1 /* 2 * Copyright 1997-2010 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 // If the deopt call site is a MethodHandle invoke call site we have 239 // to adjust the unpack_sp. 240 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 241 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 242 unpack_sp = deoptee.unextended_sp(); 243 244 #ifdef ASSERT 245 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 246 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 247 #endif 248 // This is a guarantee instead of an assert because if vframe doesn't match 249 // we will unpack the wrong deoptimized frame and wind up in strange places 250 // where it will be very difficult to figure out what went wrong. Better 251 // to die an early death here than some very obscure death later when the 252 // trail is cold. 253 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 254 // in that it will fail to detect a problem when there is one. This needs 255 // more work in tiger timeframe. 256 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 257 258 int number_of_frames = array->frames(); 259 260 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 261 // virtual activation, which is the reverse of the elements in the vframes array. 262 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 263 // +1 because we always have an interpreter return address for the final slot. 264 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 265 int callee_parameters = 0; 266 int callee_locals = 0; 267 int popframe_extra_args = 0; 268 // Create an interpreter return address for the stub to use as its return 269 // address so the skeletal frames are perfectly walkable 270 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 271 272 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 273 // activation be put back on the expression stack of the caller for reexecution 274 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 275 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 276 } 277 278 // 279 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 280 // frame_sizes/frame_pcs[1] next oldest frame (int) 281 // frame_sizes/frame_pcs[n] youngest frame (int) 282 // 283 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 284 // owns the space for the return address to it's caller). Confusing ain't it. 285 // 286 // The vframe array can address vframes with indices running from 287 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 288 // When we create the skeletal frames we need the oldest frame to be in the zero slot 289 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 290 // so things look a little strange in this loop. 291 // 292 for (int index = 0; index < array->frames(); index++ ) { 293 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 294 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 295 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 296 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 297 callee_locals, 298 index == 0, 299 popframe_extra_args); 300 // This pc doesn't have to be perfect just good enough to identify the frame 301 // as interpreted so the skeleton frame will be walkable 302 // The correct pc will be set when the skeleton frame is completely filled out 303 // The final pc we store in the loop is wrong and will be overwritten below 304 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 305 306 callee_parameters = array->element(index)->method()->size_of_parameters(); 307 callee_locals = array->element(index)->method()->max_locals(); 308 popframe_extra_args = 0; 309 } 310 311 // Compute whether the root vframe returns a float or double value. 312 BasicType return_type; 313 { 314 HandleMark hm; 315 methodHandle method(thread, array->element(0)->method()); 316 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci()); 317 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL; 318 } 319 320 // Compute information for handling adapters and adjusting the frame size of the caller. 321 int caller_adjustment = 0; 322 323 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 324 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 325 // than simply use array->sender.pc(). This requires us to walk the current set of frames 326 // 327 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 328 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 329 330 // Compute the amount the oldest interpreter frame will have to adjust 331 // its caller's stack by. If the caller is a compiled frame then 332 // we pretend that the callee has no parameters so that the 333 // extension counts for the full amount of locals and not just 334 // locals-parms. This is because without a c2i adapter the parm 335 // area as created by the compiled frame will not be usable by 336 // the interpreter. (Depending on the calling convention there 337 // may not even be enough space). 338 339 // QQQ I'd rather see this pushed down into last_frame_adjust 340 // and have it take the sender (aka caller). 341 342 if (deopt_sender.is_compiled_frame()) { 343 caller_adjustment = last_frame_adjust(0, callee_locals); 344 } else if (callee_locals > callee_parameters) { 345 // The caller frame may need extending to accommodate 346 // non-parameter locals of the first unpacked interpreted frame. 347 // Compute that adjustment. 348 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 349 } 350 351 352 // If the sender is deoptimized the we must retrieve the address of the handler 353 // since the frame will "magically" show the original pc before the deopt 354 // and we'd undo the deopt. 355 356 frame_pcs[0] = deopt_sender.raw_pc(); 357 358 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 359 360 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 361 caller_adjustment * BytesPerWord, 362 number_of_frames, 363 frame_sizes, 364 frame_pcs, 365 return_type); 366 #if defined(IA32) || defined(AMD64) 367 // We need a way to pass fp to the unpacking code so the skeletal frames 368 // come out correct. This is only needed for x86 because of c2 using ebp 369 // as an allocatable register. So this update is useless (and harmless) 370 // on the other platforms. It would be nice to do this in a different 371 // way but even the old style deoptimization had a problem with deriving 372 // this value. NEEDS_CLEANUP 373 // Note: now that c1 is using c2's deopt blob we must do this on all 374 // x86 based platforms 375 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); 376 *fp_addr = array->sender().fp(); // was adapter_caller 377 #endif /* IA32 || AMD64 */ 378 379 if (array->frames() > 1) { 380 if (VerifyStack && TraceDeoptimization) { 381 tty->print_cr("Deoptimizing method containing inlining"); 382 } 383 } 384 385 array->set_unroll_block(info); 386 return info; 387 } 388 389 // Called to cleanup deoptimization data structures in normal case 390 // after unpacking to stack and when stack overflow error occurs 391 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 392 vframeArray *array) { 393 394 // Get array if coming from exception 395 if (array == NULL) { 396 array = thread->vframe_array_head(); 397 } 398 thread->set_vframe_array_head(NULL); 399 400 // Free the previous UnrollBlock 401 vframeArray* old_array = thread->vframe_array_last(); 402 thread->set_vframe_array_last(array); 403 404 if (old_array != NULL) { 405 UnrollBlock* old_info = old_array->unroll_block(); 406 