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 #ifndef SHARE_RUNTIME_DEOPTIMIZATION_HPP 26 #define SHARE_RUNTIME_DEOPTIMIZATION_HPP 27 28 #include "memory/allocation.hpp" 29 #include "runtime/frame.hpp" 30 31 class ProfileData; 32 class vframeArray; 33 class MonitorInfo; 34 class MonitorValue; 35 class ObjectValue; 36 class AutoBoxObjectValue; 37 class ScopeValue; 38 class compiledVFrame; 39 40 template<class E> class GrowableArray; 41 42 class Deoptimization : AllStatic { 43 friend class VMStructs; 44 45 public: 46 // What condition caused the deoptimization? 47 enum DeoptReason { 48 Reason_many = -1, // indicates presence of several reasons 49 Reason_none = 0, // indicates absence of a relevant deopt. 50 // Next 8 reasons are recorded per bytecode in DataLayout::trap_bits. 51 // This is more complicated for JVMCI as JVMCI may deoptimize to *some* bytecode before the 52 // bytecode that actually caused the deopt (with inlining, JVMCI may even deoptimize to a 53 // bytecode in another method): 54 // - bytecode y in method b() causes deopt 55 // - JVMCI deoptimizes to bytecode x in method a() 56 // -> the deopt reason will be recorded for method a() at bytecode x 57 Reason_null_check, // saw unexpected null or zero divisor (@bci) 58 Reason_null_assert, // saw unexpected non-null or non-zero (@bci) 59 Reason_range_check, // saw unexpected array index (@bci) 60 Reason_class_check, // saw unexpected object class (@bci) 61 Reason_array_check, // saw unexpected array class (aastore @bci) 62 Reason_intrinsic, // saw unexpected operand to intrinsic (@bci) 63 Reason_bimorphic, // saw unexpected object class in bimorphic inlining (@bci) 64 65 #if INCLUDE_JVMCI 66 Reason_unreached0 = Reason_null_assert, 67 Reason_type_checked_inlining = Reason_intrinsic, 68 Reason_optimized_type_check = Reason_bimorphic, 69 #endif 70 71 Reason_profile_predicate, // compiler generated predicate moved from frequent branch in a loop failed 72 73 // recorded per method 74 Reason_unloaded, // unloaded class or constant pool entry 75 Reason_uninitialized, // bad class state (uninitialized) 76 Reason_initialized, // class has been fully initialized 77 Reason_unreached, // code is not reached, compiler 78 Reason_unhandled, // arbitrary compiler limitation 79 Reason_constraint, // arbitrary runtime constraint violated 80 Reason_div0_check, // a null_check due to division by zero 81 Reason_age, // nmethod too old; tier threshold reached 82 Reason_predicate, // compiler generated predicate failed 83 Reason_loop_limit_check, // compiler generated loop limits check failed 84 Reason_speculate_class_check, // saw unexpected object class from type speculation 85 Reason_speculate_null_check, // saw unexpected null from type speculation 86 Reason_speculate_null_assert, // saw unexpected null from type speculation 87 Reason_rtm_state_change, // rtm state change detected 88 Reason_unstable_if, // a branch predicted always false was taken 89 Reason_unstable_fused_if, // fused two ifs that had each one untaken branch. One is now taken. 90 #if INCLUDE_JVMCI 91 Reason_aliasing, // optimistic assumption about aliasing failed 92 Reason_transfer_to_interpreter, // explicit transferToInterpreter() 93 Reason_not_compiled_exception_handler, 94 Reason_unresolved, 95 Reason_jsr_mismatch, 96 #endif 97 98 // Reason_tenured is counted separately, add normal counted Reasons above. 99 // Related to MethodData::_trap_hist_limit where Reason_tenured isn't included 100 Reason_tenured, // age of the code has reached the limit 101 Reason_LIMIT, 102 103 // Note: Keep this enum in sync. with _trap_reason_name. 104 Reason_RECORDED_LIMIT = Reason_profile_predicate // some are not recorded per bc 105 // Note: Reason_RECORDED_LIMIT should fit into 31 bits of 106 // DataLayout::trap_bits. This dependency is enforced indirectly 107 // via asserts, to avoid excessive direct header-to-header dependencies. 108 // See Deoptimization::trap_state_reason and class DataLayout. 109 }; 110 111 // What action must be taken by the runtime? 