1 /* 2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "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 // Set the deoptee nmethod 128 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 129 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 130 131 // Create a growable array of VFrames where each VFrame represents an inlined 132 // Java frame. This storage is allocated with the usual system arena. 133 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 134 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 135 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 136 while (!vf->is_top()) { 137 assert(vf->is_compiled_frame(), "Wrong frame type"); 138 chunk->push(compiledVFrame::cast(vf)); 139 vf = vf->sender(); 140 } 141 assert(vf->is_compiled_frame(), "Wrong frame type"); 142 chunk->push(compiledVFrame::cast(vf)); 143 144 #ifdef COMPILER2 145 // Reallocate the non-escaping objects and restore their fields. Then 146 // relock objects if synchronization on them was eliminated. 147 if (DoEscapeAnalysis) { 148 if (EliminateAllocations) { 149 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 150 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 151 152 // The flag return_oop() indicates call sites which return oop 153 // in compiled code. Such sites include java method calls, 154 // runtime calls (for example, used to allocate new objects/arrays 155 // on slow code path) and any other calls generated in compiled code. 156 // It is not guaranteed that we can get such information here only 157 // by analyzing bytecode in deoptimized frames. This is why this flag 158 // is set during method compilation (see Compile::Process_OopMap_Node()). 159 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 160 Handle return_value; 161 if (save_oop_result) { 162 // Reallocation may trigger GC. If deoptimization happened on return from 163 // call which returns oop we need to save it since it is not in oopmap. 164 oop result = deoptee.saved_oop_result(&map); 165 assert(result == NULL || result->is_oop(), "must be oop"); 166 return_value = Handle(thread, result); 167 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 168 if (TraceDeoptimization) { 169 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 170 } 171 } 172 bool reallocated = false; 173 if (objects != NULL) { 174 JRT_BLOCK 175 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 176 JRT_END 177 } 178 if (reallocated) { 179 reassign_fields(&deoptee, &map, objects); 180 #ifndef PRODUCT 181 if (TraceDeoptimization) { 182 ttyLocker ttyl; 183 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 184 print_objects(objects); 185 } 186 #endif 187 } 188 if (save_oop_result) { 189 // Restore result. 190 deoptee.set_saved_oop_result(&map, return_value()); 191 } 192 } 193 if (EliminateLocks) { 194 #ifndef PRODUCT 195 bool first = true; 196 #endif 197 for (int i = 0; i < chunk->length(); i++) { 198 compiledVFrame* cvf = chunk->at(i); 199 assert (cvf->scope() != NULL,"expect only compiled java frames"); 200 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 201 if (monitors->is_nonempty()) { 202 relock_objects(monitors, thread); 203 #ifndef PRODUCT 204 if (TraceDeoptimization) { 205 ttyLocker ttyl; 206 for (int j = 0; j < monitors->length(); j++) { 207 MonitorInfo* mi = monitors->at(j); 208 if (mi->eliminated()) { 209 if (first) { 210 first = false; 211 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 212 } 213 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 214 } 215 } 216 } 217 #endif 218 } 219 } 220 } 221 } 222 #endif // COMPILER2 223 // Ensure that no safepoint is taken after pointers have been stored 224 // in fields of rematerialized objects. If a safepoint occurs from here on 225 // out the java state residing in the vframeArray will be missed. 226 No_Safepoint_Verifier no_safepoint; 227 228 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 229 230 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 231 thread->set_vframe_array_head(array); 232 233 // Now that the vframeArray has been created if we have any deferred local writes 234 // added by jvmti then we can free up that structure as the data is now in the 235 // vframeArray 236 237 if (thread->deferred_locals() != NULL) { 238 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 239 int i = 0; 240 do { 241 // Because of inlining we could have multiple vframes for a single frame 242 // and several of the vframes could have deferred writes. Find them all. 243 if (list->at(i)->id() == array->original().id()) { 244 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 245 list->remove_at(i); 246 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 247 delete dlv; 248 } else { 249 i++; 250 } 251 } while ( i < list->length() ); 252 if (list->length() == 0) { 253 thread->set_deferred_locals(NULL); 254 // free the list and elements back to C heap. 255 delete list; 256 } 257 258 } 259 260 #ifndef SHARK 261 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 262 CodeBlob* cb = stub_frame.cb(); 263 // Verify we have the right vframeArray 264 assert(cb->frame_size() >= 0, "Unexpected frame size"); 265 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 266 267 // If the deopt call site is a MethodHandle invoke call site we have 268 // to adjust the unpack_sp. 269 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 270 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 271 unpack_sp = deoptee.unextended_sp(); 272 273 #ifdef ASSERT 274 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 275 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 276 #endif 277 #else 278 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 279 #endif // !SHARK 280 281 // This is a guarantee instead of an assert because if vframe doesn't match 282 // we will unpack the wrong deoptimized frame and wind up in strange places 283 // where it will be very difficult to figure out what went wrong. Better 284 // to die an early death here than some very obscure death later when the 285 // trail is cold. 286 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 287 // in that it will fail to detect a problem when there is one. This needs 288 // more work in tiger timeframe. 289 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 290 291 int number_of_frames = array->frames(); 292 293 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 294 // virtual activation, which is the reverse of the elements in the vframes array. 295 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 296 // +1 because we always have an interpreter return address for the final slot. 297 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 298 int callee_parameters = 0; 299 int callee_locals = 0; 300 int popframe_extra_args = 0; 301 // Create an interpreter return address for the stub to use as its return 302 // address so the skeletal frames are perfectly walkable 303 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 304 305 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 306 // activation be put back on the expression stack of the caller for reexecution 307 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 308 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 309 } 310 311 // 312 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 313 // frame_sizes/frame_pcs[1] next oldest frame (int) 314 // frame_sizes/frame_pcs[n] youngest frame (int) 315 // 316 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 317 // owns the space for the return address to it's caller). Confusing ain't it. 318 // 319 // The vframe array can address vframes with indices running from 320 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 321 // When we create the skeletal frames we need the oldest frame to be in the zero slot 322 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 323 // so things look a little strange in this loop. 324 // 325 for (int index = 0; index < array->frames(); index++ ) { 326 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 327 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 328 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 329 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 330 callee_locals, 331 index == 0, 332 popframe_extra_args); 333 // This pc doesn't have to be perfect just good enough to identify the frame 334 // as interpreted so the skeleton frame will be walkable 335 // The correct pc will be set when the skeleton frame is completely filled out 336 // The final pc we store in the loop is wrong and will be overwritten below 337 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 338 339 callee_parameters = array->element(index)->method()->size_of_parameters(); 340 callee_locals = array->element(index)->method()->max_locals(); 341 popframe_extra_args = 0; 342 } 343 344 // Compute whether the root vframe returns a float or double value. 