deoptimization.cpp revision 3064:aa3d708d67c4
1/* 2 * Copyright (c) 1997, 2012, 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/debugInfoRec.hpp" 28#include "code/nmethod.hpp" 29#include "code/pcDesc.hpp" 30#include "code/scopeDesc.hpp" 31#include "interpreter/bytecode.hpp" 32#include "interpreter/interpreter.hpp" 33#include "interpreter/oopMapCache.hpp" 34#include "memory/allocation.inline.hpp" 35#include "memory/oopFactory.hpp" 36#include "memory/resourceArea.hpp" 37#include "oops/methodOop.hpp" 38#include "oops/oop.inline.hpp" 39#include "prims/jvmtiThreadState.hpp" 40#include "runtime/biasedLocking.hpp" 41#include "runtime/compilationPolicy.hpp" 42#include "runtime/deoptimization.hpp" 43#include "runtime/interfaceSupport.hpp" 44#include "runtime/sharedRuntime.hpp" 45#include "runtime/signature.hpp" 46#include "runtime/stubRoutines.hpp" 47#include "runtime/thread.hpp" 48#include "runtime/vframe.hpp" 49#include "runtime/vframeArray.hpp" 50#include "runtime/vframe_hp.hpp" 51#include "utilities/events.hpp" 52#include "utilities/xmlstream.hpp" 53#ifdef TARGET_ARCH_x86 54# include "vmreg_x86.inline.hpp" 55#endif 56#ifdef TARGET_ARCH_sparc 57# include "vmreg_sparc.inline.hpp" 58#endif 59#ifdef TARGET_ARCH_zero 60# include "vmreg_zero.inline.hpp" 61#endif 62#ifdef TARGET_ARCH_arm 63# include "vmreg_arm.inline.hpp" 64#endif 65#ifdef TARGET_ARCH_ppc 66# include "vmreg_ppc.inline.hpp" 67#endif 68#ifdef COMPILER2 69#ifdef TARGET_ARCH_MODEL_x86_32 70# include "adfiles/ad_x86_32.hpp" 71#endif 72#ifdef TARGET_ARCH_MODEL_x86_64 73# include "adfiles/ad_x86_64.hpp" 74#endif 75#ifdef TARGET_ARCH_MODEL_sparc 76# include "adfiles/ad_sparc.hpp" 77#endif 78#ifdef TARGET_ARCH_MODEL_zero 79# include "adfiles/ad_zero.hpp" 80#endif 81#ifdef TARGET_ARCH_MODEL_arm 82# include "adfiles/ad_arm.hpp" 83#endif 84#ifdef TARGET_ARCH_MODEL_ppc 85# include "adfiles/ad_ppc.hpp" 86#endif 87#endif 88 89bool DeoptimizationMarker::_is_active = false; 90 91Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 92 int caller_adjustment, 93 int caller_actual_parameters, 94 int number_of_frames, 95 intptr_t* frame_sizes, 96 address* frame_pcs, 97 BasicType return_type) { 98 _size_of_deoptimized_frame = size_of_deoptimized_frame; 99 _caller_adjustment = caller_adjustment; 100 _caller_actual_parameters = caller_actual_parameters; 101 _number_of_frames = number_of_frames; 102 _frame_sizes = frame_sizes; 103 _frame_pcs = frame_pcs; 104 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); 105 _return_type = return_type; 106 _initial_info = 0; 107 // PD (x86 only) 108 _counter_temp = 0; 109 _unpack_kind = 0; 110 _sender_sp_temp = 0; 111 112 _total_frame_sizes = size_of_frames(); 113} 114 115 116Deoptimization::UnrollBlock::~UnrollBlock() { 117 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 118 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 119 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 120} 121 122 123intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 124 assert(register_number < RegisterMap::reg_count, "checking register number"); 125 return &_register_block[register_number * 2]; 126} 127 128 129 130int Deoptimization::UnrollBlock::size_of_frames() const { 131 // Acount first for the adjustment of the initial frame 132 int result = _caller_adjustment; 133 for (int index = 0; index < number_of_frames(); index++) { 134 result += frame_sizes()[index]; 135 } 136 return result; 137} 138 139 140void Deoptimization::UnrollBlock::print() { 141 ttyLocker ttyl; 142 tty->print_cr("UnrollBlock"); 143 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 144 tty->print( " frame_sizes: "); 145 for (int index = 0; index < number_of_frames(); index++) { 146 tty->print("%d ", frame_sizes()[index]); 147 } 148 tty->cr(); 149} 150 151 152// In order to make fetch_unroll_info work properly with escape 153// analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 154// ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 155// of previously eliminated objects occurs in realloc_objects, which is 156// called from the method fetch_unroll_info_helper below. 157JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 158 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 159 // but makes the entry a little slower. There is however a little dance we have to 160 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 161 162 // fetch_unroll_info() is called at the beginning of the deoptimization 163 // handler. Note this fact before we start generating temporary frames 164 // that can confuse an asynchronous stack walker. This counter is 165 // decremented at the end of unpack_frames(). 166 thread->inc_in_deopt_handler(); 167 168 return fetch_unroll_info_helper(thread); 169JRT_END 170 171 172// This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 173Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 174 175 // Note: there is a safepoint safety issue here. No matter whether we enter 176 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 177 // the vframeArray is created. 178 // 179 180 // Allocate our special deoptimization ResourceMark 181 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 182 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 183 thread->set_deopt_mark(dmark); 184 185 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 186 RegisterMap map(thread, true); 187 RegisterMap dummy_map(thread, false); 188 // Now get the deoptee with a valid map 189 frame deoptee = stub_frame.sender(&map); 190 // Set the deoptee nmethod 191 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 192 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 193 194 if (VerifyStack) { 195 thread->validate_frame_layout(); 196 } 197 198 // Create a growable array of VFrames where each VFrame represents an inlined 199 // Java frame. This storage is allocated with the usual system arena. 200 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 201 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 202 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 203 while (!vf->is_top()) { 204 assert(vf->is_compiled_frame(), "Wrong frame type"); 205 chunk->push(compiledVFrame::cast(vf)); 206 vf = vf->sender(); 207 } 208 assert(vf->is_compiled_frame(), "Wrong frame type"); 209 chunk->push(compiledVFrame::cast(vf)); 210 211#ifdef COMPILER2 212 // Reallocate the non-escaping objects and restore their fields. Then 213 // relock objects if synchronization on them was eliminated. 214 if (DoEscapeAnalysis || EliminateNestedLocks) { 215 if (EliminateAllocations) { 216 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 217 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 218 219 // The flag return_oop() indicates call sites which return oop 220 // in compiled code. Such sites include java method calls, 221 // runtime calls (for example, used to allocate new objects/arrays 222 // on slow code path) and any other calls generated in compiled code. 223 // It is not guaranteed that we can get such information here only 224 // by analyzing bytecode in deoptimized frames. This is why this flag 225 // is set during method compilation (see Compile::Process_OopMap_Node()). 226 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 227 Handle return_value; 228 if (save_oop_result) { 229 // Reallocation may trigger GC. If deoptimization happened on return from 230 // call which returns oop we need to save it since it is not in oopmap. 231 oop result = deoptee.saved_oop_result(&map); 232 assert(result == NULL || result->is_oop(), "must be oop"); 233 return_value = Handle(thread, result); 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 235 if (TraceDeoptimization) { 236 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 237 } 238 } 239 bool reallocated = false; 240 if (objects != NULL) { 241 JRT_BLOCK 242 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 243 JRT_END 244 } 245 if (reallocated) { 246 reassign_fields(&deoptee, &map, objects); 247#ifndef PRODUCT 248 if (TraceDeoptimization) { 249 ttyLocker ttyl; 250 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 251 print_objects(objects); 252 } 253#endif 254 } 255 if (save_oop_result) { 256 // Restore result. 