1 /* 2 * Copyright (c) 1997, 2013, 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/method.hpp" 38 #include "oops/oop.inline.hpp" 39 #include "prims/jvmtiThreadState.hpp" 40 #include "runtime/biasedLocking.hpp" 41 #include "runtime/compilationPolicy.hpp" 42 #include "runtime/deoptimization.hpp" 43 #include "runtime/interfaceSupport.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/signature.hpp" 46 #include "runtime/stubRoutines.hpp" 47 #include "runtime/thread.hpp" 48 #include "runtime/vframe.hpp" 49 #include "runtime/vframeArray.hpp" 50 #include "runtime/vframe_hp.hpp" 51 #include "utilities/events.hpp" 52 #include "utilities/xmlstream.hpp" 53 #ifdef TARGET_ARCH_x86 54 # include "vmreg_x86.inline.hpp" 55 #endif 56 #ifdef TARGET_ARCH_sparc 57 # include "vmreg_sparc.inline.hpp" 58 #endif 59 #ifdef TARGET_ARCH_zero 60 # include "vmreg_zero.inline.hpp" 61 #endif 62 #ifdef TARGET_ARCH_arm 63 # include "vmreg_arm.inline.hpp" 64 #endif 65 #ifdef TARGET_ARCH_ppc 66 # include "vmreg_ppc.inline.hpp" 67 #endif 68 #ifdef COMPILER2 69 #ifdef TARGET_ARCH_MODEL_x86_32 70 # include "adfiles/ad_x86_32.hpp" 71 #endif 72 #ifdef TARGET_ARCH_MODEL_x86_64 73 # include "adfiles/ad_x86_64.hpp" 74 #endif 75 #ifdef TARGET_ARCH_MODEL_sparc 76 # include "adfiles/ad_sparc.hpp" 77 #endif 78 #ifdef TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #endif 81 #ifdef TARGET_ARCH_MODEL_arm 82 # include "adfiles/ad_arm.hpp" 83 #endif 84 #ifdef TARGET_ARCH_MODEL_ppc_32 85 # include "adfiles/ad_ppc_32.hpp" 86 #endif 87 #ifdef TARGET_ARCH_MODEL_ppc_64 88 # include "adfiles/ad_ppc_64.hpp" 89 #endif 90 #endif // COMPILER2 91 92 bool DeoptimizationMarker::_is_active = false; 93 94 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 95 int caller_adjustment, 96 int caller_actual_parameters, 97 int number_of_frames, 98 intptr_t* frame_sizes, 99 address* frame_pcs, 100 BasicType return_type) { 101 _size_of_deoptimized_frame = size_of_deoptimized_frame; 102 _caller_adjustment = caller_adjustment; 103 _caller_actual_parameters = caller_actual_parameters; 104 _number_of_frames = number_of_frames; 105 _frame_sizes = frame_sizes; 106 _frame_pcs = frame_pcs; 107 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 108 _return_type = return_type; 109 _initial_info = 0; 110 // PD (x86 only) 111 _counter_temp = 0; 112 _unpack_kind = 0; 113 _sender_sp_temp = 0; 114 115 _total_frame_sizes = size_of_frames(); 116 } 117 118 119 Deoptimization::UnrollBlock::~UnrollBlock() { 120 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); 121 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); 122 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); 123 } 124 125 126 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 127 assert(register_number < RegisterMap::reg_count, "checking register number"); 128 return &_register_block[register_number * 2]; 129 } 130 131 132 133 int Deoptimization::UnrollBlock::size_of_frames() const { 134 // Acount first for the adjustment of the initial frame 135 int result = _caller_adjustment; 136 for (int index = 0; index < number_of_frames(); index++) { 137 result += frame_sizes()[index]; 138 } 139 return result; 140 } 141 142 143 void Deoptimization::UnrollBlock::print() { 144 ttyLocker ttyl; 145 tty->print_cr("UnrollBlock"); 146 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 147 tty->print( " frame_sizes: "); 148 for (int index = 0; index < number_of_frames(); index++) { 149 tty->print("%d ", frame_sizes()[index]); 150 } 151 tty->cr(); 152 } 153 154 155 // In order to make fetch_unroll_info work properly with escape 156 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 157 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 158 // of previously eliminated objects occurs in realloc_objects, which is 159 // called from the method fetch_unroll_info_helper below. 160 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 161 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 162 // but makes the entry a little slower. There is however a little dance we have to 163 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 164 165 // fetch_unroll_info() is called at the beginning of the deoptimization 166 // handler. Note this fact before we start generating temporary frames 167 // that can confuse an asynchronous stack walker. This counter is 168 // decremented at the end of unpack_frames(). 169 thread->inc_in_deopt_handler(); 170 171 return fetch_unroll_info_helper(thread); 172 JRT_END 173 174 175 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 176 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 177 178 // Note: there is a safepoint safety issue here. No matter whether we enter 179 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 180 // the vframeArray is created. 181 // 182 183 // Allocate our special deoptimization ResourceMark 184 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 185 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 186 thread->set_deopt_mark(dmark); 187 188 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 189 RegisterMap map(thread, true); 190 RegisterMap dummy_map(thread, false); 191 // Now get the deoptee with a valid map 192 frame deoptee = stub_frame.sender(&map); 193 // Set the deoptee nmethod 194 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 195 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 196 197 if (VerifyStack) { 198 thread->validate_frame_layout(); 199 } 200 201 // Create a growable array of VFrames where each VFrame represents an inlined 202 // Java frame. This storage is allocated with the usual system arena. 203 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 204 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 205 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 206 while (!vf->is_top()) { 207 assert(vf->is_compiled_frame(), "Wrong frame type"); 208 chunk->push(compiledVFrame::cast(vf)); 209 vf = vf->sender(); 210 } 211 assert(vf->is_compiled_frame(), "Wrong frame type"); 212 chunk->push(compiledVFrame::cast(vf)); 213 214 #ifdef COMPILER2 215 // Reallocate the non-escaping objects and restore their fields. Then 216 // relock objects if synchronization on them was eliminated. 217 if (DoEscapeAnalysis || EliminateNestedLocks) { 218 if (EliminateAllocations) { 219 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 220 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 221 222 // The flag return_oop() indicates call sites which return oop 223 // in compiled code. Such sites include java method calls, 224 // runtime calls (for example, used to allocate new objects/arrays 225 // on slow code path) and any other calls generated in compiled code. 226 // It is not guaranteed that we can get such information here only 227 // by analyzing bytecode in deoptimized frames. This is why this flag 228 // is set during method compilation (see Compile::Process_OopMap_Node()). 229 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 230 Handle return_value; 231 if (save_oop_result) { 232 // Reallocation may trigger GC. If deoptimization happened on return from 233 // call which returns oop we need to save it since it is not in oopmap. 234 oop result = deoptee.saved_oop_result(&map); 235 assert(result == NULL || result->is_oop(), "must be oop"); 236 return_value = Handle(thread, result); 237 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 238 if (TraceDeoptimization) { 239 ttyLocker ttyl; 240 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread); 241 } 242 } 243 bool reallocated = false; 244 if (objects != NULL) { 245 JRT_BLOCK 246 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 247 JRT_END 248 } 249 if (reallocated) { 250 reassign_fields(&deoptee, &map, objects); 251 #ifndef PRODUCT 252 if (TraceDeoptimization) { 253 ttyLocker ttyl; 254 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 255 print_objects(objects); 256 } 257 #endif 258 } 259 if (save_oop_result) { 260 // Restore result. 261 deoptee.set_saved_oop_result(&map, return_value()); 262 } 263 } 264 if (EliminateLocks) { 265 #ifndef PRODUCT 266 bool first = true; 267 #endif 268 for (int i = 0; i < chunk->length(); i++) { 269 compiledVFrame* cvf = chunk->at(i); 270 assert (cvf->scope() != NULL,"expect only compiled java frames"); 271 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 272 if (monitors->is_nonempty()) { 273 relock_objects(monitors, thread); 274 #ifndef PRODUCT 275 if (TraceDeoptimization) { 276 ttyLocker ttyl; 277 for (int j = 0; j < monitors->length(); j++) { 278 MonitorInfo* mi = monitors->at(j); 279 if (mi->eliminated()) { 280 if (first) { 281 first = false; 282 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 283 } 284 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner()); 285 } 286 } 287 } 288 #endif 289 } 290 } 291 } 292 } 293 #endif // COMPILER2 294 // Ensure that no safepoint is taken after pointers have been stored 295 // in fields of rematerialized objects. If a safepoint occurs from here on 296 // out the java state residing in the vframeArray will be missed. 297 No_Safepoint_Verifier no_safepoint; 298 299 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 300 301 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 302 thread->set_vframe_array_head(array); 303 304 // Now that the vframeArray has been created if we have any deferred local writes 305 // added by jvmti then we can free up that structure as the data is now in the 306 // vframeArray 307 308 if (thread->deferred_locals() != NULL) { 309 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 310 int i = 0; 311 do { 312 // Because of inlining we could have multiple vframes for a single frame 313 // and several of the vframes could have deferred writes. Find them all. 314 if (list->at(i)->id() == array->original().id()) { 315 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 316 list->remove_at(i); 317 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 318 delete dlv; 319 } else { 320 i++; 321 } 322 } while ( i < list->length() ); 323 if (list->length() == 0) { 324 thread->set_deferred_locals(NULL); 325 // free the list and elements back to C heap. 326 delete list; 327 } 328 329 } 330 331 #ifndef SHARK 332 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 333 CodeBlob* cb = stub_frame.cb(); 334 // Verify we have the right vframeArray 335 assert(cb->frame_size() >= 0, "Unexpected frame size"); 336 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 337 338 // If the deopt call site is a MethodHandle invoke call site we have 339 // to adjust the unpack_sp. 340 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 341 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 342 unpack_sp = deoptee.unextended_sp(); 343 344 #ifdef ASSERT 345 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 346 #endif 347 #else 348 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 349 #endif // !SHARK 350 351 // This is a guarantee instead of an assert because if vframe doesn't match 352 // we will unpack the wrong deoptimized frame and wind up in strange places 353 // where it will be very difficult to figure out what went wrong. Better 354 // to die an early death here than some very obscure death later when the 355 // trail is cold. 356 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 357 // in that it will fail to detect a problem when there is one. This needs 358 // more work in tiger timeframe. 359 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 360 361 int number_of_frames = array->frames(); 362 363 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 364 // virtual activation, which is the reverse of the elements in the vframes array. 365 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 366 // +1 because we always have an interpreter return address for the final slot. 367 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 368 int popframe_extra_args = 0; 369 // Create an interpreter return address for the stub to use as its return 370 // address so the skeletal frames are perfectly walkable 371 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 372 373 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 374 // activation be put back on the expression stack of the caller for reexecution 375 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 376 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 377 } 378 379 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 380 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 381 // than simply use array->sender.pc(). This requires us to walk the current set of frames 382 // 383 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 384 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 385 386 // It's possible that the number of parameters at the call site is 387 // different than number of arguments in the callee when method 388 // handles are used. If the caller is interpreted get the real 389 // value so that the proper amount of space can be added to it's 390 // frame. 391 bool caller_was_method_handle = false; 392 if (deopt_sender.is_interpreted_frame()) { 393 methodHandle method = deopt_sender.interpreter_frame_method(); 394 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 395 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 396 // Method handle invokes may involve fairly arbitrary chains of 397 // calls so it's impossible to know how much actual space the 398 // caller has for locals. 399 caller_was_method_handle = true; 400 } 401 } 402 403 // 404 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 405 // frame_sizes/frame_pcs[1] next oldest frame (int) 406 // frame_sizes/frame_pcs[n] youngest frame (int) 407 // 408 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 409 // owns the space for the return address to it's caller). Confusing ain't it. 410 // 411 // The vframe array can address vframes with indices running from 412 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 413 // When we create the skeletal frames we need the oldest frame to be in the zero slot 414 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 415 // so things look a little strange in this loop. 416 // 417 int callee_parameters = 0; 418 int callee_locals = 0; 419 for (int index = 0; index < array->frames(); index++ ) { 420 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 421 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 422 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 423 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 424 callee_locals, 425 index == 0, 426 popframe_extra_args); 427 // This pc doesn't have to be perfect just good enough to identify the frame 428 // as interpreted so the skeleton frame will be walkable 429 // The correct pc will be set when the skeleton frame is completely filled out 430 // The final pc we store in the loop is wrong and will be overwritten below 431 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 432 433 callee_parameters = array->element(index)->method()->size_of_parameters(); 434 callee_locals = array->element(index)->method()->max_locals(); 435 popframe_extra_args = 0; 436 } 437 438 // Compute whether the root vframe returns a float or double value. 439 BasicType return_type; 440 { 441 HandleMark hm; 442 methodHandle method(thread, array->element(0)->method()); 443 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 444 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 445 } 446 447 // Compute information for handling adapters and adjusting the frame size of the caller. 448 int caller_adjustment = 0; 449 450 // Compute the amount the oldest interpreter frame will have to adjust 451 // its caller's stack by. If the caller is a compiled frame then 452 // we pretend that the callee has no parameters so that the 453 // extension counts for the full amount of locals and not just 454 // locals-parms. This is because without a c2i adapter the parm 455 // area as created by the compiled frame will not be usable by 456 // the interpreter. (Depending on the calling convention there 457 // may not even be enough space). 458 459 // QQQ I'd rather see this pushed down into last_frame_adjust 460 // and have it take the sender (aka caller). 461 462 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 463 caller_adjustment = last_frame_adjust(0, callee_locals); 464 } else if (callee_locals > callee_parameters) { 465 // The caller frame may need extending to accommodate 466 // non-parameter locals of the first unpacked interpreted frame. 467 // Compute that adjustment. 468 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 469 } 470 471 // If the sender is deoptimized the we must retrieve the address of the handler 472 // since the frame will "magically" show the original pc before the deopt 473 // and we'd undo the deopt. 474 475 frame_pcs[0] = deopt_sender.raw_pc(); 476 477 #ifndef SHARK 478 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 479 #endif // SHARK 480 481 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 482 caller_adjustment * BytesPerWord, 483 caller_was_method_handle ? 0 : callee_parameters, 484 number_of_frames, 485 frame_sizes, 486 frame_pcs, 487 return_type); 488 // On some platforms, we need a way to pass some platform dependent 489 // information to the unpacking code so the skeletal frames come out 490 // correct (initial fp value, unextended sp, ...) 491 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 492 493 if (array->frames() > 1) { 494 if (VerifyStack && TraceDeoptimization) { 495 ttyLocker ttyl; 496 tty->print_cr("Deoptimizing method containing inlining"); 497 } 498 } 499 500 array->set_unroll_block(info); 501 return info; 502 } 503 504 // Called to cleanup deoptimization data structures in normal case 505 // after unpacking to stack and when stack overflow error occurs 506 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 507 vframeArray *array) { 508 509 // Get array if coming from exception 510 if (array == NULL) { 511 array = thread->vframe_array_head(); 512 } 513 thread->set_vframe_array_head(NULL); 514 515 // Free the previous UnrollBlock 516 vframeArray* old_array = thread->vframe_array_last(); 517 thread->set_vframe_array_last(array); 518 519 if (old_array != NULL) { 520 UnrollBlock* old_info = old_array->unroll_block(); 521 old_array->set_unroll_block(NULL); 522 delete old_info; 523 delete old_array; 524 } 525 526 // Deallocate any resource creating in this routine and any ResourceObjs allocated 527 // inside the vframeArray (StackValueCollections) 528 529 delete thread->deopt_mark(); 530 thread->set_deopt_mark(NULL); 531 thread->set_deopt_nmethod(NULL); 532 533 534 if (JvmtiExport::can_pop_frame()) { 535 #ifndef CC_INTERP 536 // Regardless of whether we entered this routine with the pending 537 // popframe condition bit set, we should always clear it now 538 thread->clear_popframe_condition(); 539 #else 540 // C++ interpreter will clear has_pending_popframe when it enters 541 // with method_resume. For deopt_resume2 we clear it now. 542 if (thread->popframe_forcing_deopt_reexecution()) 543 thread->clear_popframe_condition(); 544 #endif /* CC_INTERP */ 545 } 546 547 // unpack_frames() is called at the end of the deoptimization handler 548 // and (in C2) at the end of the uncommon trap handler. Note this fact 549 // so that an asynchronous stack walker can work again. This counter is 550 // incremented at the beginning of fetch_unroll_info() and (in C2) at 551 // the beginning of uncommon_trap(). 552 thread->dec_in_deopt_handler(); 553 } 554 555 556 // Return BasicType of value being returned 557 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 558 559 // We are already active int he special DeoptResourceMark any ResourceObj's we 560 // allocate will be freed at the end of the routine. 561 562 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 563 // but makes the entry a little slower. There is however a little dance we have to 564 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 565 ResetNoHandleMark rnhm; // No-op in release/product versions 566 HandleMark hm; 567 568 frame stub_frame = thread->last_frame(); 569 570 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 571 // must point to the vframeArray for the unpack frame. 572 vframeArray* array = thread->vframe_array_head(); 573 574 #ifndef PRODUCT 575 if (TraceDeoptimization) { 576 ttyLocker ttyl; 577 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 578 } 579 #endif 580 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 581 stub_frame.pc(), stub_frame.sp(), exec_mode); 582 583 UnrollBlock* info = array->unroll_block(); 584 585 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 586 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 587 588 BasicType bt = info->return_type(); 589 590 // If we have an exception pending, claim that the return type is an oop 591 // so the deopt_blob does not overwrite the exception_oop. 592 593 if (exec_mode == Unpack_exception) 594 bt = T_OBJECT; 595 596 // Cleanup thread deopt data 597 cleanup_deopt_info(thread, array); 598 599 #ifndef PRODUCT 600 if (VerifyStack) { 601 ResourceMark res_mark; 602 603 thread->validate_frame_layout(); 604 605 // Verify that the just-unpacked frames match the interpreter's 606 // notions of expression stack and locals 607 vframeArray* cur_array = thread->vframe_array_last(); 608 RegisterMap rm(thread, false); 609 rm.set_include_argument_oops(false); 610 bool is_top_frame = true; 611 int callee_size_of_parameters = 0; 612 int callee_max_locals = 0; 613 for (int i = 0; i < cur_array->frames(); i++) { 614 vframeArrayElement* el = cur_array->element(i); 615 frame* iframe = el->iframe(); 616 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 617 618 // Get the oop map for this bci 619 InterpreterOopMap mask; 620 int cur_invoke_parameter_size = 0; 621 bool try_next_mask = false; 622 int next_mask_expression_stack_size = -1; 623 int top_frame_expression_stack_adjustment = 0; 624 methodHandle mh(thread, iframe->interpreter_frame_method()); 625 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 626 BytecodeStream str(mh); 627 str.set_start(iframe->interpreter_frame_bci()); 628 int max_bci = mh->code_size(); 629 // Get to the next bytecode if possible 630 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 631 // Check to see if we can grab the number of outgoing arguments 632 // at an uncommon trap for an invoke (where the compiler 633 // generates debug info before the invoke has executed) 634 Bytecodes::Code cur_code = str.next(); 635 if (cur_code == Bytecodes::_invokevirtual || 636 cur_code == Bytecodes::_invokespecial || 637 cur_code == Bytecodes::_invokestatic || 638 cur_code == Bytecodes::_invokeinterface || 639 cur_code == Bytecodes::_invokedynamic) { 640 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 641 Symbol* signature = invoke.signature(); 642 ArgumentSizeComputer asc(signature); 643 cur_invoke_parameter_size = asc.size(); 644 if (invoke.has_receiver()) { 645 // Add in receiver 646 ++cur_invoke_parameter_size; 647 } 648 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 649 callee_size_of_parameters++; 650 } 651 } 652 if (str.bci() < max_bci) { 653 Bytecodes::Code bc = str.next(); 654 if (bc >= 0) { 655 // The interpreter oop map generator reports results before 656 // the current bytecode has executed except in the case of 657 // calls. It seems to be hard to tell whether the compiler 658 // has emitted debug information matching the "state before" 659 // a given bytecode or the state after, so we try both 660 switch (cur_code) { 661 case Bytecodes::_invokevirtual: 662 case Bytecodes::_invokespecial: 663 case Bytecodes::_invokestatic: 664 case Bytecodes::_invokeinterface: 665 case Bytecodes::_invokedynamic: 666 case Bytecodes::_athrow: 667 break; 668 default: { 669 InterpreterOopMap next_mask; 670 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 671 next_mask_expression_stack_size = next_mask.expression_stack_size(); 672 // Need to subtract off the size of the result type of 673 // the bytecode because this is not described in the 674 // debug info but returned to the interpreter in the TOS 675 // caching register 676 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 677 if (bytecode_result_type != T_ILLEGAL) { 678 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 679 } 680 assert(top_frame_expression_stack_adjustment >= 0, ""); 681 try_next_mask = true; 682 break; 683 } 684 } 685 } 686 } 687 688 // Verify stack depth and oops in frame 689 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 690 if (!( 691 /* SPARC */ 692 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 693 /* x86 */ 694 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 695 (try_next_mask && 696 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 697 top_frame_expression_stack_adjustment))) || 698 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 699 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 700 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 701 )) { 702 ttyLocker ttyl; 703 704 // Print out some information that will help us debug the problem 705 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 706 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 707 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 708 iframe->interpreter_frame_expression_stack_size()); 709 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 710 tty->print_cr(" try_next_mask = %d", try_next_mask); 711 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 712 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 713 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 714 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 715 tty->print_cr(" exec_mode = %d", exec_mode); 716 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 717 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 718 tty->print_cr(" Interpreted frames:"); 719 for (int k = 0; k < cur_array->frames(); k++) { 720 vframeArrayElement* el = cur_array->element(k); 721 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 722 } 723 cur_array->print_on_2(tty); 724 guarantee(false, "wrong number of expression stack elements during deopt"); 725 } 726 VerifyOopClosure verify; 727 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 728 callee_size_of_parameters = mh->size_of_parameters(); 729 callee_max_locals = mh->max_locals(); 730 is_top_frame = false; 731 } 732 } 733 #endif /* !PRODUCT */ 734 735 736 return bt; 737 JRT_END 738 739 740 int Deoptimization::deoptimize_dependents() { 741 Threads::deoptimized_wrt_marked_nmethods(); 742 return 0; 743 } 744 745 Deoptimization::DeoptAction Deoptimization::_unloaded_action 746 = Deoptimization::Action_reinterpret; 747 748 #ifdef COMPILER2 749 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 750 Handle pending_exception(thread->pending_exception()); 751 const char* exception_file = thread->exception_file(); 752 int exception_line = thread->exception_line(); 753 thread->clear_pending_exception(); 754 755 for (int i = 0; i < objects->length(); i++) { 756 assert(objects->at(i)->is_object(), "invalid debug information"); 757 ObjectValue* sv = (ObjectValue*) objects->at(i); 758 759 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 760 oop obj = NULL; 761 762 if (k->oop_is_instance()) { 763 InstanceKlass* ik = InstanceKlass::cast(k()); 764 obj = ik->allocate_instance(CHECK_(false)); 765 } else if (k->oop_is_typeArray()) { 766 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 767 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 768 int len = sv->field_size() / type2size[ak->element_type()]; 769 obj = ak->allocate(len, CHECK_(false)); 770 } else if (k->oop_is_objArray()) { 771 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 772 obj = ak->allocate(sv->field_size(), CHECK_(false)); 773 } 774 775 assert(obj != NULL, "allocation failed"); 776 assert(sv->value().is_null(), "redundant reallocation"); 777 sv->set_value(obj); 778 } 779 780 if (pending_exception.not_null()) { 781 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 782 } 783 784 return true; 785 } 786 787 // This assumes that the fields are stored in ObjectValue in the same order 788 // they are yielded by do_nonstatic_fields. 789 class FieldReassigner: public FieldClosure { 790 frame* _fr; 791 RegisterMap* _reg_map; 792 ObjectValue* _sv; 793 InstanceKlass* _ik; 794 oop _obj; 795 796 int _i; 797 public: 798 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 799 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 800 801 int i() const { return _i; } 802 803 804 void do_field(fieldDescriptor* fd) { 805 intptr_t val; 806 StackValue* value = 807 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 808 int offset = fd->offset(); 809 switch (fd->field_type()) { 810 case T_OBJECT: case T_ARRAY: 811 assert(value->type() == T_OBJECT, "Agreement."); 812 _obj->obj_field_put(offset, value->get_obj()()); 813 break; 814 815 case T_LONG: case T_DOUBLE: { 816 assert(value->type() == T_INT, "Agreement."); 817 StackValue* low = 818 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 819 #ifdef _LP64 820 jlong res = (jlong)low->get_int(); 821 #else 822 #ifdef SPARC 823 // For SPARC we have to swap high and low words. 824 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 825 #else 826 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 827 #endif //SPARC 828 #endif 829 _obj->long_field_put(offset, res); 830 break; 831 } 832 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 833 case T_INT: case T_FLOAT: // 4 bytes. 834 assert(value->type() == T_INT, "Agreement."); 835 val = value->get_int(); 836 _obj->int_field_put(offset, (jint)*((jint*)&val)); 837 break; 838 839 case T_SHORT: case T_CHAR: // 2 bytes 840 assert(value->type() == T_INT, "Agreement."); 841 val = value->get_int(); 842 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 843 break; 844 845 case T_BOOLEAN: case T_BYTE: // 1 byte 846 assert(value->type() == T_INT, "Agreement."); 847 val = value->get_int(); 848 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 849 break; 850 851 default: 852 ShouldNotReachHere(); 853 } 854 _i++; 855 } 856 }; 857 858 // restore elements of an eliminated type array 859 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 860 int index = 0; 861 intptr_t val; 862 863 for (int i = 0; i < sv->field_size(); i++) { 864 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 865 switch(type) { 866 case T_LONG: case T_DOUBLE: { 867 assert(value->type() == T_INT, "Agreement."); 868 StackValue* low = 869 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 870 #ifdef _LP64 871 jlong res = (jlong)low->get_int(); 872 #else 873 #ifdef SPARC 874 // For SPARC we have to swap high and low words. 875 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 876 #else 877 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 878 #endif //SPARC 879 #endif 880 obj->long_at_put(index, res); 881 break; 882 } 883 884 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 885 case T_INT: case T_FLOAT: // 4 bytes. 886 assert(value->type() == T_INT, "Agreement."); 887 val = value->get_int(); 888 obj->int_at_put(index, (jint)*((jint*)&val)); 889 break; 890 891 case T_SHORT: case T_CHAR: // 2 bytes 892 assert(value->type() == T_INT, "Agreement."); 893 val = value->get_int(); 894 obj->short_at_put(index, (jshort)*((jint*)&val)); 895 break; 896 897 case T_BOOLEAN: case T_BYTE: // 1 byte 898 assert(value->type() == T_INT, "Agreement."); 899 val = value->get_int(); 900 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 901 break; 902 903 default: 904 ShouldNotReachHere(); 905 } 906 index++; 907 } 908 } 909 910 911 // restore fields of an eliminated object array 912 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 913 for (int i = 0; i < sv->field_size(); i++) { 914 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 915 assert(value->type() == T_OBJECT, "object element expected"); 916 obj->obj_at_put(i, value->get_obj()()); 917 } 918 } 919 920 921 // restore fields of all eliminated objects and arrays 922 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 923 for (int i = 0; i < objects->length(); i++) { 924 ObjectValue* sv = (ObjectValue*) objects->at(i); 925 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 926 Handle obj = sv->value(); 927 assert(obj.not_null(), "reallocation was missed"); 928 929 if (k->oop_is_instance()) { 930 InstanceKlass* ik = InstanceKlass::cast(k()); 931 FieldReassigner reassign(fr, reg_map, sv, obj()); 932 ik->do_nonstatic_fields(&reassign); 933 } else if (k->oop_is_typeArray()) { 934 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 935 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 936 } else if (k->oop_is_objArray()) { 937 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 938 } 939 } 940 } 941 942 943 // relock objects for which synchronization was eliminated 944 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 945 for (int i = 0; i < monitors->length(); i++) { 946 MonitorInfo* mon_info = monitors->at(i); 947 if (mon_info->eliminated()) { 948 assert(mon_info->owner() != NULL, "reallocation was missed"); 949 Handle obj = Handle(mon_info->owner()); 950 markOop mark = obj->mark(); 951 if (UseBiasedLocking && mark->has_bias_pattern()) { 952 // New allocated objects may have the mark set to anonymously biased. 953 // Also the deoptimized method may called methods with synchronization 954 // where the thread-local object is bias locked to the current thread. 955 assert(mark->is_biased_anonymously() || 956 mark->biased_locker() == thread, "should be locked to current thread"); 957 // Reset mark word to unbiased prototype. 958 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 959 obj->set_mark(unbiased_prototype); 960 } 961 BasicLock* lock = mon_info->lock(); 962 ObjectSynchronizer::slow_enter(obj, lock, thread); 963 } 964 assert(mon_info->owner()->is_locked(), "object must be locked now"); 965 } 966 } 967 968 969 #ifndef PRODUCT 970 // print information about reallocated objects 971 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 972 fieldDescriptor fd; 973 974 for (int i = 0; i < objects->length(); i++) { 975 ObjectValue* sv = (ObjectValue*) objects->at(i); 976 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 977 Handle obj = sv->value(); 978 979 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()()); 980 k->print_value(); 981 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 982 tty->cr(); 983 984 if (Verbose) { 985 k->oop_print_on(obj(), tty); 986 } 987 } 988 } 989 #endif 990 #endif // COMPILER2 991 992 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 993 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 994 995 #ifndef PRODUCT 996 if (TraceDeoptimization) { 997 ttyLocker ttyl; 998 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 999 fr.print_on(tty); 1000 tty->print_cr(" Virtual frames (innermost first):"); 1001 for (int index = 0; index < chunk->length(); index++) { 1002 compiledVFrame* vf = chunk->at(index); 1003 tty->print(" %2d - ", index); 1004 vf->print_value(); 1005 int bci = chunk->at(index)->raw_bci(); 1006 const char* code_name; 1007 if (bci == SynchronizationEntryBCI) { 1008 code_name = "sync entry"; 1009 } else { 1010 Bytecodes::Code code = vf->method()->code_at(bci); 1011 code_name = Bytecodes::name(code); 1012 } 1013 tty->print(" - %s", code_name); 1014 tty->print_cr(" @ bci %d ", bci); 1015 if (Verbose) { 1016 vf->print(); 1017 tty->cr(); 1018 } 1019 } 1020 } 1021 #endif 1022 1023 // Register map for next frame (used for stack crawl). We capture 1024 // the state of the deopt'ing frame's caller. Thus if we need to 1025 // stuff a C2I adapter we can properly fill in the callee-save 1026 // register locations. 1027 frame caller = fr.sender(reg_map); 1028 int frame_size = caller.sp() - fr.sp(); 1029 1030 frame sender = caller; 1031 1032 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1033 // the vframeArray containing the unpacking information is allocated in the C heap. 1034 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1035 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1036 1037 // Compare the vframeArray to the collected vframes 1038 assert(array->structural_compare(thread, chunk), "just checking"); 1039 1040 #ifndef PRODUCT 1041 if (TraceDeoptimization) { 1042 ttyLocker ttyl; 1043 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1044 } 1045 #endif // PRODUCT 1046 1047 return array; 1048 } 1049 1050 1051 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1052 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1053 for (int i = 0; i < monitors->length(); i++) { 1054 MonitorInfo* mon_info = monitors->at(i); 1055 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1056 objects_to_revoke->append(Handle(mon_info->owner())); 1057 } 1058 } 1059 } 1060 1061 1062 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1063 if (!UseBiasedLocking) { 1064 return; 1065 } 1066 1067 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1068 1069 // Unfortunately we don't have a RegisterMap available in most of 1070 // the places we want to call this routine so we need to walk the 1071 // stack again to update the register map. 1072 if (map == NULL || !map->update_map()) { 1073 StackFrameStream sfs(thread, true); 1074 bool found = false; 1075 while (!found && !sfs.is_done()) { 1076 frame* cur = sfs.current(); 1077 sfs.next(); 1078 found = cur->id() == fr.id(); 1079 } 1080 assert(found, "frame to be deoptimized not found on target thread's stack"); 1081 map = sfs.register_map(); 1082 } 1083 1084 vframe* vf = vframe::new_vframe(&fr, map, thread); 1085 compiledVFrame* cvf = compiledVFrame::cast(vf); 1086 // Revoke monitors' biases in all scopes 1087 while (!cvf->is_top()) { 1088 collect_monitors(cvf, objects_to_revoke); 1089 cvf = compiledVFrame::cast(cvf->sender()); 1090 } 1091 collect_monitors(cvf, objects_to_revoke); 1092 1093 if (SafepointSynchronize::is_at_safepoint()) { 1094 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1095 } else { 1096 BiasedLocking::revoke(objects_to_revoke); 1097 } 1098 } 1099 1100 1101 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1102 if (!UseBiasedLocking) { 1103 return; 1104 } 1105 1106 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1107 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1108 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1109 if (jt->has_last_Java_frame()) { 1110 StackFrameStream sfs(jt, true); 1111 while (!sfs.is_done()) { 1112 frame* cur = sfs.current(); 1113 if (cb->contains(cur->pc())) { 1114 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1115 compiledVFrame* cvf = compiledVFrame::cast(vf); 1116 // Revoke monitors' biases in all scopes 1117 while (!cvf->is_top()) { 1118 collect_monitors(cvf, objects_to_revoke); 1119 cvf = compiledVFrame::cast(cvf->sender()); 1120 } 1121 collect_monitors(cvf, objects_to_revoke); 1122 } 1123 sfs.next(); 1124 } 1125 } 1126 } 1127 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1128 } 1129 1130 1131 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1132 assert(fr.can_be_deoptimized(), "checking frame type"); 1133 1134 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1135 1136 // Patch the nmethod so that when execution returns to it we will 1137 // deopt the execution state and return to the interpreter. 1138 fr.deoptimize(thread); 1139 } 1140 1141 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1142 // Deoptimize only if the frame comes from compile code. 1143 // Do not deoptimize the frame which is already patched 1144 // during the execution of the loops below. 1145 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1146 return; 1147 } 1148 ResourceMark rm; 1149 DeoptimizationMarker dm; 1150 if (UseBiasedLocking) { 1151 revoke_biases_of_monitors(thread, fr, map); 1152 } 1153 deoptimize_single_frame(thread, fr); 1154 1155 } 1156 1157 1158 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1159 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1160 "can only deoptimize other thread at a safepoint"); 1161 // Compute frame and register map based on thread and sp. 1162 RegisterMap reg_map(thread, UseBiasedLocking); 1163 frame fr = thread->last_frame(); 1164 while (fr.id() != id) { 1165 fr = fr.sender(®_map); 1166 } 1167 deoptimize(thread, fr, ®_map); 1168 } 1169 1170 1171 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1172 if (thread == Thread::current()) { 1173 Deoptimization::deoptimize_frame_internal(thread, id); 1174 } else { 1175 VM_DeoptimizeFrame deopt(thread, id); 1176 VMThread::execute(&deopt); 1177 } 1178 } 1179 1180 1181 // JVMTI PopFrame support 1182 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1183 { 1184 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1185 } 1186 JRT_END 1187 1188 1189 #if defined(COMPILER2) || defined(SHARK) 1190 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1191 // in case of an unresolved klass entry, load the class. 1192 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1193 Klass* tk = constant_pool->klass_at(index, CHECK); 1194 return; 1195 } 1196 1197 if (!constant_pool->tag_at(index).is_symbol()) return; 1198 1199 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1200 Symbol* symbol = constant_pool->symbol_at(index); 1201 1202 // class name? 1203 if (symbol->byte_at(0) != '(') { 1204 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1205 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1206 return; 1207 } 1208 1209 // then it must be a signature! 1210 ResourceMark rm(THREAD); 1211 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1212 if (ss.is_object()) { 1213 Symbol* class_name = ss.as_symbol(CHECK); 1214 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1215 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1216 } 1217 } 1218 } 1219 1220 1221 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1222 EXCEPTION_MARK; 1223 load_class_by_index(constant_pool, index, THREAD); 1224 if (HAS_PENDING_EXCEPTION) { 1225 // Exception happened during classloading. We ignore the exception here, since it 1226 // is going to be rethrown since the current activation is going to be deoptimized and 1227 // the interpreter will re-execute the bytecode. 1228 CLEAR_PENDING_EXCEPTION; 1229 // Class loading called java code which may have caused a stack 1230 // overflow. If the exception was thrown right before the return 1231 // to the runtime the stack is no longer guarded. Reguard the 1232 // stack otherwise if we return to the uncommon trap blob and the 1233 // stack bang causes a stack overflow we crash. 1234 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1235 JavaThread* thread = (JavaThread*)THREAD; 1236 bool guard_pages_enabled = thread->stack_yellow_zone_enabled(); 1237 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1238 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1239 } 1240 } 1241 1242 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1243 HandleMark hm; 1244 1245 // uncommon_trap() is called at the beginning of the uncommon trap 1246 // handler. Note this fact before we start generating temporary frames 1247 // that can confuse an asynchronous stack walker. This counter is 1248 // decremented at the end of unpack_frames(). 1249 thread->inc_in_deopt_handler(); 1250 1251 // We need to update the map if we have biased locking. 1252 RegisterMap reg_map(thread, UseBiasedLocking); 1253 frame stub_frame = thread->last_frame(); 1254 frame fr = stub_frame.sender(®_map); 1255 // Make sure the calling nmethod is not getting deoptimized and removed 1256 // before we are done with it. 1257 nmethodLocker nl(fr.pc()); 1258 1259 // Log a message 1260 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT, 1261 trap_request, fr.pc()); 1262 1263 { 1264 ResourceMark rm; 1265 1266 // Revoke biases of any monitors in the frame to ensure we can migrate them 1267 revoke_biases_of_monitors(thread, fr, ®_map); 1268 1269 DeoptReason reason = trap_request_reason(trap_request); 1270 DeoptAction action = trap_request_action(trap_request); 1271 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1272 1273 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1274 compiledVFrame* cvf = compiledVFrame::cast(vf); 1275 1276 nmethod* nm = cvf->code(); 1277 1278 ScopeDesc* trap_scope = cvf->scope(); 1279 methodHandle trap_method = trap_scope->method(); 1280 int trap_bci = trap_scope->bci(); 1281 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1282 1283 // Record this event in the histogram. 1284 gather_statistics(reason, action, trap_bc); 1285 1286 // Ensure that we can record deopt. history: 1287 // Need MDO to record RTM code generation state. 1288 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 1289 1290 MethodData* trap_mdo = 1291 get_method_data(thread, trap_method, create_if_missing); 1292 1293 // Log a message 1294 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1295 trap_reason_name(reason), trap_action_name(action), fr.pc(), 1296 trap_method->name_and_sig_as_C_string(), trap_bci); 1297 1298 // Print a bunch of diagnostics, if requested. 1299 if (TraceDeoptimization || LogCompilation) { 1300 ResourceMark rm; 1301 ttyLocker ttyl; 1302 char buf[100]; 1303 if (xtty != NULL) { 1304 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1305 os::current_thread_id(), 1306 format_trap_request(buf, sizeof(buf), trap_request)); 1307 nm->log_identity(xtty); 1308 } 1309 Symbol* class_name = NULL; 1310 bool unresolved = false; 1311 if (unloaded_class_index >= 0) { 1312 constantPoolHandle constants (THREAD, trap_method->constants()); 1313 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1314 class_name = constants->klass_name_at(unloaded_class_index); 1315 unresolved = true; 1316 if (xtty != NULL) 1317 xtty->print(" unresolved='1'"); 1318 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1319 class_name = constants->symbol_at(unloaded_class_index); 1320 } 1321 if (xtty != NULL) 1322 xtty->name(class_name); 1323 } 1324 if (xtty != NULL && trap_mdo != NULL) { 1325 // Dump the relevant MDO state. 1326 // This is the deopt count for the current reason, any previous 1327 // reasons or recompiles seen at this point. 1328 int dcnt = trap_mdo->trap_count(reason); 1329 if (dcnt != 0) 1330 xtty->print(" count='%d'", dcnt); 1331 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1332 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1333 if (dos != 0) { 1334 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1335 if (trap_state_is_recompiled(dos)) { 1336 int recnt2 = trap_mdo->overflow_recompile_count(); 1337 if (recnt2 != 0) 1338 xtty->print(" recompiles2='%d'", recnt2); 1339 } 1340 } 1341 } 1342 if (xtty != NULL) { 1343 xtty->stamp(); 1344 xtty->end_head(); 1345 } 1346 if (TraceDeoptimization) { // make noise on the tty 1347 tty->print("Uncommon trap occurred in"); 1348 nm->method()->print_short_name(tty); 1349 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1350 fr.pc(), 1351 os::current_thread_id(), 1352 trap_reason_name(reason), 1353 trap_action_name(action), 1354 unloaded_class_index); 1355 if (class_name != NULL) { 1356 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1357 class_name->print_symbol_on(tty); 1358 } 1359 tty->cr(); 1360 } 1361 if (xtty != NULL) { 1362 // Log the precise location of the trap. 1363 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1364 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1365 xtty->method(sd->method()); 1366 xtty->end_elem(); 1367 if (sd->is_top()) break; 1368 } 1369 xtty->tail("uncommon_trap"); 1370 } 1371 } 1372 // (End diagnostic printout.) 1373 1374 // Load class if necessary 1375 if (unloaded_class_index >= 0) { 1376 constantPoolHandle constants(THREAD, trap_method->constants()); 1377 load_class_by_index(constants, unloaded_class_index); 1378 } 1379 1380 // Flush the nmethod if necessary and desirable. 1381 // 1382 // We need to avoid situations where we are re-flushing the nmethod 1383 // because of a hot deoptimization site. Repeated flushes at the same 1384 // point need to be detected by the compiler and avoided. If the compiler 1385 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1386 // module must take measures to avoid an infinite cycle of recompilation 1387 // and deoptimization. There are several such measures: 1388 // 1389 // 1. If a recompilation is ordered a second time at some site X 1390 // and for the same reason R, the action is adjusted to 'reinterpret', 1391 // to give the interpreter time to exercise the method more thoroughly. 1392 // If this happens, the method's overflow_recompile_count is incremented. 1393 // 1394 // 2. If the compiler fails to reduce the deoptimization rate, then 1395 // the method's overflow_recompile_count will begin to exceed the set 1396 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1397 // is adjusted to 'make_not_compilable', and the method is abandoned 1398 // to the interpreter. This is a performance hit for hot methods, 1399 // but is better than a disastrous infinite cycle of recompilations. 1400 // (Actually, only the method containing the site X is abandoned.) 1401 // 1402 // 3. In parallel with the previous measures, if the total number of 1403 // recompilations of a method exceeds the much larger set limit 1404 // PerMethodRecompilationCutoff, the method is abandoned. 1405 // This should only happen if the method is very large and has 1406 // many "lukewarm" deoptimizations. The code which enforces this 1407 // limit is elsewhere (class nmethod, class Method). 1408 // 1409 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1410 // to recompile at each bytecode independently of the per-BCI cutoff. 1411 // 1412 // The decision to update code is up to the compiler, and is encoded 1413 // in the Action_xxx code. If the compiler requests Action_none 1414 // no trap state is changed, no compiled code is changed, and the 1415 // computation suffers along in the interpreter. 1416 // 1417 // The other action codes specify various tactics for decompilation 1418 // and recompilation. Action_maybe_recompile is the loosest, and 1419 // allows the compiled code to stay around until enough traps are seen, 1420 // and until the compiler gets around to recompiling the trapping method. 1421 // 1422 // The other actions cause immediate removal of the present code. 1423 1424 bool update_trap_state = true; 1425 bool make_not_entrant = false; 1426 bool make_not_compilable = false; 1427 bool reprofile = false; 1428 switch (action) { 1429 case Action_none: 1430 // Keep the old code. 1431 update_trap_state = false; 1432 break; 1433 case Action_maybe_recompile: 1434 // Do not need to invalidate the present code, but we can 1435 // initiate another 1436 // Start compiler without (necessarily) invalidating the nmethod. 1437 // The system will tolerate the old code, but new code should be 1438 // generated when possible. 1439 break; 1440 case Action_reinterpret: 1441 // Go back into the interpreter for a while, and then consider 1442 // recompiling form scratch. 1443 make_not_entrant = true; 1444 // Reset invocation counter for outer most method. 1445 // This will allow the interpreter to exercise the bytecodes 1446 // for a while before recompiling. 1447 // By contrast, Action_make_not_entrant is immediate. 1448 // 1449 // Note that the compiler will track null_check, null_assert, 1450 // range_check, and class_check events and log them as if they 1451 // had been traps taken from compiled code. This will update 1452 // the MDO trap history so that the next compilation will 1453 // properly detect hot trap sites. 1454 reprofile = true; 1455 break; 1456 case Action_make_not_entrant: 1457 // Request immediate recompilation, and get rid of the old code. 1458 // Make them not entrant, so next time they are called they get 1459 // recompiled. Unloaded classes are loaded now so recompile before next 1460 // time they are called. Same for uninitialized. The interpreter will 1461 // link the missing class, if any. 1462 make_not_entrant = true; 1463 break; 1464 case Action_make_not_compilable: 1465 // Give up on compiling this method at all. 1466 make_not_entrant = true; 1467 make_not_compilable = true; 1468 break; 1469 default: 1470 ShouldNotReachHere(); 1471 } 1472 1473 // Setting +ProfileTraps fixes the following, on all platforms: 1474 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1475 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1476 // recompile relies on a MethodData* to record heroic opt failures. 1477 1478 // Whether the interpreter is producing MDO data or not, we also need 1479 // to use the MDO to detect hot deoptimization points and control 1480 // aggressive optimization. 1481 bool inc_recompile_count = false; 1482 ProfileData* pdata = NULL; 1483 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1484 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1485 uint this_trap_count = 0; 1486 bool maybe_prior_trap = false; 1487 bool maybe_prior_recompile = false; 1488 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1489 nm->method(), 1490 //outputs: 1491 this_trap_count, 1492 maybe_prior_trap, 1493 maybe_prior_recompile); 1494 // Because the interpreter also counts null, div0, range, and class 1495 // checks, these traps from compiled code are double-counted. 1496 // This is harmless; it just means that the PerXTrapLimit values 1497 // are in effect a little smaller than they look. 1498 1499 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1500 if (per_bc_reason != Reason_none) { 1501 // Now take action based on the partially known per-BCI history. 1502 if (maybe_prior_trap 1503 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1504 // If there are too many traps at this BCI, force a recompile. 1505 // This will allow the compiler to see the limit overflow, and 1506 // take corrective action, if possible. The compiler generally 1507 // does not use the exact PerBytecodeTrapLimit value, but instead 1508 // changes its tactics if it sees any traps at all. This provides 1509 // a little hysteresis, delaying a recompile until a trap happens 1510 // several times. 1511 // 1512 // Actually, since there is only one bit of counter per BCI, 1513 // the possible per-BCI counts are {0,1,(per-method count)}. 1514 // This produces accurate results if in fact there is only 1515 // one hot trap site, but begins to get fuzzy if there are 1516 // many sites. For example, if there are ten sites each 1517 // trapping two or more times, they each get the blame for 1518 // all of their traps. 1519 make_not_entrant = true; 1520 } 1521 1522 // Detect repeated recompilation at the same BCI, and enforce a limit. 1523 if (make_not_entrant && maybe_prior_recompile) { 1524 // More than one recompile at this point. 1525 inc_recompile_count = maybe_prior_trap; 1526 } 1527 } else { 1528 // For reasons which are not recorded per-bytecode, we simply 1529 // force recompiles unconditionally. 1530 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1531 make_not_entrant = true; 1532 } 1533 1534 // Go back to the compiler if there are too many traps in this method. 1535 if (this_trap_count >= per_method_trap_limit(reason)) { 1536 // If there are too many traps in this method, force a recompile. 1537 // This will allow the compiler to see the limit overflow, and 1538 // take corrective action, if possible. 1539 // (This condition is an unlikely backstop only, because the 1540 // PerBytecodeTrapLimit is more likely to take effect first, 1541 // if it is applicable.) 1542 make_not_entrant = true; 1543 } 1544 1545 // Here's more hysteresis: If there has been a recompile at 1546 // this trap point already, run the method in the interpreter 1547 // for a while to exercise it more thoroughly. 1548 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1549 reprofile = true; 1550 } 1551 1552 } 1553 1554 // Take requested actions on the method: 1555 1556 // Recompile 1557 if (make_not_entrant) { 1558 if (!nm->make_not_entrant()) { 1559 return; // the call did not change nmethod's state 1560 } 1561 1562 if (pdata != NULL) { 1563 // Record the recompilation event, if any. 1564 int tstate0 = pdata->trap_state(); 1565 int tstate1 = trap_state_set_recompiled(tstate0, true); 1566 if (tstate1 != tstate0) 1567 pdata->set_trap_state(tstate1); 1568 } 1569 1570 #if INCLUDE_RTM_OPT 1571 // Restart collecting RTM locking abort statistic if the method 1572 // is recompiled for a reason other than RTM state change. 1573 // Assume that in new recompiled code the statistic could be different, 1574 // for example, due to different inlining. 1575 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1576 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1577 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1578 } 1579 #endif 1580 } 1581 1582 if (inc_recompile_count) { 1583 trap_mdo->inc_overflow_recompile_count(); 1584 if ((uint)trap_mdo->overflow_recompile_count() > 1585 (uint)PerBytecodeRecompilationCutoff) { 1586 // Give up on the method containing the bad BCI. 1587 if (trap_method() == nm->method()) { 1588 make_not_compilable = true; 1589 } else { 1590 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1591 // But give grace to the enclosing nm->method(). 1592 } 1593 } 1594 } 1595 1596 // Reprofile 1597 if (reprofile) { 1598 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1599 } 1600 1601 // Give up compiling 1602 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1603 assert(make_not_entrant, "consistent"); 1604 nm->method()->set_not_compilable(CompLevel_full_optimization); 1605 } 1606 1607 } // Free marked resources 1608 1609 } 1610 JRT_END 1611 1612 MethodData* 1613 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1614 bool create_if_missing) { 1615 Thread* THREAD = thread; 1616 MethodData* mdo = m()->method_data(); 1617 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1618 // Build an MDO. Ignore errors like OutOfMemory; 1619 // that simply means we won't have an MDO to update. 1620 Method::build_interpreter_method_data(m, THREAD); 1621 if (HAS_PENDING_EXCEPTION) { 1622 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1623 CLEAR_PENDING_EXCEPTION; 1624 } 1625 mdo = m()->method_data(); 1626 } 1627 return mdo; 1628 } 1629 1630 ProfileData* 1631 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1632 int trap_bci, 1633 Deoptimization::DeoptReason reason, 1634 Method* compiled_method, 1635 //outputs: 1636 uint& ret_this_trap_count, 1637 bool& ret_maybe_prior_trap, 1638 bool& ret_maybe_prior_recompile) { 1639 uint prior_trap_count = trap_mdo->trap_count(reason); 1640 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1641 1642 // If the runtime cannot find a place to store trap history, 1643 // it is estimated based on the general condition of the method. 1644 // If the method has ever been recompiled, or has ever incurred 1645 // a trap with the present reason , then this BCI is assumed 1646 // (pessimistically) to be the culprit. 1647 bool maybe_prior_trap = (prior_trap_count != 0); 1648 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1649 ProfileData* pdata = NULL; 1650 1651 1652 // For reasons which are recorded per bytecode, we check per-BCI data. 1653 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1654 if (per_bc_reason != Reason_none) { 1655 // Find the profile data for this BCI. If there isn't one, 1656 // try to allocate one from the MDO's set of spares. 1657 // This will let us detect a repeated trap at this point. 1658 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1659 1660 if (pdata != NULL) { 1661 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1662 if (LogCompilation && xtty != NULL) { 1663 ttyLocker ttyl; 1664 // no more room for speculative traps in this MDO 1665 xtty->elem("speculative_traps_oom"); 1666 } 1667 } 1668 // Query the trap state of this profile datum. 1669 int tstate0 = pdata->trap_state(); 1670 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1671 maybe_prior_trap = false; 1672 if (!trap_state_is_recompiled(tstate0)) 1673 maybe_prior_recompile = false; 1674 1675 // Update the trap state of this profile datum. 1676 int tstate1 = tstate0; 1677 // Record the reason. 1678 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1679 // Store the updated state on the MDO, for next time. 1680 if (tstate1 != tstate0) 1681 pdata->set_trap_state(tstate1); 1682 } else { 1683 if (LogCompilation && xtty != NULL) { 1684 ttyLocker ttyl; 1685 // Missing MDP? Leave a small complaint in the log. 1686 xtty->elem("missing_mdp bci='%d'", trap_bci); 1687 } 1688 } 1689 } 1690 1691 // Return results: 1692 ret_this_trap_count = this_trap_count; 1693 ret_maybe_prior_trap = maybe_prior_trap; 1694 ret_maybe_prior_recompile = maybe_prior_recompile; 1695 return pdata; 1696 } 1697 1698 void 1699 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1700 ResourceMark rm; 1701 // Ignored outputs: 1702 uint ignore_this_trap_count; 1703 bool ignore_maybe_prior_trap; 1704 bool ignore_maybe_prior_recompile; 1705 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 1706 query_update_method_data(trap_mdo, trap_bci, 1707 (DeoptReason)reason, 1708 NULL, 1709 ignore_this_trap_count, 1710 ignore_maybe_prior_trap, 1711 ignore_maybe_prior_recompile); 1712 } 1713 1714 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1715 1716 // Still in Java no safepoints 1717 { 1718 // This enters VM and may safepoint 1719 uncommon_trap_inner(thread, trap_request); 1720 } 1721 return fetch_unroll_info_helper(thread); 1722 } 1723 1724 // Local derived constants. 1725 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1726 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1727 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1728 1729 //---------------------------trap_state_reason--------------------------------- 1730 Deoptimization::DeoptReason 1731 Deoptimization::trap_state_reason(int trap_state) { 1732 // This assert provides the link between the width of DataLayout::trap_bits 1733 // and the encoding of "recorded" reasons. It ensures there are enough 1734 // bits to store all needed reasons in the per-BCI MDO profile. 