1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_deoptimization.cpp.incl" 27 28 bool DeoptimizationMarker::_is_active = false; 29 30 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 31 int caller_adjustment, 32 int number_of_frames, 33 intptr_t* frame_sizes, 34 address* frame_pcs, 35 BasicType return_type) { 36 _size_of_deoptimized_frame = size_of_deoptimized_frame; 37 _caller_adjustment = caller_adjustment; 38 _number_of_frames = number_of_frames; 39 _frame_sizes = frame_sizes; 40 _frame_pcs = frame_pcs; 41 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); 42 _return_type = return_type; 43 // PD (x86 only) 44 _counter_temp = 0; 45 _initial_fp = 0; 46 _unpack_kind = 0; 47 _sender_sp_temp = 0; 48 49 _total_frame_sizes = size_of_frames(); 50 } 51 52 53 Deoptimization::UnrollBlock::~UnrollBlock() { 54 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 55 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 56 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 57 } 58 59 60 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 61 assert(register_number < RegisterMap::reg_count, "checking register number"); 62 return &_register_block[register_number * 2]; 63 } 64 65 66 67 int Deoptimization::UnrollBlock::size_of_frames() const { 68 // Acount first for the adjustment of the initial frame 69 int result = _caller_adjustment; 70 for (int index = 0; index < number_of_frames(); index++) { 71 result += frame_sizes()[index]; 72 } 73 return result; 74 } 75 76 77 void Deoptimization::UnrollBlock::print() { 78 ttyLocker ttyl; 79 tty->print_cr("UnrollBlock"); 80 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 81 tty->print( " frame_sizes: "); 82 for (int index = 0; index < number_of_frames(); index++) { 83 tty->print("%d ", frame_sizes()[index]); 84 } 85 tty->cr(); 86 } 87 88 89 // In order to make fetch_unroll_info work properly with escape 90 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 91 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 92 // of previously eliminated objects occurs in realloc_objects, which is 93 // called from the method fetch_unroll_info_helper below. 94 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 95 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 96 // but makes the entry a little slower. There is however a little dance we have to 97 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 98 99 // fetch_unroll_info() is called at the beginning of the deoptimization 100 // handler. Note this fact before we start generating temporary frames 101 // that can confuse an asynchronous stack walker. This counter is 102 // decremented at the end of unpack_frames(). 103 thread->inc_in_deopt_handler(); 104 105 return fetch_unroll_info_helper(thread); 106 JRT_END 107 108 109 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 110 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 111 112 // Note: there is a safepoint safety issue here. No matter whether we enter 113 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 114 // the vframeArray is created. 115 // 116 117 // Allocate our special deoptimization ResourceMark 118 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 119 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 120 thread->set_deopt_mark(dmark); 121 122 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 123 RegisterMap map(thread, true); 124 RegisterMap dummy_map(thread, false); 125 // Now get the deoptee with a valid map 126 frame deoptee = stub_frame.sender(&map); 127 128 // Create a growable array of VFrames where each VFrame represents an inlined 129 // Java frame. This storage is allocated with the usual system arena. 130 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 131 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 132 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 133 while (!vf->is_top()) { 134 assert(vf->is_compiled_frame(), "Wrong frame type"); 135 chunk->push(compiledVFrame::cast(vf)); 136 vf = vf->sender(); 137 } 138 assert(vf->is_compiled_frame(), "Wrong frame type"); 139 chunk->push(compiledVFrame::cast(vf)); 140 141 #ifdef COMPILER2 142 // Reallocate the non-escaping objects and restore their fields. Then 143 // relock objects if synchronization on them was eliminated. 144 if (DoEscapeAnalysis) { 145 if (EliminateAllocations) { 146 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 147 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 148 bool reallocated = false; 149 if (objects != NULL) { 150 JRT_BLOCK 151 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 152 JRT_END 153 } 154 if (reallocated) { 155 reassign_fields(&deoptee, &map, objects); 156 #ifndef PRODUCT 157 if (TraceDeoptimization) { 158 ttyLocker ttyl; 159 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 160 print_objects(objects); 161 } 162 #endif 163 } 164 } 165 if (EliminateLocks) { 166 #ifndef PRODUCT 167 bool first = true; 168 #endif 169 for (int i = 0; i < chunk->length(); i++) { 170 compiledVFrame* cvf = chunk->at(i); 171 assert (cvf->scope() != NULL,"expect only compiled java frames"); 172 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 173 if (monitors->is_nonempty()) { 174 relock_objects(monitors, thread); 175 #ifndef PRODUCT 176 if (TraceDeoptimization) { 177 ttyLocker ttyl; 178 for (int j = 0; j < monitors->length(); j++) { 179 MonitorInfo* mi = monitors->at(j); 180 if (mi->eliminated()) { 181 if (first) { 182 first = false; 183 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 184 } 185 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 186 } 187 } 188 } 189 #endif 190 } 191 } 192 } 193 } 194 #endif // COMPILER2 195 // Ensure that no safepoint is taken after pointers have been stored 196 // in fields of rematerialized objects. If a safepoint occurs from here on 197 // out the java state residing in the vframeArray will be missed. 198 No_Safepoint_Verifier no_safepoint; 199 200 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 201 202 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 203 thread->set_vframe_array_head(array); 204 205 // Now that the vframeArray has been created if we have any deferred local writes 206 // added by jvmti then we can free up that structure as the data is now in the 207 // vframeArray 208 209 if (thread->deferred_locals() != NULL) { 210 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 211 int i = 0; 212 do { 213 // Because of inlining we could have multiple vframes for a single frame 214 // and several of the vframes could have deferred writes. Find them all. 215 if (list->at(i)->id() == array->original().id()) { 216 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 217 list->remove_at(i); 218 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 219 delete dlv; 220 } else { 221 i++; 222 } 223 } while ( i < list->length() ); 224 if (list->length() == 0) { 225 thread->set_deferred_locals(NULL); 226 // free the list and elements back to C heap. 227 delete list; 228 } 229 230 } 231 232 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 233 CodeBlob* cb = stub_frame.cb(); 234 // Verify we have the right vframeArray 235 assert(cb->frame_size() >= 0, "Unexpected frame size"); 236 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 237 238 #ifdef ASSERT 239 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 240 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 241 #endif 242 // This is a guarantee instead of an assert because if vframe doesn't match 243 // we will unpack the wrong deoptimized frame and wind up in strange places 244 // where it will be very difficult to figure out what went wrong. Better 245 // to die an early death here than some very obscure death later when the 246 // trail is cold. 247 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 248 // in that it will fail to detect a problem when there is one. This needs 249 // more work in tiger timeframe. 250 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 251 252 int number_of_frames = array->frames(); 253 254 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 255 // virtual activation, which is the reverse of the elements in the vframes array. 