old_array->set_unroll_block(NULL); 407 delete old_info; 408 delete old_array; 409 } 410 411 // Deallocate any resource creating in this routine and any ResourceObjs allocated 412 // inside the vframeArray (StackValueCollections) 413 414 delete thread->deopt_mark(); 415 thread->set_deopt_mark(NULL); 416 417 418 if (JvmtiExport::can_pop_frame()) { 419 #ifndef CC_INTERP 420 // Regardless of whether we entered this routine with the pending 421 // popframe condition bit set, we should always clear it now 422 thread->clear_popframe_condition(); 423 #else 424 // C++ interpeter will clear has_pending_popframe when it enters 425 // with method_resume. For deopt_resume2 we clear it now. 426 if (thread->popframe_forcing_deopt_reexecution()) 427 thread->clear_popframe_condition(); 428 #endif /* CC_INTERP */ 429 } 430 431 // unpack_frames() is called at the end of the deoptimization handler 432 // and (in C2) at the end of the uncommon trap handler. Note this fact 433 // so that an asynchronous stack walker can work again. This counter is 434 // incremented at the beginning of fetch_unroll_info() and (in C2) at 435 // the beginning of uncommon_trap(). 436 thread->dec_in_deopt_handler(); 437 } 438 439 440 // Return BasicType of value being returned 441 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 442 443 // We are already active int he special DeoptResourceMark any ResourceObj's we 444 // allocate will be freed at the end of the routine. 445 446 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 447 // but makes the entry a little slower. There is however a little dance we have to 448 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 449 ResetNoHandleMark rnhm; // No-op in release/product versions 450 HandleMark hm; 451 452 frame stub_frame = thread->last_frame(); 453 454 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 455 // must point to the vframeArray for the unpack frame. 456 vframeArray* array = thread->vframe_array_head(); 457 458 #ifndef PRODUCT 459 if (TraceDeoptimization) { 460 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 461 } 462 #endif 463 464 UnrollBlock* info = array->unroll_block(); 465 466 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 467 array->unpack_to_stack(stub_frame, exec_mode); 468 469 BasicType bt = info->return_type(); 470 471 // If we have an exception pending, claim that the return type is an oop 472 // so the deopt_blob does not overwrite the exception_oop. 473 474 if (exec_mode == Unpack_exception) 475 bt = T_OBJECT; 476 477 // Cleanup thread deopt data 478 cleanup_deopt_info(thread, array); 479 480 #ifndef PRODUCT 481 if (VerifyStack) { 482 ResourceMark res_mark; 483 484 // Verify that the just-unpacked frames match the interpreter's 485 // notions of expression stack and locals 486 vframeArray* cur_array = thread->vframe_array_last(); 487 RegisterMap rm(thread, false); 488 rm.set_include_argument_oops(false); 489 bool is_top_frame = true; 490 int callee_size_of_parameters = 0; 491 int callee_max_locals = 0; 492 for (int i = 0; i < cur_array->frames(); i++) { 493 vframeArrayElement* el = cur_array->element(i); 494 frame* iframe = el->iframe(); 495 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 496 497 // Get the oop map for this bci 498 InterpreterOopMap mask; 499 int cur_invoke_parameter_size = 0; 500 bool try_next_mask = false; 501 int next_mask_expression_stack_size = -1; 502 int top_frame_expression_stack_adjustment = 0; 503 methodHandle mh(thread, iframe->interpreter_frame_method()); 504 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 505 BytecodeStream str(mh); 506 str.set_start(iframe->interpreter_frame_bci()); 507 int max_bci = mh->code_size(); 508 // Get to the next bytecode if possible 509 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 510 // Check to see if we can grab the number of outgoing arguments 511 // at an uncommon trap for an invoke (where the compiler 512 // generates debug info before the invoke has executed) 513 Bytecodes::Code cur_code = str.next(); 514 if (cur_code == Bytecodes::_invokevirtual || 515 cur_code == Bytecodes::_invokespecial || 516 cur_code == Bytecodes::_invokestatic || 517 cur_code == Bytecodes::_invokeinterface) { 518 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci()); 519 symbolHandle signature(thread, invoke->signature()); 520 ArgumentSizeComputer asc(signature); 521 cur_invoke_parameter_size = asc.size(); 522 if (cur_code != Bytecodes::_invokestatic) { 523 // Add in receiver 524 ++cur_invoke_parameter_size; 525 } 526 } 527 if (str.bci() < max_bci) { 528 Bytecodes::Code bc = str.next(); 529 if (bc >= 0) { 530 // The interpreter oop map generator reports results before 531 // the current bytecode has executed except in the case of 532 // calls. It seems to be hard to tell whether the compiler 533 // has emitted debug information matching the "state before" 534 // a given bytecode or the state after, so we try both 535 switch (cur_code) { 536 case Bytecodes::_invokevirtual: 537 case Bytecodes::_invokespecial: 538 case Bytecodes::_invokestatic: 539 case Bytecodes::_invokeinterface: 540 case Bytecodes::_athrow: 541 break; 542 default: { 543 InterpreterOopMap next_mask; 544 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 545 next_mask_expression_stack_size = next_mask.expression_stack_size(); 546 // Need to subtract off the size of the result type of 547 // the bytecode because this is not described in the 548 // debug info but returned to the interpreter in the TOS 549 // caching register 550 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 551 if (bytecode_result_type != T_ILLEGAL) { 552 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 553 } 554 assert(top_frame_expression_stack_adjustment >= 0, ""); 555 try_next_mask = true; 556 break; 557 } 558 } 559 } 560 } 561 562 // Verify stack depth and oops in frame 563 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 564 if (!( 565 /* SPARC */ 566 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 567 /* x86 */ 568 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 569 (try_next_mask && 570 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 571 top_frame_expression_stack_adjustment))) || 572 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 573 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 574 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 575 )) { 576 ttyLocker ttyl; 577 578 // Print out some information that will help us debug the problem 579 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 580 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 581 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 582 iframe->interpreter_frame_expression_stack_size()); 583 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 584 tty->print_cr(" try_next_mask = %d", try_next_mask); 585 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 586 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 587 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 588 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 589 tty->print_cr(" exec_mode = %d", exec_mode); 590 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 591 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 592 tty->print_cr(" Interpreted frames:"); 593 for (int k = 0; k < cur_array->frames(); k++) { 594 vframeArrayElement* el = cur_array->element(k); 595 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 