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 #include "utilities/powerOfTwo.hpp" 32 33 // Optimization - Graph Style 34 35 class Block; 36 class CFGLoop; 37 class MachCallNode; 38 class Matcher; 39 class RootNode; 40 class VectorSet; 41 class PhaseChaitin; 42 struct Tarjan; 43 44 //------------------------------Block_Array------------------------------------ 45 // Map dense integer indices to Blocks. Uses classic doubling-array trick. 46 // Abstractly provides an infinite array of Block*'s, initialized to NULL. 47 // Note that the constructor just zeros things, and since I use Arena 48 // allocation I do not need a destructor to reclaim storage. 49 class Block_Array : public ResourceObj { 50 friend class VMStructs; 51 uint _size; // allocated size, as opposed to formal limit 52 debug_only(uint _limit;) // limit to formal domain 53 Arena *_arena; // Arena to allocate in 54 protected: 55 Block **_blocks; 56 void grow( uint i ); // Grow array node to fit 57 58 public: Block_Array(Arena * a)59 Block_Array(Arena *a) : _size(OptoBlockListSize), _arena(a) { 60 debug_only(_limit=0); 61 _blocks = NEW_ARENA_ARRAY( a, Block *, OptoBlockListSize ); 62 for( int i = 0; i < OptoBlockListSize; i++ ) { 63 _blocks[i] = NULL; 64 } 65 } lookup(uint i) const66 Block *lookup( uint i ) const // Lookup, or NULL for not mapped 67 { return (i<Max()) ? _blocks[i] : (Block*)NULL; } operator [](uint i) const68 Block *operator[] ( uint i ) const // Lookup, or assert for not mapped 69 { assert( i < Max(), "oob" ); return _blocks[i]; } 70 // Extend the mapping: index i maps to Block *n. map(uint i,Block * n)71 void map( uint i, Block *n ) { if( i>=Max() ) grow(i); _blocks[i] = n; } Max() const72 uint Max() const { debug_only(return _limit); return _size; } 73 }; 74 75 76 class Block_List : public Block_Array { 77 friend class VMStructs; 78 public: 79 uint _cnt; Block_List()80 Block_List() : Block_Array(Thread::current()->resource_area()), _cnt(0) {} push(Block * b)81 void push( Block *b ) { map(_cnt++,b); } pop()82 Block *pop() { return _blocks[--_cnt]; } rpop()83 Block *rpop() { Block *b = _blocks[0]; _blocks[0]=_blocks[--_cnt]; return b;} 84 void remove( uint i ); 85 void insert( uint i, Block *n ); size() const86 uint size() const { return _cnt; } reset()87 void reset() { _cnt = 0; } 88 void print(); 89 }; 90 91 92 class CFGElement : public ResourceObj { 93 friend class VMStructs; 94 public: 95 double _freq; // Execution frequency (estimate) 96 CFGElement()97 CFGElement() : _freq(0.0) {} is_block()98 virtual bool is_block() { return false; } is_loop()99 virtual bool is_loop() { return false; } as_Block()100 Block* as_Block() { assert(is_block(), "must be block"); return (Block*)this; } as_CFGLoop()101 CFGLoop* as_CFGLoop() { assert(is_loop(), "must be loop"); return (CFGLoop*)this; } 102 }; 103 104 //------------------------------Block------------------------------------------ 105 // This class defines a Basic Block. 106 // Basic blocks are used during the output routines, and are not used during 107 // any optimization pass. They are created late in the game. 108 class Block : public CFGElement { 109 friend class VMStructs; 110 111 private: 112 // Nodes in this block, in order 113 Node_List _nodes; 114 115 public: 116 117 // Get the node at index 'at_index', if 'at_index' is out of bounds return NULL get_node(uint at_index) const118 Node* get_node(uint at_index) const { 119 return _nodes[at_index]; 120 } 121 122 // Get the number of nodes in this block number_of_nodes() const123 uint number_of_nodes() const { 124 return _nodes.size(); 125 } 126 127 // 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 map_node(Node * node,uint to_index)128 void map_node(Node* node, uint to_index) { 129 _nodes.map(to_index, node); 130 } 131 132 // 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 insert_node(Node * node,uint at_index)133 void insert_node(Node* node, uint at_index) { 