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 class PhaseChaitin;
  41 struct Tarjan;
  42 
  43 //------------------------------Block_Array------------------------------------
  44 // Map dense integer indices to Blocks.  Uses classic doubling-array trick.
  45 // Abstractly provides an infinite array of Block*'s, initialized to NULL.
  46 // Note that the constructor just zeros things, and since I use Arena
  47 // allocation I do not need a destructor to reclaim storage.
  48 class Block_Array : public ResourceObj {
  49   friend class VMStructs;
  50   uint _size;                   // allocated size, as opposed to formal limit
  51   debug_only(uint _limit;)      // limit to formal domain
  52   Arena *_arena;                // Arena to allocate in
  53 protected:
  54   Block **_blocks;
  55   void grow( uint i );          // Grow array node to fit
  56 
  57 public:
  58   Block_Array(Arena *a) : _arena(a), _size(OptoBlockListSize) {
  59     debug_only(_limit=0);
  60     _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize );
  61     for( int i = 0; i < OptoBlockListSize; i++ ) {
  62       _blocks[i] = NULL;
  63     }
  64   }
  65   Block *lookup( uint i ) const // Lookup, or NULL for not mapped
  66   { return (i<Max()) ? _blocks[i] : (Block*)NULL; }
  67   Block *operator[] ( uint i ) const // Lookup, or assert for not mapped
  68   { assert( i < Max(), "oob" ); return _blocks[i]; }
  69   // Extend the mapping: index i maps to Block *n.
  70   void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; }
  71   uint Max() const { debug_only(return _limit); return _size; }
  72 };
  73 
  74 
  75 class Block_List : public Block_Array {
  76   friend class VMStructs;
  77 public:
  78   uint _cnt;
  79   Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {}
  80   void push( Block *b ) {  map(_cnt++,b); }
  81   Block *pop() { return _blocks[--_cnt]; }
  82   Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;}
  83   void remove( uint i );
  84   void insert( uint i, Block *n );
  85   uint size() const { return _cnt; }
  86   void reset() { _cnt = 0; }
  87   void print();
  88 };
  89 
  90 
  91 class CFGElement : public ResourceObj {
  92   friend class VMStructs;
  93  public:
  94   double _freq; // Execution frequency (estimate)
  95 
  96   CFGElement() : _freq(0.0) {}
  97   virtual bool is_block() { return false; }
  98   virtual bool is_loop()  { return false; }
  99   Block*   as_Block() { assert(is_block(), "must be block"); return (Block*)this; }
 100   CFGLoop* as_CFGLoop()  { assert(is_loop(),  "must be loop");  return (CFGLoop*)this;  }
 101 };
 102 
 103 //------------------------------Block------------------------------------------
 104 // This class defines a Basic Block.
 105 // Basic blocks are used during the output routines, and are not used during
 106 // any optimization pass.  They are created late in the game.
 107 class Block : public CFGElement {
 108   friend class VMStructs;
 109 
 110 private:
 111   // Nodes in this block, in order
 112   Node_List _nodes;
 113 
 114 public:
 115 
 116   // Get the node at index 'at_index', if 'at_index' is out of bounds return NULL
 117   Node* get_node(uint at_index) const {
 118     return _nodes[at_index];
 119   }
 120 
 121   // Get the number of nodes in this block
 122   uint number_of_nodes() const {
 123     return _nodes.size();
 124   }
 125 
 126   // 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
 127   void map_node(Node* node, uint to_index) {
 128     _nodes.map(to_index, node);
 129   }
 130 
 131   // 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
 132   void insert_node(Node* node, uint at_index) {
 133     _nodes.insert(at_index, node);
 134   }
 135 
 136   // Remove a node at index 'at_index'
 137   void remove_node(uint at_index) {
 138     _nodes.remove(at_index);
 139   }
 140 
 141   // Push a node 'node' onto the node list
 142   void push_node(Node* node) {
 143     _nodes.push(node);
 144   }
 145 
 146   // Pop the last node off the node list
 147   Node* pop_node() {
 148     return _nodes.pop();
 149   }
 150 
 151   // Basic blocks have a Node which defines Control for all Nodes pinned in
 152   // this block.  This Node is a RegionNode.  Exception-causing Nodes
 153   // (division, subroutines) and Phi functions are always pinned.  Later,
 154   // every Node will get pinned to some block.
 155   Node *head() const { return get_node(0); }
 156 
 157   // CAUTION: num_preds() is ONE based, so that predecessor numbers match
 158   // input edges to Regions and Phis.
 159   uint num_preds() const { return head()->req(); }
 160   Node *pred(uint i) const { return head()->in(i); }
 161 
 162   // Array of successor blocks, same size as projs array
 163   Block_Array _succs;
 164 
 165   // Basic blocks have some number of Nodes which split control to all
 166   // following blocks.  These Nodes are always Projections.  The field in
 167   // the Projection and the block-ending Node determine which Block follows.
 168   uint _num_succs;
 169 
 170   // Basic blocks also carry all sorts of good old fashioned DFS information
 171   // used to find loops, loop nesting depth, dominators, etc.
 172   uint _pre_order;              // Pre-order DFS number
 173 
 174   // Dominator tree
 175   uint _dom_depth;              // Depth in dominator tree for fast LCA
 176   Block* _idom;                 // Immediate dominator block
 177 
 178   CFGLoop *_loop;               // Loop to which this block belongs
 179   uint _rpo;                    // Number in reverse post order walk
 180 
 181   virtual bool is_block() { return true; }
 182   float succ_prob(uint i);      // return probability of i'th successor
 183   int num_fall_throughs();      // How many fall-through candidate this block has
 184   void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code
 185   bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate
 186   Block* lone_fall_through();   // Return lone fall-through Block or null
 187 
 188   Block* dom_lca(Block* that);  // Compute LCA in dominator tree.
 189 #ifdef ASSERT
 190   bool dominates(Block* that) {
 191     int dom_diff = this->_dom_depth - that->_dom_depth;
 192     if (dom_diff > 0)  return false;
 193     for (; dom_diff < 0; dom_diff++)  that = that->_idom;
 194     return this == that;
 195   }
 196 #endif
 197 
 198   // Report the alignment required by this block.  Must be a power of 2.
 199   // The previous block will insert nops to get this alignment.
 200   uint code_alignment();
 201   uint compute_loop_alignment();
 202 
 203   // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.
 204   // It is currently also used to scale such frequencies relative to
 205   // FreqCountInvocations relative to the old value of 1500.
 206 #define BLOCK_FREQUENCY(f) ((f * (double) 1500) / FreqCountInvocations)
 207 
 208   // Register Pressure (estimate) for Splitting heuristic
 209   uint _reg_pressure;
 210   uint _ihrp_index;
 211   uint _freg_pressure;
 212   uint _fhrp_index;
 213 
 214   // Mark and visited bits for an LCA calculation in insert_anti_dependences.
 215   // Since they hold unique node indexes, they do not need reinitialization.
 216   node_idx_t _raise_LCA_mark;
 217   void    set_raise_LCA_mark(node_idx_t x)    { _raise_LCA_mark = x; }
 218   node_idx_t  raise_LCA_mark() const          { return _raise_LCA_mark; }
 219   node_idx_t _raise_LCA_visited;
 220   void    set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; }
 221   node_idx_t  raise_LCA_visited() const       { return _raise_LCA_visited; }
 222 
 223   // Estimated size in bytes of first instructions in a loop.
 224   uint _first_inst_size;
 225   uint first_inst_size() const     { return _first_inst_size; }
 226   void set_first_inst_size(uint s) { _first_inst_size = s; }
 227 
 228   // Compute the size of first instructions in this block.
 229   uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra);
 230 
 231   // Compute alignment padding if the block needs it.
 232   // Align a loop if loop's padding is less or equal to padding limit
 233   // or the size of first instructions in the loop > padding.
