| 1 | /* |
| 2 | * Copyright (c) 1997, 2019, 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 "jvm.h" |
| 27 | #include "classfile/symbolTable.hpp" |
| 28 | #include "classfile/systemDictionary.hpp" |
| 29 | #include "code/codeCache.hpp" |
| 30 | #include "code/debugInfoRec.hpp" |
| 31 | #include "code/nmethod.hpp" |
| 32 | #include "code/pcDesc.hpp" |
| 33 | #include "code/scopeDesc.hpp" |
| 34 | #include "interpreter/bytecode.hpp" |
| 35 | #include "interpreter/interpreter.hpp" |
| 36 | #include "interpreter/oopMapCache.hpp" |
| 37 | #include "memory/allocation.inline.hpp" |
| 38 | #include "memory/oopFactory.hpp" |
| 39 | #include "memory/resourceArea.hpp" |
| 40 | #include "memory/universe.hpp" |
| 41 | #include "oops/constantPool.hpp" |
| 42 | #include "oops/method.hpp" |
| 43 | #include "oops/objArrayKlass.hpp" |
| 44 | #include "oops/objArrayOop.inline.hpp" |
| 45 | #include "oops/oop.inline.hpp" |
| 46 | #include "oops/fieldStreams.hpp" |
| 47 | #include "oops/typeArrayOop.inline.hpp" |
| 48 | #include "oops/verifyOopClosure.hpp" |
| 49 | #include "prims/jvmtiThreadState.hpp" |
| 50 | #include "runtime/biasedLocking.hpp" |
| 51 | #include "runtime/compilationPolicy.hpp" |
| 52 | #include "runtime/deoptimization.hpp" |
| 53 | #include "runtime/fieldDescriptor.hpp" |
| 54 | #include "runtime/fieldDescriptor.inline.hpp" |
| 55 | #include "runtime/frame.inline.hpp" |
| 56 | #include "runtime/jniHandles.inline.hpp" |
| 57 | #include "runtime/handles.inline.hpp" |
| 58 | #include "runtime/interfaceSupport.inline.hpp" |
| 59 | #include "runtime/safepointVerifiers.hpp" |
| 60 | #include "runtime/sharedRuntime.hpp" |
| 61 | #include "runtime/signature.hpp" |
| 62 | #include "runtime/stubRoutines.hpp" |
| 63 | #include "runtime/thread.hpp" |
| 64 | #include "runtime/threadSMR.hpp" |
| 65 | #include "runtime/vframe.hpp" |
| 66 | #include "runtime/vframeArray.hpp" |
| 67 | #include "runtime/vframe_hp.hpp" |
| 68 | #include "utilities/events.hpp" |
| 69 | #include "utilities/preserveException.hpp" |
| 70 | #include "utilities/xmlstream.hpp" |
| 71 | |
| 72 | |
| 73 | bool DeoptimizationMarker::_is_active = false; |
| 74 | |
| 75 | Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, |
| 76 | int caller_adjustment, |
| 77 | int caller_actual_parameters, |
| 78 | int number_of_frames, |
| 79 | intptr_t* frame_sizes, |
| 80 | address* frame_pcs, |
| 81 | BasicType return_type, |
| 82 | int exec_mode) { |
| 83 | _size_of_deoptimized_frame = size_of_deoptimized_frame; |
| 84 | _caller_adjustment = caller_adjustment; |
| 85 | _caller_actual_parameters = caller_actual_parameters; |
| 86 | _number_of_frames = number_of_frames; |
| 87 | _frame_sizes = frame_sizes; |
| 88 | _frame_pcs = frame_pcs; |
| 89 | _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); |
| 90 | _return_type = return_type; |
| 91 | _initial_info = 0; |
| 92 | // PD (x86 only) |
| 93 | _counter_temp = 0; |
| 94 | _unpack_kind = exec_mode; |
| 95 | _sender_sp_temp = 0; |
| 96 | |
| 97 | _total_frame_sizes = size_of_frames(); |
| 98 | assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode" ); |
| 99 | } |
| 100 | |
| 101 | |
| 102 | Deoptimization::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 | |
| 109 | intptr_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 | |
| 116 | int 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 | |
| 126 | void 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(INTX_FORMAT " " , 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. |
| 143 | JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) |
| 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 | if (TraceDeoptimization) { |
| 153 | tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); |
| 154 | } |
| 155 | thread->inc_in_deopt_handler(); |
| 156 | |
| 157 | return fetch_unroll_info_helper(thread, exec_mode); |
| 158 | JRT_END |
| 159 | |
| 160 | |
| 161 | // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) |
| 162 | Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { |
| 163 | |
| 164 | // Note: there is a safepoint safety issue here. No matter whether we enter |
| 165 | // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once |
| 166 | // the vframeArray is created. |
| 167 | // |
| 168 | |
| 169 | // Allocate our special deoptimization ResourceMark |
| 170 | DeoptResourceMark* dmark = new DeoptResourceMark(thread); |
| 171 | assert(thread->deopt_mark() == NULL, "Pending deopt!" ); |
| 172 | thread->set_deopt_mark(dmark); |
| 173 | |
| 174 | frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect |
| 175 | RegisterMap map(thread, true); |
| 176 | RegisterMap dummy_map(thread, false); |
| 177 | // Now get the deoptee with a valid map |
| 178 | frame deoptee = stub_frame.sender(&map); |
| 179 | // Set the deoptee nmethod |
| 180 | assert(thread->deopt_compiled_method() == NULL, "Pending deopt!" ); |
| 181 | CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); |
| 182 | thread->set_deopt_compiled_method(cm); |
| 183 | |
| 184 | if (VerifyStack) { |
| 185 | thread->validate_frame_layout(); |
| 186 | } |
| 187 | |
| 188 | // Create a growable array of VFrames where each VFrame represents an inlined |
| 189 | // Java frame. This storage is allocated with the usual system arena. |
| 190 | assert(deoptee.is_compiled_frame(), "Wrong frame type" ); |
| 191 | GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); |
| 192 | vframe* vf = vframe::new_vframe(&deoptee, &map, thread); |
| 193 | while (!vf->is_top()) { |
| 194 | assert(vf->is_compiled_frame(), "Wrong frame type" ); |
| 195 | chunk->push(compiledVFrame::cast(vf)); |
| 196 | vf = vf->sender(); |
| 197 | } |
| 198 | assert(vf->is_compiled_frame(), "Wrong frame type" ); |
| 199 | chunk->push(compiledVFrame::cast(vf)); |
| 200 | |
| 201 | bool realloc_failures = false; |
| 202 | |
| 203 | #if COMPILER2_OR_JVMCI |
| 204 | // Reallocate the non-escaping objects and restore their fields. Then |
| 205 | // relock objects if synchronization on them was eliminated. |
| 206 | #if !INCLUDE_JVMCI |
| 207 | if (DoEscapeAnalysis || EliminateNestedLocks) { |
| 208 | if (EliminateAllocations) { |
| 209 | #endif // INCLUDE_JVMCI |
| 210 | assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames" ); |
| 211 | GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); |
| 212 | |
| 213 | // The flag return_oop() indicates call sites which return oop |
| 214 | // in compiled code. Such sites include java method calls, |
| 215 | // runtime calls (for example, used to allocate new objects/arrays |
| 216 | // on slow code path) and any other calls generated in compiled code. |
| 217 | // It is not guaranteed that we can get such information here only |
| 218 | // by analyzing bytecode in deoptimized frames. This is why this flag |
| 219 | // is set during method compilation (see Compile::Process_OopMap_Node()). |
| 220 | // If the previous frame was popped or if we are dispatching an exception, |
| 221 | // we don't have an oop result. |
| 222 | bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt); |
| 223 | Handle return_value; |
| 224 | if (save_oop_result) { |
| 225 | // Reallocation may trigger GC. If deoptimization happened on return from |
| 226 | // call which returns oop we need to save it since it is not in oopmap. |
| 227 | oop result = deoptee.saved_oop_result(&map); |
| 228 | assert(oopDesc::is_oop_or_null(result), "must be oop" ); |
| 229 | return_value = Handle(thread, result); |
| 230 | assert(Universe::heap()->is_in_or_null(result), "must be heap pointer" ); |
| 231 | if (TraceDeoptimization) { |
| 232 | ttyLocker ttyl; |
| 233 | tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); |
| 234 | } |
| 235 | } |
| 236 | if (objects != NULL) { |
| 237 | JRT_BLOCK |
| 238 | realloc_failures = realloc_objects(thread, &deoptee, &map, objects, THREAD); |
| 239 | JRT_END |
| 240 | bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci(); |
| 241 | reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); |
| 242 | #ifndef PRODUCT |
| 243 | if (TraceDeoptimization) { |
| 244 | ttyLocker ttyl; |
| 245 | tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); |
| 246 | print_objects(objects, realloc_failures); |
| 247 | } |
| 248 | #endif |
| 249 | } |
| 250 | if (save_oop_result) { |
| 251 | // Restore result. |
| 252 | deoptee.set_saved_oop_result(&map, return_value()); |
| 253 | } |
| 254 | #if !INCLUDE_JVMCI |
| 255 | } |
| 256 | if (EliminateLocks) { |
| 257 | #endif // INCLUDE_JVMCI |
| 258 | #ifndef PRODUCT |
| 259 | bool first = true; |
| 260 | #endif |
| 261 | for (int i = 0; i < chunk->length(); i++) { |
| 262 | compiledVFrame* cvf = chunk->at(i); |
| 263 | assert (cvf->scope() != NULL,"expect only compiled java frames" ); |
| 264 | GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
| 265 | if (monitors->is_nonempty()) { |
| 266 | relock_objects(monitors, thread, realloc_failures); |
| 267 | #ifndef PRODUCT |
| 268 | if (PrintDeoptimizationDetails) { |
| 269 | ttyLocker ttyl; |
| 270 | for (int j = 0; j < monitors->length(); j++) { |
| 271 | MonitorInfo* mi = monitors->at(j); |
| 272 | if (mi->eliminated()) { |
| 273 | if (first) { |
| 274 | first = false; |
| 275 | tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); |
| 276 | } |
| 277 | if (mi->owner_is_scalar_replaced()) { |
| 278 | Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); |
| 279 | tty->print_cr(" failed reallocation for klass %s" , k->external_name()); |
| 280 | } else { |
| 281 | tty->print_cr(" object <" INTPTR_FORMAT "> locked" , p2i(mi->owner())); |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | #endif // !PRODUCT |
| 287 | } |
| 288 | } |
| 289 | #if !INCLUDE_JVMCI |
| 290 | } |
| 291 | } |
| 292 | #endif // INCLUDE_JVMCI |
| 293 | #endif // COMPILER2_OR_JVMCI |
| 294 | |
| 295 | ScopeDesc* trap_scope = chunk->at(0)->scope(); |
| 296 | Handle exceptionObject; |
| 297 | if (trap_scope->rethrow_exception()) { |
| 298 | if (PrintDeoptimizationDetails) { |
| 299 | tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d" , trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci()); |
| 300 | } |
| 301 | GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); |
| 302 | guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw" ); |
| 303 | ScopeValue* topOfStack = expressions->top(); |
| 304 | exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); |
| 305 | guarantee(exceptionObject() != NULL, "exception oop can not be null" ); |
| 306 | } |
| 307 | |
| 308 | // Ensure that no safepoint is taken after pointers have been stored |
| 309 | // in fields of rematerialized objects. If a safepoint occurs from here on |
| 310 | // out the java state residing in the vframeArray will be missed. |
| 311 | NoSafepointVerifier no_safepoint; |
| 312 | |
| 313 | vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); |
| 314 | #if COMPILER2_OR_JVMCI |
| 315 | if (realloc_failures) { |
| 316 | pop_frames_failed_reallocs(thread, array); |
| 317 | } |
| 318 | #endif |
| 319 | |
| 320 | assert(thread->vframe_array_head() == NULL, "Pending deopt!" ); |
| 321 | thread->set_vframe_array_head(array); |
| 322 | |
| 323 | // Now that the vframeArray has been created if we have any deferred local writes |
| 324 | // added by jvmti then we can free up that structure as the data is now in the |
| 325 | // vframeArray |
| 326 | |
| 327 | if (thread->deferred_locals() != NULL) { |
| 328 | GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); |
| 329 | int i = 0; |
| 330 | do { |
| 331 | // Because of inlining we could have multiple vframes for a single frame |
| 332 | // and several of the vframes could have deferred writes. Find them all. |
| 333 | if (list->at(i)->id() == array->original().id()) { |
| 334 | jvmtiDeferredLocalVariableSet* dlv = list->at(i); |
| 335 | list->remove_at(i); |
| 336 | // individual jvmtiDeferredLocalVariableSet are CHeapObj's |
| 337 | delete dlv; |
| 338 | } else { |
| 339 | i++; |
| 340 | } |
| 341 | } while ( i < list->length() ); |
| 342 | if (list->length() == 0) { |
| 343 | thread->set_deferred_locals(NULL); |
| 344 | // free the list and elements back to C heap. |
| 345 | delete list; |
| 346 | } |
| 347 | |
| 348 | } |
| 349 | |
| 350 | // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. |
| 351 | CodeBlob* cb = stub_frame.cb(); |
| 352 | // Verify we have the right vframeArray |
| 353 | assert(cb->frame_size() >= 0, "Unexpected frame size" ); |
| 354 | intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); |
| 355 | |
| 356 | // If the deopt call site is a MethodHandle invoke call site we have |
| 357 | // to adjust the unpack_sp. |
| 358 | nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); |
| 359 | if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) |
| 360 | unpack_sp = deoptee.unextended_sp(); |
| 361 | |
| 362 | #ifdef ASSERT |
| 363 | assert(cb->is_deoptimization_stub() || |
| 364 | cb->is_uncommon_trap_stub() || |
| 365 | strcmp("Stub<DeoptimizationStub.deoptimizationHandler>" , cb->name()) == 0 || |
| 366 | strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>" , cb->name()) == 0, |
| 367 | "unexpected code blob: %s" , cb->name()); |
| 368 | #endif |
| 369 | |
| 370 | // This is a guarantee instead of an assert because if vframe doesn't match |
| 371 | // we will unpack the wrong deoptimized frame and wind up in strange places |
| 372 | // where it will be very difficult to figure out what went wrong. Better |
| 373 | // to die an early death here than some very obscure death later when the |
| 374 | // trail is cold. |
| 375 | // Note: on ia64 this guarantee can be fooled by frames with no memory stack |
| 376 | // in that it will fail to detect a problem when there is one. This needs |
| 377 | // more work in tiger timeframe. |
| 378 | guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack" ); |
| 379 | |
| 380 | int number_of_frames = array->frames(); |
| 381 | |
| 382 | // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost |
| 383 | // virtual activation, which is the reverse of the elements in the vframes array. |
| 384 | intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); |
| 385 | // +1 because we always have an interpreter return address for the final slot. |
| 386 | address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); |
| 387 | int = 0; |
| 388 | // Create an interpreter return address for the stub to use as its return |
| 389 | // address so the skeletal frames are perfectly walkable |
| 390 | frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); |
| 391 | |
| 392 | // PopFrame requires that the preserved incoming arguments from the recently-popped topmost |
| 393 | // activation be put back on the expression stack of the caller for reexecution |
| 394 | if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { |
| 395 | popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); |
| 396 | } |
| 397 | |
| 398 | // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized |
| 399 | // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather |
| 400 | // than simply use array->sender.pc(). This requires us to walk the current set of frames |
| 401 | // |
| 402 | frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame |
| 403 | deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller |
| 404 | |
| 405 | // It's possible that the number of parameters at the call site is |
| 406 | // different than number of arguments in the callee when method |
| 407 | // handles are used. If the caller is interpreted get the real |
| 408 | // value so that the proper amount of space can be added to it's |
| 409 | // frame. |
| 410 | bool caller_was_method_handle = false; |
| 411 | if (deopt_sender.is_interpreted_frame()) { |
| 412 | methodHandle method = deopt_sender.interpreter_frame_method(); |
| 413 | Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); |
| 414 | if (cur.is_invokedynamic() || cur.is_invokehandle()) { |
| 415 | // Method handle invokes may involve fairly arbitrary chains of |
| 416 | // calls so it's impossible to know how much actual space the |
| 417 | // caller has for locals. |
| 418 | caller_was_method_handle = true; |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | // |
| 423 | // frame_sizes/frame_pcs[0] oldest frame (int or c2i) |
| 424 | // frame_sizes/frame_pcs[1] next oldest frame (int) |
| 425 | // frame_sizes/frame_pcs[n] youngest frame (int) |
| 426 | // |
| 427 | // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame |
| 428 | // owns the space for the return address to it's caller). Confusing ain't it. |
| 429 | // |
| 430 | // The vframe array can address vframes with indices running from |
| 431 | // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. |
| 432 | // When we create the skeletal frames we need the oldest frame to be in the zero slot |
| 433 | // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. |
| 434 | // so things look a little strange in this loop. |
| 435 | // |
| 436 | int callee_parameters = 0; |
| 437 | int callee_locals = 0; |
| 438 | for (int index = 0; index < array->frames(); index++ ) { |
| 439 | // frame[number_of_frames - 1 ] = on_stack_size(youngest) |
| 440 | // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) |
| 441 | // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) |
| 442 | frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, |
| 443 | callee_locals, |
| 444 | index == 0, |
| 445 | popframe_extra_args); |
| 446 | // This pc doesn't have to be perfect just good enough to identify the frame |
| 447 | // as interpreted so the skeleton frame will be walkable |
| 448 | // The correct pc will be set when the skeleton frame is completely filled out |
| 449 | // The final pc we store in the loop is wrong and will be overwritten below |
| 450 | frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; |
| 451 | |
| 452 | callee_parameters = array->element(index)->method()->size_of_parameters(); |
| 453 | callee_locals = array->element(index)->method()->max_locals(); |
| 454 | popframe_extra_args = 0; |
| 455 | } |
| 456 | |
| 457 | // Compute whether the root vframe returns a float or double value. |
| 458 | BasicType return_type; |
| 459 | { |
| 460 | methodHandle method(thread, array->element(0)->method()); |
| 461 | Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); |
| 462 | return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; |
| 463 | } |
| 464 | |
| 465 | // Compute information for handling adapters and adjusting the frame size of the caller. |
| 466 | int caller_adjustment = 0; |
| 467 | |
| 468 | // Compute the amount the oldest interpreter frame will have to adjust |
| 469 | // its caller's stack by. If the caller is a compiled frame then |
| 470 | // we pretend that the callee has no parameters so that the |
| 471 | // extension counts for the full amount of locals and not just |
| 472 | // locals-parms. This is because without a c2i adapter the parm |
| 473 | // area as created by the compiled frame will not be usable by |
| 474 | // the interpreter. (Depending on the calling convention there |
| 475 | // may not even be enough space). |
| 476 | |
| 477 | // QQQ I'd rather see this pushed down into last_frame_adjust |
| 478 | // and have it take the sender (aka caller). |
| 479 | |
| 480 | if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { |
| 481 | caller_adjustment = last_frame_adjust(0, callee_locals); |
| 482 | } else if (callee_locals > callee_parameters) { |
| 483 | // The caller frame may need extending to accommodate |
| 484 | // non-parameter locals of the first unpacked interpreted frame. |
| 485 | // Compute that adjustment. |
| 486 | caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); |
| 487 | } |
| 488 | |
| 489 | // If the sender is deoptimized the we must retrieve the address of the handler |
| 490 | // since the frame will "magically" show the original pc before the deopt |
| 491 | // and we'd undo the deopt. |
| 492 | |
| 493 | frame_pcs[0] = deopt_sender.raw_pc(); |
| 494 | |
| 495 | assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc" ); |
| 496 | |
| 497 | #if INCLUDE_JVMCI |
| 498 | if (exceptionObject() != NULL) { |
| 499 | thread->set_exception_oop(exceptionObject()); |
| 500 | exec_mode = Unpack_exception; |
| 501 | } |
| 502 | #endif |
| 503 | |
| 504 | if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { |
| 505 | assert(thread->has_pending_exception(), "should have thrown OOME" ); |
| 506 | thread->set_exception_oop(thread->pending_exception()); |
| 507 | thread->clear_pending_exception(); |
| 508 | exec_mode = Unpack_exception; |
| 509 | } |
| 510 | |
| 511 | #if INCLUDE_JVMCI |
| 512 | if (thread->frames_to_pop_failed_realloc() > 0) { |
| 513 | thread->set_pending_monitorenter(false); |
| 514 | } |
| 515 | #endif |
| 516 | |
| 517 | UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, |
| 518 | caller_adjustment * BytesPerWord, |
| 519 | caller_was_method_handle ? 0 : callee_parameters, |
| 520 | number_of_frames, |
| 521 | frame_sizes, |
| 522 | frame_pcs, |
| 523 | return_type, |
| 524 | exec_mode); |
| 525 | // On some platforms, we need a way to pass some platform dependent |
| 526 | // information to the unpacking code so the skeletal frames come out |
| 527 | // correct (initial fp value, unextended sp, ...) |
| 528 | info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); |
| 529 | |
| 530 | if (array->frames() > 1) { |
| 531 | if (VerifyStack && TraceDeoptimization) { |
| 532 | ttyLocker ttyl; |
| 533 | tty->print_cr("Deoptimizing method containing inlining" ); |
| 534 | } |
| 535 | } |
| 536 | |
| 537 | array->set_unroll_block(info); |
| 538 | return info; |
| 539 | } |
| 540 | |
| 541 | // Called to cleanup deoptimization data structures in normal case |
| 542 | // after unpacking to stack and when stack overflow error occurs |
| 543 | void Deoptimization::cleanup_deopt_info(JavaThread *thread, |
| 544 | vframeArray *array) { |
| 545 | |
| 546 | // Get array if coming from exception |
| 547 | if (array == NULL) { |
| 548 | array = thread->vframe_array_head(); |
| 549 | } |
| 550 | thread->set_vframe_array_head(NULL); |
| 551 | |
| 552 | // Free the previous UnrollBlock |
| 553 | vframeArray* old_array = thread->vframe_array_last(); |
| 554 | thread->set_vframe_array_last(array); |
| 555 | |
| 556 | if (old_array != NULL) { |
| 557 | UnrollBlock* old_info = old_array->unroll_block(); |
| 558 | old_array->set_unroll_block(NULL); |
| 559 | delete old_info; |
| 560 | delete old_array; |
| 561 | } |
| 562 | |
| 563 | // Deallocate any resource creating in this routine and any ResourceObjs allocated |
| 564 | // inside the vframeArray (StackValueCollections) |
| 565 | |
| 566 | delete thread->deopt_mark(); |
| 567 | thread->set_deopt_mark(NULL); |
| 568 | thread->set_deopt_compiled_method(NULL); |
| 569 | |
| 570 | |
| 571 | if (JvmtiExport::can_pop_frame()) { |
| 572 | #ifndef CC_INTERP |
| 573 | // Regardless of whether we entered this routine with the pending |
| 574 | // popframe condition bit set, we should always clear it now |
| 575 | thread->clear_popframe_condition(); |
| 576 | #else |
| 577 | // C++ interpreter will clear has_pending_popframe when it enters |
| 578 | // with method_resume. For deopt_resume2 we clear it now. |
| 579 | if (thread->popframe_forcing_deopt_reexecution()) |
| 580 | thread->clear_popframe_condition(); |
| 581 | #endif /* CC_INTERP */ |
| 582 | } |
| 583 | |
| 584 | // unpack_frames() is called at the end of the deoptimization handler |
| 585 | // and (in C2) at the end of the uncommon trap handler. Note this fact |
| 586 | // so that an asynchronous stack walker can work again. This counter is |
| 587 | // incremented at the beginning of fetch_unroll_info() and (in C2) at |
| 588 | // the beginning of uncommon_trap(). |
| 589 | thread->dec_in_deopt_handler(); |
| 590 | } |
| 591 | |
| 592 | // Moved from cpu directories because none of the cpus has callee save values. |
| 593 | // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. |
| 594 | void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { |
| 595 | |
| 596 | // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in |
| 597 | // the days we had adapter frames. When we deoptimize a situation where a |
| 598 | // compiled caller calls a compiled caller will have registers it expects |
| 599 | // to survive the call to the callee. If we deoptimize the callee the only |
| 600 | // way we can restore these registers is to have the oldest interpreter |
| 601 | // frame that we create restore these values. That is what this routine |
| 602 | // will accomplish. |
| 603 | |
| 604 | // At the moment we have modified c2 to not have any callee save registers |
| 605 | // so this problem does not exist and this routine is just a place holder. |
| 606 | |
| 607 | assert(f->is_interpreted_frame(), "must be interpreted" ); |
| 608 | } |
| 609 | |
| 610 | // Return BasicType of value being returned |
| 611 | JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) |
| 612 | |
| 613 | // We are already active in the special DeoptResourceMark any ResourceObj's we |
| 614 | // allocate will be freed at the end of the routine. |
| 615 | |
| 616 | // It is actually ok to allocate handles in a leaf method. It causes no safepoints, |
| 617 | // but makes the entry a little slower. There is however a little dance we have to |
| 618 | // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro |
| 619 | ResetNoHandleMark rnhm; // No-op in release/product versions |
| 620 | HandleMark hm; |
| 621 | |
| 622 | frame stub_frame = thread->last_frame(); |
| 623 | |
| 624 | // Since the frame to unpack is the top frame of this thread, the vframe_array_head |
| 625 | // must point to the vframeArray for the unpack frame. |
| 626 | vframeArray* array = thread->vframe_array_head(); |
| 627 | |
| 628 | #ifndef PRODUCT |
| 629 | if (TraceDeoptimization) { |
| 630 | ttyLocker ttyl; |
| 631 | tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d" , |
| 632 | p2i(thread), p2i(array), exec_mode); |
| 633 | } |
| 634 | #endif |
| 635 | Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d" , |
| 636 | p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); |
| 637 | |
