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