Method::build_interpreter_method_data(m, THREAD); 1379 if (HAS_PENDING_EXCEPTION) { 1380 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1381 CLEAR_PENDING_EXCEPTION; 1382 } 1383 mdo = m()->method_data(); 1384 } 1385 return mdo; 1386 } 1387 1388 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI 1389 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1390 // in case of an unresolved klass entry, load the class. 1391 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1392 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1393 return; 1394 } 1395 1396 if (!constant_pool->tag_at(index).is_symbol()) return; 1397 1398 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1399 Symbol* symbol = constant_pool->symbol_at(index); 1400 1401 // class name? 1402 if (symbol->byte_at(0) != '(') { 1403 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1404 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1405 return; 1406 } 1407 1408 // then it must be a signature! 1409 ResourceMark rm(THREAD); 1410 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1411 if (ss.is_object()) { 1412 Symbol* class_name = ss.as_symbol(CHECK); 1413 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1414 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1415 } 1416 } 1417 } 1418 1419 1420 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1421 EXCEPTION_MARK; 1422 load_class_by_index(constant_pool, index, THREAD); 1423 if (HAS_PENDING_EXCEPTION) { 1424 // Exception happened during classloading. We ignore the exception here, since it 1425 // is going to be rethrown since the current activation is going to be deoptimized and 1426 // the interpreter will re-execute the bytecode. 1427 CLEAR_PENDING_EXCEPTION; 1428 // Class loading called java code which may have caused a stack 1429 // overflow. If the exception was thrown right before the return 1430 // to the runtime the stack is no longer guarded. Reguard the 1431 // stack otherwise if we return to the uncommon trap blob and the 1432 // stack bang causes a stack overflow we crash. 1433 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1434 JavaThread* thread = (JavaThread*)THREAD; 1435 bool guard_pages_enabled = thread->stack_guards_enabled(); 1436 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1437 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1438 } 1439 } 1440 1441 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1442 HandleMark hm; 1443 1444 // uncommon_trap() is called at the beginning of the uncommon trap 1445 // handler. Note this fact before we start generating temporary frames 1446 // that can confuse an asynchronous stack walker. This counter is 1447 // decremented at the end of unpack_frames(). 1448 thread->inc_in_deopt_handler(); 1449 1450 // We need to update the map if we have biased locking. 1451 #if INCLUDE_JVMCI 1452 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1453 RegisterMap reg_map(thread, true); 1454 #else 1455 RegisterMap reg_map(thread, UseBiasedLocking); 1456 #endif 1457 frame stub_frame = thread->last_frame(); 1458 frame fr = stub_frame.sender(®_map); 1459 // Make sure the calling nmethod is not getting deoptimized and removed 1460 // before we are done with it. 1461 nmethodLocker nl(fr.pc()); 1462 1463 // Log a message 1464 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1465 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1466 1467 { 1468 ResourceMark rm; 1469 1470 // Revoke biases of any monitors in the frame to ensure we can migrate them 1471 revoke_biases_of_monitors(thread, fr, ®_map); 1472 1473 DeoptReason reason = trap_request_reason(trap_request); 1474 DeoptAction action = trap_request_action(trap_request); 1475 #if INCLUDE_JVMCI 1476 int debug_id = trap_request_debug_id(trap_request); 1477 #endif 1478 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1479 1480 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1481 compiledVFrame* cvf = compiledVFrame::cast(vf); 1482 1483 nmethod* nm = cvf->code(); 1484 1485 ScopeDesc* trap_scope = cvf->scope(); 1486 1487 if (TraceDeoptimization) { 1488 ttyLocker ttyl; 1489 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() 1490 #if INCLUDE_JVMCI 1491 , debug_id 1492 #endif 1493 ); 1494 } 1495 1496 methodHandle trap_method = trap_scope->method(); 1497 int trap_bci = trap_scope->bci(); 1498 #if INCLUDE_JVMCI 1499 oop speculation = thread->pending_failed_speculation(); 