/*

 * Copyright 2005-2006 Sun Microsystems, Inc.  All Rights Reserved.

 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

 *

 * This code is free software; you can redistribute it and/or modify it

 * under the terms of the GNU General Public License version 2 only, as

 * published by the Free Software Foundation.

 *

 * This code is distributed in the hope that it will be useful, but WITHOUT

 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

 * version 2 for more details (a copy is included in the LICENSE file that

 * accompanied this code).

 *

 * You should have received a copy of the GNU General Public License version

 * 2 along with this work; if not, write to the Free Software Foundation,

 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

 *

 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,

 * CA 95054 USA or visit www.sun.com if you need additional information or

 * have any questions.

 *

 */

class DebugInfoCache;

class FpuStackAllocator;

class IRScopeDebugInfo;

class Interval;

class IntervalWalker;

class LIRGenerator;

class LinearScan;

class MoveResolver;

class Range;

define_array(IntervalArray, Interval*)

define_stack(IntervalList, IntervalArray)

define_array(IntervalsArray, IntervalList*)

define_stack(IntervalsList, IntervalsArray)

define_array(OopMapArray, OopMap*)

define_stack(OopMapList, OopMapArray)

define_array(ScopeValueArray, ScopeValue*)

define_array(LIR_OpListArray, LIR_OpList*);

define_stack(LIR_OpListStack, LIR_OpListArray);

enum IntervalUseKind {

  // priority of use kinds must be ascending

  noUse = 0,

  loopEndMarker = 1,

  shouldHaveRegister = 2,

  mustHaveRegister = 3,

  firstValidKind = 1,

  lastValidKind = 3

};

define_array(UseKindArray, IntervalUseKind)

define_stack(UseKindStack, UseKindArray)

enum IntervalKind {

  fixedKind = 0,  // interval pre-colored by LIR_Generator

  anyKind   = 1,  // no register/memory allocated by LIR_Generator

  nofKinds,

  firstKind = fixedKind

};

// during linear scan an interval is in one of four states in

enum IntervalState {

  unhandledState = 0, // unhandled state (not processed yet)

  activeState   = 1,  // life and is in a physical register

  inactiveState = 2,  // in a life time hole and is in a physical register

  handledState  = 3,  // spilled or not life again

  invalidState = -1

};

enum IntervalSpillState {

  noDefinitionFound,  // starting state of calculation: no definition found yet

  oneDefinitionFound, // one definition has already been found.

                      // Note: two consecutive definitions are treated as one (e.g. consecutive move and add because of two-operand LIR form)

                      // the position of this definition is stored in _definition_pos

  oneMoveInserted,    // one spill move has already been inserted.

  storeAtDefinition,  // the interval should be stored immediately after its definition because otherwise

                      // there would be multiple redundant stores

  startInMemory,      // the interval starts in memory (e.g. method parameter), so a store is never necessary

  noOptimization      // the interval has more then one definition (e.g. resulting from phi moves), so stores to memory are not optimized

};

#define for_each_interval_kind(kind) \

  for (IntervalKind kind = firstKind; kind < nofKinds; kind = (IntervalKind)(kind + 1))

#define for_each_visitor_mode(mode) \

  for (LIR_OpVisitState::OprMode mode = LIR_OpVisitState::firstMode; mode < LIR_OpVisitState::numModes; mode = (LIR_OpVisitState::OprMode)(mode + 1))

class LinearScan : public CompilationResourceObj {

  // declare classes used by LinearScan as friends because they

  // need a wide variety of functions declared here

  //

  // Only the small interface to the rest of the compiler is public

  friend class Interval;

  friend class IntervalWalker;

  friend class LinearScanWalker;

  friend class FpuStackAllocator;

  friend class MoveResolver;

  friend class LinearScanStatistic;

  friend class LinearScanTimers;

  friend class RegisterVerifier;

 public:

  enum {

    any_reg = -1,

    nof_cpu_regs = pd_nof_cpu_regs_linearscan,

    nof_fpu_regs = pd_nof_fpu_regs_linearscan,

    nof_xmm_regs = pd_nof_xmm_regs_linearscan,

    nof_regs = nof_cpu_regs + nof_fpu_regs + nof_xmm_regs

  };

 private:

  Compilation*              _compilation;

  IR*                       _ir;

  LIRGenerator*             _gen;

  FrameMap*                 _frame_map;

  BlockList                 _cached_blocks;     // cached list with all blocks in linear-scan order (only correct if original list keeps unchanged)

  int                       _num_virtual_regs;  // number of virtual registers (without new registers introduced because of splitting intervals)

  bool                      _has_fpu_registers; // true if this method uses any floating point registers (and so fpu stack allocation is necessary)

  int                       _num_calls;         // total number of calls in this method

  int                       _max_spills;        // number of stack slots used for intervals allocated to memory

  int                       _unused_spill_slot; // unused spill slot for a single-word value because of alignment of a double-word value

  IntervalList              _intervals;         // mapping from register number to interval

  IntervalList*             _new_intervals_from_allocation; // list with all intervals created during allocation when an existing interval is split

  IntervalArray*            _sorted_intervals;  // intervals sorted by Interval::from()

  LIR_OpArray               _lir_ops;           // mapping from LIR_Op id to LIR_Op node

  BlockBeginArray           _block_of_op;       // mapping from LIR_Op id to the BlockBegin containing this instruction

  BitMap                    _has_info;          // bit set for each LIR_Op id that has a CodeEmitInfo

  BitMap                    _has_call;          // bit set for each LIR_Op id that destroys all caller save registers

  BitMap2D                  _interval_in_loop;  // bit set for each virtual register that is contained in each loop

  // cached debug info to prevent multiple creation of same object

  // TODO: cached scope values for registers could be static

  ScopeValueArray           _scope_value_cache;

  static ConstantOopWriteValue _oop_null_scope_value;

  static ConstantIntValue    _int_m1_scope_value;

  static ConstantIntValue    _int_0_scope_value;

  static ConstantIntValue    _int_1_scope_value;

  static ConstantIntValue    _int_2_scope_value;

  // accessors

  IR*           ir() const                       { return _ir; }

  Compilation*  compilation() const              { return _compilation; }

  LIRGenerator* gen() const                      { return _gen; }

  FrameMap*     frame_map() const                { return _frame_map; }

  // unified bailout support

  void          bailout(const char* msg) const   { compilation()->bailout(msg); }

  bool          bailed_out() const               { return compilation()->bailed_out(); }

  // access to block list (sorted in linear scan order)

  int           block_count() const              { assert(_cached_blocks.length() == ir()->linear_scan_order()->length(), "invalid cached block list"); return _cached_blocks.length(); }

  BlockBegin*   block_at(int idx) const          { assert(_cached_blocks.at(idx) == ir()->linear_scan_order()->at(idx), "invalid cached block list");   return _cached_blocks.at(idx); }

  int           num_virtual_regs() const         { return _num_virtual_regs; }

  // size of live_in and live_out sets of BasicBlocks (BitMap needs rounded size for iteration)

  int           live_set_size() const            { return round_to(_num_virtual_regs, BitsPerWord); }

  bool          has_fpu_registers() const        { return _has_fpu_registers; }

  int           num_loops() const                { return ir()->num_loops(); }

  bool          is_interval_in_loop(int interval, int loop) const { return _interval_in_loop.at(interval, loop); }

  // handling of fpu stack allocation (platform dependent, needed for debug information generation)

#ifdef IA32

  FpuStackAllocator* _fpu_stack_allocator;

  bool use_fpu_stack_allocation() const          { return UseSSE < 2 && has_fpu_registers(); }

#else

  bool use_fpu_stack_allocation() const          { return false; }

#endif

  // access to interval list

  int           interval_count() const           { return _intervals.length(); }

  Interval*     interval_at(int reg_num) const   { return _intervals.at(reg_num); }

  IntervalList* new_intervals_from_allocation() const { return _new_intervals_from_allocation; }

  // access to LIR_Ops and Blocks indexed by op_id

  int          max_lir_op_id() const                { assert(_lir_ops.length() > 0, "no operations"); return (_lir_ops.length() - 1) << 1; }

