1 /*
   2  * Copyright (c) 1997, 2013, 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_VM_OPTO_BLOCK_HPP
  26 #define SHARE_VM_OPTO_BLOCK_HPP
  27 
  28 #include "opto/multnode.hpp"
  29 #include "opto/node.hpp"
  30 #include "opto/phase.hpp"
  31 
  32 // Optimization - Graph Style
  33 
  34 class Block;
  35 class CFGLoop;
  36 class MachCallNode;
  37 class Matcher;
  38 class RootNode;
  39 class VectorSet;

  40 struct Tarjan;
  41 
  42 //------------------------------Block_Array------------------------------------
  43 // Map dense integer indices to Blocks.  Uses classic doubling-array trick.
  44 // Abstractly provides an infinite array of Block*'s, initialized to NULL.
  45 // Note that the constructor just zeros things, and since I use Arena
  46 // allocation I do not need a destructor to reclaim storage.
  47 class Block_Array : public ResourceObj {
  48   friend class VMStructs;
  49   uint _size;                   // allocated size, as opposed to formal limit
  50   debug_only(uint _limit;)      // limit to formal domain
  51   Arena *_arena;                // Arena to allocate in
  52 protected:
  53   Block **_blocks;
  54   void grow( uint i );          // Grow array node to fit
  55 
  56 public:
  57   Block_Array(Arena *a) : _arena(a), _size(OptoBlockListSize) {
  58     debug_only(_limit=0);
  59     _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize );
  60     for( int i = 0; i < OptoBlockListSize; i++ ) {
  61       _blocks[i] = NULL;
  62     }
  63   }
  64   Block *lookup( uint i ) const // Lookup, or NULL for not mapped
  65   { return (i<Max()) ? _blocks[i] : (Block*)NULL; }
  66   Block *operator[] ( uint i ) const // Lookup, or assert for not mapped
  67   { assert( i < Max(), "oob" ); return _blocks[i]; }
  68   // Extend the mapping: index i maps to Block *n.
  69   void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; }
  70   uint Max() const { debug_only(return _limit); return _size; }
  71 };
  72 
  73 
  74 class Block_List : public Block_Array {
  75   friend class VMStructs;
  76 public:
  77   uint _cnt;
  78   Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {}
  79   void push( Block *b ) {  map(_cnt++,b); }
  80   Block *pop() { return _blocks[--_cnt]; }
  81   Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;}
  82   void remove( uint i );
  83   void insert( uint i, Block *n );
  84   uint size() const { return _cnt; }
  85   void reset() { _cnt = 0; }
  86   void print();
  87 };
  88 
  89 
  90 class CFGElement : public ResourceObj {
  91   friend class VMStructs;
  92  public:
  93   double _freq; // Execution frequency (estimate)
  94 
  95   CFGElement() : _freq(0.0) {}
  96   virtual bool is_block() { return false; }
  97   virtual bool is_loop()  { return false; }
  98   Block*   as_Block() { assert(is_block(), "must be block"); return (Block*)this; }
  99   CFGLoop* as_CFGLoop()  { assert(is_loop(),  "must be loop");  return (CFGLoop*)this;  }
 100 };
 101 
 102 //------------------------------Block------------------------------------------
 103 // This class defines a Basic Block.
 104 // Basic blocks are used during the output routines, and are not used during
 105 // any optimization pass.  They are created late in the game.
