1 /* 2 * Copyright (c) 2000, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "compiler/oopMap.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/block.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/chaitin.hpp" 34 #include "opto/coalesce.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/idealGraphPrinter.hpp" 37 #include "opto/indexSet.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/movenode.hpp" 41 #include "opto/opcodes.hpp" 42 #include "opto/rootnode.hpp" 43 44 #ifndef PRODUCT 45 void LRG::dump() const { 46 ttyLocker ttyl; 47 tty->print("%d ",num_regs()); 48 _mask.dump(); 49 if( _msize_valid ) { 50 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size); 51 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size()); 52 } else { 53 tty->print(", #?(%d) ",_mask.Size()); 54 } 55 56 tty->print("EffDeg: "); 57 if( _degree_valid ) tty->print( "%d ", _eff_degree ); 58 else tty->print("? "); 59 60 if( is_multidef() ) { 61 tty->print("MultiDef "); 62 if (_defs != NULL) { 63 tty->print("("); 64 for (int i = 0; i < _defs->length(); i++) { 65 tty->print("N%d ", _defs->at(i)->_idx); 66 } 67 tty->print(") "); 68 } 69 } 70 else if( _def == 0 ) tty->print("Dead "); 71 else tty->print("Def: N%d ",_def->_idx); 72 73 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score()); 74 // Flags 75 if( _is_oop ) tty->print("Oop "); 76 if( _is_float ) tty->print("Float "); 77 if( _is_vector ) tty->print("Vector "); 78 if( _was_spilled1 ) tty->print("Spilled "); 79 if( _was_spilled2 ) tty->print("Spilled2 "); 80 if( _direct_conflict ) tty->print("Direct_conflict "); 81 if( _fat_proj ) tty->print("Fat "); 82 if( _was_lo ) tty->print("Lo "); 83 if( _has_copy ) tty->print("Copy "); 84 if( _at_risk ) tty->print("Risk "); 85 86 if( _must_spill ) tty->print("Must_spill "); 87 if( _is_bound ) tty->print("Bound "); 88 if( _msize_valid ) { 89 if( _degree_valid && lo_degree() ) tty->print("Trivial "); 90 } 91 92 tty->cr(); 93 } 94 #endif 95 96 // Compute score from cost and area. Low score is best to spill. 97 static double raw_score( double cost, double area ) { 98 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; 99 } 100 101 double LRG::score() const { 102 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. 103 // Bigger area lowers score, encourages spilling this live range. 104 // Bigger cost raise score, prevents spilling this live range. 105 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer 106 // to turn a divide by a constant into a multiply by the reciprical). 107 double score = raw_score( _cost, _area); 108 109 // Account for area. Basically, LRGs covering large areas are better 110 // to spill because more other LRGs get freed up. 111 if( _area == 0.0 ) // No area? Then no progress to spill 112 return 1e35; 113 114 if( _was_spilled2 ) // If spilled once before, we are unlikely 115 return score + 1e30; // to make progress again. 116 117 if( _cost >= _area*3.0 ) // Tiny area relative to cost 118 return score + 1e17; // Probably no progress to spill 119 120 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost 121 return score + 1e10; // Likely no progress to spill 122 123 return score; 124 } 125 126 #define NUMBUCKS 3 127 128 // Straight out of Tarjan's union-find algorithm 129 uint LiveRangeMap::find_compress(uint lrg) { 130 uint cur = lrg; 131 uint next = _uf_map.at(cur); 132 while (next != cur) { // Scan chain of equivalences 133 assert( next < cur, "always union smaller"); 134 cur = next; // until find a fixed-point 135 next = _uf_map.at(cur); 136 } 137 138 // Core of union-find algorithm: update chain of 139 // equivalences to be equal to the root. 140 while (lrg != next) { 141 uint tmp = _uf_map.at(lrg); 142 _uf_map.at_put(lrg, next); 143 lrg = tmp; 144 } 145 return lrg; 146 } 147 148 // Reset the Union-Find map to identity 149 void LiveRangeMap::reset_uf_map(uint max_lrg_id) { 150 _max_lrg_id= max_lrg_id; 151 // Force the Union-Find mapping to be at least this large 152 _uf_map.at_put_grow(_max_lrg_id, 0); 153 // Initialize it to be the ID mapping. 154 for (uint i = 0; i < _max_lrg_id; ++i) { 155 _uf_map.at_put(i, i); 156 } 157 } 158 159 // Make all Nodes map directly to their final live range; no need for 160 // the Union-Find mapping after this call. 161 void LiveRangeMap::compress_uf_map_for_nodes() { 162 // For all Nodes, compress mapping 163 uint unique = _names.length(); 164 for (uint i = 0; i < unique; ++i) { 165 uint lrg = _names.at(i); 166 uint compressed_lrg = find(lrg); 167 if (lrg != compressed_lrg) { 168 _names.at_put(i, compressed_lrg); 169 } 170 } 171 } 172 173 // Like Find above, but no path compress, so bad asymptotic behavior 174 uint LiveRangeMap::find_const(uint lrg) const { 175 if (!lrg) { 176 return lrg; // Ignore the zero LRG 177 } 178 179 // Off the end? This happens during debugging dumps when you got 180 // brand new live ranges but have not told the allocator yet. 181 if (lrg >= _max_lrg_id) { 182 return lrg; 183 } 184 185 uint next = _uf_map.at(lrg); 186 while (next != lrg) { // Scan chain of equivalences 187 assert(next < lrg, "always union smaller"); 188 lrg = next; // until find a fixed-point 189 next = _uf_map.at(lrg); 190 } 191 return next; 192 } 193 194 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher) 195 : PhaseRegAlloc(unique, cfg, matcher, 196 #ifndef PRODUCT 197 print_chaitin_statistics 198 #else 199 NULL 200 #endif 201 ) 202 , _lrg_map(Thread::current()->resource_area(), unique) 203 , _live(0) 204 , _spilled_once(Thread::current()->resource_area()) 205 , _spilled_twice(Thread::current()->resource_area()) 206 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0) 207 , _oldphi(unique) 208 #ifndef PRODUCT 209 , _trace_spilling(TraceSpilling || C->method_has_option("TraceSpilling")) 210 #endif 211 { 212 NOT_PRODUCT( Compile::TracePhase t3("ctorChaitin", &_t_ctorChaitin, TimeCompiler); ) 213 214 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency()); 215 216 // Build a list of basic blocks, sorted by frequency 217 _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 218 // Experiment with sorting strategies to speed compilation 219 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket 220 Block **buckets[NUMBUCKS]; // Array of buckets 221 uint buckcnt[NUMBUCKS]; // Array of bucket counters 222 double buckval[NUMBUCKS]; // Array of bucket value cutoffs 223 for (uint i = 0; i < NUMBUCKS; i++) { 224 buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 225 buckcnt[i] = 0; 226 // Bump by three orders of magnitude each time 227 cutoff *= 0.001; 228 buckval[i] = cutoff; 229 for (uint j = 0; j < _cfg.number_of_blocks(); j++) { 230 buckets[i][j] = NULL; 231 } 232 } 233 // Sort blocks into buckets 234 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 235 for (uint j = 0; j < NUMBUCKS; j++) { 236 if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) { 237 // Assign block to end of list for appropriate bucket 238 buckets[j][buckcnt[j]++] = _cfg.get_block(i); 239 break; // kick out of inner loop 240 } 241 } 242 } 243 // Dump buckets into final block array 244 uint blkcnt = 0; 245 for (uint i = 0; i < NUMBUCKS; i++) { 246 for (uint j = 0; j < buckcnt[i]; j++) { 247 _blks[blkcnt++] = buckets[i][j]; 248 } 249 } 250 251 assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled"); 252 } 253 254 // union 2 sets together. 255 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) { 256 uint src = _lrg_map.find(src_n); 257 uint dst = _lrg_map.find(dst_n); 258 assert(src, ""); 259 assert(dst, ""); 260 assert(src < _lrg_map.max_lrg_id(), "oob"); 261 assert(dst < _lrg_map.max_lrg_id(), "oob"); 262 assert(src < dst, "always union smaller"); 263 _lrg_map.uf_map(dst, src); 264 } 265 266 void PhaseChaitin::new_lrg(const Node *x, uint lrg) { 267 // Make the Node->LRG mapping 268 _lrg_map.extend(x->_idx,lrg); 269 // Make the Union-Find mapping an identity function 270 _lrg_map.uf_extend(lrg, lrg); 271 } 272 273 274 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) { 275 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections"); 276 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); ) 277 int found_projs = 0; 278 uint cnt = orig->outcnt(); 279 for (uint i = 0; i < cnt; i++) { 280 Node* proj = orig->raw_out(i); 281 if (proj->is_MachProj()) { 282 assert(proj->outcnt() == 0, "only kill projections are expected here"); 283 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections"); 284 found_projs++; 285 // Copy kill projections after the cloned node 286 Node* kills = proj->clone(); 287 kills->set_req(0, copy); 288 b->insert_node(kills, idx++); 289 _cfg.map_node_to_block(kills, b); 290 new_lrg(kills, max_lrg_id++); 291 } 292 } 293 return found_projs; 294 } 295 296 // Renumber the live ranges to compact them. Makes the IFG smaller. 297 void PhaseChaitin::compact() { 298 // Current the _uf_map contains a series of short chains which are headed 299 // by a self-cycle. All the chains run from big numbers to little numbers. 300 // The Find() call chases the chains & shortens them for the next Find call. 301 // We are going to change this structure slightly. Numbers above a moving 302 // wave 'i' are unchanged. Numbers below 'j' point directly to their 303 // compacted live range with no further chaining. There are no chains or 304 // cycles below 'i', so the Find call no longer works. 305 uint j=1; 306 uint i; 307 for (i = 1; i < _lrg_map.max_lrg_id(); i++) { 308 uint lr = _lrg_map.uf_live_range_id(i); 309 // Ignore unallocated live ranges 310 if (!lr) { 311 continue; 312 } 313 assert(lr <= i, ""); 314 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr)); 315 } 316 // Now change the Node->LR mapping to reflect the compacted names 317 uint unique = _lrg_map.size(); 318 for (i = 0; i < unique; i++) { 319 uint lrg_id = _lrg_map.live_range_id(i); 320 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id)); 321 } 322 323 // Reset the Union-Find mapping 324 _lrg_map.reset_uf_map(j); 325 } 326 327 void PhaseChaitin::Register_Allocate() { 328 329 // Above the OLD FP (and in registers) are the incoming arguments. Stack 330 // slots in this area are called "arg_slots". Above the NEW FP (and in 331 // registers) is the outgoing argument area; above that is the spill/temp 332 // area. These are all "frame_slots". Arg_slots start at the zero 333 // stack_slots and count up to the known arg_size. Frame_slots start at 334 // the stack_slot #arg_size and go up. After allocation I map stack 335 // slots to actual offsets. Stack-slots in the arg_slot area are biased 336 // by the frame_size; stack-slots in the frame_slot area are biased by 0. 337 338 _trip_cnt = 0; 339 _alternate = 0; 340 _matcher._allocation_started = true; 341 342 ResourceArea split_arena; // Arena for Split local resources 343 ResourceArea live_arena; // Arena for liveness & IFG info 344 ResourceMark rm(&live_arena); 345 346 // Need live-ness for the IFG; need the IFG for coalescing. If the 347 // liveness is JUST for coalescing, then I can get some mileage by renaming 348 // all copy-related live ranges low and then using the max copy-related 349 // live range as a cut-off for LIVE and the IFG. In other words, I can 350 // build a subset of LIVE and IFG just for copies. 351 PhaseLive live(_cfg, _lrg_map.names(), &live_arena); 352 353 // Need IFG for coalescing and coloring 354 PhaseIFG ifg(&live_arena); 355 _ifg = &ifg; 356 357 // Come out of SSA world to the Named world. Assign (virtual) registers to 358 // Nodes. Use the same register for all inputs and the output of PhiNodes 359 // - effectively ending SSA form. This requires either coalescing live 360 // ranges or inserting copies. For the moment, we insert "virtual copies" 361 // - we pretend there is a copy prior to each Phi in predecessor blocks. 362 // We will attempt to coalesce such "virtual copies" before we manifest 363 // them for real. 364 de_ssa(); 365 366 #ifdef ASSERT 367 // Veify the graph before RA. 368 verify(&live_arena); 369 #endif 370 371 { 372 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 373 _live = NULL; // Mark live as being not available 374 rm.reset_to_mark(); // Reclaim working storage 375 IndexSet::reset_memory(C, &live_arena); 376 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG 377 gather_lrg_masks( false ); // Collect LRG masks 378 live.compute(_lrg_map.max_lrg_id()); // Compute liveness 379 _live = &live; // Mark LIVE as being available 380 } 381 382 // Base pointers are currently "used" by instructions which define new 383 // derived pointers. This makes base pointers live up to the where the 384 // derived pointer is made, but not beyond. Really, they need to be live 385 // across any GC point where the derived value is live. So this code looks 386 // at all the GC points, and "stretches" the live range of any base pointer 387 // to the GC point. 388 if (stretch_base_pointer_live_ranges(&live_arena)) { 389 NOT_PRODUCT(Compile::TracePhase t3("computeLive (sbplr)", &_t_computeLive, TimeCompiler);) 390 // Since some live range stretched, I need to recompute live 391 _live = NULL; 392 rm.reset_to_mark(); // Reclaim working storage 393 IndexSet::reset_memory(C, &live_arena); 394 ifg.init(_lrg_map.max_lrg_id()); 395 gather_lrg_masks(false); 396 live.compute(_lrg_map.max_lrg_id()); 397 _live = &live; 398 } 399 // Create the interference graph using virtual copies 400 build_ifg_virtual(); // Include stack slots this time 401 402 // Aggressive (but pessimistic) copy coalescing. 403 // This pass works on virtual copies. Any virtual copies which are not 404 // coalesced get manifested as actual copies 405 { 406 // The IFG is/was triangular. I am 'squaring it up' so Union can run 407 // faster. Union requires a 'for all' operation which is slow on the 408 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - 409 // meaning I can visit all the Nodes neighbors less than a Node in time 410 // O(# of neighbors), but I have to visit all the Nodes greater than a 411 // given Node and search them for an instance, i.e., time O(#MaxLRG)). 412 _ifg->SquareUp(); 413 414 PhaseAggressiveCoalesce coalesce(*this); 415 coalesce.coalesce_driver(); 416 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do 417 // not match the Phi itself, insert a copy. 418 coalesce.insert_copies(_matcher); 419 if (C->failing()) { 420 return; 421 } 422 } 423 424 // After aggressive coalesce, attempt a first cut at coloring. 425 // To color, we need the IFG and for that we need LIVE. 426 { 427 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 428 _live = NULL; 429 rm.reset_to_mark(); // Reclaim working storage 430 IndexSet::reset_memory(C, &live_arena); 431 ifg.init(_lrg_map.max_lrg_id()); 432 gather_lrg_masks( true ); 433 live.compute(_lrg_map.max_lrg_id()); 434 _live = &live; 435 } 436 437 // Build physical interference graph 438 uint must_spill = 0; 439 must_spill = build_ifg_physical(&live_arena); 440 // If we have a guaranteed spill, might as well spill now 441 if (must_spill) { 442 if(!_lrg_map.max_lrg_id()) { 443 return; 444 } 445 // Bail out if unique gets too large (ie - unique > MaxNodeLimit) 446 C->check_node_count(10*must_spill, "out of nodes before split"); 447 if (C->failing()) { 448 return; 449 } 450 451 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 452 _lrg_map.set_max_lrg_id(new_max_lrg_id); 453 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 454 // or we failed to split 455 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); 456 if (C->failing()) { 457 return; 458 } 459 460 NOT_PRODUCT(C->verify_graph_edges();) 461 462 compact(); // Compact LRGs; return new lower max lrg 463 464 { 465 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 466 _live = NULL; 467 rm.reset_to_mark(); // Reclaim working storage 468 IndexSet::reset_memory(C, &live_arena); 469 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph 470 gather_lrg_masks( true ); // Collect intersect mask 471 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE 472 _live = &live; 473 } 474 build_ifg_physical(&live_arena); 475 _ifg->SquareUp(); 476 _ifg->Compute_Effective_Degree(); 477 // Only do conservative coalescing if requested 478 if (OptoCoalesce) { 479 // Conservative (and pessimistic) copy coalescing of those spills 480 PhaseConservativeCoalesce coalesce(*this); 481 // If max live ranges greater than cutoff, don't color the stack. 482 // This cutoff can be larger than below since it is only done once. 483 coalesce.coalesce_driver(); 484 } 485 _lrg_map.compress_uf_map_for_nodes(); 486 487 #ifdef ASSERT 488 verify(&live_arena, true); 489 #endif 490 } else { 491 ifg.SquareUp(); 492 ifg.Compute_Effective_Degree(); 493 #ifdef ASSERT 494 set_was_low(); 495 #endif 496 } 497 498 // Prepare for Simplify & Select 499 cache_lrg_info(); // Count degree of LRGs 500 501 // Simplify the InterFerence Graph by removing LRGs of low degree. 502 // LRGs of low degree are trivially colorable. 503 Simplify(); 504 505 // Select colors by re-inserting LRGs back into the IFG in reverse order. 506 // Return whether or not something spills. 507 uint spills = Select( ); 508 509 // If we spill, split and recycle the entire thing 510 while( spills ) { 511 if( _trip_cnt++ > 24 ) { 512 DEBUG_ONLY( dump_for_spill_split_recycle(); ) 513 if( _trip_cnt > 27 ) { 514 C->record_method_not_compilable("failed spill-split-recycle sanity check"); 515 return; 516 } 517 } 518 519 if (!_lrg_map.max_lrg_id()) { 520 return; 521 } 522 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 523 _lrg_map.set_max_lrg_id(new_max_lrg_id); 524 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 525 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split"); 526 if (C->failing()) { 527 return; 528 } 529 530 compact(); // Compact LRGs; return new lower max lrg 531 532 // Nuke the live-ness and interference graph and LiveRanGe info 533 { 534 NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) 535 _live = NULL; 536 rm.reset_to_mark(); // Reclaim working storage 537 IndexSet::reset_memory(C, &live_arena); 538 ifg.init(_lrg_map.max_lrg_id()); 539 540 // Create LiveRanGe array. 541 // Intersect register masks for all USEs and DEFs 542 gather_lrg_masks(true); 543 live.compute(_lrg_map.max_lrg_id()); 544 _live = &live; 545 } 546 must_spill = build_ifg_physical(&live_arena); 547 _ifg->SquareUp(); 548 _ifg->Compute_Effective_Degree(); 549 550 // Only do conservative coalescing if requested 551 if (OptoCoalesce) { 552 // Conservative (and pessimistic) copy coalescing 553 PhaseConservativeCoalesce coalesce(*this); 554 // Check for few live ranges determines how aggressive coalesce is. 555 coalesce.coalesce_driver(); 556 } 557 _lrg_map.compress_uf_map_for_nodes(); 558 #ifdef ASSERT 559 verify(&live_arena, true); 560 #endif 561 cache_lrg_info(); // Count degree of LRGs 562 563 // Simplify the InterFerence Graph by removing LRGs of low degree. 564 // LRGs of low degree are trivially colorable. 565 Simplify(); 566 567 // Select colors by re-inserting LRGs back into the IFG in reverse order. 568 // Return whether or not something spills. 569 spills = Select(); 570 } 571 572 // Count number of Simplify-Select trips per coloring success. 573 _allocator_attempts += _trip_cnt + 1; 574 _allocator_successes += 1; 575 576 // Peephole remove copies 577 post_allocate_copy_removal(); 578 579 #ifdef ASSERT 580 // Veify the graph after RA. 581 verify(&live_arena); 582 #endif 583 584 // max_reg is past the largest *register* used. 585 // Convert that to a frame_slot number. 586 if (_max_reg <= _matcher._new_SP) { 587 _framesize = C->out_preserve_stack_slots(); 588 } 589 else { 590 _framesize = _max_reg -_matcher._new_SP; 591 } 592 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); 593 594 // This frame must preserve the required fp alignment 595 _framesize = round_to(_framesize, Matcher::stack_alignment_in_slots()); 596 assert( _framesize >= 0 && _framesize <= 1000000, "sanity check" ); 597 #ifndef PRODUCT 598 _total_framesize += _framesize; 599 if ((int)_framesize > _max_framesize) { 600 _max_framesize = _framesize; 601 } 602 #endif 603 604 // Convert CISC spills 605 fixup_spills(); 606 607 // Log regalloc results 608 CompileLog* log = Compile::current()->log(); 609 if (log != NULL) { 610 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); 611 } 612 613 if (C->failing()) { 614 return; 615 } 616 617 NOT_PRODUCT(C->verify_graph_edges();) 618 619 // Move important info out of the live_arena to longer lasting storage. 620 alloc_node_regs(_lrg_map.size()); 621 for (uint i=0; i < _lrg_map.size(); i++) { 622 if (_lrg_map.live_range_id(i)) { // Live range associated with Node? 623 LRG &lrg = lrgs(_lrg_map.live_range_id(i)); 624 if (!lrg.alive()) { 625 set_bad(i); 626 } else if (lrg.num_regs() == 1) { 627 set1(i, lrg.reg()); 628 } else { // Must be a register-set 629 if (!lrg._fat_proj) { // Must be aligned adjacent register set 630 // Live ranges record the highest register in their mask. 631 // We want the low register for the AD file writer's convenience. 632 OptoReg::Name hi = lrg.reg(); // Get hi register 633 OptoReg::Name lo = OptoReg::add(hi, (1-lrg.num_regs())); // Find lo 634 // We have to use pair [lo,lo+1] even for wide vectors because 635 // the rest of code generation works only with pairs. It is safe 636 // since for registers encoding only 'lo' is used. 637 // Second reg from pair is used in ScheduleAndBundle on SPARC where 638 // vector max size is 8 which corresponds to registers pair. 639 // It is also used in BuildOopMaps but oop operations are not 640 // vectorized. 641 set2(i, lo); 642 } else { // Misaligned; extract 2 bits 643 OptoReg::Name hi = lrg.reg(); // Get hi register 644 lrg.Remove(hi); // Yank from mask 645 int lo = lrg.mask().find_first_elem(); // Find lo 646 set_pair(i, hi, lo); 647 } 648 } 649 if( lrg._is_oop ) _node_oops.set(i); 650 } else { 651 set_bad(i); 652 } 653 } 654 655 // Done! 656 _live = NULL; 657 _ifg = NULL; 658 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope 659 } 660 661 void PhaseChaitin::de_ssa() { 662 // Set initial Names for all Nodes. Most Nodes get the virtual register 663 // number. A few get the ZERO live range number. These do not 664 // get allocated, but instead rely on correct scheduling to ensure that 665 // only one instance is simultaneously live at a time. 666 uint lr_counter = 1; 667 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 668 Block* block = _cfg.get_block(i); 669 uint cnt = block->number_of_nodes(); 670 671 // Handle all the normal Nodes in the block 672 for( uint j = 0; j < cnt; j++ ) { 673 Node *n = block->get_node(j); 674 // Pre-color to the zero live range, or pick virtual register 675 const RegMask &rm = n->out_RegMask(); 676 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0); 677 } 678 } 679 680 // Reset the Union-Find mapping to be identity 681 _lrg_map.reset_uf_map(lr_counter); 682 } 683 684 685 // Gather LiveRanGe information, including register masks. Modification of 686 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. 687 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { 688 689 // Nail down the frame pointer live range 690 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr)); 691 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite 692 693 // For all blocks 694 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 695 Block* block = _cfg.get_block(i); 696 697 // For all instructions 698 for (uint j = 1; j < block->number_of_nodes(); j++) { 699 Node* n = block->get_node(j); 700 uint input_edge_start =1; // Skip control most nodes 701 if (n->is_Mach()) { 702 input_edge_start = n->as_Mach()->oper_input_base(); 703 } 704 uint idx = n->is_Copy(); 705 706 // Get virtual register number, same as LiveRanGe index 707 uint vreg = _lrg_map.live_range_id(n); 708 LRG& lrg = lrgs(vreg); 709 if (vreg) { // No vreg means un-allocable (e.g. memory) 710 711 // Collect has-copy bit 712 if (idx) { 713 lrg._has_copy = 1; 714 uint clidx = _lrg_map.live_range_id(n->in(idx)); 715 LRG& copy_src = lrgs(clidx); 716 copy_src._has_copy = 1; 717 } 718 719 // Check for float-vs-int live range (used in register-pressure 720 // calculations) 721 const Type *n_type = n->bottom_type(); 722 if (n_type->is_floatingpoint()) { 723 lrg._is_float = 1; 724 } 725 726 // Check for twice prior spilling. Once prior spilling might have 727 // spilled 'soft', 2nd prior spill should have spilled 'hard' and 728 // further spilling is unlikely to make progress. 729 if (_spilled_once.test(n->_idx)) { 730 lrg._was_spilled1 = 1; 731 if (_spilled_twice.test(n->_idx)) { 732 lrg._was_spilled2 = 1; 733 } 734 } 735 736 #ifndef PRODUCT 737 if (trace_spilling() && lrg._def != NULL) { 738 // collect defs for MultiDef printing 739 if (lrg._defs == NULL) { 740 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL); 741 lrg._defs->append(lrg._def); 742 } 743 lrg._defs->append(n); 744 } 745 #endif 746 747 // Check for a single def LRG; these can spill nicely 748 // via rematerialization. Flag as NULL for no def found 749 // yet, or 'n' for single def or -1 for many defs. 750 lrg._def = lrg._def ? NodeSentinel : n; 751 752 // Limit result register mask to acceptable registers 753 const RegMask &rm = n->out_RegMask(); 754 lrg.AND( rm ); 755 756 int ireg = n->ideal_reg(); 757 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, 758 "oops must be in Op_RegP's" ); 759 760 // Check for vector live range (only if vector register is used). 761 // On SPARC vector uses RegD which could be misaligned so it is not 762 // processes as vector in RA. 763 if (RegMask::is_vector(ireg)) 764 lrg._is_vector = 1; 765 assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL, 766 "vector must be in vector registers"); 767 768 // Check for bound register masks 769 const RegMask &lrgmask = lrg.mask(); 770 if (lrgmask.is_bound(ireg)) { 771 lrg._is_bound = 1; 772 } 773 774 // Check for maximum frequency value 775 if (lrg._maxfreq < block->_freq) { 776 lrg._maxfreq = block->_freq; 777 } 778 779 // Check for oop-iness, or long/double 780 // Check for multi-kill projection 781 switch (ireg) { 782 case MachProjNode::fat_proj: 783 // Fat projections have size equal to number of registers killed 784 lrg.set_num_regs(rm.Size()); 785 lrg.set_reg_pressure(lrg.num_regs()); 786 lrg._fat_proj = 1; 787 lrg._is_bound = 1; 788 break; 789 case Op_RegP: 790 #ifdef _LP64 791 lrg.set_num_regs(2); // Size is 2 stack words 792 #else 793 lrg.set_num_regs(1); // Size is 1 stack word 794 #endif 795 // Register pressure is tracked relative to the maximum values 796 // suggested for that platform, INTPRESSURE and FLOATPRESSURE, 797 // and relative to other types which compete for the same regs. 798 // 799 // The following table contains suggested values based on the 800 // architectures as defined in each .ad file. 801 // INTPRESSURE and FLOATPRESSURE may be tuned differently for 802 // compile-speed or performance. 803 // Note1: 804 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 805 // since .ad registers are defined as high and low halves. 806 // These reg_pressure values remain compatible with the code 807 // in is_high_pressure() which relates get_invalid_mask_size(), 808 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. 809 // Note2: 810 // SPARC -d32 has 24 registers available for integral values, 811 // but only 10 of these are safe for 64-bit longs. 812 // Using set_reg_pressure(2) for both int and long means 813 // the allocator will believe it can fit 26 longs into 814 // registers. Using 2 for longs and 1 for ints means the 815 // allocator will attempt to put 52 integers into registers. 816 // The settings below limit this problem to methods with 817 // many long values which are being run on 32-bit SPARC. 818 // 819 // ------------------- reg_pressure -------------------- 820 // Each entry is reg_pressure_per_value,number_of_regs 821 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE 822 // IA32 2 1 1 1 1 6 6 823 // IA64 1 1 1 1 1 50 41 824 // SPARC 2 2 2 2 2 48 (24) 52 (26) 825 // SPARCV9 2 2 2 2 2 48 (24) 52 (26) 826 // AMD64 1 1 1 1 1 14 15 827 // ----------------------------------------------------- 828 #if defined(SPARC) 829 lrg.set_reg_pressure(2); // use for v9 as well 830 #else 831 lrg.set_reg_pressure(1); // normally one value per register 832 #endif 833 if( n_type->isa_oop_ptr() ) { 834 lrg._is_oop = 1; 835 } 836 break; 837 case Op_RegL: // Check for long or double 838 case Op_RegD: 839 lrg.set_num_regs(2); 840 // Define platform specific register pressure 841 #if defined(SPARC) || defined(ARM) 842 lrg.set_reg_pressure(2); 843 #elif defined(IA32) 844 if( ireg == Op_RegL ) { 845 lrg.set_reg_pressure(2); 846 } else { 847 lrg.set_reg_pressure(1); 848 } 849 #else 850 lrg.set_reg_pressure(1); // normally one value per register 851 #endif 852 // If this def of a double forces a mis-aligned double, 853 // flag as '_fat_proj' - really flag as allowing misalignment 854 // AND changes how we count interferences. A mis-aligned 855 // double can interfere with TWO aligned pairs, or effectively 856 // FOUR registers! 857 if (rm.is_misaligned_pair()) { 858 lrg._fat_proj = 1; 859 lrg._is_bound = 1; 860 } 861 break; 862 case Op_RegF: 863 case Op_RegI: 864 case Op_RegN: 865 case Op_RegFlags: 866 case 0: // not an ideal register 867 lrg.set_num_regs(1); 868 #ifdef SPARC 869 lrg.set_reg_pressure(2); 870 #else 871 lrg.set_reg_pressure(1); 872 #endif 873 break; 874 case Op_VecS: 875 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity"); 876 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity"); 877 lrg.set_num_regs(RegMask::SlotsPerVecS); 878 lrg.set_reg_pressure(1); 879 break; 880 case Op_VecD: 881 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity"); 882 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity"); 883 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned"); 884 lrg.set_num_regs(RegMask::SlotsPerVecD); 885 lrg.set_reg_pressure(1); 886 break; 887 case Op_VecX: 888 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity"); 889 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity"); 890 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned"); 891 lrg.set_num_regs(RegMask::SlotsPerVecX); 892 lrg.set_reg_pressure(1); 893 break; 894 case Op_VecY: 895 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity"); 896 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity"); 897 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned"); 898 lrg.set_num_regs(RegMask::SlotsPerVecY); 899 lrg.set_reg_pressure(1); 900 break; 901 default: 902 ShouldNotReachHere(); 903 } 904 } 905 906 // Now do the same for inputs 907 uint cnt = n->req(); 908 // Setup for CISC SPILLING 909 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 910 if( UseCISCSpill && after_aggressive ) { 911 inp = n->cisc_operand(); 912 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 913 // Convert operand number to edge index number 914 inp = n->as_Mach()->operand_index(inp); 915 } 916 // Prepare register mask for each input 917 for( uint k = input_edge_start; k < cnt; k++ ) { 918 uint vreg = _lrg_map.live_range_id(n->in(k)); 919 if (!vreg) { 920 continue; 921 } 922 923 // If this instruction is CISC Spillable, add the flags 924 // bit to its appropriate input 925 if( UseCISCSpill && after_aggressive && inp == k ) { 926 #ifndef PRODUCT 927 if( TraceCISCSpill ) { 928 tty->print(" use_cisc_RegMask: "); 929 n->dump(); 930 } 931 #endif 932 n->as_Mach()->use_cisc_RegMask(); 933 } 934 935 LRG &lrg = lrgs(vreg); 936 // // Testing for floating point code shape 937 // Node *test = n->in(k); 938 // if( test->is_Mach() ) { 939 // MachNode *m = test->as_Mach(); 940 // int op = m->ideal_Opcode(); 941 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 942 // int zzz = 1; 943 // } 944 // } 945 946 // Limit result register mask to acceptable registers. 947 // Do not limit registers from uncommon uses before 948 // AggressiveCoalesce. This effectively pre-virtual-splits 949 // around uncommon uses of common defs. 950 const RegMask &rm = n->in_RegMask(k); 951 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { 952 // Since we are BEFORE aggressive coalesce, leave the register 953 // mask untrimmed by the call. This encourages more coalescing. 954 // Later, AFTER aggressive, this live range will have to spill 955 // but the spiller handles slow-path calls very nicely. 956 } else { 957 lrg.AND( rm ); 958 } 959 960 // Check for bound register masks 961 const RegMask &lrgmask = lrg.mask(); 962 int kreg = n->in(k)->ideal_reg(); 963 bool is_vect = RegMask::is_vector(kreg); 964 assert(n->in(k)->bottom_type()->isa_vect() == NULL || 965 is_vect || kreg == Op_RegD || kreg == Op_RegL, 966 "vector must be in vector registers"); 967 if (lrgmask.is_bound(kreg)) 968 lrg._is_bound = 1; 969 970 // If this use of a double forces a mis-aligned double, 971 // flag as '_fat_proj' - really flag as allowing misalignment 972 // AND changes how we count interferences. A mis-aligned 973 // double can interfere with TWO aligned pairs, or effectively 974 // FOUR registers! 975 #ifdef ASSERT 976 if (is_vect) { 977 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned"); 978 assert(!lrg._fat_proj, "sanity"); 979 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity"); 980 } 981 #endif 982 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { 983 lrg._fat_proj = 1; 984 lrg._is_bound = 1; 985 } 986 // if the LRG is an unaligned pair, we will have to spill 987 // so clear the LRG's register mask if it is not already spilled 988 if (!is_vect && !n->is_SpillCopy() && 989 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 990 lrgmask.is_misaligned_pair()) { 991 lrg.Clear(); 992 } 993 994 // Check for maximum frequency value 995 if (lrg._maxfreq < block->_freq) { 996 lrg._maxfreq = block->_freq; 997 } 998 999 } // End for all allocated inputs 1000 } // end for all instructions 1001 } // end for all blocks 1002 1003 // Final per-liverange setup 1004 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { 1005 LRG &lrg = lrgs(i2); 1006 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); 1007 if (lrg.num_regs() > 1 && !lrg._fat_proj) { 1008 lrg.clear_to_sets(); 1009 } 1010 lrg.compute_set_mask_size(); 1011 if (lrg.not_free()) { // Handle case where we lose from the start 1012 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 1013 lrg._direct_conflict = 1; 1014 } 1015 lrg.set_degree(0); // no neighbors in IFG yet 1016 } 1017 } 1018 1019 // Set the was-lo-degree bit. Conservative coalescing should not change the 1020 // colorability of the graph. If any live range was of low-degree before 1021 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 1022 // The bit is checked in Simplify. 1023 void PhaseChaitin::set_was_low() { 1024 #ifdef ASSERT 1025 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1026 int size = lrgs(i).num_regs(); 1027 uint old_was_lo = lrgs(i)._was_lo; 1028 lrgs(i)._was_lo = 0; 1029 if( lrgs(i).lo_degree() ) { 1030 lrgs(i)._was_lo = 1; // Trivially of low degree 1031 } else { // Else check the Brigg's assertion 1032 // Brigg's observation is that the lo-degree neighbors of a 1033 // hi-degree live range will not interfere with the color choices 1034 // of said hi-degree live range. The Simplify reverse-stack-coloring 1035 // order takes care of the details. Hence you do not have to count 1036 // low-degree neighbors when determining if this guy colors. 1037 int briggs_degree = 0; 1038 IndexSet *s = _ifg->neighbors(i); 1039 IndexSetIterator elements(s); 1040 uint lidx; 1041 while((lidx = elements.next()) != 0) { 1042 if( !lrgs(lidx).lo_degree() ) 1043 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 1044 } 1045 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 1046 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 1047 } 1048 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 1049 } 1050 #endif 1051 } 1052 1053 #define REGISTER_CONSTRAINED 16 1054 1055 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 1056 void PhaseChaitin::cache_lrg_info( ) { 1057 1058 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1059 LRG &lrg = lrgs(i); 1060 1061 // Check for being of low degree: means we can be trivially colored. 1062 // Low degree, dead or must-spill guys just get to simplify right away 1063 if( lrg.lo_degree() || 1064 !lrg.alive() || 1065 lrg._must_spill ) { 1066 // Split low degree list into those guys that must get a 1067 // register and those that can go to register or stack. 1068 // The idea is LRGs that can go register or stack color first when 1069 // they have a good chance of getting a register. The register-only 1070 // lo-degree live ranges always get a register. 1071 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 1072 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 1073 lrg._next = _lo_stk_degree; 1074 _lo_stk_degree = i; 1075 } else { 1076 lrg._next = _lo_degree; 1077 _lo_degree = i; 1078 } 1079 } else { // Else high degree 1080 lrgs(_hi_degree)._prev = i; 1081 lrg._next = _hi_degree; 1082 lrg._prev = 0; 1083 _hi_degree = i; 1084 } 1085 } 1086 } 1087 1088 // Simplify the IFG by removing LRGs of low degree that have NO copies 1089 void PhaseChaitin::Pre_Simplify( ) { 1090 1091 // Warm up the lo-degree no-copy list 1092 int lo_no_copy = 0; 1093 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1094 if ((lrgs(i).lo_degree() && !lrgs(i)._has_copy) || 1095 !lrgs(i).alive() || 1096 lrgs(i)._must_spill) { 1097 lrgs(i)._next = lo_no_copy; 1098 lo_no_copy = i; 1099 } 1100 } 1101 1102 while( lo_no_copy ) { 1103 uint lo = lo_no_copy; 1104 lo_no_copy = lrgs(lo)._next; 1105 int size = lrgs(lo).num_regs(); 1106 1107 // Put the simplified guy on the simplified list. 1108 lrgs(lo)._next = _simplified; 1109 _simplified = lo; 1110 1111 // Yank this guy from the IFG. 1112 IndexSet *adj = _ifg->remove_node( lo ); 1113 1114 // If any neighbors' degrees fall below their number of 1115 // allowed registers, then put that neighbor on the low degree 1116 // list. Note that 'degree' can only fall and 'numregs' is 1117 // unchanged by this action. Thus the two are equal at most once, 1118 // so LRGs hit the lo-degree worklists at most once. 1119 IndexSetIterator elements(adj); 1120 uint neighbor; 1121 while ((neighbor = elements.next()) != 0) { 1122 LRG *n = &lrgs(neighbor); 1123 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1124 1125 // Check for just becoming of-low-degree 1126 if( n->just_lo_degree() && !n->_has_copy ) { 1127 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1128 // Put on lo-degree list 1129 n->_next = lo_no_copy; 1130 lo_no_copy = neighbor; 1131 } 1132 } 1133 } // End of while lo-degree no_copy worklist not empty 1134 1135 // No more lo-degree no-copy live ranges to simplify 1136 } 1137 1138 // Simplify the IFG by removing LRGs of low degree. 1139 void PhaseChaitin::Simplify( ) { 1140 1141 while( 1 ) { // Repeat till simplified it all 1142 // May want to explore simplifying lo_degree before _lo_stk_degree. 1143 // This might result in more spills coloring into registers during 1144 // Select(). 1145 while( _lo_degree || _lo_stk_degree ) { 1146 // If possible, pull from lo_stk first 1147 uint lo; 1148 if( _lo_degree ) { 1149 lo = _lo_degree; 1150 _lo_degree = lrgs(lo)._next; 1151 } else { 1152 lo = _lo_stk_degree; 1153 _lo_stk_degree = lrgs(lo)._next; 1154 } 1155 1156 // Put the simplified guy on the simplified list. 1157 lrgs(lo)._next = _simplified; 1158 _simplified = lo; 1159 // If this guy is "at risk" then mark his current neighbors 1160 if( lrgs(lo)._at_risk ) { 1161 IndexSetIterator elements(_ifg->neighbors(lo)); 1162 uint datum; 1163 while ((datum = elements.next()) != 0) { 1164 lrgs(datum)._risk_bias = lo; 1165 } 1166 } 1167 1168 // Yank this guy from the IFG. 1169 IndexSet *adj = _ifg->remove_node( lo ); 1170 1171 // If any neighbors' degrees fall below their number of 1172 // allowed registers, then put that neighbor on the low degree 1173 // list. Note that 'degree' can only fall and 'numregs' is 1174 // unchanged by this action. Thus the two are equal at most once, 1175 // so LRGs hit the lo-degree worklist at most once. 1176 IndexSetIterator elements(adj); 1177 uint neighbor; 1178 while ((neighbor = elements.next()) != 0) { 1179 LRG *n = &lrgs(neighbor); 1180 #ifdef ASSERT 1181 if( VerifyOpto || VerifyRegisterAllocator ) { 1182 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1183 } 1184 #endif 1185 1186 // Check for just becoming of-low-degree just counting registers. 1187 // _must_spill live ranges are already on the low degree list. 1188 if( n->just_lo_degree() && !n->_must_spill ) { 1189 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1190 // Pull from hi-degree list 1191 uint prev = n->_prev; 1192 uint next = n->_next; 1193 if( prev ) lrgs(prev)._next = next; 1194 else _hi_degree = next; 1195 lrgs(next)._prev = prev; 1196 n->_next = _lo_degree; 1197 _lo_degree = neighbor; 1198 } 1199 } 1200 } // End of while lo-degree/lo_stk_degree worklist not empty 1201 1202 // Check for got everything: is hi-degree list empty? 1203 if( !_hi_degree ) break; 1204 1205 // Time to pick a potential spill guy 1206 uint lo_score = _hi_degree; 1207 double score = lrgs(lo_score).score(); 1208 double area = lrgs(lo_score)._area; 1209 double cost = lrgs(lo_score)._cost; 1210 bool bound = lrgs(lo_score)._is_bound; 1211 1212 // Find cheapest guy 1213 debug_only( int lo_no_simplify=0; ); 1214 for( uint i = _hi_degree; i; i = lrgs(i)._next ) { 1215 assert( !(*_ifg->_yanked)[i], "" ); 1216 // It's just vaguely possible to move hi-degree to lo-degree without 1217 // going through a just-lo-degree stage: If you remove a double from 1218 // a float live range it's degree will drop by 2 and you can skip the 1219 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 1220 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1221 if( lrgs(i).lo_degree() ) { 1222 lo_score = i; 1223 break; 1224 } 1225 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1226 double iscore = lrgs(i).score(); 1227 double iarea = lrgs(i)._area; 1228 double icost = lrgs(i)._cost; 1229 bool ibound = lrgs(i)._is_bound; 1230 1231 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1232 // wins. Ties happen because all live ranges in question have spilled 1233 // a few times before and the spill-score adds a huge number which 1234 // washes out the low order bits. We are choosing the lesser of 2 1235 // evils; in this case pick largest area to spill. 1236 // Ties also happen when live ranges are defined and used only inside 1237 // one block. In which case their area is 0 and score set to max. 1238 // In such case choose bound live range over unbound to free registers 1239 // or with smaller cost to spill. 1240 if( iscore < score || 1241 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || 1242 (iscore == score && iarea == area && 1243 ( (ibound && !bound) || ibound == bound && (icost < cost) )) ) { 1244 lo_score = i; 1245 score = iscore; 1246 area = iarea; 1247 cost = icost; 1248 bound = ibound; 1249 } 1250 } 1251 LRG *lo_lrg = &lrgs(lo_score); 1252 // The live range we choose for spilling is either hi-degree, or very 1253 // rarely it can be low-degree. If we choose a hi-degree live range 1254 // there better not be any lo-degree choices. 1255 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1256 1257 // Pull from hi-degree list 1258 uint prev = lo_lrg->_prev; 1259 uint next = lo_lrg->_next; 1260 if( prev ) lrgs(prev)._next = next; 1261 else _hi_degree = next; 1262 lrgs(next)._prev = prev; 1263 // Jam him on the lo-degree list, despite his high degree. 1264 // Maybe he'll get a color, and maybe he'll spill. 1265 // Only Select() will know. 1266 lrgs(lo_score)._at_risk = true; 1267 _lo_degree = lo_score; 1268 lo_lrg->_next = 0; 1269 1270 } // End of while not simplified everything 1271 1272 } 1273 1274 // Is 'reg' register legal for 'lrg'? 1275 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { 1276 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && 1277 lrg.mask().Member(OptoReg::add(reg,-chunk))) { 1278 // RA uses OptoReg which represent the highest element of a registers set. 1279 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set 1280 // in which XMMd is used by RA to represent such vectors. A double value 1281 // uses [XMM,XMMb] pairs and XMMb is used by RA for it. 1282 // The register mask uses largest bits set of overlapping register sets. 1283 // On x86 with AVX it uses 8 bits for each XMM registers set. 1284 // 1285 // The 'lrg' already has cleared-to-set register mask (done in Select() 1286 // before calling choose_color()). Passing mask.Member(reg) check above 1287 // indicates that the size (num_regs) of 'reg' set is less or equal to 1288 // 'lrg' set size. 1289 // For set size 1 any register which is member of 'lrg' mask is legal. 1290 if (lrg.num_regs()==1) 1291 return true; 1292 // For larger sets only an aligned register with the same set size is legal. 1293 int mask = lrg.num_regs()-1; 1294 if ((reg&mask) == mask) 1295 return true; 1296 } 1297 return false; 1298 } 1299 1300 // Choose a color using the biasing heuristic 1301 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1302 1303 // Check for "at_risk" LRG's 1304 uint risk_lrg = _lrg_map.find(lrg._risk_bias); 1305 if( risk_lrg != 0 ) { 1306 // Walk the colored neighbors of the "at_risk" candidate 1307 // Choose a color which is both legal and already taken by a neighbor 1308 // of the "at_risk" candidate in order to improve the chances of the 1309 // "at_risk" candidate of coloring 1310 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1311 uint datum; 1312 while ((datum = elements.next()) != 0) { 1313 OptoReg::Name reg = lrgs(datum).reg(); 1314 // If this LRG's register is legal for us, choose it 1315 if (is_legal_reg(lrg, reg, chunk)) 1316 return reg; 1317 } 1318 } 1319 1320 uint copy_lrg = _lrg_map.find(lrg._copy_bias); 1321 if( copy_lrg != 0 ) { 1322 // If he has a color, 1323 if( !(*(_ifg->_yanked))[copy_lrg] ) { 1324 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1325 // And it is legal for you, 1326 if (is_legal_reg(lrg, reg, chunk)) 1327 return reg; 1328 } else if( chunk == 0 ) { 1329 // Choose a color which is legal for him 1330 RegMask tempmask = lrg.mask(); 1331 tempmask.AND(lrgs(copy_lrg).mask()); 1332 tempmask.clear_to_sets(lrg.num_regs()); 1333 OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); 1334 if (OptoReg::is_valid(reg)) 1335 return reg; 1336 } 1337 } 1338 1339 // If no bias info exists, just go with the register selection ordering 1340 if (lrg._is_vector || lrg.num_regs() == 2) { 1341 // Find an aligned set 1342 return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); 1343 } 1344 1345 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1346 // copy removal to remove many more copies, by preventing a just-assigned 1347 // register from being repeatedly assigned. 1348 OptoReg::Name reg = lrg.mask().find_first_elem(); 1349 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1350 // This 'Remove; find; Insert' idiom is an expensive way to find the 1351 // SECOND element in the mask. 1352 lrg.Remove(reg); 1353 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1354 lrg.Insert(reg); 1355 if( OptoReg::is_reg(reg2)) 1356 reg = reg2; 1357 } 1358 return OptoReg::add( reg, chunk ); 1359 } 1360 1361 // Choose a color in the current chunk 1362 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1363 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)"); 1364 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)"); 1365 1366 if( lrg.num_regs() == 1 || // Common Case 1367 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1368 // Use a heuristic to "bias" the color choice 1369 return bias_color(lrg, chunk); 1370 1371 assert(!lrg._is_vector, "should be not vector here" ); 1372 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1373 1374 // Fat-proj case or misaligned double argument. 1375 assert(lrg.compute_mask_size() == lrg.num_regs() || 1376 lrg.num_regs() == 2,"fat projs exactly color" ); 1377 assert( !chunk, "always color in 1st chunk" ); 1378 // Return the highest element in the set. 1379 return lrg.mask().find_last_elem(); 1380 } 1381 1382 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1383 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1384 // everything going back is guaranteed a color. Select that color. If some 1385 // hi-degree LRG cannot get a color then we record that we must spill. 1386 uint PhaseChaitin::Select( ) { 1387 uint spill_reg = LRG::SPILL_REG; 1388 _max_reg = OptoReg::Name(0); // Past max register used 1389 while( _simplified ) { 1390 // Pull next LRG from the simplified list - in reverse order of removal 1391 uint lidx = _simplified; 1392 LRG *lrg = &lrgs(lidx); 1393 _simplified = lrg->_next; 1394 1395 1396 #ifndef PRODUCT 1397 if (trace_spilling()) { 1398 ttyLocker ttyl; 1399 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1400 lrg->degrees_of_freedom()); 1401 lrg->dump(); 1402 } 1403 #endif 1404 1405 // Re-insert into the IFG 1406 _ifg->re_insert(lidx); 1407 if( !lrg->alive() ) continue; 1408 // capture allstackedness flag before mask is hacked 1409 const int is_allstack = lrg->mask().is_AllStack(); 1410 1411 // Yeah, yeah, yeah, I know, I know. I can refactor this 1412 // to avoid the GOTO, although the refactored code will not 1413 // be much clearer. We arrive here IFF we have a stack-based 1414 // live range that cannot color in the current chunk, and it 1415 // has to move into the next free stack chunk. 1416 int chunk = 0; // Current chunk is first chunk 1417 retry_next_chunk: 1418 1419 // Remove neighbor colors 1420 IndexSet *s = _ifg->neighbors(lidx); 1421 1422 debug_only(RegMask orig_mask = lrg->mask();) 1423 IndexSetIterator elements(s); 1424 uint neighbor; 1425 while ((neighbor = elements.next()) != 0) { 1426 // Note that neighbor might be a spill_reg. In this case, exclusion 1427 // of its color will be a no-op, since the spill_reg chunk is in outer 1428 // space. Also, if neighbor is in a different chunk, this exclusion 1429 // will be a no-op. (Later on, if lrg runs out of possible colors in 1430 // its chunk, a new chunk of color may be tried, in which case 1431 // examination of neighbors is started again, at retry_next_chunk.) 1432 LRG &nlrg = lrgs(neighbor); 1433 OptoReg::Name nreg = nlrg.reg(); 1434 // Only subtract masks in the same chunk 1435 if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { 1436 #ifndef PRODUCT 1437 uint size = lrg->mask().Size(); 1438 RegMask rm = lrg->mask(); 1439 #endif 1440 lrg->SUBTRACT(nlrg.mask()); 1441 #ifndef PRODUCT 1442 if (trace_spilling() && lrg->mask().Size() != size) { 1443 ttyLocker ttyl; 1444 tty->print("L%d ", lidx); 1445 rm.dump(); 1446 tty->print(" intersected L%d ", neighbor); 1447 nlrg.mask().dump(); 1448 tty->print(" removed "); 1449 rm.SUBTRACT(lrg->mask()); 1450 rm.dump(); 1451 tty->print(" leaving "); 1452 lrg->mask().dump(); 1453 tty->cr(); 1454 } 1455 #endif 1456 } 1457 } 1458 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1459 // Aligned pairs need aligned masks 1460 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1461 if (lrg->num_regs() > 1 && !lrg->_fat_proj) { 1462 lrg->clear_to_sets(); 1463 } 1464 1465 // Check if a color is available and if so pick the color 1466 OptoReg::Name reg = choose_color( *lrg, chunk ); 1467 #ifdef SPARC 1468 debug_only(lrg->compute_set_mask_size()); 1469 assert(lrg->num_regs() < 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); 1470 #endif 1471 1472 //--------------- 1473 // If we fail to color and the AllStack flag is set, trigger 1474 // a chunk-rollover event 1475 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1476 // Bump register mask up to next stack chunk 1477 chunk += RegMask::CHUNK_SIZE; 1478 lrg->Set_All(); 1479 1480 goto retry_next_chunk; 1481 } 1482 1483 //--------------- 1484 // Did we get a color? 1485 else if( OptoReg::is_valid(reg)) { 1486 #ifndef PRODUCT 1487 RegMask avail_rm = lrg->mask(); 1488 #endif 1489 1490 // Record selected register 1491 lrg->set_reg(reg); 1492 1493 if( reg >= _max_reg ) // Compute max register limit 1494 _max_reg = OptoReg::add(reg,1); 1495 // Fold reg back into normal space 1496 reg = OptoReg::add(reg,-chunk); 1497 1498 // If the live range is not bound, then we actually had some choices 1499 // to make. In this case, the mask has more bits in it than the colors 1500 // chosen. Restrict the mask to just what was picked. 1501 int n_regs = lrg->num_regs(); 1502 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1503 if (n_regs == 1 || !lrg->_fat_proj) { 1504 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecY, "sanity"); 1505 lrg->Clear(); // Clear the mask 1506 lrg->Insert(reg); // Set regmask to match selected reg 1507 // For vectors and pairs, also insert the low bit of the pair 1508 for (int i = 1; i < n_regs; i++) 1509 lrg->Insert(OptoReg::add(reg,-i)); 1510 lrg->set_mask_size(n_regs); 1511 } else { // Else fatproj 1512 // mask must be equal to fatproj bits, by definition 1513 } 1514 #ifndef PRODUCT 1515 if (trace_spilling()) { 1516 ttyLocker ttyl; 1517 tty->print("L%d selected ", lidx); 1518 lrg->mask().dump(); 1519 tty->print(" from "); 1520 avail_rm.dump(); 1521 tty->cr(); 1522 } 1523 #endif 1524 // Note that reg is the highest-numbered register in the newly-bound mask. 1525 } // end color available case 1526 1527 //--------------- 1528 // Live range is live and no colors available 1529 else { 1530 assert( lrg->alive(), "" ); 1531 assert( !lrg->_fat_proj || lrg->is_multidef() || 1532 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1533 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1534 1535 // Assign the special spillreg register 1536 lrg->set_reg(OptoReg::Name(spill_reg++)); 1537 // Do not empty the regmask; leave mask_size lying around 1538 // for use during Spilling 1539 #ifndef PRODUCT 1540 if( trace_spilling() ) { 1541 ttyLocker ttyl; 1542 tty->print("L%d spilling with neighbors: ", lidx); 1543 s->dump(); 1544 debug_only(tty->print(" original mask: ")); 1545 debug_only(orig_mask.dump()); 1546 dump_lrg(lidx); 1547 } 1548 #endif 1549 } // end spill case 1550 1551 } 1552 1553 return spill_reg-LRG::SPILL_REG; // Return number of spills 1554 } 1555 1556 // Copy 'was_spilled'-edness from the source Node to the dst Node. 1557 void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { 1558 if( _spilled_once.test(src->_idx) ) { 1559 _spilled_once.set(dst->_idx); 1560 lrgs(_lrg_map.find(dst))._was_spilled1 = 1; 1561 if( _spilled_twice.test(src->_idx) ) { 1562 _spilled_twice.set(dst->_idx); 1563 lrgs(_lrg_map.find(dst))._was_spilled2 = 1; 1564 } 1565 } 1566 } 1567 1568 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1569 void PhaseChaitin::set_was_spilled( Node *n ) { 1570 if( _spilled_once.test_set(n->_idx) ) 1571 _spilled_twice.set(n->_idx); 1572 } 1573 1574 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1575 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1576 void PhaseChaitin::fixup_spills() { 1577 // This function does only cisc spill work. 1578 if( !UseCISCSpill ) return; 1579 1580 NOT_PRODUCT( Compile::TracePhase t3("fixupSpills", &_t_fixupSpills, TimeCompiler); ) 1581 1582 // Grab the Frame Pointer 1583 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); 1584 1585 // For all blocks 1586 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1587 Block* block = _cfg.get_block(i); 1588 1589 // For all instructions in block 1590 uint last_inst = block->end_idx(); 1591 for (uint j = 1; j <= last_inst; j++) { 1592 Node* n = block->get_node(j); 1593 1594 // Dead instruction??? 