| 638 | UnrollBlock* info = array->unroll_block(); |
| 639 | |
| 640 | // Unpack the interpreter frames and any adapter frame (c2 only) we might create. |
| 641 | array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); |
| 642 | |
| 643 | BasicType bt = info->return_type(); |
| 644 | |
| 645 | // If we have an exception pending, claim that the return type is an oop |
| 646 | // so the deopt_blob does not overwrite the exception_oop. |
| 647 | |
| 648 | if (exec_mode == Unpack_exception) |
| 649 | bt = T_OBJECT; |
| 650 | |
| 651 | // Cleanup thread deopt data |
| 652 | cleanup_deopt_info(thread, array); |
| 653 | |
| 654 | #ifndef PRODUCT |
| 655 | if (VerifyStack) { |
| 656 | ResourceMark res_mark; |
| 657 | // Clear pending exception to not break verification code (restored afterwards) |
| 658 | PRESERVE_EXCEPTION_MARK; |
| 659 | |
| 660 | thread->validate_frame_layout(); |
| 661 | |
| 662 | // Verify that the just-unpacked frames match the interpreter's |
| 663 | // notions of expression stack and locals |
| 664 | vframeArray* cur_array = thread->vframe_array_last(); |
| 665 | RegisterMap rm(thread, false); |
| 666 | rm.set_include_argument_oops(false); |
| 667 | bool is_top_frame = true; |
| 668 | int callee_size_of_parameters = 0; |
| 669 | int callee_max_locals = 0; |
| 670 | for (int i = 0; i < cur_array->frames(); i++) { |
| 671 | vframeArrayElement* el = cur_array->element(i); |
| 672 | frame* iframe = el->iframe(); |
| 673 | guarantee(iframe->is_interpreted_frame(), "Wrong frame type" ); |
| 674 | |
| 675 | // Get the oop map for this bci |
| 676 | InterpreterOopMap mask; |
| 677 | int cur_invoke_parameter_size = 0; |
| 678 | bool try_next_mask = false; |
| 679 | int next_mask_expression_stack_size = -1; |
| 680 | int top_frame_expression_stack_adjustment = 0; |
| 681 | methodHandle mh(thread, iframe->interpreter_frame_method()); |
| 682 | OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); |
| 683 | BytecodeStream str(mh, iframe->interpreter_frame_bci()); |
| 684 | int max_bci = mh->code_size(); |
| 685 | // Get to the next bytecode if possible |
| 686 | assert(str.bci() < max_bci, "bci in interpreter frame out of bounds" ); |
| 687 | // Check to see if we can grab the number of outgoing arguments |
| 688 | // at an uncommon trap for an invoke (where the compiler |
| 689 | // generates debug info before the invoke has executed) |
| 690 | Bytecodes::Code cur_code = str.next(); |
| 691 | if (Bytecodes::is_invoke(cur_code)) { |
| 692 | Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); |
| 693 | cur_invoke_parameter_size = invoke.size_of_parameters(); |
| 694 | if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { |
| 695 | callee_size_of_parameters++; |
| 696 | } |
| 697 | } |
| 698 | if (str.bci() < max_bci) { |
| 699 | Bytecodes::Code next_code = str.next(); |
| 700 | if (next_code >= 0) { |
| 701 | // The interpreter oop map generator reports results before |
| 702 | // the current bytecode has executed except in the case of |
| 703 | // calls. It seems to be hard to tell whether the compiler |
| 704 | // has emitted debug information matching the "state before" |
| 705 | // a given bytecode or the state after, so we try both |
| 706 | if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { |
| 707 | // Get expression stack size for the next bytecode |
| 708 | InterpreterOopMap next_mask; |
| 709 | OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); |
| 710 | next_mask_expression_stack_size = next_mask.expression_stack_size(); |
| 711 | if (Bytecodes::is_invoke(next_code)) { |
| 712 | Bytecode_invoke invoke(mh, str.bci()); |
| 713 | next_mask_expression_stack_size += invoke.size_of_parameters(); |
| 714 | } |
| 715 | // Need to subtract off the size of the result type of |
| 716 | // the bytecode because this is not described in the |
| 717 | // debug info but returned to the interpreter in the TOS |
| 718 | // caching register |
| 719 | BasicType bytecode_result_type = Bytecodes::result_type(cur_code); |
| 720 | if (bytecode_result_type != T_ILLEGAL) { |
| 721 | top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; |
| 722 | } |
| 723 | assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive" ); |
| 724 | try_next_mask = true; |
| 725 | } |
| 726 | } |
| 727 | } |
| 728 | |
| 729 | // Verify stack depth and oops in frame |
| 730 | // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) |
| 731 | if (!( |
| 732 | /* SPARC */ |
| 733 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || |
| 734 | /* x86 */ |
| 735 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || |
| 736 | (try_next_mask && |
| 737 | (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - |
| 738 | top_frame_expression_stack_adjustment))) || |
| 739 | (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || |
| 740 | (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && |
| 741 | (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) |
| 742 | )) { |
| 743 | { |
| 744 | ttyLocker ttyl; |
| 745 | |
| 746 | // Print out some information that will help us debug the problem |
| 747 | tty->print_cr("Wrong number of expression stack elements during deoptimization" ); |
| 748 | tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)" , i, cur_array->frames() - 1); |
| 749 | tty->print_cr(" Fabricated interpreter frame had %d expression stack elements" , |
| 750 | iframe->interpreter_frame_expression_stack_size()); |
| 751 | tty->print_cr(" Interpreter oop map had %d expression stack elements" , mask.expression_stack_size()); |
| 752 | tty->print_cr(" try_next_mask = %d" , try_next_mask); |
| 753 | tty->print_cr(" next_mask_expression_stack_size = %d" , next_mask_expression_stack_size); |
| 754 | tty->print_cr(" callee_size_of_parameters = %d" , callee_size_of_parameters); |
| 755 | tty->print_cr(" callee_max_locals = %d" , callee_max_locals); |
| 756 | tty->print_cr(" top_frame_expression_stack_adjustment = %d" , top_frame_expression_stack_adjustment); |
| 757 | tty->print_cr(" exec_mode = %d" , exec_mode); |
| 758 | tty->print_cr(" cur_invoke_parameter_size = %d" , cur_invoke_parameter_size); |
| 759 | tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d" , p2i(thread), thread->osthread()->thread_id()); |
| 760 | tty->print_cr(" Interpreted frames:" ); |
| 761 | for (int k = 0; k < cur_array->frames(); k++) { |
| 762 | vframeArrayElement* el = cur_array->element(k); |
| 763 | tty->print_cr(" %s (bci %d)" , el->method()->name_and_sig_as_C_string(), el->bci()); |
| 764 | } |
| 765 | cur_array->print_on_2(tty); |
| 766 | } // release tty lock before calling guarantee |
| 767 | guarantee(false, "wrong number of expression stack elements during deopt" ); |
| 768 | } |
| 769 | VerifyOopClosure verify; |
| 770 | iframe->oops_interpreted_do(&verify, &rm, false); |
| 771 | callee_size_of_parameters = mh->size_of_parameters(); |
| 772 | callee_max_locals = mh->max_locals(); |
| 773 | is_top_frame = false; |
| 774 | } |
| 775 | } |
| 776 | #endif /* !PRODUCT */ |
| 777 | |
| 778 | |
| 779 | return bt; |
| 780 | JRT_END |
| 781 | |
| 782 | |
| 783 | int Deoptimization::deoptimize_dependents() { |
| 784 | Threads::deoptimized_wrt_marked_nmethods(); |
| 785 | return 0; |
| 786 | } |
| 787 | |
| 788 | Deoptimization::DeoptAction Deoptimization::_unloaded_action |
| 789 | = Deoptimization::Action_reinterpret; |
| 790 | |
| 791 | |
| 792 | |
| 793 | #if INCLUDE_JVMCI || INCLUDE_AOT |
| 794 | template<typename CacheType> |
| 795 | class BoxCacheBase : public CHeapObj<mtCompiler> { |
| 796 | protected: |
| 797 | static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { |
| 798 | ResourceMark rm; |
| 799 | char* klass_name_str = klass_name->as_C_string(); |
| 800 | Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); |
| 801 | guarantee(k != NULL, "%s must be loaded" , klass_name_str); |
| 802 | InstanceKlass* ik = InstanceKlass::cast(k); |
| 803 | guarantee(ik->is_initialized(), "%s must be initialized" , klass_name_str); |
| 804 | CacheType::compute_offsets(ik); |
| 805 | return ik; |
| 806 | } |
| 807 | }; |
| 808 | |
| 809 | template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { |
| 810 | PrimitiveType _low; |
| 811 | PrimitiveType _high; |
| 812 | jobject _cache; |
| 813 | protected: |
| 814 | static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; |
| 815 | BoxCache(Thread* thread) { |
| 816 | InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); |
| 817 | objArrayOop cache = CacheType::cache(ik); |
| 818 | assert(cache->length() > 0, "Empty cache" ); |
| 819 | _low = BoxType::value(cache->obj_at(0)); |
| 820 | _high = _low + cache->length() - 1; |
| 821 | _cache = JNIHandles::make_global(Handle(thread, cache)); |
| 822 | } |
| 823 | ~BoxCache() { |
| 824 | JNIHandles::destroy_global(_cache); |
| 825 | } |
| 826 | public: |
| 827 | static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { |
| 828 | if (_singleton == NULL) { |
| 829 | BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); |
| 830 | if (!Atomic::replace_if_null(s, &_singleton)) { |
| 831 | delete s; |
| 832 | } |
| 833 | } |
| 834 | return _singleton; |
| 835 | } |
| 836 | oop lookup(PrimitiveType value) { |
| 837 | if (_low <= value && value <= _high) { |
| 838 | int offset = value - _low; |
| 839 | return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); |
| 840 | } |
| 841 | return NULL; |
| 842 | } |
| 843 | oop lookup_raw(intptr_t raw_value) { |
| 844 | // Have to cast to avoid little/big-endian problems. |
| 845 | if (sizeof(PrimitiveType) > sizeof(jint)) { |
| 846 | jlong value = (jlong)raw_value; |
| 847 | return lookup(value); |
| 848 | } |
| 849 | PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); |
| 850 | return lookup(value); |
| 851 | } |
| 852 | }; |
| 853 | |
| 854 | typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; |
| 855 | typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; |
| 856 | typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; |
| 857 | typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; |
| 858 | typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; |
| 859 | |
| 860 | template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; |
| 861 | template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; |
| 862 | template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; |
| 863 | template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; |
| 864 | template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; |
| 865 | |
| 866 | class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { |
| 867 | jobject _true_cache; |
| 868 | jobject _false_cache; |
| 869 | protected: |
| 870 | static BooleanBoxCache *_singleton; |
| 871 | BooleanBoxCache(Thread *thread) { |
| 872 | InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); |
| 873 | _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); |
| 874 | _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); |
| 875 | } |
| 876 | ~BooleanBoxCache() { |
| 877 | JNIHandles::destroy_global(_true_cache); |
| 878 | JNIHandles::destroy_global(_false_cache); |
| 879 | } |
| 880 | public: |
| 881 | static BooleanBoxCache* singleton(Thread* thread) { |
| 882 | if (_singleton == NULL) { |
| 883 | BooleanBoxCache* s = new BooleanBoxCache(thread); |
| 884 | if (!Atomic::replace_if_null(s, &_singleton)) { |
| 885 | delete s; |
| 886 | } |
| 887 | } |
| 888 | return _singleton; |
| 889 | } |
| 890 | oop lookup_raw(intptr_t raw_value) { |
| 891 | // Have to cast to avoid little/big-endian problems. |
| 892 | jboolean value = (jboolean)*((jint*)&raw_value); |
| 893 | return lookup(value); |
| 894 | } |
| 895 | oop lookup(jboolean value) { |
| 896 | if (value != 0) { |
| 897 | return JNIHandles::resolve_non_null(_true_cache); |
| 898 | } |
| 899 | return JNIHandles::resolve_non_null(_false_cache); |
| 900 | } |
| 901 | }; |
| 902 | |
| 903 | BooleanBoxCache* BooleanBoxCache::_singleton = NULL; |
| 904 | |
| 905 | oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { |
| 906 | Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); |
| 907 | BasicType box_type = SystemDictionary::box_klass_type(k); |
| 908 | if (box_type != T_OBJECT) { |
| 909 | StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); |
| 910 | switch(box_type) { |
| 911 | case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 912 | case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 913 | case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 914 | case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 915 | case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 916 | case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); |
| 917 | default:; |
| 918 | } |
| 919 | } |
| 920 | return NULL; |
| 921 | } |
| 922 | #endif // INCLUDE_JVMCI || INCLUDE_AOT |
| 923 | |
| 924 | #if COMPILER2_OR_JVMCI |
| 925 | bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { |
| 926 | Handle pending_exception(THREAD, thread->pending_exception()); |
| 927 | const char* exception_file = thread->exception_file(); |
| 928 | int exception_line = thread->exception_line(); |