256 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 257 // +1 because we always have an interpreter return address for the final slot. 258 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 259 int callee_parameters = 0; 260 int callee_locals = 0; 261 int popframe_extra_args = 0; 262 // Create an interpreter return address for the stub to use as its return 263 // address so the skeletal frames are perfectly walkable 264 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 265 266 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 267 // activation be put back on the expression stack of the caller for reexecution 268 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 269 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 270 } 271 272 // 273 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 274 // frame_sizes/frame_pcs[1] next oldest frame (int) 275 // frame_sizes/frame_pcs[n] youngest frame (int) 276 // 277 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 278 // owns the space for the return address to it's caller). Confusing ain't it. 279 // 280 // The vframe array can address vframes with indices running from 281 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 282 // When we create the skeletal frames we need the oldest frame to be in the zero slot 283 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 284 // so things look a little strange in this loop. 285 // 286 for (int index = 0; index < array->frames(); index++ ) { 287 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 288 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 289 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 290 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 291 callee_locals, 292 index == 0, 293 popframe_extra_args); 294 // This pc doesn't have to be perfect just good enough to identify the frame 295 // as interpreted so the skeleton frame will be walkable 296 // The correct pc will be set when the skeleton frame is completely filled out 297 // The final pc we store in the loop is wrong and will be overwritten below 298 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 299 300 callee_parameters = array->element(index)->method()->size_of_parameters(); 301 callee_locals = array->element(index)->method()->max_locals(); 302 popframe_extra_args = 0; 303 } 304 305 // Compute whether the root vframe returns a float or double value. 306 BasicType return_type; 307 { 308 HandleMark hm; 309 methodHandle method(thread, array->element(0)->method()); 310 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci()); 311 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL; 312 } 313 314 // Compute information for handling adapters and adjusting the frame size of the caller. 315 int caller_adjustment = 0; 316 317 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 318 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 319 // than simply use array->sender.pc(). This requires us to walk the current set of frames 320 // 321 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 322 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 323 324 // Compute the amount the oldest interpreter frame will have to adjust 325 // its caller's stack by. If the caller is a compiled frame then 326 // we pretend that the callee has no parameters so that the 327 // extension counts for the full amount of locals and not just 328 // locals-parms. This is because without a c2i adapter the parm 329 // area as created by the compiled frame will not be usable by 330 // the interpreter. (Depending on the calling convention there 331 // may not even be enough space). 332 333 // QQQ I'd rather see this pushed down into last_frame_adjust 334 // and have it take the sender (aka caller). 335 336 if (deopt_sender.is_compiled_frame()) { 337 caller_adjustment = last_frame_adjust(0, callee_locals); 338 } else if (callee_locals > callee_parameters) { 339 // The caller frame may need extending to accommodate 340 // non-parameter locals of the first unpacked interpreted frame. 341 // Compute that adjustment. 342 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 343 } 344 345 346 // If the sender is deoptimized the we must retrieve the address of the handler 347 // since the frame will "magically" show the original pc before the deopt 348 // and we'd undo the deopt. 349 350 frame_pcs[0] = deopt_sender.raw_pc(); 351 352 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 353 354 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 355 caller_adjustment * BytesPerWord, 356 number_of_frames, 357 frame_sizes, 358 frame_pcs, 359 return_type); 360 #if defined(IA32) || defined(AMD64) 361 // We need a way to pass fp to the unpacking code so the skeletal frames 362 // come out correct. This is only needed for x86 because of c2 using ebp 363 // as an allocatable register. So this update is useless (and harmless) 364 // on the other platforms. It would be nice to do this in a different 365 // way but even the old style deoptimization had a problem with deriving 366 // this value. NEEDS_CLEANUP 367 // Note: now that c1 is using c2's deopt blob we must do this on all 368 // x86 based platforms 369 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); 370 *fp_addr = array->sender().fp(); // was adapter_caller 371 #endif /* IA32 || AMD64 */ 372 373 if (array->frames() > 1) { 374 if (VerifyStack && TraceDeoptimization) { 375 tty->print_cr("Deoptimizing method containing inlining"); 376 } 377 } 378 379 array->set_unroll_block(info); 380 return info; 381 } 382 383 // Called to cleanup deoptimization data structures in normal case 384 // after unpacking to stack and when stack overflow error occurs 385 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 386 vframeArray *array) { 387 388 // Get array if coming from exception 389 if (array == NULL) { 390 array = thread->vframe_array_head(); 391 } 392 thread->set_vframe_array_head(NULL); 393 394 // Free the previous UnrollBlock 395 vframeArray* old_array = thread->vframe_array_last(); 396 thread->set_vframe_array_last(array); 397 398 if (old_array != NULL) { 399 UnrollBlock* old_info = old_array->unroll_block(); 400 old_array->set_unroll_block(NULL); 401 delete old_info; 402 delete old_array; 403 } 404 405 // Deallocate any resource creating in this routine and any ResourceObjs allocated 406 // inside the vframeArray (StackValueCollections) 407 408 delete thread->deopt_mark(); 409 thread->set_deopt_mark(NULL); 410 411 412 if (JvmtiExport::can_pop_frame()) { 413 #ifndef CC_INTERP 414 // Regardless of whether we entered this routine with the pending 415 // popframe condition bit set, we should always clear it now 416 thread->clear_popframe_condition(); 417 #else 418 // C++ interpeter will clear has_pending_popframe when it enters 419 // with method_resume. For deopt_resume2 we clear it now. 420 if (thread->popframe_forcing_deopt_reexecution()) 421 thread->clear_popframe_condition(); 422 #endif /* CC_INTERP */ 423 } 424 425 // unpack_frames() is called at the end of the deoptimization handler 426 // and (in C2) at the end of the uncommon trap handler. Note this fact 427 // so that an asynchronous stack walker can work again. This counter is 428 // incremented at the beginning of fetch_unroll_info() and (in C2) at 429 // the beginning of uncommon_trap(). 430 thread->dec_in_deopt_handler(); 431 } 432 433 434 // Return BasicType of value being returned 435 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 436 437 // We are already active int he special DeoptResourceMark any ResourceObj's we 438 // allocate will be freed at the end of the routine. 439 440 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 441 // but makes the entry a little slower. There is however a little dance we have to 442 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 443 ResetNoHandleMark rnhm; // No-op in release/product versions 444 HandleMark hm; 445 446 frame stub_frame = thread->last_frame(); 447 448 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 449 // must point to the vframeArray for the unpack frame. 450 vframeArray* array = thread->vframe_array_head(); 451 452 #ifndef PRODUCT 453 if (TraceDeoptimization) { 454 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 455 } 456 #endif 457 458 UnrollBlock* info = array->unroll_block(); 459 460 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 461 array->unpack_to_stack(stub_frame, exec_mode); 462 463 BasicType bt = info->return_type(); 464 465 // If we have an exception pending, claim that the return type is an oop 466 // so the deopt_blob does not overwrite the exception_oop. 467 468 if (exec_mode == Unpack_exception) 469 bt = T_OBJECT; 470 471 // Cleanup thread deopt data 472 cleanup_deopt_info(thread, array); 473 474 #ifndef PRODUCT 475 if (VerifyStack) { 476 ResourceMark res_mark; 477 478 // Verify that the just-unpacked frames match the interpreter's 479 // notions of expression stack and locals 480 vframeArray* cur_array = thread->vframe_array_last(); 481 RegisterMap rm(thread, false); 482 rm.set_include_argument_oops(false); 483 bool is_top_frame = true; 484 int callee_size_of_parameters = 0; 485 int callee_max_locals = 0; 486 for (int i = 0; i < cur_array->frames(); i++) { 487 vframeArrayElement* el = cur_array->element(i); 488 frame* iframe = el->iframe(); 489 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 490 491 // Get the oop map for this bci 492 InterpreterOopMap mask; 493 int cur_invoke_parameter_size = 0; 494 bool try_next_mask = false; 495 int next_mask_expression_stack_size = -1; 496 int top_frame_expression_stack_adjustment = 0; 497 methodHandle mh(thread, iframe->interpreter_frame_method()); 498 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 499 BytecodeStream str(mh); 500 str.set_start(iframe->interpreter_frame_bci()); 501 int max_bci = mh->code_size(); 502 // Get to the next bytecode if possible 503 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 504 // Check to see if we can grab the number of outgoing arguments 505 // at an uncommon trap for an invoke (where the compiler 506 // generates debug info before the invoke has executed) 507 Bytecodes::Code cur_code = str.next(); 508 if (cur_code == Bytecodes::_invokevirtual || 509 cur_code == Bytecodes::_invokespecial || 510 cur_code == Bytecodes::_invokestatic || 511 cur_code == Bytecodes::_invokeinterface) { 512 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci()); 513 symbolHandle signature(thread, invoke->signature()); 514 ArgumentSizeComputer asc(signature); 515 cur_invoke_parameter_size = asc.size(); 516 if (cur_code != Bytecodes::_invokestatic) { 517 // Add in receiver 518 ++cur_invoke_parameter_size; 519 } 520 } 521 if (str.bci() < max_bci) { 522 Bytecodes::Code bc = str.next(); 523 if (bc >= 0) { 524 // The interpreter oop map generator reports results before 525 // the current bytecode has executed except in the case of 526 // calls. It seems to be hard to tell whether the compiler 527 // has emitted debug information matching the "state before" 528 // a given bytecode or the state after, so we try both 529 switch (cur_code) { 530 case Bytecodes::_invokevirtual: 531 case Bytecodes::_invokespecial: 532 case Bytecodes::_invokestatic: 533 case Bytecodes::_invokeinterface: 534 case Bytecodes::_athrow: 535 break; 536 default: { 537 InterpreterOopMap next_mask; 538 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 539 next_mask_expression_stack_size = next_mask.expression_stack_size(); 540 // Need to subtract off the size of the result type of 541 // the bytecode because this is not described in the 542 // debug info but returned to the interpreter in the TOS 543 // caching register 544 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 545 if (bytecode_result_type != T_ILLEGAL) { 546 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 547 } 548 assert(top_frame_expression_stack_adjustment >= 0, ""); 549 try_next_mask = true; 550 break; 551 } 552 } 553 } 554 } 555 556 // Verify stack depth and oops in frame 557 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 558 if (!( 559 /* SPARC */ 560 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 561 /* x86 */ 562 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 563 (try_next_mask && 564 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 565 top_frame_expression_stack_adjustment))) || 566 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 567 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 568 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 569 )) { 570 ttyLocker ttyl; 571 572 // Print out some information that will help us debug the problem 573 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 574 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 575 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 576 iframe->interpreter_frame_expression_stack_size()); 577 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 578 tty->print_cr(" try_next_mask = %d", try_next_mask); 579 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 580 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 581 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 582 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 583 tty->print_cr(" exec_mode = %d", exec_mode); 584 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 585 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 586 tty->print_cr(" Interpreted frames:"); 587 for (int k = 0; k < cur_array->frames(); k++) { 588 vframeArrayElement* el = cur_array->element(k); 589 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 590 } 591 cur_array->print_on_2(tty); 592 guarantee(false, "wrong number of expression stack elements during deopt"); 593 } 594 VerifyOopClosure verify; 595 iframe->oops_interpreted_do(&verify, &rm, false); 596 callee_size_of_parameters = mh->size_of_parameters(); 597 callee_max_locals = mh->max_locals(); 598 is_top_frame = false; 599 } 600 } 601 #endif /* !PRODUCT */ 602 603 604 return bt; 605 JRT_END 606 607 608 int Deoptimization::deoptimize_dependents() { 609 Threads::deoptimized_wrt_marked_nmethods(); 610 return 0; 611 } 612 613 614 #ifdef COMPILER2 615 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 616 Handle pending_exception(thread->pending_exception()); 617 const char* exception_file = thread->exception_file(); 618 int exception_line = thread->exception_line(); 619 thread->clear_pending_exception(); 620 621 for (int i = 0; i < objects->length(); i++) { 622 assert(objects->at(i)->is_object(), "invalid debug information"); 623 ObjectValue* sv = (ObjectValue*) objects->at(i); 624 625 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 626 oop obj = NULL; 627 628 if (k->oop_is_instance()) { 629 instanceKlass* ik = instanceKlass::cast(k()); 630 obj = ik->allocate_instance(CHECK_(false)); 631 } else if (k->oop_is_typeArray()) { 632 typeArrayKlass* ak = typeArrayKlass::cast(k()); 633 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 634 int len = sv->field_size() / type2size[ak->element_type()]; 635 obj = ak->allocate(len, CHECK_(false)); 636 } else if (k->oop_is_objArray()) { 637 objArrayKlass* ak = objArrayKlass::cast(k()); 638 obj = ak->allocate(sv->field_size(), CHECK_(false)); 639 } 640 641 assert(obj != NULL, "allocation failed"); 642 assert(sv->value().is_null(), "redundant reallocation"); 643 sv->set_value(obj); 644 } 645 646 if (pending_exception.not_null()) { 647 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 648 } 649 650 return true; 651 } 652 653 // This assumes that the fields are stored in ObjectValue in the same order 654 // they are yielded by do_nonstatic_fields. 655 class FieldReassigner: public FieldClosure { 656 frame* _fr; 657 RegisterMap* _reg_map; 658 ObjectValue* _sv; 659 instanceKlass* _ik; 660 oop _obj; 661 662 int _i; 663 public: 664 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 665 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 666 667 int i() const { return _i; } 668 669 670 void do_field(fieldDescriptor* fd) { 671 intptr_t val; 672 StackValue* value = 673 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 674 int offset = fd->offset(); 675 switch (fd->field_type()) { 676 case T_OBJECT: case T_ARRAY: 677 assert(value->type() == T_OBJECT, "Agreement."); 678 _obj->obj_field_put(offset, value->get_obj()()); 679 break; 680 681 case T_LONG: case T_DOUBLE: { 682 assert(value->type() == T_INT, "Agreement."); 683 StackValue* low = 684 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 685 #ifdef _LP64 686 jlong res = (jlong)low->get_int(); 687 #else 688 #ifdef SPARC 689 // For SPARC we have to swap high and low words. 