257 deoptee.set_saved_oop_result(&map, return_value()); 258 } 259 } 260 if (EliminateLocks) { 261#ifndef PRODUCT 262 bool first = true; 263#endif 264 for (int i = 0; i < chunk->length(); i++) { 265 compiledVFrame* cvf = chunk->at(i); 266 assert (cvf->scope() != NULL,"expect only compiled java frames"); 267 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 268 if (monitors->is_nonempty()) { 269 relock_objects(monitors, thread); 270#ifndef PRODUCT 271 if (TraceDeoptimization) { 272 ttyLocker ttyl; 273 for (int j = 0; j < monitors->length(); j++) { 274 MonitorInfo* mi = monitors->at(j); 275 if (mi->eliminated()) { 276 if (first) { 277 first = false; 278 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 279 } 280 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 281 } 282 } 283 } 284#endif 285 } 286 } 287 } 288 } 289#endif // COMPILER2 290 // Ensure that no safepoint is taken after pointers have been stored 291 // in fields of rematerialized objects. If a safepoint occurs from here on 292 // out the java state residing in the vframeArray will be missed. 293 No_Safepoint_Verifier no_safepoint; 294 295 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 296 297 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 298 thread->set_vframe_array_head(array); 299 300 // Now that the vframeArray has been created if we have any deferred local writes 301 // added by jvmti then we can free up that structure as the data is now in the 302 // vframeArray 303 304 if (thread->deferred_locals() != NULL) { 305 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 306 int i = 0; 307 do { 308 // Because of inlining we could have multiple vframes for a single frame 309 // and several of the vframes could have deferred writes. Find them all. 310 if (list->at(i)->id() == array->original().id()) { 311 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 312 list->remove_at(i); 313 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 314 delete dlv; 315 } else { 316 i++; 317 } 318 } while ( i < list->length() ); 319 if (list->length() == 0) { 320 thread->set_deferred_locals(NULL); 321 // free the list and elements back to C heap. 322 delete list; 323 } 324 325 } 326 327#ifndef SHARK 328 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 329 CodeBlob* cb = stub_frame.cb(); 330 // Verify we have the right vframeArray 331 assert(cb->frame_size() >= 0, "Unexpected frame size"); 332 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 333 334 // If the deopt call site is a MethodHandle invoke call site we have 335 // to adjust the unpack_sp. 336 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 337 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 338 unpack_sp = deoptee.unextended_sp(); 339 340#ifdef ASSERT 341 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 342#endif 343#else 344 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 345#endif // !SHARK 346 347 // This is a guarantee instead of an assert because if vframe doesn't match 348 // we will unpack the wrong deoptimized frame and wind up in strange places 349 // where it will be very difficult to figure out what went wrong. Better 350 // to die an early death here than some very obscure death later when the 351 // trail is cold. 352 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 353 // in that it will fail to detect a problem when there is one. This needs 354 // more work in tiger timeframe. 355 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 356 357 int number_of_frames = array->frames(); 358 359 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 360 // virtual activation, which is the reverse of the elements in the vframes array. 361 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 362 // +1 because we always have an interpreter return address for the final slot. 363 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 364 int popframe_extra_args = 0; 365 // Create an interpreter return address for the stub to use as its return 366 // address so the skeletal frames are perfectly walkable 367 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 368 369 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 370 // activation be put back on the expression stack of the caller for reexecution 371 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 372 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 373 } 374 375 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 376 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 377 // than simply use array->sender.pc(). This requires us to walk the current set of frames 378 // 379 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 380 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 381 382 // It's possible that the number of paramters at the call site is 383 // different than number of arguments in the callee when method 384 // handles are used. If the caller is interpreted get the real 385 // value so that the proper amount of space can be added to it's 386 // frame. 387 bool caller_was_method_handle = false; 388 if (deopt_sender.is_interpreted_frame()) { 389 methodHandle method = deopt_sender.interpreter_frame_method(); 390 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 391 if (cur.is_method_handle_invoke()) { 392 // Method handle invokes may involve fairly arbitrary chains of 393 // calls so it's impossible to know how much actual space the 394 // caller has for locals. 395 caller_was_method_handle = true; 396 } 397 } 398 399 // 400 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 401 // frame_sizes/frame_pcs[1] next oldest frame (int) 402 // frame_sizes/frame_pcs[n] youngest frame (int) 403 // 404 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 405 // owns the space for the return address to it's caller). Confusing ain't it. 406 // 407 // The vframe array can address vframes with indices running from 408 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 409 // When we create the skeletal frames we need the oldest frame to be in the zero slot 410 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 411 // so things look a little strange in this loop. 412 // 413 int callee_parameters = 0; 414 int callee_locals = 0; 415 for (int index = 0; index < array->frames(); index++ ) { 416 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 417 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 418 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 419 int caller_parms = callee_parameters; 420 if ((index == array->frames() - 1) && caller_was_method_handle) { 421 caller_parms = 0; 422 } 423 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms, 424 callee_parameters, 425 callee_locals, 426 index == 0, 427 popframe_extra_args); 428 // This pc doesn't have to be perfect just good enough to identify the frame 429 // as interpreted so the skeleton frame will be walkable 430 // The correct pc will be set when the skeleton frame is completely filled out 431 // The final pc we store in the loop is wrong and will be overwritten below 432 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 433 434 callee_parameters = array->element(index)->method()->size_of_parameters(); 435 callee_locals = array->element(index)->method()->max_locals(); 436 popframe_extra_args = 0; 437 } 438 439 // Compute whether the root vframe returns a float or double value. 440 BasicType return_type; 441 { 442 HandleMark hm; 443 methodHandle method(thread, array->element(0)->method()); 444 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 445 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 446 } 447 448 // Compute information for handling adapters and adjusting the frame size of the caller. 449 int caller_adjustment = 0; 450 451 // Compute the amount the oldest interpreter frame will have to adjust 452 // its caller's stack by. If the caller is a compiled frame then 453 // we pretend that the callee has no parameters so that the 454 // extension counts for the full amount of locals and not just 455 // locals-parms. This is because without a c2i adapter the parm 456 // area as created by the compiled frame will not be usable by 457 // the interpreter. (Depending on the calling convention there 458 // may not even be enough space). 459 460 // QQQ I'd rather see this pushed down into last_frame_adjust 461 // and have it take the sender (aka caller). 462 463 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 464 caller_adjustment = last_frame_adjust(0, callee_locals); 465 } else if (callee_locals > callee_parameters) { 466 // The caller frame may need extending to accommodate 467 // non-parameter locals of the first unpacked interpreted frame. 468 // Compute that adjustment. 469 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 470 } 471 472 // If the sender is deoptimized the we must retrieve the address of the handler 473 // since the frame will "magically" show the original pc before the deopt 474 // and we'd undo the deopt. 