 106 class Block : public CFGElement {
 107   friend class VMStructs;
 108 
 109 private:
 110   // Nodes in this block, in order
 111   Node_List _nodes;
 112 
 113 public:
 114 
 115   // Get the node at index 'at_index', if 'at_index' is out of bounds return NULL
 116   Node* get_node(uint at_index) const {
 117     return _nodes[at_index];
 118   }
 119 
 120   // Get the number of nodes in this block
 121   uint number_of_nodes() const {
 122     return _nodes.size();
 123   }
 124 
 125   // Map a node 'node' to index 'to_index' in the block, if the index is out of bounds the size of the node list is increased
 126   void map_node(Node* node, uint to_index) {
 127     _nodes.map(to_index, node);
 128   }
 129 
 130   // Insert a node 'node' at index 'at_index', moving all nodes that are on a higher index one step, if 'at_index' is out of bounds we crash
 131   void insert_node(Node* node, uint at_index) {
 132     _nodes.insert(at_index, node);
 133   }
 134 
 135   // Remove a node at index 'at_index'
 136   void remove_node(uint at_index) {
 137     _nodes.remove(at_index);
 138   }
 139 
 140   // Push a node 'node' onto the node list
 141   void push_node(Node* node) {
 142     _nodes.push(node);
 143   }
 144 
 145   // Pop the last node off the node list
 146   Node* pop_node() {
 147     return _nodes.pop();
 148   }
 149 
 150   // Basic blocks have a Node which defines Control for all Nodes pinned in
 151   // this block.  This Node is a RegionNode.  Exception-causing Nodes
 152   // (division, subroutines) and Phi functions are always pinned.  Later,
 153   // every Node will get pinned to some block.
 154   Node *head() const { return get_node(0); }
 155 
 156   // CAUTION: num_preds() is ONE based, so that predecessor numbers match
 157   // input edges to Regions and Phis.
 158   uint num_preds() const { return head()->req(); }
 159   Node *pred(uint i) const { return head()->in(i); }
 160 
 161   // Array of successor blocks, same size as projs array
 162   Block_Array _succs;
 163 
 164   // Basic blocks have some number of Nodes which split control to all
 165   // following blocks.  These Nodes are always Projections.  The field in
 166   // the Projection and the block-ending Node determine which Block follows.
 167   uint _num_succs;
 168 
 169   // Basic blocks also carry all sorts of good old fashioned DFS information
 170   // used to find loops, loop nesting depth, dominators, etc.
 171   uint _pre_order;              // Pre-order DFS number
 172 
 173   // Dominator tree
 174   uint _dom_depth;              // Depth in dominator tree for fast LCA
 175   Block* _idom;                 // Immediate dominator block
 176 
 177   CFGLoop *_loop;               // Loop to which this block belongs
 178   uint _rpo;                    // Number in reverse post order walk
 179 
 180   virtual bool is_block() { return true; }
 181   float succ_prob(uint i);      // return probability of i'th successor
 182   int num_fall_throughs();      // How many fall-through candidate this block has
 183   void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code
 184   bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate
 185   Block* lone_fall_through();   // Return lone fall-through Block or null
 186 
 187   Block* dom_lca(Block* that);  // Compute LCA in dominator tree.
 188 #ifdef ASSERT
 189   bool dominates(Block* that) {
 190     int dom_diff = this->_dom_depth - that->_dom_depth;
 191     if (dom_diff > 0)  return false;
 192     for (; dom_diff < 0; dom_diff++)  that = that->_idom;
 193     return this == that;
 194   }
 195 #endif
 196 
 197   // Report the alignment required by this block.  Must be a power of 2.
 198   // The previous block will insert nops to get this alignment.
 199   uint code_alignment();
 200   uint compute_loop_alignment();
 201 
 202   // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.
 203   // It is currently also used to scale such frequencies relative to
 204   // FreqCountInvocations relative to the old value of 1500.
 205 #define BLOCK_FREQUENCY(f) ((f * (double) 1500) / FreqCountInvocations)
 206 
 207   // Register Pressure (estimate) for Splitting heuristic
 208   uint _reg_pressure;
 209   uint _ihrp_index;
 210   uint _freg_pressure;
 211   uint _fhrp_index;
 212 
 213   // Mark and visited bits for an LCA calculation in insert_anti_dependences.
 214   // Since they hold unique node indexes, they do not need reinitialization.
 215   node_idx_t _raise_LCA_mark;
 216   void    set_raise_LCA_mark(node_idx_t x)    { _raise_LCA_mark = x; }
 217   node_idx_t  raise_LCA_mark() const          { return _raise_LCA_mark; }
 218   node_idx_t _raise_LCA_visited;
 219   void    set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; }
 220   node_idx_t  raise_LCA_visited() const       { return _raise_LCA_visited; }
 221 
 222   // Estimated size in bytes of first instructions in a loop.
 223   uint _first_inst_size;
 224   uint first_inst_size() const     { return _first_inst_size; }
 225   void set_first_inst_size(uint s) { _first_inst_size = s; }
 226 
 227   // Compute the size of first instructions in this block.
 228   uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra);
 229 
 230   // Compute alignment padding if the block needs it.
 231   // Align a loop if loop's padding is less or equal to padding limit
 232   // or the size of first instructions in the loop > padding.
 233   uint alignment_padding(int current_offset) {
 234     int block_alignment = code_alignment();
 235     int max_pad = block_alignment-relocInfo::addr_unit();
 236     if( max_pad > 0 ) {
 237       assert(is_power_of_2(max_pad+relocInfo::addr_unit()), "");
 238       int current_alignment = current_offset & max_pad;
 239       if( current_alignment != 0 ) {
 240         uint padding = (block_alignment-current_alignment) & max_pad;
 241         if( has_loop_alignment() &&
 242             padding > (uint)MaxLoopPad &&
 243             first_inst_size() <= padding ) {
 244           return 0;
 245         }
 246         return padding;
 247       }
 248     }
 249     return 0;
 250   }
 251 
 252   // Connector blocks. Connector blocks are basic blocks devoid of
 253   // instructions, but may have relevant non-instruction Nodes, such as
 254   // Phis or MergeMems. Such blocks are discovered and marked during the
 255   // RemoveEmpty phase, and elided during Output.
 256   bool _connector;
 257   void set_connector() { _connector = true; }
 258   bool is_connector() const { return _connector; };
 259 
 260   // Loop_alignment will be set for blocks which are at the top of loops.
 261   // The block layout pass may rotate loops such that the loop head may not
 262   // be the sequentially first block of the loop encountered in the linear
 263   // list of blocks.  If the layout pass is not run, loop alignment is set
 264   // for each block which is the head of a loop.
 265   uint _loop_alignment;
 266   void set_loop_alignment(Block *loop_top) {
 267     uint new_alignment = loop_top->compute_loop_alignment();
 268     if (new_alignment > _loop_alignment) {
 269       _loop_alignment = new_alignment;
 270     }
 271   }
 272   uint loop_alignment() const { return _loop_alignment; }
 273   bool has_loop_alignment() const { return loop_alignment() > 0; }
 274 
 275   // Create a new Block with given head Node.
 276   // Creates the (empty) predecessor arrays.
 277   Block( Arena *a, Node *headnode )
 278     : CFGElement(),
 279       _nodes(a),
 280       _succs(a),
 281       _num_succs(0),
 282       _pre_order(0),
 283       _idom(0),
 284       _loop(NULL),
 285       _reg_pressure(0),
 286       _ihrp_index(1),
 287       _freg_pressure(0),
 288       _fhrp_index(1),
 289       _raise_LCA_mark(0),
 290       _raise_LCA_visited(0),
 291       _first_inst_size(999999),
 292       _connector(false),
 293       _loop_alignment(0) {
 294     _nodes.push(headnode);
 295   }
 296 
 297   // Index of 'end' Node
 298   uint end_idx() const {
 299     // %%%%% add a proj after every goto
 300     // so (last->is_block_proj() != last) always, then simplify this code
 301     // This will not give correct end_idx for block 0 when it only contains root.
 302     int last_idx = _nodes.size() - 1;
 303     Node *last  = _nodes[last_idx];
 304     assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], "");
 305     return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs);
 306   }
 307 
 308   // Basic blocks have a Node which ends them.  This Node determines which
 309   // basic block follows this one in the program flow.  This Node is either an
 310   // IfNode, a GotoNode, a JmpNode, or a ReturnNode.
 311   Node *end() const { return _nodes[end_idx()]; }
 312 
 313   // Add an instruction to an existing block.  It must go after the head
 314   // instruction and before the end instruction.
 315   void add_inst( Node *n ) { insert_node(n, end_idx()); }
 316   // Find node in block. Fails if node not in block.