690 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 691 #else 692 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 693 #endif //SPARC 694 #endif 695 _obj->long_field_put(offset, res); 696 break; 697 } 698 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 699 case T_INT: case T_FLOAT: // 4 bytes. 700 assert(value->type() == T_INT, "Agreement."); 701 val = value->get_int(); 702 _obj->int_field_put(offset, (jint)*((jint*)&val)); 703 break; 704 705 case T_SHORT: case T_CHAR: // 2 bytes 706 assert(value->type() == T_INT, "Agreement."); 707 val = value->get_int(); 708 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 709 break; 710 711 case T_BOOLEAN: case T_BYTE: // 1 byte 712 assert(value->type() == T_INT, "Agreement."); 713 val = value->get_int(); 714 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 715 break; 716 717 default: 718 ShouldNotReachHere(); 719 } 720 _i++; 721 } 722 }; 723 724 // restore elements of an eliminated type array 725 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 726 int index = 0; 727 intptr_t val; 728 729 for (int i = 0; i < sv->field_size(); i++) { 730 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 731 switch(type) { 732 case T_LONG: case T_DOUBLE: { 733 assert(value->type() == T_INT, "Agreement."); 734 StackValue* low = 735 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 736 #ifdef _LP64 737 jlong res = (jlong)low->get_int(); 738 #else 739 #ifdef SPARC 740 // For SPARC we have to swap high and low words. 741 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 742 #else 743 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 744 #endif //SPARC 745 #endif 746 obj->long_at_put(index, res); 747 break; 748 } 749 750 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 751 case T_INT: case T_FLOAT: // 4 bytes. 752 assert(value->type() == T_INT, "Agreement."); 753 val = value->get_int(); 754 obj->int_at_put(index, (jint)*((jint*)&val)); 755 break; 756 757 case T_SHORT: case T_CHAR: // 2 bytes 758 assert(value->type() == T_INT, "Agreement."); 759 val = value->get_int(); 760 obj->short_at_put(index, (jshort)*((jint*)&val)); 761 break; 762 763 case T_BOOLEAN: case T_BYTE: // 1 byte 764 assert(value->type() == T_INT, "Agreement."); 765 val = value->get_int(); 766 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 767 break; 768 769 default: 770 ShouldNotReachHere(); 771 } 772 index++; 773 } 774 } 775 776 777 // restore fields of an eliminated object array 778 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 779 for (int i = 0; i < sv->field_size(); i++) { 780 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 781 assert(value->type() == T_OBJECT, "object element expected"); 782 obj->obj_at_put(i, value->get_obj()()); 783 } 784 } 785 786 787 // restore fields of all eliminated objects and arrays 788 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 789 for (int i = 0; i < objects->length(); i++) { 790 ObjectValue* sv = (ObjectValue*) objects->at(i); 791 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 792 Handle obj = sv->value(); 793 assert(obj.not_null(), "reallocation was missed"); 794 795 if (k->oop_is_instance()) { 796 instanceKlass* ik = instanceKlass::cast(k()); 797 FieldReassigner reassign(fr, reg_map, sv, obj()); 798 ik->do_nonstatic_fields(&reassign); 799 } else if (k->oop_is_typeArray()) { 800 typeArrayKlass* ak = typeArrayKlass::cast(k()); 801 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 802 } else if (k->oop_is_objArray()) { 803 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 804 } 805 } 806 } 807 808 809 // relock objects for which synchronization was eliminated 810 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 811 for (int i = 0; i < monitors->length(); i++) { 812 MonitorInfo* mon_info = monitors->at(i); 813 if (mon_info->eliminated()) { 814 assert(mon_info->owner() != NULL, "reallocation was missed"); 815 Handle obj = Handle(mon_info->owner()); 816 markOop mark = obj->mark(); 817 if (UseBiasedLocking && mark->has_bias_pattern()) { 818 // New allocated objects may have the mark set to anonymously biased. 819 // Also the deoptimized method may called methods with synchronization 820 // where the thread-local object is bias locked to the current thread. 821 assert(mark->is_biased_anonymously() || 822 mark->biased_locker() == thread, "should be locked to current thread"); 823 // Reset mark word to unbiased prototype. 824 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 825 obj->set_mark(unbiased_prototype); 826 } 827 BasicLock* lock = mon_info->lock(); 828 ObjectSynchronizer::slow_enter(obj, lock, thread); 829 } 830 assert(mon_info->owner()->is_locked(), "object must be locked now"); 831 } 832 } 833 834 835 #ifndef PRODUCT 836 // print information about reallocated objects 837 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 838 fieldDescriptor fd; 839 840 for (int i = 0; i < objects->length(); i++) { 841 ObjectValue* sv = (ObjectValue*) objects->at(i); 842 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 843 Handle obj = sv->value(); 844 845 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 846 k->as_klassOop()->print_value(); 847 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 848 tty->cr(); 849 850 if (Verbose) { 851 k->oop_print_on(obj(), tty); 852 } 853 } 854 } 855 #endif 856 #endif // COMPILER2 857 858 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 859 860 #ifndef PRODUCT 861 if (TraceDeoptimization) { 862 ttyLocker ttyl; 863 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 864 fr.print_on(tty); 865 tty->print_cr(" Virtual frames (innermost first):"); 866 for (int index = 0; index < chunk->length(); index++) { 867 compiledVFrame* vf = chunk->at(index); 868 tty->print(" %2d - ", index); 869 vf->print_value(); 870 int bci = chunk->at(index)->raw_bci(); 871 const char* code_name; 872 if (bci == SynchronizationEntryBCI) { 873 code_name = "sync entry"; 874 } else { 875 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci); 876 code_name = Bytecodes::name(code); 877 } 878 tty->print(" - %s", code_name); 879 tty->print_cr(" @ bci %d ", bci); 880 if (Verbose) { 881 vf->print(); 882 tty->cr(); 883 } 884 } 885 } 886 #endif 887 888 // Register map for next frame (used for stack crawl). We capture 889 // the state of the deopt'ing frame's caller. Thus if we need to 890 // stuff a C2I adapter we can properly fill in the callee-save 891 // register locations. 892 frame caller = fr.sender(reg_map); 893 int frame_size = caller.sp() - fr.sp(); 894 895 frame sender = caller; 896 897 // Since the Java thread being deoptimized will eventually adjust it's own stack, 898 // the vframeArray containing the unpacking information is allocated in the C heap. 899 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 900 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 901 902 // Compare the vframeArray to the collected vframes 903 assert(array->structural_compare(thread, chunk), "just checking"); 904 Events::log("# vframes = %d", (intptr_t)chunk->length()); 905 906 #ifndef PRODUCT 907 if (TraceDeoptimization) { 908 ttyLocker ttyl; 909 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 910 if (Verbose) { 911 int count = 0; 912 // this used to leak deoptimizedVFrame like it was going out of style!!! 