1500 if (nm->is_compiled_by_jvmci()) { 1501 if (speculation != NULL) { 1502 oop speculation_log = nm->speculation_log(); 1503 if (speculation_log != NULL) { 1504 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1505 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) { 1506 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1507 } 1508 } 1509 if (TraceDeoptimization) { 1510 tty->print_cr("Saving speculation to speculation log"); 1511 } 1512 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1513 } else { 1514 if (TraceDeoptimization) { 1515 tty->print_cr("Speculation present but no speculation log"); 1516 } 1517 } 1518 thread->set_pending_failed_speculation(NULL); 1519 } else { 1520 if (TraceDeoptimization) { 1521 tty->print_cr("No speculation"); 1522 } 1523 } 1524 } else { 1525 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1526 } 1527 1528 if (trap_bci == SynchronizationEntryBCI) { 1529 trap_bci = 0; 1530 thread->set_pending_monitorenter(true); 1531 } 1532 1533 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1534 thread->set_pending_transfer_to_interpreter(true); 1535 } 1536 #endif 1537 1538 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1539 // Record this event in the histogram. 1540 gather_statistics(reason, action, trap_bc); 1541 1542 // Ensure that we can record deopt. history: 1543 // Need MDO to record RTM code generation state. 1544 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1545 1546 methodHandle profiled_method; 1547 #if INCLUDE_JVMCI 1548 if (nm->is_compiled_by_jvmci()) { 1549 profiled_method = nm->method(); 1550 } else { 1551 profiled_method = trap_method; 1552 } 1553 #else 1554 profiled_method = trap_method; 1555 #endif 1556 1557 MethodData* trap_mdo = 1558 get_method_data(thread, profiled_method, create_if_missing); 1559 1560 // Log a message 1561 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1562 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1563 trap_method->name_and_sig_as_C_string(), trap_bci); 1564 1565 // Print a bunch of diagnostics, if requested. 1566 if (TraceDeoptimization || LogCompilation) { 1567 ResourceMark rm; 1568 ttyLocker ttyl; 1569 char buf[100]; 1570 if (xtty != NULL) { 1571 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1572 os::current_thread_id(), 1573 format_trap_request(buf, sizeof(buf), trap_request)); 1574 nm->log_identity(xtty); 1575 } 1576 Symbol* class_name = NULL; 1577 bool unresolved = false; 1578 if (unloaded_class_index >= 0) { 1579 constantPoolHandle constants (THREAD, trap_method->constants()); 1580 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1581 class_name = constants->klass_name_at(unloaded_class_index); 1582 unresolved = true; 1583 if (xtty != NULL) 1584 xtty->print(" unresolved='1'"); 1585 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1586 class_name = constants->symbol_at(unloaded_class_index); 1587 } 1588 if (xtty != NULL) 1589 xtty->name(class_name); 1590 } 1591 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1592 // Dump the relevant MDO state. 1593 // This is the deopt count for the current reason, any previous 1594 // reasons or recompiles seen at this point. 1595 int dcnt = trap_mdo->trap_count(reason); 1596 if (dcnt != 0) 1597 xtty->print(" count='%d'", dcnt); 1598 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1599 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1600 if (dos != 0) { 1601 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1602 if (trap_state_is_recompiled(dos)) { 1603 int recnt2 = trap_mdo->overflow_recompile_count(); 1604 if (recnt2 != 0) 1605 xtty->print(" recompiles2='%d'", recnt2); 1606 } 1607 } 1608 } 1609 if (xtty != NULL) { 1610 xtty->stamp(); 1611 xtty->end_head(); 1612 } 1613 if (TraceDeoptimization) { // make noise on the tty 1614 tty->print("Uncommon trap occurred in"); 1615 nm->method()->print_short_name(tty); 1616 tty->print(" compiler=%s compile_id=%d", nm->compiler() == NULL ? "" : nm->compiler()->name(), nm->compile_id()); 1617 #if INCLUDE_JVMCI 1618 oop installedCode = nm->jvmci_installed_code(); 1619 if (installedCode != NULL) { 1620 oop installedCodeName = NULL; 1621 if (installedCode->is_a(InstalledCode::klass())) { 1622 installedCodeName = InstalledCode::name(installedCode); 1623 } 1624 if (installedCodeName != NULL) { 1625 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName)); 1626 } else { 1627 tty->print(" (JVMCI: installed code has no name) "); 1628 } 1629 } else if (nm->is_compiled_by_jvmci()) { 1630 tty->print(" (JVMCI: no installed code) "); 1631 } 1632 #endif 1633 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1634 p2i(fr.pc()), 1635 os::current_thread_id(), 1636 trap_reason_name(reason), 1637 trap_action_name(action), 1638 unloaded_class_index 1639 #if INCLUDE_JVMCI 1640 , debug_id 1641 #endif 1642 ); 1643 if (class_name != NULL) { 1644 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1645 class_name->print_symbol_on(tty); 1646 } 1647 tty->cr(); 1648 } 1649 if (xtty != NULL) { 1650 // Log the precise location of the trap. 1651 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1652 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1653 xtty->method(sd->method()); 1654 xtty->end_elem(); 1655 if (sd->is_top()) break; 1656 } 1657 xtty->tail("uncommon_trap"); 1658 } 1659 } 1660 // (End diagnostic printout.) 1661 1662 // Load class if necessary 1663 if (unloaded_class_index >= 0) { 1664 constantPoolHandle constants(THREAD, trap_method->constants()); 1665 load_class_by_index(constants, unloaded_class_index); 1666 } 1667 1668 // Flush the nmethod if necessary and desirable. 1669 // 1670 // We need to avoid situations where we are re-flushing the nmethod 1671 // because of a hot deoptimization site. Repeated flushes at the same 1672 // point need to be detected by the compiler and avoided. If the compiler 1673 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1674 // module must take measures to avoid an infinite cycle of recompilation 1675 // and deoptimization. There are several such measures: 1676 // 1677 // 1. If a recompilation is ordered a second time at some site X 1678 // and for the same reason R, the action is adjusted to 'reinterpret', 1679 // to give the interpreter time to exercise the method more thoroughly. 1680 // If this happens, the method's overflow_recompile_count is incremented. 1681 // 1682 // 2. If the compiler fails to reduce the deoptimization rate, then 1683 // the method's overflow_recompile_count will begin to exceed the set 1684 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1685 // is adjusted to 'make_not_compilable', and the method is abandoned 1686 // to the interpreter. This is a performance hit for hot methods, 1687 // but is better than a disastrous infinite cycle of recompilations. 1688 // (Actually, only the method containing the site X is abandoned.) 1689 // 1690 // 3. In parallel with the previous measures, if the total number of 1691 // recompilations of a method exceeds the much larger set limit 1692 // PerMethodRecompilationCutoff, the method is abandoned. 1693 // This should only happen if the method is very large and has 1694 // many "lukewarm" deoptimizations. The code which enforces this 1695 // limit is elsewhere (class nmethod, class Method). 1696 // 1697 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1698 // to recompile at each bytecode independently of the per-BCI cutoff. 1699 // 1700 // The decision to update code is up to the compiler, and is encoded 1701 // in the Action_xxx code. If the compiler requests Action_none 1702 // no trap state is changed, no compiled code is changed, and the 1703 // computation suffers along in the interpreter. 1704 // 1705 // The other action codes specify various tactics for decompilation 1706 // and recompilation. Action_maybe_recompile is the loosest, and 1707 // allows the compiled code to stay around until enough traps are seen, 1708 // and until the compiler gets around to recompiling the trapping method. 1709 // 1710 // The other actions cause immediate removal of the present code. 1711 1712 // Traps caused by injected profile shouldn't pollute trap counts. 1713 bool injected_profile_trap = trap_method->has_injected_profile() && 1714 (reason == Reason_intrinsic || reason == Reason_unreached); 1715 1716 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1717 bool make_not_entrant = false; 1718 bool make_not_compilable = false; 1719 bool reprofile = false; 1720 switch (action) { 1721 case Action_none: 1722 // Keep the old code. 1723 update_trap_state = false; 1724 break; 1725 case Action_maybe_recompile: 1726 // Do not need to invalidate the present code, but we can 1727 // initiate another 1728 // Start compiler without (necessarily) invalidating the nmethod. 1729 // The system will tolerate the old code, but new code should be 1730 // generated when possible. 1731 break; 1732 case Action_reinterpret: 1733 // Go back into the interpreter for a while, and then consider 1734 // recompiling form scratch. 1735 make_not_entrant = true; 1736 // Reset invocation counter for outer most method. 1737 // This will allow the interpreter to exercise the bytecodes 1738 // for a while before recompiling. 1739 // By contrast, Action_make_not_entrant is immediate. 1740 // 1741 // Note that the compiler will track null_check, null_assert, 1742 // range_check, and class_check events and log them as if they 1743 // had been traps taken from compiled code. This will update 1744 // the MDO trap history so that the next compilation will 1745 // properly detect hot trap sites. 1746 reprofile = true; 1747 break; 1748 case Action_make_not_entrant: 1749 // Request immediate recompilation, and get rid of the old code. 1750 // Make them not entrant, so next time they are called they get 1751 // recompiled. Unloaded classes are loaded now so recompile before next 1752 // time they are called. Same for uninitialized. The interpreter will 1753 // link the missing class, if any. 1754 make_not_entrant = true; 1755 break; 1756 case Action_make_not_compilable: 1757 // Give up on compiling this method at all. 1758 make_not_entrant = true; 1759 make_not_compilable = true; 1760 break; 1761 default: 1762 ShouldNotReachHere(); 1763 } 1764 1765 // Setting +ProfileTraps fixes the following, on all platforms: 1766 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1767 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1768 // recompile relies on a MethodData* to record heroic opt failures. 1769 1770 // Whether the interpreter is producing MDO data or not, we also need 1771 // to use the MDO to detect hot deoptimization points and control 1772 // aggressive optimization. 1773 bool inc_recompile_count = false; 1774 ProfileData* pdata = NULL; 1775 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1776 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1777 uint this_trap_count = 0; 1778 bool maybe_prior_trap = false; 1779 bool maybe_prior_recompile = false; 1780 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1781 #if INCLUDE_JVMCI 1782 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1783 #endif 1784 nm->method(), 1785 //outputs: 1786 this_trap_count, 1787 maybe_prior_trap, 1788 maybe_prior_recompile); 1789 // Because the interpreter also counts null, div0, range, and class 1790 // checks, these traps from compiled code are double-counted. 1791 // This is harmless; it just means that the PerXTrapLimit values 1792 // are in effect a little smaller than they look. 1793 1794 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1795 if (per_bc_reason != Reason_none) { 1796 // Now take action based on the partially known per-BCI history. 1797 if (maybe_prior_trap 1798 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1799 // If there are too many traps at this BCI, force a recompile. 1800 // This will allow the compiler to see the limit overflow, and 1801 // take corrective action, if possible. The compiler generally 1802 // does not use the exact PerBytecodeTrapLimit value, but instead 1803 // changes its tactics if it sees any traps at all. This provides 1804 // a little hysteresis, delaying a recompile until a trap happens 1805 // several times. 1806 // 1807 // Actually, since there is only one bit of counter per BCI, 1808 // the possible per-BCI counts are {0,1,(per-method count)}. 1809 // This produces accurate results if in fact there is only 1810 // one hot trap site, but begins to get fuzzy if there are 1811 // many sites. For example, if there are ten sites each 1812 // trapping two or more times, they each get the blame for 1813 // all of their traps. 1814 make_not_entrant = true; 1815 } 1816 1817 // Detect repeated recompilation at the same BCI, and enforce a limit. 1818 if (make_not_entrant && maybe_prior_recompile) { 1819 // More than one recompile at this point. 1820 inc_recompile_count = maybe_prior_trap; 1821 } 1822 } else { 1823 // For reasons which are not recorded per-bytecode, we simply 1824 // force recompiles unconditionally. 1825 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1826 make_not_entrant = true; 1827 } 1828 1829 // Go back to the compiler if there are too many traps in this method. 1830 if (this_trap_count >= per_method_trap_limit(reason)) { 1831 // If there are too many traps in this method, force a recompile. 1832 // This will allow the compiler to see the limit overflow, and 1833 // take corrective action, if possible. 1834 // (This condition is an unlikely backstop only, because the 1835 // PerBytecodeTrapLimit is more likely to take effect first, 1836 // if it is applicable.) 1837 make_not_entrant = true; 1838 } 1839 1840 // Here's more hysteresis: If there has been a recompile at 1841 // this trap point already, run the method in the interpreter 1842 // for a while to exercise it more thoroughly. 1843 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1844 reprofile = true; 1845 } 1846 } 1847 1848 // Take requested actions on the method: 1849 1850 // Recompile 1851 if (make_not_entrant) { 1852 if (!nm->make_not_entrant()) { 1853 return; // the call did not change nmethod's state 1854 } 1855 1856 if (pdata != NULL) { 1857 // Record the recompilation event, if any. 1858 int tstate0 = pdata->trap_state(); 1859 int tstate1 = trap_state_set_recompiled(tstate0, true); 1860 if (tstate1 != tstate0) 1861 pdata->set_trap_state(tstate1); 1862 } 1863 1864 #if INCLUDE_RTM_OPT 1865 // Restart collecting RTM locking abort statistic if the method 1866 // is recompiled for a reason other than RTM state change. 1867 // Assume that in new recompiled code the statistic could be different, 1868 // for example, due to different inlining. 1869 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1870 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1871 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1872 } 1873 #endif 1874 // For code aging we count traps separately here, using make_not_entrant() 1875 // as a guard against simultaneous deopts in multiple threads. 1876 if (reason == Reason_tenured && trap_mdo != NULL) { 1877 trap_mdo->inc_tenure_traps(); 1878 } 1879 } 1880 1881 if (inc_recompile_count) { 1882 trap_mdo->inc_overflow_recompile_count(); 1883 if ((uint)trap_mdo->overflow_recompile_count() > 1884 (uint)PerBytecodeRecompilationCutoff) { 1885 // Give up on the method containing the bad BCI. 1886 if (trap_method() == nm->method()) { 1887 make_not_compilable = true; 1888 } else { 1889 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1890 // But give grace to the enclosing nm->method(). 1891 } 1892 } 1893 } 1894 1895 // Reprofile 1896 if (reprofile) { 1897 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1898 } 1899 1900 // Give up compiling 1901 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1902 assert(make_not_entrant, "consistent"); 1903 nm->method()->set_not_compilable(CompLevel_full_optimization); 1904 } 1905 1906 } // Free marked resources 1907 1908 } 1909 JRT_END 1910 1911 ProfileData* 1912 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1913 int trap_bci, 1914 Deoptimization::DeoptReason reason, 1915 bool update_total_trap_count, 1916 #if INCLUDE_JVMCI 1917 bool is_osr, 1918 #endif 1919 Method* compiled_method, 1920 //outputs: 1921 uint& ret_this_trap_count, 1922 bool& ret_maybe_prior_trap, 1923 bool& ret_maybe_prior_recompile) { 1924 bool maybe_prior_trap = false; 1925 bool maybe_prior_recompile = false; 1926 uint this_trap_count = 0; 1927 if (update_total_trap_count) { 1928 uint idx = reason; 1929 #if INCLUDE_JVMCI 1930 if (is_osr) { 1931 idx += Reason_LIMIT; 1932 } 1933 #endif 1934 uint prior_trap_count = trap_mdo->trap_count(idx); 1935 this_trap_count = trap_mdo->inc_trap_count(idx); 1936 1937 // If the runtime cannot find a place to store trap history, 1938 // it is estimated based on the general condition of the method. 1939 // If the method has ever been recompiled, or has ever incurred 1940 // a trap with the present reason , then this BCI is assumed 1941 // (pessimistically) to be the culprit. 1942 maybe_prior_trap = (prior_trap_count != 0); 1943 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1944 } 1945 ProfileData* pdata = NULL; 1946 1947 1948 // For reasons which are recorded per bytecode, we check per-BCI data. 1949 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1950 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 1951 if (per_bc_reason != Reason_none) { 1952 // Find the profile data for this BCI. If there isn't one, 1953 // try to allocate one from the MDO's set of spares. 1954 // This will let us detect a repeated trap at this point. 