345 BasicType return_type; 346 { 347 HandleMark hm; 348 methodHandle method(thread, array->element(0)->method()); 349 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci()); 350 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL; 351 } 352 353 // Compute information for handling adapters and adjusting the frame size of the caller. 354 int caller_adjustment = 0; 355 356 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 357 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 358 // than simply use array->sender.pc(). This requires us to walk the current set of frames 359 // 360 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 361 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 362 363 // Compute the amount the oldest interpreter frame will have to adjust 364 // its caller's stack by. If the caller is a compiled frame then 365 // we pretend that the callee has no parameters so that the 366 // extension counts for the full amount of locals and not just 367 // locals-parms. This is because without a c2i adapter the parm 368 // area as created by the compiled frame will not be usable by 369 // the interpreter. (Depending on the calling convention there 370 // may not even be enough space). 371 372 // QQQ I'd rather see this pushed down into last_frame_adjust 373 // and have it take the sender (aka caller). 374 375 if (deopt_sender.is_compiled_frame()) { 376 caller_adjustment = last_frame_adjust(0, callee_locals); 377 } else if (callee_locals > callee_parameters) { 378 // The caller frame may need extending to accommodate 379 // non-parameter locals of the first unpacked interpreted frame. 380 // Compute that adjustment. 381 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 382 } 383 384 385 // If the sender is deoptimized the we must retrieve the address of the handler 386 // since the frame will "magically" show the original pc before the deopt 387 // and we'd undo the deopt. 388 389 frame_pcs[0] = deopt_sender.raw_pc(); 390 391 #ifndef SHARK 392 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 393 #endif // SHARK 394 395 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 396 caller_adjustment * BytesPerWord, 397 number_of_frames, 398 frame_sizes, 399 frame_pcs, 400 return_type); 401 #if defined(IA32) || defined(AMD64) 402 // We need a way to pass fp to the unpacking code so the skeletal frames 403 // come out correct. This is only needed for x86 because of c2 using ebp 404 // as an allocatable register. So this update is useless (and harmless) 405 // on the other platforms. It would be nice to do this in a different 406 // way but even the old style deoptimization had a problem with deriving 407 // this value. NEEDS_CLEANUP 408 // Note: now that c1 is using c2's deopt blob we must do this on all 409 // x86 based platforms 410 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); 411 *fp_addr = array->sender().fp(); // was adapter_caller 412 #endif /* IA32 || AMD64 */ 413 414 if (array->frames() > 1) { 415 if (VerifyStack && TraceDeoptimization) { 416 tty->print_cr("Deoptimizing method containing inlining"); 417 } 418 } 419 420 array->set_unroll_block(info); 421 return info; 422 } 423 424 // Called to cleanup deoptimization data structures in normal case 425 // after unpacking to stack and when stack overflow error occurs 426 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 427 vframeArray *array) { 428 429 // Get array if coming from exception 430 if (array == NULL) { 431 array = thread->vframe_array_head(); 432 } 433 thread->set_vframe_array_head(NULL); 434 435 // Free the previous UnrollBlock 436 vframeArray* old_array = thread->vframe_array_last(); 437 thread->set_vframe_array_last(array); 438 439 if (old_array != NULL) { 440 UnrollBlock* old_info = old_array->unroll_block(); 441 old_array->set_unroll_block(NULL); 442 delete old_info; 443 delete old_array; 444 } 445 446 // Deallocate any resource creating in this routine and any ResourceObjs allocated 447 // inside the vframeArray (StackValueCollections) 448 449 delete thread->deopt_mark(); 450 thread->set_deopt_mark(NULL); 451 thread->set_deopt_nmethod(NULL); 452 453 454 if (JvmtiExport::can_pop_frame()) { 455 #ifndef CC_INTERP 456 // Regardless of whether we entered this routine with the pending 457 // popframe condition bit set, we should always clear it now 458 thread->clear_popframe_condition(); 459 #else 460 // C++ interpeter will clear has_pending_popframe when it enters 461 // with method_resume. For deopt_resume2 we clear it now. 462 if (thread->popframe_forcing_deopt_reexecution()) 463 thread->clear_popframe_condition(); 464 #endif /* CC_INTERP */ 465 } 466 467 // unpack_frames() is called at the end of the deoptimization handler 468 // and (in C2) at the end of the uncommon trap handler. Note this fact 469 // so that an asynchronous stack walker can work again. This counter is 470 // incremented at the beginning of fetch_unroll_info() and (in C2) at 471 // the beginning of uncommon_trap(). 472 thread->dec_in_deopt_handler(); 473 } 474 475 476 // Return BasicType of value being returned 477 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 478 479 // We are already active int he special DeoptResourceMark any ResourceObj's we 480 // allocate will be freed at the end of the routine. 481 482 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 483 // but makes the entry a little slower. There is however a little dance we have to 484 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 485 ResetNoHandleMark rnhm; // No-op in release/product versions 486 HandleMark hm; 487 488 frame stub_frame = thread->last_frame(); 489 490 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 491 // must point to the vframeArray for the unpack frame. 492 vframeArray* array = thread->vframe_array_head(); 493 494 #ifndef PRODUCT 495 if (TraceDeoptimization) { 496 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 497 } 498 #endif 499 500 UnrollBlock* info = array->unroll_block(); 501 502 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 503 array->unpack_to_stack(stub_frame, exec_mode); 504 505 BasicType bt = info->return_type(); 506 507 // If we have an exception pending, claim that the return type is an oop 508 // so the deopt_blob does not overwrite the exception_oop. 509 510 if (exec_mode == Unpack_exception) 511 bt = T_OBJECT; 512 513 // Cleanup thread deopt data 514 cleanup_deopt_info(thread, array); 515 516 #ifndef PRODUCT 517 if (VerifyStack) { 518 ResourceMark res_mark; 519 520 // Verify that the just-unpacked frames match the interpreter's 521 // notions of expression stack and locals 522 vframeArray* cur_array = thread->vframe_array_last(); 523 RegisterMap rm(thread, false); 524 rm.set_include_argument_oops(false); 525 bool is_top_frame = true; 526 int callee_size_of_parameters = 0; 527 int callee_max_locals = 0; 528 for (int i = 0; i < cur_array->frames(); i++) { 529 vframeArrayElement* el = cur_array->element(i); 530 frame* iframe = el->iframe(); 531 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 