913 for (int index = 0; index < array->frames(); index++ ) { 914 vframeArrayElement* e = array->element(index); 915 e->print(tty); 916 917 /* 918 No printing yet. 919 array->vframe_at(index)->print_activation(count++); 920 // better as... 921 array->print_activation_for(index, count++); 922 */ 923 } 924 } 925 } 926 #endif // PRODUCT 927 928 return array; 929 } 930 931 932 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 933 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 934 for (int i = 0; i < monitors->length(); i++) { 935 MonitorInfo* mon_info = monitors->at(i); 936 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 937 objects_to_revoke->append(Handle(mon_info->owner())); 938 } 939 } 940 } 941 942 943 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 944 if (!UseBiasedLocking) { 945 return; 946 } 947 948 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 949 950 // Unfortunately we don't have a RegisterMap available in most of 951 // the places we want to call this routine so we need to walk the 952 // stack again to update the register map. 953 if (map == NULL || !map->update_map()) { 954 StackFrameStream sfs(thread, true); 955 bool found = false; 956 while (!found && !sfs.is_done()) { 957 frame* cur = sfs.current(); 958 sfs.next(); 959 found = cur->id() == fr.id(); 960 } 961 assert(found, "frame to be deoptimized not found on target thread's stack"); 962 map = sfs.register_map(); 963 } 964 965 vframe* vf = vframe::new_vframe(&fr, map, thread); 966 compiledVFrame* cvf = compiledVFrame::cast(vf); 967 // Revoke monitors' biases in all scopes 968 while (!cvf->is_top()) { 969 collect_monitors(cvf, objects_to_revoke); 970 cvf = compiledVFrame::cast(cvf->sender()); 971 } 972 collect_monitors(cvf, objects_to_revoke); 973 974 if (SafepointSynchronize::is_at_safepoint()) { 975 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 976 } else { 977 BiasedLocking::revoke(objects_to_revoke); 978 } 979 } 980 981 982 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 983 if (!UseBiasedLocking) { 984 return; 985 } 986 987 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 988 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 989 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 990 if (jt->has_last_Java_frame()) { 991 StackFrameStream sfs(jt, true); 992 while (!sfs.is_done()) { 993 frame* cur = sfs.current(); 994 if (cb->contains(cur->pc())) { 995 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 996 compiledVFrame* cvf = compiledVFrame::cast(vf); 997 // Revoke monitors' biases in all scopes 998 while (!cvf->is_top()) { 999 collect_monitors(cvf, objects_to_revoke); 1000 cvf = compiledVFrame::cast(cvf->sender()); 1001 } 1002 collect_monitors(cvf, objects_to_revoke); 1003 } 1004 sfs.next(); 1005 } 1006 } 1007 } 1008 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1009 } 1010 1011 1012 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1013 assert(fr.can_be_deoptimized(), "checking frame type"); 1014 1015 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1016 1017 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1018 1019 // Patch the nmethod so that when execution returns to it we will 1020 // deopt the execution state and return to the interpreter. 1021 fr.deoptimize(thread); 1022 } 1023 1024 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1025 // Deoptimize only if the frame comes from compile code. 1026 // Do not deoptimize the frame which is already patched 1027 // during the execution of the loops below. 1028 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1029 return; 1030 } 1031 ResourceMark rm; 1032 DeoptimizationMarker dm; 1033 if (UseBiasedLocking) { 1034 revoke_biases_of_monitors(thread, fr, map); 1035 } 1036 deoptimize_single_frame(thread, fr); 1037 1038 } 1039 1040 1041 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1042 // Compute frame and register map based on thread and sp. 1043 RegisterMap reg_map(thread, UseBiasedLocking); 1044 frame fr = thread->last_frame(); 1045 while (fr.id() != id) { 1046 fr = fr.sender(®_map); 1047 } 1048 deoptimize(thread, fr, ®_map); 1049 } 1050 1051 1052 // JVMTI PopFrame support 1053 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1054 { 1055 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1056 } 1057 JRT_END 1058 1059 1060 #ifdef COMPILER2 1061 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1062 // in case of an unresolved klass entry, load the class. 1063 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1064 klassOop tk = constant_pool->klass_at(index, CHECK); 1065 return; 1066 } 1067 1068 if (!constant_pool->tag_at(index).is_symbol()) return; 1069 1070 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1071 symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); 1072 1073 // class name? 1074 if (symbol->byte_at(0) != '(') { 1075 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1076 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1077 return; 1078 } 1079 1080 // then it must be a signature! 1081 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1082 if (ss.is_object()) { 1083 symbolOop s = ss.as_symbol(CHECK); 1084 symbolHandle class_name (THREAD, s); 1085 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1086 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1087 } 1088 } 1089 } 1090 1091 1092 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1093 EXCEPTION_MARK; 1094 load_class_by_index(constant_pool, index, THREAD); 1095 if (HAS_PENDING_EXCEPTION) { 1096 // Exception happened during classloading. We ignore the exception here, since it 1097 // is going to be rethrown since the current activation is going to be deoptimzied and 1098 // the interpreter will re-execute the bytecode. 1099 CLEAR_PENDING_EXCEPTION; 1100 } 1101 } 1102 1103 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1104 HandleMark hm; 1105 1106 // uncommon_trap() is called at the beginning of the uncommon trap 1107 // handler. Note this fact before we start generating temporary frames 1108 // that can confuse an asynchronous stack walker. This counter is 1109 // decremented at the end of unpack_frames(). 1110 thread->inc_in_deopt_handler(); 1111 1112 // We need to update the map if we have biased locking. 1113 RegisterMap reg_map(thread, UseBiasedLocking); 1114 frame stub_frame = thread->last_frame(); 1115 frame fr = stub_frame.sender(®_map); 1116 // Make sure the calling nmethod is not getting deoptimized and removed 1117 // before we are done with it. 1118 nmethodLocker nl(fr.pc()); 1119 1120 { 1121 ResourceMark rm; 1122 1123 // Revoke biases of any monitors in the frame to ensure we can migrate them 1124 revoke_biases_of_monitors(thread, fr, ®_map); 1125 1126 DeoptReason reason = trap_request_reason(trap_request); 1127 DeoptAction action = trap_request_action(trap_request); 1128 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1129 1130 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1131 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1132 compiledVFrame* cvf = compiledVFrame::cast(vf); 1133 1134 nmethod* nm = cvf->code(); 1135 1136 ScopeDesc* trap_scope = cvf->scope(); 1137 methodHandle trap_method = trap_scope->method(); 1138 int trap_bci = trap_scope->bci(); 1139 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code(); 1140 1141 // Record this event in the histogram. 1142 gather_statistics(reason, action, trap_bc); 1143 1144 // Ensure that we can record deopt. history: 1145 bool create_if_missing = ProfileTraps; 1146 1147 methodDataHandle trap_mdo 1148 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1149 1150 // Print a bunch of diagnostics, if requested. 1151 if (TraceDeoptimization || LogCompilation) { 1152 ResourceMark rm; 1153 ttyLocker ttyl; 1154 char buf[100]; 1155 if (xtty != NULL) { 1156 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1157 os::current_thread_id(), 1158 format_trap_request(buf, sizeof(buf), trap_request)); 1159 nm->log_identity(xtty); 1160 } 1161 symbolHandle class_name; 1162 bool unresolved = false; 1163 if (unloaded_class_index >= 0) { 1164 constantPoolHandle constants (THREAD, trap_method->constants()); 1165 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1166 class_name = symbolHandle(THREAD, 1167 constants->klass_name_at(unloaded_class_index)); 1168 unresolved = true; 1169 if (xtty != NULL) 1170 xtty->print(" unresolved='1'"); 1171 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1172 class_name = symbolHandle(THREAD, 1173 constants->symbol_at(unloaded_class_index)); 1174 } 1175 if (xtty != NULL) 1176 xtty->name(class_name); 1177 } 1178 if (xtty != NULL && trap_mdo.not_null()) { 1179 // Dump the relevant MDO state. 1180 // This is the deopt count for the current reason, any previous 1181 // reasons or recompiles seen at this point. 1182 int dcnt = trap_mdo->trap_count(reason); 1183 if (dcnt != 0) 1184 xtty->print(" count='%d'", dcnt); 1185 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1186 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1187 if (dos != 0) { 1188 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1189 if (trap_state_is_recompiled(dos)) { 1190 int recnt2 = trap_mdo->overflow_recompile_count(); 1191 if (recnt2 != 0) 1192 xtty->print(" recompiles2='%d'", recnt2); 1193 } 1194 } 1195 } 1196 if (xtty != NULL) { 1197 xtty->stamp(); 1198 xtty->end_head(); 1199 } 1200 if (TraceDeoptimization) { // make noise on the tty 1201 tty->print("Uncommon trap occurred in"); 1202 nm->method()->print_short_name(tty); 1203 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1204 fr.pc(), 1205 (int) os::current_thread_id(), 1206 trap_reason_name(reason), 1207 trap_action_name(action), 1208 unloaded_class_index); 1209 if (class_name.not_null()) { 1210 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1211 class_name->print_symbol_on(tty); 1212 } 1213 tty->cr(); 1214 } 1215 if (xtty != NULL) { 1216 // Log the precise location of the trap. 1217 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1218 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1219 xtty->method(sd->method()); 1220 xtty->end_elem(); 1221 if (sd->is_top()) break; 1222 } 1223 xtty->tail("uncommon_trap"); 1224 } 1225 } 1226 // (End diagnostic printout.) 1227 1228 // Load class if necessary 1229 if (unloaded_class_index >= 0) { 1230 constantPoolHandle constants(THREAD, trap_method->constants()); 1231 load_class_by_index(constants, unloaded_class_index); 1232 } 1233 1234 // Flush the nmethod if necessary and desirable. 1235 // 1236 // We need to avoid situations where we are re-flushing the nmethod 1237 // because of a hot deoptimization site. Repeated flushes at the same 1238 // point need to be detected by the compiler and avoided. If the compiler 1239 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1240 // module must take measures to avoid an infinite cycle of recompilation 1241 // and deoptimization. There are several such measures: 1242 // 1243 // 1. If a recompilation is ordered a second time at some site X 1244 // and for the same reason R, the action is adjusted to 'reinterpret', 1245 // to give the interpreter time to exercise the method more thoroughly. 1246 // If this happens, the method's overflow_recompile_count is incremented. 1247 // 1248 // 2. If the compiler fails to reduce the deoptimization rate, then 1249 // the method's overflow_recompile_count will begin to exceed the set 1250 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1251 // is adjusted to 'make_not_compilable', and the method is abandoned 1252 // to the interpreter. This is a performance hit for hot methods, 1253 // but is better than a disastrous infinite cycle of recompilations. 1254 // (Actually, only the method containing the site X is abandoned.) 1255 // 1256 // 3. In parallel with the previous measures, if the total number of 1257 // recompilations of a method exceeds the much larger set limit 1258 // PerMethodRecompilationCutoff, the method is abandoned. 1259 // This should only happen if the method is very large and has 1260 // many "lukewarm" deoptimizations. The code which enforces this 1261 // limit is elsewhere (class nmethod, class methodOopDesc). 1262 // 1263 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1264 // to recompile at each bytecode independently of the per-BCI cutoff. 1265 // 1266 // The decision to update code is up to the compiler, and is encoded 1267 // in the Action_xxx code. If the compiler requests Action_none 1268 // no trap state is changed, no compiled code is changed, and the 1269 // computation suffers along in the interpreter. 1270 // 1271 // The other action codes specify various tactics for decompilation 1272 // and recompilation. Action_maybe_recompile is the loosest, and 1273 // allows the compiled code to stay around until enough traps are seen, 1274 // and until the compiler gets around to recompiling the trapping method. 1275 // 1276 // The other actions cause immediate removal of the present code. 1277 1278 bool update_trap_state = true; 1279 bool make_not_entrant = false; 1280 bool make_not_compilable = false; 1281 bool reset_counters = false; 1282 switch (action) { 1283 case Action_none: 1284 // Keep the old code. 1285 update_trap_state = false; 1286 break; 1287 case Action_maybe_recompile: 1288 // Do not need to invalidate the present code, but we can 1289 // initiate another 1290 // Start compiler without (necessarily) invalidating the nmethod. 1291 // The system will tolerate the old code, but new code should be 1292 // generated when possible. 1293 break; 1294 case Action_reinterpret: 1295 // Go back into the interpreter for a while, and then consider 1296 // recompiling form scratch. 1297 make_not_entrant = true; 1298 // Reset invocation counter for outer most method. 1299 // This will allow the interpreter to exercise the bytecodes 1300 // for a while before recompiling. 1301 // By contrast, Action_make_not_entrant is immediate. 1302 // 1303 // Note that the compiler will track null_check, null_assert, 1304 // range_check, and class_check events and log them as if they 1305 // had been traps taken from compiled code. This will update 1306 // the MDO trap history so that the next compilation will 1307 // properly detect hot trap sites. 1308 reset_counters = true; 1309 break; 1310 case Action_make_not_entrant: 1311 // Request immediate recompilation, and get rid of the old code. 1312 // Make them not entrant, so next time they are called they get 1313 // recompiled. Unloaded classes are loaded now so recompile before next 1314 // time they are called. Same for uninitialized. The interpreter will 1315 // link the missing class, if any. 1316 make_not_entrant = true; 1317 break; 1318 case Action_make_not_compilable: 1319 // Give up on compiling this method at all. 1320 make_not_entrant = true; 1321 make_not_compilable = true; 1322 break; 1323 default: 1324 ShouldNotReachHere(); 1325 } 1326 1327 // Setting +ProfileTraps fixes the following, on all platforms: 1328 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1329 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1330 // recompile relies on a methodDataOop to record heroic opt failures. 1331 1332 // Whether the interpreter is producing MDO data or not, we also need 1333 // to use the MDO to detect hot deoptimization points and control 1334 // aggressive optimization. 1335 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1336 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1337 uint this_trap_count = 0; 1338 bool maybe_prior_trap = false; 1339 bool maybe_prior_recompile = false; 1340 ProfileData* pdata 1341 = query_update_method_data(trap_mdo, trap_bci, reason, 1342 //outputs: 1343 this_trap_count, 1344 maybe_prior_trap, 1345 maybe_prior_recompile); 1346 // Because the interpreter also counts null, div0, range, and class 1347 // checks, these traps from compiled code are double-counted. 1348 // This is harmless; it just means that the PerXTrapLimit values 1349 // are in effect a little smaller than they look. 1350 1351 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1352 if (per_bc_reason != Reason_none) { 1353 // Now take action based on the partially known per-BCI history. 1354 if (maybe_prior_trap 1355 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1356 // If there are too many traps at this BCI, force a recompile. 1357 // This will allow the compiler to see the limit overflow, and 1358 // take corrective action, if possible. The compiler generally 1359 // does not use the exact PerBytecodeTrapLimit value, but instead 1360 // changes its tactics if it sees any traps at all. This provides 1361 // a little hysteresis, delaying a recompile until a trap happens 1362 // several times. 1363 // 1364 // Actually, since there is only one bit of counter per BCI, 1365 // the possible per-BCI counts are {0,1,(per-method count)}. 1366 // This produces accurate results if in fact there is only 1367 // one hot trap site, but begins to get fuzzy if there are 1368 // many sites. For example, if there are ten sites each 1369 // trapping two or more times, they each get the blame for 1370 // all of their traps. 1371 make_not_entrant = true; 1372 } 1373 1374 // Detect repeated recompilation at the same BCI, and enforce a limit. 1375 if (make_not_entrant && maybe_prior_recompile) { 1376 // More than one recompile at this point. 1377 trap_mdo->inc_overflow_recompile_count(); 1378 if (maybe_prior_trap 1379 && ((uint)trap_mdo->overflow_recompile_count() 1380 > (uint)PerBytecodeRecompilationCutoff)) { 1381 // Give up on the method containing the bad BCI. 1382 if (trap_method() == nm->method()) { 1383 make_not_compilable = true; 1384 } else { 1385 trap_method->set_not_compilable(); 1386 // But give grace to the enclosing nm->method(). 1387 } 1388 } 1389 } 1390 } else { 1391 // For reasons which are not recorded per-bytecode, we simply 1392 // force recompiles unconditionally. 1393 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1394 make_not_entrant = true; 1395 } 1396 1397 // Go back to the compiler if there are too many traps in this method. 1398 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1399 // If there are too many traps in this method, force a recompile. 1400 // This will allow the compiler to see the limit overflow, and 1401 // take corrective action, if possible. 1402 // (This condition is an unlikely backstop only, because the 1403 // PerBytecodeTrapLimit is more likely to take effect first, 1404 // if it is applicable.) 1405 make_not_entrant = true; 1406 } 1407 1408 // Here's more hysteresis: If there has been a recompile at 1409 // this trap point already, run the method in the interpreter 1410 // for a while to exercise it more thoroughly. 1411 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1412 reset_counters = true; 1413 } 1414 1415 if (make_not_entrant && pdata != NULL) { 1416 // Record the recompilation event, if any. 1417 int tstate0 = pdata->trap_state(); 1418 int tstate1 = trap_state_set_recompiled(tstate0, true); 1419 if (tstate1 != tstate0) 1420 pdata->set_trap_state(tstate1); 1421 } 1422 } 1423 1424 // Take requested actions on the method: 1425 1426 // Reset invocation counters 1427 if (reset_counters) { 1428 if (nm->is_osr_method()) 1429 reset_invocation_counter(trap_scope, CompileThreshold); 1430 else 1431 reset_invocation_counter(trap_scope); 1432 } 1433 1434 // Recompile 1435 if (make_not_entrant) { 1436 nm->make_not_entrant(); 1437 } 1438 1439 // Give up compiling 1440 if (make_not_compilable) { 1441 assert(make_not_entrant, "consistent"); 1442 nm->method()->set_not_compilable(); 1443 } 1444 1445 } // Free marked resources 1446 1447 } 1448 JRT_END 1449 1450 methodDataOop 1451 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1452 bool create_if_missing) { 1453 Thread* THREAD = thread; 1454 methodDataOop mdo = m()->method_data(); 1455 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1456 // Build an MDO. Ignore errors like OutOfMemory; 1457 // that simply means we won't have an MDO to update. 1458 methodOopDesc::build_interpreter_method_data(m, THREAD); 1459 if (HAS_PENDING_EXCEPTION) { 1460 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1461 CLEAR_PENDING_EXCEPTION; 1462 } 1463 mdo = m()->method_data(); 1464 } 1465 return mdo; 1466 } 1467 1468 ProfileData* 1469 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1470 int trap_bci, 1471 Deoptimization::DeoptReason reason, 1472 //outputs: 1473 uint& ret_this_trap_count, 1474 bool& ret_maybe_prior_trap, 1475 bool& ret_maybe_prior_recompile) { 1476 uint prior_trap_count = trap_mdo->trap_count(reason); 1477 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1478 1479 // If the runtime cannot find a place to store trap history, 1480 // it is estimated based on the general condition of the method. 1481 // If the method has ever been recompiled, or has ever incurred 1482 // a trap with the present reason , then this BCI is assumed 1483 // (pessimistically) to be the culprit. 1484 bool maybe_prior_trap = (prior_trap_count != 0); 1485 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1486 ProfileData* pdata = NULL; 1487 1488 1489 // For reasons which are recorded per bytecode, we check per-BCI data. 1490 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1491 if (per_bc_reason != Reason_none) { 1492 // Find the profile data for this BCI. If there isn't one, 1493 // try to allocate one from the MDO's set of spares. 1494 // This will let us detect a repeated trap at this point. 1495 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1496 1497 if (pdata != NULL) { 1498 // Query the trap state of this profile datum. 1499 int tstate0 = pdata->trap_state(); 1500 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1501 maybe_prior_trap = false; 1502 if (!trap_state_is_recompiled(tstate0)) 1503 maybe_prior_recompile = false; 1504 1505 // Update the trap state of this profile datum. 1506 int tstate1 = tstate0; 1507 // Record the reason. 1508 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1509 // Store the updated state on the MDO, for next time. 1510 if (tstate1 != tstate0) 1511 pdata->set_trap_state(tstate1); 1512 } else { 1513 if (LogCompilation && xtty != NULL) 1514 // Missing MDP? Leave a small complaint in the log. 1515 xtty->elem("missing_mdp bci='%d'", trap_bci); 1516 } 1517 } 1518 1519 // Return results: 1520 ret_this_trap_count = this_trap_count; 1521 ret_maybe_prior_trap = maybe_prior_trap; 1522 ret_maybe_prior_recompile = maybe_prior_recompile; 1523 return pdata; 1524 } 1525 1526 void 1527 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1528 ResourceMark rm; 1529 // Ignored outputs: 1530 uint ignore_this_trap_count; 1531 bool ignore_maybe_prior_trap; 1532 bool ignore_maybe_prior_recompile; 1533 query_update_method_data(trap_mdo, trap_bci, 1534 (DeoptReason)reason, 1535 ignore_this_trap_count, 1536 ignore_maybe_prior_trap, 1537 ignore_maybe_prior_recompile); 1538 } 1539 1540 void Deoptimization::reset_invocation_counter(ScopeDesc* trap_scope, jint top_count) { 1541 ScopeDesc* sd = trap_scope; 1542 for (; !sd->is_top(); sd = sd->sender()) { 1543 // Reset ICs of inlined methods, since they can trigger compilations also. 1544 sd->method()->invocation_counter()->reset(); 1545 } 1546 InvocationCounter* c = sd->method()->invocation_counter(); 1547 if (top_count != _no_count) { 1548 // It was an OSR method, so bump the count higher. 1549 c->set(c->state(), top_count); 1550 } else { 1551 c->reset(); 1552 } 1553 sd->method()->backedge_counter()->reset(); 1554 } 1555 1556 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1557 1558 // Still in Java no safepoints 1559 { 1560 // This enters VM and may safepoint 1561 uncommon_trap_inner(thread, trap_request); 1562 } 1563 return fetch_unroll_info_helper(thread); 1564 } 1565 1566 // Local derived constants. 1567 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1568 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1569 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1570 1571 //---------------------------trap_state_reason--------------------------------- 1572 Deoptimization::DeoptReason 1573 Deoptimization::trap_state_reason(int trap_state) { 1574 // This assert provides the link between the width of DataLayout::trap_bits 1575 // and the encoding of "recorded" reasons. It ensures there are enough 1576 // bits to store all needed reasons in the per-BCI MDO profile. 1577 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1578 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1579 trap_state -= recompile_bit; 1580 if (trap_state == DS_REASON_MASK) { 1581 return Reason_many; 1582 } else { 1583 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1584 return (DeoptReason)trap_state; 1585 } 1586 } 1587 //-------------------------trap_state_has_reason------------------------------- 1588 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1589 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1590 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1591 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1592 trap_state -= recompile_bit; 1593 if (trap_state == DS_REASON_MASK) { 1594 return -1; // true, unspecifically (bottom of state lattice) 1595 } else if (trap_state == reason) { 1596 return 1; // true, definitely 1597 } else if (trap_state == 0) { 1598 return 0; // false, definitely (top of state lattice) 1599 } else { 1600 return 0; // false, definitely 1601 } 1602 } 1603 //-------------------------trap_state_add_reason------------------------------- 1604 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1605 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1606 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1607 trap_state -= recompile_bit; 1608 if (trap_state == DS_REASON_MASK) { 1609 return trap_state + recompile_bit; // already at state lattice bottom 1610 } else if (trap_state == reason) { 1611 return trap_state + recompile_bit; // the condition is already true 1612 } else if (trap_state == 0) { 1613 return reason + recompile_bit; // no condition has yet been true 1614 } else { 1615 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1616 } 1617 } 1618 //-----------------------trap_state_is_recompiled------------------------------ 1619 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1620 return (trap_state & DS_RECOMPILE_BIT) != 0; 1621 } 1622 //-----------------------trap_state_set_recompiled----------------------------- 1623 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1624 if (z) return trap_state | DS_RECOMPILE_BIT; 1625 else return trap_state & ~DS_RECOMPILE_BIT; 1626 } 1627 //---------------------------format_trap_state--------------------------------- 1628 // This is used for debugging and diagnostics, including hotspot.log output. 1629 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1630 int trap_state) { 1631 DeoptReason reason = trap_state_reason(trap_state); 1632 bool recomp_flag = trap_state_is_recompiled(trap_state); 1633 // Re-encode the state from its decoded components. 1634 int decoded_state = 0; 1635 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1636 decoded_state = trap_state_add_reason(decoded_state, reason); 1637 if (recomp_flag) 1638 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1639 // If the state re-encodes properly, format it symbolically. 1640 // Because this routine is used for debugging and diagnostics, 1641 // be robust even if the state is a strange value. 1642 size_t len; 1643 if (decoded_state != trap_state) { 1644 // Random buggy state that doesn't decode?? 1645 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1646 } else { 1647 len = jio_snprintf(buf, buflen, "%s%s", 1648 trap_reason_name(reason), 1649 recomp_flag ? " recompiled" : ""); 1650 } 1651 if (len >= buflen) 1652 buf[buflen-1] = '\0'; 1653 return buf; 1654 } 1655 1656 1657 //--------------------------------statics-------------------------------------- 1658 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1659 = Deoptimization::Action_reinterpret; 1660 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1661 // Note: Keep this in sync. with enum DeoptReason. 1662 "none", 1663 "null_check", 1664 "null_assert", 1665 "range_check", 1666 "class_check", 1667 "array_check", 1668 "intrinsic", 1669 "unloaded", 1670 "uninitialized", 1671 "unreached", 1672 "unhandled", 1673 "constraint", 1674 "div0_check", 1675 "age", 1676 "predicate" 1677 }; 1678 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1679 // Note: Keep this in sync. with enum DeoptAction. 1680 "none", 1681 "maybe_recompile", 1682 "reinterpret", 1683 "make_not_entrant", 1684 "make_not_compilable" 1685 }; 1686 1687 const char* Deoptimization::trap_reason_name(int reason) { 1688 if (reason == Reason_many) return "many"; 1689 if ((uint)reason < Reason_LIMIT) 1690 return _trap_reason_name[reason]; 1691 static char buf[20]; 1692 sprintf(buf, "reason%d", reason); 1693 return buf; 1694 } 1695 const char* Deoptimization::trap_action_name(int action) { 1696 if ((uint)action < Action_LIMIT) 1697 return _trap_action_name[action]; 1698 static char buf[20]; 1699 sprintf(buf, "action%d", action); 1700 return buf; 1701 } 1702 1703 // This is used for debugging and diagnostics, including hotspot.log output. 1704 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1705 int trap_request) { 1706 jint unloaded_class_index = trap_request_index(trap_request); 1707 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1708 const char* action = trap_action_name(trap_request_action(trap_request)); 1709 size_t len; 1710 if (unloaded_class_index < 0) { 1711 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1712 reason, action); 1713 } else { 1714 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1715 reason, action, unloaded_class_index); 1716 } 1717 if (len >= buflen) 1718 buf[buflen-1] = '\0'; 1719 return buf; 1720 } 1721 1722 juint Deoptimization::_deoptimization_hist 1723 [Deoptimization::Reason_LIMIT] 1724 [1 + Deoptimization::Action_LIMIT] 1725 [Deoptimization::BC_CASE_LIMIT] 1726 = {0}; 1727 1728 enum { 1729 LSB_BITS = 8, 1730 LSB_MASK = right_n_bits(LSB_BITS) 1731 }; 1732 1733 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1734 Bytecodes::Code bc) { 1735 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1736 assert(action >= 0 && action < Action_LIMIT, "oob"); 1737 _deoptimization_hist[Reason_none][0][0] += 1; // total 1738 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1739 juint* cases = _deoptimization_hist[reason][1+action]; 1740 juint* bc_counter_addr = NULL; 1741 juint bc_counter = 0; 1742 // Look for an unused counter, or an exact match to this BC. 1743 if (bc != Bytecodes::_illegal) { 1744 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1745 juint* counter_addr = &cases[bc_case]; 1746 juint counter = *counter_addr; 1747 if ((counter == 0 && bc_counter_addr == NULL) 1748 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1749 // this counter is either free or is already devoted to this BC 1750 bc_counter_addr = counter_addr; 1751 bc_counter = counter | bc; 1752 } 1753 } 1754 } 1755 if (bc_counter_addr == NULL) { 1756 // Overflow, or no given bytecode. 1757 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1758 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1759 } 1760 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1761 } 1762 1763 jint Deoptimization::total_deoptimization_count() { 1764 return _deoptimization_hist[Reason_none][0][0]; 1765 } 1766 1767 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1768 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1769 return _deoptimization_hist[reason][0][0]; 1770 } 1771 1772 void Deoptimization::print_statistics() { 1773 juint total = total_deoptimization_count(); 1774 juint account = total; 1775 if (total != 0) { 1776 ttyLocker ttyl; 1777 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1778 tty->print_cr("Deoptimization traps recorded:"); 1779 #define PRINT_STAT_LINE(name, r) \ 1780 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1781 PRINT_STAT_LINE("total", total); 1782 // For each non-zero entry in the histogram, print the reason, 1783 // the action, and (if specifically known) the type of bytecode. 1784 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1785 for (int action = 0; action < Action_LIMIT; action++) { 1786 juint* cases = _deoptimization_hist[reason][1+action]; 1787 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1788 juint counter = cases[bc_case]; 1789 if (counter != 0) { 1790 char name[1*K]; 1791 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1792 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1793 bc = Bytecodes::_illegal; 1794 sprintf(name, "%s/%s/%s", 1795 trap_reason_name(reason), 1796 trap_action_name(action), 1797 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1798 juint r = counter >> LSB_BITS; 1799 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1800 account -= r; 1801 } 1802 } 1803 } 1804 } 1805 if (account != 0) { 1806 PRINT_STAT_LINE("unaccounted", account); 1807 } 1808 #undef PRINT_STAT_LINE 1809 if (xtty != NULL) xtty->tail("statistics"); 1810 } 1811 } 1812 #else // COMPILER2 1813 1814 1815 // Stubs for C1 only system. 1816 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1817 return false; 1818 } 1819 1820 const char* Deoptimization::trap_reason_name(int reason) { 1821 return "unknown"; 1822 } 1823 1824 void Deoptimization::print_statistics() { 1825 // no output 1826 } 1827 1828 void 1829 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1830 // no udpate 1831 } 1832 1833 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1834 return 0; 1835 } 1836 1837 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1838 Bytecodes::Code bc) { 1839 // no update 1840 } 1841 1842 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1843 int trap_state) { 1844 jio_snprintf(buf, buflen, "#%d", trap_state); 1845 return buf; 1846 } 1847 1848 #endif // COMPILER2