134 _nodes.insert(at_index, node); 135 } 136 137 // Remove a node at index 'at_index' remove_node(uint at_index)138 void remove_node(uint at_index) { 139 _nodes.remove(at_index); 140 } 141 142 // Push a node 'node' onto the node list push_node(Node * node)143 void push_node(Node* node) { 144 _nodes.push(node); 145 } 146 147 // Pop the last node off the node list pop_node()148 Node* pop_node() { 149 return _nodes.pop(); 150 } 151 152 // Basic blocks have a Node which defines Control for all Nodes pinned in 153 // this block. This Node is a RegionNode. Exception-causing Nodes 154 // (division, subroutines) and Phi functions are always pinned. Later, 155 // every Node will get pinned to some block. head() const156 Node *head() const { return get_node(0); } 157 158 // CAUTION: num_preds() is ONE based, so that predecessor numbers match 159 // input edges to Regions and Phis. num_preds() const160 uint num_preds() const { return head()->req(); } pred(uint i) const161 Node *pred(uint i) const { return head()->in(i); } 162 163 // Array of successor blocks, same size as projs array 164 Block_Array _succs; 165 166 // Basic blocks have some number of Nodes which split control to all 167 // following blocks. These Nodes are always Projections. The field in 168 // the Projection and the block-ending Node determine which Block follows. 169 uint _num_succs; 170 171 // Basic blocks also carry all sorts of good old fashioned DFS information 172 // used to find loops, loop nesting depth, dominators, etc. 173 uint _pre_order; // Pre-order DFS number 174 175 // Dominator tree 176 uint _dom_depth; // Depth in dominator tree for fast LCA 177 Block* _idom; // Immediate dominator block 178 179 CFGLoop *_loop; // Loop to which this block belongs 180 uint _rpo; // Number in reverse post order walk 181 is_block()182 virtual bool is_block() { return true; } 183 float succ_prob(uint i); // return probability of i'th successor 184 int num_fall_throughs(); // How many fall-through candidate this block has 185 void update_uncommon_branch(Block* un); // Lower branch prob to uncommon code 186 bool succ_fall_through(uint i); // Is successor "i" is a fall-through candidate 187 Block* lone_fall_through(); // Return lone fall-through Block or null 188 189 Block* dom_lca(Block* that); // Compute LCA in dominator tree. 190 dominates(Block * that)191 bool dominates(Block* that) { 192 int dom_diff = this->_dom_depth - that->_dom_depth; 193 if (dom_diff > 0) return false; 194 for (; dom_diff < 0; dom_diff++) that = that->_idom; 195 return this == that; 196 } 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() const; 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; set_raise_LCA_mark(node_idx_t x)217 void set_raise_LCA_mark(node_idx_t x) { _raise_LCA_mark = x; } raise_LCA_mark() const218 node_idx_t raise_LCA_mark() const { return _raise_LCA_mark; } 219 node_idx_t _raise_LCA_visited; set_raise_LCA_visited(node_idx_t x)220 void set_raise_LCA_visited(node_idx_t x) { _raise_LCA_visited = x; } raise_LCA_visited() const221 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; first_inst_size() const225 uint first_inst_size() const { return _first_inst_size; } set_first_inst_size(uint s)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. alignment_padding(int current_offset)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; set_connector()258 void set_connector() { _connector = true; } is_connector() const259 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; set_loop_alignment(Block * loop_top)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 } loop_alignment() const273 uint loop_alignment() const { return _loop_alignment; } has_loop_alignment() const274 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. Block(Arena * a,Node * headnode)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 end_idx() const299 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. end() const312 