1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/share/vm/runtime/deoptimization.hpp Wed Apr 27 01:25:04 2016 +0800 1.3 @@ -0,0 +1,385 @@ 1.4 +/* 1.5 + * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 1.23 + * or visit www.oracle.com if you need additional information or have any 1.24 + * questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#ifndef SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP 1.29 +#define SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP 1.30 + 1.31 +#include "memory/allocation.hpp" 1.32 +#include "runtime/frame.inline.hpp" 1.33 + 1.34 +class ProfileData; 1.35 +class vframeArray; 1.36 +class MonitorValue; 1.37 +class ObjectValue; 1.38 + 1.39 +class Deoptimization : AllStatic { 1.40 + friend class VMStructs; 1.41 + 1.42 + public: 1.43 + // What condition caused the deoptimization? 1.44 + enum DeoptReason { 1.45 + Reason_many = -1, // indicates presence of several reasons 1.46 + Reason_none = 0, // indicates absence of a relevant deopt. 1.47 + // Next 7 reasons are recorded per bytecode in DataLayout::trap_bits 1.48 + Reason_null_check, // saw unexpected null or zero divisor (@bci) 1.49 + Reason_null_assert, // saw unexpected non-null or non-zero (@bci) 1.50 + Reason_range_check, // saw unexpected array index (@bci) 1.51 + Reason_class_check, // saw unexpected object class (@bci) 1.52 + Reason_array_check, // saw unexpected array class (aastore @bci) 1.53 + Reason_intrinsic, // saw unexpected operand to intrinsic (@bci) 1.54 + Reason_bimorphic, // saw unexpected object class in bimorphic inlining (@bci) 1.55 + 1.56 + Reason_unloaded, // unloaded class or constant pool entry 1.57 + Reason_uninitialized, // bad class state (uninitialized) 1.58 + Reason_unreached, // code is not reached, compiler 1.59 + Reason_unhandled, // arbitrary compiler limitation 1.60 + Reason_constraint, // arbitrary runtime constraint violated 1.61 + Reason_div0_check, // a null_check due to division by zero 1.62 + Reason_age, // nmethod too old; tier threshold reached 1.63 + Reason_predicate, // compiler generated predicate failed 1.64 + Reason_loop_limit_check, // compiler generated loop limits check failed 1.65 + Reason_speculate_class_check, // saw unexpected object class from type speculation 1.66 + Reason_rtm_state_change, // rtm state change detected 1.67 + Reason_LIMIT, 1.68 + // Note: Keep this enum in sync. with _trap_reason_name. 1.69 + Reason_RECORDED_LIMIT = Reason_bimorphic // some are not recorded per bc 1.70 + // Note: Reason_RECORDED_LIMIT should be < 8 to fit into 3 bits of 1.71 + // DataLayout::trap_bits. This dependency is enforced indirectly 1.72 + // via asserts, to avoid excessive direct header-to-header dependencies. 1.73 + // See Deoptimization::trap_state_reason and class DataLayout. 1.74 + }; 1.75 + 1.76 + // What action must be taken by the runtime? 1.77 + enum DeoptAction { 1.78 + Action_none, // just interpret, do not invalidate nmethod 1.79 + Action_maybe_recompile, // recompile the nmethod; need not invalidate 1.80 + Action_reinterpret, // invalidate the nmethod, reset IC, maybe recompile 1.81 + Action_make_not_entrant, // invalidate the nmethod, recompile (probably) 1.82 + Action_make_not_compilable, // invalidate the nmethod and do not compile 1.83 + Action_LIMIT 1.84 + // Note: Keep this enum in sync. with _trap_action_name. 1.85 + }; 1.86 + 1.87 + enum { 1.88 + _action_bits = 3, 1.89 + _reason_bits = 5, 1.90 + _action_shift = 0, 1.91 + _reason_shift = _action_shift+_action_bits, 1.92 + BC_CASE_LIMIT = PRODUCT_ONLY(1) NOT_PRODUCT(4) // for _deoptimization_hist 1.93 + }; 1.94 + 1.95 + enum UnpackType { 1.96 + Unpack_deopt = 0, // normal deoptimization, use pc computed in unpack_vframe_on_stack 1.97 + Unpack_exception = 1, // exception is pending 1.98 + Unpack_uncommon_trap = 2, // redo last byte code (C2 only) 1.99 + Unpack_reexecute = 3 // reexecute bytecode (C1 only) 1.100 + }; 1.101 + 1.102 + // Checks all compiled methods. Invalid methods are deleted and 1.103 + // corresponding activations are deoptimized. 