| 929 | thread->clear_pending_exception(); |
| 930 | |
| 931 | bool failures = false; |
| 932 | |
| 933 | for (int i = 0; i < objects->length(); i++) { |
| 934 | assert(objects->at(i)->is_object(), "invalid debug information" ); |
| 935 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
| 936 | |
| 937 | Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); |
| 938 | oop obj = NULL; |
| 939 | |
| 940 | if (k->is_instance_klass()) { |
| 941 | #if INCLUDE_JVMCI || INCLUDE_AOT |
| 942 | CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); |
| 943 | if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { |
| 944 | AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; |
| 945 | obj = get_cached_box(abv, fr, reg_map, THREAD); |
| 946 | if (obj != NULL) { |
| 947 | // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. |
| 948 | abv->set_cached(true); |
| 949 | } |
| 950 | } |
| 951 | #endif // INCLUDE_JVMCI || INCLUDE_AOT |
| 952 | InstanceKlass* ik = InstanceKlass::cast(k); |
| 953 | if (obj == NULL) { |
| 954 | obj = ik->allocate_instance(THREAD); |
| 955 | } |
| 956 | } else if (k->is_typeArray_klass()) { |
| 957 | TypeArrayKlass* ak = TypeArrayKlass::cast(k); |
| 958 | assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length" ); |
| 959 | int len = sv->field_size() / type2size[ak->element_type()]; |
| 960 | obj = ak->allocate(len, THREAD); |
| 961 | } else if (k->is_objArray_klass()) { |
| 962 | ObjArrayKlass* ak = ObjArrayKlass::cast(k); |
| 963 | obj = ak->allocate(sv->field_size(), THREAD); |
| 964 | } |
| 965 | |
| 966 | if (obj == NULL) { |
| 967 | failures = true; |
| 968 | } |
| 969 | |
| 970 | assert(sv->value().is_null(), "redundant reallocation" ); |
| 971 | assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception" ); |
| 972 | CLEAR_PENDING_EXCEPTION; |
| 973 | sv->set_value(obj); |
| 974 | } |
| 975 | |
| 976 | if (failures) { |
| 977 | THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); |
| 978 | } else if (pending_exception.not_null()) { |
| 979 | thread->set_pending_exception(pending_exception(), exception_file, exception_line); |
| 980 | } |
| 981 | |
| 982 | return failures; |
| 983 | } |
| 984 | |
| 985 | // restore elements of an eliminated type array |
| 986 | void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { |
| 987 | int index = 0; |
| 988 | intptr_t val; |
| 989 | |
| 990 | for (int i = 0; i < sv->field_size(); i++) { |
| 991 | StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); |
| 992 | switch(type) { |
| 993 | case T_LONG: case T_DOUBLE: { |
| 994 | assert(value->type() == T_INT, "Agreement." ); |
| 995 | StackValue* low = |
| 996 | StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); |
| 997 | #ifdef _LP64 |
| 998 | jlong res = (jlong)low->get_int(); |
| 999 | #else |
| 1000 | #ifdef SPARC |
| 1001 | // For SPARC we have to swap high and low words. |
| 1002 | jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
| 1003 | #else |
| 1004 | jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
| 1005 | #endif //SPARC |
| 1006 | #endif |
| 1007 | obj->long_at_put(index, res); |
| 1008 | break; |
| 1009 | } |
| 1010 | |
| 1011 | // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
| 1012 | case T_INT: case T_FLOAT: { // 4 bytes. |
| 1013 | assert(value->type() == T_INT, "Agreement." ); |
| 1014 | bool big_value = false; |
| 1015 | if (i + 1 < sv->field_size() && type == T_INT) { |
| 1016 | if (sv->field_at(i)->is_location()) { |
| 1017 | Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); |
| 1018 | if (type == Location::dbl || type == Location::lng) { |
| 1019 | big_value = true; |
| 1020 | } |
| 1021 | } else if (sv->field_at(i)->is_constant_int()) { |
| 1022 | ScopeValue* next_scope_field = sv->field_at(i + 1); |
| 1023 | if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { |
| 1024 | big_value = true; |
| 1025 | } |
| 1026 | } |
| 1027 | } |
| 1028 | |
| 1029 | if (big_value) { |
| 1030 | StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); |
| 1031 | #ifdef _LP64 |
| 1032 | jlong res = (jlong)low->get_int(); |
| 1033 | #else |
| 1034 | #ifdef SPARC |
| 1035 | // For SPARC we have to swap high and low words. |
| 1036 | jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
| 1037 | #else |
| 1038 | jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
| 1039 | #endif //SPARC |
| 1040 | #endif |
| 1041 | obj->int_at_put(index, (jint)*((jint*)&res)); |
| 1042 | obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); |
| 1043 | } else { |
| 1044 | val = value->get_int(); |
| 1045 | obj->int_at_put(index, (jint)*((jint*)&val)); |
| 1046 | } |
| 1047 | break; |
| 1048 | } |
| 1049 | |
| 1050 | case T_SHORT: |
| 1051 | assert(value->type() == T_INT, "Agreement." ); |
| 1052 | val = value->get_int(); |
| 1053 | obj->short_at_put(index, (jshort)*((jint*)&val)); |
| 1054 | break; |
| 1055 | |
| 1056 | case T_CHAR: |
| 1057 | assert(value->type() == T_INT, "Agreement." ); |
| 1058 | val = value->get_int(); |
| 1059 | obj->char_at_put(index, (jchar)*((jint*)&val)); |
| 1060 | break; |
| 1061 | |
| 1062 | case T_BYTE: |
| 1063 | assert(value->type() == T_INT, "Agreement." ); |
| 1064 | val = value->get_int(); |
| 1065 | obj->byte_at_put(index, (jbyte)*((jint*)&val)); |
| 1066 | break; |
| 1067 | |
| 1068 | case T_BOOLEAN: |
| 1069 | assert(value->type() == T_INT, "Agreement." ); |
| 1070 | val = value->get_int(); |
| 1071 | obj->bool_at_put(index, (jboolean)*((jint*)&val)); |
| 1072 | break; |
| 1073 | |
| 1074 | default: |
| 1075 | ShouldNotReachHere(); |
| 1076 | } |
| 1077 | index++; |
| 1078 | } |
| 1079 | } |
| 1080 | |
| 1081 | |
| 1082 | // restore fields of an eliminated object array |
| 1083 | void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { |
| 1084 | for (int i = 0; i < sv->field_size(); i++) { |
| 1085 | StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); |
| 1086 | assert(value->type() == T_OBJECT, "object element expected" ); |
| 1087 | obj->obj_at_put(i, value->get_obj()()); |
| 1088 | } |
| 1089 | } |
| 1090 | |
| 1091 | class ReassignedField { |
| 1092 | public: |
| 1093 | int _offset; |
| 1094 | BasicType _type; |
| 1095 | public: |
| 1096 | ReassignedField() { |
| 1097 | _offset = 0; |
| 1098 | _type = T_ILLEGAL; |
| 1099 | } |
| 1100 | }; |
| 1101 | |
| 1102 | int compare(ReassignedField* left, ReassignedField* right) { |
| 1103 | return left->_offset - right->_offset; |
| 1104 | } |
| 1105 | |
| 1106 | // Restore fields of an eliminated instance object using the same field order |
| 1107 | // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) |
| 1108 | static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { |
| 1109 | if (klass->superklass() != NULL) { |
| 1110 | svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); |
| 1111 | } |
| 1112 | |
| 1113 | GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); |
| 1114 | for (AllFieldStream fs(klass); !fs.done(); fs.next()) { |
| 1115 | if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { |
| 1116 | ReassignedField field; |
| 1117 | field._offset = fs.offset(); |
| 1118 | field._type = FieldType::basic_type(fs.signature()); |
| 1119 | fields->append(field); |
| 1120 | } |
| 1121 | } |
| 1122 | fields->sort(compare); |
| 1123 | for (int i = 0; i < fields->length(); i++) { |
| 1124 | intptr_t val; |
| 1125 | ScopeValue* scope_field = sv->field_at(svIndex); |
| 1126 | StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); |
| 1127 | int offset = fields->at(i)._offset; |
| 1128 | BasicType type = fields->at(i)._type; |
| 1129 | switch (type) { |
| 1130 | case T_OBJECT: case T_ARRAY: |
| 1131 | assert(value->type() == T_OBJECT, "Agreement." ); |
| 1132 | obj->obj_field_put(offset, value->get_obj()()); |
| 1133 | break; |
| 1134 | |
| 1135 | // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
| 1136 | case T_INT: case T_FLOAT: { // 4 bytes. |
| 1137 | assert(value->type() == T_INT, "Agreement." ); |
| 1138 | bool big_value = false; |
| 1139 | if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { |
| 1140 | if (scope_field->is_location()) { |
| 1141 | Location::Type type = ((LocationValue*) scope_field)->location().type(); |
| 1142 | if (type == Location::dbl || type == Location::lng) { |
| 1143 | big_value = true; |
| 1144 | } |
| 1145 | } |
| 1146 | if (scope_field->is_constant_int()) { |
| 1147 | ScopeValue* next_scope_field = sv->field_at(svIndex + 1); |
| 1148 | if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { |
| 1149 | big_value = true; |
| 1150 | } |
| 1151 | } |
| 1152 | } |
| 1153 | |
| 1154 | if (big_value) { |
| 1155 | i++; |
| 1156 | assert(i < fields->length(), "second T_INT field needed" ); |
| 1157 | assert(fields->at(i)._type == T_INT, "T_INT field needed" ); |
| 1158 | } else { |
| 1159 | val = value->get_int(); |
| 1160 | obj->int_field_put(offset, (jint)*((jint*)&val)); |
| 1161 | break; |
| 1162 | } |
| 1163 | } |
| 1164 | /* no break */ |
| 1165 | |
| 1166 | case T_LONG: case T_DOUBLE: { |
| 1167 | assert(value->type() == T_INT, "Agreement." ); |
| 1168 | StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); |
| 1169 | #ifdef _LP64 |
| 1170 | jlong res = (jlong)low->get_int(); |
| 1171 | #else |
| 1172 | #ifdef SPARC |
| 1173 | // For SPARC we have to swap high and low words. |
| 1174 | jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
| 1175 | #else |
| 1176 | jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
| 1177 | #endif //SPARC |
| 1178 | #endif |
| 1179 | obj->long_field_put(offset, res); |
| 1180 | break; |
| 1181 | } |
| 1182 | |
| 1183 | case T_SHORT: |
| 1184 | assert(value->type() == T_INT, "Agreement." ); |
| 1185 | val = value->get_int(); |
| 1186 | obj->short_field_put(offset, (jshort)*((jint*)&val)); |
| 1187 | break; |
| 1188 | |
| 1189 | case T_CHAR: |
| 1190 | assert(value->type() == T_INT, "Agreement." ); |
| 1191 | val = value->get_int(); |
| 1192 | obj->char_field_put(offset, (jchar)*((jint*)&val)); |
| 1193 | break; |
| 1194 | |
| 1195 | case T_BYTE: |
| 1196 | assert(value->type() == T_INT, "Agreement." ); |
| 1197 | val = value->get_int(); |
| 1198 | obj->byte_field_put(offset, (jbyte)*((jint*)&val)); |
| 1199 | break; |
| 1200 | |
| 1201 | case T_BOOLEAN: |
| 1202 | assert(value->type() == T_INT, "Agreement." ); |
| 1203 | val = value->get_int(); |
| 1204 | obj->bool_field_put(offset, (jboolean)*((jint*)&val)); |
| 1205 | break; |
| 1206 | |
| 1207 | default: |
| 1208 | ShouldNotReachHere(); |
| 1209 | } |
| 1210 | svIndex++; |
| 1211 | } |
| 1212 | return svIndex; |
| 1213 | } |
| 1214 | |
| 1215 | // restore fields of all eliminated objects and arrays |
| 1216 | void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { |
| 1217 | for (int i = 0; i < objects->length(); i++) { |
| 1218 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
| 1219 | Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); |
| 1220 | Handle obj = sv->value(); |
| 1221 | assert(obj.not_null() || realloc_failures, "reallocation was missed" ); |
| 1222 | if (PrintDeoptimizationDetails) { |
| 1223 | tty->print_cr("reassign fields for object of type %s!" , k->name()->as_C_string()); |
| 1224 | } |
| 1225 | if (obj.is_null()) { |
| 1226 | continue; |
| 1227 | } |
| 1228 | #if INCLUDE_JVMCI || INCLUDE_AOT |
| 1229 | // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. |
| 1230 | if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { |
| 1231 | continue; |
| 1232 | } |
| 1233 | #endif // INCLUDE_JVMCI || INCLUDE_AOT |
| 1234 | if (k->is_instance_klass()) { |
| 1235 | InstanceKlass* ik = InstanceKlass::cast(k); |
| 1236 | reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); |
| 1237 | } else if (k->is_typeArray_klass()) { |
| 1238 | TypeArrayKlass* ak = TypeArrayKlass::cast(k); |
| 1239 | reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); |
| 1240 | } else if (k->is_objArray_klass()) { |
| 1241 | reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); |
| 1242 | } |
| 1243 | } |
| 1244 | } |
| 1245 | |
| 1246 | |
| 1247 | // relock objects for which synchronization was eliminated |
| 1248 | void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { |
| 1249 | for (int i = 0; i < monitors->length(); i++) { |
| 1250 | MonitorInfo* mon_info = monitors->at(i); |
| 1251 | if (mon_info->eliminated()) { |
| 1252 | assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed" ); |
| 1253 | if (!mon_info->owner_is_scalar_replaced()) { |
| 1254 | Handle obj(thread, mon_info->owner()); |
| 1255 | markOop mark = obj->mark(); |
| 1256 | if (UseBiasedLocking && mark->has_bias_pattern()) { |
| 1257 | // New allocated objects may have the mark set to anonymously biased. |
| 1258 | // Also the deoptimized method may called methods with synchronization |
| 1259 | // where the thread-local object is bias locked to the current thread. |
| 1260 | assert(mark->is_biased_anonymously() || |
| 1261 | mark->biased_locker() == thread, "should be locked to current thread" ); |
| 1262 | // Reset mark word to unbiased prototype. |
| 1263 | markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); |
| 1264 | obj->set_mark(unbiased_prototype); |
| 1265 | } |