475 476 frame_pcs[0] = deopt_sender.raw_pc(); 477 478#ifndef SHARK 479 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 480#endif // SHARK 481 482 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 483 caller_adjustment * BytesPerWord, 484 caller_was_method_handle ? 0 : callee_parameters, 485 number_of_frames, 486 frame_sizes, 487 frame_pcs, 488 return_type); 489 // On some platforms, we need a way to pass some platform dependent 490 // information to the unpacking code so the skeletal frames come out 491 // correct (initial fp value, unextended sp, ...) 492 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 493 494 if (array->frames() > 1) { 495 if (VerifyStack && TraceDeoptimization) { 496 tty->print_cr("Deoptimizing method containing inlining"); 497 } 498 } 499 500 array->set_unroll_block(info); 501 return info; 502} 503 504// Called to cleanup deoptimization data structures in normal case 505// after unpacking to stack and when stack overflow error occurs 506void Deoptimization::cleanup_deopt_info(JavaThread *thread, 507 vframeArray *array) { 508 509 // Get array if coming from exception 510 if (array == NULL) { 511 array = thread->vframe_array_head(); 512 } 513 thread->set_vframe_array_head(NULL); 514 515 // Free the previous UnrollBlock 516 vframeArray* old_array = thread->vframe_array_last(); 517 thread->set_vframe_array_last(array); 518 519 if (old_array != NULL) { 520 UnrollBlock* old_info = old_array->unroll_block(); 521 old_array->set_unroll_block(NULL); 522 delete old_info; 523 delete old_array; 524 } 525 526 // Deallocate any resource creating in this routine and any ResourceObjs allocated 527 // inside the vframeArray (StackValueCollections) 528 529 delete thread->deopt_mark(); 530 thread->set_deopt_mark(NULL); 531 thread->set_deopt_nmethod(NULL); 532 533 534 if (JvmtiExport::can_pop_frame()) { 535#ifndef CC_INTERP 536 // Regardless of whether we entered this routine with the pending 537 // popframe condition bit set, we should always clear it now 538 thread->clear_popframe_condition(); 539#else 540 // C++ interpeter will clear has_pending_popframe when it enters 541 // with method_resume. For deopt_resume2 we clear it now. 542 if (thread->popframe_forcing_deopt_reexecution()) 543 thread->clear_popframe_condition(); 544#endif /* CC_INTERP */ 545 } 546 547 // unpack_frames() is called at the end of the deoptimization handler 548 // and (in C2) at the end of the uncommon trap handler. Note this fact 549 // so that an asynchronous stack walker can work again. This counter is 550 // incremented at the beginning of fetch_unroll_info() and (in C2) at 551 // the beginning of uncommon_trap(). 552 thread->dec_in_deopt_handler(); 553} 554 555 556// Return BasicType of value being returned 557JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 558 559 // We are already active int he special DeoptResourceMark any ResourceObj's we 560 // allocate will be freed at the end of the routine. 561 562 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 563 // but makes the entry a little slower. There is however a little dance we have to 564 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 565 ResetNoHandleMark rnhm; // No-op in release/product versions 566 HandleMark hm; 567 568 frame stub_frame = thread->last_frame(); 569 570 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 571 // must point to the vframeArray for the unpack frame. 572 vframeArray* array = thread->vframe_array_head(); 573 574#ifndef PRODUCT 575 if (TraceDeoptimization) { 576 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 577 } 578#endif 579 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 580 stub_frame.pc(), stub_frame.sp(), exec_mode); 581 582 UnrollBlock* info = array->unroll_block(); 583 584 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 585 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 586 587 BasicType bt = info->return_type(); 588 589 // If we have an exception pending, claim that the return type is an oop 590 // so the deopt_blob does not overwrite the exception_oop. 591 592 if (exec_mode == Unpack_exception) 593 bt = T_OBJECT; 594 595 // Cleanup thread deopt data 596 cleanup_deopt_info(thread, array); 597 598#ifndef PRODUCT 599 if (VerifyStack) { 600 ResourceMark res_mark; 601 602 thread->validate_frame_layout(); 603 604 // Verify that the just-unpacked frames match the interpreter's 605 // notions of expression stack and locals 606 vframeArray* cur_array = thread->vframe_array_last(); 607 RegisterMap rm(thread, false); 608 rm.set_include_argument_oops(false); 609 bool is_top_frame = true; 610 int callee_size_of_parameters = 0; 611 int callee_max_locals = 0; 612 for (int i = 0; i < cur_array->frames(); i++) { 613 vframeArrayElement* el = cur_array->element(i); 614 frame* iframe = el->iframe(); 615 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 616 617 // Get the oop map for this bci 618 InterpreterOopMap mask; 619 int cur_invoke_parameter_size = 0; 620 bool try_next_mask = false; 621 int next_mask_expression_stack_size = -1; 622 int top_frame_expression_stack_adjustment = 0; 623 methodHandle mh(thread, iframe->interpreter_frame_method()); 624 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 625 BytecodeStream str(mh); 626 str.set_start(iframe->interpreter_frame_bci()); 627 int max_bci = mh->code_size(); 628 // Get to the next bytecode if possible 629 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 630 // Check to see if we can grab the number of outgoing arguments 631 // at an uncommon trap for an invoke (where the compiler 632 // generates debug info before the invoke has executed) 633 Bytecodes::Code cur_code = str.next(); 634 if (cur_code == Bytecodes::_invokevirtual || 635 cur_code == Bytecodes::_invokespecial || 636 cur_code == Bytecodes::_invokestatic || 637 cur_code == Bytecodes::_invokeinterface) { 638 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 639 Symbol* signature = invoke.signature(); 640 ArgumentSizeComputer asc(signature); 641 cur_invoke_parameter_size = asc.size(); 642 if (cur_code != Bytecodes::_invokestatic) { 643 // Add in receiver 644 ++cur_invoke_parameter_size; 645 } 646 } 647 if (str.bci() < max_bci) { 648 Bytecodes::Code bc = str.next(); 649 if (bc >= 0) { 650 // The interpreter oop map generator reports results before 651 // the current bytecode has executed except in the case of 652 // calls. It seems to be hard to tell whether the compiler 653 // has emitted debug information matching the "state before" 654 // a given bytecode or the state after, so we try both 655 switch (cur_code) { 656 case Bytecodes::_invokevirtual: 657 case Bytecodes::_invokespecial: 658 case Bytecodes::_invokestatic: 659 case Bytecodes::_invokeinterface: 660 case Bytecodes::_athrow: 661 break; 662 default: { 663 InterpreterOopMap next_mask; 664 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 665 next_mask_expression_stack_size = next_mask.expression_stack_size(); 666 // Need to subtract off the size of the result type of 667 // the bytecode because this is not described in the 668 // debug info but returned to the interpreter in the TOS 669 // caching register 670 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 671 if (bytecode_result_type != T_ILLEGAL) { 672 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 673 } 674 assert(top_frame_expression_stack_adjustment >= 0, ""); 675 try_next_mask = true; 676 break; 677 } 678 } 679 } 680 } 681 682 // Verify stack depth and oops in frame 683 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 684 if (!