 317   uint find_node( const Node *n ) const;
 318   // Find and remove n from block list
 319   void find_remove( const Node *n );
 320   // Check wether the node is in the block.
 321   bool contains (const Node *n) const;
 322 
 323   // Return the empty status of a block
 324   enum { not_empty, empty_with_goto, completely_empty };
 325   int is_Empty() const;
 326 
 327   // Forward through connectors
 328   Block* non_connector() {
 329     Block* s = this;
 330     while (s->is_connector()) {
 331       s = s->_succs[0];
 332     }
 333     return s;
 334   }
 335 
 336   // Return true if b is a successor of this block
 337   bool has_successor(Block* b) const {
 338     for (uint i = 0; i < _num_succs; i++ ) {
 339       if (non_connector_successor(i) == b) {
 340         return true;
 341       }
 342     }
 343     return false;
 344   }
 345 
 346   // Successor block, after forwarding through connectors
 347   Block* non_connector_successor(int i) const {
 348     return _succs[i]->non_connector();
 349   }
 350 
 351   // Examine block's code shape to predict if it is not commonly executed.
 352   bool has_uncommon_code() const;
 353 
 354 #ifndef PRODUCT
 355   // Debugging print of basic block
 356   void dump_bidx(const Block* orig, outputStream* st = tty) const;
 357   void dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st = tty) const;
 358   void dump_head(const PhaseCFG* cfg, outputStream* st = tty) const;
 359   void dump() const;
 360   void dump(const PhaseCFG* cfg) const;
 361 #endif
 362 };
 363 
 364 
 365 //------------------------------PhaseCFG---------------------------------------
 366 // Build an array of Basic Block pointers, one per Node.
 367 class PhaseCFG : public Phase {
 368   friend class VMStructs;
 369  private:
 370 
 371   // Root of whole program
 372   RootNode* _root;
 373 
 374   // The block containing the root node
 375   Block* _root_block;
 376 
 377   // List of basic blocks that are created during CFG creation
 378   Block_List _blocks;
 379 
 380   // Count of basic blocks
 381   uint _number_of_blocks;
 382 
 383   // Arena for the blocks to be stored in
 384   Arena* _block_arena;
 385 






 386   // The matcher for this compilation
 387   Matcher& _matcher;
 388 
 389   // Map nodes to owning basic block
 390   Block_Array _node_to_block_mapping;
 391 
 392   // Loop from the root
 393   CFGLoop* _root_loop;
 394 
 395   // Outmost loop frequency
 396   double _outer_loop_frequency;
 397 
 398   // Per node latency estimation, valid only during GCM
 399   GrowableArray<uint>* _node_latency;
 400 
 401   // Build a proper looking cfg.  Return count of basic blocks
 402   uint build_cfg();
 403 
 404   // Build the dominator tree so that we know where we can move instructions
 405   void build_dominator_tree();
 406 
 407   // Estimate block frequencies based on IfNode probabilities, so that we know where we want to move instructions
 408   void estimate_block_frequency();
 409 
 410   // Global Code Motion.  See Click's PLDI95 paper.  Place Nodes in specific
 411   // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block.
 412   // Move nodes to ensure correctness from GVN and also try to move nodes out of loops.
 413   void global_code_motion();
 414 
 415   // Schedule Nodes early in their basic blocks.
 416   bool schedule_early(VectorSet &visited, Node_List &roots);
 417 
 418   // For each node, find the latest block it can be scheduled into
 419   // and then select the cheapest block between the latest and earliest
 420   // block to place the node.
 421   void schedule_late(VectorSet &visited, Node_List &stack);
 422 
 423   // Compute the (backwards) latency of a node from a single use
 424   int latency_from_use(Node *n, const Node *def, Node *use);
 425 
 426   // Compute the (backwards) latency of a node from the uses of this instruction
 427   void partial_latency_of_defs(Node *n);
 428 
 429   // Compute the instruction global latency with a backwards walk
 430   void compute_latencies_backwards(VectorSet &visited, Node_List &stack);
 431 
 432   // Pick a block between early and late that is a cheaper alternative
 433   // to late. Helper for schedule_late.
 434   Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);
 435 
 436   bool schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call);
 437   void set_next_call(Block* block, Node* n, VectorSet& next_call);
 438   void needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call);
 439 
 440   // Perform basic-block local scheduling
 441   Node* select(Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, VectorSet& next_call, uint sched_slot);


 442 
 443   // Schedule a call next in the block
 444   uint sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call);
 445 
 446   // Cleanup if any code lands between a Call and his Catch
 447   void call_catch_cleanup(Block* block);
 448 
 449   Node* catch_cleanup_find_cloned_def(Block* use_blk, Node* def, Block* def_blk, int n_clone_idx);
 450   void  catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx);
 451 
 452   // Detect implicit-null-check opportunities.  Basically, find NULL checks
 453   // with suitable memory ops nearby.  Use the memory op to do the NULL check.
 454   // I can generate a memory op if there is not one nearby.
 455   void implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons);
 456 
 457   // Perform a Depth First Search (DFS).
 458   // Setup 'vertex' as DFS to vertex mapping.
 459   // Setup 'semi' as vertex to DFS mapping.
 460   // Set 'parent' to DFS parent.
 461   uint do_DFS(Tarjan* tarjan, uint rpo_counter);
 462 
 463   // Helper function to insert a node into a block
 464   void schedule_node_into_block( Node *n, Block *b );
 465 
 466   void replace_block_proj_ctrl( Node *n );
 467 
 468   // Set the basic block for pinned Nodes
 469   void schedule_pinned_nodes( VectorSet &visited );
 470 
 471   // I'll need a few machine-specific GotoNodes.  Clone from this one.
 472   // Used when building the CFG and creating end nodes for blocks.
 473   MachNode* _goto;
 474 
 475   Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
 476   void verify_anti_dependences(Block* LCA, Node* load) {
 477     assert(LCA == get_block_for_node(load), "should already be scheduled");
 478     insert_anti_dependences(LCA, load, true);
 479   }
 480 
 481   bool move_to_next(Block* bx, uint b_index);
 482   void move_to_end(Block* bx, uint b_index);
 483 
 484   void insert_goto_at(uint block_no, uint succ_no);
 485 
 486   // Check for NeverBranch at block end.  This needs to become a GOTO to the
 487   // true target.  NeverBranch are treated as a conditional branch that always
 488   // goes the same direction for most of the optimizer and are used to give a
 489   // fake exit path to infinite loops.  At this late stage they need to turn
 490   // into Goto's so that when you enter the infinite loop you indeed hang.
 491   void convert_NeverBranch_to_Goto(Block *b);
 492 
 493   CFGLoop* create_loop_tree();
 494 
 495   #ifndef PRODUCT
 496   bool _trace_opto_pipelining;  // tracing flag
 497   #endif
 498 
 499  public:
 500   PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher);
 501 
 502   void set_latency_for_node(Node* node, int latency) {
 503     _node_latency->at_put_grow(node->_idx, latency);
 504   }
 505 
 506   uint get_latency_for_node(Node* node) {
 507     return _node_latency->at_grow(node->_idx);
 508   }
 509 
 510   // Get the outer most frequency
 511   double get_outer_loop_frequency() const {
 512     return _outer_loop_frequency;
 513   }
 514 
 515   // Get the root node of the CFG
 516   RootNode* get_root_node() const {
 517     return _root;
 518   }
 519 
 520   // Get the block of the root node
 521   Block* get_root_block() const {
 522     return _root_block;
 523   }
 524 
 525   // Add a block at a position and moves the later ones one step
 526   void add_block_at(uint pos, Block* block) {
 527     _blocks.insert(pos, block);
 528     _number_of_blocks++;
 529   }
 530 
 531   // Adds a block to the top of the block list
 532   void add_block(Block* block) {
 533     _blocks.push(block);
 534     _number_of_blocks++;
 535   }
 536 
 537   // Clear the list of blocks
 538   void clear_blocks() {
 539     _blocks.reset();
 540     _number_of_blocks = 0;
 541   }
 542 
 543   // Get the block at position pos in _blocks
 544   Block* get_block(uint pos) const {
 545     return _blocks[pos];
 546   }
 547 
 548   // Number of blocks
 549   uint number_of_blocks() const {
 550     return _number_of_blocks;
 551   }
 552 
 553   // set which block this node should reside in
 554   void map_node_to_block(const Node* node, Block* block) {
 555     _node_to_block_mapping.map(node->_idx, block);
 556   }
 557 
 558   // removes the mapping from a node to a block
 559   void unmap_node_from_block(const Node* node) {
 560     _node_to_block_mapping.map(node->_idx, NULL);
 561   }
 562 
 563   // get the block in which this node resides
 564   Block* get_block_for_node(const Node* node) const {
 565     return _node_to_block_mapping[node->_idx];
 566   }
 567 
 568   // does this node reside in a block; return true
 569   bool has_block(const Node* node) const {
 570     return (_node_to_block_mapping.lookup(node->_idx) != NULL);
 571   }
 572 
 573   // Use frequency calculations and code shape to predict if the block
 574   // is uncommon.
 575   bool is_uncommon(const Block* block);
 576 
 577 #ifdef ASSERT
 578   Unique_Node_List _raw_oops;
 579 #endif
 580 
 581   // Do global code motion by first building dominator tree and estimate block frequency
 582   // Returns true on success
 583   bool do_global_code_motion();
 584 
 585   // Compute the (backwards) latency of a node from the uses
 586   void latency_from_uses(Node *n);
 587 
 588   // Set loop alignment
 589   void set_loop_alignment();
 590 
 591   // Remove empty basic blocks
 592   void remove_empty_blocks();
 593   Block *fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext);
 594   void fixup_flow();
 595 
 596   // Insert a node into a block at index and map the node to the block
 597   void insert(Block *b, uint idx, Node *n) {
 598     b->insert_node(n , idx);
 599     map_node_to_block(n, b);
 600   }
 601 
 602   // Check all nodes and postalloc_expand them if necessary.
 603   void postalloc_expand(PhaseRegAlloc* _ra);
 604 
 605 #ifndef PRODUCT
 606   bool trace_opto_pipelining() const { return _trace_opto_pipelining; }
 607 
 608   // Debugging print of CFG
 609   void dump( ) const;           // CFG only
 610   void _dump_cfg( const Node *end, VectorSet &visited  ) const;
 611   void verify() const;
 612   void dump_headers();
 613 #else
 614   bool trace_opto_pipelining() const { return false; }
 615 #endif
 616 };
 617 
 618 
 619 //------------------------------UnionFind--------------------------------------
 620 // Map Block indices to a block-index for a cfg-cover.
 621 // Array lookup in the optimized case.
 622 class UnionFind : public ResourceObj {
 623   uint _cnt, _max;
 624   uint* _indices;
 625   ReallocMark _nesting;  // assertion check for reallocations
 626 public:
 627   UnionFind( uint max );
 628   void reset( uint max );  // Reset to identity map for [0..max]
 629 
 630   uint lookup( uint nidx ) const {
 631     return _indices[nidx];
 632   }
 633   uint operator[] (uint nidx) const { return lookup(nidx); }
 634 
 635   void map( uint from_idx, uint to_idx ) {
 636     assert( from_idx < _cnt, "oob" );
 637     _indices[from_idx] = to_idx;
 638   }
 639   void extend( uint from_idx, uint to_idx );
 640 
 641   uint Size() const { return _cnt; }
 642 
 643   uint Find( uint idx ) {
 644     assert( idx < 65536, "Must fit into uint");
 645     uint uf_idx = lookup(idx);
 646     return (uf_idx == idx) ? uf_idx : Find_compress(idx);
 647   }
 648   uint Find_compress( uint idx );
 649   uint Find_const( uint idx ) const;
 650   void Union( uint idx1, uint idx2 );
 651 
 652 };
 653 
 654 //----------------------------BlockProbPair---------------------------
 655 // Ordered pair of Node*.
 656 class BlockProbPair VALUE_OBJ_CLASS_SPEC {
 657 protected:
 658   Block* _target;      // block target
 659   double  _prob;        // probability of edge to block
 660 public:
 661   BlockProbPair() : _target(NULL), _prob(0.0) {}
 662   BlockProbPair(Block* b, double p) : _target(b), _prob(p) {}
 663 
 664   Block* get_target() const { return _target; }
 665   double get_prob() const { return _prob; }
 666 };
 667 
 668 //------------------------------CFGLoop-------------------------------------------
 669 class CFGLoop : public CFGElement {
 670   friend class VMStructs;
 671   int _id;
 672   int _depth;
 673   CFGLoop *_parent;      // root of loop tree is the method level "pseudo" loop, it's parent is null
 674   CFGLoop *_sibling;     // null terminated list
 675   CFGLoop *_child;       // first child, use child's sibling to visit all immediately nested loops
 676   GrowableArray<CFGElement*> _members; // list of members of loop
 677   GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities
 678   double _exit_prob;       // probability any loop exit is taken on a single loop iteration
 679   void update_succ_freq(Block* b, double freq);
 680 
 681  public:
 682   CFGLoop(int id) :
 683     CFGElement(),
 684     _id(id),
 685     _depth(0),
 686     _parent(NULL),
 687     _sibling(NULL),
 688     _child(NULL),
 689     _exit_prob(1.0f) {}
 690   CFGLoop* parent() { return _parent; }
 691   void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg);
 692   void add_member(CFGElement *s) { _members.push(s); }
 693   void add_nested_loop(CFGLoop* cl);
 694   Block* head() {
 695     assert(_members.at(0)->is_block(), "head must be a block");
 696     Block* hd = _members.at(0)->as_Block();
 697     assert(hd->_loop == this, "just checking");
 698     assert(hd->head()->is_Loop(), "must begin with loop head node");
 699     return hd;
 700   }
 701   Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)
 702   void compute_loop_depth(int depth);
 703   void compute_freq(); // compute frequency with loop assuming head freq 1.0f
 704   void scale_freq();   // scale frequency by loop trip count (including outer loops)
 705   double outer_loop_freq() const; // frequency of outer loop
 706   bool in_loop_nest(Block* b);
 707   double trip_count() const { return 1.0 / _exit_prob; }
 708   virtual bool is_loop()  { return true; }
 709   int id() { return _id; }
 710 
 711 #ifndef PRODUCT
 712   void dump( ) const;
 713   void dump_tree() const;
 714 #endif
 715 };
 716 
 717 
 718 //----------------------------------CFGEdge------------------------------------
 719 // A edge between two basic blocks that will be embodied by a branch or a
 720 // fall-through.
 721 class CFGEdge : public ResourceObj {
 722   friend class VMStructs;
 723  private:
 724   Block * _from;        // Source basic block
 725   Block * _to;          // Destination basic block
 726   double _freq;          // Execution frequency (estimate)
 727   int   _state;
 728   bool  _infrequent;
 729   int   _from_pct;
 730   int   _to_pct;
 731 
 732   // Private accessors
 733   int  from_pct() const { return _from_pct; }
 734   int  to_pct()   const { return _to_pct;   }
 735   int  from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
 736   int  to_infrequent()   const { return to_pct()   < BlockLayoutMinDiamondPercentage; }
 737 
 738  public:
 739   enum {
 740     open,               // initial edge state; unprocessed
 741     connected,          // edge used to connect two traces together
 742     interior            // edge is interior to trace (could be backedge)
 743   };
 744 
 745   CFGEdge(Block *from, Block *to, double freq, int from_pct, int to_pct) :
 746     _from(from), _to(to), _freq(freq),
 747     _from_pct(from_pct), _to_pct(to_pct), _state(open) {
 748     _infrequent = from_infrequent() || to_infrequent();
 749   }
 750 
 751   double  freq() const { return _freq; }
 752   Block* from() const { return _from; }
 753   Block* to  () const { return _to;   }
 754   int  infrequent() const { return _infrequent; }
 755   int state() const { return _state; }
 756 
 757   void set_state(int state) { _state = state; }
 758 
 759 #ifndef PRODUCT
 760   void dump( ) const;
 761 #endif
 762 };
 763 
 764 
 765 //-----------------------------------Trace-------------------------------------
 766 // An ordered list of basic blocks.