| 1266 | BasicLock* lock = mon_info->lock(); |
| 1267 | ObjectSynchronizer::slow_enter(obj, lock, thread); |
| 1268 | assert(mon_info->owner()->is_locked(), "object must be locked now" ); |
| 1269 | } |
| 1270 | } |
| 1271 | } |
| 1272 | } |
| 1273 | |
| 1274 | |
| 1275 | #ifndef PRODUCT |
| 1276 | // print information about reallocated objects |
| 1277 | void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { |
| 1278 | fieldDescriptor fd; |
| 1279 | |
| 1280 | for (int i = 0; i < objects->length(); i++) { |
| 1281 | ObjectValue* sv = (ObjectValue*) objects->at(i); |
| 1282 | Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); |
| 1283 | Handle obj = sv->value(); |
| 1284 | |
| 1285 | tty->print(" object <" INTPTR_FORMAT "> of type " , p2i(sv->value()())); |
| 1286 | k->print_value(); |
| 1287 | assert(obj.not_null() || realloc_failures, "reallocation was missed" ); |
| 1288 | if (obj.is_null()) { |
| 1289 | tty->print(" allocation failed" ); |
| 1290 | } else { |
| 1291 | tty->print(" allocated (%d bytes)" , obj->size() * HeapWordSize); |
| 1292 | } |
| 1293 | tty->cr(); |
| 1294 | |
| 1295 | if (Verbose && !obj.is_null()) { |
| 1296 | k->oop_print_on(obj(), tty); |
| 1297 | } |
| 1298 | } |
| 1299 | } |
| 1300 | #endif |
| 1301 | #endif // COMPILER2_OR_JVMCI |
| 1302 | |
| 1303 | vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { |
| 1304 | Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); |
| 1305 | |
| 1306 | #ifndef PRODUCT |
| 1307 | if (PrintDeoptimizationDetails) { |
| 1308 | ttyLocker ttyl; |
| 1309 | tty->print("DEOPT PACKING thread " INTPTR_FORMAT " " , p2i(thread)); |
| 1310 | fr.print_on(tty); |
| 1311 | tty->print_cr(" Virtual frames (innermost first):" ); |
| 1312 | for (int index = 0; index < chunk->length(); index++) { |
| 1313 | compiledVFrame* vf = chunk->at(index); |
| 1314 | tty->print(" %2d - " , index); |
| 1315 | vf->print_value(); |
| 1316 | int bci = chunk->at(index)->raw_bci(); |
| 1317 | const char* code_name; |
| 1318 | if (bci == SynchronizationEntryBCI) { |
| 1319 | code_name = "sync entry" ; |
| 1320 | } else { |
| 1321 | Bytecodes::Code code = vf->method()->code_at(bci); |
| 1322 | code_name = Bytecodes::name(code); |
| 1323 | } |
| 1324 | tty->print(" - %s" , code_name); |
| 1325 | tty->print_cr(" @ bci %d " , bci); |
| 1326 | if (Verbose) { |
| 1327 | vf->print(); |
| 1328 | tty->cr(); |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | #endif |
| 1333 | |
| 1334 | // Register map for next frame (used for stack crawl). We capture |
| 1335 | // the state of the deopt'ing frame's caller. Thus if we need to |
| 1336 | // stuff a C2I adapter we can properly fill in the callee-save |
| 1337 | // register locations. |
| 1338 | frame caller = fr.sender(reg_map); |
| 1339 | int frame_size = caller.sp() - fr.sp(); |
| 1340 | |
| 1341 | frame sender = caller; |
| 1342 | |
| 1343 | // Since the Java thread being deoptimized will eventually adjust it's own stack, |
| 1344 | // the vframeArray containing the unpacking information is allocated in the C heap. |
| 1345 | // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). |
| 1346 | vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); |
| 1347 | |
| 1348 | // Compare the vframeArray to the collected vframes |
| 1349 | assert(array->structural_compare(thread, chunk), "just checking" ); |
| 1350 | |
| 1351 | #ifndef PRODUCT |
| 1352 | if (PrintDeoptimizationDetails) { |
| 1353 | ttyLocker ttyl; |
| 1354 | tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); |
| 1355 | } |
| 1356 | #endif // PRODUCT |
| 1357 | |
| 1358 | return array; |
| 1359 | } |
| 1360 | |
| 1361 | #if COMPILER2_OR_JVMCI |
| 1362 | void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { |
| 1363 | // Reallocation of some scalar replaced objects failed. Record |
| 1364 | // that we need to pop all the interpreter frames for the |
| 1365 | // deoptimized compiled frame. |
| 1366 | assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?" ); |
| 1367 | thread->set_frames_to_pop_failed_realloc(array->frames()); |
| 1368 | // Unlock all monitors here otherwise the interpreter will see a |
| 1369 | // mix of locked and unlocked monitors (because of failed |
| 1370 | // reallocations of synchronized objects) and be confused. |
| 1371 | for (int i = 0; i < array->frames(); i++) { |
| 1372 | MonitorChunk* monitors = array->element(i)->monitors(); |
| 1373 | if (monitors != NULL) { |
| 1374 | for (int j = 0; j < monitors->number_of_monitors(); j++) { |
| 1375 | BasicObjectLock* src = monitors->at(j); |
| 1376 | if (src->obj() != NULL) { |
| 1377 | ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); |
| 1378 | } |
| 1379 | } |
| 1380 | array->element(i)->free_monitors(thread); |
| 1381 | #ifdef ASSERT |
| 1382 | array->element(i)->set_removed_monitors(); |
| 1383 | #endif |
| 1384 | } |
| 1385 | } |
| 1386 | } |
| 1387 | #endif |
| 1388 | |
| 1389 | static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { |
| 1390 | GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
| 1391 | Thread* thread = Thread::current(); |
| 1392 | for (int i = 0; i < monitors->length(); i++) { |
| 1393 | MonitorInfo* mon_info = monitors->at(i); |
| 1394 | if (!mon_info->eliminated() && mon_info->owner() != NULL) { |
| 1395 | objects_to_revoke->append(Handle(thread, mon_info->owner())); |
| 1396 | } |
| 1397 | } |
| 1398 | } |
| 1399 | |
| 1400 | |
| 1401 | void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { |
| 1402 | if (!UseBiasedLocking) { |
| 1403 | return; |
| 1404 | } |
| 1405 | |
| 1406 | GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); |
| 1407 | |
| 1408 | // Unfortunately we don't have a RegisterMap available in most of |
| 1409 | // the places we want to call this routine so we need to walk the |
| 1410 | // stack again to update the register map. |
| 1411 | if (map == NULL || !map->update_map()) { |
| 1412 | StackFrameStream sfs(thread, true); |
| 1413 | bool found = false; |
| 1414 | while (!found && !sfs.is_done()) { |
| 1415 | frame* cur = sfs.current(); |
| 1416 | sfs.next(); |
| 1417 | found = cur->id() == fr.id(); |
| 1418 | } |
| 1419 | assert(found, "frame to be deoptimized not found on target thread's stack" ); |
| 1420 | map = sfs.register_map(); |
| 1421 | } |
| 1422 | |
| 1423 | vframe* vf = vframe::new_vframe(&fr, map, thread); |
| 1424 | compiledVFrame* cvf = compiledVFrame::cast(vf); |
| 1425 | // Revoke monitors' biases in all scopes |
| 1426 | while (!cvf->is_top()) { |
| 1427 | collect_monitors(cvf, objects_to_revoke); |
| 1428 | cvf = compiledVFrame::cast(cvf->sender()); |
| 1429 | } |
| 1430 | collect_monitors(cvf, objects_to_revoke); |
| 1431 | |
| 1432 | if (SafepointSynchronize::is_at_safepoint()) { |
| 1433 | BiasedLocking::revoke_at_safepoint(objects_to_revoke); |
| 1434 | } else { |
| 1435 | BiasedLocking::revoke(objects_to_revoke); |
| 1436 | } |
| 1437 | } |
| 1438 | |
| 1439 | |
| 1440 | void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { |
| 1441 | assert(fr.can_be_deoptimized(), "checking frame type" ); |
| 1442 | |
| 1443 | gather_statistics(reason, Action_none, Bytecodes::_illegal); |
| 1444 | |
| 1445 | if (LogCompilation && xtty != NULL) { |
| 1446 | CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); |
| 1447 | assert(cm != NULL, "only compiled methods can deopt" ); |
| 1448 | |
| 1449 | ttyLocker ttyl; |
| 1450 | xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'" ,(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); |
| 1451 | cm->log_identity(xtty); |
| 1452 | xtty->end_head(); |
| 1453 | for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { |
| 1454 | xtty->begin_elem("jvms bci='%d'" , sd->bci()); |
| 1455 | xtty->method(sd->method()); |
| 1456 | xtty->end_elem(); |
| 1457 | if (sd->is_top()) break; |
| 1458 | } |
| 1459 | xtty->tail("deoptimized" ); |
| 1460 | } |
| 1461 | |
| 1462 | // Patch the compiled method so that when execution returns to it we will |
| 1463 | // deopt the execution state and return to the interpreter. |
| 1464 | fr.deoptimize(thread); |
| 1465 | } |
| 1466 | |
| 1467 | void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { |
| 1468 | deoptimize(thread, fr, map, Reason_constraint); |
| 1469 | } |
| 1470 | |
| 1471 | void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { |
| 1472 | // Deoptimize only if the frame comes from compile code. |
| 1473 | // Do not deoptimize the frame which is already patched |
| 1474 | // during the execution of the loops below. |
| 1475 | if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { |
| 1476 | return; |
| 1477 | } |
| 1478 | ResourceMark rm; |
| 1479 | DeoptimizationMarker dm; |
| 1480 | if (UseBiasedLocking) { |
| 1481 | revoke_biases_of_monitors(thread, fr, map); |
| 1482 | } |
| 1483 | deoptimize_single_frame(thread, fr, reason); |
| 1484 | |
| 1485 | } |
| 1486 | |
| 1487 | #if INCLUDE_JVMCI |
| 1488 | address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { |
| 1489 | // there is no exception handler for this pc => deoptimize |
| 1490 | cm->make_not_entrant(); |
| 1491 | |
| 1492 | // Use Deoptimization::deoptimize for all of its side-effects: |
| 1493 | // revoking biases of monitors, gathering traps statistics, logging... |
| 1494 | // it also patches the return pc but we do not care about that |
| 1495 | // since we return a continuation to the deopt_blob below. |
| 1496 | JavaThread* thread = JavaThread::current(); |
| 1497 | RegisterMap reg_map(thread, UseBiasedLocking); |
| 1498 | frame runtime_frame = thread->last_frame(); |
| 1499 | frame caller_frame = runtime_frame.sender(®_map); |
| 1500 | assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method" ); |
| 1501 | Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); |
| 1502 | |
| 1503 | MethodData* trap_mdo = get_method_data(thread, cm->method(), true); |
| 1504 | if (trap_mdo != NULL) { |
| 1505 | trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); |
| 1506 | } |
| 1507 | |
| 1508 | return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); |
| 1509 | } |
| 1510 | #endif |
| 1511 | |
| 1512 | void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { |
| 1513 | assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), |
| 1514 | "can only deoptimize other thread at a safepoint" ); |
| 1515 | // Compute frame and register map based on thread and sp. |
| 1516 | RegisterMap reg_map(thread, UseBiasedLocking); |
| 1517 | frame fr = thread->last_frame(); |
| 1518 | while (fr.id() != id) { |
| 1519 | fr = fr.sender(®_map); |
| 1520 | } |
| 1521 | deoptimize(thread, fr, ®_map, reason); |
| 1522 | } |
| 1523 | |
| 1524 | |
| 1525 | void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { |
| 1526 | if (thread == Thread::current()) { |
| 1527 | Deoptimization::deoptimize_frame_internal(thread, id, reason); |
| 1528 | } else { |
| 1529 | VM_DeoptimizeFrame deopt(thread, id, reason); |
| 1530 | VMThread::execute(&deopt); |
| 1531 | } |
| 1532 | } |
| 1533 | |
| 1534 | void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { |
| 1535 | deoptimize_frame(thread, id, Reason_constraint); |
| 1536 | } |
| 1537 | |
| 1538 | // JVMTI PopFrame support |
| 1539 | JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) |
| 1540 | { |
| 1541 | thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); |
| 1542 | } |
| 1543 | JRT_END |
| 1544 | |
| 1545 | MethodData* |
| 1546 | Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, |
| 1547 | bool create_if_missing) { |
| 1548 | Thread* THREAD = thread; |
| 1549 | MethodData* mdo = m()->method_data(); |
| 1550 | if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { |
| 1551 | // Build an MDO. Ignore errors like OutOfMemory; |
| 1552 | // that simply means we won't have an MDO to update. |
| 1553 | Method::build_interpreter_method_data(m, THREAD); |
| 1554 | if (HAS_PENDING_EXCEPTION) { |
| 1555 | assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here" ); |
| 1556 | CLEAR_PENDING_EXCEPTION; |
| 1557 | } |
| 1558 | mdo = m()->method_data(); |
| 1559 | } |
| 1560 | return mdo; |
| 1561 | } |
| 1562 | |
| 1563 | #if COMPILER2_OR_JVMCI |
| 1564 | void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { |
| 1565 | // in case of an unresolved klass entry, load the class. |
| 1566 | if (constant_pool->tag_at(index).is_unresolved_klass()) { |
| 1567 | Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); |
| 1568 | return; |
| 1569 | } |
| 1570 | |
| 1571 | if (!constant_pool->tag_at(index).is_symbol()) return; |
| 1572 | |
| 1573 | Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); |
| 1574 | Symbol* symbol = constant_pool->symbol_at(index); |
| 1575 | |
| 1576 | // class name? |
| 1577 | if (symbol->char_at(0) != '(') { |
| 1578 | Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); |
| 1579 | SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); |
| 1580 | return; |
| 1581 | } |
| 1582 | |
| 1583 | // then it must be a signature! |
| 1584 | ResourceMark rm(THREAD); |
| 1585 | for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { |
| 1586 | if (ss.is_object()) { |
| 1587 | Symbol* class_name = ss.as_symbol(); |
