aoqi@0: /* aoqi@0: * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. aoqi@0: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. aoqi@0: * aoqi@0: * This code is free software; you can redistribute it and/or modify it aoqi@0: * under the terms of the GNU General Public License version 2 only, as aoqi@0: * published by the Free Software Foundation. aoqi@0: * aoqi@0: * This code is distributed in the hope that it will be useful, but WITHOUT aoqi@0: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or aoqi@0: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License aoqi@0: * version 2 for more details (a copy is included in the LICENSE file that aoqi@0: * accompanied this code). aoqi@0: * aoqi@0: * You should have received a copy of the GNU General Public License version aoqi@0: * 2 along with this work; if not, write to the Free Software Foundation, aoqi@0: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. aoqi@0: * aoqi@0: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA aoqi@0: * or visit www.oracle.com if you need additional information or have any aoqi@0: * questions. aoqi@0: * aoqi@0: */ aoqi@0: aoqi@0: #ifndef SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP aoqi@0: #define SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP aoqi@0: aoqi@0: #include "memory/allocation.hpp" aoqi@0: #include "runtime/frame.inline.hpp" aoqi@0: aoqi@0: class ProfileData; aoqi@0: class vframeArray; aoqi@0: class MonitorValue; aoqi@0: class ObjectValue; aoqi@0: aoqi@0: class Deoptimization : AllStatic { aoqi@0: friend class VMStructs; aoqi@0: aoqi@0: public: aoqi@0: // What condition caused the deoptimization? aoqi@0: enum DeoptReason { aoqi@0: Reason_many = -1, // indicates presence of several reasons aoqi@0: Reason_none = 0, // indicates absence of a relevant deopt. aoqi@0: // Next 7 reasons are recorded per bytecode in DataLayout::trap_bits aoqi@0: Reason_null_check, // saw unexpected null or zero divisor (@bci) aoqi@0: Reason_null_assert, // saw unexpected non-null or non-zero (@bci) aoqi@0: Reason_range_check, // saw unexpected array index (@bci) aoqi@0: Reason_class_check, // saw unexpected object class (@bci) aoqi@0: Reason_array_check, // saw unexpected array class (aastore @bci) aoqi@0: Reason_intrinsic, // saw unexpected operand to intrinsic (@bci) aoqi@0: Reason_bimorphic, // saw unexpected object class in bimorphic inlining (@bci) aoqi@0: aoqi@0: Reason_unloaded, // unloaded class or constant pool entry aoqi@0: Reason_uninitialized, // bad class state (uninitialized) aoqi@0: Reason_unreached, // code is not reached, compiler aoqi@0: Reason_unhandled, // arbitrary compiler limitation aoqi@0: Reason_constraint, // arbitrary runtime constraint violated aoqi@0: Reason_div0_check, // a null_check due to division by zero aoqi@0: Reason_age, // nmethod too old; tier threshold reached aoqi@0: Reason_predicate, // compiler generated predicate failed aoqi@0: Reason_loop_limit_check, // compiler generated loop limits check failed aoqi@0: Reason_speculate_class_check, // saw unexpected object class from type speculation aoqi@0: Reason_rtm_state_change, // rtm state change detected rbackman@7153: Reason_unstable_if, // a branch predicted always false was taken aoqi@0: Reason_LIMIT, aoqi@0: // Note: Keep this enum in sync. with _trap_reason_name. aoqi@0: Reason_RECORDED_LIMIT = Reason_bimorphic // some are not recorded per bc aoqi@0: // Note: Reason_RECORDED_LIMIT should be < 8 to fit into 3 bits of aoqi@0: // DataLayout::trap_bits. This dependency is enforced indirectly aoqi@0: // via asserts, to avoid excessive direct header-to-header dependencies. aoqi@0: // See Deoptimization::trap_state_reason and class DataLayout. aoqi@0: }; aoqi@0: aoqi@0: // What action must be taken by the runtime? aoqi@0: enum DeoptAction { aoqi@0: Action_none, // just interpret, do not invalidate nmethod aoqi@0: Action_maybe_recompile, // recompile