596 } 597 cur_array->print_on_2(tty); 598 guarantee(false, "wrong number of expression stack elements during deopt"); 599 } 600 VerifyOopClosure verify; 601 iframe->oops_interpreted_do(&verify, &rm, false); 602 callee_size_of_parameters = mh->size_of_parameters(); 603 callee_max_locals = mh->max_locals(); 604 is_top_frame = false; 605 } 606 } 607 #endif /* !PRODUCT */ 608 609 610 return bt; 611 JRT_END 612 613 614 int Deoptimization::deoptimize_dependents() { 615 Threads::deoptimized_wrt_marked_nmethods(); 616 return 0; 617 } 618 619 620 #ifdef COMPILER2 621 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 622 Handle pending_exception(thread->pending_exception()); 623 const char* exception_file = thread->exception_file(); 624 int exception_line = thread->exception_line(); 625 thread->clear_pending_exception(); 626 627 for (int i = 0; i < objects->length(); i++) { 628 assert(objects->at(i)->is_object(), "invalid debug information"); 629 ObjectValue* sv = (ObjectValue*) objects->at(i); 630 631 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 632 oop obj = NULL; 633 634 if (k->oop_is_instance()) { 635 instanceKlass* ik = instanceKlass::cast(k()); 636 obj = ik->allocate_instance(CHECK_(false)); 637 } else if (k->oop_is_typeArray()) { 638 typeArrayKlass* ak = typeArrayKlass::cast(k()); 639 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 640 int len = sv->field_size() / type2size[ak->element_type()]; 641 obj = ak->allocate(len, CHECK_(false)); 642 } else if (k->oop_is_objArray()) { 643 objArrayKlass* ak = objArrayKlass::cast(k()); 644 obj = ak->allocate(sv->field_size(), CHECK_(false)); 645 } 646 647 assert(obj != NULL, "allocation failed"); 648 assert(sv->value().is_null(), "redundant reallocation"); 649 sv->set_value(obj); 650 } 651 652 if (pending_exception.not_null()) { 653 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 654 } 655 656 return true; 657 } 658 659 // This assumes that the fields are stored in ObjectValue in the same order 660 // they are yielded by do_nonstatic_fields. 661 class FieldReassigner: public FieldClosure { 662 frame* _fr; 663 RegisterMap* _reg_map; 664 ObjectValue* _sv; 665 instanceKlass* _ik; 666 oop _obj; 667 668 int _i; 669 public: 670 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 671 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 672 673 int i() const { return _i; } 674 675 676 void do_field(fieldDescriptor* fd) { 677 intptr_t val; 678 StackValue* value = 679 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 680 int offset = fd->offset(); 681 switch (fd->field_type()) { 682 case T_OBJECT: case T_ARRAY: 683 assert(value->type() == T_OBJECT, "Agreement."); 684 _obj->obj_field_put(offset, value->get_obj()()); 685 break; 686 687 case T_LONG: case T_DOUBLE: { 688 assert(value->type() == T_INT, "Agreement."); 689 StackValue* low = 690 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 691 #ifdef _LP64 692 jlong res = (jlong)low->get_int(); 693 #else 694 #ifdef SPARC 695 // For SPARC we have to swap high and low words. 696 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 697 #else 698 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 699 #endif //SPARC 700 #endif 701 _obj->long_field_put(offset, res); 702 break; 703 } 704 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 705 case T_INT: case T_FLOAT: // 4 bytes. 706 assert(value->type() == T_INT, "Agreement."); 707 val = value->get_int(); 708 _obj->int_field_put(offset, (jint)*((jint*)&val)); 709 break; 710 711 case T_SHORT: case T_CHAR: // 2 bytes 712 assert(value->type() == T_INT, "Agreement."); 713 val = value->get_int(); 714 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 715 break; 716 717 case T_BOOLEAN: case T_BYTE: // 1 byte 718 assert(value->type() == T_INT, "Agreement."); 719 val = value->get_int(); 720 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 721 break; 722 723 default: 724 ShouldNotReachHere(); 725 } 726 _i++; 727 } 728 }; 729 730 // restore elements of an eliminated type array 731 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 732 int index = 0; 733 intptr_t val; 734 735 for (int i = 0; i < sv->field_size(); i++) { 736 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 737 switch(type) { 738 case T_LONG: case T_DOUBLE: { 739 assert(value->type() == T_INT, "Agreement."); 740 StackValue* low = 741 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 742 #ifdef _LP64 743 jlong res = (jlong)low->get_int(); 744 #else 745 #ifdef SPARC 746 // For SPARC we have to swap high and low words. 747 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 748 #else 749 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 750 #endif //SPARC 751 #endif 752 obj->long_at_put(index, res); 753 break; 754 } 755 756 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 757 case T_INT: case T_FLOAT: // 4 bytes. 758 assert(value->type() == T_INT, "Agreement."); 759 val = value->get_int(); 760 obj->int_at_put(index, (jint)*((jint*)&val)); 761 break; 762 763 case T_SHORT: case T_CHAR: // 2 bytes 764 assert(value->type() == T_INT, "Agreement."); 765 val = value->get_int(); 766 obj->short_at_put(index, (jshort)*((jint*)&val)); 767 break; 768 769 case T_BOOLEAN: case T_BYTE: // 1 byte 770 assert(value->type() == T_INT, "Agreement."); 771 val = value->get_int(); 772 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 773 break; 774 775 default: 776 ShouldNotReachHere(); 777 } 778 index++; 779 } 780 } 781 782 783 // restore fields of an eliminated object array 784 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 785 for (int i = 0; i < sv->field_size(); i++) { 786 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 787 assert(value->type() == T_OBJECT, "object element expected"); 788 obj->obj_at_put(i, value->get_obj()()); 789 } 790 } 791 792 793 // restore fields of all eliminated objects and arrays 794 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 795 for (int i = 0; i < objects->length(); i++) { 796 ObjectValue* sv = (ObjectValue*) objects->at(i); 797 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 798 Handle obj = sv->value(); 799 assert(obj.not_null(), "reallocation was missed"); 800 801 if (k->oop_is_instance()) { 802 instanceKlass* ik = instanceKlass::cast(k()); 803 FieldReassigner reassign(fr, reg_map, sv, obj()); 804 ik->do_nonstatic_fields(&reassign); 805 } else if (k->oop_is_typeArray()) { 806 typeArrayKlass* ak = typeArrayKlass::cast(k()); 807 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 808 } else if (k->oop_is_objArray()) { 809 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 810 } 811 } 812 } 813 814 815 // relock objects for which synchronization was eliminated 816 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 817 for (int i = 0; i < monitors->length(); i++) { 818 MonitorInfo* mon_info = monitors->at(i); 819 if (mon_info->eliminated()) { 820 assert(mon_info->owner() != NULL, "reallocation was missed"); 821 Handle obj = Handle(mon_info->owner()); 822 markOop mark = obj->mark(); 823 if (UseBiasedLocking && mark->has_bias_pattern()) { 824 // New allocated objects may have the mark set to anonymously biased. 825 // Also the deoptimized method may called methods with synchronization 826 // where the thread-local object is bias locked to the current thread. 827 assert(mark->is_biased_anonymously() || 828 mark->biased_locker() == thread, "should be locked to current thread"); 829 // Reset mark word to unbiased prototype. 