 234   uint alignment_padding(int current_offset) {
 235     int block_alignment = code_alignment();
 236     int max_pad = block_alignment-relocInfo::addr_unit();
 237     if( max_pad > 0 ) {
 238       assert(is_power_of_2(max_pad+relocInfo::addr_unit()), "");
 239       int current_alignment = current_offset & max_pad;
 240       if( current_alignment != 0 ) {
 241         uint padding = (block_alignment-current_alignment) & max_pad;
 242         if( has_loop_alignment() &&
 243             padding > (uint)MaxLoopPad &&
 244             first_inst_size() <= padding ) {
 245           return 0;
 246         }
 247         return padding;
 248       }
 249     }
 250     return 0;
 251   }
 252 
 253   // Connector blocks. Connector blocks are basic blocks devoid of
 254   // instructions, but may have relevant non-instruction Nodes, such as
 255   // Phis or MergeMems. Such blocks are discovered and marked during the
 256   // RemoveEmpty phase, and elided during Output.
 257   bool _connector;
 258   void set_connector() { _connector = true; }
 259   bool is_connector() const { return _connector; };
 260 
 261   // Loop_alignment will be set for blocks which are at the top of loops.
 262   // The block layout pass may rotate loops such that the loop head may not
 263   // be the sequentially first block of the loop encountered in the linear
 264   // list of blocks.  If the layout pass is not run, loop alignment is set
 265   // for each block which is the head of a loop.
 266   uint _loop_alignment;
 267   void set_loop_alignment(Block *loop_top) {
 268     uint new_alignment = loop_top->compute_loop_alignment();
 269     if (new_alignment > _loop_alignment) {
 270       _loop_alignment = new_alignment;
 271     }
 272   }
 273   uint loop_alignment() const { return _loop_alignment; }
 274   bool has_loop_alignment() const { return loop_alignment() > 0; }
 275 
 276   // Create a new Block with given head Node.
 277   // Creates the (empty) predecessor arrays.
 278   Block( Arena *a, Node *headnode )
 279     : CFGElement(),
 280       _nodes(a),
 281       _succs(a),
 282       _num_succs(0),
 283       _pre_order(0),
 284       _idom(0),
 285       _loop(NULL),
 286       _reg_pressure(0),
 287       _ihrp_index(1),
 288       _freg_pressure(0),
 289       _fhrp_index(1),
 290       _raise_LCA_mark(0),
 291       _raise_LCA_visited(0),
 292       _first_inst_size(999999),
 293       _connector(false),
 294       _loop_alignment(0) {
 295     _nodes.push(headnode);
 296   }
 297 
 298   // Index of 'end' Node
 299   uint end_idx() const {
 300     // %%%%% add a proj after every goto
 301     // so (last->is_block_proj() != last) always, then simplify this code
 302     // This will not give correct end_idx for block 0 when it only contains root.
 303     int last_idx = _nodes.size() - 1;
 304     Node *last  = _nodes[last_idx];
 305     assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], "");
 306     return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs);
 307   }
 308 
 309   // Basic blocks have a Node which ends them.  This Node determines which
 310   // basic block follows this one in the program flow.  This Node is either an
 311   // IfNode, a GotoNode, a JmpNode, or a ReturnNode.
 312   Node *end() const { return _nodes[end_idx()]; }
 313 
 314   // Add an instruction to an existing block.  It must go after the head
 315   // instruction and before the end instruction.
 316   void add_inst( Node *n ) { insert_node(n, end_idx()); }
 317   // Find node in block. Fails if node not in block.
 318   uint find_node( const Node *n ) const;
 319   // Find and remove n from block list
 320   void find_remove( const Node *n );
 321   // Check wether the node is in the block.
 322   bool contains (const Node *n) const;
 323 
 324   // Return the empty status of a block
 325   enum { not_empty, empty_with_goto, completely_empty };
 326   int is_Empty() const;
 327 
 328   // Forward through connectors
 329   Block* non_connector() {
 330     Block* s = this;
 331     while (s->is_connector()) {
 332       s = s->_succs[0];
 333     }
 334     return s;
 335   }
 336 
 337   // Return true if b is a successor of this block
 338   bool has_successor(Block* b) const {
 339     for (uint i = 0; i < _num_succs; i++ ) {
 340       if (non_connector_successor(i) == b) {
 341         return true;
 342       }
 343     }
 344     return false;
 345   }
 346 
 347   // Successor block, after forwarding through connectors
 348   Block* non_connector_successor(int i) const {
 349     return _succs[i]->non_connector();
 350   }
 351 
 352   // Examine block's code shape to predict if it is not commonly executed.
