deoptimization.cpp revision 2062:3582bf76420e
1/* 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "precompiled.hpp" 26#include "classfile/systemDictionary.hpp" 27#include "code/debugInfoRec.hpp" 28#include "code/nmethod.hpp" 29#include "code/pcDesc.hpp" 30#include "code/scopeDesc.hpp" 31#include "interpreter/bytecode.hpp" 32#include "interpreter/interpreter.hpp" 33#include "interpreter/oopMapCache.hpp" 34#include "memory/allocation.inline.hpp" 35#include "memory/oopFactory.hpp" 36#include "memory/resourceArea.hpp" 37#include "oops/methodOop.hpp" 38#include "oops/oop.inline.hpp" 39#include "prims/jvmtiThreadState.hpp" 40#include "runtime/biasedLocking.hpp" 41#include "runtime/compilationPolicy.hpp" 42#include "runtime/deoptimization.hpp" 43#include "runtime/interfaceSupport.hpp" 44#include "runtime/sharedRuntime.hpp" 45#include "runtime/signature.hpp" 46#include "runtime/stubRoutines.hpp" 47#include "runtime/thread.hpp" 48#include "runtime/vframe.hpp" 49#include "runtime/vframeArray.hpp" 50#include "runtime/vframe_hp.hpp" 51#include "utilities/events.hpp" 52#include "utilities/xmlstream.hpp" 53#ifdef TARGET_ARCH_x86 54# include "vmreg_x86.inline.hpp" 55#endif 56#ifdef TARGET_ARCH_sparc 57# include "vmreg_sparc.inline.hpp" 58#endif 59#ifdef TARGET_ARCH_zero 60# include "vmreg_zero.inline.hpp" 61#endif 62#ifdef COMPILER2 63#ifdef TARGET_ARCH_MODEL_x86_32 64# include "adfiles/ad_x86_32.hpp" 65#endif 66#ifdef TARGET_ARCH_MODEL_x86_64 67# include "adfiles/ad_x86_64.hpp" 68#endif 69#ifdef TARGET_ARCH_MODEL_sparc 70# include "adfiles/ad_sparc.hpp" 71#endif 72#ifdef TARGET_ARCH_MODEL_zero 73# include "adfiles/ad_zero.hpp" 74#endif 75#endif 76 77bool DeoptimizationMarker::_is_active = false; 78 79Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 80 int caller_adjustment, 81 int number_of_frames, 82 intptr_t* frame_sizes, 83 address* frame_pcs, 84 BasicType return_type) { 85 _size_of_deoptimized_frame = size_of_deoptimized_frame; 86 _caller_adjustment = caller_adjustment; 87 _number_of_frames = number_of_frames; 88 _frame_sizes = frame_sizes; 89 _frame_pcs = frame_pcs; 90 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); 91 _return_type = return_type; 92 // PD (x86 only) 93 _counter_temp = 0; 94 _initial_fp = 0; 95 _unpack_kind = 0; 96 _sender_sp_temp = 0; 97 98 _total_frame_sizes = size_of_frames(); 99} 100 101 102Deoptimization::UnrollBlock::~UnrollBlock() { 103 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 104 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 105 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 106} 107 108 109intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 110 assert(register_number < RegisterMap::reg_count, "checking register number"); 111 return &_register_block[register_number * 2]; 112} 113 114 115 116int Deoptimization::UnrollBlock::size_of_frames() const { 117 // Acount first for the adjustment of the initial frame 118 int result = _caller_adjustment; 119 for (int index = 0; index < number_of_frames(); index++) { 120 result += frame_sizes()[index]; 121 } 122 return result; 123} 124 125 126void Deoptimization::UnrollBlock::print() { 127 ttyLocker ttyl; 128 tty->print_cr("UnrollBlock"); 129 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 130 tty->print( " frame_sizes: "); 131 for (int index = 0; index < number_of_frames(); index++) { 132 tty->print("%d ", frame_sizes()[index]); 133 } 134 tty->cr(); 135} 136 137 138// In order to make fetch_unroll_info work properly with escape 139// analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 140// ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 141// of previously eliminated objects occurs in realloc_objects, which is 142// called from the method fetch_unroll_info_helper below. 143JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 144 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 145 // but makes the entry a little slower. There is however a little dance we have to 146 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 147 148 // fetch_unroll_info() is called at the beginning of the deoptimization 149 // handler. Note this fact before we start generating temporary frames 150 // that can confuse an asynchronous stack walker. This counter is 151 // decremented at the end of unpack_frames(). 152 thread->inc_in_deopt_handler(); 153 154 return fetch_unroll_info_helper(thread); 155JRT_END 156 157 158// This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 159Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 160 161 // Note: there is a safepoint safety issue here. No matter whether we enter 162 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 163 // the vframeArray is created. 164 // 165 166 // Allocate our special deoptimization ResourceMark 167 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 168 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 169 thread->set_deopt_mark(dmark); 170 171 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 172 RegisterMap map(thread, true); 173 RegisterMap dummy_map(thread, false); 174 // Now get the deoptee with a valid map 175 frame deoptee = stub_frame.sender(&map); 176 // Set the deoptee nmethod 177 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 178 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 179 180 // Create a growable array of VFrames where each VFrame represents an inlined 181 // Java frame. This storage is allocated with the usual system arena. 182 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 183 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 184 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 185 while (!vf->is_top()) { 186 assert(vf->is_compiled_frame(), "Wrong frame type"); 187 chunk->push(compiledVFrame::cast(vf)); 188 vf = vf->sender(); 189 } 190 assert(vf->is_compiled_frame(), "Wrong frame type"); 191 chunk->push(compiledVFrame::cast(vf)); 192 193#ifdef COMPILER2 194 // Reallocate the non-escaping objects and restore their fields. Then 195 // relock objects if synchronization on them was eliminated. 196 if (DoEscapeAnalysis) { 197 if (EliminateAllocations) { 198 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 199 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 200 201 // The flag return_oop() indicates call sites which return oop 202 // in compiled code. Such sites include java method calls, 203 // runtime calls (for example, used to allocate new objects/arrays 204 // on slow code path) and any other calls generated in compiled code. 205 // It is not guaranteed that we can get such information here only 206 // by analyzing bytecode in deoptimized frames. This is why this flag 207 // is set during method compilation (see Compile::Process_OopMap_Node()). 208 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 209 Handle return_value; 210 if (save_oop_result) { 211 // Reallocation may trigger GC. If deoptimization happened on return from 212 // call which returns oop we need to save it since it is not in oopmap. 213 oop result = deoptee.saved_oop_result(&map); 214 assert(result == NULL || result->is_oop(), "must be oop"); 215 return_value = Handle(thread, result); 216 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 217 if (TraceDeoptimization) { 218 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 219 } 220 } 221 bool reallocated = false; 222 if (objects != NULL) { 223 JRT_BLOCK 224 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 225 JRT_END 226 } 227 if (reallocated) { 228 reassign_fields(&deoptee, &map, objects); 229#ifndef PRODUCT 230 if (TraceDeoptimization) { 231 ttyLocker ttyl; 232 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 233 print_objects(objects); 234 } 235#endif 236 } 237 if (save_oop_result) { 238 // Restore result. 239 deoptee.set_saved_oop_result(&map, return_value()); 240 } 241 } 242 if (EliminateLocks) { 243#ifndef PRODUCT 244 bool first = true; 245#endif 246 for (int i = 0; i < chunk->length(); i++) { 247 compiledVFrame* cvf = chunk->at(i); 248 assert (cvf->scope() != NULL,"expect only compiled java frames"); 249 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 250 if (monitors->is_nonempty()) { 251 relock_objects(monitors, thread); 252#ifndef PRODUCT 253 if (TraceDeoptimization) { 254 ttyLocker ttyl; 255 for (int j = 0; j < monitors->length(); j++) { 256 MonitorInfo* mi = monitors->at(j); 257 if (mi->eliminated()) { 258 if (first) { 259 first = false; 260 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 261 } 262 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 263 } 264 } 265 } 266#endif 267 } 268 } 269 } 270 } 271#endif // COMPILER2 272 // Ensure that no safepoint is taken after pointers have been stored 273 // in fields of rematerialized objects. If a safepoint occurs from here on 274 // out the java state residing in the vframeArray will be missed. 