1 /*
   2  * Copyright (c) 1997, 2019, 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_OPTO_BLOCK_HPP
  26 #define SHARE_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) : _size(OptoBlockListSize), _arena(a) {
  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 
 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 
 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() const;
 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 whether 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   // Root of whole program
 371   RootNode* _root;
 372 
 373   // The block containing the root node
 374   Block* _root_block;
 375 
 376   // List of basic blocks that are created during CFG creation
 377   Block_List _blocks;
 378 
 379   // Count of basic blocks
 380   uint _number_of_blocks;
 381 
 382   // Arena for the blocks to be stored in
 383   Arena* _block_arena;
 384 
 385   // Info used for scheduling
 386   PhaseChaitin* _regalloc;
 387 
 388   // Register pressure heuristic used?
 389   bool _scheduling_for_pressure;
 390 
 391   // The matcher for this compilation
 392   Matcher& _matcher;
 393 
 394   // Map nodes to owning basic block
 395   Block_Array _node_to_block_mapping;
 396 
 397   // Loop from the root
 398   CFGLoop* _root_loop;
 399 
 400   // Outmost loop frequency
 401   double _outer_loop_frequency;
 402 
 403   // Per node latency estimation, valid only during GCM
 404   GrowableArray<uint>* _node_latency;
 405 
 406   // Build a proper looking cfg.  Return count of basic blocks
 407   uint build_cfg();
 408 
 409   // Build the dominator tree so that we know where we can move instructions
 410   void build_dominator_tree();
 411 
 412   // Estimate block frequencies based on IfNode probabilities, so that we know where we want to move instructions
 413   void estimate_block_frequency();
 414 
 415   // Global Code Motion.  See Click's PLDI95 paper.  Place Nodes in specific
 416   // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block.
 417   // Move nodes to ensure correctness from GVN and also try to move nodes out of loops.
 418   void global_code_motion();
 419 
 420   // Schedule Nodes early in their basic blocks.
 421   bool schedule_early(VectorSet &visited, Node_Stack &roots);
 422 
 423   // For each node, find the latest block it can be scheduled into
 424   // and then select the cheapest block between the latest and earliest
 425   // block to place the node.
 426   void schedule_late(VectorSet &visited, Node_Stack &stack);
 427 
 428   // Compute the (backwards) latency of a node from a single use
 429   int latency_from_use(Node *n, const Node *def, Node *use);
 430 
 431   // Compute the (backwards) latency of a node from the uses of this instruction
 432   void partial_latency_of_defs(Node *n);
 433 
 434   // Compute the instruction global latency with a backwards walk
 435   void compute_latencies_backwards(VectorSet &visited, Node_Stack &stack);
 436 
 437   // Pick a block between early and late that is a cheaper alternative
 438   // to late. Helper for schedule_late.
 439   Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);
 440 
 441   bool schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t* recacl_pressure_nodes);
 442   void set_next_call(Block* block, Node* n, VectorSet& next_call);
 443   void needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call);
 444 
 445   // Perform basic-block local scheduling
 446   Node* select(Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, VectorSet& next_call, uint sched_slot,
 447                intptr_t* recacl_pressure_nodes);
 448   void adjust_register_pressure(Node* n, Block* block, intptr_t *recalc_pressure_nodes, bool finalize_mode);
 449 
 450   // Schedule a call next in the block
 451   uint sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call);
 452 
 453   // Cleanup if any code lands between a Call and his Catch
 454   void call_catch_cleanup(Block* block);
 455 
 456   Node* catch_cleanup_find_cloned_def(Block* use_blk, Node* def, Block* def_blk, int n_clone_idx);
 457   void  catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx);
 458 
 459   // Detect implicit-null-check opportunities.  Basically, find NULL checks
 460   // with suitable memory ops nearby.  Use the memory op to do the NULL check.
 461   // I can generate a memory op if there is not one nearby.
 462   void implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons);
 463 
 464   // Perform a Depth First Search (DFS).
 465   // Setup 'vertex' as DFS to vertex mapping.
 466   // Setup 'semi' as vertex to DFS mapping.
 467   // Set 'parent' to DFS parent.
 468   uint do_DFS(Tarjan* tarjan, uint rpo_counter);
 469 
 470   // Helper function to insert a node into a block
 471   void schedule_node_into_block( Node *n, Block *b );
 472 
 473   void replace_block_proj_ctrl( Node *n );
 474 
 475   // Set the basic block for pinned Nodes
 476   void schedule_pinned_nodes( VectorSet &visited );
 477 
 478   // I'll need a few machine-specific GotoNodes.  Clone from this one.
 479   // Used when building the CFG and creating end nodes for blocks.
 480   MachNode* _goto;
 481 
 482   Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);
 483   void verify_anti_dependences(Block* LCA, Node* load) const {
 484     assert(LCA == get_block_for_node(load), "should already be scheduled");
 485     const_cast<PhaseCFG*>(this)->insert_anti_dependences(LCA, load, true);
 486   }
 487 
 488   bool move_to_next(Block* bx, uint b_index);
 489   void move_to_end(Block* bx, uint b_index);
 490 
 491   void insert_goto_at(uint block_no, uint succ_no);
 492 
 493   // Check for NeverBranch at block end.  This needs to become a GOTO to the
 494   // true target.  NeverBranch are treated as a conditional branch that always
 495   // goes the same direction for most of the optimizer and are used to give a
 496   // fake exit path to infinite loops.  At this late stage they need to turn
 497   // into Goto's so that when you enter the infinite loop you indeed hang.
 498   void convert_NeverBranch_to_Goto(Block *b);
 499 
 500   CFGLoop* create_loop_tree();
 501   bool is_dominator(Node* dom_node, Node* node);
 502 
 503   #ifndef PRODUCT
 504   bool _trace_opto_pipelining;  // tracing flag
 505   #endif
 506 
 507  public:
 508   PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher);
 509 
 510   void set_latency_for_node(Node* node, int latency) {
 511     _node_latency->at_put_grow(node->_idx, latency);
 512   }
 513 
 514   uint get_latency_for_node(Node* node) {
 515     return _node_latency->at_grow(node->_idx);
 516   }
 517 
 518   // Get the outer most frequency
 519   double get_outer_loop_frequency() const {
 520     return _outer_loop_frequency;
 521   }
 522 
 523   // Get the root node of the CFG
 524   RootNode* get_root_node() const {
 525     return _root;
 526   }
 527 
 528   // Get the block of the root node
 529   Block* get_root_block() const {
 530     return _root_block;
 531   }
 532 
 533   // Add a block at a position and moves the later ones one step
 534   void add_block_at(uint pos, Block* block) {
 535     _blocks.insert(pos, block);
 536     _number_of_blocks++;
 537   }
 538 
 539   // Adds a block to the top of the block list
 540   void add_block(Block* block) {
 541     _blocks.push(block);
 542     _number_of_blocks++;
 543   }
 544 
 545   // Clear the list of blocks
 546   void clear_blocks() {
 547     _blocks.reset();
 548     _number_of_blocks = 0;
 549   }
 550 
 551   // Get the block at position pos in _blocks
 552   Block* get_block(uint pos) const {
 553     return _blocks[pos];
 554   }
 555 
 556   // Number of blocks
 557   uint number_of_blocks() const {
 558     return _number_of_blocks;
 559   }
 560 
 561   // set which block this node should reside in
 562   void map_node_to_block(const Node* node, Block* block) {
 563     _node_to_block_mapping.map(node->_idx, block);
 564   }
 565 
 566   // removes the mapping from a node to a block
 567   void unmap_node_from_block(const Node* node) {
 568     _node_to_block_mapping.map(node->_idx, NULL);
 569   }
 570 
 571   // get the block in which this node resides
 572   Block* get_block_for_node(const Node* node) const {
 573     return _node_to_block_mapping[node->_idx];
 574   }
 575 
 576   // does this node reside in a block; return true
 577   bool has_block(const Node* node) const {
 578     return (_node_to_block_mapping.lookup(node->_idx) != NULL);
 579   }
 580 
 581   // Use frequency calculations and code shape to predict if the block
 582   // is uncommon.
 583   bool is_uncommon(const Block* block);
 584 
 585 #ifdef ASSERT
 586   Unique_Node_List _raw_oops;
 587 #endif
 588 
 589   // Do global code motion by first building dominator tree and estimate block frequency
 590   // Returns true on success
 591   bool do_global_code_motion();
 592 
 593   // Compute the (backwards) latency of a node from the uses
 594   void latency_from_uses(Node *n);
 595 
 596   // Set loop alignment
 597   void set_loop_alignment();
 598 
 599   // Remove empty basic blocks
 600   void remove_empty_blocks();
 601   Block *fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext);
 602   void fixup_flow();
 603 
 604   // Insert a node into a block at index and map the node to the block
 605   void insert(Block *b, uint idx, Node *n) {
 606     b->insert_node(n , idx);
 607     map_node_to_block(n, b);
 608   }
 609 
 610   // Check all nodes and postalloc_expand them if necessary.
 611   void postalloc_expand(PhaseRegAlloc* _ra);
 612 
 613 #ifndef PRODUCT
 614   bool trace_opto_pipelining() const { return _trace_opto_pipelining; }
 615 
 616   // Debugging print of CFG
 617   void dump( ) const;           // CFG only
 618   void _dump_cfg( const Node *end, VectorSet &visited  ) const;
 619   void verify() const;
 620   void dump_headers();
 621 #else
 622   bool trace_opto_pipelining() const { return false; }
 623 #endif
 624 };
 625 
 626 
 627 //------------------------------UnionFind--------------------------------------
 628 // Map Block indices to a block-index for a cfg-cover.
 629 // Array lookup in the optimized case.
 630 class UnionFind : public ResourceObj {
 631   uint _cnt, _max;
 632   uint* _indices;
 633   ReallocMark _nesting;  // assertion check for reallocations
 634 public:
 635   UnionFind( uint max );
 636   void reset( uint max );  // Reset to identity map for [0..max]
 637 
 638   uint lookup( uint nidx ) const {
 639     return _indices[nidx];
 640   }
 641   uint operator[] (uint nidx) const { return lookup(nidx); }
 642 
 643   void map( uint from_idx, uint to_idx ) {
 644     assert( from_idx < _cnt, "oob" );
 645     _indices[from_idx] = to_idx;
 646   }
 647   void extend( uint from_idx, uint to_idx );
 648 
 649   uint Size() const { return _cnt; }
 650 
 651   uint Find( uint idx ) {
 652     assert( idx < 65536, "Must fit into uint");
 653     uint uf_idx = lookup(idx);
 654     return (uf_idx == idx) ? uf_idx : Find_compress(idx);
 655   }
 656   uint Find_compress( uint idx );
 657   uint Find_const( uint idx ) const;
 658   void Union( uint idx1, uint idx2 );
 659 
 660 };
 661 
 662 //----------------------------BlockProbPair---------------------------
 663 // Ordered pair of Node*.
 664 class BlockProbPair {
 665 protected:
 666   Block* _target;      // block target
 667   double  _prob;        // probability of edge to block
 668 public:
 669   BlockProbPair() : _target(NULL), _prob(0.0) {}
 670   BlockProbPair(Block* b, double p) : _target(b), _prob(p) {}
 671 
 672   Block* get_target() const { return _target; }
 673   double get_prob() const { return _prob; }
 674 };
 675 
 676 //------------------------------CFGLoop-------------------------------------------
 677 class CFGLoop : public CFGElement {
 678   friend class VMStructs;
 679   int _id;
 680   int _depth;
 681   CFGLoop *_parent;      // root of loop tree is the method level "pseudo" loop, it's parent is null
 682   CFGLoop *_sibling;     // null terminated list
 683   CFGLoop *_child;       // first child, use child's sibling to visit all immediately nested loops
 684   GrowableArray<CFGElement*> _members; // list of members of loop
 685   GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities
 686   double _exit_prob;       // probability any loop exit is taken on a single loop iteration
 687   void update_succ_freq(Block* b, double freq);
 688 
 689  public:
 690   CFGLoop(int id) :
 691     CFGElement(),
 692     _id(id),
 693     _depth(0),
 694     _parent(NULL),
 695     _sibling(NULL),
 696     _child(NULL),
 697     _exit_prob(1.0f) {}
 698   CFGLoop* parent() { return _parent; }
 699   void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg);
 700   void add_member(CFGElement *s) { _members.push(s); }
 701   void add_nested_loop(CFGLoop* cl);
 702   Block* head() {
 703     assert(_members.at(0)->is_block(), "head must be a block");
 704     Block* hd = _members.at(0)->as_Block();
 705     assert(hd->_loop == this, "just checking");
 706     assert(hd->head()->is_Loop(), "must begin with loop head node");
 707     return hd;
 708   }
 709   Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)
 710   void compute_loop_depth(int depth);
 711   void compute_freq(); // compute frequency with loop assuming head freq 1.0f
 712   void scale_freq();   // scale frequency by loop trip count (including outer loops)
 713   double outer_loop_freq() const; // frequency of outer loop
 714   bool in_loop_nest(Block* b);
 715   double trip_count() const { return 1.0 / _exit_prob; }
 716   virtual bool is_loop()  { return true; }
 717   int id() { return _id; }
 718 
 719 #ifndef PRODUCT
 720   void dump( ) const;
 721   void dump_tree() const;
 722 #endif
 723 };
 724 
 725 
 726 //----------------------------------CFGEdge------------------------------------
 727 // A edge between two basic blocks that will be embodied by a branch or a
 728 // fall-through.
 729 class CFGEdge : public ResourceObj {
 730   friend class VMStructs;
 731  private:
 732   Block * _from;        // Source basic block
 733   Block * _to;          // Destination basic block
 734   double _freq;          // Execution frequency (estimate)
 735   int   _state;
 736   bool  _infrequent;
 737   int   _from_pct;
 738   int   _to_pct;
 739 
 740   // Private accessors
 741   int  from_pct() const { return _from_pct; }
 742   int  to_pct()   const { return _to_pct;   }
 743   int  from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; }
 744   int  to_infrequent()   const { return to_pct()   < BlockLayoutMinDiamondPercentage; }
 745 
 746  public:
 747   enum {
 748     open,               // initial edge state; unprocessed
 749     connected,          // edge used to connect two traces together
 750     interior            // edge is interior to trace (could be backedge)
 751   };
 752 
 753   CFGEdge(Block *from, Block *to, double freq, int from_pct, int to_pct) :
 754     _from(from), _to(to), _freq(freq),
 755     _state(open), _from_pct(from_pct), _to_pct(to_pct) {
 756     _infrequent = from_infrequent() || to_infrequent();
 757   }
 758 
 759   double  freq() const { return _freq; }
 760   Block* from() const { return _from; }
 761   Block* to  () const { return _to;   }
 762   int  infrequent() const { return _infrequent; }
 763   int state() const { return _state; }
 764 
 765   void set_state(int state) { _state = state; }
 766 
 767 #ifndef PRODUCT
 768   void dump( ) const;
 769 #endif
 770 };
 771 
 772 
 773 //-----------------------------------Trace-------------------------------------
 774 // An ordered list of basic blocks.
 775 class Trace : public ResourceObj {
 776  private:
 777   uint _id;             // Unique Trace id (derived from initial block)
 778   Block ** _next_list;  // Array mapping index to next block
 779   Block ** _prev_list;  // Array mapping index to previous block
 780   Block * _first;       // First block in the trace
 781   Block * _last;        // Last block in the trace
 782 
 783   // Return the block that follows "b" in the trace.
 784   Block * next(Block *b) const { return _next_list[b->_pre_order]; }
 785   void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; }
 786 
 787   // Return the block that precedes "b" in the trace.
 788   Block * prev(Block *b) const { return _prev_list[b->_pre_order]; }
 789   void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; }
 790 
 791   // We've discovered a loop in this trace. Reset last to be "b", and first as
 792   // the block following "b
 793   void break_loop_after(Block *b) {
 794     _last = b;
 795     _first = next(b);
 796     set_prev(_first, NULL);
 797     set_next(_last, NULL);
 798   }
 799 
 800  public:
 801 
 802   Trace(Block *b, Block **next_list, Block **prev_list) :
 803     _id(b->_pre_order),
 804     _next_list(next_list),
 805     _prev_list(prev_list),
 806     _first(b),
 807     _last(b) {
 808     set_next(b, NULL);
 809     set_prev(b, NULL);
 810   };
 811 
 812   // Return the id number
 813   uint id() const { return _id; }
 814   void set_id(uint id) { _id = id; }
 815 
 816   // Return the first block in the trace
 817   Block * first_block() const { return _first; }
 818 
 819   // Return the last block in the trace
 820   Block * last_block() const { return _last; }
 821 
 822   // Insert a trace in the middle of this one after b
 823   void insert_after(Block *b, Trace *tr) {
 824     set_next(tr->last_block(), next(b));
 825     if (next(b) != NULL) {
 826       set_prev(next(b), tr->last_block());
 827     }
 828 
 829     set_next(b, tr->first_block());
 830     set_prev(tr->first_block(), b);
 831 
 832     if (b == _last) {
 833       _last = tr->last_block();
 834     }
 835   }
 836 
 837   void insert_before(Block *b, Trace *tr) {
 838     Block *p = prev(b);
 839     assert(p != NULL, "use append instead");
 840     insert_after(p, tr);
 841   }
 842 
 843   // Append another trace to this one.
 844   void append(Trace *tr) {
 845     insert_after(_last, tr);
 846   }
 847 
 848   // Append a block at the end of this trace
 849   void append(Block *b) {
 850     set_next(_last, b);
 851     set_prev(b, _last);
 852     _last = b;
 853   }
 854 
 855   // Adjust the the blocks in this trace
 856   void fixup_blocks(PhaseCFG &cfg);
 857   bool backedge(CFGEdge *e);
 858 
 859 #ifndef PRODUCT
 860   void dump( ) const;
 861 #endif
 862 };
 863 
 864 //------------------------------PhaseBlockLayout-------------------------------
 865 // Rearrange blocks into some canonical order, based on edges and their frequencies
 866 class PhaseBlockLayout : public Phase {
 867   friend class VMStructs;
 868   PhaseCFG &_cfg;               // Control flow graph
 869 
 870   GrowableArray<CFGEdge *> *edges;
 871   Trace **traces;
 872   Block **next;
 873   Block **prev;
 874   UnionFind *uf;
 875 
 876   // Given a block, find its encompassing Trace
 877   Trace * trace(Block *b) {
 878     return traces[uf->Find_compress(b->_pre_order)];
 879   }
 880  public:
 881   PhaseBlockLayout(PhaseCFG &cfg);
 882 
 883   void find_edges();
 884   void grow_traces();
 885   void merge_traces(bool loose_connections);
 886   void reorder_traces(int count);
 887   void union_traces(Trace* from, Trace* to);
 888 };
 889 
 890 #endif // SHARE_OPTO_BLOCK_HPP