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. add_inst(Node * n)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 whether 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 non_connector()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 has_successor(Block * b) const338 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 non_connector_successor(int i) const348 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 // Root of whole program 372 RootNode* _root; 373 374 // The block containing the root node 375 Block* _root_block; 376 377 // List of basic blocks that are created during CFG creation 378 Block_List _blocks; 379 380 // Count of basic blocks 381 uint _number_of_blocks; 382 383 // Arena for the blocks to be stored in 384 Arena* _block_arena; 385 386 // Info used for scheduling 387 PhaseChaitin* _regalloc; 388 389 // Register pressure heuristic used? 390 bool _scheduling_for_pressure; 391 392 // The matcher for this compilation 393 Matcher& _matcher; 394 395 // Map nodes to owning basic block 396 Block_Array _node_to_block_mapping; 397 398 // Loop from the root 399 CFGLoop* _root_loop; 400 401 // Outmost loop frequency 402 double _outer_loop_frequency; 403 404 // Per node latency estimation, valid only during GCM 405 GrowableArray<uint>* _node_latency; 406 407 // Build a proper looking cfg. Return count of basic blocks 408 uint build_cfg(); 409 410 // Build the dominator tree so that we know where we can move instructions 411 void build_dominator_tree(); 412 413 // Estimate block frequencies based on IfNode probabilities, so that we know where we want to move instructions 414 void estimate_block_frequency(); 415 416 // Global Code Motion. See Click's PLDI95 paper. Place Nodes in specific 417 // basic blocks; i.e. _node_to_block_mapping now maps _idx for all Nodes to some Block. 418 // Move nodes to ensure correctness from GVN and also try to move nodes out of loops. 419 void global_code_motion(); 420 421 // Schedule Nodes early in their basic blocks. 422 bool schedule_early(VectorSet &visited, Node_Stack &roots); 423 424 // For each node, find the latest block it can be scheduled into 425 // and then select the cheapest block between the latest and earliest 426 // block to place the node. 427 void schedule_late(VectorSet &visited, Node_Stack &stack); 428 429 // Compute the (backwards) latency of a node from a single use 430 int latency_from_use(Node *n, const Node *def, Node *use); 431 432 // Compute the (backwards) latency of a node from the uses of this instruction 433 void partial_latency_of_defs(Node *n); 434 435 // Compute the instruction global latency with a backwards walk 436 void compute_latencies_backwards(VectorSet &visited, Node_Stack &stack); 437 438 // Pick a block between early and late that is a cheaper alternative 439 // to late. Helper for schedule_late. 440 Block* hoist_to_cheaper_block(Block* LCA, Block* early, Node* self); 441 442 bool schedule_local(Block* block, GrowableArray<int>& ready_cnt, VectorSet& next_call, intptr_t* recacl_pressure_nodes); 443 void set_next_call(Block* block, Node* n, VectorSet& next_call); 444 void needed_for_next_call(Block* block, Node* this_call, VectorSet& next_call); 445 446 // Perform basic-block local scheduling 447 Node* select(Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, VectorSet& next_call, uint sched_slot, 448 intptr_t* recacl_pressure_nodes); 449 void adjust_register_pressure(Node* n, Block* block, intptr_t *recalc_pressure_nodes, bool finalize_mode); 450 451 // Schedule a call next in the block 452 uint sched_call(Block* block, uint node_cnt, Node_List& worklist, GrowableArray<int>& ready_cnt, MachCallNode* mcall, VectorSet& next_call); 453 454 // Cleanup if any code lands between a Call and his Catch 455 void call_catch_cleanup(Block* block); 456 457 Node* catch_cleanup_find_cloned_def(Block* use_blk, Node* def, Block* def_blk, int n_clone_idx); 458 void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, int n_clone_idx); 459 460 // Detect implicit-null-check opportunities. Basically, find NULL checks 461 // with suitable memory ops nearby. Use the memory op to do the NULL check. 462 // I can generate a memory op if there is not one nearby. 463 void implicit_null_check(Block* block, Node *proj, Node *val, int allowed_reasons); 464 465 // Perform a Depth First Search (DFS). 466 // Setup 'vertex' as DFS to vertex mapping. 467 // Setup 'semi' as vertex to DFS mapping. 468 // Set 'parent' to DFS parent. 469 uint do_DFS(Tarjan* tarjan, uint rpo_counter); 470 471 // Helper function to insert a node into a block 472 void schedule_node_into_block( Node *n, Block *b ); 473 474 void replace_block_proj_ctrl( Node *n ); 475 476 // Set the basic block for pinned Nodes 477 void schedule_pinned_nodes( VectorSet &visited ); 478 479 // I'll need a few machine-specific GotoNodes. Clone from this one. 480 // Used when building the CFG and creating end nodes for blocks. 481 MachNode* _goto; 482 483 Block* insert_anti_dependences(Block* LCA, Node* load, bool verify = false); verify_anti_dependences(Block * LCA,Node * load) const484 void verify_anti_dependences(Block* LCA, Node* load) const { 485 assert(LCA == get_block_for_node(load), "should already be scheduled"); 486 const_cast<PhaseCFG*>(this)->insert_anti_dependences(LCA, load, true); 487 } 488 489 bool move_to_next(Block* bx, uint b_index); 490 void move_to_end(Block* bx, uint b_index); 491 492 void insert_goto_at(uint block_no, uint succ_no); 493 494 // Check for NeverBranch at block end. This needs to become a GOTO to the 495 // true target. NeverBranch are treated as a conditional branch that always 496 // goes the same direction for most of the optimizer and are used to give a 497 // fake exit path to infinite loops. At this late stage they need to turn 498 // into Goto's so that when you enter the infinite loop you indeed hang. 499 void convert_NeverBranch_to_Goto(Block *b); 500 501 CFGLoop* create_loop_tree(); 502 bool is_dominator(Node* dom_node, Node* node); 503 bool is_CFG(Node* n); 504 bool is_control_proj_or_safepoint(Node* n) const; 505 Block* find_block_for_node(Node* n) const; 506 bool is_dominating_control(Node* dom_ctrl, Node* n); 507 #ifndef PRODUCT 508 bool _trace_opto_pipelining; // tracing flag 509 #endif 510 511 public: 512 PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher); 513 set_latency_for_node(Node * node,int latency)514 void set_latency_for_node(Node* node, int latency) { 515 _node_latency->at_put_grow(node->_idx, latency); 516 } 517 get_latency_for_node(Node * node)518 uint get_latency_for_node(Node* node) { 519 return _node_latency->at_grow(node->_idx); 520 } 521 522 // Get the outer most frequency get_outer_loop_frequency() const523 double get_outer_loop_frequency() const { 524 return _outer_loop_frequency; 525 } 526 527 // Get the root node of the CFG get_root_node() const528 RootNode* get_root_node() const { 529 return _root; 530 } 531 532 // Get the block of the root node get_root_block() const533 Block* get_root_block() const { 534 return _root_block; 535 } 536 537 // Add a block at a position and moves the later ones one step add_block_at(uint pos,Block * block)538 void add_block_at(uint pos, Block* block) { 539 _blocks.insert(pos, block); 540 _number_of_blocks++; 541 } 542 543 // Adds a block to the top of the block list add_block(Block * block)544 void add_block(Block* block) { 545 _blocks.push(block); 546 _number_of_blocks++; 547 } 548 549 // Clear the list of blocks clear_blocks()550 void clear_blocks() { 551 _blocks.reset(); 552 _number_of_blocks = 0; 553 } 554 555 // Get the block at position pos in _blocks get_block(uint pos) const556 Block* get_block(uint pos) const { 557 return _blocks[pos]; 558 } 559 560 // Number of blocks number_of_blocks() const561 uint number_of_blocks() const { 562 