1.104 + static int deoptimize_dependents(); 1.105 + 1.106 + // Deoptimizes a frame lazily. nmethod gets patched deopt happens on return to the frame 1.107 + static void deoptimize(JavaThread* thread, frame fr, RegisterMap *reg_map); 1.108 + 1.109 + private: 1.110 + // Does the actual work for deoptimizing a single frame 1.111 + static void deoptimize_single_frame(JavaThread* thread, frame fr); 1.112 + 1.113 + // Helper function to revoke biases of all monitors in frame if UseBiasedLocking 1.114 + // is enabled 1.115 + static void revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map); 1.116 + // Helper function to revoke biases of all monitors in frames 1.117 + // executing in a particular CodeBlob if UseBiasedLocking is enabled 1.118 + static void revoke_biases_of_monitors(CodeBlob* cb); 1.119 + 1.120 +#ifdef COMPILER2 1.121 + // Support for restoring non-escaping objects 1.122 + static bool realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS); 1.123 + static void reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type); 1.124 + static void reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj); 1.125 + static void reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects); 1.126 + static void relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread); 1.127 + NOT_PRODUCT(static void print_objects(GrowableArray<ScopeValue*>* objects);) 1.128 +#endif // COMPILER2 1.129 + 1.130 + public: 1.131 + static vframeArray* create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk); 1.132 + 1.133 + // Interface used for unpacking deoptimized frames 1.134 + 1.135 + // UnrollBlock is returned by fetch_unroll_info() to the deoptimization handler (blob). 1.136 + // This is only a CheapObj to ease debugging after a deopt failure 1.137 + class UnrollBlock : public CHeapObj<mtCompiler> { 1.138 + private: 1.139 + int _size_of_deoptimized_frame; // Size, in bytes, of current deoptimized frame 1.140 + int _caller_adjustment; // Adjustment, in bytes, to caller's SP by initial interpreted frame 1.141 + int _number_of_frames; // Number frames to unroll 1.142 + int _total_frame_sizes; // Total of number*sizes frames 1.143 + intptr_t* _frame_sizes; // Array of frame sizes, in bytes, for unrolling the stack 1.144 + address* _frame_pcs; // Array of frame pc's, in bytes, for unrolling the stack 1.145 + intptr_t* _register_block; // Block for storing callee-saved registers. 1.146 + BasicType _return_type; // Tells if we have to restore double or long return value 1.147 + intptr_t _initial_info; // Platform dependent data for the sender frame (was FP on x86) 1.148 + int _caller_actual_parameters; // The number of actual arguments at the 1.149 + // interpreted caller of the deoptimized frame 1.150 + 1.151 + // The following fields are used as temps during the unpacking phase 1.152 + // (which is tight on registers, especially on x86). They really ought 1.153 + // to be PD variables but that involves moving this class into its own 1.154 + // file to use the pd include mechanism. Maybe in a later cleanup ... 