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