1955 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1956 1957 if (pdata != NULL) { 1958 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1959 if (LogCompilation && xtty != NULL) { 1960 ttyLocker ttyl; 1961 // no more room for speculative traps in this MDO 1962 xtty->elem("speculative_traps_oom"); 1963 } 1964 } 1965 // Query the trap state of this profile datum. 1966 int tstate0 = pdata->trap_state(); 1967 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1968 maybe_prior_trap = false; 1969 if (!trap_state_is_recompiled(tstate0)) 1970 maybe_prior_recompile = false; 1971 1972 // Update the trap state of this profile datum. 1973 int tstate1 = tstate0; 1974 // Record the reason. 1975 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1976 // Store the updated state on the MDO, for next time. 1977 if (tstate1 != tstate0) 1978 pdata->set_trap_state(tstate1); 1979 } else { 1980 if (LogCompilation && xtty != NULL) { 1981 ttyLocker ttyl; 1982 // Missing MDP? Leave a small complaint in the log. 1983 xtty->elem("missing_mdp bci='%d'", trap_bci); 1984 } 1985 } 1986 } 1987 1988 // Return results: 1989 ret_this_trap_count = this_trap_count; 1990 ret_maybe_prior_trap = maybe_prior_trap; 1991 ret_maybe_prior_recompile = maybe_prior_recompile; 1992 return pdata; 1993 } 1994 1995 void 1996 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1997 ResourceMark rm; 1998 // Ignored outputs: 1999 uint ignore_this_trap_count; 2000 bool ignore_maybe_prior_trap; 2001 bool ignore_maybe_prior_recompile; 2002 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2003 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2004 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true); 2005 query_update_method_data(trap_mdo, trap_bci, 2006 (DeoptReason)reason, 2007 update_total_counts, 2008 #if INCLUDE_JVMCI 2009 false, 2010 #endif 2011 NULL, 2012 ignore_this_trap_count, 2013 ignore_maybe_prior_trap, 2014 ignore_maybe_prior_recompile); 2015 } 2016 2017 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2018 if (TraceDeoptimization) { 2019 tty->print("Uncommon trap "); 2020 } 2021 // Still in Java no safepoints 2022 { 2023 // This enters VM and may safepoint 2024 uncommon_trap_inner(thread, trap_request); 2025 } 2026 return fetch_unroll_info_helper(thread, exec_mode); 2027 } 2028 2029 // Local derived constants. 2030 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2031 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 2032 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2033 2034 //---------------------------trap_state_reason--------------------------------- 2035 Deoptimization::DeoptReason 2036 Deoptimization::trap_state_reason(int trap_state) { 2037 // This assert provides the link between the width of DataLayout::trap_bits 2038 // and the encoding of "recorded" reasons. It ensures there are enough 2039 // bits to store all needed reasons in the per-BCI MDO profile. 2040 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2041 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2042 trap_state -= recompile_bit; 2043 if (trap_state == DS_REASON_MASK) { 2044 return Reason_many; 2045 } else { 2046 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2047 return (DeoptReason)trap_state; 2048 } 2049 } 2050 //-------------------------trap_state_has_reason------------------------------- 2051 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2052 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2053 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2054 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2055 trap_state -= recompile_bit; 2056 if (trap_state == DS_REASON_MASK) { 2057 return -1; // true, unspecifically (bottom of state lattice) 2058 } else if (trap_state == reason) { 2059 return 1; // true, definitely 2060 } else if (trap_state == 0) { 2061 return 0; // false, definitely (top of state lattice) 2062 } else { 2063 return 0; // false, definitely 2064 } 2065 } 2066 //-------------------------trap_state_add_reason------------------------------- 2067 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2068 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2069 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2070 trap_state -= recompile_bit; 2071 if (trap_state == DS_REASON_MASK) { 2072 return trap_state + recompile_bit; // already at state lattice bottom 2073 } else if (trap_state == reason) { 2074 return trap_state + recompile_bit; // the condition is already true 2075 } else if (trap_state == 0) { 2076 return reason + recompile_bit; // no condition has yet been true 2077 } else { 2078 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2079 } 2080 } 2081 //-----------------------trap_state_is_recompiled------------------------------ 2082 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2083 return (trap_state & DS_RECOMPILE_BIT) != 0; 2084 } 2085 //-----------------------trap_state_set_recompiled----------------------------- 2086 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2087 if (z) return trap_state | DS_RECOMPILE_BIT; 2088 else return trap_state & ~DS_RECOMPILE_BIT; 2089 } 2090 //---------------------------format_trap_state--------------------------------- 2091 // This is used for debugging and diagnostics, including LogFile output. 2092 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2093 int trap_state) { 2094 assert(buflen > 0, "sanity"); 2095 DeoptReason reason = trap_state_reason(trap_state); 2096 bool recomp_flag = trap_state_is_recompiled(trap_state); 2097 // Re-encode the state from its decoded components. 2098 int decoded_state = 0; 2099 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2100 decoded_state = trap_state_add_reason(decoded_state, reason); 2101 if (recomp_flag) 2102 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2103 // If the state re-encodes properly, format it symbolically. 2104 // Because this routine is used for debugging and diagnostics, 2105 // be robust even if the state is a strange value. 2106 size_t len; 2107 if (decoded_state != trap_state) { 2108 // Random buggy state that doesn't decode?? 2109 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2110 } else { 2111 len = jio_snprintf(buf, buflen, "%s%s", 2112 trap_reason_name(reason), 2113 recomp_flag ? " recompiled" : ""); 2114 } 2115 return buf; 2116 } 2117 2118 2119 //--------------------------------statics-------------------------------------- 2120 const char* Deoptimization::_trap_reason_name[] = { 2121 // Note: Keep this in sync. with enum DeoptReason. 2122 "none", 2123 "null_check", 2124 "null_assert" JVMCI_ONLY("_or_unreached0"), 2125 "range_check", 2126 "class_check", 2127 "array_check", 2128 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2129 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2130 "unloaded", 2131 "uninitialized", 2132 "unreached", 2133 "unhandled", 2134 "constraint", 2135 "div0_check", 2136 "age", 2137 "predicate", 2138 "loop_limit_check", 2139 "speculate_class_check", 2140 "speculate_null_check", 2141 "rtm_state_change", 2142 "unstable_if", 2143 "unstable_fused_if", 2144 #if INCLUDE_JVMCI 2145 "aliasing", 2146 "transfer_to_interpreter", 2147 "not_compiled_exception_handler", 2148 "unresolved", 2149 "jsr_mismatch", 2150 #endif 2151 "tenured" 2152 }; 2153 const char* Deoptimization::_trap_action_name[] = { 2154 // Note: Keep this in sync. with enum DeoptAction. 2155 "none", 2156 "maybe_recompile", 2157 "reinterpret", 2158 "make_not_entrant", 2159 "make_not_compilable" 2160 }; 2161 2162 const char* Deoptimization::trap_reason_name(int reason) { 2163 // Check that every reason has a name 2164 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2165 2166 if (reason == Reason_many) return "many"; 2167 if ((uint)reason < Reason_LIMIT) 2168 return _trap_reason_name[reason]; 2169 static char buf[20]; 2170 sprintf(buf, "reason%d", reason); 2171 return buf; 2172 } 2173 const char* Deoptimization::trap_action_name(int action) { 2174 // Check that every action has a name 2175 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2176 2177 if ((uint)action < Action_LIMIT) 2178 return _trap_action_name[action]; 2179 static char buf[20]; 2180 sprintf(buf, "action%d", action); 2181 return buf; 2182 } 2183 2184 // This is used for debugging and diagnostics, including LogFile output. 