  LIR_Op*      lir_op_with_id(int op_id) const      { assert(op_id >= 0 && op_id <= max_lir_op_id() && op_id % 2 == 0, "op_id out of range or not even"); return _lir_ops.at(op_id >> 1); }

  BlockBegin*  block_of_op_with_id(int op_id) const { assert(_block_of_op.length() > 0 && op_id >= 0 && op_id <= max_lir_op_id() + 1, "op_id out of range"); return _block_of_op.at(op_id >> 1); }

  bool is_block_begin(int op_id)                    { return op_id == 0 || block_of_op_with_id(op_id) != block_of_op_with_id(op_id - 1); }

  bool covers_block_begin(int op_id_1, int op_id_2) { return block_of_op_with_id(op_id_1) != block_of_op_with_id(op_id_2); }

  bool has_call(int op_id)                          { assert(op_id % 2 == 0, "must be even"); return _has_call.at(op_id >> 1); }

  bool has_info(int op_id)                          { assert(op_id % 2 == 0, "must be even"); return _has_info.at(op_id >> 1); }

  // functions for converting LIR-Operands to register numbers

  static bool is_valid_reg_num(int reg_num)         { return reg_num >= 0; }

  static int  reg_num(LIR_Opr opr);

  static int  reg_numHi(LIR_Opr opr);

  // functions for classification of intervals

  static bool is_precolored_interval(const Interval* i);

  static bool is_virtual_interval(const Interval* i);

  static bool is_precolored_cpu_interval(const Interval* i);

  static bool is_virtual_cpu_interval(const Interval* i);

  static bool is_precolored_fpu_interval(const Interval* i);

  static bool is_virtual_fpu_interval(const Interval* i);

  static bool is_in_fpu_register(const Interval* i);

  static bool is_oop_interval(const Interval* i);

  // General helper functions

  int         allocate_spill_slot(bool double_word);

  void        assign_spill_slot(Interval* it);

  void        propagate_spill_slots();

  Interval*   create_interval(int reg_num);

  void        append_interval(Interval* it);

  void        copy_register_flags(Interval* from, Interval* to);

  // platform dependent functions

  static bool is_processed_reg_num(int reg_num);

  static int  num_physical_regs(BasicType type);

  static bool requires_adjacent_regs(BasicType type);

  static bool is_caller_save(int assigned_reg);

  // spill move optimization: eliminate moves from register to stack if

  // stack slot is known to be correct

  void        change_spill_definition_pos(Interval* interval, int def_pos);

  void        change_spill_state(Interval* interval, int spill_pos);

  static bool must_store_at_definition(const Interval* i);

  void        eliminate_spill_moves();

  // Phase 1: number all instructions in all blocks

  void number_instructions();

  // Phase 2: compute local live sets separately for each block

  // (sets live_gen and live_kill for each block)

  //

  // helper methods used by compute_local_live_sets()

  void set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill);

  void compute_local_live_sets();

  // Phase 3: perform a backward dataflow analysis to compute global live sets

  // (sets live_in and live_out for each block)

  void compute_global_live_sets();

  // Phase 4: build intervals

  // (fills the list _intervals)

  //

  // helper methods used by build_intervals()

  void add_use (Value value, int from, int to, IntervalUseKind use_kind);

  void add_def (LIR_Opr opr, int def_pos,      IntervalUseKind use_kind);

  void add_use (LIR_Opr opr, int from, int to, IntervalUseKind use_kind);

  void add_temp(LIR_Opr opr, int temp_pos,     IntervalUseKind use_kind);

  void add_def (int reg_num, int def_pos,      IntervalUseKind use_kind, BasicType type);

  void add_use (int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type);

  void add_temp(int reg_num, int temp_pos,     IntervalUseKind use_kind, BasicType type);

  // Add platform dependent kills for particular LIR ops.  Can be used

  // to add platform dependent behaviour for some operations.

  void pd_add_temps(LIR_Op* op);

  IntervalUseKind use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr);

  IntervalUseKind use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr);

  void handle_method_arguments(LIR_Op* op);

  void handle_doubleword_moves(LIR_Op* op);

  void add_register_hints(LIR_Op* op);

  void build_intervals();