275 No_Safepoint_Verifier no_safepoint; 276 277 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 278 279 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 280 thread->set_vframe_array_head(array); 281 282 // Now that the vframeArray has been created if we have any deferred local writes 283 // added by jvmti then we can free up that structure as the data is now in the 284 // vframeArray 285 286 if (thread->deferred_locals() != NULL) { 287 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 288 int i = 0; 289 do { 290 // Because of inlining we could have multiple vframes for a single frame 291 // and several of the vframes could have deferred writes. Find them all. 292 if (list->at(i)->id() == array->original().id()) { 293 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 294 list->remove_at(i); 295 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 296 delete dlv; 297 } else { 298 i++; 299 } 300 } while ( i < list->length() ); 301 if (list->length() == 0) { 302 thread->set_deferred_locals(NULL); 303 // free the list and elements back to C heap. 304 delete list; 305 } 306 307 } 308 309#ifndef SHARK 310 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 311 CodeBlob* cb = stub_frame.cb(); 312 // Verify we have the right vframeArray 313 assert(cb->frame_size() >= 0, "Unexpected frame size"); 314 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 315 316 // If the deopt call site is a MethodHandle invoke call site we have 317 // to adjust the unpack_sp. 318 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 319 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 320 unpack_sp = deoptee.unextended_sp(); 321 322#ifdef ASSERT 323 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 324 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 325#endif 326#else 327 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 328#endif // !SHARK 329 330 // This is a guarantee instead of an assert because if vframe doesn't match 331 // we will unpack the wrong deoptimized frame and wind up in strange places 332 // where it will be very difficult to figure out what went wrong. Better 333 // to die an early death here than some very obscure death later when the 334 // trail is cold. 335 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 336 // in that it will fail to detect a problem when there is one. This needs 337 // more work in tiger timeframe. 338 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 339 340 int number_of_frames = array->frames(); 341 342 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 343 // virtual activation, which is the reverse of the elements in the vframes array. 344 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 345 // +1 because we always have an interpreter return address for the final slot. 346 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 347 int callee_parameters = 0; 348 int callee_locals = 0; 349 int popframe_extra_args = 0; 350 // Create an interpreter return address for the stub to use as its return 351 // address so the skeletal frames are perfectly walkable 352 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 353 354 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 355 // activation be put back on the expression stack of the caller for reexecution 356 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 357 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 358 } 359 360 // 361 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 362 // frame_sizes/frame_pcs[1] next oldest frame (int) 363 // frame_sizes/frame_pcs[n] youngest frame (int) 364 // 365 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 366 // owns the space for the return address to it's caller). Confusing ain't it. 367 // 368 // The vframe array can address vframes with indices running from 369 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 370 // When we create the skeletal frames we need the oldest frame to be in the zero slot 371 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 372 // so things look a little strange in this loop. 373 // 374 for (int index = 0; index < array->frames(); index++ ) { 375 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 376 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 377 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 378 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 379 callee_locals, 380 index == 0, 381 popframe_extra_args); 382 // This pc doesn't have to be perfect just good enough to identify the frame 383 // as interpreted so the skeleton frame will be walkable 384 // The correct pc will be set when the skeleton frame is completely filled out 385 // The final pc we store in the loop is wrong and will be overwritten below 386 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 387 388 callee_parameters = array->element(index)->method()->size_of_parameters(); 389 callee_locals = array->element(index)->method()->max_locals(); 390 popframe_extra_args = 0; 391 } 392 393 // Compute whether the root vframe returns a float or double value. 394 BasicType return_type; 395 { 396 HandleMark hm; 397 methodHandle method(thread, array->element(0)->method()); 398 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 399 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 400 } 401 402 // Compute information for handling adapters and adjusting the frame size of the caller. 403 int caller_adjustment = 0; 404 405 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 406 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 407 // than simply use array->sender.pc(). This requires us to walk the current set of frames 408 // 409 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 410 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 411 412 // Compute the amount the oldest interpreter frame will have to adjust 413 // its caller's stack by. If the caller is a compiled frame then 414 // we pretend that the callee has no parameters so that the 415 // extension counts for the full amount of locals and not just 416 // locals-parms. This is because without a c2i adapter the parm 417 // area as created by the compiled frame will not be usable by 418 // the interpreter. (Depending on the calling convention there 419 // may not even be enough space). 420 421 // QQQ I'd rather see this pushed down into last_frame_adjust 422 // and have it take the sender (aka caller). 423 424 if (deopt_sender.is_compiled_frame()) { 425 caller_adjustment = last_frame_adjust(0, callee_locals); 426 } else if (callee_locals > callee_parameters) { 427 // The caller frame may need extending to accommodate 428 // non-parameter locals of the first unpacked interpreted frame. 429 // Compute that adjustment. 430 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 431 } 432 433 434 // If the sender is deoptimized the we must retrieve the address of the handler 435 // since the frame will "magically" show the original pc before the deopt 436 // and we'd undo the deopt. 437 438 frame_pcs[0] = deopt_sender.raw_pc(); 439 440#ifndef SHARK 441 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 442#endif // SHARK 443 444 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 445 caller_adjustment * BytesPerWord, 446 number_of_frames, 447 frame_sizes, 448 frame_pcs, 449 return_type); 450#if defined(IA32) || defined(AMD64) 451 // We need a way to pass fp to the unpacking code so the skeletal frames 452 // come out correct. This is only needed for x86 because of c2 using ebp 453 // as an allocatable register. So this update is useless (and harmless) 454 // on the other platforms. It would be nice to do this in a different 455 // way but even the old style deoptimization had a problem with deriving 456 // this value. NEEDS_CLEANUP 457 // Note: now that c1 is using c2's deopt blob we must do this on all 458 // x86 based platforms 459 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); 460 *fp_addr = array->sender().fp(); // was adapter_caller 461#endif /* IA32 || AMD64 */ 462 463 if (array->frames() > 1) { 464 if (VerifyStack && TraceDeoptimization) { 465 tty->print_cr("Deoptimizing method containing inlining"); 466 } 467 } 468 469 array->set_unroll_block(info); 470 return info; 471} 472 473// Called to cleanup deoptimization data structures in normal case 474// after unpacking to stack and when stack overflow error occurs 475void Deoptimization::cleanup_deopt_info(JavaThread *thread, 476 vframeArray *array) { 477 478 // Get array if coming from exception 479 if (array == NULL) { 480 array = thread->vframe_array_head(); 481 } 482 thread->set_vframe_array_head(NULL); 483 484 // Free the previous UnrollBlock 485 vframeArray* old_array = thread->vframe_array_last(); 486 thread->set_vframe_array_last(array); 487 488 if (old_array != NULL) { 489 UnrollBlock* old_info = old_array->unroll_block(); 490 old_array->set_unroll_block(NULL); 491 delete old_info; 492 delete old_array; 493 } 494 495 // Deallocate any resource creating in this routine and any ResourceObjs allocated 496 // inside the vframeArray (StackValueCollections) 497 498 delete thread->deopt_mark(); 499 thread->set_deopt_mark(NULL); 500 thread->set_deopt_nmethod(NULL); 501 502 503 if (JvmtiExport::can_pop_frame()) { 504#ifndef CC_INTERP 505 // Regardless of whether we entered this routine with the pending 506 // popframe condition bit set, we should always clear it now 507 thread->clear_popframe_condition(); 508#else 509 // C++ interpeter will clear has_pending_popframe when it enters 510 // with method_resume. For deopt_resume2 we clear it now. 511 if (thread->popframe_forcing_deopt_reexecution()) 512 thread->clear_popframe_condition(); 513#endif /* CC_INTERP */ 514 } 515 516 // unpack_frames() is called at the end of the deoptimization handler 517 // and (in C2) at the end of the uncommon trap handler. Note this fact 518 // so that an asynchronous stack walker can work again. This counter is 519 // incremented at the beginning of fetch_unroll_info() and (in C2) at 520 // the beginning of uncommon_trap(). 521 thread->dec_in_deopt_handler(); 522} 523 524 525// Return BasicType of value being returned 526JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 527 528 // We are already active int he special DeoptResourceMark any ResourceObj's we 529 // allocate will be freed at the end of the routine. 530 531 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 532 // but makes the entry a little slower. There is however a little dance we have to 533 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 534 ResetNoHandleMark rnhm; // No-op in release/product versions 535 HandleMark hm; 536 537 frame stub_frame = thread->last_frame(); 538 539 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 540 // must point to the vframeArray for the unpack frame. 541 vframeArray* array = thread->vframe_array_head(); 542 543#ifndef PRODUCT 544 if (TraceDeoptimization) { 545 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 546 } 547#endif 548 549 UnrollBlock* info = array->unroll_block(); 550 551 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 552 array->unpack_to_stack(stub_frame, exec_mode); 553 554 BasicType bt = info->return_type(); 555 556 // If we have an exception pending, claim that the return type is an oop 557 // so the deopt_blob does not overwrite the exception_oop. 558 559 if (exec_mode == Unpack_exception) 560 bt = T_OBJECT; 561 562 // Cleanup thread deopt data 563 cleanup_deopt_info(thread, array); 564 565#ifndef PRODUCT 566 if (VerifyStack) { 567 ResourceMark res_mark; 568 569 // Verify that the just-unpacked frames match the interpreter's 570 // notions of expression stack and locals 571 vframeArray* cur_array = thread->vframe_array_last(); 572 RegisterMap rm(thread, false); 573 rm.set_include_argument_oops(false); 574 bool is_top_frame = true; 575 int callee_size_of_parameters = 0; 576 int callee_max_locals = 0; 577 for (int i = 0; i < cur_array->frames(); i++) { 578 vframeArrayElement* el = cur_array->element(i); 579 frame* iframe = el->iframe(); 580 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 581 582 // Get the oop map for this bci 583 InterpreterOopMap mask; 584 int cur_invoke_parameter_size = 0; 585 bool try_next_mask = false; 586 int next_mask_expression_stack_size = -1; 587 int top_frame_expression_stack_adjustment = 0; 588 methodHandle mh(thread, iframe->interpreter_frame_method()); 589 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 590 BytecodeStream str(mh); 591 str.set_start(iframe->interpreter_frame_bci()); 592 int max_bci = mh->code_size(); 593 // Get to the next bytecode if possible 594 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 595 // Check to see if we can grab the number of outgoing arguments 596 // at an uncommon trap for an invoke (where the compiler 597 // generates debug info before the invoke has executed) 598 Bytecodes::Code cur_code = str.next(); 599 if (cur_code == Bytecodes::_invokevirtual || 600 cur_code == Bytecodes::_invokespecial || 601 cur_code == Bytecodes::_invokestatic || 602 cur_code == Bytecodes::_invokeinterface) { 603 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 604 Symbol* signature = invoke.signature(); 605 ArgumentSizeComputer asc(signature); 606 cur_invoke_parameter_size = asc.size(); 607 if (cur_code != Bytecodes::_invokestatic) { 608 // Add in receiver 609 ++cur_invoke_parameter_size; 610 } 611 } 612 if (str.bci() < max_bci) { 613 Bytecodes::Code bc = str.next(); 614 if (bc >= 0) { 615 // The interpreter oop map generator reports results before 616 // the current bytecode has executed except in the case of 617 // calls. It seems to be hard to tell whether the compiler 618 // has emitted debug information matching the "state before" 619 // a given bytecode or the state after, so we try both 620 switch (cur_code) { 621 case Bytecodes::_invokevirtual: 622 case Bytecodes::_invokespecial: 623 case Bytecodes::_invokestatic: 624 case Bytecodes::_invokeinterface: 625 case Bytecodes::_athrow: 626 break; 627 default: { 628 InterpreterOopMap next_mask; 629 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 630 next_mask_expression_stack_size = next_mask.expression_stack_size(); 631 // Need to subtract off the size of the result type of 632 // the bytecode because this is not described in the 633 // debug info but returned to the interpreter in the TOS 634 // caching register 635 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 636 if (bytecode_result_type != T_ILLEGAL) { 637 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 638 } 639 assert(top_frame_expression_stack_adjustment >= 0, ""); 640 try_next_mask = true; 641 break; 642 } 643 } 644 } 645 } 646 647 // Verify stack depth and oops in frame 648 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 649 if (!( 650 /* SPARC */ 651 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 652 /* x86 */ 653 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 654 (try_next_mask && 655 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 656 top_frame_expression_stack_adjustment))) || 657 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 658 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 659 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 660 )) { 661 ttyLocker ttyl; 662 663 // Print out some information that will help us debug the problem 664 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 665 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 666 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 667 iframe->interpreter_frame_expression_stack_size()); 668 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 669 tty->print_cr(" try_next_mask = %d", try_next_mask); 670 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 671 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 672 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 673 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 674 tty->print_cr(" exec_mode = %d", exec_mode); 675 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 676 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 677 tty->print_cr(" Interpreted frames:"); 678 for (int k = 0; k < cur_array->frames(); k++) { 679 vframeArrayElement* el = cur_array->element(k); 680 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 681 } 682 cur_array->print_on_2(tty); 683 guarantee(false, "wrong number of expression stack elements during deopt"); 684 } 685 VerifyOopClosure verify; 686 iframe->oops_interpreted_do(&verify, &rm, false); 687 callee_size_of_parameters = mh->size_of_parameters(); 688 callee_max_locals = mh->max_locals(); 689 is_top_frame = false; 690 } 691 } 692#endif /* !PRODUCT */ 693 694 695 return bt; 696JRT_END 697 698 699int Deoptimization::deoptimize_dependents() { 700 Threads::deoptimized_wrt_marked_nmethods(); 701 return 0; 702} 703 704 705#ifdef COMPILER2 706bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 707 Handle pending_exception(thread->pending_exception()); 708 const char* exception_file = thread->exception_file(); 