the nmethod; need not invalidate aoqi@0: Action_reinterpret, // invalidate the nmethod, reset IC, maybe recompile aoqi@0: Action_make_not_entrant, // invalidate the nmethod, recompile (probably) aoqi@0: Action_make_not_compilable, // invalidate the nmethod and do not compile aoqi@0: Action_LIMIT aoqi@0: // Note: Keep this enum in sync. with _trap_action_name. aoqi@0: }; aoqi@0: aoqi@0: enum { aoqi@0: _action_bits = 3, aoqi@0: _reason_bits = 5, aoqi@0: _action_shift = 0, aoqi@0: _reason_shift = _action_shift+_action_bits, aoqi@0: BC_CASE_LIMIT = PRODUCT_ONLY(1) NOT_PRODUCT(4) // for _deoptimization_hist aoqi@0: }; aoqi@0: aoqi@0: enum UnpackType { aoqi@0: Unpack_deopt = 0, // normal deoptimization, use pc computed in unpack_vframe_on_stack aoqi@0: Unpack_exception = 1, // exception is pending aoqi@0: Unpack_uncommon_trap = 2, // redo last byte code (C2 only) aoqi@0: Unpack_reexecute = 3 // reexecute bytecode (C1 only) aoqi@0: }; aoqi@0: aoqi@0: // Checks all compiled methods. Invalid methods are deleted and aoqi@0: // corresponding activations are deoptimized. aoqi@0: static int deoptimize_dependents(); aoqi@0: aoqi@0: // Deoptimizes a frame lazily. nmethod gets patched deopt happens on return to the frame aoqi@0: static void deoptimize(JavaThread* thread, frame fr, RegisterMap *reg_map); aoqi@0: aoqi@0: private: aoqi@0: // Does the actual work for deoptimizing a single frame aoqi@0: static void deoptimize_single_frame(JavaThread* thread, frame fr); aoqi@0: aoqi@0: // Helper function to revoke biases of all monitors in frame if UseBiasedLocking aoqi@0: // is enabled aoqi@0: static void revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map); aoqi@0: // Helper function to revoke biases of all monitors in frames aoqi@0: // executing in a particular CodeBlob if UseBiasedLocking is enabled aoqi@0: static void revoke_biases_of_monitors(CodeBlob* cb); aoqi@0: aoqi@0: #ifdef COMPILER2 aoqi@0: // Support for restoring non-escaping objects aoqi@0: static bool realloc_objects(JavaThread* thread, frame* fr, GrowableArray* objects, TRAPS); aoqi@0: static void reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type); aoqi@0: static void reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj); roland@7419: static void reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray* objects, bool realloc_failures); roland@7419: static void relock_objects(GrowableArray* monitors, JavaThread* thread, bool realloc_failures); roland@7419: static void pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array); roland@7419: NOT_PRODUCT(static void print_objects(GrowableArray* objects, bool realloc_failures);) aoqi@0: #endif // COMPILER2 aoqi@0: aoqi@0: public: roland@7419: static vframeArray* create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray* chunk, bool realloc_failures); aoqi@0: aoqi@0: // Interface used for unpacking deoptimized frames aoqi@0: aoqi@0: // UnrollBlock is returned by fetch_unroll_info() to the deoptimization handler (blob). aoqi@0: // This is only a CheapObj to ease debugging after a deopt failure aoqi@0: class UnrollBlock : public CHeapObj { aoqi@0: private: aoqi@0: int _size_of_deoptimized_frame; // Size, in bytes, of current deoptimized frame aoqi@0: int _caller_adjustment; // Adjustment, in bytes, to caller's SP by initial interpreted frame aoqi@0: int _number_of_frames; // Number frames to unroll aoqi@0: int _total_frame_sizes; // Total of number*sizes frames aoqi@0: intptr_t* _frame_sizes; // Array of frame sizes, in bytes, for unrolling the stack aoqi@0: address* _frame_pcs; // Array of frame pc's, in bytes, for unrolling the stack aoqi@0: intptr_t* _register_block; // Block for storing callee-saved registers. aoqi@0: BasicType _return_type; // Tells if we have to restore double or long return value aoqi@0: intptr_t _initial_info; // Platform dependent data for the sender frame (was FP on x86) aoqi@0: int _caller_actual_parameters; // The number of actual arguments at the aoqi@0: // interpreted caller of the deoptimized frame aoqi@0: aoqi@0: // The following fields are used as temps during the unpacking phase aoqi@0: // (which is tight on registers, especially on x86). They really ought aoqi@0: // to be PD variables but that involves moving this class into its own aoqi@0: // file to use the pd include mechanism. Maybe in a later cleanup ... aoqi@0: intptr_t _counter_temp; // SHOULD BE PD VARIABLE (x86 frame count temp) aoqi@0: intptr_t _unpack_kind; // SHOULD BE PD VARIABLE (x86 unpack kind) aoqi@0: intptr_t _sender_sp_temp; // SHOULD BE PD VARIABLE (x86 sender_sp) aoqi@0: public: aoqi@0: // Constructor aoqi@0: UnrollBlock(int size_of_deoptimized_frame, aoqi@0: int caller_adjustment, aoqi@0: int caller_actual_parameters, aoqi@0: int number_of_frames, aoqi@0: intptr_t* frame_sizes, aoqi@0: address* frames_pcs, aoqi@0: BasicType return_type); aoqi@0: ~UnrollBlock(); aoqi@0: aoqi@0: // Returns where a register is located. aoqi@0: intptr_t* value_addr_at(int register_number) const; aoqi@0: aoqi@0: // Accessors aoqi@0: intptr_t* frame_sizes() const { return _frame_sizes; } aoqi@0: int number_of_frames() const { return _number_of_frames; } aoqi@0: address* frame_pcs() const { return _frame_pcs ; } aoqi@0: aoqi@0: // Returns the total size of frames aoqi@0: int size_of_frames() const; aoqi@0: aoqi@0: void set_initial_info(intptr_t info) { _initial_info = info; } aoqi@0: aoqi@0: int caller_actual_parameters() const { return _caller_actual_parameters; } aoqi@0: aoqi@0: // Accessors used by the code generator for the unpack stub. aoqi@0: static int size_of_deoptimized_frame_offset_in_bytes() { return offset_of(UnrollBlock, _size_of_deoptimized_frame); } aoqi@0: static int caller_adjustment_offset_in_bytes() { return offset_of(UnrollBlock, _caller_adjustment); } aoqi@0: static int number_of_frames_offset_in_bytes() { return offset_of(UnrollBlock, _number_of_frames); } aoqi@0: static int frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _frame_sizes); } aoqi@0: static int total_frame_sizes_offset_in_bytes() { return offset_of(UnrollBlock, _total_frame_sizes); } aoqi@0: static int frame_pcs_offset_in_bytes() { return offset_of(UnrollBlock, _frame_pcs); } aoqi@0: static int register_block_offset_in_bytes() { return offset_of(UnrollBlock, _register_block); } aoqi@0: static int return_type_offset_in_bytes() { return offset_of(UnrollBlock, _return_type); } aoqi@0: static int counter_temp_offset_in_bytes() { return offset_of(UnrollBlock, _counter_temp); } aoqi@0: static int initial_info_offset_in_bytes() { return offset_of(UnrollBlock, _initial_info); } aoqi@0: static int unpack_kind_offset_in_bytes() { return offset_of(UnrollBlock, _unpack_kind); } aoqi@0: static int sender_sp_temp_offset_in_bytes() { return offset_of(UnrollBlock, _sender_sp_temp); } aoqi@0: aoqi@0: BasicType return_type() const { return _return_type; } aoqi@0: void print(); aoqi@0: }; aoqi@0: aoqi@0: //** Returns an UnrollBlock continuing information aoqi@0: // how to make room for the resulting interpreter frames. aoqi@0: // Called by assembly stub after execution has returned to aoqi@0: // deoptimized frame. aoqi@0: // @argument thread. Thread where stub_frame resides. aoqi@0: // @see OptoRuntime::deoptimization_fetch_unroll_info_C aoqi@0: static UnrollBlock* fetch_unroll_info(JavaThread* thread); aoqi@0: aoqi@0: //** Unpacks vframeArray onto execution stack aoqi@0: // Called by assembly stub after execution has returned to aoqi@0: // deoptimized frame and after the stack unrolling. aoqi@0: // @argument thread. Thread where stub_frame resides. aoqi@0: // @argument exec_mode. Determines how execution should be continuted in top frame. aoqi@0: // 0 means continue after current byte code aoqi@0: // 1 means exception has happened, handle exception aoqi@0: // 2 means reexecute current