  // Phase 5: actual register allocation

  // (Uses LinearScanWalker)

  //

  // helper functions for building a sorted list of intervals

  NOT_PRODUCT(bool is_sorted(IntervalArray* intervals);)

  static int interval_cmp(Interval** a, Interval** b);

  void add_to_list(Interval** first, Interval** prev, Interval* interval);

  void create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i));

  void sort_intervals_before_allocation();

  void sort_intervals_after_allocation();

  void allocate_registers();

  // Phase 6: resolve data flow

  // (insert moves at edges between blocks if intervals have been split)

  //

  // helper functions for resolve_data_flow()

  Interval* split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode);

  Interval* interval_at_block_begin(BlockBegin* block, int reg_num);

  Interval* interval_at_block_end(BlockBegin* block, int reg_num);

  Interval* interval_at_op_id(int reg_num, int op_id);

  void resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver);

  void resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver);

  void resolve_data_flow();

  void resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver);

  void resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver);

  void resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver);

  void resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver);

  void resolve_exception_handlers();

  // Phase 7: assign register numbers back to LIR

  // (includes computation of debug information and oop maps)

  //

  // helper functions for assign_reg_num()

  VMReg vm_reg_for_interval(Interval* interval);

  VMReg vm_reg_for_operand(LIR_Opr opr);

  static LIR_Opr operand_for_interval(Interval* interval);

  static LIR_Opr calc_operand_for_interval(const Interval* interval);

  LIR_Opr       canonical_spill_opr(Interval* interval);

  LIR_Opr color_lir_opr(LIR_Opr opr, int id, LIR_OpVisitState::OprMode);

  // methods used for oop map computation

  IntervalWalker* init_compute_oop_maps();

  OopMap*         compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site);

  void            compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op);

  // methods used for debug information computation

  void init_compute_debug_info();

  MonitorValue*  location_for_monitor_index(int monitor_index);

  LocationValue* location_for_name(int name, Location::Type loc_type);

  int append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values);

  int append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values);

  int append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values);

  IRScopeDebugInfo* compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state, int cur_bci, int stack_end, int locks_end);

  void compute_debug_info(CodeEmitInfo* info, int op_id);

  void assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw);

  void assign_reg_num();

  // Phase 8: fpu stack allocation

  // (Used only on x86 when fpu operands are present)

  void allocate_fpu_stack();

  // helper functions for printing state

#ifndef PRODUCT

  static void print_bitmap(BitMap& bitmap);

  void        print_intervals(const char* label);

  void        print_lir(int level, const char* label, bool hir_valid = true);

#endif

#ifdef ASSERT

  // verification functions for allocation

  // (check that all intervals have a correct register and that no registers are overwritten)

  void verify();

  void verify_intervals();

  void verify_no_oops_in_fixed_intervals();

  void verify_constants();

  void verify_registers();

#endif

 public:

  // creation

  LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map);

  // main entry function: perform linear scan register allocation

  void             do_linear_scan();

  // accessors used by Compilation

  int         max_spills()  const { return _max_spills; }

  int         num_calls() const   { assert(_num_calls >= 0, "not set"); return _num_calls; }

  // entry functions for printing

#ifndef PRODUCT

  static void print_statistics();

  static void print_timers(double total);

#endif

};

// Helper class for ordering moves that are inserted at the same position in the LIR

// When moves between registers are inserted, it is important that the moves are

// ordered such that no register is overwritten. So moves from register to stack

// are processed prior to moves from stack to register. When moves have circular

// dependencies, a temporary stack slot is used to break the circle.

// The same logic is used in the LinearScanWalker and in LinearScan during resolve_data_flow

// and therefore factored out in a separate class

class MoveResolver: public StackObj {

 private:

  LinearScan*      _allocator;

  LIR_List*        _insert_list;

  int              _insert_idx;

  LIR_InsertionBuffer _insertion_buffer; // buffer where moves are inserted

  IntervalList     _mapping_from;

  LIR_OprList      _mapping_from_opr;

  IntervalList     _mapping_to;

  bool             _multiple_reads_allowed;

  int              _register_blocked[LinearScan::nof_regs];