709 int exception_line = thread->exception_line(); 710 thread->clear_pending_exception(); 711 712 for (int i = 0; i < objects->length(); i++) { 713 assert(objects->at(i)->is_object(), "invalid debug information"); 714 ObjectValue* sv = (ObjectValue*) objects->at(i); 715 716 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 717 oop obj = NULL; 718 719 if (k->oop_is_instance()) { 720 instanceKlass* ik = instanceKlass::cast(k()); 721 obj = ik->allocate_instance(CHECK_(false)); 722 } else if (k->oop_is_typeArray()) { 723 typeArrayKlass* ak = typeArrayKlass::cast(k()); 724 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 725 int len = sv->field_size() / type2size[ak->element_type()]; 726 obj = ak->allocate(len, CHECK_(false)); 727 } else if (k->oop_is_objArray()) { 728 objArrayKlass* ak = objArrayKlass::cast(k()); 729 obj = ak->allocate(sv->field_size(), CHECK_(false)); 730 } 731 732 assert(obj != NULL, "allocation failed"); 733 assert(sv->value().is_null(), "redundant reallocation"); 734 sv->set_value(obj); 735 } 736 737 if (pending_exception.not_null()) { 738 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 739 } 740 741 return true; 742} 743 744// This assumes that the fields are stored in ObjectValue in the same order 745// they are yielded by do_nonstatic_fields. 746class FieldReassigner: public FieldClosure { 747 frame* _fr; 748 RegisterMap* _reg_map; 749 ObjectValue* _sv; 750 instanceKlass* _ik; 751 oop _obj; 752 753 int _i; 754public: 755 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 756 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 757 758 int i() const { return _i; } 759 760 761 void do_field(fieldDescriptor* fd) { 762 intptr_t val; 763 StackValue* value = 764 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 765 int offset = fd->offset(); 766 switch (fd->field_type()) { 767 case T_OBJECT: case T_ARRAY: 768 assert(value->type() == T_OBJECT, "Agreement."); 769 _obj->obj_field_put(offset, value->get_obj()()); 770 break; 771 772 case T_LONG: case T_DOUBLE: { 773 assert(value->type() == T_INT, "Agreement."); 774 StackValue* low = 775 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 776#ifdef _LP64 777 jlong res = (jlong)low->get_int(); 778#else 779#ifdef SPARC 780 // For SPARC we have to swap high and low words. 781 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 782#else 783 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 784#endif //SPARC 785#endif 786 _obj->long_field_put(offset, res); 787 break; 788 } 789 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 790 case T_INT: case T_FLOAT: // 4 bytes. 791 assert(value->type() == T_INT, "Agreement."); 792 val = value->get_int(); 793 _obj->int_field_put(offset, (jint)*((jint*)&val)); 794 break; 795 796 case T_SHORT: case T_CHAR: // 2 bytes 797 assert(value->type() == T_INT, "Agreement."); 798 val = value->get_int(); 799 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 800 break; 801 802 case T_BOOLEAN: case T_BYTE: // 1 byte 803 assert(value->type() == T_INT, "Agreement."); 804 val = value->get_int(); 805 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 806 break; 807 808 default: 809 ShouldNotReachHere(); 810 } 811 _i++; 812 } 813}; 814 815// restore elements of an eliminated type array 816void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 817 int index = 0; 818 intptr_t val; 819 820 for (int i = 0; i < sv->field_size(); i++) { 821 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 822 switch(type) { 823 case T_LONG: case T_DOUBLE: { 824 assert(value->type() == T_INT, "Agreement."); 825 StackValue* low = 826 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 827#ifdef _LP64 828 jlong res = (jlong)low->get_int(); 829#else 830#ifdef SPARC 831 // For SPARC we have to swap high and low words. 832 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 833#else 834 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 835#endif //SPARC 836#endif 837 obj->long_at_put(index, res); 838 break; 839 } 840 841 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 842 case T_INT: case T_FLOAT: // 4 bytes. 843 assert(value->type() == T_INT, "Agreement."); 844 val = value->get_int(); 845 obj->int_at_put(index, (jint)*((jint*)&val)); 846 break; 847 848 case T_SHORT: case T_CHAR: // 2 bytes 849 assert(value->type() == T_INT, "Agreement."); 850 val = value->get_int(); 851 obj->short_at_put(index, (jshort)*((jint*)&val)); 852 break; 853 854 case T_BOOLEAN: case T_BYTE: // 1 byte 855 assert(value->type() == T_INT, "Agreement."); 856 val = value->get_int(); 857 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 858 break; 859 860 default: 861 ShouldNotReachHere(); 862 } 863 index++; 864 } 865} 866 867 868// restore fields of an eliminated object array 869void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 870 for (int i = 0; i < sv->field_size(); i++) { 871 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 872 assert(value->type() == T_OBJECT, "object element expected"); 873 obj->obj_at_put(i, value->get_obj()()); 874 } 875} 876 877 878// restore fields of all eliminated objects and arrays 879void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 880 for (int i = 0; i < objects->length(); i++) { 881 ObjectValue* sv = (ObjectValue*) objects->at(i); 882 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 883 Handle obj = sv->value(); 884 assert(obj.not_null(), "reallocation was missed"); 885 886 if (k->oop_is_instance()) { 887 instanceKlass* ik = instanceKlass::cast(k()); 888 FieldReassigner reassign(fr, reg_map, sv, obj()); 889 ik->do_nonstatic_fields(&reassign); 890 } else if (k->oop_is_typeArray()) { 891 typeArrayKlass* ak = typeArrayKlass::cast(k()); 892 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 893 } else if (k->oop_is_objArray()) { 894 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 895 } 896 } 897} 898 899 900// relock objects for which synchronization was eliminated 901void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 902 for (int i = 0; i < monitors->length(); i++) { 903 MonitorInfo* mon_info = monitors->at(i); 904 if (mon_info->eliminated()) { 905 assert(mon_info->owner() != NULL, "reallocation was missed"); 906 Handle obj = Handle(mon_info->owner()); 907 markOop mark = obj->mark(); 908 if (UseBiasedLocking && mark->has_bias_pattern()) { 909 // New allocated objects may have the mark set to anonymously biased. 910 // Also the deoptimized method may called methods with synchronization 911 // where the thread-local object is bias locked to the current thread. 912 assert(mark->is_biased_anonymously() || 913 mark->biased_locker() == thread, "should be locked to current thread"); 914 // Reset mark word to unbiased prototype. 915 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 916 obj->set_mark(unbiased_prototype); 917 } 918 BasicLock* lock = mon_info->lock(); 919 ObjectSynchronizer::slow_enter(obj, lock, thread); 920 } 921 assert(mon_info->owner()->is_locked(), "object must be locked now"); 922 } 923} 924 925 926#ifndef PRODUCT 927// print information about reallocated objects 928void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 929 fieldDescriptor fd; 930 931 for (int i = 0; i < objects->length(); i++) { 932 ObjectValue* sv = (ObjectValue*) objects->at(i); 933 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 934 Handle obj = sv->value(); 935 936 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 937 k->as_klassOop()->print_value(); 938 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 939 tty->cr(); 940 941 if (Verbose) { 942 k->oop_print_on(obj(), tty); 943 } 944 } 945} 946#endif 947#endif // COMPILER2 948 949vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 950 951#ifndef PRODUCT 952 if (TraceDeoptimization) { 953 ttyLocker ttyl; 954 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 955 fr.print_on(tty); 956 tty->print_cr(" Virtual frames (innermost first):"); 957 for (int index = 0; index < chunk->length(); index++) { 958 compiledVFrame* vf = chunk->at(index); 959 tty->print(" %2d - ", index); 960 vf->print_value(); 961 int bci = chunk->at(index)->raw_bci(); 962 const char* code_name; 963 if (bci == SynchronizationEntryBCI) { 964 code_name = "sync entry"; 965 } else { 966 Bytecodes::Code code = vf->method()->code_at(bci); 967 code_name = Bytecodes::name(code); 968 } 969 tty->print(" - %s", code_name); 970 tty->print_cr(" @ bci %d ", bci); 971 if (Verbose) { 972 vf->print(); 973 tty->cr(); 974 } 975 } 976 } 977#endif 978 979 // Register map for next frame (used for stack crawl). We capture 980 // the state of the deopt'ing frame's caller. Thus if we need to 981 // stuff a C2I adapter we can properly fill in the callee-save 982 // register locations. 