| 1588 | Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); |
| 1589 | SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); |
| 1590 | } |
| 1591 | } |
| 1592 | } |
| 1593 | |
| 1594 | |
| 1595 | void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { |
| 1596 | EXCEPTION_MARK; |
| 1597 | load_class_by_index(constant_pool, index, THREAD); |
| 1598 | if (HAS_PENDING_EXCEPTION) { |
| 1599 | // Exception happened during classloading. We ignore the exception here, since it |
| 1600 | // is going to be rethrown since the current activation is going to be deoptimized and |
| 1601 | // the interpreter will re-execute the bytecode. |
| 1602 | CLEAR_PENDING_EXCEPTION; |
| 1603 | // Class loading called java code which may have caused a stack |
| 1604 | // overflow. If the exception was thrown right before the return |
| 1605 | // to the runtime the stack is no longer guarded. Reguard the |
| 1606 | // stack otherwise if we return to the uncommon trap blob and the |
| 1607 | // stack bang causes a stack overflow we crash. |
| 1608 | assert(THREAD->is_Java_thread(), "only a java thread can be here" ); |
| 1609 | JavaThread* thread = (JavaThread*)THREAD; |
| 1610 | bool guard_pages_enabled = thread->stack_guards_enabled(); |
| 1611 | if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); |
| 1612 | assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash" ); |
| 1613 | } |
| 1614 | } |
| 1615 | |
| 1616 | JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { |
| 1617 | HandleMark hm; |
| 1618 | |
| 1619 | // uncommon_trap() is called at the beginning of the uncommon trap |
| 1620 | // handler. Note this fact before we start generating temporary frames |
| 1621 | // that can confuse an asynchronous stack walker. This counter is |
| 1622 | // decremented at the end of unpack_frames(). |
| 1623 | thread->inc_in_deopt_handler(); |
| 1624 | |
| 1625 | // We need to update the map if we have biased locking. |
| 1626 | #if INCLUDE_JVMCI |
| 1627 | // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid |
| 1628 | RegisterMap reg_map(thread, true); |
| 1629 | #else |
| 1630 | RegisterMap reg_map(thread, UseBiasedLocking); |
| 1631 | #endif |
| 1632 | frame stub_frame = thread->last_frame(); |
| 1633 | frame fr = stub_frame.sender(®_map); |
| 1634 | // Make sure the calling nmethod is not getting deoptimized and removed |
| 1635 | // before we are done with it. |
| 1636 | nmethodLocker nl(fr.pc()); |
| 1637 | |
| 1638 | // Log a message |
| 1639 | Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, |
| 1640 | trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); |
| 1641 | |
| 1642 | { |
| 1643 | ResourceMark rm; |
| 1644 | |
| 1645 | // Revoke biases of any monitors in the frame to ensure we can migrate them |
| 1646 | revoke_biases_of_monitors(thread, fr, ®_map); |
| 1647 | |
| 1648 | DeoptReason reason = trap_request_reason(trap_request); |
| 1649 | DeoptAction action = trap_request_action(trap_request); |
| 1650 | #if INCLUDE_JVMCI |
| 1651 | int debug_id = trap_request_debug_id(trap_request); |
| 1652 | #endif |
| 1653 | jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 |
| 1654 | |
| 1655 | vframe* vf = vframe::new_vframe(&fr, ®_map, thread); |
| 1656 | compiledVFrame* cvf = compiledVFrame::cast(vf); |
| 1657 | |
| 1658 | CompiledMethod* nm = cvf->code(); |
| 1659 | |
| 1660 | ScopeDesc* trap_scope = cvf->scope(); |
| 1661 | |
| 1662 | if (TraceDeoptimization) { |
| 1663 | ttyLocker ttyl; |
| 1664 | tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d" ), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() |
| 1665 | #if INCLUDE_JVMCI |
| 1666 | , debug_id |
| 1667 | #endif |
| 1668 | ); |
| 1669 | } |
| 1670 | |
| 1671 | methodHandle trap_method = trap_scope->method(); |
| 1672 | int trap_bci = trap_scope->bci(); |
| 1673 | #if INCLUDE_JVMCI |
| 1674 | jlong speculation = thread->pending_failed_speculation(); |
| 1675 | if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods |
| 1676 | nm->as_nmethod()->update_speculation(thread); |
| 1677 | } else { |
| 1678 | assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers" ); |
| 1679 | } |
| 1680 | |
| 1681 | if (trap_bci == SynchronizationEntryBCI) { |
| 1682 | trap_bci = 0; |
| 1683 | thread->set_pending_monitorenter(true); |
| 1684 | } |
| 1685 | |
| 1686 | if (reason == Deoptimization::Reason_transfer_to_interpreter) { |
| 1687 | thread->set_pending_transfer_to_interpreter(true); |
| 1688 | } |
| 1689 | #endif |
| 1690 | |
| 1691 | Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); |
| 1692 | // Record this event in the histogram. |
| 1693 | gather_statistics(reason, action, trap_bc); |
| 1694 | |
| 1695 | // Ensure that we can record deopt. history: |
| 1696 | // Need MDO to record RTM code generation state. |
| 1697 | bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); |
| 1698 | |
| 1699 | methodHandle profiled_method; |
| 1700 | #if INCLUDE_JVMCI |
| 1701 | if (nm->is_compiled_by_jvmci()) { |
| 1702 | profiled_method = nm->method(); |
| 1703 | } else { |
| 1704 | profiled_method = trap_method; |
| 1705 | } |
| 1706 | #else |
| 1707 | profiled_method = trap_method; |
| 1708 | #endif |
| 1709 | |
| 1710 | MethodData* trap_mdo = |
| 1711 | get_method_data(thread, profiled_method, create_if_missing); |
| 1712 | |
| 1713 | // Log a message |
| 1714 | Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s" , |
| 1715 | trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), |
| 1716 | trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); |
| 1717 | |
| 1718 | // Print a bunch of diagnostics, if requested. |
| 1719 | if (TraceDeoptimization || LogCompilation) { |
| 1720 | ResourceMark rm; |
| 1721 | ttyLocker ttyl; |
| 1722 | char buf[100]; |
| 1723 | if (xtty != NULL) { |
| 1724 | xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s" , |
| 1725 | os::current_thread_id(), |
| 1726 | format_trap_request(buf, sizeof(buf), trap_request)); |
| 1727 | #if INCLUDE_JVMCI |
| 1728 | if (speculation != 0) { |
| 1729 | xtty->print(" speculation='" JLONG_FORMAT "'" , speculation); |
| 1730 | } |
| 1731 | #endif |
| 1732 | nm->log_identity(xtty); |
| 1733 | } |
| 1734 | Symbol* class_name = NULL; |
| 1735 | bool unresolved = false; |
| 1736 | if (unloaded_class_index >= 0) { |
| 1737 | constantPoolHandle constants (THREAD, trap_method->constants()); |
| 1738 | if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { |
| 1739 | class_name = constants->klass_name_at(unloaded_class_index); |
| 1740 | unresolved = true; |
| 1741 | if (xtty != NULL) |
| 1742 | xtty->print(" unresolved='1'" ); |
| 1743 | } else if (constants->tag_at(unloaded_class_index).is_symbol()) { |
| 1744 | class_name = constants->symbol_at(unloaded_class_index); |
| 1745 | } |
| 1746 | if (xtty != NULL) |
| 1747 | xtty->name(class_name); |
| 1748 | } |
| 1749 | if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { |
| 1750 | // Dump the relevant MDO state. |
| 1751 | // This is the deopt count for the current reason, any previous |
| 1752 | // reasons or recompiles seen at this point. |
| 1753 | int dcnt = trap_mdo->trap_count(reason); |
| 1754 | if (dcnt != 0) |
| 1755 | xtty->print(" count='%d'" , dcnt); |
| 1756 | ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); |
| 1757 | int dos = (pdata == NULL)? 0: pdata->trap_state(); |
| 1758 | if (dos != 0) { |
| 1759 | xtty->print(" state='%s'" , format_trap_state(buf, sizeof(buf), dos)); |
| 1760 | if (trap_state_is_recompiled(dos)) { |
| 1761 | int recnt2 = trap_mdo->overflow_recompile_count(); |
| 1762 | if (recnt2 != 0) |
| 1763 | xtty->print(" recompiles2='%d'" , recnt2); |
| 1764 | } |
| 1765 | } |
| 1766 | } |
| 1767 | if (xtty != NULL) { |
| 1768 | xtty->stamp(); |
| 1769 | xtty->end_head(); |
| 1770 | } |
| 1771 | if (TraceDeoptimization) { // make noise on the tty |
| 1772 | tty->print("Uncommon trap occurred in" ); |
| 1773 | nm->method()->print_short_name(tty); |
| 1774 | tty->print(" compiler=%s compile_id=%d" , nm->compiler_name(), nm->compile_id()); |
| 1775 | #if INCLUDE_JVMCI |
| 1776 | if (nm->is_nmethod()) { |
| 1777 | const char* installed_code_name = nm->as_nmethod()->jvmci_name(); |
| 1778 | if (installed_code_name != NULL) { |
| 1779 | tty->print(" (JVMCI: installed code name=%s) " , installed_code_name); |
| 1780 | } |
| 1781 | } |
| 1782 | #endif |
| 1783 | tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d" ), |
| 1784 | p2i(fr.pc()), |
| 1785 | os::current_thread_id(), |
| 1786 | trap_reason_name(reason), |
| 1787 | trap_action_name(action), |
| 1788 | unloaded_class_index |
| 1789 | #if INCLUDE_JVMCI |
| 1790 | , debug_id |
| 1791 | #endif |
| 1792 | ); |
| 1793 | if (class_name != NULL) { |
| 1794 | tty->print(unresolved ? " unresolved class: " : " symbol: " ); |
| 1795 | class_name->print_symbol_on(tty); |
| 1796 | } |
| 1797 | tty->cr(); |
| 1798 | } |
| 1799 | if (xtty != NULL) { |
| 1800 | // Log the precise location of the trap. |
| 1801 | for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { |
| 1802 | xtty->begin_elem("jvms bci='%d'" , sd->bci()); |
| 1803 | xtty->method(sd->method()); |
| 1804 | xtty->end_elem(); |
| 1805 | if (sd->is_top()) break; |
| 1806 | } |
| 1807 | xtty->tail("uncommon_trap" ); |
| 1808 | } |
| 1809 | } |
| 1810 | // (End diagnostic printout.) |
| 1811 | |
| 1812 | // Load class if necessary |
| 1813 | if (unloaded_class_index >= 0) { |
| 1814 | constantPoolHandle constants(THREAD, trap_method->constants()); |
| 1815 | load_class_by_index(constants, unloaded_class_index); |
| 1816 | } |
| 1817 | |
| 1818 | // Flush the nmethod if necessary and desirable. |
| 1819 | // |
| 1820 | // We need to avoid situations where we are re-flushing the nmethod |
| 1821 | // because of a hot deoptimization site. Repeated flushes at the same |
| 1822 | // point need to be detected by the compiler and avoided. If the compiler |
| 1823 | // cannot avoid them (or has a bug and "refuses" to avoid them), this |
| 1824 | // module must take measures to avoid an infinite cycle of recompilation |
| 1825 | // and deoptimization. There are several such measures: |
| 1826 | // |
| 1827 | // 1. If a recompilation is ordered a second time at some site X |
| 1828 | // and for the same reason R, the action is adjusted to 'reinterpret', |
| 1829 | // to give the interpreter time to exercise the method more thoroughly. |
| 1830 | // If this happens, the method's overflow_recompile_count is incremented. |
| 1831 | // |
| 1832 | // 2. If the compiler fails to reduce the deoptimization rate, then |
| 1833 | // the method's overflow_recompile_count will begin to exceed the set |
| 1834 | // limit PerBytecodeRecompilationCutoff. If this happens, the action |
| 1835 | // is adjusted to 'make_not_compilable', and the method is abandoned |
| 1836 | // to the interpreter. This is a performance hit for hot methods, |
| 1837 | // but is better than a disastrous infinite cycle of recompilations. |
| 1838 | // (Actually, only the method containing the site X is abandoned.) |
| 1839 | // |
| 1840 | // 3. In parallel with the previous measures, if the total number of |
| 1841 | // recompilations of a method exceeds the much larger set limit |
| 1842 | // PerMethodRecompilationCutoff, the method is abandoned. |
| 1843 | // This should only happen if the method is very large and has |
| 1844 | // many "lukewarm" deoptimizations. The code which enforces this |
| 1845 | // limit is elsewhere (class nmethod, class Method). |
| 1846 | // |
| 1847 | // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance |
| 1848 | // to recompile at each bytecode independently of the per-BCI cutoff. |
| 1849 | // |
| 1850 | // The decision to update code is up to the compiler, and is encoded |
| 1851 | // in the Action_xxx code. If the compiler requests Action_none |
| 1852 | // no trap state is changed, no compiled code is changed, and the |
| 1853 | // computation suffers along in the interpreter. |
| 1854 | // |
| 1855 | // The other action codes specify various tactics for decompilation |
| 1856 | // and recompilation. Action_maybe_recompile is the loosest, and |
| 1857 | // allows the compiled code to stay around until enough traps are seen, |
| 1858 | // and until the compiler gets around to recompiling the trapping method. |
| 1859 | // |
| 1860 | // The other actions cause immediate removal of the present code. |
| 1861 | |
| 1862 | // Traps caused by injected profile shouldn't pollute trap counts. |
| 1863 | bool injected_profile_trap = trap_method->has_injected_profile() && |
| 1864 | (reason == Reason_intrinsic || reason == Reason_unreached); |
| 1865 | |
| 1866 | bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; |
| 1867 | bool make_not_entrant = false; |
| 1868 | bool make_not_compilable = false; |
| 1869 | bool reprofile = false; |
| 1870 | switch (action) { |
| 1871 | case Action_none: |
| 1872 | // Keep the old code. |
| 1873 | update_trap_state = false; |
| 1874 | break; |
| 1875 | case Action_maybe_recompile: |
| 1876 | // Do not need to invalidate the present code, but we can |
| 1877 | // initiate another |
| 1878 | // Start compiler without (necessarily) invalidating the nmethod. |
| 1879 | // The system will tolerate the old code, but new code should be |
| 1880 | // generated when possible. |
| 1881 | break; |
| 1882 | case Action_reinterpret: |
| 1883 | // Go back into the interpreter for a while, and then consider |
| 1884 | // recompiling form scratch. |
| 1885 | make_not_entrant = true; |
| 1886 | // Reset invocation counter for outer most method. |
| 1887 | // This will allow the interpreter to exercise the bytecodes |
| 1888 | // for a while before recompiling. |
| 1889 | // By contrast, Action_make_not_entrant is immediate. |
| 1890 | // |
| 1891 | // Note that the compiler will track null_check, null_assert, |
| 1892 | // range_check, and class_check events and log them as if they |
| 1893 | // had been traps taken from compiled code. This will update |
| 1894 | // the MDO trap history so that the next compilation will |
| 1895 | // properly detect hot trap sites. |
| 1896 | reprofile = true; |
| 1897 | break; |
| 1898 | case Action_make_not_entrant: |
| 1899 | // Request immediate recompilation, and get rid of the old code. |
| 1900 | // Make them not entrant, so next time they are called they get |
| 1901 | // recompiled. Unloaded classes are loaded now so recompile before next |
| 1902 | // time they are called. Same for uninitialized. The interpreter will |
| 1903 | // link the missing class, if any. |
| 1904 | make_not_entrant = true; |
| 1905 | break; |
| 1906 | case Action_make_not_compilable: |
| 1907 | // Give up on compiling this method at all. |
| 1908 | make_not_entrant = true; |
| 1909 | make_not_compilable = true; |
| 1910 | break; |
| 1911 | default: |
| 1912 | ShouldNotReachHere(); |
| 1913 | } |
| 1914 | |
| 1915 | // Setting +ProfileTraps fixes the following, on all platforms: |
| 1916 | // 4852688: ProfileInterpreter is off by default for ia64. The result is |
| 1917 | // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the |
| 1918 | // recompile relies on a MethodData* to record heroic opt failures. |
| 1919 | |
| 1920 | // Whether the interpreter is producing MDO data or not, we also need |
| 1921 | // to use the MDO to detect hot deoptimization points and control |
| 1922 | // aggressive optimization. |
| 1923 | bool inc_recompile_count = false; |
| 1924 | ProfileData* pdata = NULL; |
| 1925 | if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { |
| 1926 | assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity" ); |
| 1927 | uint this_trap_count = 0; |
| 1928 | bool maybe_prior_trap = false; |
| 1929 | bool maybe_prior_recompile = false; |
| 1930 | pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, |
| 1931 | #if INCLUDE_JVMCI |
| 1932 | nm->is_compiled_by_jvmci() && nm->is_osr_method(), |
| 1933 | #endif |
| 1934 | nm->method(), |
| 1935 | //outputs: |
| 1936 | this_trap_count, |
| 1937 | maybe_prior_trap, |
| 1938 | maybe_prior_recompile); |
| 1939 | // Because the interpreter also counts null, div0, range, and class |
| 1940 | // checks, these traps from compiled code are double-counted. |
| 1941 | // This is harmless; it just means that the PerXTrapLimit values |
| 1942 | // are in effect a little smaller than they look. |
| 1943 | |
| 1944 | DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); |
| 1945 | if (per_bc_reason != Reason_none) { |
| 1946 | // Now take action based on the partially known per-BCI history. |
| 1947 | if (maybe_prior_trap |
| 1948 | && this_trap_count >= (uint)PerBytecodeTrapLimit) { |
| 1949 | // If there are too many traps at this BCI, force a recompile. |
| 1950 | // This will allow the compiler to see the limit overflow, and |
| 1951 | // take corrective action, if possible. The compiler generally |
| 1952 | // does not use the exact PerBytecodeTrapLimit value, but instead |
| 1953 | // changes its tactics if it sees any traps at all. This provides |
| 1954 | // a little hysteresis, delaying a recompile until a trap happens |
| 1955 | // several times. |
| 1956 | // |
| 1957 | // Actually, since there is only one bit of counter per BCI, |
| 1958 | // the possible per-BCI counts are {0,1,(per-method count)}. |
| 1959 | // This produces accurate results if in fact there is only |
| 1960 | // one hot trap site, but begins to get fuzzy if there are |
| 1961 | // many sites. For example, if there are ten sites each |
| 1962 | // trapping two or more times, they each get the blame for |
| 1963 | // all of their traps. |
| 1964 | make_not_entrant = true; |
| 1965 | } |
| 1966 | |
| 1967 | // Detect repeated recompilation at the same BCI, and enforce a limit. |
| 1968 | if (make_not_entrant && maybe_prior_recompile) { |
| 1969 | // More than one recompile at this point. |
| 1970 | inc_recompile_count = maybe_prior_trap; |
| 1971 | } |
| 1972 | } else { |
| 1973 | // For reasons which are not recorded per-bytecode, we simply |
| 1974 | // force recompiles unconditionally. |
| 1975 | // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) |
| 1976 | make_not_entrant = true; |
| 1977 | } |
| 1978 | |
| 1979 | // Go back to the compiler if there are too many traps in this method. |
| 1980 | if (this_trap_count >= per_method_trap_limit(reason)) { |
| 1981 | // If there are too many traps in this method, force a recompile. |
| 1982 | // This will allow the compiler to see the limit overflow, and |
| 1983 | // take corrective action, if possible. |
| 1984 | // (This condition is an unlikely backstop only, because the |
| 1985 | // PerBytecodeTrapLimit is more likely to take effect first, |
| 1986 | // if it is applicable.) |
| 1987 | make_not_entrant = true; |
| 1988 | } |
| 1989 | |
| 1990 | // Here's more hysteresis: If there has been a recompile at |
| 1991 | // this trap point already, run the method in the interpreter |
| 1992 | // for a while to exercise it more thoroughly. |
| 1993 | if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { |
| 1994 | reprofile = true; |
| 1995 | } |
| 1996 | } |
| 1997 | |
| 1998 | // Take requested actions on the method: |
| 1999 | |
| 2000 | // Recompile |
| 2001 | if (make_not_entrant) { |
| 2002 | if (!nm->make_not_entrant()) { |
| 2003 | return; // the call did not change nmethod's state |
| 2004 | } |
| 2005 | |
| 2006 | if (pdata != NULL) { |
| 2007 | // Record the recompilation event, if any. |
| 2008 | int tstate0 = pdata->trap_state(); |
| 2009 | int tstate1 = trap_state_set_recompiled(tstate0, true); |
| 2010 | if (tstate1 != tstate0) |
| 2011 | pdata->set_trap_state(tstate1); |
| 2012 | } |
| 2013 | |
| 2014 | #if INCLUDE_RTM_OPT |
| 2015 | // Restart collecting RTM locking abort statistic if the method |
| 2016 | // is recompiled for a reason other than RTM state change. |
| 2017 | // Assume that in new recompiled code the statistic could be different, |
| 2018 | // for example, due to different inlining. |
| 2019 | if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && |
| 2020 | UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { |
| 2021 | trap_mdo->atomic_set_rtm_state(ProfileRTM); |
| 2022 | } |
| 2023 | #endif |
| 2024 | // For code aging we count traps separately here, using make_not_entrant() |
| 2025 | // as a guard against simultaneous deopts in multiple threads. |
| 2026 | if (reason == Reason_tenured && trap_mdo != NULL) { |
| 2027 | trap_mdo->inc_tenure_traps(); |
| 2028 | } |
| 2029 | } |
| 2030 | |
| 2031 | if (inc_recompile_count) { |
| 2032 | trap_mdo->inc_overflow_recompile_count(); |
| 2033 | if ((uint)trap_mdo->overflow_recompile_count() > |
| 2034 | (uint)PerBytecodeRecompilationCutoff) { |
| 2035 | // Give up on the method containing the bad BCI. |
| 2036 | if (trap_method() == nm->method()) { |
| 2037 | make_not_compilable = true; |
| 2038 | } else { |
| 2039 | trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff" , CompLevel_full_optimization); |
| 2040 | // But give grace to the enclosing nm->method(). |
| 2041 | } |
| 2042 | } |
| 2043 | } |
| 2044 | |
| 2045 | // Reprofile |
| 2046 | if (reprofile) { |
| 2047 | CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); |
| 2048 | } |
| 2049 | |
| 2050 | // Give up compiling |
| 2051 | if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { |
| 2052 | assert(make_not_entrant, "consistent" ); |
| 2053 | nm->method()->set_not_compilable("give up compiling" , CompLevel_full_optimization); |
| 2054 | } |
| 2055 | |
| 2056 | } // Free marked resources |
| 2057 | |
| 2058 | } |
| 2059 | JRT_END |
| 2060 | |
| 2061 | ProfileData* |
| 2062 | Deoptimization::query_update_method_data(MethodData* trap_mdo, |
| 2063 | int trap_bci, |
| 2064 | Deoptimization::DeoptReason reason, |
| 2065 | bool update_total_trap_count, |
| 2066 | #if INCLUDE_JVMCI |
| 2067 | bool is_osr, |
| 2068 | #endif |
| 2069 | Method* compiled_method, |
| 2070 | //outputs: |
| 2071 | uint& ret_this_trap_count, |
| 2072 | bool& ret_maybe_prior_trap, |
| 2073 | bool& ret_maybe_prior_recompile) { |
| 2074 | bool maybe_prior_trap = false; |
| 2075 | bool maybe_prior_recompile = false; |
| 2076 | uint this_trap_count = 0; |
| 2077 | if (update_total_trap_count) { |
| 2078 | uint idx = reason; |
| 2079 | #if INCLUDE_JVMCI |
| 2080 | if (is_osr) { |
| 2081 | idx += Reason_LIMIT; |
| 2082 | } |
| 2083 | #endif |
| 2084 | uint prior_trap_count = trap_mdo->trap_count(idx); |
| 2085 | this_trap_count = trap_mdo->inc_trap_count(idx); |
| 2086 | |
| 2087 | // If the runtime cannot find a place to store trap history, |
| 2088 | // it is estimated based on the general condition of the method. |
| 2089 | // If the method has ever been recompiled, or has ever incurred |
| 2090 | // a trap with the present reason , then this BCI is assumed |
| 2091 | // (pessimistically) to be the culprit. |
| 2092 | maybe_prior_trap = (prior_trap_count != 0); |
| 2093 | maybe_prior_recompile = (trap_mdo->decompile_count() != 0); |
| 2094 | } |
| 2095 | ProfileData* pdata = NULL; |
| 2096 | |
| 2097 | |
| 2098 | // For reasons which are recorded per bytecode, we check per-BCI data. |
| 2099 | DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); |
| 2100 | assert(per_bc_reason != Reason_none || update_total_trap_count, "must be" ); |
| 2101 | if (per_bc_reason != Reason_none) { |
| 2102 | // Find the profile data for this BCI. If there isn't one, |
| 2103 | // try to allocate one from the MDO's set of spares. |
| 2104 | // This will let us detect a repeated trap at this point. |
| 2105 | pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); |
| 2106 | |
| 2107 | if (pdata != NULL) { |
| 2108 | if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { |
| 2109 | if (LogCompilation && xtty != NULL) { |
| 2110 | ttyLocker ttyl; |
| 2111 | // no more room for speculative traps in this MDO |
| 2112 | xtty->elem("speculative_traps_oom" ); |
| 2113 | } |
| 2114 | } |
| 2115 | // Query the trap state of this profile datum. |
| 2116 | int tstate0 = pdata->trap_state(); |
| 2117 | if (!trap_state_has_reason(tstate0, per_bc_reason)) |
| 2118 | maybe_prior_trap = false; |
| 2119 | if (!trap_state_is_recompiled(tstate0)) |
| 2120 | maybe_prior_recompile = false; |
| 2121 | |
| 2122 | // Update the trap state of this profile datum. |
| 2123 | int tstate1 = tstate0; |
| 2124 | // Record the reason. |
| 2125 | tstate1 = trap_state_add_reason(tstate1, per_bc_reason); |
| 2126 | // Store the updated state on the MDO, for next time. |
| 2127 | if (tstate1 != tstate0) |
| 2128 | pdata->set_trap_state(tstate1); |
| 2129 | } else { |
| 2130 | if (LogCompilation && xtty != NULL) { |
| 2131 | ttyLocker ttyl; |
| 2132 | // Missing MDP? Leave a small complaint in the log. |
| 2133 | xtty->elem("missing_mdp bci='%d'" , trap_bci); |
| 2134 | } |
| 2135 | } |
| 2136 | } |
| 2137 | |
| 2138 | // Return results: |
| 2139 | ret_this_trap_count = this_trap_count; |
| 2140 | ret_maybe_prior_trap = maybe_prior_trap; |
| 2141 | ret_maybe_prior_recompile = maybe_prior_recompile; |
| 2142 | return pdata; |
| 2143 | } |
| 2144 | |
| 2145 | void |
| 2146 | Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { |
| 2147 | ResourceMark rm; |
| 2148 | // Ignored outputs: |
| 2149 | uint ignore_this_trap_count; |
| 2150 | bool ignore_maybe_prior_trap; |
| 2151 | bool ignore_maybe_prior_recompile; |
| 2152 | assert(!reason_is_speculate(reason), "reason speculate only used by compiler" ); |
| 2153 | // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts |
| 2154 | bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); |
| 2155 | query_update_method_data(trap_mdo, trap_bci, |
| 2156 | (DeoptReason)reason, |
| 2157 | update_total_counts, |
| 2158 | #if INCLUDE_JVMCI |
| 2159 | false, |
| 2160 | #endif |
| 2161 | NULL, |
| 2162 | ignore_this_trap_count, |
| 2163 | ignore_maybe_prior_trap, |
| 2164 | ignore_maybe_prior_recompile); |
| 2165 | } |
| 2166 | |
| 2167 | Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { |
| 2168 | if (TraceDeoptimization) { |
| 2169 | tty->print("Uncommon trap " ); |
| 2170 | } |
| 2171 | // Still in Java no safepoints |
| 2172 | { |
| 2173 | // This enters VM and may safepoint |
| 2174 | uncommon_trap_inner(thread, trap_request); |
| 2175 | } |
| 2176 | return fetch_unroll_info_helper(thread, exec_mode); |
| 2177 | } |
| 2178 | |
| 2179 | // Local derived constants. |
| 2180 | // Further breakdown of DataLayout::trap_state, as promised by DataLayout. |
| 2181 | const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; |
| 2182 | const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; |
| 2183 | |
| 2184 | //---------------------------trap_state_reason--------------------------------- |
| 2185 | Deoptimization::DeoptReason |
| 2186 | Deoptimization::trap_state_reason(int trap_state) { |
| 2187 | // This assert provides the link between the width of DataLayout::trap_bits |
| 2188 | // and the encoding of "recorded" reasons. It ensures there are enough |
| 2189 | // bits to store all needed reasons in the per-BCI MDO profile. |
| 2190 | assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits" ); |
| 2191 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
| 2192 | trap_state -= recompile_bit; |
| 2193 | if (trap_state == DS_REASON_MASK) { |
| 2194 | return Reason_many; |
| 2195 | } else { |
| 2196 | assert((int)Reason_none == 0, "state=0 => Reason_none" ); |
| 2197 | return (DeoptReason)trap_state; |
| 2198 | } |
| 2199 | } |
| 2200 | //-------------------------trap_state_has_reason------------------------------- |
| 2201 | int Deoptimization::trap_state_has_reason(int trap_state, int reason) { |
| 2202 | assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason" ); |
| 2203 | assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits" ); |
| 2204 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
| 2205 | trap_state -= recompile_bit; |
| 2206 | if (trap_state == DS_REASON_MASK) { |
| 2207 | return -1; // true, unspecifically (bottom of state lattice) |
| 2208 | } else if (trap_state == reason) { |
| 2209 | return 1; // true, definitely |
| 2210 | } else if (trap_state == 0) { |
| 2211 | return 0; // false, definitely (top of state lattice) |
| 2212 | } else { |
| 2213 | return 0; // false, definitely |
| 2214 | } |
| 2215 | } |
| 2216 | //-------------------------trap_state_add_reason------------------------------- |
| 2217 | int Deoptimization::trap_state_add_reason(int trap_state, int reason) { |
| 2218 | assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason" ); |
| 2219 | int recompile_bit = (trap_state & DS_RECOMPILE_BIT); |
| 2220 | trap_state -= recompile_bit; |
| 2221 | if (trap_state == DS_REASON_MASK) { |
| 2222 | return trap_state + recompile_bit; // already at state lattice bottom |
| 2223 | } else if (trap_state == reason) { |
| 2224 | return trap_state + recompile_bit; // the condition is already true |
| 2225 | } else if (trap_state == 0) { |
| 2226 | return reason + recompile_bit; // no condition has yet been true |
| 2227 | } else { |
| 2228 | return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom |
| 2229 | } |
| 2230 | } |
| 2231 | //-----------------------trap_state_is_recompiled------------------------------ |
| 2232 | bool Deoptimization::trap_state_is_recompiled(int trap_state) { |
| 2233 | return (trap_state & DS_RECOMPILE_BIT) != 0; |
| 2234 | } |
| 2235 | //-----------------------trap_state_set_recompiled----------------------------- |
| 2236 | int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { |
| 2237 | if (z) return trap_state | DS_RECOMPILE_BIT; |
| 2238 | else return trap_state & ~DS_RECOMPILE_BIT; |
| 2239 | } |
| 2240 | //---------------------------format_trap_state--------------------------------- |
| 2241 | // This is used for debugging and diagnostics, including LogFile output. |
| 2242 | const char* Deoptimization::format_trap_state(char* buf, size_t buflen, |
| 2243 | int trap_state) { |
| 2244 | assert(buflen > 0, "sanity" ); |
| 2245 | DeoptReason reason = trap_state_reason(trap_state); |
| 2246 | bool recomp_flag = trap_state_is_recompiled(trap_state); |
| 2247 | // Re-encode the state from its decoded components. |
| 2248 | int decoded_state = 0; |
| 2249 | if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) |
| 2250 | decoded_state = trap_state_add_reason(decoded_state, reason); |
| 2251 | if (recomp_flag) |
| 2252 | decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); |
| 2253 | // If the state re-encodes properly, format it symbolically. |
| 2254 | // Because this routine is used for debugging and diagnostics, |
| 2255 | // be robust even if the state is a strange value. |
| 2256 | size_t len; |
| 2257 | if (decoded_state != trap_state) { |
| 2258 | // Random buggy state that doesn't decode?? |
| 2259 | len = jio_snprintf(buf, buflen, "#%d" , trap_state); |
| 2260 | } else { |
| 2261 | len = jio_snprintf(buf, buflen, "%s%s" , |
| 2262 | trap_reason_name(reason), |
| 2263 | recomp_flag ? " recompiled" : "" ); |
| 2264 | } |
| 2265 | return buf; |
| 2266 | } |
| 2267 | |
| 2268 | |
| 2269 | //--------------------------------statics-------------------------------------- |
| 2270 | const char* Deoptimization::_trap_reason_name[] = { |
| 2271 | // Note: Keep this in sync. with enum DeoptReason. |
| 2272 | "none" , |
| 2273 | "null_check" , |
| 2274 | "null_assert" JVMCI_ONLY("_or_unreached0" ), |
| 2275 | "range_check" , |
| 2276 | "class_check" , |
| 2277 | "array_check" , |
| 2278 | "intrinsic" JVMCI_ONLY("_or_type_checked_inlining" ), |
| 2279 | "bimorphic" JVMCI_ONLY("_or_optimized_type_check" ), |
| 2280 | "profile_predicate" , |
| 2281 | "unloaded" , |
| 2282 | "uninitialized" , |
| 2283 | "initialized" , |
| 2284 | "unreached" , |
| 2285 | "unhandled" , |
| 2286 | "constraint" , |
| 2287 | "div0_check" , |
| 2288 | "age" , |
| 2289 | "predicate" , |
| 2290 | "loop_limit_check" , |
| 2291 | "speculate_class_check" , |
| 2292 | "speculate_null_check" , |
| 2293 | "speculate_null_assert" , |
| 2294 | "rtm_state_change" , |
| 2295 | "unstable_if" , |
| 2296 | "unstable_fused_if" , |
| 2297 | #if INCLUDE_JVMCI |
| 2298 | "aliasing" , |
| 2299 | "transfer_to_interpreter" , |
| 2300 | "not_compiled_exception_handler" , |
| 2301 | "unresolved" , |
| 2302 | "jsr_mismatch" , |
| 2303 | #endif |
| 2304 | "tenured" |
| 2305 | }; |
| 2306 | const char* Deoptimization::_trap_action_name[] = { |
| 2307 | // Note: Keep this in sync. with enum DeoptAction. |
| 2308 | "none" , |
| 2309 | "maybe_recompile" , |
| 2310 | "reinterpret" , |
| 2311 | "make_not_entrant" , |
| 2312 | "make_not_compilable" |
| 2313 | }; |
| 2314 | |
| 2315 | const char* Deoptimization::trap_reason_name(int reason) { |
| 2316 | // Check that every reason has a name |
| 2317 | STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); |
| 2318 | |
| 2319 | if (reason == Reason_many) return "many" ; |
| 2320 | if ((uint)reason < Reason_LIMIT) |
| 2321 | return _trap_reason_name[reason]; |
| 2322 | static char buf[20]; |
| 2323 | sprintf(buf, "reason%d" , reason); |
| 2324 | return buf; |
| 2325 | } |
| 2326 | const char* Deoptimization::trap_action_name(int action) { |
| 2327 | // Check that every action has a name |
| 2328 | STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); |
| 2329 | |
| 2330 | if ((uint)action < Action_LIMIT) |
| 2331 | return _trap_action_name[action]; |
| 2332 | static char buf[20]; |
| 2333 | sprintf(buf, "action%d" , action); |
| 2334 | return buf; |
| 2335 | } |
| 2336 | |
| 2337 | // This is used for debugging and diagnostics, including LogFile output. |
| 2338 | const char* Deoptimization::format_trap_request(char* buf, size_t buflen, |
| 2339 | int trap_request) { |
| 2340 | jint unloaded_class_index = trap_request_index(trap_request); |
| 2341 | const char* reason = trap_reason_name(trap_request_reason(trap_request)); |
| 2342 | const char* action = trap_action_name(trap_request_action(trap_request)); |
| 2343 | #if INCLUDE_JVMCI |
| 2344 | int debug_id = trap_request_debug_id(trap_request); |
| 2345 | #endif |
| 2346 | size_t len; |
| 2347 | if (unloaded_class_index < 0) { |
| 2348 | len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'" ), |
| 2349 | reason, action |
| 2350 | #if INCLUDE_JVMCI |
| 2351 | ,debug_id |
| 2352 | #endif |
| 2353 | ); |
| 2354 | } else { |
| 2355 | len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'" ), |
| 2356 | reason, action, unloaded_class_index |
| 2357 | #if INCLUDE_JVMCI |
| 2358 | ,debug_id |
| 2359 | #endif |
| 2360 | ); |
| 2361 | } |
| 2362 | return buf; |
| 2363 | } |
| 2364 | |
| 2365 | juint Deoptimization::_deoptimization_hist |
| 2366 | [Deoptimization::Reason_LIMIT] |
| 2367 | [1 + Deoptimization::Action_LIMIT] |
| 2368 | [Deoptimization::BC_CASE_LIMIT] |
| 2369 | = {0}; |
| 2370 | |
| 2371 | enum { |
| 2372 | LSB_BITS = 8, |
| 2373 | LSB_MASK = right_n_bits(LSB_BITS) |
| 2374 | }; |
| 2375 | |
| 2376 | void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, |
| 2377 | Bytecodes::Code bc) { |
| 2378 | assert(reason >= 0 && reason < Reason_LIMIT, "oob" ); |
| 2379 | assert(action >= 0 && action < Action_LIMIT, "oob" ); |
| 2380 | _deoptimization_hist[Reason_none][0][0] += 1; // total |
| 2381 | _deoptimization_hist[reason][0][0] += 1; // per-reason total |
| 2382 | juint* cases = _deoptimization_hist[reason][1+action]; |
| 2383 | juint* bc_counter_addr = NULL; |
| 2384 | juint bc_counter = 0; |
| 2385 | // Look for an unused counter, or an exact match to this BC. |
| 2386 | if (bc != Bytecodes::_illegal) { |
| 2387 | for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { |
| 2388 | juint* counter_addr = &cases[bc_case]; |
| 2389 | juint counter = *counter_addr; |
| 2390 | if ((counter == 0 && bc_counter_addr == NULL) |
| 2391 | || (Bytecodes::Code)(counter & LSB_MASK) == bc) { |
| 2392 | // this counter is either free or is already devoted to this BC |
| 2393 | bc_counter_addr = counter_addr; |
| 2394 | bc_counter = counter | bc; |
| 2395 | } |
| 2396 | } |
| 2397 | } |
| 2398 | if (bc_counter_addr == NULL) { |
| 2399 | // Overflow, or no given bytecode. |
| 2400 | bc_counter_addr = &cases[BC_CASE_LIMIT-1]; |
| 2401 | bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB |
| 2402 | } |
| 2403 | *bc_counter_addr = bc_counter + (1 << LSB_BITS); |
| 2404 | } |
| 2405 | |
| 2406 | jint Deoptimization::total_deoptimization_count() { |
| 2407 | return _deoptimization_hist[Reason_none][0][0]; |
| 2408 | } |
| 2409 | |
| 2410 | void Deoptimization::print_statistics() { |
| 2411 | juint total = total_deoptimization_count(); |
| 2412 | juint account = total; |
| 2413 | if (total != 0) { |
| 2414 | ttyLocker ttyl; |
| 2415 | if (xtty != NULL) xtty->head("statistics type='deoptimization'" ); |
| 2416 | tty->print_cr("Deoptimization traps recorded:" ); |
| 2417 | #define PRINT_STAT_LINE(name, r) \ |
| 2418 | tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); |
| 2419 | PRINT_STAT_LINE("total" , total); |
| 2420 | // For each non-zero entry in the histogram, print the reason, |
| 2421 | // the action, and (if specifically known) the type of bytecode. |
| 2422 | for (int reason = 0; reason < Reason_LIMIT; reason++) { |
| 2423 | for (int action = 0; action < Action_LIMIT; action++) { |
| 2424 | juint* cases = _deoptimization_hist[reason][1+action]; |
| 2425 | for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { |
| 2426 | juint counter = cases[bc_case]; |
| 2427 | if (counter != 0) { |
| 2428 | char name[1*K]; |
| 2429 | Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); |
| 2430 | if (bc_case == BC_CASE_LIMIT && (int)bc == 0) |
| 2431 | bc = Bytecodes::_illegal; |
| 2432 | sprintf(name, "%s/%s/%s" , |
| 2433 | trap_reason_name(reason), |
| 2434 | trap_action_name(action), |
| 2435 | Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other" ); |
| 2436 | juint r = counter >> LSB_BITS; |
| 2437 | tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)" , name, r, (r * 100.0) / total); |
| 2438 | account -= r; |
| 2439 | } |
| 2440 | } |
| 2441 | } |
| 2442 | } |
| 2443 | if (account != 0) { |
| 2444 | PRINT_STAT_LINE("unaccounted" , account); |
| 2445 | } |
| 2446 | #undef PRINT_STAT_LINE |
| 2447 | if (xtty != NULL) xtty->tail("statistics" ); |
| 2448 | } |
| 2449 | } |
| 2450 | #else // COMPILER2_OR_JVMCI |
| 2451 | |
| 2452 | |
| 2453 | // Stubs for C1 only system. |
| 2454 | bool Deoptimization::trap_state_is_recompiled(int trap_state) { |
| 2455 | return false; |
| 2456 | } |
| 2457 | |
| 2458 | const char* Deoptimization::trap_reason_name(int reason) { |
| 2459 | return "unknown" ; |
| 2460 | } |
| 2461 | |
| 2462 | void Deoptimization::print_statistics() { |
| 2463 | // no output |
| 2464 | } |
| 2465 | |
| 2466 | void |
| 2467 | Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { |
| 2468 | // no udpate |
| 2469 | } |
| 2470 | |
| 2471 | int Deoptimization::trap_state_has_reason(int trap_state, int reason) { |
| 2472 | return 0; |
| 2473 | } |
| 2474 | |
| 2475 | void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, |
| 2476 | Bytecodes::Code bc) { |
| 2477 | // no update |
| 2478 | } |
| 2479 | |
| 2480 | const char* Deoptimization::format_trap_state(char* buf, size_t buflen, |
| 2481 | int trap_state) { |
| 2482 | jio_snprintf(buf, buflen, "#%d" , trap_state); |
| 2483 | return buf; |
| 2484 | } |
| 2485 | |
| 2486 | #endif // COMPILER2_OR_JVMCI |
| 2487 | |