1.155 + intptr_t _counter_temp; // SHOULD BE PD VARIABLE (x86 frame count temp) 1.156 + intptr_t _unpack_kind; // SHOULD BE PD VARIABLE (x86 unpack kind) 1.157 + intptr_t _sender_sp_temp; // SHOULD BE PD VARIABLE (x86 sender_sp) 1.158 + public: 1.159 + // Constructor 1.160 + UnrollBlock(int size_of_deoptimized_frame, 1.161 + int caller_adjustment, 1.162 + int caller_actual_parameters, 1.163 + int number_of_frames, 1.164 + intptr_t* frame_sizes, 1.165 + address* frames_pcs, 1.166 + BasicType return_type); 1.167 + ~UnrollBlock(); 1.168 + 1.169 + // Returns where a register is located. 1.170 + intptr_t* value_addr_at(int register_number) const; 1.171 + 1.172 + // Accessors 1.173 + intptr_t* frame_sizes() const { return _frame_sizes; } 1.174 + int number_of_frames() const { return _number_of_frames; } 1.175 + address* frame_pcs() const { return _frame_pcs ; } 1.176 + 1.177 + // Returns the total size of frames 1.178 + int size_of_frames() const; 1.179 + 1.180 + void set_initial_info(intptr_t info) { _initial_info = info; } 1.181 + 1.182 + int caller_actual_parameters() const { return _caller_actual_parameters; } 1.183 + 1.184 + // Accessors used by the code generator for the unpack stub. 1.185 + static int size_of_deoptimized_frame_offset_in_bytes() { return offset_of(UnrollBlock, _size_of_deoptimized_frame); } 1.186 + static int caller_adjustment_offset_in_bytes() { return offset_of(UnrollBlock, _caller_adjustment); } 1.187 + static int number_of_frames_offset_in_bytes() { return offset_of(UnrollBlock, _number_of_frames); } 1.188 + static int frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _frame_sizes); } 1.189 + static int total_frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _total_frame_sizes); } 1.190 + static int frame_pcs_offset_in_bytes() { return offset_of(UnrollBlock, _frame_pcs); } 1.191 + static int register_block_offset_in_bytes() { return offset_of(UnrollBlock, _register_block); } 1.192 + static int return_type_offset_in_bytes() { return offset_of(UnrollBlock, _return_type); } 1.193 + static int counter_temp_offset_in_bytes() { return offset_of(UnrollBlock, _counter_temp); } 1.194 + static int initial_info_offset_in_bytes() { return offset_of(UnrollBlock, _initial_info); } 1.195 + static int unpack_kind_offset_in_bytes() { return offset_of(UnrollBlock, _unpack_kind); } 1.196 + static int sender_sp_temp_offset_in_bytes() { return offset_of(UnrollBlock, _sender_sp_temp); } 1.197 + 1.198 + BasicType return_type() const { return _return_type; } 1.199 + void print(); 1.200 + }; 1.201 + 1.202 + //** Returns an UnrollBlock continuing information 1.203 + // how to make room for the resulting interpreter frames. 1.204 + // Called by assembly stub after execution has returned to 1.205 + // deoptimized frame. 1.206 + // @argument thread. Thread where stub_frame resides. 1.207 + // @see OptoRuntime::deoptimization_fetch_unroll_info_C 1.208 + static UnrollBlock* fetch_unroll_info(JavaThread* thread); 1.209 + 1.210 + //** Unpacks vframeArray onto execution stack 1.211 + // Called by assembly stub after execution has returned to 1.212 + // deoptimized frame and after the stack unrolling. 1.213 + // @argument thread. Thread where stub_frame resides. 1.214 + // @argument exec_mode. Determines how execution should be continuted in top frame. 1.215 + // 0 means continue after current byte code 1.216 + // 1 means exception has happened, handle exception 1.217 + // 2 means reexecute current bytecode (for uncommon traps). 1.218 + // @see OptoRuntime::deoptimization_unpack_frames_C 1.219 + // Return BasicType of call return type, if any 1.220 + static BasicType unpack_frames(JavaThread* thread, int exec_mode); 1.221 + 1.222 + // Cleans up deoptimization bits on thread after unpacking or in the 1.223 + // case of an exception. 