983 frame caller = fr.sender(reg_map); 984 int frame_size = caller.sp() - fr.sp(); 985 986 frame sender = caller; 987 988 // Since the Java thread being deoptimized will eventually adjust it's own stack, 989 // the vframeArray containing the unpacking information is allocated in the C heap. 990 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 991 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 992 993 // Compare the vframeArray to the collected vframes 994 assert(array->structural_compare(thread, chunk), "just checking"); 995 Events::log("# vframes = %d", (intptr_t)chunk->length()); 996 997#ifndef PRODUCT 998 if (TraceDeoptimization) { 999 ttyLocker ttyl; 1000 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1001 } 1002#endif // PRODUCT 1003 1004 return array; 1005} 1006 1007 1008static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1009 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1010 for (int i = 0; i < monitors->length(); i++) { 1011 MonitorInfo* mon_info = monitors->at(i); 1012 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1013 objects_to_revoke->append(Handle(mon_info->owner())); 1014 } 1015 } 1016} 1017 1018 1019void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1020 if (!UseBiasedLocking) { 1021 return; 1022 } 1023 1024 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1025 1026 // Unfortunately we don't have a RegisterMap available in most of 1027 // the places we want to call this routine so we need to walk the 1028 // stack again to update the register map. 1029 if (map == NULL || !map->update_map()) { 1030 StackFrameStream sfs(thread, true); 1031 bool found = false; 1032 while (!found && !sfs.is_done()) { 1033 frame* cur = sfs.current(); 1034 sfs.next(); 1035 found = cur->id() == fr.id(); 1036 } 1037 assert(found, "frame to be deoptimized not found on target thread's stack"); 1038 map = sfs.register_map(); 1039 } 1040 1041 vframe* vf = vframe::new_vframe(&fr, map, thread); 1042 compiledVFrame* cvf = compiledVFrame::cast(vf); 1043 // Revoke monitors' biases in all scopes 1044 while (!cvf->is_top()) { 1045 collect_monitors(cvf, objects_to_revoke); 1046 cvf = compiledVFrame::cast(cvf->sender()); 1047 } 1048 collect_monitors(cvf, objects_to_revoke); 1049 1050 if (SafepointSynchronize::is_at_safepoint()) { 1051 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1052 } else { 1053 BiasedLocking::revoke(objects_to_revoke); 1054 } 1055} 1056 1057 1058void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1059 if (!UseBiasedLocking) { 1060 return; 1061 } 1062 1063 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1064 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1065 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1066 if (jt->has_last_Java_frame()) { 1067 StackFrameStream sfs(jt, true); 1068 while (!sfs.is_done()) { 1069 frame* cur = sfs.current(); 1070 if (cb->contains(cur->pc())) { 1071 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1072 compiledVFrame* cvf = compiledVFrame::cast(vf); 1073 // Revoke monitors' biases in all scopes 1074 while (!cvf->is_top()) { 1075 collect_monitors(cvf, objects_to_revoke); 1076 cvf = compiledVFrame::cast(cvf->sender()); 1077 } 1078 collect_monitors(cvf, objects_to_revoke); 1079 } 1080 sfs.next(); 1081 } 1082 } 1083 } 1084 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1085} 1086 1087 1088void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1089 assert(fr.can_be_deoptimized(), "checking frame type"); 1090 1091 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1092 1093 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1094 1095 // Patch the nmethod so that when execution returns to it we will 1096 // deopt the execution state and return to the interpreter. 1097 fr.deoptimize(thread); 1098} 1099 1100void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1101 // Deoptimize only if the frame comes from compile code. 1102 // Do not deoptimize the frame which is already patched 1103 // during the execution of the loops below. 1104 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1105 return; 1106 } 1107 ResourceMark rm; 1108 DeoptimizationMarker dm; 1109 if (UseBiasedLocking) { 1110 revoke_biases_of_monitors(thread, fr, map); 1111 } 1112 deoptimize_single_frame(thread, fr); 1113 1114} 1115 1116 1117void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1118 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1119 "can only deoptimize other thread at a safepoint"); 1120 // Compute frame and register map based on thread and sp. 1121 RegisterMap reg_map(thread, UseBiasedLocking); 1122 frame fr = thread->last_frame(); 1123 while (fr.id() != id) { 1124 fr = fr.sender(®_map); 1125 } 1126 deoptimize(thread, fr, ®_map); 1127} 1128 1129 1130void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1131 if (thread == Thread::current()) { 1132 Deoptimization::deoptimize_frame_internal(thread, id); 1133 } else { 1134 VM_DeoptimizeFrame deopt(thread, id); 1135 VMThread::execute(&deopt); 1136 } 1137} 1138 1139 1140// JVMTI PopFrame support 1141JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1142{ 1143 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1144} 1145JRT_END 1146 1147 1148#if defined(COMPILER2) || defined(SHARK) 1149void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1150 // in case of an unresolved klass entry, load the class. 1151 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1152 klassOop tk = constant_pool->klass_at(index, CHECK); 1153 return; 1154 } 1155 1156 if (!constant_pool->tag_at(index).is_symbol()) return; 1157 1158 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1159 Symbol* symbol = constant_pool->symbol_at(index); 1160 1161 // class name? 1162 if (symbol->byte_at(0) != '(') { 1163 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1164 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1165 return; 1166 } 1167 1168 // then it must be a signature! 1169 ResourceMark rm(THREAD); 1170 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1171 if (ss.is_object()) { 1172 Symbol* class_name = ss.as_symbol(CHECK); 1173 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1174 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1175 } 1176 } 1177} 1178 1179 1180void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1181 EXCEPTION_MARK; 1182 load_class_by_index(constant_pool, index, THREAD); 1183 if (HAS_PENDING_EXCEPTION) { 1184 // Exception happened during classloading. We ignore the exception here, since it 1185 // is going to be rethrown since the current activation is going to be deoptimzied and 1186 // the interpreter will re-execute the bytecode. 1187 CLEAR_PENDING_EXCEPTION; 1188 } 1189} 1190 1191JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1192 HandleMark hm; 1193 1194 // uncommon_trap() is called at the beginning of the uncommon trap 1195 // handler. Note this fact before we start generating temporary frames 1196 // that can confuse an asynchronous stack walker. This counter is 1197 // decremented at the end of unpack_frames(). 1198 thread->inc_in_deopt_handler(); 1199 1200 // We need to update the map if we have biased locking. 1201 RegisterMap reg_map(thread, UseBiasedLocking); 1202 frame stub_frame = thread->last_frame(); 1203 frame fr = stub_frame.sender(®_map); 1204 // Make sure the calling nmethod is not getting deoptimized and removed 1205 // before we are done with it. 1206 nmethodLocker nl(fr.pc()); 1207 1208 { 1209 ResourceMark rm; 1210 1211 // Revoke biases of any monitors in the frame to ensure we can migrate them 1212 revoke_biases_of_monitors(thread, fr, ®_map); 1213 1214 DeoptReason reason = trap_request_reason(trap_request); 1215 DeoptAction action = trap_request_action(trap_request); 1216 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1217 1218 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1219 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1220 compiledVFrame* cvf = compiledVFrame::cast(vf); 1221 1222 nmethod* nm = cvf->code(); 1223 1224 ScopeDesc* trap_scope = cvf->scope(); 1225 methodHandle trap_method = trap_scope->method(); 1226 int trap_bci = trap_scope->bci(); 1227 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1228 1229 // Record this event in the histogram. 1230 gather_statistics(reason, action, trap_bc); 1231 1232 // Ensure that we can record deopt. history: 1233 bool create_if_missing = ProfileTraps; 1234 1235 methodDataHandle trap_mdo 1236 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1237 1238 // Print a bunch of diagnostics, if requested. 