112 enum DeoptAction { 113 Action_none, // just interpret, do not invalidate nmethod 114 Action_maybe_recompile, // recompile the nmethod; need not invalidate 115 Action_reinterpret, // invalidate the nmethod, reset IC, maybe recompile 116 Action_make_not_entrant, // invalidate the nmethod, recompile (probably) 117 Action_make_not_compilable, // invalidate the nmethod and do not compile 118 Action_LIMIT 119 // Note: Keep this enum in sync. with _trap_action_name. 120 }; 121 122 enum { 123 _action_bits = 3, 124 _reason_bits = 5, 125 _debug_id_bits = 23, 126 _action_shift = 0, 127 _reason_shift = _action_shift+_action_bits, 128 _debug_id_shift = _reason_shift+_reason_bits, 129 BC_CASE_LIMIT = PRODUCT_ONLY(1) NOT_PRODUCT(4) // for _deoptimization_hist 130 }; 131 132 enum UnpackType { 133 Unpack_deopt = 0, // normal deoptimization, use pc computed in unpack_vframe_on_stack 134 Unpack_exception = 1, // exception is pending 135 Unpack_uncommon_trap = 2, // redo last byte code (C2 only) 136 Unpack_reexecute = 3, // reexecute bytecode (C1 only) 137 Unpack_LIMIT = 4 138 }; 139 140 static void deoptimize_all_marked(); 141 142 private: 143 // Checks all compiled methods. Invalid methods are deleted and 144 // corresponding activations are deoptimized. 145 static int deoptimize_dependents(); 146 static void revoke_using_handshake(JavaThread* thread, frame fr, RegisterMap* map); 147 static void revoke_using_safepoint(JavaThread* thread, frame fr, RegisterMap* map); 148 static void deopt_thread(bool in_handshake, JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason); 149 150 public: 151 // Deoptimizes a frame lazily. nmethod gets patched deopt happens on return to the frame 152 static void deoptimize(JavaThread* thread, frame fr, RegisterMap *map, bool in_handshake = false); 153 static void deoptimize(JavaThread* thread, frame fr, RegisterMap *reg_map, DeoptReason reason); 154 155 #if INCLUDE_JVMCI 156 static address deoptimize_for_missing_exception_handler(CompiledMethod* cm); 157 static oop get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS); 158 #endif 159 160 private: 161 // Does the actual work for deoptimizing a single frame 162 static void deoptimize_single_frame(JavaThread* thread, frame fr, DeoptReason reason); 163 164 // Helper function to revoke biases of all monitors in frame if UseBiasedLocking 165 // is enabled 166 static void revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 167 revoke_using_safepoint(thread, fr, map); 168 } 169 170 #if COMPILER2_OR_JVMCI 171 JVMCI_ONLY(public:) 172 173 // Support for restoring non-escaping objects 174 static bool realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS); 175 static void reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type); 176 static void reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj); 177 static void reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal); 178 static void relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures); 179 static void pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array); 180 NOT_PRODUCT(static void print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures);) 181 #endif // COMPILER2_OR_JVMCI 182 183 public: 184 static vframeArray* create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures); 185 186 // Interface used for unpacking deoptimized frames 187 188 // UnrollBlock is returned by fetch_unroll_info() to the deoptimization handler (blob). 189 // This is only a CheapObj to ease debugging after a deopt failure 190 class UnrollBlock : public CHeapObj<mtCompiler> { 191 friend class VMStructs; 192 friend class JVMCIVMStructs; 193 private: 194 int _size_of_deoptimized_frame; // Size, in bytes, of current deoptimized frame 195 int _caller_adjustment; // Adjustment, in bytes, to caller's SP by initial interpreted frame 196 int _number_of_frames; // Number frames to unroll 197 int _total_frame_sizes; // Total of number*sizes frames 198 intptr_t* _frame_sizes; // Array of frame sizes, in bytes, for unrolling the stack 199 address* _frame_pcs; // Array of frame pc's, in bytes, for unrolling the stack 200 intptr_t* _register_block; // Block for storing callee-saved registers. 