( 685 /* SPARC */ 686 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 687 /* x86 */ 688 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 689 (try_next_mask && 690 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 691 top_frame_expression_stack_adjustment))) || 692 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 693 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 694 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 695 )) { 696 ttyLocker ttyl; 697 698 // Print out some information that will help us debug the problem 699 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 700 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 701 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 702 iframe->interpreter_frame_expression_stack_size()); 703 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 704 tty->print_cr(" try_next_mask = %d", try_next_mask); 705 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 706 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 707 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 708 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 709 tty->print_cr(" exec_mode = %d", exec_mode); 710 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 711 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 712 tty->print_cr(" Interpreted frames:"); 713 for (int k = 0; k < cur_array->frames(); k++) { 714 vframeArrayElement* el = cur_array->element(k); 715 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 716 } 717 cur_array->print_on_2(tty); 718 guarantee(false, "wrong number of expression stack elements during deopt"); 719 } 720 VerifyOopClosure verify; 721 iframe->oops_interpreted_do(&verify, &rm, false); 722 callee_size_of_parameters = mh->size_of_parameters(); 723 callee_max_locals = mh->max_locals(); 724 is_top_frame = false; 725 } 726 } 727#endif /* !PRODUCT */ 728 729 730 return bt; 731JRT_END 732 733 734int Deoptimization::deoptimize_dependents() { 735 Threads::deoptimized_wrt_marked_nmethods(); 736 return 0; 737} 738 739 740#ifdef COMPILER2 741bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 742 Handle pending_exception(thread->pending_exception()); 743 const char* exception_file = thread->exception_file(); 744 int exception_line = thread->exception_line(); 745 thread->clear_pending_exception(); 746 747 for (int i = 0; i < objects->length(); i++) { 748 assert(objects->at(i)->is_object(), "invalid debug information"); 749 ObjectValue* sv = (ObjectValue*) objects->at(i); 750 751 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 752 oop obj = NULL; 753 754 if (k->oop_is_instance()) { 755 instanceKlass* ik = instanceKlass::cast(k()); 756 obj = ik->allocate_instance(CHECK_(false)); 757 } else if (k->oop_is_typeArray()) { 758 typeArrayKlass* ak = typeArrayKlass::cast(k()); 759 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 760 int len = sv->field_size() / type2size[ak->element_type()]; 761 obj = ak->allocate(len, CHECK_(false)); 762 } else if (k->oop_is_objArray()) { 763 objArrayKlass* ak = objArrayKlass::cast(k()); 764 obj = ak->allocate(sv->field_size(), CHECK_(false)); 765 } 766 767 assert(obj != NULL, "allocation failed"); 768 assert(sv->value().is_null(), "redundant reallocation"); 769 sv->set_value(obj); 770 } 771 772 if (pending_exception.not_null()) { 773 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 774 } 775 776 return true; 777} 778 779// This assumes that the fields are stored in ObjectValue in the same order 780// they are yielded by do_nonstatic_fields. 781class FieldReassigner: public FieldClosure { 782 frame* _fr; 783 RegisterMap* _reg_map; 784 ObjectValue* _sv; 785 instanceKlass* _ik; 786 oop _obj; 787 788 int _i; 789public: 790 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 791 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 792 793 int i() const { return _i; } 794 795 796 void do_field(fieldDescriptor* fd) { 797 intptr_t val; 798 StackValue* value = 799 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 800 int offset = fd->offset(); 801 switch (fd->field_type()) { 802 case T_OBJECT: case T_ARRAY: 803 assert(value->type() == T_OBJECT, "Agreement."); 804 _obj->obj_field_put(offset, value->get_obj()()); 805 break; 806 807 case T_LONG: case T_DOUBLE: { 808 assert(value->type() == T_INT, "Agreement."); 809 StackValue* low = 810 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 811#ifdef _LP64 812 jlong res = (jlong)low->get_int(); 813#else 814#ifdef SPARC 815 // For SPARC we have to swap high and low words. 816 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 817#else 818 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 819#endif //SPARC 820#endif 821 _obj->long_field_put(offset, res); 822 break; 823 } 824 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 825 case T_INT: case T_FLOAT: // 4 bytes. 826 assert(value->type() == T_INT, "Agreement."); 827 val = value->get_int(); 828 _obj->int_field_put(offset, (jint)*((jint*)&val)); 829 break; 830 831 case T_SHORT: case T_CHAR: // 2 bytes 832 assert(value->type() == T_INT, "Agreement."); 833 val = value->get_int(); 834 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 835 break; 836 837 case T_BOOLEAN: case T_BYTE: // 1 byte 838 assert(value->type() == T_INT, "Agreement."); 839 val = value->get_int(); 840 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 841 break; 842 843 default: 844 ShouldNotReachHere(); 845 } 846 _i++; 847 } 848}; 849 850// restore elements of an eliminated type array 851void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 852 int index = 0; 853 intptr_t val; 854 855 for (int i = 0; i < sv->field_size(); i++) { 856 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 857 switch(type) { 858 case T_LONG: case T_DOUBLE: { 859 assert(value->type() == T_INT, "Agreement."); 860 StackValue* low = 861 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 862#ifdef _LP64 863 jlong res = (jlong)low->get_int(); 864#else 865#ifdef SPARC 866 // For SPARC we have to swap high and low words. 867 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 868#else 869 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 870#endif //SPARC 871#endif 872 obj->long_at_put(index, res); 873 break; 874 } 875 876 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 877 case T_INT: case T_FLOAT: // 4 bytes. 878 assert(value->type() == T_INT, "Agreement."); 879 val = value->get_int(); 880 obj->int_at_put(index, (jint)*((jint*)&val)); 881 break; 882 883 case T_SHORT: case T_CHAR: // 2 bytes 884 assert(value->type() == T_INT, "Agreement."); 885 val = value->get_int(); 886 obj->short_at_put(index, (jshort)*((jint*)&val)); 887 break; 888 889 case T_BOOLEAN: case T_BYTE: // 1 byte 890 assert(value->type() == T_INT, "Agreement."); 891 val = value->get_int(); 892 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 893 break; 894 895 default: 896 ShouldNotReachHere(); 897 } 898 index++; 899 } 900} 901 902 903// restore fields of an eliminated object array 904void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 905 for (int i = 0; i < sv->field_size(); i++) { 906 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 907 assert(value->type() == T_OBJECT, "object element expected"); 908 obj->obj_at_put(i, value->get_obj()()); 909 } 910} 911 912 913// restore fields of all eliminated objects and arrays 914void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 915 for (int i = 0; i < objects->length(); i++) { 916 ObjectValue* sv = (ObjectValue*) objects->at(i); 917 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 918 Handle obj = sv->value(); 919 assert(obj.not_null(), "reallocation was missed"); 920 921 if (k->oop_is_instance()) { 922 instanceKlass* ik = instanceKlass::cast(k()); 923 FieldReassigner reassign(fr, reg_map, sv, obj()); 924 ik->do_nonstatic_fields(&reassign); 925 } else if (k->oop_is_typeArray()) { 926 typeArrayKlass* ak = typeArrayKlass::cast(k()); 927 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 928 } else if (k->oop_is_objArray()) { 929 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 930 } 931 } 932} 933 934 935// relock objects for which synchronization was eliminated 936void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 937 for (int i = 0; i < monitors->length(); i++) { 938 MonitorInfo* mon_info = monitors->at(i); 939 if (mon_info->eliminated()) { 940 assert(mon_info->owner() != NULL, "reallocation was missed"); 941 Handle obj = Handle(mon_info->owner()); 942 markOop mark = obj->mark(); 943 if (UseBiasedLocking && mark->has_bias_pattern()) { 944 // New allocated objects may have the mark set to anonymously biased. 945 // Also the deoptimized method may called methods with synchronization 946 // where the thread-local object is bias locked to the current thread. 947 assert(mark->is_biased_anonymously() || 948 mark->biased_locker() == thread, "should be locked to current thread"); 949 // Reset mark word to unbiased prototype. 