 353   bool has_uncommon_code() const;
 354 
 355 #ifndef PRODUCT
 356   // Debugging print of basic block
 357   void dump_bidx(const Block* orig, outputStream* st = tty) const;
 358   void dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st = tty) const;
 359   void dump_head(const PhaseCFG* cfg, outputStream* st = tty) const;
 360   void dump() const;
 361   void dump(const PhaseCFG* cfg) const;
 362 #endif
 363 };
 364 
 365 
 366 //------------------------------PhaseCFG---------------------------------------
 367 // Build an array of Basic Block pointers, one per Node.
 368 class PhaseCFG : public Phase {
 369   friend class VMStructs;
 370  private:
 371   // ciEnv for retrieving flags
 372   ciEnv* _ciEnv;
 373 
 374   // Root of whole program
 375   RootNode* _root;
 376 
 377   // The block containing the root node
 378   Block* _root_block;
 379 
 380   // List of basic blocks that are created during CFG creation
 381   Block_List _blocks;
 382 
 383   // Count of basic blocks
 384   uint _number_of_blocks;
 385 
 386   // Arena for the blocks to be stored in
 387   Arena* _block_arena;
 388 
 389   // Info used for scheduling
 390   PhaseChaitin* _regalloc;
 391 
 392   // Register pressure heuristic used?
 393   bool _scheduling_for_pressure;
 394 
 395   // The matcher for this compilation
 396   Matcher& _matcher;
 397 
 398   // Map nodes to owning basic block
 399   Block_Array _node_to_block_mapping;
 400 
 401   // Loop from the root
 402   CFGLoop* _root_loop;
 403 
 404   // Outmost loop frequency
 405   double _outer_loop_frequency;
 406 
 407   // Per node latency estimation, valid only during GCM
 408   GrowableArray<uint>* _node_latency;
 409 
 410   // Build a proper looking cfg.  Return count of basic blocks
 411   uint build_cfg();
 412 
 413   // Build the dominator tree so that we know where we can move instructions
 414   void build_dominator_tree();
 415 
 416   // Estimate block frequencies based on IfNode probabilities, so that we know where we want to move instructions
 417   void estimate_block_frequency();
 418 
 419   // Global Code Motion.  See Click's PLDI95 paper.  Place Nodes in specific
 420   // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block.
 421   // Move nodes to ensure correctness from GVN and also try to move nodes out of loops.
 422   void global_code_motion();
 423 
 424   // Schedule Nodes early in their basic blocks.
 425   bool schedule_early(VectorSet &visited, Node_List &roots);
 426 
 427   // For each node, find the latest block it can be scheduled into
 428   // and then select the cheapest block between the latest and earliest
 429   // block to place the node.
 430   void schedule_late(VectorSet &visited, Node_List &stack);
 431 
 432   // Compute the (backwards) latency of a node from a single use
 433   int latency_from_use(Node *n, const Node *def, Node *use);
 434 
 435   // Compute the (backwards) latency of a node from the uses of this instruction
 436   void partial_latency_of_defs(Node *n);
 437 
 438   // Compute the instruction global latency with a backwards walk
 439   void compute_latencies_backwards(VectorSet &visited, Node_List &stack);
 440 
 441   // Pick a block between early and late that is a cheaper alternative
 442   // to late. Helper for schedule_late.
 443   Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);
 444 
 445   bool schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t* recacl_pressure_nodes);
 446   void set_next_call(Block* block, Node* n, VectorSet& next_call);
 447   void needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call);
 448 
 449   // Perform basic-block local scheduling
 450   Node* select(Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, VectorSet& next_call, uint sched_slot,
 451                intptr_t* recacl_pressure_nodes);
 452   void adjust_register_pressure(Node* n, Block* block, intptr_t *recalc_pressure_nodes, bool finalize_mode);
 453 
 454   // Schedule a call next in the block
 455   uint sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call);
 456 
 457   // Cleanup if any code lands between a Call and his Catch
 458   void call_catch_cleanup(Block* block);
 459 
 460   Node* catch_cleanup_find_cloned_def(Block* use_blk, Node* def, Block* def_blk, int n_clone_idx);
 461   void  catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx);
 462 
 463   // Detect implicit-null-check opportunities.  Basically, find NULL checks
 464   // with suitable memory ops nearby.  Use the memory op to do the NULL check.
 465   // I can generate a memory op if there is not one nearby.
 466   void implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons);
 467 
 468   // Perform a Depth First Search (DFS).
 469   // Setup 'vertex' as DFS to vertex mapping.
 470   // Setup 'semi' as vertex to DFS mapping.
 471   // Set 'parent' to DFS parent.
 472   uint do_DFS(Tarjan* tarjan, uint rpo_counter);
 473 
 474   // Helper function to insert a node into a block
 475   void schedule_node_into_block( Node *n, Block *b );
 476 
 477   void replace_block_proj_ctrl( Node *n );
 478 
 479   // Set the basic block for pinned Nodes
 480   void schedule_pinned_nodes( VectorSet &visited );
 481 
 482   // I'll need a few machine-specific GotoNodes.  Clone from this one.
 483   // Used when building the CFG and creating end nodes for blocks.
 484   MachNode* _goto;
 485 
 486   Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
 487   void verify_anti_dependences(Block* LCA, Node* load) {
 488     assert(LCA == get_block_for_node(load), "should already be scheduled");
 489     insert_anti_dependences(LCA, load, true);
 490   }
 491 
 492   bool move_to_next(Block* bx, uint b_index);
 493   void move_to_end(Block* bx, uint b_index);
 494 
 495   void insert_goto_at(uint block_no, uint succ_no);
 496 
 497   // Check for NeverBranch at block end.  This needs to become a GOTO to the
 498   // true target.  NeverBranch are treated as a conditional branch that always
 499   // goes the same direction for most of the optimizer and are used to give a
 500   // fake exit path to infinite loops.  At this late stage they need to turn
 501   // into Goto's so that when you enter the infinite loop you indeed hang.
 502   void convert_NeverBranch_to_Goto(Block *b);
 503 
 504   CFGLoop* create_loop_tree();
 505 
 506   #ifndef PRODUCT
 507   bool _trace_opto_pipelining;  // tracing flag
 508   #endif
 509 
 510  public:
 511   PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher);
 512 
 513   void set_latency_for_node(Node* node, int latency) {
 514     _node_latency->at_put_grow(node->_idx, latency);
 515   }
 516 
 517   uint get_latency_for_node(Node* node) {
 518     return _node_latency->at_grow(node->_idx);
 519   }
 520 
 521   // Get the outer most frequency
 522   double get_outer_loop_frequency() const {
 523     return _outer_loop_frequency;
 524   }
 525 
 526   // Get the root node of the CFG
 527   RootNode* get_root_node() const {
 528     return _root;
 529   }
 530 
 531   // Get the block of the root node
 532   Block* get_root_block() const {
 533     return _root_block;
 534   }
 535 
 536   // Add a block at a position and moves the later ones one step
 537   void add_block_at(uint pos, Block* block) {
 538     _blocks.insert(pos, block);
 539     _number_of_blocks++;
 540   }
 541 
 542   // Adds a block to the top of the block list
 543   void add_block(Block* block) {
 544     _blocks.push(block);
 545     _number_of_blocks++;
 546   }
 547 
 548   // Clear the list of blocks
 549   void clear_blocks() {
 550     _blocks.reset();
 551     _number_of_blocks = 0;
 552   }
 553 
 554   // Get the block at position pos in _blocks
 555   Block* get_block(uint pos) const {
 556     return _blocks[pos];
 557   }
 558 
 559   // Number of blocks
 560   uint number_of_blocks() const {
 561     return _number_of_blocks;
 562   }
 563 
 564   // set which block this node should reside in
 565   void map_node_to_block(const Node* node, Block* block) {
 566     _node_to_block_mapping.map(node->_idx, block);
 567   }
 568 
 569   // removes the mapping from a node to a block
 570   void unmap_node_from_block(const Node* node) {
 571     _node_to_block_mapping.map(node->_idx, NULL);
 572   }
 573 
 574   // get the block in which this node resides
 575   Block* get_block_for_node(const Node* node) const {
 576     return _node_to_block_mapping[node->_idx];
 577   }
 578 
 579   // does this node reside in a block; return true
 580   bool has_block(const Node* node) const {
 581     return (_node_to_block_mapping.lookup(node->_idx) != NULL);
 582   }
 583 
 584   // Use frequency calculations and code shape to predict if the block
 585   // is uncommon.
 586   bool is_uncommon(const Block* block);
 587 
 588 #ifdef ASSERT
 589   Unique_Node_List _raw_oops;
 590 #endif
 591 
 592   // Do global code motion by first building dominator tree and estimate block frequency
 593   // Returns true on success
 594   bool do_global_code_motion();
 595 
 596   // Compute the (backwards) latency of a node from the uses
 597   void latency_from_uses(Node *n);
 598 
 599   // Set loop alignment
 600   void set_loop_alignment();
 601 
 602   // Remove empty basic blocks
 603   void remove_empty_blocks();
 604   Block *fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext);
 605   void fixup_flow();
 606 
 607   // Insert a node into a block at index and map the node to the block
 608   void insert(Block *b, uint idx, Node *n) {
 609     b->insert_node(n , idx);
 610     map_node_to_block(n, b);
 611   }
 612 
 613   // Check all nodes and postalloc_expand them if necessary.
 614   void postalloc_expand(PhaseRegAlloc* _ra);
 615 
 616 #ifndef PRODUCT
 617   bool trace_opto_pipelining() const { return _trace_opto_pipelining; }
 618 
 619   // Debugging print of CFG
 620   void dump( ) const;           // CFG only
 621   void _dump_cfg( const Node *end, VectorSet &visited  ) const;
 622   void verify() const;
 623   void dump_headers();
 624 #else
 625   bool trace_opto_pipelining() const { return false; }
 626 #endif
 627 };
 628 
 629 
 630 //------------------------------UnionFind--------------------------------------
 631 // Map Block indices to a block-index for a cfg-cover.
 632 // Array lookup in the optimized case.
 633 class UnionFind : public ResourceObj {
 634   uint _cnt, _max;
 635   uint* _indices;
 636   ReallocMark _nesting;  // assertion check for reallocations
 637 public:
 638   UnionFind( uint max );
 639   void reset( uint max );  // Reset to identity map for [0..max]
 640 
 641   uint lookup( uint nidx ) const {
 642     return _indices[nidx];
 643   }
 644   uint operator[] (uint nidx) const { return lookup(nidx); }
 645 
 646   void map( uint from_idx, uint to_idx ) {
 647     assert( from_idx < _cnt, "oob" );
 648     _indices[from_idx] = to_idx;
 649   }
 650   void extend( uint from_idx, uint to_idx );
 651 
 652   uint Size() const { return _cnt; }
 653 
 654   uint Find( uint idx ) {
 655     assert( idx < 65536, "Must fit into uint");
 656     uint uf_idx = lookup(idx);
 657     return (uf_idx == idx) ? uf_idx : Find_compress(idx);
 658   }
 659   uint Find_compress( uint idx );
 660   uint Find_const( uint idx ) const;
 661   void Union( uint idx1, uint idx2 );
 662 
 663 };
 664 
 665 //----------------------------BlockProbPair---------------------------
 666 // Ordered pair of Node*.