1.224 + static void cleanup_deopt_info(JavaThread *thread, 1.225 + vframeArray * array); 1.226 + 1.227 + // Restores callee saved values from deoptimized frame into oldest interpreter frame 1.228 + // so caller of the deoptimized frame will get back the values it expects. 1.229 + static void unwind_callee_save_values(frame* f, vframeArray* vframe_array); 1.230 + 1.231 + //** Performs an uncommon trap for compiled code. 1.232 + // The top most compiler frame is converted into interpreter frames 1.233 + static UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); 1.234 + // Helper routine that enters the VM and may block 1.235 + static void uncommon_trap_inner(JavaThread* thread, jint unloaded_class_index); 1.236 + 1.237 + //** Deoptimizes the frame identified by id. 1.238 + // Only called from VMDeoptimizeFrame 1.239 + // @argument thread. Thread where stub_frame resides. 1.240 + // @argument id. id of frame that should be deoptimized. 1.241 + static void deoptimize_frame_internal(JavaThread* thread, intptr_t* id); 1.242 + 1.243 + // If thread is not the current thread then execute 1.244 + // VM_DeoptimizeFrame otherwise deoptimize directly. 1.245 + static void deoptimize_frame(JavaThread* thread, intptr_t* id); 1.246 + 1.247 + // Statistics 1.248 + static void gather_statistics(DeoptReason reason, DeoptAction action, 1.249 + Bytecodes::Code bc = Bytecodes::_illegal); 1.250 + static void print_statistics(); 1.251 + 1.252 + // How much room to adjust the last frame's SP by, to make space for 1.253 + // the callee's interpreter frame (which expects locals to be next to 1.254 + // incoming arguments) 1.255 + static int last_frame_adjust(int callee_parameters, int callee_locals); 1.256 + 1.257 + // trap_request codes 1.258 + static DeoptReason trap_request_reason(int trap_request) { 1.259 + if (trap_request < 0) 1.260 + return (DeoptReason) 1.261 + ((~(trap_request) >> _reason_shift) & right_n_bits(_reason_bits)); 1.262 + else 1.263 + // standard reason for unloaded CP entry 1.264 + return Reason_unloaded; 1.265 + } 1.266 + static DeoptAction trap_request_action(int trap_request) { 1.267 + if (trap_request < 0) 1.268 + return (DeoptAction) 1.269 + ((~(trap_request) >> _action_shift) & right_n_bits(_action_bits)); 1.270 + else 1.271 + // standard action for unloaded CP entry 1.272 + return _unloaded_action; 1.273 + } 1.274 + static int trap_request_index(int trap_request) { 1.275 + if (trap_request < 0) 1.276 + return -1; 1.277 + else 1.278 + return trap_request; 1.279 + } 1.280 + static int make_trap_request(DeoptReason reason, DeoptAction action, 1.281 + int index = -1) { 1.282 + assert((1 << _reason_bits) >= Reason_LIMIT, "enough bits"); 1.283 + assert((1 << _action_bits) >= Action_LIMIT, "enough bits"); 1.284 + int trap_request; 1.285 + if (index != -1) 1.286 + trap_request = index; 1.287 + else 1.288 + trap_request = (~(((reason) << _reason_shift) 1.289 + + ((action) << _action_shift))); 1.290 + assert(reason == trap_request_reason(trap_request), "valid reason"); 1.291 + assert(action == trap_request_action(trap_request), "valid action"); 1.292 + assert(index == trap_request_index(trap_request), "valid index"); 1.293 + return trap_request; 1.294 + } 1.295 + 1.296 + // The trap_state stored in a MDO is decoded here. 1.297 + // It records two items of information. 1.298 + // reason: If a deoptimization happened here, what its reason was, 1.299 + // or if there were multiple deopts with differing reasons. 1.300 + // recompiled: If a deoptimization here triggered a recompilation. 1.301 + // Note that not all reasons are recorded per-bci. 1.302 + static DeoptReason trap_state_reason(int trap_state); 1.303 + static int trap_state_has_reason(int trap_state, int reason); 1.304 + static int trap_state_add_reason(int trap_state, int reason); 1.305 + static bool trap_state_is_recompiled(int trap_state); 1.306 + static int trap_state_set_recompiled(int trap_state, bool z); 1.307 + static const char* format_trap_state(char* buf, size_t buflen, 1.308 + int trap_state); 1.309 + 1.310 + static bool reason_is_recorded_per_bytecode(DeoptReason reason) { 1.311 + return reason > Reason_none && reason <= Reason_RECORDED_LIMIT; 1.312 + } 1.313 + 1.314 + static DeoptReason reason_recorded_per_bytecode_if_any(DeoptReason reason) { 1.315 + if (reason_is_recorded_per_bytecode(reason)) 1.316 + return reason; 1.317 + else if (reason == Reason_div0_check) // null check due to divide-by-zero? 1.318 + return Reason_null_check; // recorded per BCI as a null check 1.319 + else if (reason == Reason_speculate_class_check) 1.320 + return Reason_class_check; 1.321 + else 1.322 + return Reason_none; 1.323 + } 1.324 + 1.325 + static bool reason_is_speculate(int reason) { 1.326 + if (reason == Reason_speculate_class_check) { 1.327 + return true; 1.328 + } 1.329 + return false; 1.330 + } 1.331 + 1.332 + static uint per_method_trap_limit(int reason) { 1.333 + return reason_is_speculate(reason) ? (uint)PerMethodSpecTrapLimit : (uint)PerMethodTrapLimit; 1.334 + } 1.335 + 1.336 + static const char* trap_reason_name(int reason); 1.337 + static const char* trap_action_name(int action); 1.338 + // Format like reason='foo' action='bar' index='123'. 1.339 + // This is suitable both for XML and for tty output. 1.340 + static const char* format_trap_request(char* buf, size_t buflen, 1.341 + int trap_request); 1.342 + 1.343 + static jint total_deoptimization_count(); 1.344 + static jint deoptimization_count(DeoptReason reason); 1.345 + 1.346 + // JVMTI PopFrame support 1.347 + 1.348 + // Preserves incoming arguments to the popped frame when it is 1.349 + // returning to a deoptimized caller 1.350 + static void popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address); 1.351 + 1.352 + private: 1.353 + static MethodData* get_method_data(JavaThread* thread, methodHandle m, bool create_if_missing); 1.354 + // Update the mdo's count and per-BCI reason bits, returning previous state: 1.355 + static ProfileData* query_update_method_data(MethodData* trap_mdo, 1.356 + int trap_bci, 1.357 + DeoptReason reason, 1.358 + Method* compiled_method, 1.359 + //outputs: 1.360 + uint& ret_this_trap_count, 1.361 + bool& ret_maybe_prior_trap, 1.362 + bool& ret_maybe_prior_recompile); 1.363 + // class loading support for uncommon trap 1.364 + static void load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS); 1.365 + static void load_class_by_index(constantPoolHandle constant_pool, int index); 1.366 + 1.367 + static UnrollBlock* fetch_unroll_info_helper(JavaThread* thread); 1.368 + 1.369 + static DeoptAction _unloaded_action; // == Action_reinterpret; 1.370 + static const char* _trap_reason_name[Reason_LIMIT]; 1.371 + static const char* _trap_action_name[Action_LIMIT]; 1.372 + 1.373 + static juint _deoptimization_hist[Reason_LIMIT][1+Action_LIMIT][BC_CASE_LIMIT]; 1.374 + // Note: Histogram array size is 1-2 Kb. 1.375 + 1.376 + public: 1.377 + static void update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason); 1.378 +}; 1.379 + 1.380 +class DeoptimizationMarker : StackObj { // for profiling 1.381 + static bool _is_active; 1.382 +public: 1.383 + DeoptimizationMarker() { _is_active = true; } 1.384 + ~DeoptimizationMarker() { _is_active = false; } 1.385 + static bool is_active() { return _is_active; } 1.386 +}; 1.387 + 1.388 +#endif // SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP