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_CHAITIN_HPP 26 #define SHARE_VM_OPTO_CHAITIN_HPP 27 28 #include "code/vmreg.hpp" 29 #include "libadt/port.hpp" 30 #include "memory/resourceArea.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/live.hpp" 33 #include "opto/matcher.hpp" 34 #include "opto/phase.hpp" 35 #include "opto/regalloc.hpp" 36 #include "opto/regmask.hpp" 37 #include "opto/machnode.hpp" 38 39 class LoopTree; 40 class Matcher; 41 class PhaseCFG; 42 class PhaseLive; 43 class PhaseRegAlloc; 44 class PhaseChaitin; 45 46 #define OPTO_DEBUG_SPLIT_FREQ BLOCK_FREQUENCY(0.001) 47 #define OPTO_LRG_HIGH_FREQ BLOCK_FREQUENCY(0.25) 48 49 //------------------------------LRG-------------------------------------------- 50 // Live-RanGe structure. 51 class LRG : public ResourceObj { 52 friend class VMStructs; 53 public: 54 static const uint AllStack_size = 0xFFFFF; // This mask size is used to tell that the mask of this LRG supports stack positions 55 enum { SPILL_REG=29999 }; // Register number of a spilled LRG 56 57 double _cost; // 2 for loads/1 for stores times block freq 58 double _area; // Sum of all simultaneously live values 59 double score() const; // Compute score from cost and area 60 double _maxfreq; // Maximum frequency of any def or use 61 62 Node *_def; // Check for multi-def live ranges 63 #ifndef PRODUCT 64 GrowableArray<Node*>* _defs; 65 #endif 66 67 uint _risk_bias; // Index of LRG which we want to avoid color 68 uint _copy_bias; // Index of LRG which we want to share color 69 70 uint _next; // Index of next LRG in linked list 71 uint _prev; // Index of prev LRG in linked list 72 private: 73 uint _reg; // Chosen register; undefined if mask is plural 74 public: 75 // Return chosen register for this LRG. Error if the LRG is not bound to 76 // a single register. 77 OptoReg::Name reg() const { return OptoReg::Name(_reg); } 78 void set_reg( OptoReg::Name r ) { _reg = r; } 79 80 private: 81 uint _eff_degree; // Effective degree: Sum of neighbors _num_regs 82 public: 83 int degree() const { assert( _degree_valid , "" ); return _eff_degree; } 84 // Degree starts not valid and any change to the IFG neighbor 85 // set makes it not valid. 86 void set_degree( uint degree ) { 87 _eff_degree = degree; 88 debug_only(_degree_valid = 1;) 89 assert(!_mask.is_AllStack() || (_mask.is_AllStack() && lo_degree()), "_eff_degree can't be bigger than AllStack_size - _num_regs if the mask supports stack registers"); 90 } 91 // Made a change that hammered degree 92 void invalid_degree() { debug_only(_degree_valid=0;) } 93 // Incrementally modify degree. If it was correct, it should remain correct 94 void inc_degree( uint mod ) { 95 _eff_degree += mod; 96 assert(!_mask.is_AllStack() || (_mask.is_AllStack() && lo_degree()), "_eff_degree can't be bigger than AllStack_size - _num_regs if the mask supports stack registers"); 97 } 98 // Compute the degree between 2 live ranges 99 int compute_degree( LRG &l ) const; 100 bool mask_is_nonempty_and_up() const { 101 return mask().is_UP() && mask_size(); 102 } 103 bool is_float_or_vector() const { 104 return _is_float || _is_vector; 105 } 106 107 private: 108 RegMask _mask; // Allowed registers for this LRG 109 uint _mask_size; // cache of _mask.Size(); 110 public: 111 int compute_mask_size() const { return _mask.is_AllStack() ? AllStack_size : _mask.Size(); } 112 void set_mask_size( int size ) { 113 assert((size == (int)AllStack_size) || (size == (int)_mask.Size()), ""); 114 _mask_size = size; 115 #ifdef ASSERT 116 _msize_valid=1; 117 if (_is_vector) { 118 assert(!_fat_proj, "sanity"); 119 _mask.verify_sets(_num_regs); 120 } else if (_num_regs == 2 && !_fat_proj) { 121 _mask.verify_pairs(); 122 } 123 #endif 124 } 125 void compute_set_mask_size() { set_mask_size(compute_mask_size()); } 126 int mask_size() const { assert( _msize_valid, "mask size not valid" ); 127 return _mask_size; } 128 // Get the last mask size computed, even if it does not match the 129 // count of bits in the current mask. 130 int get_invalid_mask_size() const { return _mask_size; } 131 const RegMask &mask() const { return _mask; } 132 void set_mask( const RegMask &rm ) { _mask = rm; debug_only(_msize_valid=0;)} 133 void AND( const RegMask &rm ) { _mask.AND(rm); debug_only(_msize_valid=0;)} 134 void SUBTRACT( const RegMask &rm ) { _mask.SUBTRACT(rm); debug_only(_msize_valid=0;)} 135 void Clear() { _mask.Clear() ; debug_only(_msize_valid=1); _mask_size = 0; } 136 void Set_All() { _mask.Set_All(); debug_only(_msize_valid=1); _mask_size = RegMask::CHUNK_SIZE; } 137 138 void Insert( OptoReg::Name reg ) { _mask.Insert(reg); debug_only(_msize_valid=0;) } 139 void Remove( OptoReg::Name reg ) { _mask.Remove(reg); debug_only(_msize_valid=0;) } 140 void clear_to_pairs() { _mask.clear_to_pairs(); debug_only(_msize_valid=0;) } 141 void clear_to_sets() { _mask.clear_to_sets(_num_regs); debug_only(_msize_valid=0;) } 142 143 // Number of registers this live range uses when it colors 144 private: 145 uint8 _num_regs; // 2 for Longs and Doubles, 1 for all else 146 // except _num_regs is kill count for fat_proj 147 public: 148 int num_regs() const { return _num_regs; } 149 void set_num_regs( int reg ) { assert( _num_regs == reg || !_num_regs, "" ); _num_regs = reg; } 150 151 private: 152 // Number of physical registers this live range uses when it colors 153 // Architecture and register-set dependent 154 uint8 _reg_pressure; 155 public: 156 void set_reg_pressure(int i) { _reg_pressure = i; } 157 int reg_pressure() const { return _reg_pressure; } 158 159 // How much 'wiggle room' does this live range have? 160 // How many color choices can it make (scaled by _num_regs)? 161 int degrees_of_freedom() const { return mask_size() - _num_regs; } 162 // Bound LRGs have ZERO degrees of freedom. We also count 163 // must_spill as bound. 164 bool is_bound () const { return _is_bound; } 165 // Negative degrees-of-freedom; even with no neighbors this 166 // live range must spill. 167 bool not_free() const { return degrees_of_freedom() < 0; } 168 // Is this live range of "low-degree"? Trivially colorable? 169 bool lo_degree () const { return degree() <= degrees_of_freedom(); } 170 // Is this live range just barely "low-degree"? Trivially colorable? 171 bool just_lo_degree () const { return degree() == degrees_of_freedom(); } 172 173 uint _is_oop:1, // Live-range holds an oop 174 _is_float:1, // True if in float registers 175 _is_vector:1, // True if in vector registers 176 _was_spilled1:1, // True if prior spilling on def 177 _was_spilled2:1, // True if twice prior spilling on def 178 _is_bound:1, // live range starts life with no 179 // degrees of freedom. 180 _direct_conflict:1, // True if def and use registers in conflict 181 _must_spill:1, // live range has lost all degrees of freedom 182 // If _fat_proj is set, live range does NOT require aligned, adjacent 183 // registers and has NO interferences. 184 // If _fat_proj is clear, live range requires num_regs() to be a power of 185 // 2, and it requires registers to form an aligned, adjacent set. 186 _fat_proj:1, // 187 _was_lo:1, // Was lo-degree prior to coalesce 188 _msize_valid:1, // _mask_size cache valid 189 _degree_valid:1, // _degree cache valid 190 _has_copy:1, // Adjacent to some copy instruction 191 _at_risk:1; // Simplify says this guy is at risk to spill 192 193 194 // Alive if non-zero, dead if zero 195 bool alive() const { return _def != NULL; } 196 bool is_multidef() const { return _def == NodeSentinel; } 197 bool is_singledef() const { return _def != NodeSentinel; } 198 199 #ifndef PRODUCT 200 void dump( ) const; 201 #endif 202 }; 203 204 //------------------------------IFG-------------------------------------------- 205 // InterFerence Graph 206 // An undirected graph implementation. Created with a fixed number of 207 // vertices. Edges can be added & tested. Vertices can be removed, then 208 // added back later with all edges intact. Can add edges between one vertex 209 // and a list of other vertices. Can union vertices (and their edges) 210 // together. The IFG needs to be really really fast, and also fairly 211 // abstract! It needs abstraction so I can fiddle with the implementation to 212 // get even more speed. 213 class PhaseIFG : public Phase { 214 friend class VMStructs; 215 // Current implementation: a triangular adjacency list. 216 217 // Array of adjacency-lists, indexed by live-range number 218 IndexSet *_adjs; 219 220 // Assertion bit for proper use of Squaring 221 bool _is_square; 222 223 // Live range structure goes here 224 LRG *_lrgs; // Array of LRG structures 225 226 public: 227 // Largest live-range number 228 uint _maxlrg; 229 230 Arena *_arena; 231 232 // Keep track of inserted and deleted Nodes 233 VectorSet *_yanked; 234 235 PhaseIFG( Arena *arena ); 236 void init( uint maxlrg ); 237 238 // Add edge between a and b. Returns true if actually addded. 239 int add_edge( uint a, uint b ); 240 241 // Add edge between a and everything in the vector 242 void add_vector( uint a, IndexSet *vec ); 243 244 // Test for edge existance 245 int test_edge( uint a, uint b ) const; 246 247 // Square-up matrix for faster Union 248 void SquareUp(); 249 250 // Return number of LRG neighbors 251 uint neighbor_cnt( uint a ) const { return _adjs[a].count(); } 252 // Union edges of b into a on Squared-up matrix 253 void Union( uint a, uint b ); 254 // Test for edge in Squared-up matrix 255 int test_edge_sq( uint a, uint b ) const; 256 // Yank a Node and all connected edges from the IFG. Be prepared to 257 // re-insert the yanked Node in reverse order of yanking. Return a 258 // list of neighbors (edges) yanked. 259 IndexSet *remove_node( uint a ); 260 // Reinsert a yanked Node 261 void re_insert( uint a ); 262 // Return set of neighbors 263 IndexSet *neighbors( uint a ) const { return &_adjs[a]; } 264 265 #ifndef PRODUCT 266 // Dump the IFG 267 void dump() const; 268 void stats() const; 269 void verify( const PhaseChaitin * ) const; 270 #endif 271 272 //--------------- Live Range Accessors 273 LRG &lrgs(uint idx) const { assert(idx < _maxlrg, "oob"); return _lrgs[idx]; } 274 275 // Compute and set effective degree. Might be folded into SquareUp(). 276 void Compute_Effective_Degree(); 277 278 // Compute effective degree as the sum of neighbors' _sizes. 279 int effective_degree( uint lidx ) const; 280 }; 281 282 // The LiveRangeMap class is responsible for storing node to live range id mapping. 283 // Each node is mapped to a live range id (a virtual register). Nodes that are 284 // not considered for register allocation are given live range id 0. 285 class LiveRangeMap VALUE_OBJ_CLASS_SPEC { 286 287 private: 288 289 uint _max_lrg_id; 290 291 // Union-find map. Declared as a short for speed. 292 // Indexed by live-range number, it returns the compacted live-range number 293 LRG_List _uf_map; 294 295 // Map from Nodes to live ranges 296 LRG_List _names; 297 298 // Straight out of Tarjan's union-find algorithm 299 uint find_compress(const Node *node) { 300 uint lrg_id = find_compress(_names.at(node->_idx)); 301 _names.at_put(node->_idx, lrg_id); 302 return lrg_id; 303 } 304 305 uint find_compress(uint lrg); 306 307 public: 308 309 const LRG_List& names() { 310 return _names; 311 } 312 313 uint max_lrg_id() const { 314 return _max_lrg_id; 315 } 316 317 void set_max_lrg_id(uint max_lrg_id) { 318 _max_lrg_id = max_lrg_id; 319 } 320 321 uint size() const { 322 return _names.length(); 323 } 324 325 uint live_range_id(uint idx) const { 326 return _names.at(idx); 327 } 328 329 uint live_range_id(const Node *node) const { 330 return _names.at(node->_idx); 331 } 332 333 uint uf_live_range_id(uint lrg_id) const { 334 return _uf_map.at(lrg_id); 335 } 336 337 void map(uint idx, uint lrg_id) { 338 _names.at_put(idx, lrg_id); 339 } 340 341 void uf_map(uint dst_lrg_id, uint src_lrg_id) { 342 _uf_map.at_put(dst_lrg_id, src_lrg_id); 343 } 344 345 void extend(uint idx, uint lrg_id) { 346 _names.at_put_grow(idx, lrg_id); 347 } 348 349 void uf_extend(uint dst_lrg_id, uint src_lrg_id) { 350 _uf_map.at_put_grow(dst_lrg_id, src_lrg_id); 351 } 352 353 LiveRangeMap(Arena* arena, uint unique) 354 : _names(arena, unique, unique, 0) 355 , _uf_map(arena, unique, unique, 0) 356 , _max_lrg_id(0) {} 357 358 uint find_id( const Node *n ) { 359 uint retval = live_range_id(n); 360 assert(retval == find(n),"Invalid node to lidx mapping"); 361 return retval; 362 } 363 364 // Reset the Union-Find map to identity 365 void reset_uf_map(uint max_lrg_id); 366 367 // Make all Nodes map directly to their final live range; no need for 368 // the Union-Find mapping after this call. 369 void compress_uf_map_for_nodes(); 370 371 uint find(uint lidx) { 372 uint uf_lidx = _uf_map.at(lidx); 373 return (uf_lidx == lidx) ? uf_lidx : find_compress(lidx); 374 } 375 376 // Convert a Node into a Live Range Index - a lidx 377 uint find(const Node *node) { 378 uint lidx = live_range_id(node); 379 uint uf_lidx = _uf_map.at(lidx); 380 return (uf_lidx == lidx) ? uf_lidx : find_compress(node); 381 } 382 383 // Like Find above, but no path compress, so bad asymptotic behavior 384 uint find_const(uint lrg) const; 385 386 // Like Find above, but no path compress, so bad asymptotic behavior 387 uint find_const(const Node *node) const { 388 if(node->_idx >= (uint)_names.length()) { 389 return 0; // not mapped, usual for debug dump 390 } 391 return find_const(_names.at(node->_idx)); 392 } 393 }; 394 395 //------------------------------Chaitin---------------------------------------- 396 // Briggs-Chaitin style allocation, mostly. 397 class PhaseChaitin : public PhaseRegAlloc { 398 friend class VMStructs; 399 400 int _trip_cnt; 401 int _alternate; 402 403 LRG &lrgs(uint idx) const { return _ifg->lrgs(idx); } 404 PhaseLive *_live; // Liveness, used in the interference graph 405 PhaseIFG *_ifg; // Interference graph (for original chunk) 406 Node_List **_lrg_nodes; // Array of node; lists for lrgs which spill 407 VectorSet _spilled_once; // Nodes that have been spilled 408 VectorSet _spilled_twice; // Nodes that have been spilled twice 409 410 // Combine the Live Range Indices for these 2 Nodes into a single live 411 // range. Future requests for any Node in either live range will 412 // return the live range index for the combined live range. 413 void Union( const Node *src, const Node *dst ); 414 415 void new_lrg( const Node *x, uint lrg ); 416 417 // Compact live ranges, removing unused ones. Return new maxlrg. 418 void compact(); 419 420 uint _lo_degree; // Head of lo-degree LRGs list 421 uint _lo_stk_degree; // Head of lo-stk-degree LRGs list 422 uint _hi_degree; // Head of hi-degree LRGs list 423 uint _simplified; // Linked list head of simplified LRGs 424 425 // Helper functions for Split() 426 uint split_DEF(Node *def, Block *b, int loc, uint max, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx ); 427 uint split_USE(MachSpillCopyNode::SpillType spill_type, Node *def, Block *b, Node *use, uint useidx, uint max, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx ); 428 429 //------------------------------clone_projs------------------------------------ 430 // After cloning some rematerialized instruction, clone any MachProj's that 431 // follow it. Example: Intel zero is XOR, kills flags. Sparc FP constants 432 // use G3 as an address temp. 433 int clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id); 434 435 int clone_projs(Block* b, uint idx, Node* orig, Node* copy, LiveRangeMap& lrg_map) { 436 uint max_lrg_id = lrg_map.max_lrg_id(); 437 int found_projs = clone_projs(b, idx, orig, copy, max_lrg_id); 438 if (found_projs > 0) { 439 // max_lrg_id is updated during call above 440 lrg_map.set_max_lrg_id(max_lrg_id); 441 } 442 return found_projs; 443 } 444 445 Node *split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits, 446 int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru); 447 // True if lidx is used before any real register is def'd in the block 448 bool prompt_use( Block *b, uint lidx ); 449 Node *get_spillcopy_wide(MachSpillCopyNode::SpillType spill_type, Node *def, Node *use, uint uidx ); 450 // Insert the spill at chosen location. Skip over any intervening Proj's or 451 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block 452 // instead. Update high-pressure indices. Create a new live range. 453 void insert_proj( Block *b, uint i, Node *spill, uint maxlrg ); 454 455 bool is_high_pressure( Block *b, LRG *lrg, uint insidx ); 456 457 uint _oldphi; // Node index which separates pre-allocation nodes 458 459 Block **_blks; // Array of blocks sorted by frequency for coalescing 460 461 float _high_frequency_lrg; // Frequency at which LRG will be spilled for debug info 462 463 #ifndef PRODUCT 464 bool _trace_spilling; 465 #endif 466 467 public: 468 PhaseChaitin( uint unique, PhaseCFG &cfg, Matcher &matcher ); 469 ~PhaseChaitin() {} 470 471 LiveRangeMap _lrg_map; 472 473 // Do all the real work of allocate 474 void Register_Allocate(); 475 476 float high_frequency_lrg() const { return _high_frequency_lrg; } 477 478 #ifndef PRODUCT 479 bool trace_spilling() const { return _trace_spilling; } 480 #endif 481 482 private: 483 // De-SSA the world. Assign registers to Nodes. Use the same register for 484 // all inputs to a PhiNode, effectively coalescing live ranges. Insert 485 // copies as needed. 486 void de_ssa(); 487 488 // Add edge between reg and everything in the vector. 489 // Same as _ifg->add_vector(reg,live) EXCEPT use the RegMask 490 // information to trim the set of interferences. Return the 491 // count of edges added. 492 void interfere_with_live(uint lid, IndexSet* liveout); 493 #ifdef ASSERT 494 // Count register pressure for asserts 495 uint count_int_pressure(IndexSet* liveout); 496 uint count_float_pressure(IndexSet* liveout); 497 #endif 498 499 // Build the interference graph using virtual registers only. 500 // Used for aggressive coalescing. 501 void build_ifg_virtual( ); 502 503 // used when computing the register pressure for each block in the CFG. This 504 // is done during IFG creation. 505 class Pressure { 506 // keeps track of the register pressure at the current 507 // instruction (used when stepping backwards in the block) 508 uint _current_pressure; 509 510 // keeps track of the instruction index of the first low to high register pressure 511 // transition (starting from the top) in the block 512 // if high_pressure_index == 0 then the whole block is high pressure 513 // if high_pressure_index = b.end_idx() + 1 then the whole block is low pressure 514 uint _high_pressure_index; 515 516 // stores the highest pressure we find 517 uint _final_pressure; 518 519 // number of live ranges that constitute high register pressure 520 const uint _high_pressure_limit; 521 public: 522 523 // lower the register pressure and look for a low to high pressure 524 // transition 525 void lower(LRG& lrg, uint& location) { 526 _current_pressure -= lrg.reg_pressure(); 527 if (_current_pressure == _high_pressure_limit) { 528 _high_pressure_index = location; 529 } 530 } 531 532 // raise the pressure and store the pressure if it's the biggest 533 // pressure so far 534 void raise(LRG &lrg) { 535 _current_pressure += lrg.reg_pressure(); 536 if (_current_pressure > _final_pressure) { 537 _final_pressure = _current_pressure; 538 } 539 } 540 541 uint high_pressure_index() const { 542 return _high_pressure_index; 543 } 544 545 uint final_pressure() const { 546 return _final_pressure; 547 } 548 549 uint current_pressure() const { 550 return _current_pressure; 551 } 552 553 uint high_pressure_limit() const { 554 return _high_pressure_limit; 555 } 556 557 void lower_high_pressure_index() { 558 _high_pressure_index--; 559 } 560 561 void set_high_pressure_index_to_block_start() { 562 _high_pressure_index = 0; 563 } 564 565 void check_pressure_at_fatproj(uint fatproj_location, RegMask& fatproj_mask) { 566 // this pressure is only valid at this instruction, i.e. we don't need to lower 567 // the register pressure since the fat proj was never live before (going backwards) 568 uint new_pressure = current_pressure() + fatproj_mask.Size(); 569 if (new_pressure > final_pressure()) { 570 _final_pressure = new_pressure; 571 } 572 573 // if we were at a low pressure and now and the fat proj is at high pressure, record the fat proj location 574 // as coming from a low to high (to low again) 575 if (current_pressure() <= high_pressure_limit() && new_pressure > high_pressure_limit()) { 576 _high_pressure_index = fatproj_location; 577 } 578 } 579 580 Pressure(uint high_pressure_index, uint high_pressure_limit) 581 : _current_pressure(0) 582 , _high_pressure_index(high_pressure_index) 583 , _high_pressure_limit(high_pressure_limit) 584 , _final_pressure(0) {} 585 }; 586 587 void lower_pressure(Block* b, uint location, LRG& lrg, IndexSet* liveout, Pressure& int_pressure, Pressure& float_pressure); 588 void raise_pressure(Block* b, LRG& lrg, Pressure& int_pressure, Pressure& float_pressure); 589 void check_for_high_pressure_transition_at_fatproj(uint& block_reg_pressure, uint location, LRG& lrg, Pressure& pressure, const int op_regtype); 590 void add_input_to_liveout(Block* b, Node* n, IndexSet* liveout, double cost, Pressure& int_pressure, Pressure& float_pressure); 591 void compute_initial_block_pressure(Block* b, IndexSet* liveout, Pressure& int_pressure, Pressure& float_pressure, double cost); 592 bool remove_node_if_not_used(Block* b, uint location, Node* n, uint lid, IndexSet* liveout); 593 void assign_high_score_to_immediate_copies(Block* b, Node* n, LRG& lrg, uint next_inst, uint last_inst); 594 void remove_interference_from_copy(Block* b, uint location, uint lid_copy, IndexSet* liveout, double cost, Pressure& int_pressure, Pressure& float_pressure); 595 void remove_bound_register_from_interfering_live_ranges(LRG& lrg, IndexSet* liveout, uint& must_spill); 596 void check_for_high_pressure_block(Pressure& pressure); 597 void adjust_high_pressure_index(Block* b, uint& hrp_index, Pressure& pressure); 598 599 // Build the interference graph using physical registers when available. 600 // That is, if 2 live ranges are simultaneously alive but in their 601 // acceptable register sets do not overlap, then they do not interfere. 602 uint build_ifg_physical( ResourceArea *a ); 603 604 // Gather LiveRanGe information, including register masks and base pointer/ 605 // derived pointer relationships. 606 void gather_lrg_masks( bool mod_cisc_masks ); 607 608 // Force the bases of derived pointers to be alive at GC points. 609 bool stretch_base_pointer_live_ranges( ResourceArea *a ); 610 // Helper to stretch above; recursively discover the base Node for 611 // a given derived Node. Easy for AddP-related machine nodes, but 612 // needs to be recursive for derived Phis. 613 Node *find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ); 614 615 // Set the was-lo-degree bit. Conservative coalescing should not change the 616 // colorability of the graph. If any live range was of low-degree before 617 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 618 void set_was_low(); 619 620 // Split live-ranges that must spill due to register conflicts (as opposed 621 // to capacity spills). Typically these are things def'd in a register 622 // and used on the stack or vice-versa. 623 void pre_spill(); 624 625 // Init LRG caching of degree, numregs. Init lo_degree list. 626 void cache_lrg_info( ); 627 628 // Simplify the IFG by removing LRGs of low degree with no copies 629 void Pre_Simplify(); 630 631 // Simplify the IFG by removing LRGs of low degree 632 void Simplify(); 633 634 // Select colors by re-inserting edges into the IFG. 635 // Return TRUE if any spills occurred. 636 uint Select( ); 637 // Helper function for select which allows biased coloring 638 OptoReg::Name choose_color( LRG &lrg, int chunk ); 639 // Helper function which implements biasing heuristic 640 OptoReg::Name bias_color( LRG &lrg, int chunk ); 641 642 // Split uncolorable live ranges 643 // Return new number of live ranges 644 uint Split(uint maxlrg, ResourceArea* split_arena); 645 646 // Copy 'was_spilled'-edness from one Node to another. 647 void copy_was_spilled( Node *src, Node *dst ); 648 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 649 void set_was_spilled( Node *n ); 650 651 // Convert ideal spill-nodes into machine loads & stores 652 // Set C->failing when fixup spills could not complete, node limit exceeded. 653 void fixup_spills(); 654 655 // Post-Allocation peephole copy removal 656 void post_allocate_copy_removal(); 657 Node *skip_copies( Node *c ); 658 // Replace the old node with the current live version of that value 659 // and yank the old value if it's dead. 660 int replace_and_yank_if_dead( Node *old, OptoReg::Name nreg, 661 Block *current_block, Node_List& value, Node_List& regnd ) { 662 Node* v = regnd[nreg]; 663 assert(v->outcnt() != 0, "no dead values"); 664 old->replace_by(v); 665 return yank_if_dead(old, current_block, &value, ®nd); 666 } 667 668 int yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { 669 return yank_if_dead_recurse(old, old, current_block, value, regnd); 670 } 671 int yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block, 672 Node_List *value, Node_List *regnd); 673 int yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ); 674 int elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ); 675 int use_prior_register( Node *copy, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ); 676 bool may_be_copy_of_callee( Node *def ) const; 677 678 // If nreg already contains the same constant as val then eliminate it 679 bool eliminate_copy_of_constant(Node* val, Node* n, 680 Block *current_block, Node_List& value, Node_List ®nd, 681 OptoReg::Name nreg, OptoReg::Name nreg2); 682 // Extend the node to LRG mapping 683 void add_reference( const Node *node, const Node *old_node); 684 685 private: 686 687 static int _final_loads, _final_stores, _final_copies, _final_memoves; 688 static double _final_load_cost, _final_store_cost, _final_copy_cost, _final_memove_cost; 689 static int _conserv_coalesce, _conserv_coalesce_pair; 690 static int _conserv_coalesce_trie, _conserv_coalesce_quad; 691 static int _post_alloc; 692 static int _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce; 693 static int _used_cisc_instructions, _unused_cisc_instructions; 694 static int _allocator_attempts, _allocator_successes; 695 696 #ifndef PRODUCT 697 static uint _high_pressure, _low_pressure; 698 699 void dump() const; 700 void dump( const Node *n ) const; 701 void dump( const Block * b ) const; 702 void dump_degree_lists() const; 703 void dump_simplified() const; 704 void dump_lrg( uint lidx, bool defs_only) const; 705 void dump_lrg( uint lidx) const { 706 // dump defs and uses by default 707 dump_lrg(lidx, false); 708 } 709 void dump_bb( uint pre_order ) const; 710 711 // Verify that base pointers and derived pointers are still sane 712 void verify_base_ptrs( ResourceArea *a ) const; 713 714 void verify( ResourceArea *a, bool verify_ifg = false ) const; 715 716 void dump_for_spill_split_recycle() const; 717 718 public: 719 void dump_frame() const; 720 char *dump_register( const Node *n, char *buf ) const; 721 private: 722 static void print_chaitin_statistics(); 723 #endif 724 friend class PhaseCoalesce; 725 friend class PhaseAggressiveCoalesce; 726 friend class PhaseConservativeCoalesce; 727 }; 728 729 #endif // SHARE_VM_OPTO_CHAITIN_HPP