950 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 951 obj->set_mark(unbiased_prototype); 952 } 953 BasicLock* lock = mon_info->lock(); 954 ObjectSynchronizer::slow_enter(obj, lock, thread); 955 } 956 assert(mon_info->owner()->is_locked(), "object must be locked now"); 957 } 958} 959 960 961#ifndef PRODUCT 962// print information about reallocated objects 963void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 964 fieldDescriptor fd; 965 966 for (int i = 0; i < objects->length(); i++) { 967 ObjectValue* sv = (ObjectValue*) objects->at(i); 968 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 969 Handle obj = sv->value(); 970 971 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 972 k->as_klassOop()->print_value(); 973 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 974 tty->cr(); 975 976 if (Verbose) { 977 k->oop_print_on(obj(), tty); 978 } 979 } 980} 981#endif 982#endif // COMPILER2 983 984vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 985 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 986 987#ifndef PRODUCT 988 if (TraceDeoptimization) { 989 ttyLocker ttyl; 990 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 991 fr.print_on(tty); 992 tty->print_cr(" Virtual frames (innermost first):"); 993 for (int index = 0; index < chunk->length(); index++) { 994 compiledVFrame* vf = chunk->at(index); 995 tty->print(" %2d - ", index); 996 vf->print_value(); 997 int bci = chunk->at(index)->raw_bci(); 998 const char* code_name; 999 if (bci == SynchronizationEntryBCI) { 1000 code_name = "sync entry"; 1001 } else { 1002 Bytecodes::Code code = vf->method()->code_at(bci); 1003 code_name = Bytecodes::name(code); 1004 } 1005 tty->print(" - %s", code_name); 1006 tty->print_cr(" @ bci %d ", bci); 1007 if (Verbose) { 1008 vf->print(); 1009 tty->cr(); 1010 } 1011 } 1012 } 1013#endif 1014 1015 // Register map for next frame (used for stack crawl). We capture 1016 // the state of the deopt'ing frame's caller. Thus if we need to 1017 // stuff a C2I adapter we can properly fill in the callee-save 1018 // register locations. 1019 frame caller = fr.sender(reg_map); 1020 int frame_size = caller.sp() - fr.sp(); 1021 1022 frame sender = caller; 1023 1024 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1025 // the vframeArray containing the unpacking information is allocated in the C heap. 1026 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1027 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1028 1029 // Compare the vframeArray to the collected vframes 1030 assert(array->structural_compare(thread, chunk), "just checking"); 1031 1032#ifndef PRODUCT 1033 if (TraceDeoptimization) { 1034 ttyLocker ttyl; 1035 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1036 } 1037#endif // PRODUCT 1038 1039 return array; 1040} 1041 1042 1043static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1044 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1045 for (int i = 0; i < monitors->length(); i++) { 1046 MonitorInfo* mon_info = monitors->at(i); 1047 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1048 objects_to_revoke->append(Handle(mon_info->owner())); 1049 } 1050 } 1051} 1052 1053 1054void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1055 if (!UseBiasedLocking) { 1056 return; 1057 } 1058 1059 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1060 1061 // Unfortunately we don't have a RegisterMap available in most of 1062 // the places we want to call this routine so we need to walk the 1063 // stack again to update the register map. 1064 if (map == NULL || !map->update_map()) { 1065 StackFrameStream sfs(thread, true); 1066 bool found = false; 1067 while (!found && !sfs.is_done()) { 1068 frame* cur = sfs.current(); 1069 sfs.next(); 1070 found = cur->id() == fr.id(); 1071 } 1072 assert(found, "frame to be deoptimized not found on target thread's stack"); 1073 map = sfs.register_map(); 1074 } 1075 1076 vframe* vf = vframe::new_vframe(&fr, map, thread); 1077 compiledVFrame* cvf = compiledVFrame::cast(vf); 1078 // Revoke monitors' biases in all scopes 1079 while (!cvf->is_top()) { 1080 collect_monitors(cvf, objects_to_revoke); 1081 cvf = compiledVFrame::cast(cvf->sender()); 1082 } 1083 collect_monitors(cvf, objects_to_revoke); 1084 1085 if (SafepointSynchronize::is_at_safepoint()) { 1086 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1087 } else { 1088 BiasedLocking::revoke(objects_to_revoke); 1089 } 1090} 1091 1092 1093void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1094 if (!UseBiasedLocking) { 1095 return; 1096 } 1097 1098 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1099 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1100 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1101 if (jt->has_last_Java_frame()) { 1102 StackFrameStream sfs(jt, true); 1103 while (!sfs.is_done()) { 1104 frame* cur = sfs.current(); 1105 if (cb->contains(cur->pc())) { 1106 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 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 sfs.next(); 1116 } 1117 } 1118 } 1119 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1120} 1121 1122 1123void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1124 assert(fr.can_be_deoptimized(), "checking frame type"); 1125 1126 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1127 1128 // Patch the nmethod so that when execution returns to it we will 1129 // deopt the execution state and return to the interpreter. 1130 fr.deoptimize(thread); 1131} 1132 1133void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1134 // Deoptimize only if the frame comes from compile code. 1135 // Do not deoptimize the frame which is already patched 1136 // during the execution of the loops below. 1137 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1138 return; 1139 } 1140 ResourceMark rm; 1141 DeoptimizationMarker dm; 1142 if (UseBiasedLocking) { 1143 revoke_biases_of_monitors(thread, fr, map); 1144 } 1145 deoptimize_single_frame(thread, fr); 1146 1147} 1148 1149 1150void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1151 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1152 "can only deoptimize other thread at a safepoint"); 1153 // Compute frame and register map based on thread and sp. 1154 RegisterMap reg_map(thread, UseBiasedLocking); 1155 frame fr = thread->last_frame(); 1156 while (fr.id() != id) { 1157 fr = fr.sender(®_map); 1158 } 1159 deoptimize(thread, fr, ®_map); 1160} 1161 1162 1163void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1164 if (thread == Thread::current()) { 1165 Deoptimization::deoptimize_frame_internal(thread, id); 1166 } else { 1167 VM_DeoptimizeFrame deopt(thread, id); 1168 VMThread::execute(&deopt); 1169 } 1170} 1171 1172 1173// JVMTI PopFrame support 1174JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1175{ 1176 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1177} 1178JRT_END 1179 1180 1181#if defined(COMPILER2) || defined(SHARK) 1182void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1183 // in case of an unresolved klass entry, load the class. 1184 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1185 klassOop tk = constant_pool->klass_at(index, CHECK); 1186 return; 1187 } 1188 1189 if (!constant_pool->tag_at(index).is_symbol()) return; 1190 1191 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1192 Symbol* symbol = constant_pool->symbol_at(index); 1193 1194 // class name? 1195 if (symbol->byte_at(0) != '(') { 1196 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1197 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1198 return; 1199 } 1200 1201 // then it must be a signature! 1202 ResourceMark rm(THREAD); 1203 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1204 if (ss.is_object()) { 1205 Symbol* class_name = ss.as_symbol(CHECK); 1206 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1207 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1208 } 1209 } 1210} 1211 1212 1213void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1214 EXCEPTION_MARK; 1215 load_class_by_index(constant_pool, index, THREAD); 1216 if (HAS_PENDING_EXCEPTION) { 1217 // Exception happened during classloading. We ignore the exception here, since it 1218 // is going to be rethrown since the current activation is going to be deoptimzied and 1219 // the interpreter will re-execute the bytecode. 1220 CLEAR_PENDING_EXCEPTION; 1221 } 1222} 1223 1224JRT_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_deopt_message(thread, "Uncommon trap %d 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 bool create_if_missing = ProfileTraps; 1270 1271 methodDataHandle trap_mdo 1272 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1273 1274 // Print a bunch of diagnostics, if requested. 