bytecode (for uncommon traps). aoqi@0: // @see OptoRuntime::deoptimization_unpack_frames_C aoqi@0: // Return BasicType of call return type, if any aoqi@0: static BasicType unpack_frames(JavaThread* thread, int exec_mode); aoqi@0: aoqi@0: // Cleans up deoptimization bits on thread after unpacking or in the aoqi@0: // case of an exception. aoqi@0: static void cleanup_deopt_info(JavaThread *thread, aoqi@0: vframeArray * array); aoqi@0: aoqi@0: // Restores callee saved values from deoptimized frame into oldest interpreter frame aoqi@0: // so caller of the deoptimized frame will get back the values it expects. aoqi@0: static void unwind_callee_save_values(frame* f, vframeArray* vframe_array); aoqi@0: aoqi@0: //** Performs an uncommon trap for compiled code. aoqi@0: // The top most compiler frame is converted into interpreter frames aoqi@0: static UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index); aoqi@0: // Helper routine that enters the VM and may block aoqi@0: static void uncommon_trap_inner(JavaThread* thread, jint unloaded_class_index); aoqi@0: aoqi@0: //** Deoptimizes the frame identified by id. aoqi@0: // Only called from VMDeoptimizeFrame aoqi@0: // @argument thread. Thread where stub_frame resides. aoqi@0: // @argument id. id of frame that should be deoptimized. aoqi@0: static void deoptimize_frame_internal(JavaThread* thread, intptr_t* id); aoqi@0: aoqi@0: // If thread is not the current thread then execute aoqi@0: // VM_DeoptimizeFrame otherwise deoptimize directly. aoqi@0: static void deoptimize_frame(JavaThread* thread, intptr_t* id); aoqi@0: aoqi@0: // Statistics aoqi@0: static void gather_statistics(DeoptReason reason, DeoptAction action, aoqi@0: Bytecodes::Code bc = Bytecodes::_illegal); aoqi@0: static void print_statistics(); aoqi@0: aoqi@0: // How much room to adjust the last frame's SP by, to make space for aoqi@0: // the callee's interpreter frame (which expects locals to be next to aoqi@0: // incoming arguments) aoqi@0: static int last_frame_adjust(int callee_parameters, int callee_locals); aoqi@0: aoqi@0: // trap_request codes aoqi@0: static DeoptReason trap_request_reason(int trap_request) { aoqi@0: if (trap_request < 0) aoqi@0: return (DeoptReason) aoqi@0: ((~(trap_request) >> _reason_shift) & right_n_bits(_reason_bits)); aoqi@0: else aoqi@0: // standard reason for unloaded CP entry aoqi@0: return Reason_unloaded; aoqi@0: } aoqi@0: static DeoptAction trap_request_action(int trap_request) { aoqi@0: if (trap_request < 0) aoqi@0: return (DeoptAction) aoqi@0: ((~(trap_request) >> _action_shift) & right_n_bits(_action_bits)); aoqi@0: else aoqi@0: // standard action for unloaded CP entry aoqi@0: return _unloaded_action; aoqi@0: } aoqi@0: static int trap_request_index(int trap_request) { aoqi@0: if (trap_request < 0) aoqi@0: return -1; aoqi@0: else aoqi@0: return trap_request; aoqi@0: } aoqi@0: static int make_trap_request(DeoptReason reason, DeoptAction action, aoqi@0: int index = -1) { aoqi@0: assert((1 << _reason_bits) >= Reason_LIMIT, "enough bits"); aoqi@0: assert((1 << _action_bits) >= Action_LIMIT, "enough bits"); aoqi@0: int trap_request; aoqi@0: if (index != -1) aoqi@0: trap_request = index; aoqi@0: else aoqi@0: trap_request = (~(((reason) << _reason_shift) aoqi@0: + ((action) << _action_shift))); aoqi@0: assert(reason == trap_request_reason(trap_request), "valid reason"); aoqi@0: assert(action == trap_request_action(trap_request), "valid action"); aoqi@0: assert(index == trap_request_index(trap_request), "valid index"); aoqi@0: return trap_request; aoqi@0: } aoqi@0: aoqi@0: // The trap_state stored in a MDO is decoded here. aoqi@0: // It records two items of information. aoqi@0: // reason: If a deoptimization happened here, what its reason was, aoqi@0: // or if there were multiple deopts with differing reasons. aoqi@0: // recompiled: If a deoptimization here triggered a recompilation. aoqi@0: // Note