201 BasicType _return_type; // Tells if we have to restore double or long return value 202 intptr_t _initial_info; // Platform dependent data for the sender frame (was FP on x86) 203 int _caller_actual_parameters; // The number of actual arguments at the 204 // interpreted caller of the deoptimized frame 205 int _unpack_kind; // exec_mode that can be changed during fetch_unroll_info 206 207 // The following fields are used as temps during the unpacking phase 208 // (which is tight on registers, especially on x86). They really ought 209 // to be PD variables but that involves moving this class into its own 210 // file to use the pd include mechanism. Maybe in a later cleanup ... 211 intptr_t _counter_temp; // SHOULD BE PD VARIABLE (x86 frame count temp) 212 intptr_t _sender_sp_temp; // SHOULD BE PD VARIABLE (x86 sender_sp) 213 public: 214 // Constructor 215 UnrollBlock(int size_of_deoptimized_frame, 216 int caller_adjustment, 217 int caller_actual_parameters, 218 int number_of_frames, 219 intptr_t* frame_sizes, 220 address* frames_pcs, 221 BasicType return_type, 222 int unpack_kind); 223 ~UnrollBlock(); 224 225 // Returns where a register is located. 226 intptr_t* value_addr_at(int register_number) const; 227 228 // Accessors 229 intptr_t* frame_sizes() const { return _frame_sizes; } 230 int number_of_frames() const { return _number_of_frames; } 231 address* frame_pcs() const { return _frame_pcs ; } 232 int unpack_kind() const { return _unpack_kind; } 233 234 // Returns the total size of frames 235 int size_of_frames() const; 236 237 void set_initial_info(intptr_t info) { _initial_info = info; } 238 239 int caller_actual_parameters() const { return _caller_actual_parameters; } 240 241 // Accessors used by the code generator for the unpack stub. 242 static int size_of_deoptimized_frame_offset_in_bytes() { return offset_of(UnrollBlock, _size_of_deoptimized_frame); } 243 static int caller_adjustment_offset_in_bytes() { return offset_of(UnrollBlock, _caller_adjustment); } 244 static int number_of_frames_offset_in_bytes() { return offset_of(UnrollBlock, _number_of_frames); } 245 static int frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _frame_sizes); } 246 static int total_frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _total_frame_sizes); } 247 static int frame_pcs_offset_in_bytes() { return offset_of(UnrollBlock, _frame_pcs); } 248 static int register_block_offset_in_bytes() { return offset_of(UnrollBlock, _register_block); } 249 static int return_type_offset_in_bytes() { return offset_of(UnrollBlock, _return_type); } 250 static int counter_temp_offset_in_bytes() { return offset_of(UnrollBlock, _counter_temp); } 251 static int initial_info_offset_in_bytes() { return offset_of(UnrollBlock, _initial_info); } 252 static int unpack_kind_offset_in_bytes() { return offset_of(UnrollBlock, _unpack_kind); } 253 static int sender_sp_temp_offset_in_bytes() { return offset_of(UnrollBlock, _sender_sp_temp); } 254 255 BasicType return_type() const { return _return_type; } 256 void print(); 257 }; 258 259 //** Returns an UnrollBlock continuing information 260 // how to make room for the resulting interpreter frames. 261 // Called by assembly stub after execution has returned to 262 // deoptimized frame. 263 // @argument thread. Thread where stub_frame resides. 264 // @see OptoRuntime::deoptimization_fetch_unroll_info_C 265 static UnrollBlock* fetch_unroll_info(JavaThread* thread, int exec_mode); 266 267 //** Unpacks vframeArray onto execution stack 268 // Called by assembly stub after execution has returned to 269 // deoptimized frame and after the stack unrolling. 270 // @argument thread. Thread where stub_frame resides. 271 // @argument exec_mode. Determines how execution should be continued in top frame. 272 // 0 means continue after current byte code 273 // 1 means exception has happened, handle exception 274 // 2 means reexecute current bytecode (for uncommon traps). 275 // @see OptoRuntime::deoptimization_unpack_frames_C 276 // Return BasicType of call return type, if any 277 static BasicType unpack_frames(JavaThread* thread, int exec_mode); 278 279 // Cleans up deoptimization bits on thread after unpacking or in the 280 // case of an exception. 281 static void cleanup_deopt_info(JavaThread *thread, 282 vframeArray * array); 283 284 // Restores callee saved values from deoptimized frame into oldest interpreter frame 285 // so caller of the deoptimized frame will get back the values it expects. 286 static void unwind_callee_save_values(frame* f, vframeArray* vframe_array); 287 288 //** Performs an uncommon trap for compiled code. 289 // The top most compiler frame is converted into interpreter frames 290 static UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index, jint exec_mode); 291 // Helper routine that enters the VM and may block 292 static void uncommon_trap_inner(JavaThread* thread, jint unloaded_class_index); 293 294 //** Deoptimizes the frame identified by id. 295 // Only called from VMDeoptimizeFrame 296 // @argument thread. Thread where stub_frame resides. 297 // @argument id. id of frame that should be deoptimized. 298 static void deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason); 299 300 // if thread is not the current thread then execute 301 // VM_DeoptimizeFrame otherwise deoptimize directly. 302 static void deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason); 303 static void deoptimize_frame(JavaThread* thread, intptr_t* id); 304 305 // Statistics 306 static void gather_statistics(DeoptReason reason, DeoptAction action, 307 Bytecodes::Code bc = Bytecodes::_illegal); 308 static void print_statistics(); 309 310 // How much room to adjust the last frame's SP by, to make space for 311 // the callee's interpreter frame (which expects locals to be next to 312 // incoming arguments) 313 static int last_frame_adjust(int callee_parameters, int callee_locals); 314 315 // trap_request codes 316 static DeoptReason trap_request_reason(int trap_request) { 317 if (trap_request < 0) 318 return (DeoptReason) 319 ((~(trap_request) >> _reason_shift) & right_n_bits(_reason_bits)); 320 else 321 // standard reason for unloaded CP entry 322 return Reason_unloaded; 323 } 324 static DeoptAction trap_request_action(int trap_request) { 325 if (trap_request < 0) 326 return (DeoptAction) 327 ((~(trap_request) >> _action_shift) & right_n_bits(_action_bits)); 328 else 329 // standard action for unloaded CP entry 330 return _unloaded_action; 331 } 332 static int trap_request_debug_id(int trap_request) { 333 if (trap_request < 0) { 334 return ((~(trap_request) >> _debug_id_shift) & right_n_bits(_debug_id_bits)); 335 } else { 336 // standard action for unloaded CP entry 337 return 0; 338 } 339 } 340 static int trap_request_index(int trap_request) { 341 if (trap_request < 0) 342 return -1; 343 else 344 return trap_request; 345 } 346 static int make_trap_request(DeoptReason reason, DeoptAction action, 347 int index = -1) { 348 assert((1 << _reason_bits) >= Reason_LIMIT, "enough bits"); 349 assert((1 << _action_bits) >= Action_LIMIT, "enough bits"); 350 int trap_request; 351 if (index != -1) 352 trap_request = index; 353 else 354 trap_request = (~(((reason) << _reason_shift) 355 + ((action) << _action_shift))); 356 assert(reason == trap_request_reason(trap_request), "valid reason"); 357 assert(action == trap_request_action(trap_request), "valid action"); 358 assert(index == trap_request_index(trap_request), "valid index"); 359 return trap_request; 360 } 361 362 // The trap_state stored in a MDO is decoded here. 363 // It records two items of information. 364 // reason: If a deoptimization happened here, what its reason was, 365 // or if there were multiple deopts with differing reasons. 366 // recompiled: If a deoptimization here triggered a recompilation. 