1 #ifdef USE_PRAGMA_IDENT_SRC 2 #pragma ident "@(#)chaitin.cpp 1.116 07/09/28 10:23:12 JVM" 3 #endif 4 /* 5 * Copyright 2000-2008 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 #include "incls/_precompiled.incl" 29 #include "incls/_chaitin.cpp.incl" 30 31 //============================================================================= 32 33 #ifndef PRODUCT 34 void LRG::dump( ) const { 35 ttyLocker ttyl; 36 tty->print("%d ",num_regs()); 37 _mask.dump(); 38 if( _msize_valid ) { 39 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size); 40 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size()); 41 } else { 42 tty->print(", #?(%d) ",_mask.Size()); 43 } 44 45 tty->print("EffDeg: "); 46 if( _degree_valid ) tty->print( "%d ", _eff_degree ); 47 else tty->print("? "); 48 49 if( is_multidef() ) { 50 tty->print("MultiDef "); 51 if (_defs != NULL) { 52 tty->print("("); 53 for (int i = 0; i < _defs->length(); i++) { 54 tty->print("N%d ", _defs->at(i)->_idx); 55 } 56 tty->print(") "); 57 } 58 } 59 else if( _def == 0 ) tty->print("Dead "); 60 else tty->print("Def: N%d ",_def->_idx); 61 62 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score()); 63 // Flags 64 if( _is_oop ) tty->print("Oop "); 65 if( _is_float ) tty->print("Float "); 66 if( _was_spilled1 ) tty->print("Spilled "); 67 if( _was_spilled2 ) tty->print("Spilled2 "); 68 if( _direct_conflict ) tty->print("Direct_conflict "); 69 if( _fat_proj ) tty->print("Fat "); 70 if( _was_lo ) tty->print("Lo "); 71 if( _has_copy ) tty->print("Copy "); 72 if( _at_risk ) tty->print("Risk "); 73 74 if( _must_spill ) tty->print("Must_spill "); 75 if( _is_bound ) tty->print("Bound "); 76 if( _msize_valid ) { 77 if( _degree_valid && lo_degree() ) tty->print("Trivial "); 78 } 79 80 tty->cr(); 81 } 82 #endif 83 84 //------------------------------score------------------------------------------ 85 // Compute score from cost and area. Low score is best to spill. 86 static double raw_score( double cost, double area ) { 87 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; 88 } 89 90 double LRG::score() const { 91 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. 92 // Bigger area lowers score, encourages spilling this live range. 93 // Bigger cost raise score, prevents spilling this live range. 94 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer 95 // to turn a divide by a constant into a multiply by the reciprical). 96 double score = raw_score( _cost, _area); 97 98 // Account for area. Basically, LRGs covering large areas are better 99 // to spill because more other LRGs get freed up. 100 if( _area == 0.0 ) // No area? Then no progress to spill 101 return 1e35; 102 103 if( _was_spilled2 ) // If spilled once before, we are unlikely 104 return score + 1e30; // to make progress again. 105 106 if( _cost >= _area*3.0 ) // Tiny area relative to cost 107 return score + 1e17; // Probably no progress to spill 108 109 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost 110 return score + 1e10; // Likely no progress to spill 111 112 return score; 113 } 114 115 //------------------------------LRG_List--------------------------------------- 116 LRG_List::LRG_List( uint max ) : _cnt(max), _max(max), _lidxs(NEW_RESOURCE_ARRAY(uint,max)) { 117 memset( _lidxs, 0, sizeof(uint)*max ); 118 } 119 120 void LRG_List::extend( uint nidx, uint lidx ) { 121 _nesting.check(); 122 if( nidx >= _max ) { 123 uint size = 16; 124 while( size <= nidx ) size <<=1; 125 _lidxs = REALLOC_RESOURCE_ARRAY( uint, _lidxs, _max, size ); 126 _max = size; 127 } 128 while( _cnt <= nidx ) 129 _lidxs[_cnt++] = 0; 130 _lidxs[nidx] = lidx; 131 } 132 133 #define NUMBUCKS 3 134 135 //------------------------------Chaitin---------------------------------------- 136 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher) 137 : PhaseRegAlloc(unique, cfg, matcher, 138 #ifndef PRODUCT 139 print_chaitin_statistics 140 #else 141 NULL 142 #endif 143 ), 144 _names(unique), _uf_map(unique), 145 _maxlrg(0), _live(0), 146 _spilled_once(Thread::current()->resource_area()), 147 _spilled_twice(Thread::current()->resource_area()), 148 _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0), 149 _oldphi(unique) 150 #ifndef PRODUCT 151 , _trace_spilling(TraceSpilling || C->method_has_option("TraceSpilling")) 152 #endif 153 { 154 NOT_PRODUCT( Compile::TracePhase t3("ctorChaitin", &_t_ctorChaitin, TimeCompiler); ) 155 uint i,j; 156 // Build a list of basic blocks, sorted by frequency 157 _blks = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks ); 158 // Experiment with sorting strategies to speed compilation 159 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket 160 Block **buckets[NUMBUCKS]; // Array of buckets 161 uint buckcnt[NUMBUCKS]; // Array of bucket counters 162 double buckval[NUMBUCKS]; // Array of bucket value cutoffs 163 for( i = 0; i < NUMBUCKS; i++ ) { 164 buckets[i] = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks ); 165 buckcnt[i] = 0; 166 // Bump by three orders of magnitude each time 167 cutoff *= 0.001; 168 buckval[i] = cutoff; 169 for( j = 0; j < _cfg._num_blocks; j++ ) { 170 buckets[i][j] = NULL; 171 } 172 } 173 // Sort blocks into buckets 174 for( i = 0; i < _cfg._num_blocks; i++ ) { 175 for( j = 0; j < NUMBUCKS; j++ ) { 176 if( (j == NUMBUCKS-1) || (_cfg._blocks[i]->_freq > buckval[j]) ) { 177 // Assign block to end of list for appropriate bucket 178 buckets[j][buckcnt[j]++] = _cfg._blocks[i]; 179 break; // kick out of inner loop 180 } 181 } 182 } 183 // Dump buckets into final block array 184 uint blkcnt = 0; 185 for( i = 0; i < NUMBUCKS; i++ ) { 186 for( j = 0; j < buckcnt[i]; j++ ) { 187 _blks[blkcnt++] = buckets[i][j]; 188 } 189 } 190 191 assert(blkcnt == _cfg._num_blocks, "Block array not totally filled"); 192 } 193 194 void PhaseChaitin::Register_Allocate() { 195 196 // Above the OLD FP (and in registers) are the incoming arguments. Stack 197 // slots in this area are called "arg_slots". Above the NEW FP (and in 198 // registers) is the outgoing argument area; above that is the spill/temp 199 // area. These are all "frame_slots". Arg_slots start at the zero 200 // stack_slots and count up to the known arg_size. Frame_slots start at 201 // the stack_slot #arg_size and go up. After allocation I map stack 202 // slots to actual offsets. Stack-slots in the arg_slot area are biased 203 // by the frame_size; stack-slots in the frame_slot area are biased by 0. 204 205 _trip_cnt = 0; 206 _alternate = 0; 207 _matcher._allocation_started = true; 208 209 ResourceArea live_arena; // Arena for liveness & IFG info 210 ResourceMark rm(&live_arena); 211 212 // Need live-ness for the IFG; need the IFG for coalescing. If the 213 // liveness is JUST for coalescing, then I can get some mileage by renaming 214 // all copy-related live ranges low and then using the max copy-related 215 // live range as a cut-off for LIVE and the IFG. In other words, I can 216 // build a subset of LIVE and IFG just for copies. 217 PhaseLive live(_cfg,_names,&live_arena); 218 219 // Need IFG for coalescing and coloring 220 PhaseIFG ifg( &live_arena ); 221 _ifg = &ifg; 222 223 if (C->unique() > _names.Size()) _names.extend(C->unique()-1, 0); 224 225 // Come out of SSA world to the Named world. Assign (virtual) registers to 226 // Nodes. Use the same register for all inputs and the output of PhiNodes 227 // - effectively ending SSA form. This requires either coalescing live 228 // ranges or inserting copies. For the moment, we insert "virtual copies" 229 // - we pretend there is a copy prior to each Phi in predecessor blocks. 230 // We will attempt to coalesce such "virtual copies" before we manifest 231 // them for real. 232 de_ssa(); 233 234 { 235 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 236 _live = NULL; // Mark live as being not available 237 rm.reset_to_mark(); // Reclaim working storage 238 IndexSet::reset_memory(C, &live_arena); 239 ifg.init(_maxlrg); // Empty IFG 240 gather_lrg_masks( false ); // Collect LRG masks 241 live.compute( _maxlrg ); // Compute liveness 242 _live = &live; // Mark LIVE as being available 243 } 244 245 // Base pointers are currently "used" by instructions which define new 246 // derived pointers. This makes base pointers live up to the where the 247 // derived pointer is made, but not beyond. Really, they need to be live 248 // across any GC point where the derived value is live. So this code looks 249 // at all the GC points, and "stretches" the live range of any base pointer 250 // to the GC point. 251 if( stretch_base_pointer_live_ranges(&live_arena) ) { 252 NOT_PRODUCT( Compile::TracePhase t3("computeLive (sbplr)", &_t_computeLive, TimeCompiler); ) 253 // Since some live range stretched, I need to recompute live 254 _live = NULL; 255 rm.reset_to_mark(); // Reclaim working storage 256 IndexSet::reset_memory(C, &live_arena); 257 ifg.init(_maxlrg); 258 gather_lrg_masks( false ); 259 live.compute( _maxlrg ); 260 _live = &live; 261 } 262 // Create the interference graph using virtual copies 263 build_ifg_virtual( ); // Include stack slots this time 264 265 // Aggressive (but pessimistic) copy coalescing. 266 // This pass works on virtual copies. Any virtual copies which are not 267 // coalesced get manifested as actual copies 268 { 269 // The IFG is/was triangular. I am 'squaring it up' so Union can run 270 // faster. Union requires a 'for all' operation which is slow on the 271 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - 272 // meaning I can visit all the Nodes neighbors less than a Node in time 273 // O(# of neighbors), but I have to visit all the Nodes greater than a 274 // given Node and search them for an instance, i.e., time O(#MaxLRG)). 275 _ifg->SquareUp(); 276 277 PhaseAggressiveCoalesce coalesce( *this ); 278 coalesce.coalesce_driver( ); 279 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do 280 // not match the Phi itself, insert a copy. 281 coalesce.insert_copies(_matcher); 282 } 283 284 // After aggressive coalesce, attempt a first cut at coloring. 285 // To color, we need the IFG and for that we need LIVE. 286 { 287 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 288 _live = NULL; 289 rm.reset_to_mark(); // Reclaim working storage 290 IndexSet::reset_memory(C, &live_arena); 291 ifg.init(_maxlrg); 292 gather_lrg_masks( true ); 293 live.compute( _maxlrg ); 294 _live = &live; 295 } 296 297 // Build physical interference graph 298 uint must_spill = 0; 299 must_spill = build_ifg_physical( &live_arena ); 300 // If we have a guaranteed spill, might as well spill now 301 if( must_spill ) { 302 if( !_maxlrg ) return; 303 // Bail out if unique gets too large (ie - unique > MaxNodeLimit) 304 C->check_node_count(10*must_spill, "out of nodes before split"); 305 if (C->failing()) return; 306 _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere 307 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 308 // or we failed to split 309 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); 310 if (C->failing()) return; 311 312 #ifdef ASSERT 313 if( VerifyOpto ) { 314 _cfg.verify(); 315 verify_base_ptrs(&live_arena); 316 } 317 #endif 318 NOT_PRODUCT( C->verify_graph_edges(); ) 319 320 compact(); // Compact LRGs; return new lower max lrg 321 322 { 323 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 324 _live = NULL; 325 rm.reset_to_mark(); // Reclaim working storage 326 IndexSet::reset_memory(C, &live_arena); 327 ifg.init(_maxlrg); // Build a new interference graph 328 gather_lrg_masks( true ); // Collect intersect mask 329 live.compute( _maxlrg ); // Compute LIVE 330 _live = &live; 331 } 332 build_ifg_physical( &live_arena ); 333 _ifg->SquareUp(); 334 _ifg->Compute_Effective_Degree(); 335 // Only do conservative coalescing if requested 336 if( OptoCoalesce ) { 337 // Conservative (and pessimistic) copy coalescing of those spills 338 PhaseConservativeCoalesce coalesce( *this ); 339 // If max live ranges greater than cutoff, don't color the stack. 340 // This cutoff can be larger than below since it is only done once. 341 coalesce.coalesce_driver( ); 342 } 343 compress_uf_map_for_nodes(); 344 345 #ifdef ASSERT 346 if( VerifyOpto ) _ifg->verify(this); 347 #endif 348 } else { 349 ifg.SquareUp(); 350 ifg.Compute_Effective_Degree(); 351 #ifdef ASSERT 352 set_was_low(); 353 #endif 354 } 355 356 // Prepare for Simplify & Select 357 cache_lrg_info(); // Count degree of LRGs 358 359 // Simplify the InterFerence Graph by removing LRGs of low degree. 360 // LRGs of low degree are trivially colorable. 361 Simplify(); 362 363 // Select colors by re-inserting LRGs back into the IFG in reverse order. 364 // Return whether or not something spills. 365 uint spills = Select( ); 366 367 // If we spill, split and recycle the entire thing 368 while( spills ) { 369 if( _trip_cnt++ > 24 ) { 370 DEBUG_ONLY( dump_for_spill_split_recycle(); ) 371 if( _trip_cnt > 27 ) { 372 C->record_method_not_compilable("failed spill-split-recycle sanity check"); 373 return; 374 } 375 } 376 377 if( !_maxlrg ) return; 378 _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere 379 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 380 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after split"); 381 if (C->failing()) return; 382 #ifdef ASSERT 383 if( VerifyOpto ) { 384 _cfg.verify(); 385 verify_base_ptrs(&live_arena); 386 } 387 #endif 388 389 compact(); // Compact LRGs; return new lower max lrg 390 391 // Nuke the live-ness and interference graph and LiveRanGe info 392 { 393 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 394 _live = NULL; 395 rm.reset_to_mark(); // Reclaim working storage 396 IndexSet::reset_memory(C, &live_arena); 397 ifg.init(_maxlrg); 398 399 // Create LiveRanGe array. 400 // Intersect register masks for all USEs and DEFs 401 gather_lrg_masks( true ); 402 live.compute( _maxlrg ); 403 _live = &live; 404 } 405 must_spill = build_ifg_physical( &live_arena ); 406 _ifg->SquareUp(); 407 _ifg->Compute_Effective_Degree(); 408 409 // Only do conservative coalescing if requested 410 if( OptoCoalesce ) { 411 // Conservative (and pessimistic) copy coalescing 412 PhaseConservativeCoalesce coalesce( *this ); 413 // Check for few live ranges determines how aggressive coalesce is. 414 coalesce.coalesce_driver( ); 415 } 416 compress_uf_map_for_nodes(); 417 #ifdef ASSERT 418 if( VerifyOpto ) _ifg->verify(this); 419 #endif 420 cache_lrg_info(); // Count degree of LRGs 421 422 // Simplify the InterFerence Graph by removing LRGs of low degree. 423 // LRGs of low degree are trivially colorable. 424 Simplify(); 425 426 // Select colors by re-inserting LRGs back into the IFG in reverse order. 427 // Return whether or not something spills. 428 spills = Select( ); 429 } 430 431 // Count number of Simplify-Select trips per coloring success. 432 _allocator_attempts += _trip_cnt + 1; 433 _allocator_successes += 1; 434 435 // Peephole remove copies 436 post_allocate_copy_removal(); 437 438 // max_reg is past the largest *register* used. 439 // Convert that to a frame_slot number. 440 if( _max_reg <= _matcher._new_SP ) 441 _framesize = C->out_preserve_stack_slots(); 442 else _framesize = _max_reg -_matcher._new_SP; 443 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); 444 445 // This frame must preserve the required fp alignment 446 _framesize = round_to(_framesize, Matcher::stack_alignment_in_slots()); 447 assert( _framesize >= 0 && _framesize <= 1000000, "sanity check" ); 448 #ifndef PRODUCT 449 _total_framesize += _framesize; 450 if( (int)_framesize > _max_framesize ) 451 _max_framesize = _framesize; 452 #endif 453 454 // Convert CISC spills 455 fixup_spills(); 456 457 // Log regalloc results 458 CompileLog* log = Compile::current()->log(); 459 if (log != NULL) { 460 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); 461 } 462 463 if (C->failing()) return; 464 465 NOT_PRODUCT( C->verify_graph_edges(); ) 466 467 // Move important info out of the live_arena to longer lasting storage. 468 alloc_node_regs(_names.Size()); 469 for( uint i=0; i < _names.Size(); i++ ) { 470 if( _names[i] ) { // Live range associated with Node? 471 LRG &lrg = lrgs( _names[i] ); 472 if( lrg.num_regs() == 1 ) { 473 _node_regs[i].set1( lrg.reg() ); 474 } else { // Must be a register-pair 475 if( !lrg._fat_proj ) { // Must be aligned adjacent register pair 476 // Live ranges record the highest register in their mask. 477 // We want the low register for the AD file writer's convenience. 478 _node_regs[i].set2( OptoReg::add(lrg.reg(),-1) ); 479 } else { // Misaligned; extract 2 bits 480 OptoReg::Name hi = lrg.reg(); // Get hi register 481 lrg.Remove(hi); // Yank from mask 482 int lo = lrg.mask().find_first_elem(); // Find lo 483 _node_regs[i].set_pair( hi, lo ); 484 } 485 } 486 if( lrg._is_oop ) _node_oops.set(i); 487 } else { 488 _node_regs[i].set_bad(); 489 } 490 } 491 492 // Done! 493 _live = NULL; 494 _ifg = NULL; 495 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope 496 } 497 498 //------------------------------de_ssa----------------------------------------- 499 void PhaseChaitin::de_ssa() { 500 // Set initial Names for all Nodes. Most Nodes get the virtual register 501 // number. A few get the ZERO live range number. These do not 502 // get allocated, but instead rely on correct scheduling to ensure that 503 // only one instance is simultaneously live at a time. 504 uint lr_counter = 1; 505 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 506 Block *b = _cfg._blocks[i]; 507 uint cnt = b->_nodes.size(); 508 509 // Handle all the normal Nodes in the block 510 for( uint j = 0; j < cnt; j++ ) { 511 Node *n = b->_nodes[j]; 512 // Pre-color to the zero live range, or pick virtual register 513 const RegMask &rm = n->out_RegMask(); 514 _names.map( n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0 ); 515 } 516 } 517 // Reset the Union-Find mapping to be identity 518 reset_uf_map(lr_counter); 519 } 520 521 522 //------------------------------gather_lrg_masks------------------------------- 523 // Gather LiveRanGe information, including register masks. Modification of 524 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. 