 667 class BlockProbPair VALUE_OBJ_CLASS_SPEC {
 668 protected:
 669   Block* _target;      // block target
 670   double  _prob;        // probability of edge to block
 671 public:
 672   BlockProbPair() : _target(NULL), _prob(0.0) {}
 673   BlockProbPair(Block* b, double p) : _target(b), _prob(p) {}
 674 
 675   Block* get_target() const { return _target; }
 676   double get_prob() const { return _prob; }
 677 };
 678 
 679 //------------------------------CFGLoop-------------------------------------------
 680 class CFGLoop : public CFGElement {
 681   friend class VMStructs;
 682   int _id;
 683   int _depth;
 684   CFGLoop *_parent;      // root of loop tree is the method level "pseudo" loop, it's parent is null
 685   CFGLoop *_sibling;     // null terminated list
 686   CFGLoop *_child;       // first child, use child's sibling to visit all immediately nested loops
 687   GrowableArray<CFGElement*> _members; // list of members of loop
 688   GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities
 689   double _exit_prob;       // probability any loop exit is taken on a single loop iteration
 690   void update_succ_freq(Block* b, double freq);
 691 
 692  public:
 693   CFGLoop(int id) :
 694     CFGElement(),
 695     _id(id),
 696     _depth(0),
 697     _parent(NULL),
 698     _sibling(NULL),
 699     _child(NULL),
 700     _exit_prob(1.0f) {}
 701   CFGLoop* parent() { return _parent; }
 702   void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg);
 703   void add_member(CFGElement *s) { _members.push(s); }
 704   void add_nested_loop(CFGLoop* cl);
 705   Block* head() {
 706     assert(_members.at(0)->is_block(), "head must be a block");
 707     Block* hd = _members.at(0)->as_Block();
 708     assert(hd->_loop == this, "just checking");
 709     assert(hd->head()->is_Loop(), "must begin with loop head node");
 710     return hd;
 711   }
 712   Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)
 713   void compute_loop_depth(int depth);
 714   void compute_freq(); // compute frequency with loop assuming head freq 1.0f
 715   void scale_freq();   // scale frequency by loop trip count (including outer loops)
 716   double outer_loop_freq() const; // frequency of outer loop
 717   bool in_loop_nest(Block* b);
 718   double trip_count() const { return 1.0 / _exit_prob; }
 719   virtual bool is_loop()  { return true; }
 720   int id() { return _id; }
 721 
 722 #ifndef PRODUCT
 723   void dump( ) const;
 724   void dump_tree() const;
 725 #endif
 726 };
 727 
 728 
 729 //----------------------------------CFGEdge------------------------------------
 730 // A edge between two basic blocks that will be embodied by a branch or a
 731 // fall-through.
 732 class CFGEdge : public ResourceObj {
 733   friend class VMStructs;
 734  private:
 735   Block * _from;        // Source basic block
 736   Block * _to;          // Destination basic block
 737   double _freq;          // Execution frequency (estimate)
 738   int   _state;
 739   bool  _infrequent;
 740   int   _from_pct;
 741   int   _to_pct;
 742 
 743   // Private accessors
 744   int  from_pct() const { return _from_pct; }
 745   int  to_pct()   const { return _to_pct;   }
 746   int  from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
 747   int  to_infrequent()   const { return to_pct()   < BlockLayoutMinDiamondPercentage; }
 748 
 749  public:
 750   enum {
 751     open,               // initial edge state; unprocessed
 752     connected,          // edge used to connect two traces together
 753     interior            // edge is interior to trace (could be backedge)
 754   };
 755 
 756   CFGEdge(Block *from, Block *to, double freq, int from_pct, int to_pct) :
 757     _from(from), _to(to), _freq(freq),
 758     _from_pct(from_pct), _to_pct(to_pct), _state(open) {
 759     _infrequent = from_infrequent() || to_infrequent();
 760   }
 761 
 762   double  freq() const { return _freq; }
 763   Block* from() const { return _from; }
 764   Block* to  () const { return _to;   }
 765   int  infrequent() const { return _infrequent; }
 766   int state() const { return _state; }
 767 
 768   void set_state(int state) { _state = state; }
 769 
 770 #ifndef PRODUCT
 771   void dump( ) const;
 772 #endif
 773 };
 774 
 775 
 776 //-----------------------------------Trace-------------------------------------
 777 // An ordered list of basic blocks.