2185 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2186 int trap_request) { 2187 jint unloaded_class_index = trap_request_index(trap_request); 2188 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2189 const char* action = trap_action_name(trap_request_action(trap_request)); 2190 #if INCLUDE_JVMCI 2191 int debug_id = trap_request_debug_id(trap_request); 2192 #endif 2193 size_t len; 2194 if (unloaded_class_index < 0) { 2195 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2196 reason, action 2197 #if INCLUDE_JVMCI 2198 ,debug_id 2199 #endif 2200 ); 2201 } else { 2202 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2203 reason, action, unloaded_class_index 2204 #if INCLUDE_JVMCI 2205 ,debug_id 2206 #endif 2207 ); 2208 } 2209 return buf; 2210 } 2211 2212 juint Deoptimization::_deoptimization_hist 2213 [Deoptimization::Reason_LIMIT] 2214 [1 + Deoptimization::Action_LIMIT] 2215 [Deoptimization::BC_CASE_LIMIT] 2216 = {0}; 2217 2218 enum { 2219 LSB_BITS = 8, 2220 LSB_MASK = right_n_bits(LSB_BITS) 2221 }; 2222 2223 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2224 Bytecodes::Code bc) { 2225 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2226 assert(action >= 0 && action < Action_LIMIT, "oob"); 2227 _deoptimization_hist[Reason_none][0][0] += 1; // total 2228 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2229 juint* cases = _deoptimization_hist[reason][1+action]; 2230 juint* bc_counter_addr = NULL; 2231 juint bc_counter = 0; 2232 // Look for an unused counter, or an exact match to this BC. 2233 if (bc != Bytecodes::_illegal) { 2234 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2235 juint* counter_addr = &cases[bc_case]; 2236 juint counter = *counter_addr; 2237 if ((counter == 0 && bc_counter_addr == NULL) 2238 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2239 // this counter is either free or is already devoted to this BC 2240 bc_counter_addr = counter_addr; 2241 bc_counter = counter | bc; 2242 } 2243 } 2244 } 2245 if (bc_counter_addr == NULL) { 2246 // Overflow, or no given bytecode. 2247 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2248 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2249 } 2250 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2251 } 2252 2253 jint Deoptimization::total_deoptimization_count() { 2254 return _deoptimization_hist[Reason_none][0][0]; 2255 } 2256 2257 jint Deoptimization::deoptimization_count(DeoptReason reason) { 2258 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2259 return _deoptimization_hist[reason][0][0]; 2260 } 2261 2262 void Deoptimization::print_statistics() { 2263 juint total = total_deoptimization_count(); 2264 juint account = total; 2265 if (total != 0) { 2266 ttyLocker ttyl; 2267 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2268 tty->print_cr("Deoptimization traps recorded:"); 2269 #define PRINT_STAT_LINE(name, r) \ 2270 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2271 PRINT_STAT_LINE("total", total); 2272 // For each non-zero entry in the histogram, print the reason, 2273 // the action, and (if specifically known) the type of bytecode. 2274 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2275 for (int action = 0; action < Action_LIMIT; action++) { 2276 juint* cases = _deoptimization_hist[reason][1+action]; 2277 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2278 juint counter = cases[bc_case]; 2279 if (counter != 0) { 2280 char name[1*K]; 2281 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2282 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2283 bc = Bytecodes::_illegal; 2284 sprintf(name, "%s/%s/%s", 2285 trap_reason_name(reason), 2286 trap_action_name(action), 2287 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2288 juint r = counter >> LSB_BITS; 2289 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2290 account -= r; 2291 } 2292 } 2293 } 2294 } 2295 if (account != 0) { 2296 PRINT_STAT_LINE("unaccounted", account); 2297 } 2298 #undef PRINT_STAT_LINE 2299 if (xtty != NULL) xtty->tail("statistics"); 2300 } 2301 } 2302 #else // COMPILER2 || SHARK || INCLUDE_JVMCI 2303 2304 2305 // Stubs for C1 only system. 2306 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2307 return false; 2308 } 2309 2310 const char* Deoptimization::trap_reason_name(int reason) { 2311 return "unknown"; 2312 } 2313 2314 void Deoptimization::print_statistics() { 2315 // no output 2316 } 2317 2318 void 2319 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2320 // no udpate 2321 } 2322 2323 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2324 return 0; 2325 } 2326 2327 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2328 Bytecodes::Code bc) { 2329 // no update 2330 } 2331 2332 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2333 int trap_state) { 2334 jio_snprintf(buf, buflen, "#%d", trap_state); 2335 return buf; 2336 } 2337 2338 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI