1 /*

     2  * Copyright 1997-2007 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,

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

    21  * have any questions.

    22  *

    23  */

    24 

    25 // Optimization - Graph Style

    26 

    27 class Block;

    28 class CFGLoop;

    29 class MachCallNode;

    30 class Matcher;

    31 class RootNode;

    32 class VectorSet;

    33 struct Tarjan;

    34 

    35 //------------------------------Block_Array------------------------------------

    36 // Map dense integer indices to Blocks.  Uses classic doubling-array trick.

    37 // Abstractly provides an infinite array of Block*'s, initialized to NULL.

    38 // Note that the constructor just zeros things, and since I use Arena

    39 // allocation I do not need a destructor to reclaim storage.

    40 class Block_Array : public ResourceObj {

    41   uint _size;                   // allocated size, as opposed to formal limit

    42   debug_only(uint _limit;)      // limit to formal domain

    43 protected:

    44   Block **_blocks;

    45   void grow( uint i );          // Grow array node to fit

    46 

    47 public:

    48   Arena *_arena;                // Arena to allocate in

    49 

    50   Block_Array(Arena *a) : _arena(a), _size(OptoBlockListSize) {

    51     debug_only(_limit=0);

    52     _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize );

    53     for( int i = 0; i < OptoBlockListSize; i++ ) {

    54       _blocks[i] = NULL;

    55     }

    56   }

    57   Block *lookup( uint i ) const // Lookup, or NULL for not mapped

    58   { return (i<Max()) ? _blocks[i] : (Block*)NULL; }

    59   Block *operator[] ( uint i ) const // Lookup, or assert for not mapped

    60   { assert( i < Max(), "oob" ); return _blocks[i]; }

    61   // Extend the mapping: index i maps to Block *n.

    62   void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; }

    63   uint Max() const { debug_only(return _limit); return _size; }

    64 };

    65 

    66 

    67 class Block_List : public Block_Array {

    68 public:

    69   uint _cnt;

    70   Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {}

    71   void push( Block *b ) { map(_cnt++,b); }

    72   Block *pop() { return _blocks[--_cnt]; }

    73   Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;}

    74   void remove( uint i );

    75   void insert( uint i, Block *n );

    76   uint size() const { return _cnt; }

    77   void reset() { _cnt = 0; }

    78 };

    79 

    80 

    81 class CFGElement : public ResourceObj {

    82  public:

    83   float _freq; // Execution frequency (estimate)

    84 

    85   CFGElement() : _freq(0.0f) {}

    86   virtual bool is_block() { return false; }

    87   virtual bool is_loop()  { return false; }

    88   Block*   as_Block() { assert(is_block(), "must be block"); return (Block*)this; }

    89   CFGLoop* as_CFGLoop()  { assert(is_loop(),  "must be loop");  return (CFGLoop*)this;  }

    90 };

    91 

    92 //------------------------------Block------------------------------------------

    93 // This class defines a Basic Block.

    94 // Basic blocks are used during the output routines, and are not used during

    95 // any optimization pass.  They are created late in the game.

    96 class Block : public CFGElement {

    97  public:

    98   // Nodes in this block, in order

    99   Node_List _nodes;

   100 

   101   // Basic blocks have a Node which defines Control for all Nodes pinned in

   102   // this block.  This Node is a RegionNode.  Exception-causing Nodes

   103   // (division, subroutines) and Phi functions are always pinned.  Later,

   104   // every Node will get pinned to some block.

   105   Node *head() const { return _nodes[0]; }

   106 

   107   // CAUTION: num_preds() is ONE based, so that predecessor numbers match

   108   // input edges to Regions and Phis.

   109   uint num_preds() const { return head()->req(); }

   110   Node *pred(uint i) const { return head()->in(i); }

   111 

   112   // Array of successor blocks, same size as projs array

   113   Block_Array _succs;

   114 

   115   // Basic blocks have some number of Nodes which split control to all

   116   // following blocks.  These Nodes are always Projections.  The field in

   117   // the Projection and the block-ending Node determine which Block follows.

   118   uint _num_succs;

   119 

   120   // Basic blocks also carry all sorts of good old fashioned DFS information

   121   // used to find loops, loop nesting depth, dominators, etc.

   122   uint _pre_order;              // Pre-order DFS number

   123 

   124   // Dominator tree

   125   uint _dom_depth;              // Depth in dominator tree for fast LCA

   126   Block* _idom;                 // Immediate dominator block

   127 

   128   CFGLoop *_loop;               // Loop to which this block belongs

   129   uint _rpo;                    // Number in reverse post order walk

   130 

   131   virtual bool is_block() { return true; }

   132   float succ_prob(uint i); // return probability of i'th successor

   133 

   134   Block* dom_lca(Block* that);  // Compute LCA in dominator tree.

   135 #ifdef ASSERT

   136   bool dominates(Block* that) {

   137     int dom_diff = this->_dom_depth - that->_dom_depth;

   138     if (dom_diff > 0)  return false;

   139     for (; dom_diff < 0; dom_diff++)  that = that->_idom;

   140     return this == that;

   141   }

   142 #endif

   143 

   144   // Report the alignment required by this block.  Must be a power of 2.

   145   // The previous block will insert nops to get this alignment.

   146   uint code_alignment();

   147 

   148   // BLOCK_FREQUENCY is a sentinel to mark uses of constant block frequencies.

   149   // It is currently also used to scale such frequencies relative to

   150   // FreqCountInvocations relative to the old value of 1500.

   151 #define BLOCK_FREQUENCY(f) ((f * (float) 1500) / FreqCountInvocations)

   152 

   153   // Register Pressure (estimate) for Splitting heuristic

   154   uint _reg_pressure;

   155   uint _ihrp_index;

   156   uint _freg_pressure;

   157   uint _fhrp_index;

   158 

   159   // Mark and visited bits for an LCA calculation in insert_anti_dependences.

   160   // Since they hold unique node indexes, they do not need reinitialization.

   161   node_idx_t _raise_LCA_mark;

   162   void    set_raise_LCA_mark(node_idx_t x)    { _raise_LCA_mark = x; }

   163   node_idx_t  raise_LCA_mark() const          { return _raise_LCA_mark; }

   164   node_idx_t _raise_LCA_visited;

   165   void    set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; }

   166   node_idx_t  raise_LCA_visited() const       { return _raise_LCA_visited; }

   167 

   168   // Estimated size in bytes of first instructions in a loop.

   169   uint _first_inst_size;

   170   uint first_inst_size() const     { return _first_inst_size; }

   171   void set_first_inst_size(uint s) { _first_inst_size = s; }

   172 

   173   // Compute the size of first instructions in this block.

   174   uint compute_first_inst_size(uint& sum_size, uint inst_cnt, PhaseRegAlloc* ra);

   175 

   176   // Compute alignment padding if the block needs it.

   177   // Align a loop if loop's padding is less or equal to padding limit

   178   // or the size of first instructions in the loop > padding.

   179   uint alignment_padding(int current_offset) {

   180     int block_alignment = code_alignment();

   181     int max_pad = block_alignment-relocInfo::addr_unit();

   182     if( max_pad > 0 ) {

   183       assert(is_power_of_2(max_pad+relocInfo::addr_unit()), "");

   184       int current_alignment = current_offset & max_pad;

   185       if( current_alignment != 0 ) {

   186         uint padding = (block_alignment-current_alignment) & max_pad;

   187         if( !head()->is_Loop() ||

   188             padding <= (uint)MaxLoopPad ||

   189             first_inst_size() > padding ) {

   190           return padding;

   191         }

   192       }

   193     }

   194     return 0;

   195   }

   196 

   197   // Connector blocks. Connector blocks are basic blocks devoid of

   198   // instructions, but may have relevant non-instruction Nodes, such as

   199   // Phis or MergeMems. Such blocks are discovered and marked during the

   200   // RemoveEmpty phase, and elided during Output.

   201   bool _connector;

   202   void set_connector() { _connector = true; }

   203   bool is_connector() const { return _connector; };

   204 

   205   // Create a new Block with given head Node.

   206   // Creates the (empty) predecessor arrays.

   207   Block( Arena *a, Node *headnode )

   208     : CFGElement(),

   209       _nodes(a),

   210       _succs(a),

   211       _num_succs(0),

   212       _pre_order(0),

   213       _idom(0),

   214       _loop(NULL),

   215       _reg_pressure(0),

   216       _ihrp_index(1),

   217       _freg_pressure(0),

   218       _fhrp_index(1),

   219       _raise_LCA_mark(0),

   220       _raise_LCA_visited(0),

   221       _first_inst_size(999999),

   222       _connector(false) {

   223     _nodes.push(headnode);

   224   }

   225 

   226   // Index of 'end' Node

   227   uint end_idx() const {

   228     // %%%%% add a proj after every goto

   229     // so (last->is_block_proj() != last) always, then simplify this code

   230     // This will not give correct end_idx for block 0 when it only contains root.

   231     int last_idx = _nodes.size() - 1;

   232     Node *last  = _nodes[last_idx];

   233     assert(last->is_block_proj() == last || last->is_block_proj() == _nodes[last_idx - _num_succs], "");

   234     return (last->is_block_proj() == last) ? last_idx : (last_idx - _num_succs);

   235   }

   236 

   237   // Basic blocks have a Node which ends them.  This Node determines which

   238   // basic block follows this one in the program flow.  This Node is either an

   239   // IfNode, a GotoNode, a JmpNode, or a ReturnNode.

   240   Node *end() const { return _nodes[end_idx()]; }

   241 

   242   // Add an instruction to an existing block.  It must go after the head

   243   // instruction and before the end instruction.

   244   void add_inst( Node *n ) { _nodes.insert(end_idx(),n); }

   245   // Find node in block

   246   uint find_node( const Node *n ) const;

   247   // Find and remove n from block list

   248   void find_remove( const Node *n );

   249 

   250   // Schedule a call next in the block

   251   uint sched_call(Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call);

   252 

   253   // Perform basic-block local scheduling

   254   Node *select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot);

   255   void set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs );

   256   void needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs);

   257   bool schedule_local(PhaseCFG *cfg, Matcher &m, int *ready_cnt, VectorSet &next_call);

   258   // Cleanup if any code lands between a Call and his Catch

   259   void call_catch_cleanup(Block_Array &bbs);

   260   // Detect implicit-null-check opportunities.  Basically, find NULL checks

   261   // with suitable memory ops nearby.  Use the memory op to do the NULL check.

   262   // I can generate a memory op if there is not one nearby.

   263   void implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons);

   264 

   265   // Return the empty status of a block

   266   enum { not_empty, empty_with_goto, completely_empty };

   267   int is_Empty() const;

   268 

   269   // Forward through connectors

   270   Block* non_connector() {

   271     Block* s = this;

   272     while (s->is_connector()) {

   273       s = s->_succs[0];

   274     }

   275     return s;

   276   }

   277 

   278   // Successor block, after forwarding through connectors

   279   Block* non_connector_successor(int i) const {

   280     return _succs[i]->non_connector();

   281   }

   282 

   283   // Examine block's code shape to predict if it is not commonly executed.

   284   bool has_uncommon_code() const;

   285 

   286   // Use frequency calculations and code shape to predict if the block

   287   // is uncommon.

   288   bool is_uncommon( Block_Array &bbs ) const;

   289 

   290 #ifndef PRODUCT

   291   // Debugging print of basic block

   292   void dump_bidx(const Block* orig) const;

   293   void dump_pred(const Block_Array *bbs, Block* orig) const;

   294   void dump_head( const Block_Array *bbs ) const;

   295   void dump( ) const;

   296   void dump( const Block_Array *bbs ) const;

   297 #endif

   298 };

   299 

   300 

   301 //------------------------------PhaseCFG---------------------------------------

   302 // Build an array of Basic Block pointers, one per Node.

   303 class PhaseCFG : public Phase {

   304  private:

   305   // Build a proper looking cfg.  Return count of basic blocks

   306   uint build_cfg();

   307 

   308   // Perform DFS search.

   309   // Setup 'vertex' as DFS to vertex mapping.

   310   // Setup 'semi' as vertex to DFS mapping.

   311   // Set 'parent' to DFS parent.

   312   uint DFS( Tarjan *tarjan );

   313 

   314   // Helper function to insert a node into a block

   315   void schedule_node_into_block( Node *n, Block *b );

   316 

   317   // Set the basic block for pinned Nodes

   318   void schedule_pinned_nodes( VectorSet &visited );

   319 

   320   // I'll need a few machine-specific GotoNodes.  Clone from this one.

   321   MachNode *_goto;

   322   void insert_goto_at(uint block_no, uint succ_no);

   323 

   324   Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false);

   325   void verify_anti_dependences(Block* LCA, Node* load) {

   326     assert(LCA == _bbs[load->_idx], "should already be scheduled");

   327     insert_anti_dependences(LCA, load, true);

   328   }

   329 

   330  public:

   331   PhaseCFG( Arena *a, RootNode *r, Matcher &m );

   332 

   333   uint _num_blocks;             // Count of basic blocks

   334   Block_List _blocks;           // List of basic blocks

   335   RootNode *_root;              // Root of whole program

   336   Block_Array _bbs;             // Map Nodes to owning Basic Block

   337   Block *_broot;                // Basic block of root

   338   uint _rpo_ctr;

   339   CFGLoop* _root_loop;

   340 

   341   // Per node latency estimation, valid only during GCM

   342   GrowableArray<uint> _node_latency;

   343 

   344 #ifndef PRODUCT

   345   bool _trace_opto_pipelining;  // tracing flag

   346 #endif

   347 

   348   // Build dominators

   349   void Dominators();

   350 

   351   // Estimate block frequencies based on IfNode probabilities

   352   void Estimate_Block_Frequency();

   353 

   354   // Global Code Motion.  See Click's PLDI95 paper.  Place Nodes in specific

   355   // basic blocks; i.e. _bbs now maps _idx for all Nodes to some Block.

   356   void GlobalCodeMotion( Matcher &m, uint unique, Node_List &proj_list );

   357 

   358   // Compute the (backwards) latency of a node from the uses

   359   void latency_from_uses(Node *n);

   360 

   361   // Compute the (backwards) latency of a node from a single use

   362   int latency_from_use(Node *n, const Node *def, Node *use);

   363 

   364   // Compute the (backwards) latency of a node from the uses of this instruction

   365   void partial_latency_of_defs(Node *n);

   366 

   367   // Schedule Nodes early in their basic blocks.

   368   bool schedule_early(VectorSet &visited, Node_List &roots);

   369 

   370   // For each node, find the latest block it can be scheduled into

   371   // and then select the cheapest block between the latest and earliest

   372   // block to place the node.

   373   void schedule_late(VectorSet &visited, Node_List &stack);

   374 

   375   // Pick a block between early and late that is a cheaper alternative

   376   // to late. Helper for schedule_late.

   377   Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self);

   378 

   379   // Compute the instruction global latency with a backwards walk

   380   void ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack);

   381 

   382   // Remove empty basic blocks

   383   void RemoveEmpty();

   384   bool MoveToNext(Block* bx, uint b_index);

   385   void MoveToEnd(Block* bx, uint b_index);

   386 

   387   // Check for NeverBranch at block end.  This needs to become a GOTO to the

   388   // true target.  NeverBranch are treated as a conditional branch that always

   389   // goes the same direction for most of the optimizer and are used to give a

   390   // fake exit path to infinite loops.  At this late stage they need to turn

   391   // into Goto's so that when you enter the infinite loop you indeed hang.

   392   void convert_NeverBranch_to_Goto(Block *b);

   393 

   394   CFGLoop* create_loop_tree();

   395 

   396   // Insert a node into a block, and update the _bbs

   397   void insert( Block *b, uint idx, Node *n ) {

   398     b->_nodes.insert( idx, n );

   399     _bbs.map( n->_idx, b );

   400   }

   401 

   402 #ifndef PRODUCT

   403   bool trace_opto_pipelining() const { return _trace_opto_pipelining; }

   404 

   405   // Debugging print of CFG

   406   void dump( ) const;           // CFG only

   407   void _dump_cfg( const Node *end, VectorSet &visited  ) const;

   408   void verify() const;

   409   void dump_headers();

   410 #else

   411   bool trace_opto_pipelining() const { return false; }

   412 #endif

   413 };

   414 

   415 

   416 //------------------------------UnionFindInfo----------------------------------

   417 // Map Block indices to a block-index for a cfg-cover.

   418 // Array lookup in the optimized case.

   419 class UnionFind : public ResourceObj {

   420   uint _cnt, _max;

   421   uint* _indices;

   422   ReallocMark _nesting;  // assertion check for reallocations

   423 public:

   424   UnionFind( uint max );

   425   void reset( uint max );  // Reset to identity map for [0..max]

   426 

   427   uint lookup( uint nidx ) const {

   428     return _indices[nidx];

   429   }

   430   uint operator[] (uint nidx) const { return lookup(nidx); }

   431 

   432   void map( uint from_idx, uint to_idx ) {

   433     assert( from_idx < _cnt, "oob" );

   434     _indices[from_idx] = to_idx;

   435   }

   436   void extend( uint from_idx, uint to_idx );

   437 

   438   uint Size() const { return _cnt; }

   439 

   440   uint Find( uint idx ) {

   441     assert( idx < 65536, "Must fit into uint");

   442     uint uf_idx = lookup(idx);

   443     return (uf_idx == idx) ? uf_idx : Find_compress(idx);

   444   }

   445   uint Find_compress( uint idx );

   446   uint Find_const( uint idx ) const;

   447   void Union( uint idx1, uint idx2 );

   448 

   449 };

   450 

   451 //----------------------------BlockProbPair---------------------------

   452 // Ordered pair of Node*.

   453 class BlockProbPair VALUE_OBJ_CLASS_SPEC {

   454 protected:

   455   Block* _target;      // block target

   456   float  _prob;        // probability of edge to block

   457 public:

   458   BlockProbPair() : _target(NULL), _prob(0.0) {}

   459   BlockProbPair(Block* b, float p) : _target(b), _prob(p) {}

   460 

   461   Block* get_target() const { return _target; }

   462   float get_prob() const { return _prob; }

   463 };

   464 

   465 //------------------------------CFGLoop-------------------------------------------

   466 class CFGLoop : public CFGElement {

   467   int _id;

   468   int _depth;

   469   CFGLoop *_parent;      // root of loop tree is the method level "pseudo" loop, it's parent is null

   470   CFGLoop *_sibling;     // null terminated list

   471   CFGLoop *_child;       // first child, use child's sibling to visit all immediately nested loops

   472   GrowableArray<CFGElement*> _members; // list of members of loop

   473   GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities

   474   float _exit_prob;       // probability any loop exit is taken on a single loop iteration

   475   void update_succ_freq(Block* b, float freq);

   476 

   477  public:

   478   CFGLoop(int id) :

   479     CFGElement(),

   480     _id(id),

   481     _depth(0),

   482     _parent(NULL),

   483     _sibling(NULL),

   484     _child(NULL),

   485     _exit_prob(1.0f) {}

   486   CFGLoop* parent() { return _parent; }

   487   void push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk);

   488   void add_member(CFGElement *s) { _members.push(s); }

   489   void add_nested_loop(CFGLoop* cl);

   490   Block* head() {

   491     assert(_members.at(0)->is_block(), "head must be a block");

   492     Block* hd = _members.at(0)->as_Block();

   493     assert(hd->_loop == this, "just checking");

   494     assert(hd->head()->is_Loop(), "must begin with loop head node");

   495     return hd;

   496   }

   497   Block* backedge_block(); // Return the block on the backedge of the loop (else NULL)

   498   void compute_loop_depth(int depth);

   499   void compute_freq(); // compute frequency with loop assuming head freq 1.0f

   500   void scale_freq();   // scale frequency by loop trip count (including outer loops)

   501   bool in_loop_nest(Block* b);

   502   float trip_count() const { return 1.0f / _exit_prob; }

   503   virtual bool is_loop()  { return true; }

   504   int id() { return _id; }

   505 

   506 #ifndef PRODUCT

   507   void dump( ) const;

   508   void dump_tree() const;

   509 #endif

   510 };