1735 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1736 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1737 trap_state -= recompile_bit; 1738 if (trap_state == DS_REASON_MASK) { 1739 return Reason_many; 1740 } else { 1741 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1742 return (DeoptReason)trap_state; 1743 } 1744 } 1745 //-------------------------trap_state_has_reason------------------------------- 1746 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1747 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1748 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1749 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1750 trap_state -= recompile_bit; 1751 if (trap_state == DS_REASON_MASK) { 1752 return -1; // true, unspecifically (bottom of state lattice) 1753 } else if (trap_state == reason) { 1754 return 1; // true, definitely 1755 } else if (trap_state == 0) { 1756 return 0; // false, definitely (top of state lattice) 1757 } else { 1758 return 0; // false, definitely 1759 } 1760 } 1761 //-------------------------trap_state_add_reason------------------------------- 1762 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1763 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1764 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1765 trap_state -= recompile_bit; 1766 if (trap_state == DS_REASON_MASK) { 1767 return trap_state + recompile_bit; // already at state lattice bottom 1768 } else if (trap_state == reason) { 1769 return trap_state + recompile_bit; // the condition is already true 1770 } else if (trap_state == 0) { 1771 return reason + recompile_bit; // no condition has yet been true 1772 } else { 1773 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1774 } 1775 } 1776 //-----------------------trap_state_is_recompiled------------------------------ 1777 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1778 return (trap_state & DS_RECOMPILE_BIT) != 0; 1779 } 1780 //-----------------------trap_state_set_recompiled----------------------------- 1781 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1782 if (z) return trap_state | DS_RECOMPILE_BIT; 1783 else return trap_state & ~DS_RECOMPILE_BIT; 1784 } 1785 //---------------------------format_trap_state--------------------------------- 1786 // This is used for debugging and diagnostics, including LogFile output. 1787 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1788 int trap_state) { 1789 DeoptReason reason = trap_state_reason(trap_state); 1790 bool recomp_flag = trap_state_is_recompiled(trap_state); 1791 // Re-encode the state from its decoded components. 1792 int decoded_state = 0; 1793 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1794 decoded_state = trap_state_add_reason(decoded_state, reason); 1795 if (recomp_flag) 1796 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1797 // If the state re-encodes properly, format it symbolically. 1798 // Because this routine is used for debugging and diagnostics, 1799 // be robust even if the state is a strange value. 1800 size_t len; 1801 if (decoded_state != trap_state) { 1802 // Random buggy state that doesn't decode?? 1803 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1804 } else { 1805 len = jio_snprintf(buf, buflen, "%s%s", 1806 trap_reason_name(reason), 1807 recomp_flag ? " recompiled" : ""); 1808 } 1809 if (len >= buflen) 1810 buf[buflen-1] = '\0'; 1811 return buf; 1812 } 1813 1814 1815 //--------------------------------statics-------------------------------------- 1816 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1817 // Note: Keep this in sync. with enum DeoptReason. 1818 "none", 1819 "null_check", 1820 "null_assert", 1821 "range_check", 1822 "class_check", 1823 "array_check", 1824 "intrinsic", 1825 "bimorphic", 1826 "unloaded", 1827 "uninitialized", 1828 "unreached", 1829 "unhandled", 1830 "constraint", 1831 "div0_check", 1832 "age", 1833 "predicate", 1834 "loop_limit_check", 1835 "speculate_class_check", 1836 "speculate_null_check", 1837 "rtm_state_change" 1838 }; 1839 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1840 // Note: Keep this in sync. with enum DeoptAction. 1841 "none", 1842 "maybe_recompile", 1843 "reinterpret", 1844 "make_not_entrant", 1845 "make_not_compilable" 1846 }; 1847 1848 const char* Deoptimization::trap_reason_name(int reason) { 1849 if (reason == Reason_many) return "many"; 1850 if ((uint)reason < Reason_LIMIT) 1851 return _trap_reason_name[reason]; 1852 static char buf[20]; 1853 sprintf(buf, "reason%d", reason); 1854 return buf; 1855 } 1856 const char* Deoptimization::trap_action_name(int action) { 1857 if ((uint)action < Action_LIMIT) 1858 return _trap_action_name[action]; 1859 static char buf[20]; 1860 sprintf(buf, "action%d", action); 1861 return buf; 1862 } 1863 1864 // This is used for debugging and diagnostics, including LogFile output. 1865 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1866 int trap_request) { 1867 jint unloaded_class_index = trap_request_index(trap_request); 1868 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1869 const char* action = trap_action_name(trap_request_action(trap_request)); 1870 size_t len; 1871 if (unloaded_class_index < 0) { 1872 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1873 reason, action); 1874 } else { 1875 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1876 reason, action, unloaded_class_index); 1877 } 1878 if (len >= buflen) 1879 buf[buflen-1] = '\0'; 1880 return buf; 1881 } 1882 1883 juint Deoptimization::_deoptimization_hist 1884 [Deoptimization::Reason_LIMIT] 1885 [1 + Deoptimization::Action_LIMIT] 1886 [Deoptimization::BC_CASE_LIMIT] 1887 = {0}; 1888 1889 enum { 1890 LSB_BITS = 8, 1891 LSB_MASK = right_n_bits(LSB_BITS) 1892 }; 1893 1894 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1895 Bytecodes::Code bc) { 1896 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1897 assert(action >= 0 && action < Action_LIMIT, "oob"); 1898 _deoptimization_hist[Reason_none][0][0] += 1; // total 1899 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1900 juint* cases = _deoptimization_hist[reason][1+action]; 1901 juint* bc_counter_addr = NULL; 1902 juint bc_counter = 0; 1903 // Look for an unused counter, or an exact match to this BC. 1904 if (bc != Bytecodes::_illegal) { 1905 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1906 juint* counter_addr = &cases[bc_case]; 1907 juint counter = *counter_addr; 1908 if ((counter == 0 && bc_counter_addr == NULL) 1909 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1910 // this counter is either free or is already devoted to this BC 1911 bc_counter_addr = counter_addr; 1912 bc_counter = counter | bc; 1913 } 1914 } 1915 } 1916 if (bc_counter_addr == NULL) { 1917 // Overflow, or no given bytecode. 1918 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1919 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1920 } 1921 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1922 } 1923 1924 jint Deoptimization::total_deoptimization_count() { 1925 return _deoptimization_hist[Reason_none][0][0]; 1926 } 1927 1928 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1929 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1930 return _deoptimization_hist[reason][0][0]; 1931 } 1932 1933 void Deoptimization::print_statistics() { 1934 juint total = total_deoptimization_count(); 1935 juint account = total; 1936 if (total != 0) { 1937 ttyLocker ttyl; 1938 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1939 tty->print_cr("Deoptimization traps recorded:"); 1940 #define PRINT_STAT_LINE(name, r) \ 1941 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1942 PRINT_STAT_LINE("total", total); 1943 // For each non-zero entry in the histogram, print the reason, 1944 // the action, and (if specifically known) the type of bytecode. 1945 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1946 for (int action = 0; action < Action_LIMIT; action++) { 1947 juint* cases = _deoptimization_hist[reason][1+action]; 1948 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1949 juint counter = cases[bc_case]; 1950 if (counter != 0) { 1951 char name[1*K]; 1952 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1953 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1954 bc = Bytecodes::_illegal; 1955 sprintf(name, "%s/%s/%s", 1956 trap_reason_name(reason), 1957 trap_action_name(action), 1958 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1959 juint r = counter >> LSB_BITS; 1960 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1961 account -= r; 1962 } 1963 } 1964 } 1965 } 1966 if (account != 0) { 1967 PRINT_STAT_LINE("unaccounted", account); 1968 } 1969 #undef PRINT_STAT_LINE 1970 if (xtty != NULL) xtty->tail("statistics"); 1971 } 1972 } 1973 #else // COMPILER2 || SHARK 1974 1975 1976 // Stubs for C1 only system. 1977 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1978 return false; 1979 } 1980 1981 const char* Deoptimization::trap_reason_name(int reason) { 1982 return "unknown"; 1983 } 1984 1985 void Deoptimization::print_statistics() { 1986 // no output 1987 } 1988 1989 void 1990 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1991 // no udpate 1992 } 1993 1994 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1995 return 0; 1996 } 1997 1998 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1999 Bytecodes::Code bc) { 2000 // no update 2001 } 2002 2003 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2004 int trap_state) { 2005 jio_snprintf(buf, buflen, "#%d", trap_state); 2006 return buf; 2007 } 2008 2009 #endif // COMPILER2 || SHARK