1239 if (TraceDeoptimization || LogCompilation) { 1240 ResourceMark rm; 1241 ttyLocker ttyl; 1242 char buf[100]; 1243 if (xtty != NULL) { 1244 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1245 os::current_thread_id(), 1246 format_trap_request(buf, sizeof(buf), trap_request)); 1247 nm->log_identity(xtty); 1248 } 1249 Symbol* class_name = NULL; 1250 bool unresolved = false; 1251 if (unloaded_class_index >= 0) { 1252 constantPoolHandle constants (THREAD, trap_method->constants()); 1253 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1254 class_name = constants->klass_name_at(unloaded_class_index); 1255 unresolved = true; 1256 if (xtty != NULL) 1257 xtty->print(" unresolved='1'"); 1258 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1259 class_name = constants->symbol_at(unloaded_class_index); 1260 } 1261 if (xtty != NULL) 1262 xtty->name(class_name); 1263 } 1264 if (xtty != NULL && trap_mdo.not_null()) { 1265 // Dump the relevant MDO state. 1266 // This is the deopt count for the current reason, any previous 1267 // reasons or recompiles seen at this point. 1268 int dcnt = trap_mdo->trap_count(reason); 1269 if (dcnt != 0) 1270 xtty->print(" count='%d'", dcnt); 1271 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1272 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1273 if (dos != 0) { 1274 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1275 if (trap_state_is_recompiled(dos)) { 1276 int recnt2 = trap_mdo->overflow_recompile_count(); 1277 if (recnt2 != 0) 1278 xtty->print(" recompiles2='%d'", recnt2); 1279 } 1280 } 1281 } 1282 if (xtty != NULL) { 1283 xtty->stamp(); 1284 xtty->end_head(); 1285 } 1286 if (TraceDeoptimization) { // make noise on the tty 1287 tty->print("Uncommon trap occurred in"); 1288 nm->method()->print_short_name(tty); 1289 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1290 fr.pc(), 1291 (int) os::current_thread_id(), 1292 trap_reason_name(reason), 1293 trap_action_name(action), 1294 unloaded_class_index); 1295 if (class_name != NULL) { 1296 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1297 class_name->print_symbol_on(tty); 1298 } 1299 tty->cr(); 1300 } 1301 if (xtty != NULL) { 1302 // Log the precise location of the trap. 1303 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1304 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1305 xtty->method(sd->method()); 1306 xtty->end_elem(); 1307 if (sd->is_top()) break; 1308 } 1309 xtty->tail("uncommon_trap"); 1310 } 1311 } 1312 // (End diagnostic printout.) 1313 1314 // Load class if necessary 1315 if (unloaded_class_index >= 0) { 1316 constantPoolHandle constants(THREAD, trap_method->constants()); 1317 load_class_by_index(constants, unloaded_class_index); 1318 } 1319 1320 // Flush the nmethod if necessary and desirable. 1321 // 1322 // We need to avoid situations where we are re-flushing the nmethod 1323 // because of a hot deoptimization site. Repeated flushes at the same 1324 // point need to be detected by the compiler and avoided. If the compiler 1325 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1326 // module must take measures to avoid an infinite cycle of recompilation 1327 // and deoptimization. There are several such measures: 1328 // 1329 // 1. If a recompilation is ordered a second time at some site X 1330 // and for the same reason R, the action is adjusted to 'reinterpret', 1331 // to give the interpreter time to exercise the method more thoroughly. 1332 // If this happens, the method's overflow_recompile_count is incremented. 1333 // 1334 // 2. If the compiler fails to reduce the deoptimization rate, then 1335 // the method's overflow_recompile_count will begin to exceed the set 1336 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1337 // is adjusted to 'make_not_compilable', and the method is abandoned 1338 // to the interpreter. This is a performance hit for hot methods, 1339 // but is better than a disastrous infinite cycle of recompilations. 1340 // (Actually, only the method containing the site X is abandoned.) 1341 // 1342 // 3. In parallel with the previous measures, if the total number of 1343 // recompilations of a method exceeds the much larger set limit 1344 // PerMethodRecompilationCutoff, the method is abandoned. 1345 // This should only happen if the method is very large and has 1346 // many "lukewarm" deoptimizations. The code which enforces this 1347 // limit is elsewhere (class nmethod, class methodOopDesc). 1348 // 1349 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1350 // to recompile at each bytecode independently of the per-BCI cutoff. 1351 // 1352 // The decision to update code is up to the compiler, and is encoded 1353 // in the Action_xxx code. If the compiler requests Action_none 1354 // no trap state is changed, no compiled code is changed, and the 1355 // computation suffers along in the interpreter. 1356 // 1357 // The other action codes specify various tactics for decompilation 1358 // and recompilation. Action_maybe_recompile is the loosest, and 1359 // allows the compiled code to stay around until enough traps are seen, 1360 // and until the compiler gets around to recompiling the trapping method. 1361 // 1362 // The other actions cause immediate removal of the present code. 1363 1364 bool update_trap_state = true; 1365 bool make_not_entrant = false; 1366 bool make_not_compilable = false; 1367 bool reprofile = false; 1368 switch (action) { 1369 case Action_none: 1370 // Keep the old code. 1371 update_trap_state = false; 1372 break; 1373 case Action_maybe_recompile: 1374 // Do not need to invalidate the present code, but we can 1375 // initiate another 1376 // Start compiler without (necessarily) invalidating the nmethod. 1377 // The system will tolerate the old code, but new code should be 1378 // generated when possible. 1379 break; 1380 case Action_reinterpret: 1381 // Go back into the interpreter for a while, and then consider 1382 // recompiling form scratch. 1383 make_not_entrant = true; 1384 // Reset invocation counter for outer most method. 1385 // This will allow the interpreter to exercise the bytecodes 1386 // for a while before recompiling. 1387 // By contrast, Action_make_not_entrant is immediate. 1388 // 1389 // Note that the compiler will track null_check, null_assert, 1390 // range_check, and class_check events and log them as if they 1391 // had been traps taken from compiled code. This will update 1392 // the MDO trap history so that the next compilation will 1393 // properly detect hot trap sites. 1394 reprofile = true; 1395 break; 1396 case Action_make_not_entrant: 1397 // Request immediate recompilation, and get rid of the old code. 1398 // Make them not entrant, so next time they are called they get 1399 // recompiled. Unloaded classes are loaded now so recompile before next 1400 // time they are called. Same for uninitialized. The interpreter will 1401 // link the missing class, if any. 1402 make_not_entrant = true; 1403 break; 1404 case Action_make_not_compilable: 1405 // Give up on compiling this method at all. 1406 make_not_entrant = true; 1407 make_not_compilable = true; 1408 break; 1409 default: 1410 ShouldNotReachHere(); 1411 } 1412 1413 // Setting +ProfileTraps fixes the following, on all platforms: 1414 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1415 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1416 // recompile relies on a methodDataOop to record heroic opt failures. 1417 1418 // Whether the interpreter is producing MDO data or not, we also need 1419 // to use the MDO to detect hot deoptimization points and control 1420 // aggressive optimization. 1421 bool inc_recompile_count = false; 1422 ProfileData* pdata = NULL; 1423 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1424 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1425 uint this_trap_count = 0; 1426 bool maybe_prior_trap = false; 1427 bool maybe_prior_recompile = false; 1428 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1429 //outputs: 1430 this_trap_count, 1431 maybe_prior_trap, 1432 maybe_prior_recompile); 1433 // Because the interpreter also counts null, div0, range, and class 1434 // checks, these traps from compiled code are double-counted. 1435 // This is harmless; it just means that the PerXTrapLimit values 1436 // are in effect a little smaller than they look. 1437 1438 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1439 if (per_bc_reason != Reason_none) { 1440 // Now take action based on the partially known per-BCI history. 1441 if (maybe_prior_trap 1442 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1443 // If there are too many traps at this BCI, force a recompile. 1444 // This will allow the compiler to see the limit overflow, and 1445 // take corrective action, if possible. The compiler generally 1446 // does not use the exact PerBytecodeTrapLimit value, but instead 1447 // changes its tactics if it sees any traps at all. This provides 1448 // a little hysteresis, delaying a recompile until a trap happens 1449 // several times. 1450 // 1451 // Actually, since there is only one bit of counter per BCI, 1452 // the possible per-BCI counts are {0,1,(per-method count)}. 1453 // This produces accurate results if in fact there is only 1454 // one hot trap site, but begins to get fuzzy if there are 1455 // many sites. For example, if there are ten sites each 1456 // trapping two or more times, they each get the blame for 1457 // all of their traps. 1458 make_not_entrant = true; 1459 } 1460 1461 // Detect repeated recompilation at the same BCI, and enforce a limit. 1462 if (make_not_entrant && maybe_prior_recompile) { 1463 // More than one recompile at this point. 1464 inc_recompile_count = maybe_prior_trap; 1465 } 1466 } else { 1467 // For reasons which are not recorded per-bytecode, we simply 1468 // force recompiles unconditionally. 1469 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1470 make_not_entrant = true; 1471 } 1472 1473 // Go back to the compiler if there are too many traps in this method. 1474 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1475 // If there are too many traps in this method, force a recompile. 1476 // This will allow the compiler to see the limit overflow, and 1477 // take corrective action, if possible. 1478 // (This condition is an unlikely backstop only, because the 1479 // PerBytecodeTrapLimit is more likely to take effect first, 1480 // if it is applicable.) 1481 make_not_entrant = true; 1482 } 1483 1484 // Here's more hysteresis: If there has been a recompile at 1485 // this trap point already, run the method in the interpreter 1486 // for a while to exercise it more thoroughly. 1487 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1488 reprofile = true; 1489 } 1490 1491 } 1492 1493 // Take requested actions on the method: 1494 1495 // Recompile 1496 if (make_not_entrant) { 1497 if (!nm->make_not_entrant()) { 1498 return; // the call did not change nmethod's state 1499 } 1500 1501 if (pdata != NULL) { 1502 // Record the recompilation event, if any. 1503 int tstate0 = pdata->trap_state(); 1504 int tstate1 = trap_state_set_recompiled(tstate0, true); 1505 if (tstate1 != tstate0) 1506 pdata->set_trap_state(tstate1); 1507 } 1508 } 1509 1510 if (inc_recompile_count) { 1511 trap_mdo->inc_overflow_recompile_count(); 1512 if ((uint)trap_mdo->overflow_recompile_count() > 1513 (uint)PerBytecodeRecompilationCutoff) { 1514 // Give up on the method containing the bad BCI. 1515 if (trap_method() == nm->method()) { 1516 make_not_compilable = true; 1517 } else { 1518 trap_method->set_not_compilable(CompLevel_full_optimization); 1519 // But give grace to the enclosing nm->method(). 1520 } 1521 } 1522 } 1523 1524 // Reprofile 1525 if (reprofile) { 1526 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1527 } 1528 1529 // Give up compiling 1530 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1531 assert(make_not_entrant, "consistent"); 1532 nm->method()->set_not_compilable(CompLevel_full_optimization); 1533 } 1534 1535 } // Free marked resources 1536 1537} 1538JRT_END 1539 1540methodDataOop 1541Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1542 bool create_if_missing) { 1543 Thread* THREAD = thread; 1544 methodDataOop mdo = m()->method_data(); 1545 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1546 // Build an MDO. Ignore errors like OutOfMemory; 1547 // that simply means we won't have an MDO to update. 1548 methodOopDesc::build_interpreter_method_data(m, THREAD); 1549 if (HAS_PENDING_EXCEPTION) { 1550 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1551 CLEAR_PENDING_EXCEPTION; 1552 } 1553 mdo = m()->method_data(); 1554 } 1555 return mdo; 1556} 1557 1558ProfileData* 1559Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1560 int trap_bci, 1561 Deoptimization::DeoptReason reason, 1562 //outputs: 1563 uint& ret_this_trap_count, 1564 bool& ret_maybe_prior_trap, 1565 bool& ret_maybe_prior_recompile) { 1566 uint prior_trap_count = trap_mdo->trap_count(reason); 1567 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1568 1569 // If the runtime cannot find a place to store trap history, 1570 // it is estimated based on the general condition of the method. 1571 // If the method has ever been recompiled, or has ever incurred 1572 // a trap with the present reason , then this BCI is assumed 1573 // (pessimistically) to be the culprit. 1574 bool maybe_prior_trap = (prior_trap_count != 0); 1575 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1576 ProfileData* pdata = NULL; 1577 1578 1579 // For reasons which are recorded per bytecode, we check per-BCI data. 1580 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1581 if (per_bc_reason != Reason_none) { 1582 // Find the profile data for this BCI. If there isn't one, 1583 // try to allocate one from the MDO's set of spares. 1584 // This will let us detect a repeated trap at this point. 1585 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1586 1587 if (pdata != NULL) { 1588 // Query the trap state of this profile datum. 1589 int tstate0 = pdata->trap_state(); 1590 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1591 maybe_prior_trap = false; 1592 if (!trap_state_is_recompiled(tstate0)) 1593 maybe_prior_recompile = false; 1594 1595 // Update the trap state of this profile datum. 1596 int tstate1 = tstate0; 1597 // Record the reason. 1598 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1599 // Store the updated state on the MDO, for next time. 1600 if (tstate1 != tstate0) 1601 pdata->set_trap_state(tstate1); 1602 } else { 1603 if (LogCompilation && xtty != NULL) { 1604 ttyLocker ttyl; 1605 // Missing MDP? Leave a small complaint in the log. 1606 xtty->elem("missing_mdp bci='%d'", trap_bci); 1607 } 1608 } 1609 } 1610 1611 // Return results: 1612 ret_this_trap_count = this_trap_count; 1613 ret_maybe_prior_trap = maybe_prior_trap; 1614 ret_maybe_prior_recompile = maybe_prior_recompile; 1615 return pdata; 1616} 1617 1618void 1619Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1620 ResourceMark rm; 1621 // Ignored outputs: 1622 uint ignore_this_trap_count; 1623 bool ignore_maybe_prior_trap; 1624 bool ignore_maybe_prior_recompile; 1625 query_update_method_data(trap_mdo, trap_bci, 1626 (DeoptReason)reason, 1627 ignore_this_trap_count, 1628 ignore_maybe_prior_trap, 1629 ignore_maybe_prior_recompile); 1630} 1631 1632Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1633 1634 // Still in Java no safepoints 1635 { 1636 // This enters VM and may safepoint 1637 uncommon_trap_inner(thread, trap_request); 1638 } 1639 return fetch_unroll_info_helper(thread); 1640} 1641 1642// Local derived constants. 1643// Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1644const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1645const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1646 1647//---------------------------trap_state_reason--------------------------------- 1648Deoptimization::DeoptReason 1649Deoptimization::trap_state_reason(int trap_state) { 1650 // This assert provides the link between the width of DataLayout::trap_bits 1651 // and the encoding of "recorded" reasons. It ensures there are enough 1652 // bits to store all needed reasons in the per-BCI MDO profile. 1653 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1654 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1655 trap_state -= recompile_bit; 1656 if (trap_state == DS_REASON_MASK) { 1657 return Reason_many; 1658 } else { 1659 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1660 return (DeoptReason)trap_state; 1661 } 1662} 1663//-------------------------trap_state_has_reason------------------------------- 1664int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1665 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1666 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1667 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1668 trap_state -= recompile_bit; 1669 if (trap_state == DS_REASON_MASK) { 1670 return -1; // true, unspecifically (bottom of state lattice) 1671 } else if (trap_state == reason) { 1672 return 1; // true, definitely 1673 } else if (trap_state == 0) { 1674 return 0; // false, definitely (top of state lattice) 1675 } else { 1676 return 0; // false, definitely 1677 } 1678} 1679//-------------------------trap_state_add_reason------------------------------- 1680int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1681 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1682 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1683 trap_state -= recompile_bit; 1684 if (trap_state == DS_REASON_MASK) { 1685 return trap_state + recompile_bit; // already at state lattice bottom 1686 } else if (trap_state == reason) { 1687 return trap_state + recompile_bit; // the condition is already true 1688 } else if (trap_state == 0) { 1689 return reason + recompile_bit; // no condition has yet been true 1690 } else { 1691 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1692 } 1693} 1694//-----------------------trap_state_is_recompiled------------------------------ 1695bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1696 return (trap_state & DS_RECOMPILE_BIT) != 0; 1697} 1698//-----------------------trap_state_set_recompiled----------------------------- 1699int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1700 if (z) return trap_state | DS_RECOMPILE_BIT; 1701 else return trap_state & ~DS_RECOMPILE_BIT; 1702} 1703//---------------------------format_trap_state--------------------------------- 1704// This is used for debugging and diagnostics, including hotspot.log output. 1705const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1706 int trap_state) { 1707 DeoptReason reason = trap_state_reason(trap_state); 1708 bool recomp_flag = trap_state_is_recompiled(trap_state); 1709 // Re-encode the state from its decoded components. 1710 int decoded_state = 0; 1711 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1712 decoded_state = trap_state_add_reason(decoded_state, reason); 1713 if (recomp_flag) 1714 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1715 // If the state re-encodes properly, format it symbolically. 1716 // Because this routine is used for debugging and diagnostics, 1717 // be robust even if the state is a strange value. 1718 size_t len; 1719 if (decoded_state != trap_state) { 1720 // Random buggy state that doesn't decode?? 1721 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1722 } else { 1723 len = jio_snprintf(buf, buflen, "%s%s", 1724 trap_reason_name(reason), 1725 recomp_flag ? " recompiled" : ""); 1726 } 1727 if (len >= buflen) 1728 buf[buflen-1] = '\0'; 1729 return buf; 1730} 1731 1732 1733//--------------------------------statics-------------------------------------- 1734Deoptimization::DeoptAction Deoptimization::_unloaded_action 1735 = Deoptimization::Action_reinterpret; 1736const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1737 // Note: Keep this in sync. with enum DeoptReason. 1738 "none", 1739 "null_check", 1740 "null_assert", 1741 "range_check", 1742 "class_check", 1743 "array_check", 1744 "intrinsic", 1745 "bimorphic", 1746 "unloaded", 1747 "uninitialized", 1748 "unreached", 1749 "unhandled", 1750 "constraint", 1751 "div0_check", 1752 "age", 1753 "predicate" 1754}; 1755const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1756 // Note: Keep this in sync. with enum DeoptAction. 1757 "none", 1758 "maybe_recompile", 1759 "reinterpret", 1760 "make_not_entrant", 1761 "make_not_compilable" 1762}; 1763 1764const char* Deoptimization::trap_reason_name(int reason) { 1765 if (reason == Reason_many) return "many"; 1766 if ((uint)reason < Reason_LIMIT) 1767 return _trap_reason_name[reason]; 1768 static char buf[20]; 1769 sprintf(buf, "reason%d", reason); 1770 return buf; 1771} 1772const char* Deoptimization::trap_action_name(int action) { 1773 if ((uint)action < Action_LIMIT) 1774 return _trap_action_name[action]; 1775 static char buf[20]; 1776 sprintf(buf, "action%d", action); 1777 return buf; 1778} 1779 1780// This is used for debugging and diagnostics, including hotspot.log output. 1781const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1782 int trap_request) { 1783 jint unloaded_class_index = trap_request_index(trap_request); 1784 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1785 const char* action = trap_action_name(trap_request_action(trap_request)); 1786 size_t len; 1787 if (unloaded_class_index < 0) { 1788 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1789 reason, action); 1790 } else { 1791 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1792 reason, action, unloaded_class_index); 1793 } 1794 if (len >= buflen) 1795 buf[buflen-1] = '\0'; 1796 return buf; 1797} 1798 1799juint Deoptimization::_deoptimization_hist 1800 [Deoptimization::Reason_LIMIT] 1801 [1 + Deoptimization::Action_LIMIT] 1802 [Deoptimization::BC_CASE_LIMIT] 1803 = {0}; 1804 1805enum { 1806 LSB_BITS = 8, 1807 LSB_MASK = right_n_bits(LSB_BITS) 1808}; 1809 1810void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1811 Bytecodes::Code bc) { 1812 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1813 assert(action >= 0 && action < Action_LIMIT, "oob"); 1814 _deoptimization_hist[Reason_none][0][0] += 1; // total 1815 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1816 juint* cases = _deoptimization_hist[reason][1+action]; 1817 juint* bc_counter_addr = NULL; 1818 juint bc_counter = 0; 1819 // Look for an unused counter, or an exact match to this BC. 1820 if (bc != Bytecodes::_illegal) { 1821 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1822 juint* counter_addr = &cases[bc_case]; 1823 juint counter = *counter_addr; 1824 if ((counter == 0 && bc_counter_addr == NULL) 1825 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1826 // this counter is either free or is already devoted to this BC 1827 bc_counter_addr = counter_addr; 1828 bc_counter = counter | bc; 1829 } 1830 } 1831 } 1832 if (bc_counter_addr == NULL) { 1833 // Overflow, or no given bytecode. 1834 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1835 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1836 } 1837 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1838} 1839 1840jint Deoptimization::total_deoptimization_count() { 1841 return _deoptimization_hist[Reason_none][0][0]; 1842} 1843 1844jint Deoptimization::deoptimization_count(DeoptReason reason) { 1845 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1846 return _deoptimization_hist[reason][0][0]; 1847} 1848 1849void Deoptimization::print_statistics() { 1850 juint total = total_deoptimization_count(); 1851 juint account = total; 1852 if (total != 0) { 1853 ttyLocker ttyl; 1854 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1855 tty->print_cr("Deoptimization traps recorded:"); 1856 #define PRINT_STAT_LINE(name, r) \ 1857 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1858 PRINT_STAT_LINE("total", total); 1859 // For each non-zero entry in the histogram, print the reason, 1860 // the action, and (if specifically known) the type of bytecode. 1861 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1862 for (int action = 0; action < Action_LIMIT; action++) { 1863 juint* cases = _deoptimization_hist[reason][1+action]; 1864 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1865 juint counter = cases[bc_case]; 1866 if (counter != 0) { 1867 char name[1*K]; 1868 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1869 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1870 bc = Bytecodes::_illegal; 1871 sprintf(name, "%s/%s/%s", 1872 trap_reason_name(reason), 1873 trap_action_name(action), 1874 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1875 juint r = counter >> LSB_BITS; 1876 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1877 account -= r; 1878 } 1879 } 1880 } 1881 } 1882 if (account != 0) { 1883 PRINT_STAT_LINE("unaccounted", account); 1884 } 1885 #undef PRINT_STAT_LINE 1886 if (xtty != NULL) xtty->tail("statistics"); 1887 } 1888} 1889#else // COMPILER2 || SHARK 1890 1891 1892// Stubs for C1 only system. 1893bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1894 return false; 1895} 1896 1897const char* Deoptimization::trap_reason_name(int reason) { 1898 return "unknown"; 1899} 1900 1901void Deoptimization::print_statistics() { 1902 // no output 1903} 1904 1905void 1906Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1907 // no udpate 1908} 1909 1910int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1911 return 0; 1912} 1913 1914void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1915 Bytecodes::Code bc) { 1916 // no update 1917} 1918 1919const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1920 int trap_state) { 1921 jio_snprintf(buf, buflen, "#%d", trap_state); 1922 return buf; 1923} 1924 1925#endif // COMPILER2 || SHARK 1926