532 533 // Get the oop map for this bci 534 InterpreterOopMap mask; 535 int cur_invoke_parameter_size = 0; 536 bool try_next_mask = false; 537 int next_mask_expression_stack_size = -1; 538 int top_frame_expression_stack_adjustment = 0; 539 methodHandle mh(thread, iframe->interpreter_frame_method()); 540 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 541 BytecodeStream str(mh); 542 str.set_start(iframe->interpreter_frame_bci()); 543 int max_bci = mh->code_size(); 544 // Get to the next bytecode if possible 545 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 546 // Check to see if we can grab the number of outgoing arguments 547 // at an uncommon trap for an invoke (where the compiler 548 // generates debug info before the invoke has executed) 549 Bytecodes::Code cur_code = str.next(); 550 if (cur_code == Bytecodes::_invokevirtual || 551 cur_code == Bytecodes::_invokespecial || 552 cur_code == Bytecodes::_invokestatic || 553 cur_code == Bytecodes::_invokeinterface) { 554 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci()); 555 symbolHandle signature(thread, invoke->signature()); 556 ArgumentSizeComputer asc(signature); 557 cur_invoke_parameter_size = asc.size(); 558 if (cur_code != Bytecodes::_invokestatic) { 559 // Add in receiver 560 ++cur_invoke_parameter_size; 561 } 562 } 563 if (str.bci() < max_bci) { 564 Bytecodes::Code bc = str.next(); 565 if (bc >= 0) { 566 // The interpreter oop map generator reports results before 567 // the current bytecode has executed except in the case of 568 // calls. It seems to be hard to tell whether the compiler 569 // has emitted debug information matching the "state before" 570 // a given bytecode or the state after, so we try both 571 switch (cur_code) { 572 case Bytecodes::_invokevirtual: 573 case Bytecodes::_invokespecial: 574 case Bytecodes::_invokestatic: 575 case Bytecodes::_invokeinterface: 576 case Bytecodes::_athrow: 577 break; 578 default: { 579 InterpreterOopMap next_mask; 580 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 581 next_mask_expression_stack_size = next_mask.expression_stack_size(); 582 // Need to subtract off the size of the result type of 583 // the bytecode because this is not described in the 584 // debug info but returned to the interpreter in the TOS 585 // caching register 586 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 587 if (bytecode_result_type != T_ILLEGAL) { 588 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 589 } 590 assert(top_frame_expression_stack_adjustment >= 0, ""); 591 try_next_mask = true; 592 break; 593 } 594 } 595 } 596 } 597 598 // Verify stack depth and oops in frame 599 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 600 if (!( 601 /* SPARC */ 602 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 603 /* x86 */ 604 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 605 (try_next_mask && 606 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 607 top_frame_expression_stack_adjustment))) || 608 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 609 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 610 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 611 )) { 612 ttyLocker ttyl; 613 614 // Print out some information that will help us debug the problem 615 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 616 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 617 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 618 iframe->interpreter_frame_expression_stack_size()); 619 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 620 tty->print_cr(" try_next_mask = %d", try_next_mask); 621 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 622 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 623 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 624 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 625 tty->print_cr(" exec_mode = %d", exec_mode); 626 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 627 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 628 tty->print_cr(" Interpreted frames:"); 629 for (int k = 0; k < cur_array->frames(); k++) { 630 vframeArrayElement* el = cur_array->element(k); 631 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 632 } 633 cur_array->print_on_2(tty); 634 guarantee(false, "wrong number of expression stack elements during deopt"); 635 } 636 VerifyOopClosure verify; 637 iframe->oops_interpreted_do(&verify, &rm, false); 638 callee_size_of_parameters = mh->size_of_parameters(); 639 callee_max_locals = mh->max_locals(); 640 is_top_frame = false; 641 } 642 } 643 #endif /* !PRODUCT */ 644 645 646 return bt; 647 JRT_END 648 649 650 int Deoptimization::deoptimize_dependents() { 651 Threads::deoptimized_wrt_marked_nmethods(); 652 return 0; 653 } 654 655 656 #ifdef COMPILER2 657 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 658 Handle pending_exception(thread->pending_exception()); 659 const char* exception_file = thread->exception_file(); 660 int exception_line = thread->exception_line(); 661 thread->clear_pending_exception(); 662 663 for (int i = 0; i < objects->length(); i++) { 664 assert(objects->at(i)->is_object(), "invalid debug information"); 665 ObjectValue* sv = (ObjectValue*) objects->at(i); 666 667 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 668 oop obj = NULL; 669 670 if (k->oop_is_instance()) { 671 instanceKlass* ik = instanceKlass::cast(k()); 672 obj = ik->allocate_instance(CHECK_(false)); 673 } else if (k->oop_is_typeArray()) { 674 typeArrayKlass* ak = typeArrayKlass::cast(k()); 675 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 676 int len = sv->field_size() / type2size[ak->element_type()]; 677 obj = ak->allocate(len, CHECK_(false)); 678 } else if (k->oop_is_objArray()) { 679 objArrayKlass* ak = objArrayKlass::cast(k()); 680 obj = ak->allocate(sv->field_size(), CHECK_(false)); 681 } 682 683 assert(obj != NULL, "allocation failed"); 684 assert(sv->value().is_null(), "redundant reallocation"); 685 sv->set_value(obj); 686 } 687 688 if (pending_exception.not_null()) { 689 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 690 } 691 692 return true; 693 } 694 695 // This assumes that the fields are stored in ObjectValue in the same order 696 // they are yielded by do_nonstatic_fields. 697 class FieldReassigner: public FieldClosure { 698 frame* _fr; 699 RegisterMap* _reg_map; 700 ObjectValue* _sv; 701 instanceKlass* _ik; 702 oop _obj; 703 704 int _i; 705 public: 706 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 707 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 708 709 int i() const { return _i; } 710 711 712 void do_field(fieldDescriptor* fd) { 713 intptr_t val; 714 StackValue* value = 715 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 716 int offset = fd->offset(); 717 switch (fd->field_type()) { 718 case T_OBJECT: case T_ARRAY: 719 assert(value->type() == T_OBJECT, "Agreement."); 720 _obj->obj_field_put(offset, value->get_obj()()); 721 break; 722 723 case T_LONG: case T_DOUBLE: { 724 assert(value->type() == T_INT, "Agreement."); 725 StackValue* low = 726 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 727 #ifdef _LP64 728 jlong res = (jlong)low->get_int(); 729 #else 730 #ifdef SPARC 731 // For SPARC we have to swap high and low words. 732 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 733 #else 734 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 735 #endif //SPARC 736 #endif 737 _obj->long_field_put(offset, res); 738 break; 739 } 740 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 741 case T_INT: case T_FLOAT: // 4 bytes. 