that not all reasons are recorded per-bci. aoqi@0: static DeoptReason trap_state_reason(int trap_state); aoqi@0: static int trap_state_has_reason(int trap_state, int reason); aoqi@0: static int trap_state_add_reason(int trap_state, int reason); aoqi@0: static bool trap_state_is_recompiled(int trap_state); aoqi@0: static int trap_state_set_recompiled(int trap_state, bool z); aoqi@0: static const char* format_trap_state(char* buf, size_t buflen, aoqi@0: int trap_state); aoqi@0: aoqi@0: static bool reason_is_recorded_per_bytecode(DeoptReason reason) { aoqi@0: return reason > Reason_none && reason <= Reason_RECORDED_LIMIT; aoqi@0: } aoqi@0: aoqi@0: static DeoptReason reason_recorded_per_bytecode_if_any(DeoptReason reason) { aoqi@0: if (reason_is_recorded_per_bytecode(reason)) aoqi@0: return reason; aoqi@0: else if (reason == Reason_div0_check) // null check due to divide-by-zero? aoqi@0: return Reason_null_check; // recorded per BCI as a null check aoqi@0: else if (reason == Reason_speculate_class_check) aoqi@0: return Reason_class_check; rbackman@7153: else if (reason == Reason_unstable_if) rbackman@7153: return Reason_intrinsic; aoqi@0: else aoqi@0: return Reason_none; aoqi@0: } aoqi@0: aoqi@0: static bool reason_is_speculate(int reason) { aoqi@0: if (reason == Reason_speculate_class_check) { aoqi@0: return true; aoqi@0: } aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: static uint per_method_trap_limit(int reason) { aoqi@0: return reason_is_speculate(reason) ? (uint)PerMethodSpecTrapLimit : (uint)PerMethodTrapLimit; aoqi@0: } aoqi@0: aoqi@0: static const char* trap_reason_name(int reason); aoqi@0: static const char* trap_action_name(int action); aoqi@0: // Format like reason='foo' action='bar' index='123'. aoqi@0: // This is suitable both for XML and for tty output. aoqi@0: static const char* format_trap_request(char* buf, size_t buflen, aoqi@0: int trap_request); aoqi@0: aoqi@0: static jint total_deoptimization_count(); aoqi@0: static jint deoptimization_count(DeoptReason reason); aoqi@0: aoqi@0: // JVMTI PopFrame support aoqi@0: aoqi@0: // Preserves incoming arguments to the popped frame when it is aoqi@0: // returning to a deoptimized caller aoqi@0: static void popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address); aoqi@0: aoqi@0: private: aoqi@0: static MethodData* get_method_data(JavaThread* thread, methodHandle m, bool create_if_missing); aoqi@0: // Update the mdo's count and per-BCI reason bits, returning previous state: aoqi@0: static ProfileData* query_update_method_data(MethodData* trap_mdo, aoqi@0: int trap_bci, aoqi@0: DeoptReason reason, aoqi@0: Method* compiled_method, aoqi@0: //outputs: aoqi@0: uint& ret_this_trap_count, aoqi@0: bool& ret_maybe_prior_trap, aoqi@0: bool& ret_maybe_prior_recompile); aoqi@0: // class loading support for uncommon trap aoqi@0: static void load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS); aoqi@0: static void load_class_by_index(constantPoolHandle constant_pool, int index); aoqi@0: aoqi@0: static UnrollBlock* fetch_unroll_info_helper(JavaThread* thread); aoqi@0: aoqi@0: static DeoptAction _unloaded_action; // == Action_reinterpret; aoqi@0: static const char* _trap_reason_name[Reason_LIMIT]; aoqi@0: static const char* _trap_action_name[Action_LIMIT]; aoqi@0: aoqi@0: static juint _deoptimization_hist[Reason_LIMIT][1+Action_LIMIT][BC_CASE_LIMIT]; aoqi@0: // Note: Histogram array size is 1-2 Kb. aoqi@0: aoqi@0: public: aoqi@0: static void update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason); aoqi@0: }; aoqi@0: aoqi@0: class DeoptimizationMarker : StackObj { // for profiling aoqi@0: static bool _is_active; aoqi@0: public: aoqi@0: DeoptimizationMarker() { _is_active = true; } aoqi@0: ~DeoptimizationMarker() { _is_active = false; } aoqi@0: static bool is_active() { return _is_active; } aoqi@0: }; aoqi@0: aoqi@0: #endif // SHARE_VM_RUNTIME_DEOPTIMIZATION_HPP