830 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 831 obj->set_mark(unbiased_prototype); 832 } 833 BasicLock* lock = mon_info->lock(); 834 ObjectSynchronizer::slow_enter(obj, lock, thread); 835 } 836 assert(mon_info->owner()->is_locked(), "object must be locked now"); 837 } 838 } 839 840 841 #ifndef PRODUCT 842 // print information about reallocated objects 843 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 844 fieldDescriptor fd; 845 846 for (int i = 0; i < objects->length(); i++) { 847 ObjectValue* sv = (ObjectValue*) objects->at(i); 848 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 849 Handle obj = sv->value(); 850 851 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 852 k->as_klassOop()->print_value(); 853 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 854 tty->cr(); 855 856 if (Verbose) { 857 k->oop_print_on(obj(), tty); 858 } 859 } 860 } 861 #endif 862 #endif // COMPILER2 863 864 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 865 866 #ifndef PRODUCT 867 if (TraceDeoptimization) { 868 ttyLocker ttyl; 869 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 870 fr.print_on(tty); 871 tty->print_cr(" Virtual frames (innermost first):"); 872 for (int index = 0; index < chunk->length(); index++) { 873 compiledVFrame* vf = chunk->at(index); 874 tty->print(" %2d - ", index); 875 vf->print_value(); 876 int bci = chunk->at(index)->raw_bci(); 877 const char* code_name; 878 if (bci == SynchronizationEntryBCI) { 879 code_name = "sync entry"; 880 } else { 881 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci); 882 code_name = Bytecodes::name(code); 883 } 884 tty->print(" - %s", code_name); 885 tty->print_cr(" @ bci %d ", bci); 886 if (Verbose) { 887 vf->print(); 888 tty->cr(); 889 } 890 } 891 } 892 #endif 893 894 // Register map for next frame (used for stack crawl). We capture 895 // the state of the deopt'ing frame's caller. Thus if we need to 896 // stuff a C2I adapter we can properly fill in the callee-save 897 // register locations. 898 frame caller = fr.sender(reg_map); 899 int frame_size = caller.sp() - fr.sp(); 900 901 frame sender = caller; 902 903 // Since the Java thread being deoptimized will eventually adjust it's own stack, 904 // the vframeArray containing the unpacking information is allocated in the C heap. 905 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 906 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 907 908 // Compare the vframeArray to the collected vframes 909 assert(array->structural_compare(thread, chunk), "just checking"); 910 Events::log("# vframes = %d", (intptr_t)chunk->length()); 911 912 #ifndef PRODUCT 913 if (TraceDeoptimization) { 914 ttyLocker ttyl; 915 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 916 if (Verbose) { 917 int count = 0; 918 // this used to leak deoptimizedVFrame like it was going out of style!!! 919 for (int index = 0; index < array->frames(); index++ ) { 920 vframeArrayElement* e = array->element(index); 921 e->print(tty); 922 923 /* 924 No printing yet. 925 array->vframe_at(index)->print_activation(count++); 926 // better as... 927 array->print_activation_for(index, count++); 928 */ 929 } 930 } 931 } 932 #endif // PRODUCT 933 934 return array; 935 } 936 937 938 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 939 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 940 for (int i = 0; i < monitors->length(); i++) { 941 MonitorInfo* mon_info = monitors->at(i); 942 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 943 objects_to_revoke->append(Handle(mon_info->owner())); 944 } 945 } 946 } 947 948 949 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 950 if (!UseBiasedLocking) { 951 return; 952 } 953 954 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 955 956 // Unfortunately we don't have a RegisterMap available in most of 957 // the places we want to call this routine so we need to walk the 958 // stack again to update the register map. 959 if (map == NULL || !map->update_map()) { 960 StackFrameStream sfs(thread, true); 961 bool found = false; 962 while (!found && !sfs.is_done()) { 963 frame* cur = sfs.current(); 964 sfs.next(); 965 found = cur->id() == fr.id(); 966 } 967 assert(found, "frame to be deoptimized not found on target thread's stack"); 968 map = sfs.register_map(); 969 } 970 971 vframe* vf = vframe::new_vframe(&fr, map, thread); 972 compiledVFrame* cvf = compiledVFrame::cast(vf); 973 // Revoke monitors' biases in all scopes 974 while (!cvf->is_top()) { 975 collect_monitors(cvf, objects_to_revoke); 976 cvf = compiledVFrame::cast(cvf->sender()); 977 } 978 collect_monitors(cvf, objects_to_revoke); 979 980 if (SafepointSynchronize::is_at_safepoint()) { 981 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 982 } else { 983 BiasedLocking::revoke(objects_to_revoke); 984 } 985 } 986 987 988 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 989 if (!UseBiasedLocking) { 990 return; 991 } 992 993 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 994 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 995 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 996 if (jt->has_last_Java_frame()) { 997 StackFrameStream sfs(jt, true); 998 while (!sfs.is_done()) { 999 frame* cur = sfs.current(); 1000 if (cb->contains(cur->pc())) { 1001 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1002 compiledVFrame* cvf = compiledVFrame::cast(vf); 1003 // Revoke monitors' biases in all scopes 1004 while (!cvf->is_top()) { 1005 collect_monitors(cvf, objects_to_revoke); 1006 cvf = compiledVFrame::cast(cvf->sender()); 1007 } 1008 collect_monitors(cvf, objects_to_revoke); 1009 } 1010 sfs.next(); 1011 } 1012 } 1013 } 1014 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1015 } 1016 1017 1018 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1019 assert(fr.can_be_deoptimized(), "checking frame type"); 1020 1021 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1022 1023 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1024 1025 // Patch the nmethod so that when execution returns to it we will 1026 // deopt the execution state and return to the interpreter. 1027 fr.deoptimize(thread); 1028 } 1029 1030 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1031 // Deoptimize only if the frame comes from compile code. 1032 // Do not deoptimize the frame which is already patched 1033 // during the execution of the loops below. 1034 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1035 return; 1036 } 1037 ResourceMark rm; 1038 DeoptimizationMarker dm; 1039 if (UseBiasedLocking) { 1040 revoke_biases_of_monitors(thread, fr, map); 1041 } 1042 deoptimize_single_frame(thread, fr); 1043 1044 } 1045 1046 1047 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1048 // Compute frame and register map based on thread and sp. 1049 RegisterMap reg_map(thread, UseBiasedLocking); 1050 frame fr = thread->last_frame(); 1051 while (fr.id() != id) { 1052 fr = fr.sender(®_map); 1053 } 1054 deoptimize(thread, fr, ®_map); 1055 } 1056 1057 1058 // JVMTI PopFrame support 1059 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1060 { 1061 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1062 } 1063 JRT_END 1064 1065 1066 #ifdef COMPILER2 1067 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1068 // in case of an unresolved klass entry, load the class. 1069 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1070 klassOop tk = constant_pool->klass_at(index, CHECK); 1071 return; 1072 } 1073 1074 if (!constant_pool->tag_at(index).is_symbol()) return; 1075 1076 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1077 symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); 1078 1079 // class name? 1080 if (symbol->byte_at(0) != '(') { 1081 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1082 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1083 return; 1084 } 1085 1086 // then it must be a signature! 1087 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1088 if (ss.is_object()) { 1089 symbolOop s = ss.as_symbol(CHECK); 1090 symbolHandle class_name (THREAD, s); 1091 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1092 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1093 } 1094 } 1095 } 1096 1097 1098 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1099 EXCEPTION_MARK; 1100 load_class_by_index(constant_pool, index, THREAD); 1101 if (HAS_PENDING_EXCEPTION) { 1102 // Exception happened during classloading. We ignore the exception here, since it 1103 // is going to be rethrown since the current activation is going to be deoptimzied and 1104 // the interpreter will re-execute the bytecode. 1105 CLEAR_PENDING_EXCEPTION; 1106 } 1107 } 1108 1109 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1110 HandleMark hm; 1111 1112 // uncommon_trap() is called at the beginning of the uncommon trap 1113 // handler. Note this fact before we start generating temporary frames 1114 // that can confuse an asynchronous stack walker. This counter is 1115 // decremented at the end of unpack_frames(). 1116 thread->inc_in_deopt_handler(); 1117 1118 // We need to update the map if we have biased locking. 1119 RegisterMap reg_map(thread, UseBiasedLocking); 1120 frame stub_frame = thread->last_frame(); 1121 frame fr = stub_frame.sender(®_map); 1122 // Make sure the calling nmethod is not getting deoptimized and removed 1123 // before we are done with it. 1124 nmethodLocker nl(fr.pc()); 1125 1126 { 1127 ResourceMark rm; 1128 1129 // Revoke biases of any monitors in the frame to ensure we can migrate them 1130 revoke_biases_of_monitors(thread, fr, ®_map); 1131 1132 DeoptReason reason = trap_request_reason(trap_request); 1133 DeoptAction action = trap_request_action(trap_request); 1134 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1135 1136 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1137 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1138 compiledVFrame* cvf = compiledVFrame::cast(vf); 1139 1140 nmethod* nm = cvf->code(); 1141 1142 ScopeDesc* trap_scope = cvf->scope(); 1143 methodHandle trap_method = trap_scope->method(); 1144 int trap_bci = trap_scope->bci(); 1145 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code(); 1146 1147 // Record this event in the histogram. 1148 gather_statistics(reason, action, trap_bc); 1149 1150 // Ensure that we can record deopt. history: 1151 bool create_if_missing = ProfileTraps; 1152 1153 methodDataHandle trap_mdo 1154 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1155 1156 // Print a bunch of diagnostics, if requested. 1157 if (TraceDeoptimization || LogCompilation) { 1158 ResourceMark rm; 1159 ttyLocker ttyl; 1160 char buf[100]; 1161 if (xtty != NULL) { 1162 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1163 os::current_thread_id(), 1164 format_trap_request(buf, sizeof(buf), trap_request)); 1165 nm->log_identity(xtty); 1166 } 1167 symbolHandle class_name; 1168 bool unresolved = false; 1169 if (unloaded_class_index >= 0) { 1170 constantPoolHandle constants (THREAD, trap_method->constants()); 1171 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1172 class_name = symbolHandle(THREAD, 1173 constants->klass_name_at(unloaded_class_index)); 1174 unresolved = true; 1175 if (xtty != NULL) 1176 xtty->print(" unresolved='1'"); 1177 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1178 class_name = symbolHandle(THREAD, 1179 constants->symbol_at(unloaded_class_index)); 1180 } 1181 if (xtty != NULL) 1182 xtty->name(class_name); 1183 } 1184 if (xtty != NULL && trap_mdo.not_null()) { 1185 // Dump the relevant MDO state. 1186 // This is the deopt count for the current reason, any previous 1187 // reasons or recompiles seen at this point. 1188 int dcnt = trap_mdo->trap_count(reason); 1189 if (dcnt != 0) 1190 xtty->print(" count='%d'", dcnt); 1191 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1192 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1193 if (dos != 0) { 1194 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1195 if (trap_state_is_recompiled(dos)) { 1196 int recnt2 = trap_mdo->overflow_recompile_count(); 1197 if (recnt2 != 0) 1198 xtty->print(" recompiles2='%d'", recnt2); 1199 } 1200 } 1201 } 1202 if (xtty != NULL) { 1203 xtty->stamp(); 1204 xtty->end_head(); 1205 } 1206 if (TraceDeoptimization) { // make noise on the tty 1207 tty->print("Uncommon trap occurred in"); 1208 nm->method()->print_short_name(tty); 1209 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1210 fr.pc(), 1211 (int) os::current_thread_id(), 1212 trap_reason_name(reason), 1213 trap_action_name(action), 1214 unloaded_class_index); 1215 if (class_name.not_null()) { 1216 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1217 class_name->print_symbol_on(tty); 1218 } 1219 tty->cr(); 1220 } 1221 if (xtty != NULL) { 1222 // Log the precise location of the trap. 1223 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1224 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1225 xtty->method(sd->method()); 1226 xtty->end_elem(); 1227 if (sd->is_top()) break; 1228 } 1229 xtty->tail("uncommon_trap"); 1230 } 1231 } 1232 // (End diagnostic printout.) 1233 1234 // Load class if necessary 1235 if (unloaded_class_index >= 0) { 1236 constantPoolHandle constants(THREAD, trap_method->constants()); 1237 load_class_by_index(constants, unloaded_class_index); 1238 } 1239 1240 // Flush the nmethod if necessary and desirable. 1241 // 1242 // We need to avoid situations where we are re-flushing the nmethod 1243 // because of a hot deoptimization site. Repeated flushes at the same 1244 // point need to be detected by the compiler and avoided. If the compiler 1245 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1246 // module must take measures to avoid an infinite cycle of recompilation 1247 // and deoptimization. There are several such measures: 1248 // 1249 // 1. If a recompilation is ordered a second time at some site X 1250 // and for the same reason R, the action is adjusted to 'reinterpret', 1251 // to give the interpreter time to exercise the method more thoroughly. 1252 // If this happens, the method's overflow_recompile_count is incremented. 1253 // 1254 // 2. If the compiler fails to reduce the deoptimization rate, then 1255 // the method's overflow_recompile_count will begin to exceed the set 1256 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1257 // is adjusted to 'make_not_compilable', and the method is abandoned 1258 // to the interpreter. This is a performance hit for hot methods, 1259 // but is better than a disastrous infinite cycle of recompilations. 1260 // (Actually, only the method containing the site X is abandoned.) 1261 // 1262 // 3. In parallel with the previous measures, if the total number of 1263 // recompilations of a method exceeds the much larger set limit 1264 // PerMethodRecompilationCutoff, the method is abandoned. 1265 // This should only happen if the method is very large and has 1266 // many "lukewarm" deoptimizations. The code which enforces this 1267 // limit is elsewhere (class nmethod, class methodOopDesc). 1268 // 1269 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1270 // to recompile at each bytecode independently of the per-BCI cutoff. 1271 // 1272 // The decision to update code is up to the compiler, and is encoded 1273 // in the Action_xxx code. If the compiler requests Action_none 1274 // no trap state is changed, no compiled code is changed, and the 1275 // computation suffers along in the interpreter. 1276 // 1277 // The other action codes specify various tactics for decompilation 1278 // and recompilation. Action_maybe_recompile is the loosest, and 1279 // allows the compiled code to stay around until enough traps are seen, 1280 // and until the compiler gets around to recompiling the trapping method. 1281 // 1282 // The other actions cause immediate removal of the present code. 1283 1284 bool update_trap_state = true; 1285 bool make_not_entrant = false; 1286 bool make_not_compilable = false; 1287 bool reset_counters = false; 1288 switch (action) { 1289 case Action_none: 1290 // Keep the old code. 1291 update_trap_state = false; 1292 break; 1293 case Action_maybe_recompile: 1294 // Do not need to invalidate the present code, but we can 1295 // initiate another 1296 // Start compiler without (necessarily) invalidating the nmethod. 1297 // The system will tolerate the old code, but new code should be 1298 // generated when possible. 1299 break; 1300 case Action_reinterpret: 1301 // Go back into the interpreter for a while, and then consider 1302 // recompiling form scratch. 1303 make_not_entrant = true; 1304 // Reset invocation counter for outer most method. 1305 // This will allow the interpreter to exercise the bytecodes 1306 // for a while before recompiling. 1307 // By contrast, Action_make_not_entrant is immediate. 1308 // 1309 // Note that the compiler will track null_check, null_assert, 1310 // range_check, and class_check events and log them as if they 1311 // had been traps taken from compiled code. This will update 1312 // the MDO trap history so that the next compilation will 1313 // properly detect hot trap sites. 1314 reset_counters = true; 1315 break; 1316 case Action_make_not_entrant: 1317 // Request immediate recompilation, and get rid of the old code. 1318 // Make them not entrant, so next time they are called they get 1319 // recompiled. Unloaded classes are loaded now so recompile before next 1320 // time they are called. Same for uninitialized. The interpreter will 1321 // link the missing class, if any. 1322 make_not_entrant = true; 1323 break; 1324 case Action_make_not_compilable: 1325 // Give up on compiling this method at all. 1326 make_not_entrant = true; 1327 make_not_compilable = true; 1328 break; 1329 default: 1330 ShouldNotReachHere(); 1331 } 1332 1333 // Setting +ProfileTraps fixes the following, on all platforms: 1334 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1335 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1336 // recompile relies on a methodDataOop to record heroic opt failures. 1337 1338 // Whether the interpreter is producing MDO data or not, we also need 1339 // to use the MDO to detect hot deoptimization points and control 1340 // aggressive optimization. 1341 bool inc_recompile_count = false; 1342 ProfileData* pdata = NULL; 1343 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1344 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1345 uint this_trap_count = 0; 1346 bool maybe_prior_trap = false; 1347 bool maybe_prior_recompile = false; 1348 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1349 //outputs: 1350 this_trap_count, 1351 maybe_prior_trap, 1352 maybe_prior_recompile); 1353 // Because the interpreter also counts null, div0, range, and class 1354 // checks, these traps from compiled code are double-counted. 1355 // This is harmless; it just means that the PerXTrapLimit values 1356 // are in effect a little smaller than they look. 1357 1358 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1359 if (per_bc_reason != Reason_none) { 1360 // Now take action based on the partially known per-BCI history. 1361 if (maybe_prior_trap 1362 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1363 // If there are too many traps at this BCI, force a recompile. 1364 // This will allow the compiler to see the limit overflow, and 1365 // take corrective action, if possible. The compiler generally 1366 // does not use the exact PerBytecodeTrapLimit value, but instead 1367 // changes its tactics if it sees any traps at all. This provides 1368 // a little hysteresis, delaying a recompile until a trap happens 1369 // several times. 1370 // 1371 // Actually, since there is only one bit of counter per BCI, 1372 // the possible per-BCI counts are {0,1,(per-method count)}. 1373 // This produces accurate results if in fact there is only 1374 // one hot trap site, but begins to get fuzzy if there are 1375 // many sites. For example, if there are ten sites each 1376 // trapping two or more times, they each get the blame for 1377 // all of their traps. 1378 make_not_entrant = true; 1379 } 1380 1381 // Detect repeated recompilation at the same BCI, and enforce a limit. 1382 if (make_not_entrant && maybe_prior_recompile) { 1383 // More than one recompile at this point. 1384 inc_recompile_count = maybe_prior_trap; 1385 } 1386 } else { 1387 // For reasons which are not recorded per-bytecode, we simply 1388 // force recompiles unconditionally. 1389 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1390 make_not_entrant = true; 1391 } 1392 1393 // Go back to the compiler if there are too many traps in this method. 1394 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1395 // If there are too many traps in this method, force a recompile. 1396 // This will allow the compiler to see the limit overflow, and 1397 // take corrective action, if possible. 1398 // (This condition is an unlikely backstop only, because the 1399 // PerBytecodeTrapLimit is more likely to take effect first, 1400 // if it is applicable.) 1401 make_not_entrant = true; 1402 } 1403 1404 // Here's more hysteresis: If there has been a recompile at 1405 // this trap point already, run the method in the interpreter 1406 // for a while to exercise it more thoroughly. 1407 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1408 reset_counters = true; 1409 } 1410 1411 } 1412 1413 // Take requested actions on the method: 1414 1415 // Recompile 1416 if (make_not_entrant) { 1417 if (!nm->make_not_entrant()) { 1418 return; // the call did not change nmethod's state 1419 } 1420 1421 if (pdata != NULL) { 1422 // Record the recompilation event, if any. 1423 int tstate0 = pdata->trap_state(); 1424 int tstate1 = trap_state_set_recompiled(tstate0, true); 1425 if (tstate1 != tstate0) 1426 pdata->set_trap_state(tstate1); 1427 } 1428 } 1429 1430 if (inc_recompile_count) { 1431 trap_mdo->inc_overflow_recompile_count(); 1432 if ((uint)trap_mdo->overflow_recompile_count() > 1433 (uint)PerBytecodeRecompilationCutoff) { 1434 // Give up on the method containing the bad BCI. 1435 if (trap_method() == nm->method()) { 1436 make_not_compilable = true; 1437 } else { 1438 trap_method->set_not_compilable(); 1439 // But give grace to the enclosing nm->method(). 1440 } 1441 } 1442 } 1443 1444 // Reset invocation counters 1445 if (reset_counters) { 1446 if (nm->is_osr_method()) 1447 reset_invocation_counter(trap_scope, CompileThreshold); 1448 else 1449 reset_invocation_counter(trap_scope); 1450 } 1451 1452 // Give up compiling 1453 if (make_not_compilable && !nm->method()->is_not_compilable()) { 1454 assert(make_not_entrant, "consistent"); 1455 nm->method()->set_not_compilable(); 1456 } 1457 1458 } // Free marked resources 1459 1460 } 1461 JRT_END 1462 1463 methodDataOop 1464 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1465 bool create_if_missing) { 1466 Thread* THREAD = thread; 1467 methodDataOop mdo = m()->method_data(); 1468 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1469 // Build an MDO. Ignore errors like OutOfMemory; 1470 // that simply means we won't have an MDO to update. 1471 methodOopDesc::build_interpreter_method_data(m, THREAD); 1472 if (HAS_PENDING_EXCEPTION) { 1473 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1474 CLEAR_PENDING_EXCEPTION; 1475 } 1476 mdo = m()->method_data(); 1477 } 1478 return mdo; 1479 } 1480 1481 ProfileData* 1482 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1483 int trap_bci, 1484 Deoptimization::DeoptReason reason, 1485 //outputs: 1486 uint& ret_this_trap_count, 1487 bool& ret_maybe_prior_trap, 1488 bool& ret_maybe_prior_recompile) { 1489 uint prior_trap_count = trap_mdo->trap_count(reason); 1490 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1491 1492 // If the runtime cannot find a place to store trap history, 1493 // it is estimated based on the general condition of the method. 1494 // If the method has ever been recompiled, or has ever incurred 1495 // a trap with the present reason , then this BCI is assumed 1496 // (pessimistically) to be the culprit. 1497 bool maybe_prior_trap = (prior_trap_count != 0); 1498 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1499 ProfileData* pdata = NULL; 1500 1501 1502 // For reasons which are recorded per bytecode, we check per-BCI data. 1503 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1504 if (per_bc_reason != Reason_none) { 1505 // Find the profile data for this BCI. If there isn't one, 1506 // try to allocate one from the MDO's set of spares. 1507 // This will let us detect a repeated trap at this point. 1508 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1509 1510 if (pdata != NULL) { 1511 // Query the trap state of this profile datum. 1512 int tstate0 = pdata->trap_state(); 1513 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1514 maybe_prior_trap = false; 1515 if (!trap_state_is_recompiled(tstate0)) 1516 maybe_prior_recompile = false; 1517 1518 // Update the trap state of this profile datum. 1519 int tstate1 = tstate0; 1520 // Record the reason. 1521 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1522 // Store the updated state on the MDO, for next time. 1523 if (tstate1 != tstate0) 1524 pdata->set_trap_state(tstate1); 1525 } else { 1526 if (LogCompilation && xtty != NULL) { 1527 ttyLocker ttyl; 1528 // Missing MDP? Leave a small complaint in the log. 1529 xtty->elem("missing_mdp bci='%d'", trap_bci); 1530 } 1531 } 1532 } 1533 1534 // Return results: 1535 ret_this_trap_count = this_trap_count; 1536 ret_maybe_prior_trap = maybe_prior_trap; 1537 ret_maybe_prior_recompile = maybe_prior_recompile; 1538 return pdata; 1539 } 1540 1541 void 1542 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1543 ResourceMark rm; 1544 // Ignored outputs: 1545 uint ignore_this_trap_count; 1546 bool ignore_maybe_prior_trap; 1547 bool ignore_maybe_prior_recompile; 1548 query_update_method_data(trap_mdo, trap_bci, 1549 (DeoptReason)reason, 1550 ignore_this_trap_count, 1551 ignore_maybe_prior_trap, 1552 ignore_maybe_prior_recompile); 1553 } 1554 1555 void Deoptimization::reset_invocation_counter(ScopeDesc* trap_scope, jint top_count) { 1556 ScopeDesc* sd = trap_scope; 1557 for (; !sd->is_top(); sd = sd->sender()) { 1558 // Reset ICs of inlined methods, since they can trigger compilations also. 1559 sd->method()->invocation_counter()->reset(); 1560 } 1561 InvocationCounter* c = sd->method()->invocation_counter(); 1562 if (top_count != _no_count) { 1563 // It was an OSR method, so bump the count higher. 1564 c->set(c->state(), top_count); 1565 } else { 1566 c->reset(); 1567 } 1568 sd->method()->backedge_counter()->reset(); 1569 } 1570 1571 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1572 1573 // Still in Java no safepoints 1574 { 1575 // This enters VM and may safepoint 1576 uncommon_trap_inner(thread, trap_request); 1577 } 1578 return fetch_unroll_info_helper(thread); 1579 } 1580 1581 // Local derived constants. 1582 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1583 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1584 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1585 1586 //---------------------------trap_state_reason--------------------------------- 1587 Deoptimization::DeoptReason 1588 Deoptimization::trap_state_reason(int trap_state) { 1589 // This assert provides the link between the width of DataLayout::trap_bits 1590 // and the encoding of "recorded" reasons. It ensures there are enough 1591 // bits to store all needed reasons in the per-BCI MDO profile. 1592 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1593 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1594 trap_state -= recompile_bit; 1595 if (trap_state == DS_REASON_MASK) { 1596 return Reason_many; 1597 } else { 1598 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1599 return (DeoptReason)trap_state; 1600 } 1601 } 1602 //-------------------------trap_state_has_reason------------------------------- 1603 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1604 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1605 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1606 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1607 trap_state -= recompile_bit; 1608 if (trap_state == DS_REASON_MASK) { 1609 return -1; // true, unspecifically (bottom of state lattice) 1610 } else if (trap_state == reason) { 1611 return 1; // true, definitely 1612 } else if (trap_state == 0) { 1613 return 0; // false, definitely (top of state lattice) 1614 } else { 1615 return 0; // false, definitely 1616 } 1617 } 1618 //-------------------------trap_state_add_reason------------------------------- 1619 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1620 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1621 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1622 trap_state -= recompile_bit; 1623 if (trap_state == DS_REASON_MASK) { 1624 return trap_state + recompile_bit; // already at state lattice bottom 1625 } else if (trap_state == reason) { 1626 return trap_state + recompile_bit; // the condition is already true 1627 } else if (trap_state == 0) { 1628 return reason + recompile_bit; // no condition has yet been true 1629 } else { 1630 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1631 } 1632 } 1633 //-----------------------trap_state_is_recompiled------------------------------ 1634 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1635 return (trap_state & DS_RECOMPILE_BIT) != 0; 1636 } 1637 //-----------------------trap_state_set_recompiled----------------------------- 1638 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1639 if (z) return trap_state | DS_RECOMPILE_BIT; 1640 else return trap_state & ~DS_RECOMPILE_BIT; 1641 } 1642 //---------------------------format_trap_state--------------------------------- 1643 // This is used for debugging and diagnostics, including hotspot.log output. 1644 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1645 int trap_state) { 1646 DeoptReason reason = trap_state_reason(trap_state); 1647 bool recomp_flag = trap_state_is_recompiled(trap_state); 1648 // Re-encode the state from its decoded components. 1649 int decoded_state = 0; 1650 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1651 decoded_state = trap_state_add_reason(decoded_state, reason); 1652 if (recomp_flag) 1653 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1654 // If the state re-encodes properly, format it symbolically. 1655 // Because this routine is used for debugging and diagnostics, 1656 // be robust even if the state is a strange value. 1657 size_t len; 1658 if (decoded_state != trap_state) { 1659 // Random buggy state that doesn't decode?? 1660 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1661 } else { 1662 len = jio_snprintf(buf, buflen, "%s%s", 1663 trap_reason_name(reason), 1664 recomp_flag ? " recompiled" : ""); 1665 } 1666 if (len >= buflen) 1667 buf[buflen-1] = '\0'; 1668 return buf; 1669 } 1670 1671 1672 //--------------------------------statics-------------------------------------- 1673 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1674 = Deoptimization::Action_reinterpret; 1675 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1676 // Note: Keep this in sync. with enum DeoptReason. 1677 "none", 1678 "null_check", 1679 "null_assert", 1680 "range_check", 1681 "class_check", 1682 "array_check", 1683 "intrinsic", 1684 "bimorphic", 1685 "unloaded", 1686 "uninitialized", 1687 "unreached", 1688 "unhandled", 1689 "constraint", 1690 "div0_check", 1691 "age", 1692 "predicate" 1693 }; 1694 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1695 // Note: Keep this in sync. with enum DeoptAction. 1696 "none", 1697 "maybe_recompile", 1698 "reinterpret", 1699 "make_not_entrant", 1700 "make_not_compilable" 1701 }; 1702 1703 const char* Deoptimization::trap_reason_name(int reason) { 1704 if (reason == Reason_many) return "many"; 1705 if ((uint)reason < Reason_LIMIT) 1706 return _trap_reason_name[reason]; 1707 static char buf[20]; 1708 sprintf(buf, "reason%d", reason); 1709 return buf; 1710 } 1711 const char* Deoptimization::trap_action_name(int action) { 1712 if ((uint)action < Action_LIMIT) 1713 return _trap_action_name[action]; 1714 static char buf[20]; 1715 sprintf(buf, "action%d", action); 1716 return buf; 1717 } 1718 1719 // This is used for debugging and diagnostics, including hotspot.log output. 1720 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1721 int trap_request) { 1722 jint unloaded_class_index = trap_request_index(trap_request); 1723 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1724 const char* action = trap_action_name(trap_request_action(trap_request)); 1725 size_t len; 1726 if (unloaded_class_index < 0) { 1727 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1728 reason, action); 1729 } else { 1730 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1731 reason, action, unloaded_class_index); 1732 } 1733 if (len >= buflen) 1734 buf[buflen-1] = '\0'; 1735 return buf; 1736 } 1737 1738 juint Deoptimization::_deoptimization_hist 1739 [Deoptimization::Reason_LIMIT] 1740 [1 + Deoptimization::Action_LIMIT] 1741 [Deoptimization::BC_CASE_LIMIT] 1742 = {0}; 1743 1744 enum { 1745 LSB_BITS = 8, 1746 LSB_MASK = right_n_bits(LSB_BITS) 1747 }; 1748 1749 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1750 Bytecodes::Code bc) { 1751 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1752 assert(action >= 0 && action < Action_LIMIT, "oob"); 1753 _deoptimization_hist[Reason_none][0][0] += 1; // total 1754 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1755 juint* cases = _deoptimization_hist[reason][1+action]; 1756 juint* bc_counter_addr = NULL; 1757 juint bc_counter = 0; 1758 // Look for an unused counter, or an exact match to this BC. 1759 if (bc != Bytecodes::_illegal) { 1760 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1761 juint* counter_addr = &cases[bc_case]; 1762 juint counter = *counter_addr; 1763 if ((counter == 0 && bc_counter_addr == NULL) 1764 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1765 // this counter is either free or is already devoted to this BC 1766 bc_counter_addr = counter_addr; 1767 bc_counter = counter | bc; 1768 } 1769 } 1770 } 1771 if (bc_counter_addr == NULL) { 1772 // Overflow, or no given bytecode. 1773 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1774 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1775 } 1776 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1777 } 1778 1779 jint Deoptimization::total_deoptimization_count() { 1780 return _deoptimization_hist[Reason_none][0][0]; 1781 } 1782 1783 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1784 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1785 return _deoptimization_hist[reason][0][0]; 1786 } 1787 1788 void Deoptimization::print_statistics() { 1789 juint total = total_deoptimization_count(); 1790 juint account = total; 1791 if (total != 0) { 1792 ttyLocker ttyl; 1793 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1794 tty->print_cr("Deoptimization traps recorded:"); 1795 #define PRINT_STAT_LINE(name, r) \ 1796 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1797 PRINT_STAT_LINE("total", total); 1798 // For each non-zero entry in the histogram, print the reason, 1799 // the action, and (if specifically known) the type of bytecode. 1800 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1801 for (int action = 0; action < Action_LIMIT; action++) { 1802 juint* cases = _deoptimization_hist[reason][1+action]; 1803 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1804 juint counter = cases[bc_case]; 1805 if (counter != 0) { 1806 char name[1*K]; 1807 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1808 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1809 bc = Bytecodes::_illegal; 1810 sprintf(name, "%s/%s/%s", 1811 trap_reason_name(reason), 1812 trap_action_name(action), 1813 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1814 juint r = counter >> LSB_BITS; 1815 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1816 account -= r; 1817 } 1818 } 1819 } 1820 } 1821 if (account != 0) { 1822 PRINT_STAT_LINE("unaccounted", account); 1823 } 1824 #undef PRINT_STAT_LINE 1825 if (xtty != NULL) xtty->tail("statistics"); 1826 } 1827 } 1828 #else // COMPILER2 1829 1830 1831 // Stubs for C1 only system. 1832 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1833 return false; 1834 } 1835 1836 const char* Deoptimization::trap_reason_name(int reason) { 1837 return "unknown"; 1838 } 1839 1840 void Deoptimization::print_statistics() { 1841 // no output 1842 } 1843 1844 void 1845 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1846 // no udpate 1847 } 1848 1849 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1850 return 0; 1851 } 1852 1853 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1854 Bytecodes::Code bc) { 1855 // no update 1856 } 1857 1858 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1859 int trap_state) { 1860 jio_snprintf(buf, buflen, "#%d", trap_state); 1861 return buf; 1862 } 1863 1864 #endif // COMPILER2