742 assert(value->type() == T_INT, "Agreement."); 743 val = value->get_int(); 744 _obj->int_field_put(offset, (jint)*((jint*)&val)); 745 break; 746 747 case T_SHORT: case T_CHAR: // 2 bytes 748 assert(value->type() == T_INT, "Agreement."); 749 val = value->get_int(); 750 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 751 break; 752 753 case T_BOOLEAN: case T_BYTE: // 1 byte 754 assert(value->type() == T_INT, "Agreement."); 755 val = value->get_int(); 756 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 757 break; 758 759 default: 760 ShouldNotReachHere(); 761 } 762 _i++; 763 } 764 }; 765 766 // restore elements of an eliminated type array 767 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 768 int index = 0; 769 intptr_t val; 770 771 for (int i = 0; i < sv->field_size(); i++) { 772 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 773 switch(type) { 774 case T_LONG: case T_DOUBLE: { 775 assert(value->type() == T_INT, "Agreement."); 776 StackValue* low = 777 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 778 #ifdef _LP64 779 jlong res = (jlong)low->get_int(); 780 #else 781 #ifdef SPARC 782 // For SPARC we have to swap high and low words. 783 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 784 #else 785 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 786 #endif //SPARC 787 #endif 788 obj->long_at_put(index, res); 789 break; 790 } 791 792 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 793 case T_INT: case T_FLOAT: // 4 bytes. 794 assert(value->type() == T_INT, "Agreement."); 795 val = value->get_int(); 796 obj->int_at_put(index, (jint)*((jint*)&val)); 797 break; 798 799 case T_SHORT: case T_CHAR: // 2 bytes 800 assert(value->type() == T_INT, "Agreement."); 801 val = value->get_int(); 802 obj->short_at_put(index, (jshort)*((jint*)&val)); 803 break; 804 805 case T_BOOLEAN: case T_BYTE: // 1 byte 806 assert(value->type() == T_INT, "Agreement."); 807 val = value->get_int(); 808 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 809 break; 810 811 default: 812 ShouldNotReachHere(); 813 } 814 index++; 815 } 816 } 817 818 819 // restore fields of an eliminated object array 820 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 821 for (int i = 0; i < sv->field_size(); i++) { 822 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 823 assert(value->type() == T_OBJECT, "object element expected"); 824 obj->obj_at_put(i, value->get_obj()()); 825 } 826 } 827 828 829 // restore fields of all eliminated objects and arrays 830 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 831 for (int i = 0; i < objects->length(); i++) { 832 ObjectValue* sv = (ObjectValue*) objects->at(i); 833 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 834 Handle obj = sv->value(); 835 assert(obj.not_null(), "reallocation was missed"); 836 837 if (k->oop_is_instance()) { 838 instanceKlass* ik = instanceKlass::cast(k()); 839 FieldReassigner reassign(fr, reg_map, sv, obj()); 840 ik->do_nonstatic_fields(&reassign); 841 } else if (k->oop_is_typeArray()) { 842 typeArrayKlass* ak = typeArrayKlass::cast(k()); 843 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 844 } else if (k->oop_is_objArray()) { 845 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 846 } 847 } 848 } 849 850 851 // relock objects for which synchronization was eliminated 852 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 853 for (int i = 0; i < monitors->length(); i++) { 854 MonitorInfo* mon_info = monitors->at(i); 855 if (mon_info->eliminated()) { 856 assert(mon_info->owner() != NULL, "reallocation was missed"); 857 Handle obj = Handle(mon_info->owner()); 858 markOop mark = obj->mark(); 859 if (UseBiasedLocking && mark->has_bias_pattern()) { 860 // New allocated objects may have the mark set to anonymously biased. 861 // Also the deoptimized method may called methods with synchronization 862 // where the thread-local object is bias locked to the current thread. 863 assert(mark->is_biased_anonymously() || 864 mark->biased_locker() == thread, "should be locked to current thread"); 865 // Reset mark word to unbiased prototype. 866 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 867 obj->set_mark(unbiased_prototype); 868 } 869 BasicLock* lock = mon_info->lock(); 870 ObjectSynchronizer::slow_enter(obj, lock, thread); 871 } 872 assert(mon_info->owner()->is_locked(), "object must be locked now"); 873 } 874 } 875 876 877 #ifndef PRODUCT 878 // print information about reallocated objects 879 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 880 fieldDescriptor fd; 881 882 for (int i = 0; i < objects->length(); i++) { 883 ObjectValue* sv = (ObjectValue*) objects->at(i); 884 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 885 Handle obj = sv->value(); 886 887 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 888 k->as_klassOop()->print_value(); 889 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 890 tty->cr(); 891 892 if (Verbose) { 893 k->oop_print_on(obj(), tty); 894 } 895 } 896 } 897 #endif 898 #endif // COMPILER2 899 900 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 901 902 #ifndef PRODUCT 903 if (TraceDeoptimization) { 904 ttyLocker ttyl; 905 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 906 fr.print_on(tty); 907 tty->print_cr(" Virtual frames (innermost first):"); 908 for (int index = 0; index < chunk->length(); index++) { 909 compiledVFrame* vf = chunk->at(index); 910 tty->print(" %2d - ", index); 911 vf->print_value(); 912 int bci = chunk->at(index)->raw_bci(); 913 const char* code_name; 914 if (bci == SynchronizationEntryBCI) { 915 code_name = "sync entry"; 916 } else { 917 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci); 918 code_name = Bytecodes::name(code); 919 } 920 tty->print(" - %s", code_name); 921 tty->print_cr(" @ bci %d ", bci); 922 if (Verbose) { 923 vf->print(); 924 tty->cr(); 925 } 926 } 927 } 928 #endif 929 930 // Register map for next frame (used for stack crawl). We capture 931 // the state of the deopt'ing frame's caller. Thus if we need to 932 // stuff a C2I adapter we can properly fill in the callee-save 933 // register locations. 934 frame caller = fr.sender(reg_map); 935 int frame_size = caller.sp() - fr.sp(); 936 937 frame sender = caller; 938 939 // Since the Java thread being deoptimized will eventually adjust it's own stack, 940 // the vframeArray containing the unpacking information is allocated in the C heap. 941 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 942 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 943 944 // Compare the vframeArray to the collected vframes 945 assert(array->structural_compare(thread, chunk), "just checking"); 946 Events::log("# vframes = %d", (intptr_t)chunk->length()); 947 948 #ifndef PRODUCT 949 if (TraceDeoptimization) { 950 ttyLocker ttyl; 951 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 952 } 953 #endif // PRODUCT 954 955 return array; 956 } 957 958 959 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 960 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 961 for (int i = 0; i < monitors->length(); i++) { 962 MonitorInfo* mon_info = monitors->at(i); 963 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 964 objects_to_revoke->append(Handle(mon_info->owner())); 965 } 966 } 967 } 968 969 970 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 971 if (!UseBiasedLocking) { 972 return; 973 } 974 975 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 976 977 // Unfortunately we don't have a RegisterMap available in most of 978 // the places we want to call this routine so we need to walk the 979 // stack again to update the register map. 980 if (map == NULL || !map->update_map()) { 981 StackFrameStream sfs(thread, true); 982 bool found = false; 983 while (!found && !sfs.is_done()) { 984 frame* cur = sfs.current(); 985 sfs.next(); 986 found = cur->id() == fr.id(); 987 } 988 assert(found, "frame to be deoptimized not found on target thread's stack"); 989 map = sfs.register_map(); 990 } 991 992 vframe* vf = vframe::new_vframe(&fr, map, thread); 993 compiledVFrame* cvf = compiledVFrame::cast(vf); 994 // Revoke monitors' biases in all scopes 995 while (!cvf->is_top()) { 996 collect_monitors(cvf, objects_to_revoke); 997 cvf = compiledVFrame::cast(cvf->sender()); 998 } 999 collect_monitors(cvf, objects_to_revoke); 1000 1001 if (SafepointSynchronize::is_at_safepoint()) { 1002 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1003 } else { 1004 BiasedLocking::revoke(objects_to_revoke); 1005 } 1006 } 1007 1008 1009 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1010 if (!UseBiasedLocking) { 1011 return; 1012 } 1013 1014 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1015 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1016 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1017 if (jt->has_last_Java_frame()) { 1018 StackFrameStream sfs(jt, true); 1019 while (!sfs.is_done()) { 1020 frame* cur = sfs.current(); 1021 if (cb->contains(cur->pc())) { 1022 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1023 compiledVFrame* cvf = compiledVFrame::cast(vf); 1024 // Revoke monitors' biases in all scopes 1025 while (!cvf->is_top()) { 1026 collect_monitors(cvf, objects_to_revoke); 1027 cvf = compiledVFrame::cast(cvf->sender()); 1028 } 1029 collect_monitors(cvf, objects_to_revoke); 1030 } 1031 sfs.next(); 1032 } 1033 } 1034 } 1035 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1036 } 1037 1038 1039 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1040 assert(fr.can_be_deoptimized(), "checking frame type"); 1041 1042 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1043 1044 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1045 1046 // Patch the nmethod so that when execution returns to it we will 1047 // deopt the execution state and return to the interpreter. 1048 fr.deoptimize(thread); 1049 } 1050 1051 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1052 // Deoptimize only if the frame comes from compile code. 1053 // Do not deoptimize the frame which is already patched 1054 // during the execution of the loops below. 1055 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1056 return; 1057 } 1058 ResourceMark rm; 1059 DeoptimizationMarker dm; 1060 if (UseBiasedLocking) { 1061 revoke_biases_of_monitors(thread, fr, map); 1062 } 1063 deoptimize_single_frame(thread, fr); 1064 1065 } 1066 1067 1068 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1069 // Compute frame and register map based on thread and sp. 1070 RegisterMap reg_map(thread, UseBiasedLocking); 1071 frame fr = thread->last_frame(); 1072 while (fr.id() != id) { 1073 fr = fr.sender(®_map); 1074 } 1075 deoptimize(thread, fr, ®_map); 1076 } 1077 1078 1079 // JVMTI PopFrame support 1080 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1081 { 1082 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1083 } 1084 JRT_END 1085 1086 1087 #if defined(COMPILER2) || defined(SHARK) 1088 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1089 // in case of an unresolved klass entry, load the class. 1090 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1091 klassOop tk = constant_pool->klass_at(index, CHECK); 1092 return; 1093 } 1094 1095 if (!constant_pool->tag_at(index).is_symbol()) return; 1096 1097 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1098 symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); 1099 1100 // class name? 1101 if (symbol->byte_at(0) != '(') { 1102 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1103 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1104 return; 1105 } 1106 1107 // then it must be a signature! 1108 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1109 if (ss.is_object()) { 1110 symbolOop s = ss.as_symbol(CHECK); 1111 symbolHandle class_name (THREAD, s); 1112 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1113 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1114 } 1115 } 1116 } 1117 1118 1119 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1120 EXCEPTION_MARK; 1121 load_class_by_index(constant_pool, index, THREAD); 1122 if (HAS_PENDING_EXCEPTION) { 1123 // Exception happened during classloading. We ignore the exception here, since it 1124 // is going to be rethrown since the current activation is going to be deoptimzied and 1125 // the interpreter will re-execute the bytecode. 1126 CLEAR_PENDING_EXCEPTION; 1127 } 1128 } 1129 1130 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1131 HandleMark hm; 1132 1133 // uncommon_trap() is called at the beginning of the uncommon trap 1134 // handler. Note this fact before we start generating temporary frames 1135 // that can confuse an asynchronous stack walker. This counter is 1136 // decremented at the end of unpack_frames(). 1137 thread->inc_in_deopt_handler(); 1138 1139 // We need to update the map if we have biased locking. 1140 RegisterMap reg_map(thread, UseBiasedLocking); 1141 frame stub_frame = thread->last_frame(); 1142 frame fr = stub_frame.sender(®_map); 1143 // Make sure the calling nmethod is not getting deoptimized and removed 1144 // before we are done with it. 1145 nmethodLocker nl(fr.pc()); 1146 1147 { 1148 ResourceMark rm; 1149 1150 // Revoke biases of any monitors in the frame to ensure we can migrate them 1151 revoke_biases_of_monitors(thread, fr, ®_map); 1152 1153 DeoptReason reason = trap_request_reason(trap_request); 1154 DeoptAction action = trap_request_action(trap_request); 1155 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1156 1157 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1158 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1159 compiledVFrame* cvf = compiledVFrame::cast(vf); 1160 1161 nmethod* nm = cvf->code(); 1162 1163 ScopeDesc* trap_scope = cvf->scope(); 1164 methodHandle trap_method = trap_scope->method(); 1165 int trap_bci = trap_scope->bci(); 1166 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code(); 1167 1168 // Record this event in the histogram. 1169 gather_statistics(reason, action, trap_bc); 1170 1171 // Ensure that we can record deopt. history: 1172 bool create_if_missing = ProfileTraps; 1173 1174 methodDataHandle trap_mdo 1175 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1176 1177 // Print a bunch of diagnostics, if requested. 1178 if (TraceDeoptimization || LogCompilation) { 1179 ResourceMark rm; 1180 ttyLocker ttyl; 1181 char buf[100]; 1182 if (xtty != NULL) { 1183 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1184 os::current_thread_id(), 1185 format_trap_request(buf, sizeof(buf), trap_request)); 1186 nm->log_identity(xtty); 1187 } 1188 symbolHandle class_name; 1189 bool unresolved = false; 1190 if (unloaded_class_index >= 0) { 1191 constantPoolHandle constants (THREAD, trap_method->constants()); 1192 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1193 class_name = symbolHandle(THREAD, 1194 constants->klass_name_at(unloaded_class_index)); 1195 unresolved = true; 1196 if (xtty != NULL) 1197 xtty->print(" unresolved='1'"); 1198 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1199 class_name = symbolHandle(THREAD, 1200 constants->symbol_at(unloaded_class_index)); 1201 } 1202 if (xtty != NULL) 1203 xtty->name(class_name); 1204 } 1205 if (xtty != NULL && trap_mdo.not_null()) { 1206 // Dump the relevant MDO state. 1207 // This is the deopt count for the current reason, any previous 1208 // reasons or recompiles seen at this point. 1209 int dcnt = trap_mdo->trap_count(reason); 1210 if (dcnt != 0) 1211 xtty->print(" count='%d'", dcnt); 1212 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1213 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1214 if (dos != 0) { 1215 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1216 if (trap_state_is_recompiled(dos)) { 1217 int recnt2 = trap_mdo->overflow_recompile_count(); 1218 if (recnt2 != 0) 1219 xtty->print(" recompiles2='%d'", recnt2); 1220 } 1221 } 1222 } 1223 if (xtty != NULL) { 1224 xtty->stamp(); 1225 xtty->end_head(); 1226 } 1227 if (TraceDeoptimization) { // make noise on the tty 1228 tty->print("Uncommon trap occurred in"); 1229 nm->method()->print_short_name(tty); 1230 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1231 fr.pc(), 1232 (int) os::current_thread_id(), 1233 trap_reason_name(reason), 1234 trap_action_name(action), 1235 unloaded_class_index); 1236 if (class_name.not_null()) { 1237 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1238 class_name->print_symbol_on(tty); 1239 } 1240 tty->cr(); 1241 } 1242 if (xtty != NULL) { 1243 // Log the precise location of the trap. 1244 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1245 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1246 xtty->method(sd->method()); 1247 xtty->end_elem(); 1248 if (sd->is_top()) break; 1249 } 1250 xtty->tail("uncommon_trap"); 1251 } 1252 } 1253 // (End diagnostic printout.) 1254 1255 // Load class if necessary 1256 if (unloaded_class_index >= 0) { 1257 constantPoolHandle constants(THREAD, trap_method->constants()); 1258 load_class_by_index(constants, unloaded_class_index); 1259 } 1260 1261 // Flush the nmethod if necessary and desirable. 1262 // 1263 // We need to avoid situations where we are re-flushing the nmethod 1264 // because of a hot deoptimization site. Repeated flushes at the same 1265 // point need to be detected by the compiler and avoided. If the compiler 1266 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1267 // module must take measures to avoid an infinite cycle of recompilation 1268 // and deoptimization. There are several such measures: 1269 // 1270 // 1. If a recompilation is ordered a second time at some site X 1271 // and for the same reason R, the action is adjusted to 'reinterpret', 1272 // to give the interpreter time to exercise the method more thoroughly. 1273 // If this happens, the method's overflow_recompile_count is incremented. 1274 // 1275 // 2. If the compiler fails to reduce the deoptimization rate, then 1276 // the method's overflow_recompile_count will begin to exceed the set 1277 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1278 // is adjusted to 'make_not_compilable', and the method is abandoned 1279 // to the interpreter. This is a performance hit for hot methods, 1280 // but is better than a disastrous infinite cycle of recompilations. 1281 // (Actually, only the method containing the site X is abandoned.) 1282 // 1283 // 3. In parallel with the previous measures, if the total number of 1284 // recompilations of a method exceeds the much larger set limit 1285 // PerMethodRecompilationCutoff, the method is abandoned. 1286 // This should only happen if the method is very large and has 1287 // many "lukewarm" deoptimizations. The code which enforces this 1288 // limit is elsewhere (class nmethod, class methodOopDesc). 1289 // 1290 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1291 // to recompile at each bytecode independently of the per-BCI cutoff. 1292 // 1293 // The decision to update code is up to the compiler, and is encoded 1294 // in the Action_xxx code. If the compiler requests Action_none 1295 // no trap state is changed, no compiled code is changed, and the 1296 // computation suffers along in the interpreter. 1297 // 1298 // The other action codes specify various tactics for decompilation 1299 // and recompilation. Action_maybe_recompile is the loosest, and 1300 // allows the compiled code to stay around until enough traps are seen, 1301 // and until the compiler gets around to recompiling the trapping method. 1302 // 1303 // The other actions cause immediate removal of the present code. 1304 1305 bool update_trap_state = true; 1306 bool make_not_entrant = false; 1307 bool make_not_compilable = false; 1308 bool reprofile = false; 1309 switch (action) { 1310 case Action_none: 1311 // Keep the old code. 1312 update_trap_state = false; 1313 break; 1314 case Action_maybe_recompile: 1315 // Do not need to invalidate the present code, but we can 1316 // initiate another 1317 // Start compiler without (necessarily) invalidating the nmethod. 1318 // The system will tolerate the old code, but new code should be 1319 // generated when possible. 1320 break; 1321 case Action_reinterpret: 1322 // Go back into the interpreter for a while, and then consider 1323 // recompiling form scratch. 1324 make_not_entrant = true; 1325 // Reset invocation counter for outer most method. 1326 // This will allow the interpreter to exercise the bytecodes 1327 // for a while before recompiling. 1328 // By contrast, Action_make_not_entrant is immediate. 1329 // 1330 // Note that the compiler will track null_check, null_assert, 1331 // range_check, and class_check events and log them as if they 1332 // had been traps taken from compiled code. This will update 1333 // the MDO trap history so that the next compilation will 1334 // properly detect hot trap sites. 1335 reprofile = true; 1336 break; 1337 case Action_make_not_entrant: 1338 // Request immediate recompilation, and get rid of the old code. 1339 // Make them not entrant, so next time they are called they get 1340 // recompiled. Unloaded classes are loaded now so recompile before next 1341 // time they are called. Same for uninitialized. The interpreter will 1342 // link the missing class, if any. 1343 make_not_entrant = true; 1344 break; 1345 case Action_make_not_compilable: 1346 // Give up on compiling this method at all. 1347 make_not_entrant = true; 1348 make_not_compilable = true; 1349 break; 1350 default: 1351 ShouldNotReachHere(); 1352 } 1353 1354 // Setting +ProfileTraps fixes the following, on all platforms: 1355 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1356 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1357 // recompile relies on a methodDataOop to record heroic opt failures. 1358 1359 // Whether the interpreter is producing MDO data or not, we also need 1360 // to use the MDO to detect hot deoptimization points and control 1361 // aggressive optimization. 1362 bool inc_recompile_count = false; 1363 ProfileData* pdata = NULL; 1364 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1365 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1366 uint this_trap_count = 0; 1367 bool maybe_prior_trap = false; 1368 bool maybe_prior_recompile = false; 1369 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1370 //outputs: 1371 this_trap_count, 1372 maybe_prior_trap, 1373 maybe_prior_recompile); 1374 // Because the interpreter also counts null, div0, range, and class 1375 // checks, these traps from compiled code are double-counted. 1376 // This is harmless; it just means that the PerXTrapLimit values 1377 // are in effect a little smaller than they look. 1378 1379 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1380 if (per_bc_reason != Reason_none) { 1381 // Now take action based on the partially known per-BCI history. 1382 if (maybe_prior_trap 1383 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1384 // If there are too many traps at this BCI, force a recompile. 1385 // This will allow the compiler to see the limit overflow, and 1386 // take corrective action, if possible. The compiler generally 1387 // does not use the exact PerBytecodeTrapLimit value, but instead 1388 // changes its tactics if it sees any traps at all. This provides 1389 // a little hysteresis, delaying a recompile until a trap happens 1390 // several times. 1391 // 1392 // Actually, since there is only one bit of counter per BCI, 1393 // the possible per-BCI counts are {0,1,(per-method count)}. 1394 // This produces accurate results if in fact there is only 1395 // one hot trap site, but begins to get fuzzy if there are 1396 // many sites. For example, if there are ten sites each 1397 // trapping two or more times, they each get the blame for 1398 // all of their traps. 1399 make_not_entrant = true; 1400 } 1401 1402 // Detect repeated recompilation at the same BCI, and enforce a limit. 1403 if (make_not_entrant && maybe_prior_recompile) { 1404 // More than one recompile at this point. 1405 inc_recompile_count = maybe_prior_trap; 1406 } 1407 } else { 1408 // For reasons which are not recorded per-bytecode, we simply 1409 // force recompiles unconditionally. 1410 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1411 make_not_entrant = true; 1412 } 1413 1414 // Go back to the compiler if there are too many traps in this method. 1415 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1416 // If there are too many traps in this method, force a recompile. 1417 // This will allow the compiler to see the limit overflow, and 1418 // take corrective action, if possible. 1419 // (This condition is an unlikely backstop only, because the 1420 // PerBytecodeTrapLimit is more likely to take effect first, 1421 // if it is applicable.) 1422 make_not_entrant = true; 1423 } 1424 1425 // Here's more hysteresis: If there has been a recompile at 1426 // this trap point already, run the method in the interpreter 1427 // for a while to exercise it more thoroughly. 1428 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1429 reprofile = true; 1430 } 1431 1432 } 1433 1434 // Take requested actions on the method: 1435 1436 // Recompile 1437 if (make_not_entrant) { 1438 if (!nm->make_not_entrant()) { 1439 return; // the call did not change nmethod's state 1440 } 1441 1442 if (pdata != NULL) { 1443 // Record the recompilation event, if any. 1444 int tstate0 = pdata->trap_state(); 1445 int tstate1 = trap_state_set_recompiled(tstate0, true); 1446 if (tstate1 != tstate0) 1447 pdata->set_trap_state(tstate1); 1448 } 1449 } 1450 1451 if (inc_recompile_count) { 1452 trap_mdo->inc_overflow_recompile_count(); 1453 if ((uint)trap_mdo->overflow_recompile_count() > 1454 (uint)PerBytecodeRecompilationCutoff) { 1455 // Give up on the method containing the bad BCI. 1456 if (trap_method() == nm->method()) { 1457 make_not_compilable = true; 1458 } else { 1459 trap_method->set_not_compilable(CompLevel_full_optimization); 1460 // But give grace to the enclosing nm->method(). 1461 } 1462 } 1463 } 1464 1465 // Reprofile 1466 if (reprofile) { 1467 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1468 } 1469 1470 // Give up compiling 1471 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1472 assert(make_not_entrant, "consistent"); 1473 nm->method()->set_not_compilable(CompLevel_full_optimization); 1474 } 1475 1476 } // Free marked resources 1477 1478 } 1479 JRT_END 1480 1481 methodDataOop 1482 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1483 bool create_if_missing) { 1484 Thread* THREAD = thread; 1485 methodDataOop mdo = m()->method_data(); 1486 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1487 // Build an MDO. Ignore errors like OutOfMemory; 1488 // that simply means we won't have an MDO to update. 1489 methodOopDesc::build_interpreter_method_data(m, THREAD); 1490 if (HAS_PENDING_EXCEPTION) { 1491 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1492 CLEAR_PENDING_EXCEPTION; 1493 } 1494 mdo = m()->method_data(); 1495 } 1496 return mdo; 1497 } 1498 1499 ProfileData* 1500 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1501 int trap_bci, 1502 Deoptimization::DeoptReason reason, 1503 //outputs: 1504 uint& ret_this_trap_count, 1505 bool& ret_maybe_prior_trap, 1506 bool& ret_maybe_prior_recompile) { 1507 uint prior_trap_count = trap_mdo->trap_count(reason); 1508 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1509 1510 // If the runtime cannot find a place to store trap history, 1511 // it is estimated based on the general condition of the method. 1512 // If the method has ever been recompiled, or has ever incurred 1513 // a trap with the present reason , then this BCI is assumed 1514 // (pessimistically) to be the culprit. 1515 bool maybe_prior_trap = (prior_trap_count != 0); 1516 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1517 ProfileData* pdata = NULL; 1518 1519 1520 // For reasons which are recorded per bytecode, we check per-BCI data. 1521 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1522 if (per_bc_reason != Reason_none) { 1523 // Find the profile data for this BCI. If there isn't one, 1524 // try to allocate one from the MDO's set of spares. 1525 // This will let us detect a repeated trap at this point. 1526 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1527 1528 if (pdata != NULL) { 1529 // Query the trap state of this profile datum. 1530 int tstate0 = pdata->trap_state(); 1531 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1532 maybe_prior_trap = false; 1533 if (!trap_state_is_recompiled(tstate0)) 1534 maybe_prior_recompile = false; 1535 1536 // Update the trap state of this profile datum. 1537 int tstate1 = tstate0; 1538 // Record the reason. 1539 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1540 // Store the updated state on the MDO, for next time. 1541 if (tstate1 != tstate0) 1542 pdata->set_trap_state(tstate1); 1543 } else { 1544 if (LogCompilation && xtty != NULL) { 1545 ttyLocker ttyl; 1546 // Missing MDP? Leave a small complaint in the log. 1547 xtty->elem("missing_mdp bci='%d'", trap_bci); 1548 } 1549 } 1550 } 1551 1552 // Return results: 1553 ret_this_trap_count = this_trap_count; 1554 ret_maybe_prior_trap = maybe_prior_trap; 1555 ret_maybe_prior_recompile = maybe_prior_recompile; 1556 return pdata; 1557 } 1558 1559 void 1560 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1561 ResourceMark rm; 1562 // Ignored outputs: 1563 uint ignore_this_trap_count; 1564 bool ignore_maybe_prior_trap; 1565 bool ignore_maybe_prior_recompile; 1566 query_update_method_data(trap_mdo, trap_bci, 1567 (DeoptReason)reason, 1568 ignore_this_trap_count, 1569 ignore_maybe_prior_trap, 1570 ignore_maybe_prior_recompile); 1571 } 1572 1573 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1574 1575 // Still in Java no safepoints 1576 { 1577 // This enters VM and may safepoint 1578 uncommon_trap_inner(thread, trap_request); 1579 } 1580 return fetch_unroll_info_helper(thread); 1581 } 1582 1583 // Local derived constants. 1584 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1585 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1586 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1587 1588 //---------------------------trap_state_reason--------------------------------- 1589 Deoptimization::DeoptReason 1590 Deoptimization::trap_state_reason(int trap_state) { 1591 // This assert provides the link between the width of DataLayout::trap_bits 1592 // and the encoding of "recorded" reasons. It ensures there are enough 1593 // bits to store all needed reasons in the per-BCI MDO profile. 1594 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1595 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1596 trap_state -= recompile_bit; 1597 if (trap_state == DS_REASON_MASK) { 1598 return Reason_many; 1599 } else { 1600 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1601 return (DeoptReason)trap_state; 1602 } 1603 } 1604 //-------------------------trap_state_has_reason------------------------------- 1605 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1606 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1607 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1608 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1609 trap_state -= recompile_bit; 1610 if (trap_state == DS_REASON_MASK) { 1611 return -1; // true, unspecifically (bottom of state lattice) 1612 } else if (trap_state == reason) { 1613 return 1; // true, definitely 1614 } else if (trap_state == 0) { 1615 return 0; // false, definitely (top of state lattice) 1616 } else { 1617 return 0; // false, definitely 1618 } 1619 } 1620 //-------------------------trap_state_add_reason------------------------------- 1621 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1622 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1623 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1624 trap_state -= recompile_bit; 1625 if (trap_state == DS_REASON_MASK) { 1626 return trap_state + recompile_bit; // already at state lattice bottom 1627 } else if (trap_state == reason) { 1628 return trap_state + recompile_bit; // the condition is already true 1629 } else if (trap_state == 0) { 1630 return reason + recompile_bit; // no condition has yet been true 1631 } else { 1632 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1633 } 1634 } 1635 //-----------------------trap_state_is_recompiled------------------------------ 1636 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1637 return (trap_state & DS_RECOMPILE_BIT) != 0; 1638 } 1639 //-----------------------trap_state_set_recompiled----------------------------- 1640 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1641 if (z) return trap_state | DS_RECOMPILE_BIT; 1642 else return trap_state & ~DS_RECOMPILE_BIT; 1643 } 1644 //---------------------------format_trap_state--------------------------------- 1645 // This is used for debugging and diagnostics, including hotspot.log output. 