return _number_of_blocks; 563 } 564 565 // set which block this node should reside in map_node_to_block(const Node * node,Block * block)566 void map_node_to_block(const Node* node, Block* block) { 567 _node_to_block_mapping.map(node->_idx, block); 568 } 569 570 // removes the mapping from a node to a block unmap_node_from_block(const Node * node)571 void unmap_node_from_block(const Node* node) { 572 _node_to_block_mapping.map(node->_idx, NULL); 573 } 574 575 // get the block in which this node resides get_block_for_node(const Node * node) const576 Block* get_block_for_node(const Node* node) const { 577 return _node_to_block_mapping[node->_idx]; 578 } 579 580 // does this node reside in a block; return true has_block(const Node * node) const581 bool has_block(const Node* node) const { 582 return (_node_to_block_mapping.lookup(node->_idx) != NULL); 583 } 584 585 // Use frequency calculations and code shape to predict if the block 586 // is uncommon. 587 bool is_uncommon(const Block* block); 588 589 #ifdef ASSERT 590 Unique_Node_List _raw_oops; 591 #endif 592 593 // Do global code motion by first building dominator tree and estimate block frequency 594 // Returns true on success 595 bool do_global_code_motion(); 596 597 // Compute the (backwards) latency of a node from the uses 598 void latency_from_uses(Node *n); 599 600 // Set loop alignment 601 void set_loop_alignment(); 602 603 // Remove empty basic blocks 604 void remove_empty_blocks(); 605 Block *fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext); 606 void fixup_flow(); 607 608 // Insert a node into a block at index and map the node to the block insert(Block * b,uint idx,Node * n)609 void insert(Block *b, uint idx, Node *n) { 610 b->insert_node(n , idx); 611 map_node_to_block(n, b); 612 } 613 614 // Check all nodes and postalloc_expand them if necessary. 615 void postalloc_expand(PhaseRegAlloc* _ra); 616 617 #ifndef PRODUCT trace_opto_pipelining() const618 bool trace_opto_pipelining() const { return _trace_opto_pipelining; } 619 620 // Debugging print of CFG 621 void dump( ) const; // CFG only 622 void _dump_cfg( const Node *end, VectorSet &visited ) const; 623 void verify() const; 624 void dump_headers(); 625 #else trace_opto_pipelining() const626 bool trace_opto_pipelining() const { return false; } 627 #endif 628 }; 629 630 631 //------------------------------UnionFind-------------------------------------- 632 // Map Block indices to a block-index for a cfg-cover. 633 // Array lookup in the optimized case. 634 class UnionFind : public ResourceObj { 635 uint _cnt, _max; 636 uint* _indices; 637 ReallocMark _nesting; // assertion check for reallocations 638 public: 639 UnionFind( uint max ); 640 void reset( uint max ); // Reset to identity map for [0..max] 641 lookup(uint nidx) const642 uint lookup( uint nidx ) const { 643 return _indices[nidx]; 644 } operator [](uint nidx) const645 uint operator[] (uint nidx) const { return lookup(nidx); } 646 map(uint from_idx,uint to_idx)647 void map( uint from_idx, uint to_idx ) { 648 assert( from_idx < _cnt, "oob" ); 649 _indices[from_idx] = to_idx; 650 } 651 void extend( uint from_idx, uint to_idx ); 652 Size() const653 uint Size() const { return _cnt; } 654 Find(uint idx)655 uint Find( uint idx ) { 656 assert( idx < 65536, "Must fit into uint"); 657 uint uf_idx = lookup(idx); 658 return (uf_idx == idx) ? uf_idx : Find_compress(idx); 659 } 660 uint Find_compress( uint idx ); 661 uint Find_const( uint idx ) const; 662 void Union( uint idx1, uint idx2 ); 663 664 }; 665 666 //----------------------------BlockProbPair--------------------------- 667 // Ordered pair of Node*. 668 class BlockProbPair { 669 protected: 670 Block* _target; // block target 671 double _prob; // probability of edge to block 672 public: BlockProbPair()673 BlockProbPair() : _target(NULL), _prob(0.0) {} BlockProbPair(Block * b,double p)674 BlockProbPair(Block* b, double p) : _target(b), _prob(p) {} 675 get_target() const676 Block* get_target() const { return _target; } get_prob() const677 double get_prob() const { return _prob; } 678 }; 679 680 //------------------------------CFGLoop------------------------------------------- 681 class CFGLoop : public CFGElement { 682 friend class VMStructs; 683 int _id; 684 int _depth; 685 CFGLoop *_parent; // root of loop tree is the method level "pseudo" loop, it's parent is null 686 CFGLoop *_sibling; // null terminated list 687 CFGLoop *_child; // first child, use child's sibling to visit all immediately nested loops 688 GrowableArray<CFGElement*> _members; // list of members of loop 689 GrowableArray<BlockProbPair> _exits; // list of successor blocks and their probabilities 690 double _exit_prob; // probability any loop exit is taken on a single loop iteration 691 void update_succ_freq(Block* b, double freq); 692 693 public: CFGLoop(int id)694 CFGLoop(int id) : 695 CFGElement(), 696 _id(id), 697 _depth(0), 698 _parent(NULL), 699 _sibling(NULL), 700 _child(NULL), 701 _exit_prob(1.0f) {} parent()702 CFGLoop* parent() { return _parent; } 703 void push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg); add_member(CFGElement * s)704 void add_member(CFGElement *s) { _members.push(s); } 705 void add_nested_loop(CFGLoop* cl); head()706 Block* head() { 707 assert(_members.at(0)->is_block(), "head must be a block"); 708 Block* hd = _members.at(0)->as_Block(); 709 assert(hd->_loop == this, "just checking"); 710 assert(hd->head()->is_Loop(), "must begin with loop head node"); 711 return hd; 712 } 713 Block* backedge_block(); // Return the block on the backedge of the loop (else NULL) 714 void compute_loop_depth(int depth); 715 void compute_freq(); // compute frequency with loop assuming head freq 1.0f 716 void scale_freq(); // scale frequency by loop trip count (including outer loops) 717 double outer_loop_freq() const; // frequency of outer loop 718 bool in_loop_nest(Block* b); trip_count() const719 double trip_count() const { return 1.0 / _exit_prob; } is_loop()720 virtual bool is_loop() { return true; } id()721 int id() { return _id; } 722 723 #ifndef PRODUCT 724 void dump( ) const; 725 void dump_tree() const; 726 #endif 727 }; 728 729 730 //----------------------------------CFGEdge------------------------------------ 731 // A edge between two basic blocks that will be embodied by a branch or a 732 // fall-through. 733 class CFGEdge : public ResourceObj { 734 friend class VMStructs; 735 private: 736 Block * _from; // Source basic block 737 Block * _to; // Destination basic block 738 double _freq; // Execution frequency (estimate) 739 int _state; 740 bool _infrequent; 741 int _from_pct; 742 int _to_pct; 743 744 // Private accessors from_pct() const745 int from_pct() const { return _from_pct; } to_pct() const746 int to_pct() const { return _to_pct; } from_infrequent() const747 int from_infrequent() const { return from_pct() < BlockLayoutMinDiamondPercentage; } to_infrequent() const748 int to_infrequent() const { return to_pct() < BlockLayoutMinDiamondPercentage; } 749 750 public: 751 enum { 752 open, // initial edge state; unprocessed 753 connected, // edge used to connect two traces together 754 interior // edge is interior to trace (could be backedge) 755 }; 756 CFGEdge(Block * from,Block * to,double freq,int from_pct,int to_pct)757 CFGEdge(Block *from, Block *to, double freq, int from_pct, int to_pct) : 758 _from(from), _to(to), _freq(freq), 759 _state(open), _from_pct(from_pct), _to_pct(to_pct) { 760 _infrequent = from_infrequent() || to_infrequent(); 761 } 762 freq() const763 double freq() const { return _freq; } from() const764 Block* from() const { return _from; } to() const765 Block* to () const { return _to; } infrequent() const766 int infrequent() const { return _infrequent; } state() const767 int state() const { return _state; } 768 set_state(int state)769 void set_state(int state) { _state = state; } 770 771 #ifndef PRODUCT 772 void dump( ) const; 773 #endif 774 }; 775 776 777 //-----------------------------------Trace------------------------------------- 778 // An ordered list of basic blocks. 779 class Trace : public ResourceObj { 780 private: 781 uint _id; // Unique Trace id (derived from initial block) 782 Block ** _next_list; // Array mapping index to next block 783 Block ** _prev_list; // Array mapping index to previous block 784 Block * _first; // First block in the trace 785 Block * _last; // Last block in the trace 786 787 // Return the block that follows "b" in the trace. next(Block * b) const788 Block * next(Block *b) const { return _next_list[b->_pre_order]; } set_next(Block * b,Block * n) const789 void set_next(Block *b, Block *n) const { _next_list[b->_pre_order] = n; } 790 791 // Return the block that precedes "b" in the trace. prev(Block * b) const792 Block * prev(Block *b) const { return _prev_list[b->_pre_order]; } set_prev(Block * b,Block * p) const793 void set_prev(Block *b, Block *p) const { _prev_list[b->_pre_order] = p; } 794 795 // We've discovered a loop in this trace. Reset last to be "b", and first as 796 // the block following "b break_loop_after(Block * b)797 void break_loop_after(Block *b) { 798 _last = b; 799 _first = next(b); 800 set_prev(_first, NULL); 801 set_next(_last, NULL); 802 } 803 804 public: 805 Trace(Block * b,Block ** next_list,Block ** prev_list)806 Trace(Block *b, Block **next_list, Block **prev_list) : 807 _id(b->_pre_order), 808 _next_list(next_list), 809 _prev_list(prev_list), 810 _first(b), 811 _last(b) { 812 set_next(b, NULL); 813 set_prev(b, NULL); 814 }; 815 816 // Return the id number id() const817 uint id() const { return _id; } set_id(uint id)818 void set_id(uint id) { _id = id; } 819 820 // Return the first block in the trace first_block() const821 Block * first_block() const { return _first; } 822 823 // Return the last block in the trace last_block() const824 Block * last_block() const { return _last; } 825 826 // Insert a trace in the middle of this one after b insert_after(Block * b,Trace * tr)827 void insert_after(Block *b, Trace *tr) { 828 set_next(tr->last_block(), next(b)); 829 if (next(b) != NULL) { 830 set_prev(next(b), tr->last_block()); 831 } 832 833 set_next(b, tr->first_block()); 834 set_prev(tr->first_block(), b); 835 836 if (b == _last) { 837 _last = tr->last_block(); 838 } 839 } 840 insert_before(Block * b,Trace * tr)841 void insert_before(Block *b, Trace *tr) { 842 Block *p = prev(b); 843 assert(p != NULL, "use append instead"); 844 insert_after(p, tr); 845 } 846 847 // Append another trace to this one. append(Trace * tr)848 void append(Trace *tr) { 849 insert_after(_last, tr); 850 } 851 852 // Append a block at the end of this trace append(Block * b)853 void append(Block *b) { 854 set_next(_last, b); 855 set_prev(b, _last); 856 _last = b; 857 } 858 859 // Adjust the the blocks in this trace 860 void fixup_blocks(PhaseCFG &cfg); 861 bool backedge(CFGEdge *e); 862 863 #ifndef PRODUCT 864 void dump( ) const; 865 #endif 866 }; 867 868 //------------------------------PhaseBlockLayout------------------------------- 869 // Rearrange blocks into some canonical order, based on edges and their frequencies 870 class PhaseBlockLayout : public Phase { 871 friend class VMStructs; 872 PhaseCFG &_cfg; // Control flow graph 873 874 GrowableArray<CFGEdge *> *edges; 875 Trace **traces; 876 Block **next; 877 Block **prev; 878 UnionFind *uf; 879 880 // Given a block, find its encompassing Trace trace(Block * b)881 Trace * trace(Block *b) { 882 return traces[uf->Find_compress(b->_pre_order)]; 883 } 884 public: 885 PhaseBlockLayout(PhaseCFG &cfg); 886 887 void find_edges(); 888 void grow_traces(); 889 void merge_traces(bool loose_connections); 890 void reorder_traces(int count); 891 void union_traces(Trace* from, Trace* to); 892 }; 893 894 #endif // SHARE_OPTO_BLOCK_HPP 895