 767 class Trace : public ResourceObj {
 768  private:
 769   uint _id;             // Unique Trace id (derived from initial block)
 770   Block ** _next_list;  // Array mapping index to next block
 771   Block ** _prev_list;  // Array mapping index to previous block
 772   Block * _first;       // First block in the trace
 773   Block * _last;        // Last block in the trace
 774 
 775   // Return the block that follows "b" in the trace.
 776   Block * next(Block *b) const { return _next_list[b->_pre_order]; }
 777   void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
 778 
 779   // Return the block that precedes "b" in the trace.
 780   Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
 781   void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
 782 
 783   // We've discovered a loop in this trace. Reset last to be "b", and first as
 784   // the block following "b
 785   void break_loop_after(Block *b) {
 786     _last = b;
 787     _first = next(b);
 788     set_prev(_first, NULL);
 789     set_next(_last, NULL);
 790   }
 791 
 792  public:
 793 
 794   Trace(Block *b, Block **next_list, Block **prev_list) :
 795     _first(b),
 796     _last(b),
 797     _next_list(next_list),
 798     _prev_list(prev_list),
 799     _id(b->_pre_order) {
 800     set_next(b, NULL);
 801     set_prev(b, NULL);
 802   };
 803 
 804   // Return the id number
 805   uint id() const { return _id; }
 806   void set_id(uint id) { _id = id; }
 807 
 808   // Return the first block in the trace
 809   Block * first_block() const { return _first; }
 810 
 811   // Return the last block in the trace
 812   Block * last_block() const { return _last; }
 813 
 814   // Insert a trace in the middle of this one after b
 815   void insert_after(Block *b, Trace *tr) {
 816     set_next(tr->last_block(), next(b));
 817     if (next(b) != NULL) {
 818       set_prev(next(b), tr->last_block());
 819     }
 820 
 821     set_next(b, tr->first_block());
 822     set_prev(tr->first_block(), b);
 823 
 824     if (b == _last) {
 825       _last = tr->last_block();
 826     }
 827   }
 828 
 829   void insert_before(Block *b, Trace *tr) {
 830     Block *p = prev(b);
 831     assert(p != NULL, "use append instead");
 832     insert_after(p, tr);
 833   }
 834 
 835   // Append another trace to this one.
 836   void append(Trace *tr) {
 837     insert_after(_last, tr);
 838   }
 839 
 840   // Append a block at the end of this trace
 841   void append(Block *b) {
 842     set_next(_last, b);
 843     set_prev(b, _last);
 844     _last = b;
 845   }
 846 
 847   // Adjust the the blocks in this trace
 848   void fixup_blocks(PhaseCFG &cfg);
 849   bool backedge(CFGEdge *e);
 850 
 851 #ifndef PRODUCT
 852   void dump( ) const;
 853 #endif
 854 };
 855 
 856 //------------------------------PhaseBlockLayout-------------------------------
 857 // Rearrange blocks into some canonical order, based on edges and their frequencies
 858 class PhaseBlockLayout : public Phase {
 859   friend class VMStructs;
 860   PhaseCFG &_cfg;               // Control flow graph
 861 
 862   GrowableArray<CFGEdge *> *edges;
 863   Trace **traces;
 864   Block **next;
 865   Block **prev;
 866   UnionFind *uf;
 867 
 868   // Given a block, find its encompassing Trace
 869   Trace * trace(Block *b) {
 870     return traces[uf->Find_compress(b->_pre_order)];
 871   }
 872  public:
 873   PhaseBlockLayout(PhaseCFG &cfg);
 874 
 875   void find_edges();
 876   void grow_traces();
 877   void merge_traces(bool loose_connections);
 878   void reorder_traces(int count);
 879   void union_traces(Trace* from, Trace* to);
 880 };
 881 
 882 #endif // SHARE_VM_OPTO_BLOCK_HPP
--- EOF ---