1275 if (TraceDeoptimization || LogCompilation) { 1276 ResourceMark rm; 1277 ttyLocker ttyl; 1278 char buf[100]; 1279 if (xtty != NULL) { 1280 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1281 os::current_thread_id(), 1282 format_trap_request(buf, sizeof(buf), trap_request)); 1283 nm->log_identity(xtty); 1284 } 1285 Symbol* class_name = NULL; 1286 bool unresolved = false; 1287 if (unloaded_class_index >= 0) { 1288 constantPoolHandle constants (THREAD, trap_method->constants()); 1289 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1290 class_name = constants->klass_name_at(unloaded_class_index); 1291 unresolved = true; 1292 if (xtty != NULL) 1293 xtty->print(" unresolved='1'"); 1294 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1295 class_name = constants->symbol_at(unloaded_class_index); 1296 } 1297 if (xtty != NULL) 1298 xtty->name(class_name); 1299 } 1300 if (xtty != NULL && trap_mdo.not_null()) { 1301 // Dump the relevant MDO state. 1302 // This is the deopt count for the current reason, any previous 1303 // reasons or recompiles seen at this point. 1304 int dcnt = trap_mdo->trap_count(reason); 1305 if (dcnt != 0) 1306 xtty->print(" count='%d'", dcnt); 1307 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1308 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1309 if (dos != 0) { 1310 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1311 if (trap_state_is_recompiled(dos)) { 1312 int recnt2 = trap_mdo->overflow_recompile_count(); 1313 if (recnt2 != 0) 1314 xtty->print(" recompiles2='%d'", recnt2); 1315 } 1316 } 1317 } 1318 if (xtty != NULL) { 1319 xtty->stamp(); 1320 xtty->end_head(); 1321 } 1322 if (TraceDeoptimization) { // make noise on the tty 1323 tty->print("Uncommon trap occurred in"); 1324 nm->method()->print_short_name(tty); 1325 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1326 fr.pc(), 1327 (int) os::current_thread_id(), 1328 trap_reason_name(reason), 1329 trap_action_name(action), 1330 unloaded_class_index); 1331 if (class_name != NULL) { 1332 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1333 class_name->print_symbol_on(tty); 1334 } 1335 tty->cr(); 1336 } 1337 if (xtty != NULL) { 1338 // Log the precise location of the trap. 1339 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1340 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1341 xtty->method(sd->method()); 1342 xtty->end_elem(); 1343 if (sd->is_top()) break; 1344 } 1345 xtty->tail("uncommon_trap"); 1346 } 1347 } 1348 // (End diagnostic printout.) 1349 1350 // Load class if necessary 1351 if (unloaded_class_index >= 0) { 1352 constantPoolHandle constants(THREAD, trap_method->constants()); 1353 load_class_by_index(constants, unloaded_class_index); 1354 } 1355 1356 // Flush the nmethod if necessary and desirable. 1357 // 1358 // We need to avoid situations where we are re-flushing the nmethod 1359 // because of a hot deoptimization site. Repeated flushes at the same 1360 // point need to be detected by the compiler and avoided. If the compiler 1361 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1362 // module must take measures to avoid an infinite cycle of recompilation 1363 // and deoptimization. There are several such measures: 1364 // 1365 // 1. If a recompilation is ordered a second time at some site X 1366 // and for the same reason R, the action is adjusted to 'reinterpret', 1367 // to give the interpreter time to exercise the method more thoroughly. 1368 // If this happens, the method's overflow_recompile_count is incremented. 1369 // 1370 // 2. If the compiler fails to reduce the deoptimization rate, then 1371 // the method's overflow_recompile_count will begin to exceed the set 1372 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1373 // is adjusted to 'make_not_compilable', and the method is abandoned 1374 // to the interpreter. This is a performance hit for hot methods, 1375 // but is better than a disastrous infinite cycle of recompilations. 1376 // (Actually, only the method containing the site X is abandoned.) 1377 // 1378 // 3. In parallel with the previous measures, if the total number of 1379 // recompilations of a method exceeds the much larger set limit 1380 // PerMethodRecompilationCutoff, the method is abandoned. 1381 // This should only happen if the method is very large and has 1382 // many "lukewarm" deoptimizations. The code which enforces this 1383 // limit is elsewhere (class nmethod, class methodOopDesc). 1384 // 1385 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1386 // to recompile at each bytecode independently of the per-BCI cutoff. 1387 // 1388 // The decision to update code is up to the compiler, and is encoded 1389 // in the Action_xxx code. If the compiler requests Action_none 1390 // no trap state is changed, no compiled code is changed, and the 1391 // computation suffers along in the interpreter. 1392 // 1393 // The other action codes specify various tactics for decompilation 1394 // and recompilation. Action_maybe_recompile is the loosest, and 1395 // allows the compiled code to stay around until enough traps are seen, 1396 // and until the compiler gets around to recompiling the trapping method. 1397 // 1398 // The other actions cause immediate removal of the present code. 1399 1400 bool update_trap_state = true; 1401 bool make_not_entrant = false; 1402 bool make_not_compilable = false; 1403 bool reprofile = false; 1404 switch (action) { 1405 case Action_none: 1406 // Keep the old code. 1407 update_trap_state = false; 1408 break; 1409 case Action_maybe_recompile: 1410 // Do not need to invalidate the present code, but we can 1411 // initiate another 1412 // Start compiler without (necessarily) invalidating the nmethod. 1413 // The system will tolerate the old code, but new code should be 1414 // generated when possible. 1415 break; 1416 case Action_reinterpret: 1417 // Go back into the interpreter for a while, and then consider 1418 // recompiling form scratch. 1419 make_not_entrant = true; 1420 // Reset invocation counter for outer most method. 1421 // This will allow the interpreter to exercise the bytecodes 1422 // for a while before recompiling. 1423 // By contrast, Action_make_not_entrant is immediate. 1424 // 1425 // Note that the compiler will track null_check, null_assert, 1426 // range_check, and class_check events and log them as if they 1427 // had been traps taken from compiled code. This will update 1428 // the MDO trap history so that the next compilation will 1429 // properly detect hot trap sites. 1430 reprofile = true; 1431 break; 1432 case Action_make_not_entrant: 1433 // Request immediate recompilation, and get rid of the old code. 1434 // Make them not entrant, so next time they are called they get 1435 // recompiled. Unloaded classes are loaded now so recompile before next 1436 // time they are called. Same for uninitialized. The interpreter will 1437 // link the missing class, if any. 1438 make_not_entrant = true; 1439 break; 1440 case Action_make_not_compilable: 1441 // Give up on compiling this method at all. 1442 make_not_entrant = true; 1443 make_not_compilable = true; 1444 break; 1445 default: 1446 ShouldNotReachHere(); 1447 } 1448 1449 // Setting +ProfileTraps fixes the following, on all platforms: 1450 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1451 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1452 // recompile relies on a methodDataOop to record heroic opt failures. 1453 1454 // Whether the interpreter is producing MDO data or not, we also need 1455 // to use the MDO to detect hot deoptimization points and control 1456 // aggressive optimization. 1457 bool inc_recompile_count = false; 1458 ProfileData* pdata = NULL; 1459 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1460 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1461 uint this_trap_count = 0; 1462 bool maybe_prior_trap = false; 1463 bool maybe_prior_recompile = false; 1464 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1465 //outputs: 1466 this_trap_count, 1467 maybe_prior_trap, 1468 maybe_prior_recompile); 1469 // Because the interpreter also counts null, div0, range, and class 1470 // checks, these traps from compiled code are double-counted. 1471 // This is harmless; it just means that the PerXTrapLimit values 1472 // are in effect a little smaller than they look. 1473 1474 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1475 if (per_bc_reason != Reason_none) { 1476 // Now take action based on the partially known per-BCI history. 1477 if (maybe_prior_trap 1478 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1479 // If there are too many traps at this BCI, force a recompile. 1480 // This will allow the compiler to see the limit overflow, and 1481 // take corrective action, if possible. The compiler generally 1482 // does not use the exact PerBytecodeTrapLimit value, but instead 1483 // changes its tactics if it sees any traps at all. This provides 1484 // a little hysteresis, delaying a recompile until a trap happens 1485 // several times. 1486 // 1487 // Actually, since there is only one bit of counter per BCI, 1488 // the possible per-BCI counts are {0,1,(per-method count)}. 1489 // This produces accurate results if in fact there is only 1490 // one hot trap site, but begins to get fuzzy if there are 1491 // many sites. For example, if there are ten sites each 1492 // trapping two or more times, they each get the blame for 1493 // all of their traps. 1494 make_not_entrant = true; 1495 } 1496 1497 // Detect repeated recompilation at the same BCI, and enforce a limit. 1498 if (make_not_entrant && maybe_prior_recompile) { 1499 // More than one recompile at this point. 1500 inc_recompile_count = maybe_prior_trap; 1501 } 1502 } else { 1503 // For reasons which are not recorded per-bytecode, we simply 1504 // force recompiles unconditionally. 1505 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1506 make_not_entrant = true; 1507 } 1508 1509 // Go back to the compiler if there are too many traps in this method. 1510 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1511 // If there are too many traps in this method, force a recompile. 1512 // This will allow the compiler to see the limit overflow, and 1513 // take corrective action, if possible. 1514 // (This condition is an unlikely backstop only, because the 1515 // PerBytecodeTrapLimit is more likely to take effect first, 1516 // if it is applicable.) 1517 make_not_entrant = true; 1518 } 1519 1520 // Here's more hysteresis: If there has been a recompile at 1521 // this trap point already, run the method in the interpreter 1522 // for a while to exercise it more thoroughly. 1523 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1524 reprofile = true; 1525 } 1526 1527 } 1528 1529 // Take requested actions on the method: 1530 1531 // Recompile 1532 if (make_not_entrant) { 1533 if (!nm->make_not_entrant()) { 1534 return; // the call did not change nmethod's state 1535 } 1536 1537 if (pdata != NULL) { 1538 // Record the recompilation event, if any. 1539 int tstate0 = pdata->trap_state(); 1540 int tstate1 = trap_state_set_recompiled(tstate0, true); 1541 if (tstate1 != tstate0) 1542 pdata->set_trap_state(tstate1); 1543 } 1544 } 1545 1546 if (inc_recompile_count) { 1547 trap_mdo->inc_overflow_recompile_count(); 1548 if ((uint)trap_mdo->overflow_recompile_count() > 1549 (uint)PerBytecodeRecompilationCutoff) { 1550 // Give up on the method containing the bad BCI. 1551 if (trap_method() == nm->method()) { 1552 make_not_compilable = true; 1553 } else { 1554 trap_method->set_not_compilable(CompLevel_full_optimization); 1555 // But give grace to the enclosing nm->method(). 1556 } 1557 } 1558 } 1559 1560 // Reprofile 1561 if (reprofile) { 1562 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1563 } 1564 1565 // Give up compiling 1566 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1567 assert(make_not_entrant, "consistent"); 1568 nm->method()->set_not_compilable(CompLevel_full_optimization); 1569 } 1570 1571 } // Free marked resources 1572 1573} 1574JRT_END 1575 1576methodDataOop 1577Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1578 bool create_if_missing) { 1579 Thread* THREAD = thread; 1580 methodDataOop mdo = m()->method_data(); 1581 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1582 // Build an MDO. Ignore errors like OutOfMemory; 1583 // that simply means we won't have an MDO to update. 1584 methodOopDesc::build_interpreter_method_data(m, THREAD); 1585 if (HAS_PENDING_EXCEPTION) { 1586 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1587 CLEAR_PENDING_EXCEPTION; 1588 } 1589 mdo = m()->method_data(); 1590 } 1591 return mdo; 1592} 1593 1594ProfileData* 1595Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1596 int trap_bci, 1597 Deoptimization::DeoptReason reason, 1598 //outputs: 1599 uint& ret_this_trap_count, 1600 bool& ret_maybe_prior_trap, 1601 bool& ret_maybe_prior_recompile) { 1602 uint prior_trap_count = trap_mdo->trap_count(reason); 1603 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1604 1605 // If the runtime cannot find a place to store trap history, 1606 // it is estimated based on the general condition of the method. 1607 // If the method has ever been recompiled, or has ever incurred 1608 // a trap with the present reason , then this BCI is assumed 1609 // (pessimistically) to be the culprit. 1610 bool maybe_prior_trap = (prior_trap_count != 0); 1611 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1612 ProfileData* pdata = NULL; 1613 1614 1615 // For reasons which are recorded per bytecode, we check per-BCI data. 1616 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1617 if (per_bc_reason != Reason_none) { 1618 // Find the profile data for this BCI. If there isn't one, 1619 // try to allocate one from the MDO's set of spares. 1620 // This will let us detect a repeated trap at this point. 1621 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1622 1623 if (pdata != NULL) { 1624 // Query the trap state of this profile datum. 1625 int tstate0 = pdata->trap_state(); 1626 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1627 maybe_prior_trap = false; 1628 if (!trap_state_is_recompiled(tstate0)) 1629 maybe_prior_recompile = false; 1630 1631 // Update the trap state of this profile datum. 1632 int tstate1 = tstate0; 1633 // Record the reason. 1634 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1635 // Store the updated state on the MDO, for next time. 1636 if (tstate1 != tstate0) 1637 pdata->set_trap_state(tstate1); 1638 } else { 1639 if (LogCompilation && xtty != NULL) { 1640 ttyLocker ttyl; 1641 // Missing MDP? Leave a small complaint in the log. 1642 xtty->elem("missing_mdp bci='%d'", trap_bci); 1643 } 1644 } 1645 } 1646 1647 // Return results: 1648 ret_this_trap_count = this_trap_count; 1649 ret_maybe_prior_trap = maybe_prior_trap; 1650 ret_maybe_prior_recompile = maybe_prior_recompile; 1651 return pdata; 1652} 1653 1654void 1655Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1656 ResourceMark rm; 1657 // Ignored outputs: 1658 uint ignore_this_trap_count; 1659 bool ignore_maybe_prior_trap; 1660 bool ignore_maybe_prior_recompile; 1661 query_update_method_data(trap_mdo, trap_bci, 1662 (DeoptReason)reason, 1663 ignore_this_trap_count, 1664 ignore_maybe_prior_trap, 1665 ignore_maybe_prior_recompile); 1666} 1667 1668Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1669 1670 // Still in Java no safepoints 1671 { 1672 // This enters VM and may safepoint 1673 uncommon_trap_inner(thread, trap_request); 1674 } 1675 return fetch_unroll_info_helper(thread); 1676} 1677 1678// Local derived constants. 1679// Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1680const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1681const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1682 1683//---------------------------trap_state_reason--------------------------------- 1684Deoptimization::DeoptReason 1685Deoptimization::trap_state_reason(int trap_state) { 1686 // This assert provides the link between the width of DataLayout::trap_bits 1687 // and the encoding of "recorded" reasons. It ensures there are enough 1688 // bits to store all needed reasons in the per-BCI MDO profile. 1689 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1690 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1691 trap_state -= recompile_bit; 1692 if (trap_state == DS_REASON_MASK) { 1693 return Reason_many; 1694 } else { 1695 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1696 return (DeoptReason)trap_state; 1697 } 1698} 1699//-------------------------trap_state_has_reason------------------------------- 1700int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1701 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1702 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1703 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1704 trap_state -= recompile_bit; 1705 if (trap_state == DS_REASON_MASK) { 1706 return -1; // true, unspecifically (bottom of state lattice) 1707 } else if (trap_state == reason) { 1708 return 1; // true, definitely 1709 } else if (trap_state == 0) { 1710 return 0; // false, definitely (top of state lattice) 1711 } else { 1712 return 0; // false, definitely 1713 } 1714} 1715//-------------------------trap_state_add_reason------------------------------- 1716int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1717 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1718 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1719 trap_state -= recompile_bit; 1720 if (trap_state == DS_REASON_MASK) { 1721 return trap_state + recompile_bit; // already at state lattice bottom 1722 } else if (trap_state == reason) { 1723 return trap_state + recompile_bit; // the condition is already true 1724 } else if (trap_state == 0) { 1725 return reason + recompile_bit; // no condition has yet been true 1726 } else { 1727 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1728 } 1729} 1730//-----------------------trap_state_is_recompiled------------------------------ 1731bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1732 return (trap_state & DS_RECOMPILE_BIT) != 0; 1733} 1734//-----------------------trap_state_set_recompiled----------------------------- 1735int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1736 if (z) return trap_state | DS_RECOMPILE_BIT; 1737 else return trap_state & ~DS_RECOMPILE_BIT; 1738} 1739//---------------------------format_trap_state--------------------------------- 1740// This is used for debugging and diagnostics, including hotspot.log output. 1741const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1742 int trap_state) { 1743 DeoptReason reason = trap_state_reason(trap_state); 1744 bool recomp_flag = trap_state_is_recompiled(trap_state); 1745 // Re-encode the state from its decoded components. 1746 int decoded_state = 0; 1747 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1748 decoded_state = trap_state_add_reason(decoded_state, reason); 1749 if (recomp_flag) 1750 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1751 // If the state re-encodes properly, format it symbolically. 1752 // Because this routine is used for debugging and diagnostics, 1753 // be robust even if the state is a strange value. 1754 size_t len; 1755 if (decoded_state != trap_state) { 1756 // Random buggy state that doesn't decode?? 1757 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1758 } else { 1759 len = jio_snprintf(buf, buflen, "%s%s", 1760 trap_reason_name(reason), 1761 recomp_flag ? " recompiled" : ""); 1762 } 1763 if (len >= buflen) 1764 buf[buflen-1] = '\0'; 1765 return buf; 1766} 1767 1768 1769//--------------------------------statics-------------------------------------- 1770Deoptimization::DeoptAction Deoptimization::_unloaded_action 1771 = Deoptimization::Action_reinterpret; 1772const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1773 // Note: Keep this in sync. with enum DeoptReason. 1774 "none", 1775 "null_check", 1776 "null_assert", 1777 "range_check", 1778 "class_check", 1779 "array_check", 1780 "intrinsic", 1781 "bimorphic", 1782 "unloaded", 1783 "uninitialized", 1784 "unreached", 1785 "unhandled", 1786 "constraint", 1787 "div0_check", 1788 "age", 1789 "predicate", 1790 "loop_limit_check" 1791}; 1792const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1793 // Note: Keep this in sync. with enum DeoptAction. 1794 "none", 1795 "maybe_recompile", 1796 "reinterpret", 1797 "make_not_entrant", 1798 "make_not_compilable" 1799}; 1800 1801const char* Deoptimization::trap_reason_name(int reason) { 1802 if (reason == Reason_many) return "many"; 1803 if ((uint)reason < Reason_LIMIT) 1804 return _trap_reason_name[reason]; 1805 static char buf[20]; 1806 sprintf(buf, "reason%d", reason); 1807 return buf; 1808} 1809const char* Deoptimization::trap_action_name(int action) { 1810 if ((uint)action < Action_LIMIT) 1811 return _trap_action_name[action]; 1812 static char buf[20]; 1813 sprintf(buf, "action%d", action); 1814 return buf; 1815} 1816 1817// This is used for debugging and diagnostics, including hotspot.log output. 1818const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1819 int trap_request) { 1820 jint unloaded_class_index = trap_request_index(trap_request); 1821 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1822 const char* action = trap_action_name(trap_request_action(trap_request)); 1823 size_t len; 1824 if (unloaded_class_index < 0) { 1825 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1826 reason, action); 1827 } else { 1828 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1829 reason, action, unloaded_class_index); 1830 } 1831 if (len >= buflen) 1832 buf[buflen-1] = '\0'; 1833 return buf; 1834} 1835 1836juint Deoptimization::_deoptimization_hist 1837 [Deoptimization::Reason_LIMIT] 1838 [1 + Deoptimization::Action_LIMIT] 1839 [Deoptimization::BC_CASE_LIMIT] 1840 = {0}; 1841 1842enum { 1843 LSB_BITS = 8, 1844 LSB_MASK = right_n_bits(LSB_BITS) 1845}; 1846 1847void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1848 Bytecodes::Code bc) { 1849 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1850 assert(action >= 0 && action < Action_LIMIT, "oob"); 1851 _deoptimization_hist[Reason_none][0][0] += 1; // total 1852 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1853 juint* cases = _deoptimization_hist[reason][1+action]; 1854 juint* bc_counter_addr = NULL; 1855 juint bc_counter = 0; 1856 // Look for an unused counter, or an exact match to this BC. 1857 if (bc != Bytecodes::_illegal) { 1858 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1859 juint* counter_addr = &cases[bc_case]; 1860 juint counter = *counter_addr; 1861 if ((counter == 0 && bc_counter_addr == NULL) 1862 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1863 // this counter is either free or is already devoted to this BC 1864 bc_counter_addr = counter_addr; 1865 bc_counter = counter | bc; 1866 } 1867 } 1868 } 1869 if (bc_counter_addr == NULL) { 1870 // Overflow, or no given bytecode. 1871 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1872 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1873 } 1874 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1875} 1876 1877jint Deoptimization::total_deoptimization_count() { 1878 return _deoptimization_hist[Reason_none][0][0]; 1879} 1880 1881jint Deoptimization::deoptimization_count(DeoptReason reason) { 1882 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1883 return _deoptimization_hist[reason][0][0]; 1884} 1885 1886void Deoptimization::print_statistics() { 1887 juint total = total_deoptimization_count(); 1888 juint account = total; 1889 if (total != 0) { 1890 ttyLocker ttyl; 1891 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1892 tty->print_cr("Deoptimization traps recorded:"); 1893 #define PRINT_STAT_LINE(name, r) \ 1894 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1895 PRINT_STAT_LINE("total", total); 1896 // For each non-zero entry in the histogram, print the reason, 1897 // the action, and (if specifically known) the type of bytecode. 1898 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1899 for (int action = 0; action < Action_LIMIT; action++) { 1900 juint* cases = _deoptimization_hist[reason][1+action]; 1901 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1902 juint counter = cases[bc_case]; 1903 if (counter != 0) { 1904 char name[1*K]; 1905 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1906 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1907 bc = Bytecodes::_illegal; 1908 sprintf(name, "%s/%s/%s", 1909 trap_reason_name(reason), 1910 trap_action_name(action), 1911 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1912 juint r = counter >> LSB_BITS; 1913 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1914 account -= r; 1915 } 1916 } 1917 } 1918 } 1919 if (account != 0) { 1920 PRINT_STAT_LINE("unaccounted", account); 1921 } 1922 #undef PRINT_STAT_LINE 1923 if (xtty != NULL) xtty->tail("statistics"); 1924 } 1925} 1926#else // COMPILER2 || SHARK 1927 1928 1929// Stubs for C1 only system. 1930bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1931 return false; 1932} 1933 1934const char* Deoptimization::trap_reason_name(int reason) { 1935 return "unknown"; 1936} 1937 1938void Deoptimization::print_statistics() { 1939 // no output 1940} 1941 1942void 1943Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1944 // no udpate 1945} 1946 1947int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1948 return 0; 1949} 1950 1951void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1952 Bytecodes::Code bc) { 1953 // no update 1954} 1955 1956const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1957 int trap_state) { 1958 jio_snprintf(buf, buflen, "#%d", trap_state); 1959 return buf; 1960} 1961 1962#endif // COMPILER2 || SHARK 1963