367 // Note that not all reasons are recorded per-bci. 368 static DeoptReason trap_state_reason(int trap_state); 369 static int trap_state_has_reason(int trap_state, int reason); 370 static int trap_state_add_reason(int trap_state, int reason); 371 static bool trap_state_is_recompiled(int trap_state); 372 static int trap_state_set_recompiled(int trap_state, bool z); 373 static const char* format_trap_state(char* buf, size_t buflen, 374 int trap_state); 375 376 static bool reason_is_recorded_per_bytecode(DeoptReason reason) { 377 return reason > Reason_none && reason <= Reason_RECORDED_LIMIT; 378 } 379 380 static DeoptReason reason_recorded_per_bytecode_if_any(DeoptReason reason) { 381 if (reason_is_recorded_per_bytecode(reason)) 382 return reason; 383 else if (reason == Reason_div0_check) // null check due to divide-by-zero? 384 return Reason_null_check; // recorded per BCI as a null check 385 else if (reason == Reason_speculate_class_check) 386 return Reason_class_check; 387 else if (reason == Reason_speculate_null_check) 388 return Reason_null_check; 389 else if (reason == Reason_speculate_null_assert) 390 return Reason_null_assert; 391 else if (reason == Reason_unstable_if) 392 return Reason_intrinsic; 393 else if (reason == Reason_unstable_fused_if) 394 return Reason_range_check; 395 else 396 return Reason_none; 397 } 398 399 static bool reason_is_speculate(int reason) { 400 if (reason == Reason_speculate_class_check || 401 reason == Reason_speculate_null_check || 402 reason == Reason_speculate_null_assert) { 403 return true; 404 } 405 return false; 406 } 407 408 static DeoptReason reason_null_check(bool speculative) { 409 return speculative ? Deoptimization::Reason_speculate_null_check : Deoptimization::Reason_null_check; 410 } 411 412 static DeoptReason reason_class_check(bool speculative) { 413 return speculative ? Deoptimization::Reason_speculate_class_check : Deoptimization::Reason_class_check; 414 } 415 416 static DeoptReason reason_null_assert(bool speculative) { 417 return speculative ? Deoptimization::Reason_speculate_null_assert : Deoptimization::Reason_null_assert; 418 } 419 420 static uint per_method_trap_limit(int reason) { 421 return reason_is_speculate(reason) ? (uint)PerMethodSpecTrapLimit : (uint)PerMethodTrapLimit; 422 } 423 424 static const char* trap_reason_name(int reason); 425 static const char* trap_action_name(int action); 426 // Format like reason='foo' action='bar' index='123'. 427 // This is suitable both for XML and for tty output. 428 static const char* format_trap_request(char* buf, size_t buflen, 429 int trap_request); 430 431 static jint total_deoptimization_count(); 432 433 // JVMTI PopFrame support 434 435 // Preserves incoming arguments to the popped frame when it is 436 // returning to a deoptimized caller 437 static void popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address); 438 439 static MethodData* get_method_data(JavaThread* thread, const methodHandle& m, bool create_if_missing); 440 private: 441 // Update the mdo's count and per-BCI reason bits, returning previous state: 442 static ProfileData* query_update_method_data(MethodData* trap_mdo, 443 int trap_bci, 444 DeoptReason reason, 445 bool update_total_trap_count, 446 #if INCLUDE_JVMCI 447 bool is_osr, 448 #endif 449 Method* compiled_method, 450 //outputs: 451 uint& ret_this_trap_count, 452 bool& ret_maybe_prior_trap, 453 bool& ret_maybe_prior_recompile); 454 // class loading support for uncommon trap 455 static void load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS); 456 static void load_class_by_index(const constantPoolHandle& constant_pool, int index); 457 458 static UnrollBlock* fetch_unroll_info_helper(JavaThread* thread, int exec_mode); 459 460 static DeoptAction _unloaded_action; // == Action_reinterpret; 461 static const char* _trap_reason_name[]; 462 static const char* _trap_action_name[]; 463 464 static juint _deoptimization_hist[Reason_LIMIT][1+Action_LIMIT][BC_CASE_LIMIT]; 465 // Note: Histogram array size is 1-2 Kb. 466 467 public: 468 static void update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason); 469 }; 470 471 class DeoptimizationMarker : StackObj { // for profiling 472 static bool _is_active; 473 public: 474 DeoptimizationMarker() { _is_active = true; } 475 ~DeoptimizationMarker() { _is_active = false; } 476 static bool is_active() { return _is_active; } 477 }; 478 479 #endif // SHARE_RUNTIME_DEOPTIMIZATION_HPP