525 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { 526 527 // Nail down the frame pointer live range 528 uint fp_lrg = n2lidx(_cfg._root->in(1)->in(TypeFunc::FramePtr)); 529 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite 530 531 // For all blocks 532 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 533 Block *b = _cfg._blocks[i]; 534 535 // For all instructions 536 for( uint j = 1; j < b->_nodes.size(); j++ ) { 537 Node *n = b->_nodes[j]; 538 uint input_edge_start =1; // Skip control most nodes 539 if( n->is_Mach() ) input_edge_start = n->as_Mach()->oper_input_base(); 540 uint idx = n->is_Copy(); 541 542 // Get virtual register number, same as LiveRanGe index 543 uint vreg = n2lidx(n); 544 LRG &lrg = lrgs(vreg); 545 if( vreg ) { // No vreg means un-allocable (e.g. memory) 546 547 // Collect has-copy bit 548 if( idx ) { 549 lrg._has_copy = 1; 550 uint clidx = n2lidx(n->in(idx)); 551 LRG ©_src = lrgs(clidx); 552 copy_src._has_copy = 1; 553 } 554 555 // Check for float-vs-int live range (used in register-pressure 556 // calculations) 557 const Type *n_type = n->bottom_type(); 558 if( n_type->is_floatingpoint() ) 559 lrg._is_float = 1; 560 561 // Check for twice prior spilling. Once prior spilling might have 562 // spilled 'soft', 2nd prior spill should have spilled 'hard' and 563 // further spilling is unlikely to make progress. 564 if( _spilled_once.test(n->_idx) ) { 565 lrg._was_spilled1 = 1; 566 if( _spilled_twice.test(n->_idx) ) 567 lrg._was_spilled2 = 1; 568 } 569 570 #ifndef PRODUCT 571 if (trace_spilling() && lrg._def != NULL) { 572 // collect defs for MultiDef printing 573 if (lrg._defs == NULL) { 574 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(); 575 lrg._defs->append(lrg._def); 576 } 577 lrg._defs->append(n); 578 } 579 #endif 580 581 // Check for a single def LRG; these can spill nicely 582 // via rematerialization. Flag as NULL for no def found 583 // yet, or 'n' for single def or -1 for many defs. 584 lrg._def = lrg._def ? NodeSentinel : n; 585 586 // Limit result register mask to acceptable registers 587 const RegMask &rm = n->out_RegMask(); 588 lrg.AND( rm ); 589 // Check for bound register masks 590 const RegMask &lrgmask = lrg.mask(); 591 if( lrgmask.is_bound1() || lrgmask.is_bound2() ) 592 lrg._is_bound = 1; 593 594 // Check for maximum frequency value 595 if( lrg._maxfreq < b->_freq ) 596 lrg._maxfreq = b->_freq; 597 598 int ireg = n->ideal_reg(); 599 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, 600 "oops must be in Op_RegP's" ); 601 // Check for oop-iness, or long/double 602 // Check for multi-kill projection 603 switch( ireg ) { 604 case MachProjNode::fat_proj: 605 // Fat projections have size equal to number of registers killed 606 lrg.set_num_regs(rm.Size()); 607 lrg.set_reg_pressure(lrg.num_regs()); 608 lrg._fat_proj = 1; 609 lrg._is_bound = 1; 610 break; 611 case Op_RegP: 612 #ifdef _LP64 613 lrg.set_num_regs(2); // Size is 2 stack words 614 #else 615 lrg.set_num_regs(1); // Size is 1 stack word 616 #endif 617 // Register pressure is tracked relative to the maximum values 618 // suggested for that platform, INTPRESSURE and FLOATPRESSURE, 619 // and relative to other types which compete for the same regs. 620 // 621 // The following table contains suggested values based on the 622 // architectures as defined in each .ad file. 623 // INTPRESSURE and FLOATPRESSURE may be tuned differently for 624 // compile-speed or performance. 625 // Note1: 626 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 627 // since .ad registers are defined as high and low halves. 628 // These reg_pressure values remain compatible with the code 629 // in is_high_pressure() which relates get_invalid_mask_size(), 630 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. 631 // Note2: 632 // SPARC -d32 has 24 registers available for integral values, 633 // but only 10 of these are safe for 64-bit longs. 634 // Using set_reg_pressure(2) for both int and long means 635 // the allocator will believe it can fit 26 longs into 636 // registers. Using 2 for longs and 1 for ints means the 637 // allocator will attempt to put 52 integers into registers. 638 // The settings below limit this problem to methods with 639 // many long values which are being run on 32-bit SPARC. 640 // 641 // ------------------- reg_pressure -------------------- 642 // Each entry is reg_pressure_per_value,number_of_regs 643 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE 644 // IA32 2 1 1 1 1 6 6 645 // IA64 1 1 1 1 1 50 41 646 // SPARC 2 2 2 2 2 48 (24) 52 (26) 647 // SPARCV9 2 2 2 2 2 48 (24) 52 (26) 648 // AMD64 1 1 1 1 1 14 15 649 // ----------------------------------------------------- 650 #if defined(SPARC) 651 lrg.set_reg_pressure(2); // use for v9 as well 652 #else 653 lrg.set_reg_pressure(1); // normally one value per register 654 #endif 655 if( n_type->isa_oop_ptr() ) { 656 lrg._is_oop = 1; 657 } 658 break; 659 case Op_RegL: // Check for long or double 660 case Op_RegD: 661 lrg.set_num_regs(2); 662 // Define platform specific register pressure 663 #ifdef SPARC 664 lrg.set_reg_pressure(2); 665 #elif defined(IA32) 666 if( ireg == Op_RegL ) { 667 lrg.set_reg_pressure(2); 668 } else { 669 lrg.set_reg_pressure(1); 670 } 671 #else 672 lrg.set_reg_pressure(1); // normally one value per register 673 #endif 674 // If this def of a double forces a mis-aligned double, 675 // flag as '_fat_proj' - really flag as allowing misalignment 676 // AND changes how we count interferences. A mis-aligned 677 // double can interfere with TWO aligned pairs, or effectively 678 // FOUR registers! 679 if( rm.is_misaligned_Pair() ) { 680 lrg._fat_proj = 1; 681 lrg._is_bound = 1; 682 } 683 break; 684 case Op_RegF: 685 case Op_RegI: 686 case Op_RegN: 687 case Op_RegFlags: 688 case 0: // not an ideal register 689 lrg.set_num_regs(1); 690 #ifdef SPARC 691 lrg.set_reg_pressure(2); 692 #else 693 lrg.set_reg_pressure(1); 694 #endif 695 break; 696 default: 697 ShouldNotReachHere(); 698 } 699 } 700 701 // Now do the same for inputs 702 uint cnt = n->req(); 703 // Setup for CISC SPILLING 704 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 705 if( UseCISCSpill && after_aggressive ) { 706 inp = n->cisc_operand(); 707 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 708 // Convert operand number to edge index number 709 inp = n->as_Mach()->operand_index(inp); 710 } 711 // Prepare register mask for each input 712 for( uint k = input_edge_start; k < cnt; k++ ) { 713 uint vreg = n2lidx(n->in(k)); 714 if( !vreg ) continue; 715 716 // If this instruction is CISC Spillable, add the flags 717 // bit to its appropriate input 718 if( UseCISCSpill && after_aggressive && inp == k ) { 719 #ifndef PRODUCT 720 if( TraceCISCSpill ) { 721 tty->print(" use_cisc_RegMask: "); 722 n->dump(); 723 } 724 #endif 725 n->as_Mach()->use_cisc_RegMask(); 726 } 727 728 LRG &lrg = lrgs(vreg); 729 // // Testing for floating point code shape 730 // Node *test = n->in(k); 731 // if( test->is_Mach() ) { 732 // MachNode *m = test->as_Mach(); 733 // int op = m->ideal_Opcode(); 734 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 735 // int zzz = 1; 736 // } 737 // } 738 739 // Limit result register mask to acceptable registers. 740 // Do not limit registers from uncommon uses before 741 // AggressiveCoalesce. This effectively pre-virtual-splits 742 // around uncommon uses of common defs. 743 const RegMask &rm = n->in_RegMask(k); 744 if( !after_aggressive && 745 _cfg._bbs[n->in(k)->_idx]->_freq > 1000*b->_freq ) { 746 // Since we are BEFORE aggressive coalesce, leave the register 747 // mask untrimmed by the call. This encourages more coalescing. 748 // Later, AFTER aggressive, this live range will have to spill 749 // but the spiller handles slow-path calls very nicely. 750 } else { 751 lrg.AND( rm ); 752 } 753 // Check for bound register masks 754 const RegMask &lrgmask = lrg.mask(); 755 if( lrgmask.is_bound1() || lrgmask.is_bound2() ) 756 lrg._is_bound = 1; 757 // If this use of a double forces a mis-aligned double, 758 // flag as '_fat_proj' - really flag as allowing misalignment 759 // AND changes how we count interferences. A mis-aligned 760 // double can interfere with TWO aligned pairs, or effectively 761 // FOUR registers! 762 if( lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_Pair() ) { 763 lrg._fat_proj = 1; 764 lrg._is_bound = 1; 765 } 766 // if the LRG is an unaligned pair, we will have to spill 767 // so clear the LRG's register mask if it is not already spilled 768 if ( !n->is_SpillCopy() && 769 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 770 lrgmask.is_misaligned_Pair()) { 771 lrg.Clear(); 772 } 773 774 // Check for maximum frequency value 775 if( lrg._maxfreq < b->_freq ) 776 lrg._maxfreq = b->_freq; 777 778 } // End for all allocated inputs 779 } // end for all instructions 780 } // end for all blocks 781 782 // Final per-liverange setup 783 for( uint i2=0; i2<_maxlrg; i2++ ) { 784 LRG &lrg = lrgs(i2); 785 if( lrg.num_regs() == 2 && !lrg._fat_proj ) 786 lrg.ClearToPairs(); 787 lrg.compute_set_mask_size(); 788 if( lrg.not_free() ) { // Handle case where we lose from the start 789 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 790 lrg._direct_conflict = 1; 791 } 792 lrg.set_degree(0); // no neighbors in IFG yet 793 } 794 } 795 796 //------------------------------set_was_low------------------------------------ 797 // Set the was-lo-degree bit. Conservative coalescing should not change the 798 // colorability of the graph. If any live range was of low-degree before 799 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 800 // The bit is checked in Simplify. 801 void PhaseChaitin::set_was_low() { 802 #ifdef ASSERT 803 for( uint i = 1; i < _maxlrg; i++ ) { 804 int size = lrgs(i).num_regs(); 805 uint old_was_lo = lrgs(i)._was_lo; 806 lrgs(i)._was_lo = 0; 807 if( lrgs(i).lo_degree() ) { 808 lrgs(i)._was_lo = 1; // Trivially of low degree 809 } else { // Else check the Brigg's assertion 810 // Brigg's observation is that the lo-degree neighbors of a 811 // hi-degree live range will not interfere with the color choices 812 // of said hi-degree live range. The Simplify reverse-stack-coloring 813 // order takes care of the details. Hence you do not have to count 814 // low-degree neighbors when determining if this guy colors. 815 int briggs_degree = 0; 816 IndexSet *s = _ifg->neighbors(i); 817 IndexSetIterator elements(s); 818 uint lidx; 819 while((lidx = elements.next()) != 0) { 820 if( !lrgs(lidx).lo_degree() ) 821 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 822 } 823 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 824 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 825 } 826 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 827 } 828 #endif 829 } 830 831 #define REGISTER_CONSTRAINED 16 832 833 //------------------------------cache_lrg_info--------------------------------- 834 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 835 void PhaseChaitin::cache_lrg_info( ) { 836 837 for( uint i = 1; i < _maxlrg; i++ ) { 838 LRG &lrg = lrgs(i); 839 840 // Check for being of low degree: means we can be trivially colored. 841 // Low degree, dead or must-spill guys just get to simplify right away 842 if( lrg.lo_degree() || 843 !lrg.alive() || 844 lrg._must_spill ) { 845 // Split low degree list into those guys that must get a 846 // register and those that can go to register or stack. 847 // The idea is LRGs that can go register or stack color first when 848 // they have a good chance of getting a register. The register-only 849 // lo-degree live ranges always get a register. 850 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 851 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 852 lrg._next = _lo_stk_degree; 853 _lo_stk_degree = i; 854 } else { 855 lrg._next = _lo_degree; 856 _lo_degree = i; 857 } 858 } else { // Else high degree 859 lrgs(_hi_degree)._prev = i; 860 lrg._next = _hi_degree; 861 lrg._prev = 0; 862 _hi_degree = i; 863 } 864 } 865 } 866 867 //------------------------------Pre-Simplify----------------------------------- 868 // Simplify the IFG by removing LRGs of low degree that have NO copies 869 void PhaseChaitin::Pre_Simplify( ) { 870 871 // Warm up the lo-degree no-copy list 872 int lo_no_copy = 0; 873 for( uint i = 1; i < _maxlrg; i++ ) { 874 if( (lrgs(i).lo_degree() && !lrgs(i)._has_copy) || 875 !lrgs(i).alive() || 876 lrgs(i)._must_spill ) { 877 lrgs(i)._next = lo_no_copy; 878 lo_no_copy = i; 879 } 880 } 881 882 while( lo_no_copy ) { 883 uint lo = lo_no_copy; 884 lo_no_copy = lrgs(lo)._next; 885 int size = lrgs(lo).num_regs(); 886 887 // Put the simplified guy on the simplified list. 888 lrgs(lo)._next = _simplified; 889 _simplified = lo; 890 891 // Yank this guy from the IFG. 892 IndexSet *adj = _ifg->remove_node( lo ); 893 894 // If any neighbors' degrees fall below their number of 895 // allowed registers, then put that neighbor on the low degree 896 // list. Note that 'degree' can only fall and 'numregs' is 897 // unchanged by this action. Thus the two are equal at most once, 898 // so LRGs hit the lo-degree worklists at most once. 899 IndexSetIterator elements(adj); 900 uint neighbor; 901 while ((neighbor = elements.next()) != 0) { 902 LRG *n = &lrgs(neighbor); 903 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 904 905 // Check for just becoming of-low-degree 906 if( n->just_lo_degree() && !n->_has_copy ) { 907 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 908 // Put on lo-degree list 909 n->_next = lo_no_copy; 910 lo_no_copy = neighbor; 911 } 912 } 913 } // End of while lo-degree no_copy worklist not empty 914 915 // No more lo-degree no-copy live ranges to simplify 916 } 917 918 //------------------------------Simplify--------------------------------------- 919 // Simplify the IFG by removing LRGs of low degree. 920 void PhaseChaitin::Simplify( ) { 921 922 while( 1 ) { // Repeat till simplified it all 923 // May want to explore simplifying lo_degree before _lo_stk_degree. 924 // This might result in more spills coloring into registers during 925 // Select(). 926 while( _lo_degree || _lo_stk_degree ) { 927 // If possible, pull from lo_stk first 928 uint lo; 929 if( _lo_degree ) { 930 lo = _lo_degree; 931 _lo_degree = lrgs(lo)._next; 932 } else { 933 lo = _lo_stk_degree; 934 _lo_stk_degree = lrgs(lo)._next; 935 } 936 937 // Put the simplified guy on the simplified list. 938 lrgs(lo)._next = _simplified; 939 _simplified = lo; 940 // If this guy is "at risk" then mark his current neighbors 941 if( lrgs(lo)._at_risk ) { 942 IndexSetIterator elements(_ifg->neighbors(lo)); 943 uint datum; 944 while ((datum = elements.next()) != 0) { 945 lrgs(datum)._risk_bias = lo; 946 } 947 } 948 949 // Yank this guy from the IFG. 950 IndexSet *adj = _ifg->remove_node( lo ); 951 952 // If any neighbors' degrees fall below their number of 953 // allowed registers, then put that neighbor on the low degree 954 // list. Note that 'degree' can only fall and 'numregs' is 955 // unchanged by this action. Thus the two are equal at most once, 956 // so LRGs hit the lo-degree worklist at most once. 957 IndexSetIterator elements(adj); 958 uint neighbor; 959 while ((neighbor = elements.next()) != 0) { 960 LRG *n = &lrgs(neighbor); 961 #ifdef ASSERT 962 if( VerifyOpto ) { 963 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 964 } 965 #endif 966 967 // Check for just becoming of-low-degree just counting registers. 968 // _must_spill live ranges are already on the low degree list. 969 if( n->just_lo_degree() && !n->_must_spill ) { 970 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 971 // Pull from hi-degree list 972 uint prev = n->_prev; 973 uint next = n->_next; 974 if( prev ) lrgs(prev)._next = next; 975 else _hi_degree = next; 976 lrgs(next)._prev = prev; 977 n->_next = _lo_degree; 978 _lo_degree = neighbor; 979 } 980 } 981 } // End of while lo-degree/lo_stk_degree worklist not empty 982 983 // Check for got everything: is hi-degree list empty? 984 if( !_hi_degree ) break; 985 986 // Time to pick a potential spill guy 987 uint lo_score = _hi_degree; 988 double score = lrgs(lo_score).score(); 989 double area = lrgs(lo_score)._area; 990 991 // Find cheapest guy 992 debug_only( int lo_no_simplify=0; ); 993 for( uint i = _hi_degree; i; i = lrgs(i)._next ) { 994 assert( !(*_ifg->_yanked)[i], "" ); 995 // It's just vaguely possible to move hi-degree to lo-degree without 996 // going through a just-lo-degree stage: If you remove a double from 997 // a float live range it's degree will drop by 2 and you can skip the 998 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 999 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1000 if( lrgs(i).lo_degree() ) { 1001 lo_score = i; 1002 break; 1003 } 1004 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1005 double iscore = lrgs(i).score(); 1006 double iarea = lrgs(i)._area; 1007 1008 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1009 // wins. Ties happen because all live ranges in question have spilled 1010 // a few times before and the spill-score adds a huge number which 1011 // washes out the low order bits. We are choosing the lesser of 2 1012 // evils; in this case pick largest area to spill. 1013 if( iscore < score || 1014 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) ) { 1015 lo_score = i; 1016 score = iscore; 1017 area = iarea; 1018 } 1019 } 1020 LRG *lo_lrg = &lrgs(lo_score); 1021 // The live range we choose for spilling is either hi-degree, or very 1022 // rarely it can be low-degree. If we choose a hi-degree live range 1023 // there better not be any lo-degree choices. 1024 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1025 1026 // Pull from hi-degree list 1027 uint prev = lo_lrg->_prev; 1028 uint next = lo_lrg->_next; 1029 if( prev ) lrgs(prev)._next = next; 1030 else _hi_degree = next; 1031 lrgs(next)._prev = prev; 1032 // Jam him on the lo-degree list, despite his high degree. 1033 // Maybe he'll get a color, and maybe he'll spill. 1034 // Only Select() will know. 1035 lrgs(lo_score)._at_risk = true; 1036 _lo_degree = lo_score; 1037 lo_lrg->_next = 0; 1038 1039 } // End of while not simplified everything 1040 1041 } 1042 1043 //------------------------------bias_color------------------------------------- 1044 // Choose a color using the biasing heuristic 1045 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1046 1047 // Check for "at_risk" LRG's 1048 uint risk_lrg = Find(lrg._risk_bias); 1049 if( risk_lrg != 0 ) { 1050 // Walk the colored neighbors of the "at_risk" candidate 1051 // Choose a color which is both legal and already taken by a neighbor 1052 // of the "at_risk" candidate in order to improve the chances of the 1053 // "at_risk" candidate of coloring 1054 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1055 uint datum; 1056 while ((datum = elements.next()) != 0) { 1057 OptoReg::Name reg = lrgs(datum).reg(); 1058 // If this LRG's register is legal for us, choose it 1059 if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE && 1060 lrg.mask().Member(OptoReg::add(reg,-chunk)) && 1061 (lrg.num_regs()==1 || // either size 1 1062 (reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs) 1063 return reg; 1064 } 1065 } 1066 1067 uint copy_lrg = Find(lrg._copy_bias); 1068 if( copy_lrg != 0 ) { 1069 // If he has a color, 1070 if( !(*(_ifg->_yanked))[copy_lrg] ) { 1071 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1072 // And it is legal for you, 1073 if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE && 1074 lrg.mask().Member(OptoReg::add(reg,-chunk)) && 1075 (lrg.num_regs()==1 || // either size 1 1076 (reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs) 1077 return reg; 1078 } else if( chunk == 0 ) { 1079 // Choose a color which is legal for him 1080 RegMask tempmask = lrg.mask(); 1081 tempmask.AND(lrgs(copy_lrg).mask()); 1082 OptoReg::Name reg; 1083 if( lrg.num_regs() == 1 ) { 1084 reg = tempmask.find_first_elem(); 1085 } else { 1086 tempmask.ClearToPairs(); 1087 reg = tempmask.find_first_pair(); 1088 } 1089 if( OptoReg::is_valid(reg) ) 1090 return reg; 1091 } 1092 } 1093 1094 // If no bias info exists, just go with the register selection ordering 1095 if( lrg.num_regs() == 2 ) { 1096 // Find an aligned pair 1097 return OptoReg::add(lrg.mask().find_first_pair(),chunk); 1098 } 1099 1100 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1101 // copy removal to remove many more copies, by preventing a just-assigned 1102 // register from being repeatedly assigned. 1103 OptoReg::Name reg = lrg.mask().find_first_elem(); 1104 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1105 // This 'Remove; find; Insert' idiom is an expensive way to find the 1106 // SECOND element in the mask. 1107 lrg.Remove(reg); 1108 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1109 lrg.Insert(reg); 1110 if( OptoReg::is_reg(reg2)) 1111 reg = reg2; 1112 } 1113 return OptoReg::add( reg, chunk ); 1114 } 1115 1116 //------------------------------choose_color----------------------------------- 1117 // Choose a color in the current chunk 1118 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1119 assert( C->in_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP-1)), "must not allocate stack0 (inside preserve area)"); 1120 assert(C->out_preserve_stack_slots() == 0 || chunk != 0 || lrg._is_bound || lrg.mask().is_bound1() || !lrg.mask().Member(OptoReg::Name(_matcher._old_SP+0)), "must not allocate stack0 (inside preserve area)"); 1121 1122 if( lrg.num_regs() == 1 || // Common Case 1123 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1124 // Use a heuristic to "bias" the color choice 1125 return bias_color(lrg, chunk); 1126 1127 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1128 1129 // Fat-proj case or misaligned double argument. 1130 assert(lrg.compute_mask_size() == lrg.num_regs() || 1131 lrg.num_regs() == 2,"fat projs exactly color" ); 1132 assert( !chunk, "always color in 1st chunk" ); 1133 // Return the highest element in the set. 1134 return lrg.mask().find_last_elem(); 1135 } 1136 1137 //------------------------------Select----------------------------------------- 1138 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1139 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1140 // everything going back is guaranteed a color. Select that color. If some 1141 // hi-degree LRG cannot get a color then we record that we must spill. 1142 uint PhaseChaitin::Select( ) { 1143 uint spill_reg = LRG::SPILL_REG; 1144 _max_reg = OptoReg::Name(0); // Past max register used 1145 while( _simplified ) { 1146 // Pull next LRG from the simplified list - in reverse order of removal 1147 uint lidx = _simplified; 1148 LRG *lrg = &lrgs(lidx); 1149 _simplified = lrg->_next; 1150 1151 1152 #ifndef PRODUCT 1153 if (trace_spilling()) { 1154 ttyLocker ttyl; 1155 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1156 lrg->degrees_of_freedom()); 1157 lrg->dump(); 1158 } 1159 #endif 1160 1161 // Re-insert into the IFG 1162 _ifg->re_insert(lidx); 1163 if( !lrg->alive() ) continue; 1164 // capture allstackedness flag before mask is hacked 1165 const int is_allstack = lrg->mask().is_AllStack(); 1166 1167 // Yeah, yeah, yeah, I know, I know. I can refactor this 1168 // to avoid the GOTO, although the refactored code will not 1169 // be much clearer. We arrive here IFF we have a stack-based 1170 // live range that cannot color in the current chunk, and it 1171 // has to move into the next free stack chunk. 1172 int chunk = 0; // Current chunk is first chunk 1173 retry_next_chunk: 1174 1175 // Remove neighbor colors 1176 IndexSet *s = _ifg->neighbors(lidx); 1177 1178 debug_only(RegMask orig_mask = lrg->mask();) 1179 IndexSetIterator elements(s); 1180 uint neighbor; 1181 while ((neighbor = elements.next()) != 0) { 1182 // Note that neighbor might be a spill_reg. In this case, exclusion 1183 // of its color will be a no-op, since the spill_reg chunk is in outer 1184 // space. Also, if neighbor is in a different chunk, this exclusion 1185 // will be a no-op. (Later on, if lrg runs out of possible colors in 1186 // its chunk, a new chunk of color may be tried, in which case 1187 // examination of neighbors is started again, at retry_next_chunk.) 1188 LRG &nlrg = lrgs(neighbor); 1189 OptoReg::Name nreg = nlrg.reg(); 1190 // Only subtract masks in the same chunk 1191 if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { 1192 #ifndef PRODUCT 1193 uint size = lrg->mask().Size(); 1194 RegMask rm = lrg->mask(); 1195 #endif 1196 lrg->SUBTRACT(nlrg.mask()); 1197 #ifndef PRODUCT 1198 if (trace_spilling() && lrg->mask().Size() != size) { 1199 ttyLocker ttyl; 1200 tty->print("L%d ", lidx); 1201 rm.dump(); 1202 tty->print(" intersected L%d ", neighbor); 1203 nlrg.mask().dump(); 1204 tty->print(" removed "); 1205 rm.SUBTRACT(lrg->mask()); 1206 rm.dump(); 1207 tty->print(" leaving "); 1208 lrg->mask().dump(); 1209 tty->cr(); 1210 } 1211 #endif 1212 } 1213 } 1214 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1215 // Aligned pairs need aligned masks 1216 if( lrg->num_regs() == 2 && !lrg->_fat_proj ) 1217 lrg->ClearToPairs(); 1218 1219 // Check if a color is available and if so pick the color 1220 OptoReg::Name reg = choose_color( *lrg, chunk ); 1221 #ifdef SPARC 1222 debug_only(lrg->compute_set_mask_size()); 1223 assert(lrg->num_regs() != 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); 1224 #endif 1225 1226 //--------------- 1227 // If we fail to color and the AllStack flag is set, trigger 1228 // a chunk-rollover event 1229 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1230 // Bump register mask up to next stack chunk 1231 chunk += RegMask::CHUNK_SIZE; 1232 lrg->Set_All(); 1233 1234 goto retry_next_chunk; 1235 } 1236 1237 //--------------- 1238 // Did we get a color? 1239 else if( OptoReg::is_valid(reg)) { 1240 #ifndef PRODUCT 1241 RegMask avail_rm = lrg->mask(); 1242 #endif 1243 1244 // Record selected register 1245 lrg->set_reg(reg); 1246 1247 if( reg >= _max_reg ) // Compute max register limit 1248 _max_reg = OptoReg::add(reg,1); 1249 // Fold reg back into normal space 1250 reg = OptoReg::add(reg,-chunk); 1251 1252 // If the live range is not bound, then we actually had some choices 1253 // to make. In this case, the mask has more bits in it than the colors 1254 // choosen. Restrict the mask to just what was picked. 1255 if( lrg->num_regs() == 1 ) { // Size 1 live range 1256 lrg->Clear(); // Clear the mask 1257 lrg->Insert(reg); // Set regmask to match selected reg 1258 lrg->set_mask_size(1); 1259 } else if( !lrg->_fat_proj ) { 1260 // For pairs, also insert the low bit of the pair 1261 assert( lrg->num_regs() == 2, "unbound fatproj???" ); 1262 lrg->Clear(); // Clear the mask 1263 lrg->Insert(reg); // Set regmask to match selected reg 1264 lrg->Insert(OptoReg::add(reg,-1)); 1265 lrg->set_mask_size(2); 1266 } else { // Else fatproj 1267 // mask must be equal to fatproj bits, by definition 1268 } 1269 #ifndef PRODUCT 1270 if (trace_spilling()) { 1271 ttyLocker ttyl; 1272 tty->print("L%d selected ", lidx); 1273 lrg->mask().dump(); 1274 tty->print(" from "); 1275 avail_rm.dump(); 1276 tty->cr(); 1277 } 1278 #endif 1279 // Note that reg is the highest-numbered register in the newly-bound mask. 1280 } // end color available case 1281 1282 //--------------- 1283 // Live range is live and no colors available 1284 else { 1285 assert( lrg->alive(), "" ); 1286 assert( !lrg->_fat_proj || lrg->is_multidef() || 1287 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1288 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1289 1290 // Assign the special spillreg register 1291 lrg->set_reg(OptoReg::Name(spill_reg++)); 1292 // Do not empty the regmask; leave mask_size lying around 1293 // for use during Spilling 1294 #ifndef PRODUCT 1295 if( trace_spilling() ) { 1296 ttyLocker ttyl; 1297 tty->print("L%d spilling with neighbors: ", lidx); 1298 s->dump(); 1299 debug_only(tty->print(" original mask: ")); 1300 debug_only(orig_mask.dump()); 1301 dump_lrg(lidx); 1302 } 1303 #endif 1304 } // end spill case 1305 1306 } 1307 1308 return spill_reg-LRG::SPILL_REG; // Return number of spills 1309 } 1310 1311 1312 //------------------------------copy_was_spilled------------------------------- 1313 // Copy 'was_spilled'-edness from the source Node to the dst Node. 1314 void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { 1315 if( _spilled_once.test(src->_idx) ) { 1316 _spilled_once.set(dst->_idx); 1317 lrgs(Find(dst))._was_spilled1 = 1; 1318 if( _spilled_twice.test(src->_idx) ) { 1319 _spilled_twice.set(dst->_idx); 1320 lrgs(Find(dst))._was_spilled2 = 1; 1321 } 1322 } 1323 } 1324 1325 //------------------------------set_was_spilled-------------------------------- 1326 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1327 void PhaseChaitin::set_was_spilled( Node *n ) { 1328 if( _spilled_once.test_set(n->_idx) ) 1329 _spilled_twice.set(n->_idx); 1330 } 1331 1332 //------------------------------fixup_spills----------------------------------- 1333 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1334 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1335 void PhaseChaitin::fixup_spills() { 1336 // This function does only cisc spill work. 1337 if( !UseCISCSpill ) return; 1338 1339 NOT_PRODUCT( Compile::TracePhase t3("fixupSpills", &_t_fixupSpills, TimeCompiler); ) 1340 1341 // Grab the Frame Pointer 1342 Node *fp = _cfg._broot->head()->in(1)->in(TypeFunc::FramePtr); 1343 1344 // For all blocks 1345 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 1346 Block *b = _cfg._blocks[i]; 1347 1348 // For all instructions in block 1349 uint last_inst = b->end_idx(); 1350 for( uint j = 1; j <= last_inst; j++ ) { 1351 Node *n = b->_nodes[j]; 1352 1353 // Dead instruction??? 1354 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1355 C->top() == n || // Or the random TOP node 1356 n->is_Proj(), // Or a fat-proj kill node 1357 "No dead instructions after post-alloc" ); 1358 1359 int inp = n->cisc_operand(); 1360 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1361 // Convert operand number to edge index number 1362 MachNode *mach = n->as_Mach(); 1363 inp = mach->operand_index(inp); 1364 Node *src = n->in(inp); // Value to load or store 1365 LRG &lrg_cisc = lrgs( Find_const(src) ); 1366 OptoReg::Name src_reg = lrg_cisc.reg(); 1367 // Doubles record the HIGH register of an adjacent pair. 1368 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1369 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1370 // This is a CISC Spill, get stack offset and construct new node 1371 #ifndef PRODUCT 1372 if( TraceCISCSpill ) { 1373 tty->print(" reg-instr: "); 1374 n->dump(); 1375 } 1376 #endif 1377 int stk_offset = reg2offset(src_reg); 1378 // Bailout if we might exceed node limit when spilling this instruction 1379 C->check_node_count(0, "out of nodes fixing spills"); 1380 if (C->failing()) return; 1381 // Transform node 1382 MachNode *cisc = mach->cisc_version(stk_offset, C)->as_Mach(); 1383 cisc->set_req(inp,fp); // Base register is frame pointer 1384 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1385 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1386 cisc->ins_req(1,src); // Requires a memory edge 1387 } 1388 b->_nodes.map(j,cisc); // Insert into basic block 1389 n->subsume_by(cisc); // Correct graph 1390 // 1391 ++_used_cisc_instructions; 1392 #ifndef PRODUCT 1393 if( TraceCISCSpill ) { 1394 tty->print(" cisc-instr: "); 1395 cisc->dump(); 1396 } 1397 #endif 1398 } else { 1399 #ifndef PRODUCT 1400 if( TraceCISCSpill ) { 1401 tty->print(" using reg-instr: "); 1402 n->dump(); 1403 } 1404 #endif 1405 ++_unused_cisc_instructions; // input can be on stack 1406 } 1407 } 1408 1409 } // End of for all instructions 1410 1411 } // End of for all blocks 1412 } 1413 1414 //------------------------------find_base_for_derived-------------------------- 1415 // Helper to stretch above; recursively discover the base Node for a 1416 // given derived Node. Easy for AddP-related machine nodes, but needs 1417 // to be recursive for derived Phis. 1418 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1419 // See if already computed; if so return it 1420 if( derived_base_map[derived->_idx] ) 1421 return derived_base_map[derived->_idx]; 1422 1423 // See if this happens to be a base. 1424 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1425 // pointers derived from NULL! These are always along paths that 1426 // can't happen at run-time but the optimizer cannot deduce it so 1427 // we have to handle it gracefully. 1428 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1429 // If its an OOP with a non-zero offset, then it is derived. 1430 if( tj->_offset == 0 ) { 1431 derived_base_map[derived->_idx] = derived; 1432 return derived; 1433 } 1434 // Derived is NULL+offset? Base is NULL! 1435 if( derived->is_Con() ) { 1436 Node *base = new (C, 1) ConPNode( TypePtr::NULL_PTR ); 1437 uint no_lidx = 0; // an unmatched constant in debug info has no LRG 1438 _names.extend(base->_idx, no_lidx); 1439 derived_base_map[derived->_idx] = base; 1440 return base; 1441 } 1442 1443 // Check for AddP-related opcodes 1444 if( !derived->is_Phi() ) { 1445 assert( derived->as_Mach()->ideal_Opcode() == Op_AddP, "" ); 1446 Node *base = derived->in(AddPNode::Base); 1447 derived_base_map[derived->_idx] = base; 1448 return base; 1449 } 1450 1451 // Recursively find bases for Phis. 1452 // First check to see if we can avoid a base Phi here. 1453 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1454 uint i; 1455 for( i = 2; i < derived->req(); i++ ) 1456 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1457 break; 1458 // Went to the end without finding any different bases? 1459 if( i == derived->req() ) { // No need for a base Phi here 1460 derived_base_map[derived->_idx] = base; 1461 return base; 1462 } 1463 1464 // Now we see we need a base-Phi here to merge the bases 1465 base = new (C, derived->req()) PhiNode( derived->in(0), base->bottom_type() ); 1466 for( i = 1; i < derived->req(); i++ ) 1467 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1468 1469 // Search the current block for an existing base-Phi 1470 Block *b = _cfg._bbs[derived->_idx]; 1471 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1472 Node *phi = b->_nodes[i]; 1473 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1474 b->_nodes.insert( i, base ); // Must insert created Phi here as base 1475 _cfg._bbs.map( base->_idx, b ); 1476 new_lrg(base,maxlrg++); 1477 break; 1478 } 1479 // See if Phi matches. 1480 uint j; 1481 for( j = 1; j < base->req(); j++ ) 1482 if( phi->in(j) != base->in(j) && 1483 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1484 break; 1485 if( j == base->req() ) { // All inputs match? 1486 base = phi; // Then use existing 'phi' and drop 'base' 1487 break; 1488 } 1489 } 1490 1491 1492 // Cache info for later passes 1493 derived_base_map[derived->_idx] = base; 1494 return base; 1495 } 1496 1497 1498 //------------------------------stretch_base_pointer_live_ranges--------------- 1499 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1500 // base pointer that is live across the Safepoint for oopmap building. The 1501 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1502 // required edge set. 1503 bool PhaseChaitin::stretch_base_pointer_live_ranges( ResourceArea *a ) { 1504 int must_recompute_live = false; 1505 uint maxlrg = _maxlrg; 1506 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1507 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1508 1509 // For all blocks in RPO do... 1510 for( uint i=0; i<_cfg._num_blocks; i++ ) { 1511 Block *b = _cfg._blocks[i]; 1512 // Note use of deep-copy constructor. I cannot hammer the original 1513 // liveout bits, because they are needed by the following coalesce pass. 1514 IndexSet liveout(_live->live(b)); 1515 1516 for( uint j = b->end_idx() + 1; j > 1; j-- ) { 1517 Node *n = b->_nodes[j-1]; 1518 1519 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1520 // like to see in the same register. Compare uses the loop-phi and so 1521 // extends its live range BUT cannot be part of the cycle. If this 1522 // extended live range overlaps with the update of the loop-phi value 1523 // we need both alive at the same time -- which requires at least 1 1524 // copy. But because Intel has only 2-address registers we end up with 1525 // at least 2 copies, one before the loop-phi update instruction and 1526 // one after. Instead we split the input to the compare just after the 1527 // phi. 1528 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1529 Node *phi = n->in(1); 1530 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1531 Block *phi_block = _cfg._bbs[phi->_idx]; 1532 if( _cfg._bbs[phi_block->pred(2)->_idx] == b ) { 1533 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1534 Node *spill = new (C) MachSpillCopyNode( phi, *mask, *mask ); 1535 insert_proj( phi_block, 1, spill, maxlrg++ ); 1536 n->set_req(1,spill); 1537 must_recompute_live = true; 1538 } 1539 } 1540 } 1541 1542 // Get value being defined 1543 uint lidx = n2lidx(n); 1544 if( lidx && lidx < _maxlrg /* Ignore the occasional brand-new live range */) { 1545 // Remove from live-out set 1546 liveout.remove(lidx); 1547 1548 // Copies do not define a new value and so do not interfere. 1549 // Remove the copies source from the liveout set before interfering. 1550 uint idx = n->is_Copy(); 1551 if( idx ) liveout.remove( n2lidx(n->in(idx)) ); 1552 } 1553 1554 // Found a safepoint? 1555 JVMState *jvms = n->jvms(); 1556 if( jvms ) { 1557 // Now scan for a live derived pointer 1558 IndexSetIterator elements(&liveout); 1559 uint neighbor; 1560 while ((neighbor = elements.next()) != 0) { 1561 // Find reaching DEF for base and derived values 1562 // This works because we are still in SSA during this call. 1563 Node *derived = lrgs(neighbor)._def; 1564 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1565 // If its an OOP with a non-zero offset, then it is derived. 1566 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { 1567 Node *base = find_base_for_derived( derived_base_map, derived, maxlrg ); 1568 assert( base->_idx < _names.Size(), "" ); 1569 // Add reaching DEFs of derived pointer and base pointer as a 1570 // pair of inputs 1571 n->add_req( derived ); 1572 n->add_req( base ); 1573 1574 // See if the base pointer is already live to this point. 1575 // Since I'm working on the SSA form, live-ness amounts to 1576 // reaching def's. So if I find the base's live range then 1577 // I know the base's def reaches here. 1578 if( (n2lidx(base) >= _maxlrg ||// (Brand new base (hence not live) or 1579 !liveout.member( n2lidx(base) ) ) && // not live) AND 1580 (n2lidx(base) > 0) && // not a constant 1581 _cfg._bbs[base->_idx] != b ) { // base not def'd in blk) 1582 // Base pointer is not currently live. Since I stretched 1583 // the base pointer to here and it crosses basic-block 1584 // boundaries, the global live info is now incorrect. 1585 // Recompute live. 1586 must_recompute_live = true; 1587 } // End of if base pointer is not live to debug info 1588 } 1589 } // End of scan all live data for derived ptrs crossing GC point 1590 } // End of if found a GC point 1591 1592 // Make all inputs live 1593 if( !n->is_Phi() ) { // Phi function uses come from prior block 1594 for( uint k = 1; k < n->req(); k++ ) { 1595 uint lidx = n2lidx(n->in(k)); 1596 if( lidx < _maxlrg ) 1597 liveout.insert( lidx ); 1598 } 1599 } 1600 1601 } // End of forall instructions in block 1602 liveout.clear(); // Free the memory used by liveout. 1603 1604 } // End of forall blocks 1605 _maxlrg = maxlrg; 1606 1607 // If I created a new live range I need to recompute live 1608 if( maxlrg != _ifg->_maxlrg ) 1609 must_recompute_live = true; 1610 1611 return must_recompute_live != 0; 1612 } 1613 1614 1615 //------------------------------add_reference---------------------------------- 1616 // Extend the node to LRG mapping 1617 void PhaseChaitin::add_reference( const Node *node, const Node *old_node ) { 1618 _names.extend( node->_idx, n2lidx(old_node) ); 1619 } 1620 1621 //------------------------------dump------------------------------------------- 1622 #ifndef PRODUCT 1623 void PhaseChaitin::dump( const Node *n ) const { 1624 uint r = (n->_idx < _names.Size() ) ? Find_const(n) : 0; 1625 tty->print("L%d",r); 1626 if( r && n->Opcode() != Op_Phi ) { 1627 if( _node_regs ) { // Got a post-allocation copy of allocation? 1628 tty->print("["); 1629 OptoReg::Name second = get_reg_second(n); 1630 if( OptoReg::is_valid(second) ) { 1631 if( OptoReg::is_reg(second) ) 1632 tty->print("%s:",Matcher::regName[second]); 1633 else 1634 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1635 } 1636 OptoReg::Name first = get_reg_first(n); 1637 if( OptoReg::is_reg(first) ) 1638 tty->print("%s]",Matcher::regName[first]); 1639 else 1640 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1641 } else 1642 n->out_RegMask().dump(); 1643 } 1644 tty->print("/N%d\t",n->_idx); 1645 tty->print("%s === ", n->Name()); 1646 uint k; 1647 for( k = 0; k < n->req(); k++) { 1648 Node *m = n->in(k); 1649 if( !m ) tty->print("_ "); 1650 else { 1651 uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0; 1652 tty->print("L%d",r); 1653 // Data MultiNode's can have projections with no real registers. 1654 // Don't die while dumping them. 1655 int op = n->Opcode(); 1656 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1657 if( _node_regs ) { 1658 tty->print("["); 1659 OptoReg::Name second = get_reg_second(n->in(k)); 1660 if( OptoReg::is_valid(second) ) { 1661 if( OptoReg::is_reg(second) ) 1662 tty->print("%s:",Matcher::regName[second]); 1663 else 1664 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 1665 reg2offset_unchecked(second)); 1666 } 1667 OptoReg::Name first = get_reg_first(n->in(k)); 1668 if( OptoReg::is_reg(first) ) 1669 tty->print("%s]",Matcher::regName[first]); 1670 else 1671 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 1672 reg2offset_unchecked(first)); 1673 } else 1674 n->in_RegMask(k).dump(); 1675 } 1676 tty->print("/N%d ",m->_idx); 1677 } 1678 } 1679 if( k < n->len() && n->in(k) ) tty->print("| "); 1680 for( ; k < n->len(); k++ ) { 1681 Node *m = n->in(k); 1682 if( !m ) break; 1683 uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0; 1684 tty->print("L%d",r); 1685 tty->print("/N%d ",m->_idx); 1686 } 1687 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 1688 else n->dump_spec(tty); 1689 if( _spilled_once.test(n->_idx ) ) { 1690 tty->print(" Spill_1"); 1691 if( _spilled_twice.test(n->_idx ) ) 1692 tty->print(" Spill_2"); 1693 } 1694 tty->print("\n"); 1695 } 1696 1697 void PhaseChaitin::dump( const Block * b ) const { 1698 b->dump_head( &_cfg._bbs ); 1699 1700 // For all instructions 1701 for( uint j = 0; j < b->_nodes.size(); j++ ) 1702 dump(b->_nodes[j]); 1703 // Print live-out info at end of block 1704 if( _live ) { 1705 tty->print("Liveout: "); 1706 IndexSet *live = _live->live(b); 1707 IndexSetIterator elements(live); 1708 tty->print("{"); 1709 uint i; 1710 while ((i = elements.next()) != 0) { 1711 tty->print("L%d ", Find_const(i)); 1712 } 1713 tty->print_cr("}"); 1714 } 1715 tty->print("\n"); 1716 } 1717 1718 void PhaseChaitin::dump() const { 1719 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 1720 _matcher._new_SP, _framesize ); 1721 1722 // For all blocks 1723 for( uint i = 0; i < _cfg._num_blocks; i++ ) 1724 dump(_cfg._blocks[i]); 1725 // End of per-block dump 1726 tty->print("\n"); 1727 1728 if (!_ifg) { 1729 tty->print("(No IFG.)\n"); 1730 return; 1731 } 1732 1733 // Dump LRG array 1734 tty->print("--- Live RanGe Array ---\n"); 1735 for(uint i2 = 1; i2 < _maxlrg; i2++ ) { 1736 tty->print("L%d: ",i2); 1737 if( i2 < _ifg->_maxlrg ) lrgs(i2).dump( ); 1738 else tty->print("new LRG"); 1739 } 1740 tty->print_cr(""); 1741 1742 // Dump lo-degree list 1743 tty->print("Lo degree: "); 1744 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 1745 tty->print("L%d ",i3); 1746 tty->print_cr(""); 1747 1748 // Dump lo-stk-degree list 1749 tty->print("Lo stk degree: "); 1750 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 1751 tty->print("L%d ",i4); 1752 tty->print_cr(""); 1753 1754 // Dump lo-degree list 1755 tty->print("Hi degree: "); 1756 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 1757 tty->print("L%d ",i5); 1758 tty->print_cr(""); 1759 } 1760 1761 //------------------------------dump_degree_lists------------------------------ 1762 void PhaseChaitin::dump_degree_lists() const { 1763 // Dump lo-degree list 1764 tty->print("Lo degree: "); 1765 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 1766 tty->print("L%d ",i); 1767 tty->print_cr(""); 1768 1769 // Dump lo-stk-degree list 1770 tty->print("Lo stk degree: "); 1771 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 1772 tty->print("L%d ",i2); 1773 tty->print_cr(""); 1774 1775 // Dump lo-degree list 1776 tty->print("Hi degree: "); 1777 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 1778 tty->print("L%d ",i3); 1779 tty->print_cr(""); 1780 } 1781 1782 //------------------------------dump_simplified-------------------------------- 1783 void PhaseChaitin::dump_simplified() const { 1784 tty->print("Simplified: "); 1785 for( uint i = _simplified; i; i = lrgs(i)._next ) 1786 tty->print("L%d ",i); 1787 tty->print_cr(""); 1788 } 1789 1790 static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { 1791 if ((int)reg < 0) 1792 sprintf(buf, "<OptoReg::%d>", (int)reg); 1793 else if (OptoReg::is_reg(reg)) 1794 strcpy(buf, Matcher::regName[reg]); 1795 else 1796 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 1797 pc->reg2offset(reg)); 1798 return buf+strlen(buf); 1799 } 1800 1801 //------------------------------dump_register---------------------------------- 1802 // Dump a register name into a buffer. Be intelligent if we get called 1803 // before allocation is complete. 1804 char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { 1805 if( !this ) { // Not got anything? 1806 sprintf(buf,"N%d",n->_idx); // Then use Node index 1807 } else if( _node_regs ) { 1808 // Post allocation, use direct mappings, no LRG info available 1809 print_reg( get_reg_first(n), this, buf ); 1810 } else { 1811 uint lidx = Find_const(n); // Grab LRG number 1812 if( !_ifg ) { 1813 sprintf(buf,"L%d",lidx); // No register binding yet 1814 } else if( !lidx ) { // Special, not allocated value 1815 strcpy(buf,"Special"); 1816 } else if( (lrgs(lidx).num_regs() == 1) 1817 ? !lrgs(lidx).mask().is_bound1() 1818 : !lrgs(lidx).mask().is_bound2() ) { 1819 sprintf(buf,"L%d",lidx); // No register binding yet 1820 } else { // Hah! We have a bound machine register 1821 print_reg( lrgs(lidx).reg(), this, buf ); 1822 } 1823 } 1824 return buf+strlen(buf); 1825 } 1826 1827 //----------------------dump_for_spill_split_recycle-------------------------- 1828 void PhaseChaitin::dump_for_spill_split_recycle() const { 1829 if( WizardMode && (PrintCompilation || PrintOpto) ) { 1830 // Display which live ranges need to be split and the allocator's state 1831 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 1832 for( uint bidx = 1; bidx < _maxlrg; bidx++ ) { 1833 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 1834 tty->print("L%d: ", bidx); 1835 lrgs(bidx).dump(); 1836 } 1837 } 1838 tty->cr(); 1839 dump(); 1840 } 1841 } 1842 1843 //------------------------------dump_frame------------------------------------ 1844 void PhaseChaitin::dump_frame() const { 1845 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 1846 const TypeTuple *domain = C->tf()->domain(); 1847 const int argcnt = domain->cnt() - TypeFunc::Parms; 1848 1849 // Incoming arguments in registers dump 1850 for( int k = 0; k < argcnt; k++ ) { 1851 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 1852 if( OptoReg::is_reg(parmreg)) { 1853 const char *reg_name = OptoReg::regname(parmreg); 1854 tty->print("#r%3.3d %s", parmreg, reg_name); 1855 parmreg = _matcher._parm_regs[k].second(); 1856 if( OptoReg::is_reg(parmreg)) { 1857 tty->print(":%s", OptoReg::regname(parmreg)); 1858 } 1859 tty->print(" : parm %d: ", k); 1860 domain->field_at(k + TypeFunc::Parms)->dump(); 1861 tty->print_cr(""); 1862 } 1863 } 1864 1865 // Check for un-owned padding above incoming args 1866 OptoReg::Name reg = _matcher._new_SP; 1867 if( reg > _matcher._in_arg_limit ) { 1868 reg = OptoReg::add(reg, -1); 1869 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 1870 } 1871 1872 // Incoming argument area dump 1873 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 1874 while( reg > begin_in_arg ) { 1875 reg = OptoReg::add(reg, -1); 1876 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 1877 int j; 1878 for( j = 0; j < argcnt; j++) { 1879 if( _matcher._parm_regs[j].first() == reg || 1880 _matcher._parm_regs[j].second() == reg ) { 1881 tty->print("parm %d: ",j); 1882 domain->field_at(j + TypeFunc::Parms)->dump(); 1883 tty->print_cr(""); 1884 break; 1885 } 1886 } 1887 if( j >= argcnt ) 1888 tty->print_cr("HOLE, owned by SELF"); 1889 } 1890 1891 // Old outgoing preserve area 1892 while( reg > _matcher._old_SP ) { 1893 reg = OptoReg::add(reg, -1); 1894 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 1895 } 1896 1897 // Old SP 1898 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 1899 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 1900 1901 // Preserve area dump 1902 reg = OptoReg::add(reg, -1); 1903 while( OptoReg::is_stack(reg)) { 1904 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 1905 if( _matcher.return_addr() == reg ) 1906 tty->print_cr("return address"); 1907 else if( _matcher.return_addr() == OptoReg::add(reg,1) && 1908 VerifyStackAtCalls ) 1909 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 1910 else if ((int)OptoReg::reg2stack(reg) < C->fixed_slots()) 1911 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 1912 else 1913 tty->print_cr("pad2, in_preserve"); 1914 reg = OptoReg::add(reg, -1); 1915 } 1916 1917 // Spill area dump 1918 reg = OptoReg::add(_matcher._new_SP, _framesize ); 1919 while( reg > _matcher._out_arg_limit ) { 1920 reg = OptoReg::add(reg, -1); 1921 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 1922 } 1923 1924 // Outgoing argument area dump 1925 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 1926 reg = OptoReg::add(reg, -1); 1927 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 1928 } 1929 1930 // Outgoing new preserve area 1931 while( reg > _matcher._new_SP ) { 1932 reg = OptoReg::add(reg, -1); 1933 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 1934 } 1935 tty->print_cr("#"); 1936 } 1937 1938 //------------------------------dump_bb---------------------------------------- 1939 void PhaseChaitin::dump_bb( uint pre_order ) const { 1940 tty->print_cr("---dump of B%d---",pre_order); 1941 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 1942 Block *b = _cfg._blocks[i]; 1943 if( b->_pre_order == pre_order ) 1944 dump(b); 1945 } 1946 } 1947 1948 //------------------------------dump_lrg--------------------------------------- 1949 void PhaseChaitin::dump_lrg( uint lidx ) const { 1950 tty->print_cr("---dump of L%d---",lidx); 1951 1952 if( _ifg ) { 1953 if( lidx >= _maxlrg ) { 1954 tty->print("Attempt to print live range index beyond max live range.\n"); 1955 return; 1956 } 1957 tty->print("L%d: ",lidx); 1958 lrgs(lidx).dump( ); 1959 } 1960 if( _ifg ) { tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 1961 _ifg->neighbors(lidx)->dump(); 1962 tty->cr(); 1963 } 1964 // For all blocks 1965 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 1966 Block *b = _cfg._blocks[i]; 1967 int dump_once = 0; 1968 1969 // For all instructions 1970 for( uint j = 0; j < b->_nodes.size(); j++ ) { 1971 Node *n = b->_nodes[j]; 1972 if( Find_const(n) == lidx ) { 1973 if( !dump_once++ ) { 1974 tty->cr(); 1975 b->dump_head( &_cfg._bbs ); 1976 } 1977 dump(n); 1978 continue; 1979 } 1980 uint cnt = n->req(); 1981 for( uint k = 1; k < cnt; k++ ) { 1982 Node *m = n->in(k); 1983 if (!m) continue; // be robust in the dumper 1984 if( Find_const(m) == lidx ) { 1985 if( !dump_once++ ) { 1986 tty->cr(); 1987 b->dump_head( &_cfg._bbs ); 1988 } 1989 dump(n); 1990 } 1991 } 1992 } 1993 } // End of per-block dump 1994 tty->cr(); 1995 } 1996 #endif // not PRODUCT 1997 1998 //------------------------------print_chaitin_statistics------------------------------- 1999 int PhaseChaitin::_final_loads = 0; 2000 int PhaseChaitin::_final_stores = 0; 2001 int PhaseChaitin::_final_memoves= 0; 2002 int PhaseChaitin::_final_copies = 0; 2003 double PhaseChaitin::_final_load_cost = 0; 2004 double PhaseChaitin::_final_store_cost = 0; 2005 double PhaseChaitin::_final_memove_cost= 0; 2006 double PhaseChaitin::_final_copy_cost = 0; 2007 int PhaseChaitin::_conserv_coalesce = 0; 2008 int PhaseChaitin::_conserv_coalesce_pair = 0; 2009 int PhaseChaitin::_conserv_coalesce_trie = 0; 2010 int PhaseChaitin::_conserv_coalesce_quad = 0; 2011 int PhaseChaitin::_post_alloc = 0; 2012 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2013 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2014 int PhaseChaitin::_used_cisc_instructions = 0; 2015 int PhaseChaitin::_unused_cisc_instructions = 0; 2016 int PhaseChaitin::_allocator_attempts = 0; 2017 int PhaseChaitin::_allocator_successes = 0; 2018 2019 #ifndef PRODUCT 2020 uint PhaseChaitin::_high_pressure = 0; 2021 uint PhaseChaitin::_low_pressure = 0; 2022 2023 void PhaseChaitin::print_chaitin_statistics() { 2024 tty->print_cr("Inserted %d spill loads, %d spill stores, %d mem-mem moves and %d copies.", _final_loads, _final_stores, _final_memoves, _final_copies); 2025 tty->print_cr("Total load cost= %6.0f, store cost = %6.0f, mem-mem cost = %5.2f, copy cost = %5.0f.", _final_load_cost, _final_store_cost, _final_memove_cost, _final_copy_cost); 2026 tty->print_cr("Adjusted spill cost = %7.0f.", 2027 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2028 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2029 tty->print("Conservatively coalesced %d copies, %d pairs", 2030 _conserv_coalesce, _conserv_coalesce_pair); 2031 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2032 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2033 tty->print_cr(", %d post alloc.", _post_alloc); 2034 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2035 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2036 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2037 if( _used_cisc_instructions || _unused_cisc_instructions ) 2038 tty->print_cr("Used cisc instruction %d, remained in register %d", 2039 _used_cisc_instructions, _unused_cisc_instructions); 2040 if( _allocator_successes != 0 ) 2041 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2042 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2043 } 2044 #endif // not PRODUCT