annotate src/share/vm/opto/chaitin.hpp @ 23286:dd9cc155639c

Merge with jdk8u66-b17
author Gilles Duboscq <gilles.m.duboscq@oracle.com>
date Thu, 07 Jan 2016 17:28:46 +0100
parents 89152779163c 9df0d8f65fea
children
rev   line source
duke@0 1 /*
mikael@14223 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1552 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1552 20 * or visit www.oracle.com if you need additional information or have any
trims@1552 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1972 25 #ifndef SHARE_VM_OPTO_CHAITIN_HPP
stefank@1972 26 #define SHARE_VM_OPTO_CHAITIN_HPP
stefank@1972 27
stefank@1972 28 #include "code/vmreg.hpp"
stefank@1972 29 #include "libadt/port.hpp"
stefank@1972 30 #include "memory/resourceArea.hpp"
stefank@1972 31 #include "opto/connode.hpp"
stefank@1972 32 #include "opto/live.hpp"
stefank@1972 33 #include "opto/matcher.hpp"
stefank@1972 34 #include "opto/phase.hpp"
stefank@1972 35 #include "opto/regalloc.hpp"
stefank@1972 36 #include "opto/regmask.hpp"
stefank@1972 37
duke@0 38 class LoopTree;
duke@0 39 class MachCallNode;
duke@0 40 class MachSafePointNode;
duke@0 41 class Matcher;
duke@0 42 class PhaseCFG;
duke@0 43 class PhaseLive;
duke@0 44 class PhaseRegAlloc;
duke@0 45 class PhaseChaitin;
duke@0 46
duke@0 47 #define OPTO_DEBUG_SPLIT_FREQ BLOCK_FREQUENCY(0.001)
duke@0 48 #define OPTO_LRG_HIGH_FREQ BLOCK_FREQUENCY(0.25)
duke@0 49
duke@0 50 //------------------------------LRG--------------------------------------------
duke@0 51 // Live-RanGe structure.
duke@0 52 class LRG : public ResourceObj {
never@3939 53 friend class VMStructs;
duke@0 54 public:
adlertz@12877 55 static const uint AllStack_size = 0xFFFFF; // This mask size is used to tell that the mask of this LRG supports stack positions
duke@0 56 enum { SPILL_REG=29999 }; // Register number of a spilled LRG
duke@0 57
duke@0 58 double _cost; // 2 for loads/1 for stores times block freq
duke@0 59 double _area; // Sum of all simultaneously live values
duke@0 60 double score() const; // Compute score from cost and area
duke@0 61 double _maxfreq; // Maximum frequency of any def or use
duke@0 62
duke@0 63 Node *_def; // Check for multi-def live ranges
duke@0 64 #ifndef PRODUCT
duke@0 65 GrowableArray<Node*>* _defs;
duke@0 66 #endif
duke@0 67
duke@0 68 uint _risk_bias; // Index of LRG which we want to avoid color
duke@0 69 uint _copy_bias; // Index of LRG which we want to share color
duke@0 70
duke@0 71 uint _next; // Index of next LRG in linked list
duke@0 72 uint _prev; // Index of prev LRG in linked list
duke@0 73 private:
duke@0 74 uint _reg; // Chosen register; undefined if mask is plural
duke@0 75 public:
duke@0 76 // Return chosen register for this LRG. Error if the LRG is not bound to
duke@0 77 // a single register.
duke@0 78 OptoReg::Name reg() const { return OptoReg::Name(_reg); }
duke@0 79 void set_reg( OptoReg::Name r ) { _reg = r; }
duke@0 80
duke@0 81 private:
duke@0 82 uint _eff_degree; // Effective degree: Sum of neighbors _num_regs
duke@0 83 public:
adlertz@12877 84 int degree() const { assert( _degree_valid , "" ); return _eff_degree; }
duke@0 85 // Degree starts not valid and any change to the IFG neighbor
duke@0 86 // set makes it not valid.
adlertz@12877 87 void set_degree( uint degree ) {
adlertz@12877 88 _eff_degree = degree;
adlertz@12877 89 debug_only(_degree_valid = 1;)
adlertz@12877 90 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");
adlertz@12877 91 }
duke@0 92 // Made a change that hammered degree
duke@0 93 void invalid_degree() { debug_only(_degree_valid=0;) }
duke@0 94 // Incrementally modify degree. If it was correct, it should remain correct
adlertz@12877 95 void inc_degree( uint mod ) {
adlertz@12877 96 _eff_degree += mod;
adlertz@12877 97 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");
adlertz@12877 98 }
duke@0 99 // Compute the degree between 2 live ranges
duke@0 100 int compute_degree( LRG &l ) const;
duke@0 101
duke@0 102 private:
duke@0 103 RegMask _mask; // Allowed registers for this LRG
duke@0 104 uint _mask_size; // cache of _mask.Size();
duke@0 105 public:
adlertz@12877 106 int compute_mask_size() const { return _mask.is_AllStack() ? AllStack_size : _mask.Size(); }
duke@0 107 void set_mask_size( int size ) {
adlertz@12877 108 assert((size == (int)AllStack_size) || (size == (int)_mask.Size()), "");
duke@0 109 _mask_size = size;
kvn@6179 110 #ifdef ASSERT
kvn@6179 111 _msize_valid=1;
kvn@6179 112 if (_is_vector) {
kvn@6179 113 assert(!_fat_proj, "sanity");
kvn@6179 114 _mask.verify_sets(_num_regs);
kvn@6179 115 } else if (_num_regs == 2 && !_fat_proj) {
kvn@6179 116 _mask.verify_pairs();
kvn@6179 117 }
kvn@6179 118 #endif
duke@0 119 }
duke@0 120 void compute_set_mask_size() { set_mask_size(compute_mask_size()); }
duke@0 121 int mask_size() const { assert( _msize_valid, "mask size not valid" );
duke@0 122 return _mask_size; }
duke@0 123 // Get the last mask size computed, even if it does not match the
duke@0 124 // count of bits in the current mask.
duke@0 125 int get_invalid_mask_size() const { return _mask_size; }
duke@0 126 const RegMask &mask() const { return _mask; }
duke@0 127 void set_mask( const RegMask &rm ) { _mask = rm; debug_only(_msize_valid=0;)}
duke@0 128 void AND( const RegMask &rm ) { _mask.AND(rm); debug_only(_msize_valid=0;)}
duke@0 129 void SUBTRACT( const RegMask &rm ) { _mask.SUBTRACT(rm); debug_only(_msize_valid=0;)}
duke@0 130 void Clear() { _mask.Clear() ; debug_only(_msize_valid=1); _mask_size = 0; }
duke@0 131 void Set_All() { _mask.Set_All(); debug_only(_msize_valid=1); _mask_size = RegMask::CHUNK_SIZE; }
duke@0 132 void Insert( OptoReg::Name reg ) { _mask.Insert(reg); debug_only(_msize_valid=0;) }
duke@0 133 void Remove( OptoReg::Name reg ) { _mask.Remove(reg); debug_only(_msize_valid=0;) }
kvn@6179 134 void clear_to_pairs() { _mask.clear_to_pairs(); debug_only(_msize_valid=0;) }
kvn@6179 135 void clear_to_sets() { _mask.clear_to_sets(_num_regs); debug_only(_msize_valid=0;) }
duke@0 136
duke@0 137 // Number of registers this live range uses when it colors
duke@0 138 private:
duke@0 139 uint8 _num_regs; // 2 for Longs and Doubles, 1 for all else
duke@0 140 // except _num_regs is kill count for fat_proj
duke@0 141 public:
duke@0 142 int num_regs() const { return _num_regs; }
duke@0 143 void set_num_regs( int reg ) { assert( _num_regs == reg || !_num_regs, "" ); _num_regs = reg; }
duke@0 144
duke@0 145 private:
duke@0 146 // Number of physical registers this live range uses when it colors
duke@0 147 // Architecture and register-set dependent
duke@0 148 uint8 _reg_pressure;
duke@0 149 public:
duke@0 150 void set_reg_pressure(int i) { _reg_pressure = i; }
duke@0 151 int reg_pressure() const { return _reg_pressure; }
duke@0 152
duke@0 153 // How much 'wiggle room' does this live range have?
duke@0 154 // How many color choices can it make (scaled by _num_regs)?
duke@0 155 int degrees_of_freedom() const { return mask_size() - _num_regs; }
duke@0 156 // Bound LRGs have ZERO degrees of freedom. We also count
duke@0 157 // must_spill as bound.
duke@0 158 bool is_bound () const { return _is_bound; }
duke@0 159 // Negative degrees-of-freedom; even with no neighbors this
duke@0 160 // live range must spill.
duke@0 161 bool not_free() const { return degrees_of_freedom() < 0; }
duke@0 162 // Is this live range of "low-degree"? Trivially colorable?
duke@0 163 bool lo_degree () const { return degree() <= degrees_of_freedom(); }
duke@0 164 // Is this live range just barely "low-degree"? Trivially colorable?
duke@0 165 bool just_lo_degree () const { return degree() == degrees_of_freedom(); }
duke@0 166
duke@0 167 uint _is_oop:1, // Live-range holds an oop
duke@0 168 _is_float:1, // True if in float registers
kvn@6179 169 _is_vector:1, // True if in vector registers
duke@0 170 _was_spilled1:1, // True if prior spilling on def
duke@0 171 _was_spilled2:1, // True if twice prior spilling on def
duke@0 172 _is_bound:1, // live range starts life with no
duke@0 173 // degrees of freedom.
duke@0 174 _direct_conflict:1, // True if def and use registers in conflict
duke@0 175 _must_spill:1, // live range has lost all degrees of freedom
duke@0 176 // If _fat_proj is set, live range does NOT require aligned, adjacent
duke@0 177 // registers and has NO interferences.
duke@0 178 // If _fat_proj is clear, live range requires num_regs() to be a power of
duke@0 179 // 2, and it requires registers to form an aligned, adjacent set.
duke@0 180 _fat_proj:1, //
duke@0 181 _was_lo:1, // Was lo-degree prior to coalesce
duke@0 182 _msize_valid:1, // _mask_size cache valid
duke@0 183 _degree_valid:1, // _degree cache valid
duke@0 184 _has_copy:1, // Adjacent to some copy instruction
duke@0 185 _at_risk:1; // Simplify says this guy is at risk to spill
duke@0 186
duke@0 187
duke@0 188 // Alive if non-zero, dead if zero
duke@0 189 bool alive() const { return _def != NULL; }
never@295 190 bool is_multidef() const { return _def == NodeSentinel; }
never@295 191 bool is_singledef() const { return _def != NodeSentinel; }
duke@0 192
duke@0 193 #ifndef PRODUCT
duke@0 194 void dump( ) const;
duke@0 195 #endif
duke@0 196 };
duke@0 197
duke@0 198 //------------------------------IFG--------------------------------------------
duke@0 199 // InterFerence Graph
duke@0 200 // An undirected graph implementation. Created with a fixed number of
duke@0 201 // vertices. Edges can be added & tested. Vertices can be removed, then
duke@0 202 // added back later with all edges intact. Can add edges between one vertex
duke@0 203 // and a list of other vertices. Can union vertices (and their edges)
duke@0 204 // together. The IFG needs to be really really fast, and also fairly
duke@0 205 // abstract! It needs abstraction so I can fiddle with the implementation to
duke@0 206 // get even more speed.
duke@0 207 class PhaseIFG : public Phase {
never@3939 208 friend class VMStructs;
duke@0 209 // Current implementation: a triangular adjacency list.
duke@0 210
duke@0 211 // Array of adjacency-lists, indexed by live-range number
duke@0 212 IndexSet *_adjs;
duke@0 213
duke@0 214 // Assertion bit for proper use of Squaring
duke@0 215 bool _is_square;
duke@0 216
duke@0 217 // Live range structure goes here
duke@0 218 LRG *_lrgs; // Array of LRG structures
duke@0 219
duke@0 220 public:
duke@0 221 // Largest live-range number
duke@0 222 uint _maxlrg;
duke@0 223
duke@0 224 Arena *_arena;
duke@0 225
duke@0 226 // Keep track of inserted and deleted Nodes
duke@0 227 VectorSet *_yanked;
duke@0 228
duke@0 229 PhaseIFG( Arena *arena );
duke@0 230 void init( uint maxlrg );
duke@0 231
duke@0 232 // Add edge between a and b. Returns true if actually addded.
duke@0 233 int add_edge( uint a, uint b );
duke@0 234
duke@0 235 // Add edge between a and everything in the vector
duke@0 236 void add_vector( uint a, IndexSet *vec );
duke@0 237
duke@0 238 // Test for edge existance
duke@0 239 int test_edge( uint a, uint b ) const;
duke@0 240
duke@0 241 // Square-up matrix for faster Union
duke@0 242 void SquareUp();
duke@0 243
duke@0 244 // Return number of LRG neighbors
duke@0 245 uint neighbor_cnt( uint a ) const { return _adjs[a].count(); }
duke@0 246 // Union edges of b into a on Squared-up matrix
duke@0 247 void Union( uint a, uint b );
duke@0 248 // Test for edge in Squared-up matrix
duke@0 249 int test_edge_sq( uint a, uint b ) const;
duke@0 250 // Yank a Node and all connected edges from the IFG. Be prepared to
duke@0 251 // re-insert the yanked Node in reverse order of yanking. Return a
duke@0 252 // list of neighbors (edges) yanked.
duke@0 253 IndexSet *remove_node( uint a );
duke@0 254 // Reinsert a yanked Node
duke@0 255 void re_insert( uint a );
duke@0 256 // Return set of neighbors
duke@0 257 IndexSet *neighbors( uint a ) const { return &_adjs[a]; }
duke@0 258
duke@0 259 #ifndef PRODUCT
duke@0 260 // Dump the IFG
duke@0 261 void dump() const;
duke@0 262 void stats() const;
duke@0 263 void verify( const PhaseChaitin * ) const;
duke@0 264 #endif
duke@0 265
duke@0 266 //--------------- Live Range Accessors
duke@0 267 LRG &lrgs(uint idx) const { assert(idx < _maxlrg, "oob"); return _lrgs[idx]; }
duke@0 268
duke@0 269 // Compute and set effective degree. Might be folded into SquareUp().
duke@0 270 void Compute_Effective_Degree();
duke@0 271
duke@0 272 // Compute effective degree as the sum of neighbors' _sizes.
duke@0 273 int effective_degree( uint lidx ) const;
duke@0 274 };
duke@0 275
neliasso@10111 276 // The LiveRangeMap class is responsible for storing node to live range id mapping.
neliasso@10111 277 // Each node is mapped to a live range id (a virtual register). Nodes that are
neliasso@10111 278 // not considered for register allocation are given live range id 0.
neliasso@10111 279 class LiveRangeMap VALUE_OBJ_CLASS_SPEC {
duke@0 280
neliasso@10111 281 private:
neliasso@10111 282
neliasso@10111 283 uint _max_lrg_id;
neliasso@10111 284
neliasso@10111 285 // Union-find map. Declared as a short for speed.
neliasso@10111 286 // Indexed by live-range number, it returns the compacted live-range number
neliasso@10111 287 LRG_List _uf_map;
neliasso@10111 288
neliasso@10111 289 // Map from Nodes to live ranges
neliasso@10111 290 LRG_List _names;
neliasso@10111 291
neliasso@10111 292 // Straight out of Tarjan's union-find algorithm
neliasso@10111 293 uint find_compress(const Node *node) {
adlertz@12254 294 uint lrg_id = find_compress(_names.at(node->_idx));
adlertz@12254 295 _names.at_put(node->_idx, lrg_id);
neliasso@10111 296 return lrg_id;
neliasso@10111 297 }
neliasso@10111 298
neliasso@10111 299 uint find_compress(uint lrg);
neliasso@10111 300
neliasso@10111 301 public:
neliasso@10111 302
neliasso@10111 303 const LRG_List& names() {
neliasso@10111 304 return _names;
neliasso@10111 305 }
neliasso@10111 306
neliasso@10111 307 uint max_lrg_id() const {
neliasso@10111 308 return _max_lrg_id;
neliasso@10111 309 }
neliasso@10111 310
neliasso@10111 311 void set_max_lrg_id(uint max_lrg_id) {
neliasso@10111 312 _max_lrg_id = max_lrg_id;
neliasso@10111 313 }
neliasso@10111 314
neliasso@10111 315 uint size() const {
adlertz@12254 316 return _names.length();
neliasso@10111 317 }
neliasso@10111 318
neliasso@10111 319 uint live_range_id(uint idx) const {
adlertz@12254 320 return _names.at(idx);
neliasso@10111 321 }
neliasso@10111 322
neliasso@10111 323 uint live_range_id(const Node *node) const {
adlertz@12254 324 return _names.at(node->_idx);
neliasso@10111 325 }
neliasso@10111 326
neliasso@10111 327 uint uf_live_range_id(uint lrg_id) const {
adlertz@12254 328 return _uf_map.at(lrg_id);
neliasso@10111 329 }
neliasso@10111 330
neliasso@10111 331 void map(uint idx, uint lrg_id) {
adlertz@12254 332 _names.at_put(idx, lrg_id);
neliasso@10111 333 }
neliasso@10111 334
neliasso@10111 335 void uf_map(uint dst_lrg_id, uint src_lrg_id) {
adlertz@12254 336 _uf_map.at_put(dst_lrg_id, src_lrg_id);
neliasso@10111 337 }
neliasso@10111 338
neliasso@10111 339 void extend(uint idx, uint lrg_id) {
adlertz@12254 340 _names.at_put_grow(idx, lrg_id);
neliasso@10111 341 }
neliasso@10111 342
neliasso@10111 343 void uf_extend(uint dst_lrg_id, uint src_lrg_id) {
adlertz@12254 344 _uf_map.at_put_grow(dst_lrg_id, src_lrg_id);
neliasso@10111 345 }
neliasso@10111 346
adlertz@12254 347 LiveRangeMap(Arena* arena, uint unique)
adlertz@12254 348 : _names(arena, unique, unique, 0)
adlertz@12254 349 , _uf_map(arena, unique, unique, 0)
neliasso@10111 350 , _max_lrg_id(0) {}
neliasso@10111 351
neliasso@10111 352 uint find_id( const Node *n ) {
neliasso@10111 353 uint retval = live_range_id(n);
neliasso@10111 354 assert(retval == find(n),"Invalid node to lidx mapping");
neliasso@10111 355 return retval;
neliasso@10111 356 }
neliasso@10111 357
neliasso@10111 358 // Reset the Union-Find map to identity
neliasso@10111 359 void reset_uf_map(uint max_lrg_id);
neliasso@10111 360
neliasso@10111 361 // Make all Nodes map directly to their final live range; no need for
neliasso@10111 362 // the Union-Find mapping after this call.
neliasso@10111 363 void compress_uf_map_for_nodes();
neliasso@10111 364
neliasso@10111 365 uint find(uint lidx) {
adlertz@12254 366 uint uf_lidx = _uf_map.at(lidx);
neliasso@10111 367 return (uf_lidx == lidx) ? uf_lidx : find_compress(lidx);
neliasso@10111 368 }
neliasso@10111 369
neliasso@10111 370 // Convert a Node into a Live Range Index - a lidx
neliasso@10111 371 uint find(const Node *node) {
neliasso@10111 372 uint lidx = live_range_id(node);
adlertz@12254 373 uint uf_lidx = _uf_map.at(lidx);
neliasso@10111 374 return (uf_lidx == lidx) ? uf_lidx : find_compress(node);
neliasso@10111 375 }
neliasso@10111 376
neliasso@10111 377 // Like Find above, but no path compress, so bad asymptotic behavior
neliasso@10111 378 uint find_const(uint lrg) const;
neliasso@10111 379
neliasso@10111 380 // Like Find above, but no path compress, so bad asymptotic behavior
neliasso@10111 381 uint find_const(const Node *node) const {
adlertz@12254 382 if(node->_idx >= (uint)_names.length()) {
neliasso@10111 383 return 0; // not mapped, usual for debug dump
neliasso@10111 384 }
adlertz@12254 385 return find_const(_names.at(node->_idx));
neliasso@10111 386 }
neliasso@10111 387 };
duke@0 388
duke@0 389 //------------------------------Chaitin----------------------------------------
duke@0 390 // Briggs-Chaitin style allocation, mostly.
duke@0 391 class PhaseChaitin : public PhaseRegAlloc {
never@3939 392 friend class VMStructs;
duke@0 393
duke@0 394 int _trip_cnt;
duke@0 395 int _alternate;
duke@0 396
duke@0 397 LRG &lrgs(uint idx) const { return _ifg->lrgs(idx); }
duke@0 398 PhaseLive *_live; // Liveness, used in the interference graph
duke@0 399 PhaseIFG *_ifg; // Interference graph (for original chunk)
duke@0 400 Node_List **_lrg_nodes; // Array of node; lists for lrgs which spill
duke@0 401 VectorSet _spilled_once; // Nodes that have been spilled
duke@0 402 VectorSet _spilled_twice; // Nodes that have been spilled twice
duke@0 403
duke@0 404 // Combine the Live Range Indices for these 2 Nodes into a single live
duke@0 405 // range. Future requests for any Node in either live range will
duke@0 406 // return the live range index for the combined live range.
duke@0 407 void Union( const Node *src, const Node *dst );
duke@0 408
duke@0 409 void new_lrg( const Node *x, uint lrg );
duke@0 410
duke@0 411 // Compact live ranges, removing unused ones. Return new maxlrg.
duke@0 412 void compact();
duke@0 413
duke@0 414 uint _lo_degree; // Head of lo-degree LRGs list
duke@0 415 uint _lo_stk_degree; // Head of lo-stk-degree LRGs list
duke@0 416 uint _hi_degree; // Head of hi-degree LRGs list
duke@0 417 uint _simplified; // Linked list head of simplified LRGs
duke@0 418
duke@0 419 // Helper functions for Split()
duke@0 420 uint split_DEF( Node *def, Block *b, int loc, uint max, Node **Reachblock, Node **debug_defs, GrowableArray<uint> splits, int slidx );
duke@0 421 uint split_USE( Node *def, Block *b, Node *use, uint useidx, uint max, bool def_down, bool cisc_sp, GrowableArray<uint> splits, int slidx );
neliasso@10111 422
neliasso@10111 423 //------------------------------clone_projs------------------------------------
neliasso@10111 424 // After cloning some rematerialized instruction, clone any MachProj's that
neliasso@10111 425 // follow it. Example: Intel zero is XOR, kills flags. Sparc FP constants
neliasso@10111 426 // use G3 as an address temp.
kvn@12075 427 int clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id);
neliasso@10111 428
kvn@12075 429 int clone_projs(Block* b, uint idx, Node* orig, Node* copy, LiveRangeMap& lrg_map) {
kvn@12075 430 uint max_lrg_id = lrg_map.max_lrg_id();
kvn@12075 431 int found_projs = clone_projs(b, idx, orig, copy, max_lrg_id);
kvn@12075 432 if (found_projs > 0) {
kvn@12075 433 // max_lrg_id is updated during call above
kvn@12075 434 lrg_map.set_max_lrg_id(max_lrg_id);
neliasso@10111 435 }
neliasso@10111 436 return found_projs;
neliasso@10111 437 }
neliasso@10111 438
never@295 439 Node *split_Rematerialize(Node *def, Block *b, uint insidx, uint &maxlrg, GrowableArray<uint> splits,
never@295 440 int slidx, uint *lrg2reach, Node **Reachblock, bool walkThru);
duke@0 441 // True if lidx is used before any real register is def'd in the block
duke@0 442 bool prompt_use( Block *b, uint lidx );
duke@0 443 Node *get_spillcopy_wide( Node *def, Node *use, uint uidx );
twisti@605 444 // Insert the spill at chosen location. Skip over any intervening Proj's or
duke@0 445 // Phis. Skip over a CatchNode and projs, inserting in the fall-through block
duke@0 446 // instead. Update high-pressure indices. Create a new live range.
duke@0 447 void insert_proj( Block *b, uint i, Node *spill, uint maxlrg );
duke@0 448
duke@0 449 bool is_high_pressure( Block *b, LRG *lrg, uint insidx );
duke@0 450
duke@0 451 uint _oldphi; // Node index which separates pre-allocation nodes
duke@0 452
duke@0 453 Block **_blks; // Array of blocks sorted by frequency for coalescing
duke@0 454
kvn@673 455 float _high_frequency_lrg; // Frequency at which LRG will be spilled for debug info
kvn@673 456
duke@0 457 #ifndef PRODUCT
duke@0 458 bool _trace_spilling;
duke@0 459 #endif
duke@0 460
duke@0 461 public:
duke@0 462 PhaseChaitin( uint unique, PhaseCFG &cfg, Matcher &matcher );
duke@0 463 ~PhaseChaitin() {}
duke@0 464
neliasso@10111 465 LiveRangeMap _lrg_map;
duke@0 466
duke@0 467 // Do all the real work of allocate
duke@0 468 void Register_Allocate();
duke@0 469
kvn@673 470 float high_frequency_lrg() const { return _high_frequency_lrg; }
kvn@673 471
duke@0 472 #ifndef PRODUCT
duke@0 473 bool trace_spilling() const { return _trace_spilling; }
duke@0 474 #endif
duke@0 475
duke@0 476 private:
duke@0 477 // De-SSA the world. Assign registers to Nodes. Use the same register for
duke@0 478 // all inputs to a PhiNode, effectively coalescing live ranges. Insert
duke@0 479 // copies as needed.
duke@0 480 void de_ssa();
duke@0 481
duke@0 482 // Add edge between reg and everything in the vector.
duke@0 483 // Same as _ifg->add_vector(reg,live) EXCEPT use the RegMask
duke@0 484 // information to trim the set of interferences. Return the
duke@0 485 // count of edges added.
duke@0 486 void interfere_with_live( uint reg, IndexSet *live );
duke@0 487 // Count register pressure for asserts
duke@0 488 uint count_int_pressure( IndexSet *liveout );
duke@0 489 uint count_float_pressure( IndexSet *liveout );
duke@0 490
duke@0 491 // Build the interference graph using virtual registers only.
duke@0 492 // Used for aggressive coalescing.
duke@0 493 void build_ifg_virtual( );
duke@0 494
duke@0 495 // Build the interference graph using physical registers when available.
duke@0 496 // That is, if 2 live ranges are simultaneously alive but in their
duke@0 497 // acceptable register sets do not overlap, then they do not interfere.
duke@0 498 uint build_ifg_physical( ResourceArea *a );
duke@0 499
duke@0 500 // Gather LiveRanGe information, including register masks and base pointer/
duke@0 501 // derived pointer relationships.
duke@0 502 void gather_lrg_masks( bool mod_cisc_masks );
duke@0 503
duke@0 504 // Force the bases of derived pointers to be alive at GC points.
duke@0 505 bool stretch_base_pointer_live_ranges( ResourceArea *a );
duke@0 506 // Helper to stretch above; recursively discover the base Node for
duke@0 507 // a given derived Node. Easy for AddP-related machine nodes, but
duke@0 508 // needs to be recursive for derived Phis.
duke@0 509 Node *find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg );
duke@0 510
duke@0 511 // Set the was-lo-degree bit. Conservative coalescing should not change the
duke@0 512 // colorability of the graph. If any live range was of low-degree before
duke@0 513 // coalescing, it should Simplify. This call sets the was-lo-degree bit.
duke@0 514 void set_was_low();
duke@0 515
duke@0 516 // Split live-ranges that must spill due to register conflicts (as opposed
duke@0 517 // to capacity spills). Typically these are things def'd in a register
duke@0 518 // and used on the stack or vice-versa.
duke@0 519 void pre_spill();
duke@0 520
duke@0 521 // Init LRG caching of degree, numregs. Init lo_degree list.
duke@0 522 void cache_lrg_info( );
duke@0 523
duke@0 524 // Simplify the IFG by removing LRGs of low degree with no copies
duke@0 525 void Pre_Simplify();
duke@0 526
duke@0 527 // Simplify the IFG by removing LRGs of low degree
duke@0 528 void Simplify();
duke@0 529
duke@0 530 // Select colors by re-inserting edges into the IFG.
twisti@605 531 // Return TRUE if any spills occurred.
duke@0 532 uint Select( );
duke@0 533 // Helper function for select which allows biased coloring
duke@0 534 OptoReg::Name choose_color( LRG &lrg, int chunk );
duke@0 535 // Helper function which implements biasing heuristic
duke@0 536 OptoReg::Name bias_color( LRG &lrg, int chunk );
duke@0 537
duke@0 538 // Split uncolorable live ranges
duke@0 539 // Return new number of live ranges
kvn@6632 540 uint Split(uint maxlrg, ResourceArea* split_arena);
duke@0 541
duke@0 542 // Copy 'was_spilled'-edness from one Node to another.
duke@0 543 void copy_was_spilled( Node *src, Node *dst );
duke@0 544 // Set the 'spilled_once' or 'spilled_twice' flag on a node.
duke@0 545 void set_was_spilled( Node *n );
duke@0 546
duke@0 547 // Convert ideal spill-nodes into machine loads & stores
duke@0 548 // Set C->failing when fixup spills could not complete, node limit exceeded.
duke@0 549 void fixup_spills();
duke@0 550
duke@0 551 // Post-Allocation peephole copy removal
duke@0 552 void post_allocate_copy_removal();
duke@0 553 Node *skip_copies( Node *c );
never@923 554 // Replace the old node with the current live version of that value
never@923 555 // and yank the old value if it's dead.
never@923 556 int replace_and_yank_if_dead( Node *old, OptoReg::Name nreg,
never@923 557 Block *current_block, Node_List& value, Node_List& regnd ) {
never@923 558 Node* v = regnd[nreg];
never@923 559 assert(v->outcnt() != 0, "no dead values");
never@923 560 old->replace_by(v);
never@923 561 return yank_if_dead(old, current_block, &value, &regnd);
never@923 562 }
never@923 563
kvn@4776 564 int yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) {
kvn@4776 565 return yank_if_dead_recurse(old, old, current_block, value, regnd);
kvn@4776 566 }
kvn@4776 567 int yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block,
kvn@4776 568 Node_List *value, Node_List *regnd);
roland@3934 569 int yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd );
duke@0 570 int elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List &regnd, bool can_change_regs );
duke@0 571 int use_prior_register( Node *copy, uint idx, Node *def, Block *current_block, Node_List &value, Node_List &regnd );
duke@0 572 bool may_be_copy_of_callee( Node *def ) const;
duke@0 573
duke@0 574 // If nreg already contains the same constant as val then eliminate it
never@70 575 bool eliminate_copy_of_constant(Node* val, Node* n,
never@70 576 Block *current_block, Node_List& value, Node_List &regnd,
duke@0 577 OptoReg::Name nreg, OptoReg::Name nreg2);
duke@0 578 // Extend the node to LRG mapping
duke@0 579 void add_reference( const Node *node, const Node *old_node);
duke@0 580
iveresov@22815 581 // Record the first use of a def in the block for a register.
iveresov@22815 582 class RegDefUse {
iveresov@22815 583 Node* _def;
iveresov@22815 584 Node* _first_use;
iveresov@22815 585 public:
iveresov@22815 586 RegDefUse() : _def(NULL), _first_use(NULL) { }
iveresov@22815 587 Node* def() const { return _def; }
iveresov@22815 588 Node* first_use() const { return _first_use; }
iveresov@22815 589
iveresov@22815 590 void update(Node* def, Node* use) {
iveresov@22815 591 if (_def != def) {
iveresov@22815 592 _def = def;
iveresov@22815 593 _first_use = use;
iveresov@22815 594 }
iveresov@22815 595 }
iveresov@22815 596 void clear() {
iveresov@22815 597 _def = NULL;
iveresov@22815 598 _first_use = NULL;
iveresov@22815 599 }
iveresov@22815 600 };
iveresov@22815 601 typedef GrowableArray<RegDefUse> RegToDefUseMap;
iveresov@22815 602 int possibly_merge_multidef(Node *n, uint k, Block *block, RegToDefUseMap& reg2defuse);
iveresov@22815 603
iveresov@22815 604 // Merge nodes that are a part of a multidef lrg and produce the same value within a block.
iveresov@22815 605 void merge_multidefs();
iveresov@22815 606
duke@0 607 private:
duke@0 608
duke@0 609 static int _final_loads, _final_stores, _final_copies, _final_memoves;
duke@0 610 static double _final_load_cost, _final_store_cost, _final_copy_cost, _final_memove_cost;
duke@0 611 static int _conserv_coalesce, _conserv_coalesce_pair;
duke@0 612 static int _conserv_coalesce_trie, _conserv_coalesce_quad;
duke@0 613 static int _post_alloc;
duke@0 614 static int _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce;
duke@0 615 static int _used_cisc_instructions, _unused_cisc_instructions;
duke@0 616 static int _allocator_attempts, _allocator_successes;
duke@0 617
duke@0 618 #ifndef PRODUCT
duke@0 619 static uint _high_pressure, _low_pressure;
duke@0 620
duke@0 621 void dump() const;
duke@0 622 void dump( const Node *n ) const;
duke@0 623 void dump( const Block * b ) const;
duke@0 624 void dump_degree_lists() const;
duke@0 625 void dump_simplified() const;
never@2016 626 void dump_lrg( uint lidx, bool defs_only) const;
never@2016 627 void dump_lrg( uint lidx) const {
never@2016 628 // dump defs and uses by default
never@2016 629 dump_lrg(lidx, false);
never@2016 630 }
duke@0 631 void dump_bb( uint pre_order ) const;
duke@0 632
duke@0 633 // Verify that base pointers and derived pointers are still sane
duke@0 634 void verify_base_ptrs( ResourceArea *a ) const;
duke@0 635
kvn@566 636 void verify( ResourceArea *a, bool verify_ifg = false ) const;
kvn@566 637
duke@0 638 void dump_for_spill_split_recycle() const;
duke@0 639
duke@0 640 public:
duke@0 641 void dump_frame() const;
duke@0 642 char *dump_register( const Node *n, char *buf ) const;
duke@0 643 private:
duke@0 644 static void print_chaitin_statistics();
duke@0 645 #endif
duke@0 646 friend class PhaseCoalesce;
duke@0 647 friend class PhaseAggressiveCoalesce;
duke@0 648 friend class PhaseConservativeCoalesce;
duke@0 649 };
stefank@1972 650
stefank@1972 651 #endif // SHARE_VM_OPTO_CHAITIN_HPP