 778 class Trace : public ResourceObj {
 779  private:
 780   uint _id;             // Unique Trace id (derived from initial block)
 781   Block ** _next_list;  // Array mapping index to next block
 782   Block ** _prev_list;  // Array mapping index to previous block
 783   Block * _first;       // First block in the trace
 784   Block * _last;        // Last block in the trace
 785 
 786   // Return the block that follows "b" in the trace.
 787   Block * next(Block *b) const { return _next_list[b->_pre_order]; }
 788   void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
 789 
 790   // Return the block that precedes "b" in the trace.
 791   Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
 792   void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
 793 
 794   // We've discovered a loop in this trace. Reset last to be "b", and first as
 795   // the block following "b
 796   void break_loop_after(Block *b) {
 797     _last = b;
 798     _first = next(b);
 799     set_prev(_first, NULL);
 800     set_next(_last, NULL);
 801   }
 802 
 803  public:
 804 
 805   Trace(Block *b, Block **next_list, Block **prev_list) :
 806     _first(b),
 807     _last(b),
 808     _next_list(next_list),
 809     _prev_list(prev_list),
 810     _id(b->_pre_order) {
 811     set_next(b, NULL);
 812     set_prev(b, NULL);
 813   };
 814 
 815   // Return the id number
 816   uint id() const { return _id; }
 817   void set_id(uint id) { _id = id; }
 818 
 819   // Return the first block in the trace
 820   Block * first_block() const { return _first; }
 821 
 822   // Return the last block in the trace
 823   Block * last_block() const { return _last; }
 824 
 825   // Insert a trace in the middle of this one after b
 826   void insert_after(Block *b, Trace *tr) {
 827     set_next(tr->last_block(), next(b));
 828     if (next(b) != NULL) {
 829       set_prev(next(b), tr->last_block());
 830     }
 831 
 832     set_next(b, tr->first_block());
 833     set_prev(tr->first_block(), b);
 834 
 835     if (b == _last) {
 836       _last = tr->last_block();
 837     }
 838   }
 839 
 840   void insert_before(Block *b, Trace *tr) {
 841     Block *p = prev(b);
 842     assert(p != NULL, "use append instead");
 843     insert_after(p, tr);
 844   }
 845 
 846   // Append another trace to this one.
 847   void append(Trace *tr) {
 848     insert_after(_last, tr);
 849   }
 850 
 851   // Append a block at the end of this trace
 852   void append(Block *b) {
 853     set_next(_last, b);
 854     set_prev(b, _last);
 855     _last = b;
 856   }
 857 
 858   // Adjust the the blocks in this trace
 859   void fixup_blocks(PhaseCFG &cfg);
 860   bool backedge(CFGEdge *e);
 861 
 862 #ifndef PRODUCT
 863   void dump( ) const;
 864 #endif
 865 };
 866 
 867 //------------------------------PhaseBlockLayout-------------------------------
 868 // Rearrange blocks into some canonical order, based on edges and their frequencies
 869 class PhaseBlockLayout : public Phase {
 870   friend class VMStructs;
 871   PhaseCFG &_cfg;               // Control flow graph
 872 
 873   GrowableArray<CFGEdge *> *edges;
 874   Trace **traces;
 875   Block **next;
 876   Block **prev;
 877   UnionFind *uf;
 878 
 879   // Given a block, find its encompassing Trace
 880   Trace * trace(Block *b) {
 881     return traces[uf->Find_compress(b->_pre_order)];
 882   }
 883  public:
 884   PhaseBlockLayout(PhaseCFG &cfg);
 885 
 886   void find_edges();
 887   void grow_traces();
 888   void merge_traces(bool loose_connections);
 889   void reorder_traces(int count);
 890   void union_traces(Trace* from, Trace* to);
 891 };
 892 
 893 #endif // SHARE_VM_OPTO_BLOCK_HPP