1646 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1647 int trap_state) { 1648 DeoptReason reason = trap_state_reason(trap_state); 1649 bool recomp_flag = trap_state_is_recompiled(trap_state); 1650 // Re-encode the state from its decoded components. 1651 int decoded_state = 0; 1652 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1653 decoded_state = trap_state_add_reason(decoded_state, reason); 1654 if (recomp_flag) 1655 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1656 // If the state re-encodes properly, format it symbolically. 1657 // Because this routine is used for debugging and diagnostics, 1658 // be robust even if the state is a strange value. 1659 size_t len; 1660 if (decoded_state != trap_state) { 1661 // Random buggy state that doesn't decode?? 1662 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1663 } else { 1664 len = jio_snprintf(buf, buflen, "%s%s", 1665 trap_reason_name(reason), 1666 recomp_flag ? " recompiled" : ""); 1667 } 1668 if (len >= buflen) 1669 buf[buflen-1] = '\0'; 1670 return buf; 1671 } 1672 1673 1674 //--------------------------------statics-------------------------------------- 1675 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1676 = Deoptimization::Action_reinterpret; 1677 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1678 // Note: Keep this in sync. with enum DeoptReason. 1679 "none", 1680 "null_check", 1681 "null_assert", 1682 "range_check", 1683 "class_check", 1684 "array_check", 1685 "intrinsic", 1686 "bimorphic", 1687 "unloaded", 1688 "uninitialized", 1689 "unreached", 1690 "unhandled", 1691 "constraint", 1692 "div0_check", 1693 "age", 1694 "predicate" 1695 }; 1696 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1697 // Note: Keep this in sync. with enum DeoptAction. 1698 "none", 1699 "maybe_recompile", 1700 "reinterpret", 1701 "make_not_entrant", 1702 "make_not_compilable" 1703 }; 1704 1705 const char* Deoptimization::trap_reason_name(int reason) { 1706 if (reason == Reason_many) return "many"; 1707 if ((uint)reason < Reason_LIMIT) 1708 return _trap_reason_name[reason]; 1709 static char buf[20]; 1710 sprintf(buf, "reason%d", reason); 1711 return buf; 1712 } 1713 const char* Deoptimization::trap_action_name(int action) { 1714 if ((uint)action < Action_LIMIT) 1715 return _trap_action_name[action]; 1716 static char buf[20]; 1717 sprintf(buf, "action%d", action); 1718 return buf; 1719 } 1720 1721 // This is used for debugging and diagnostics, including hotspot.log output. 1722 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1723 int trap_request) { 1724 jint unloaded_class_index = trap_request_index(trap_request); 1725 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1726 const char* action = trap_action_name(trap_request_action(trap_request)); 1727 size_t len; 1728 if (unloaded_class_index < 0) { 1729 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1730 reason, action); 1731 } else { 1732 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1733 reason, action, unloaded_class_index); 1734 } 1735 if (len >= buflen) 1736 buf[buflen-1] = '\0'; 1737 return buf; 1738 } 1739 1740 juint Deoptimization::_deoptimization_hist 1741 [Deoptimization::Reason_LIMIT] 1742 [1 + Deoptimization::Action_LIMIT] 1743 [Deoptimization::BC_CASE_LIMIT] 1744 = {0}; 1745 1746 enum { 1747 LSB_BITS = 8, 1748 LSB_MASK = right_n_bits(LSB_BITS) 1749 }; 1750 1751 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1752 Bytecodes::Code bc) { 1753 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1754 assert(action >= 0 && action < Action_LIMIT, "oob"); 1755 _deoptimization_hist[Reason_none][0][0] += 1; // total 1756 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1757 juint* cases = _deoptimization_hist[reason][1+action]; 1758 juint* bc_counter_addr = NULL; 1759 juint bc_counter = 0; 1760 // Look for an unused counter, or an exact match to this BC. 1761 if (bc != Bytecodes::_illegal) { 1762 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1763 juint* counter_addr = &cases[bc_case]; 1764 juint counter = *counter_addr; 1765 if ((counter == 0 && bc_counter_addr == NULL) 1766 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1767 // this counter is either free or is already devoted to this BC 1768 bc_counter_addr = counter_addr; 1769 bc_counter = counter | bc; 1770 } 1771 } 1772 } 1773 if (bc_counter_addr == NULL) { 1774 // Overflow, or no given bytecode. 1775 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1776 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1777 } 1778 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1779 } 1780 1781 jint Deoptimization::total_deoptimization_count() { 1782 return _deoptimization_hist[Reason_none][0][0]; 1783 } 1784 1785 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1786 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1787 return _deoptimization_hist[reason][0][0]; 1788 } 1789 1790 void Deoptimization::print_statistics() { 1791 juint total = total_deoptimization_count(); 1792 juint account = total; 1793 if (total != 0) { 1794 ttyLocker ttyl; 1795 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1796 tty->print_cr("Deoptimization traps recorded:"); 1797 #define PRINT_STAT_LINE(name, r) \ 1798 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1799 PRINT_STAT_LINE("total", total); 1800 // For each non-zero entry in the histogram, print the reason, 1801 // the action, and (if specifically known) the type of bytecode. 1802 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1803 for (int action = 0; action < Action_LIMIT; action++) { 1804 juint* cases = _deoptimization_hist[reason][1+action]; 1805 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1806 juint counter = cases[bc_case]; 1807 if (counter != 0) { 1808 char name[1*K]; 1809 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1810 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1811 bc = Bytecodes::_illegal; 1812 sprintf(name, "%s/%s/%s", 1813 trap_reason_name(reason), 1814 trap_action_name(action), 1815 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1816 juint r = counter >> LSB_BITS; 1817 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1818 account -= r; 1819 } 1820 } 1821 } 1822 } 1823 if (account != 0) { 1824 PRINT_STAT_LINE("unaccounted", account); 1825 } 1826 #undef PRINT_STAT_LINE 1827 if (xtty != NULL) xtty->tail("statistics"); 1828 } 1829 } 1830 #else // COMPILER2 || SHARK 1831 1832 1833 // Stubs for C1 only system. 1834 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1835 return false; 1836 } 1837 1838 const char* Deoptimization::trap_reason_name(int reason) { 1839 return "unknown"; 1840 } 1841 1842 void Deoptimization::print_statistics() { 1843 // no output 1844 } 1845 1846 void 1847 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1848 // no udpate 1849 } 1850 1851 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1852 return 0; 1853 } 1854 1855 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1856 Bytecodes::Code bc) { 1857 // no update 1858 } 1859 1860 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1861 int trap_state) { 1862 jio_snprintf(buf, buflen, "#%d", trap_state); 1863 return buf; 1864 } 1865 1866 #endif // COMPILER2 || SHARK