1595 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1596 C->top() == n || // Or the random TOP node 1597 n->is_Proj(), // Or a fat-proj kill node 1598 "No dead instructions after post-alloc" ); 1599 1600 int inp = n->cisc_operand(); 1601 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1602 // Convert operand number to edge index number 1603 MachNode *mach = n->as_Mach(); 1604 inp = mach->operand_index(inp); 1605 Node *src = n->in(inp); // Value to load or store 1606 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); 1607 OptoReg::Name src_reg = lrg_cisc.reg(); 1608 // Doubles record the HIGH register of an adjacent pair. 1609 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1610 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1611 // This is a CISC Spill, get stack offset and construct new node 1612 #ifndef PRODUCT 1613 if( TraceCISCSpill ) { 1614 tty->print(" reg-instr: "); 1615 n->dump(); 1616 } 1617 #endif 1618 int stk_offset = reg2offset(src_reg); 1619 // Bailout if we might exceed node limit when spilling this instruction 1620 C->check_node_count(0, "out of nodes fixing spills"); 1621 if (C->failing()) return; 1622 // Transform node 1623 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); 1624 cisc->set_req(inp,fp); // Base register is frame pointer 1625 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1626 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1627 cisc->ins_req(1,src); // Requires a memory edge 1628 } 1629 block->map_node(cisc, j); // Insert into basic block 1630 n->subsume_by(cisc, C); // Correct graph 1631 // 1632 ++_used_cisc_instructions; 1633 #ifndef PRODUCT 1634 if( TraceCISCSpill ) { 1635 tty->print(" cisc-instr: "); 1636 cisc->dump(); 1637 } 1638 #endif 1639 } else { 1640 #ifndef PRODUCT 1641 if( TraceCISCSpill ) { 1642 tty->print(" using reg-instr: "); 1643 n->dump(); 1644 } 1645 #endif 1646 ++_unused_cisc_instructions; // input can be on stack 1647 } 1648 } 1649 1650 } // End of for all instructions 1651 1652 } // End of for all blocks 1653 } 1654 1655 // Helper to stretch above; recursively discover the base Node for a 1656 // given derived Node. Easy for AddP-related machine nodes, but needs 1657 // to be recursive for derived Phis. 1658 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1659 // See if already computed; if so return it 1660 if( derived_base_map[derived->_idx] ) 1661 return derived_base_map[derived->_idx]; 1662 1663 // See if this happens to be a base. 1664 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1665 // pointers derived from NULL! These are always along paths that 1666 // can't happen at run-time but the optimizer cannot deduce it so 1667 // we have to handle it gracefully. 1668 assert(!derived->bottom_type()->isa_narrowoop() || 1669 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1670 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1671 // If its an OOP with a non-zero offset, then it is derived. 1672 if( tj == NULL || tj->_offset == 0 ) { 1673 derived_base_map[derived->_idx] = derived; 1674 return derived; 1675 } 1676 // Derived is NULL+offset? Base is NULL! 1677 if( derived->is_Con() ) { 1678 Node *base = _matcher.mach_null(); 1679 assert(base != NULL, "sanity"); 1680 if (base->in(0) == NULL) { 1681 // Initialize it once and make it shared: 1682 // set control to _root and place it into Start block 1683 // (where top() node is placed). 1684 base->init_req(0, _cfg.get_root_node()); 1685 Block *startb = _cfg.get_block_for_node(C->top()); 1686 uint node_pos = startb->find_node(C->top()); 1687 startb->insert_node(base, node_pos); 1688 _cfg.map_node_to_block(base, startb); 1689 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet"); 1690 1691 // The loadConP0 might have projection nodes depending on architecture 1692 // Add the projection nodes to the CFG 1693 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { 1694 Node* use = base->fast_out(i); 1695 if (use->is_MachProj()) { 1696 startb->insert_node(use, ++node_pos); 1697 _cfg.map_node_to_block(use, startb); 1698 new_lrg(use, maxlrg++); 1699 } 1700 } 1701 } 1702 if (_lrg_map.live_range_id(base) == 0) { 1703 new_lrg(base, maxlrg++); 1704 } 1705 assert(base->in(0) == _cfg.get_root_node() && _cfg.get_block_for_node(base) == _cfg.get_block_for_node(C->top()), "base NULL should be shared"); 1706 derived_base_map[derived->_idx] = base; 1707 return base; 1708 } 1709 1710 // Check for AddP-related opcodes 1711 if (!derived->is_Phi()) { 1712 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, err_msg_res("but is: %s", derived->Name())); 1713 Node *base = derived->in(AddPNode::Base); 1714 derived_base_map[derived->_idx] = base; 1715 return base; 1716 } 1717 1718 // Recursively find bases for Phis. 1719 // First check to see if we can avoid a base Phi here. 1720 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1721 uint i; 1722 for( i = 2; i < derived->req(); i++ ) 1723 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1724 break; 1725 // Went to the end without finding any different bases? 1726 if( i == derived->req() ) { // No need for a base Phi here 1727 derived_base_map[derived->_idx] = base; 1728 return base; 1729 } 1730 1731 // Now we see we need a base-Phi here to merge the bases 1732 const Type *t = base->bottom_type(); 1733 base = new PhiNode( derived->in(0), t ); 1734 for( i = 1; i < derived->req(); i++ ) { 1735 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1736 t = t->meet(base->in(i)->bottom_type()); 1737 } 1738 base->as_Phi()->set_type(t); 1739 1740 // Search the current block for an existing base-Phi 1741 Block *b = _cfg.get_block_for_node(derived); 1742 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1743 Node *phi = b->get_node(i); 1744 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1745 b->insert_node(base, i); // Must insert created Phi here as base 1746 _cfg.map_node_to_block(base, b); 1747 new_lrg(base,maxlrg++); 1748 break; 1749 } 1750 // See if Phi matches. 1751 uint j; 1752 for( j = 1; j < base->req(); j++ ) 1753 if( phi->in(j) != base->in(j) && 1754 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1755 break; 1756 if( j == base->req() ) { // All inputs match? 1757 base = phi; // Then use existing 'phi' and drop 'base' 1758 break; 1759 } 1760 } 1761 1762 1763 // Cache info for later passes 1764 derived_base_map[derived->_idx] = base; 1765 return base; 1766 } 1767 1768 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1769 // base pointer that is live across the Safepoint for oopmap building. The 1770 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1771 // required edge set. 1772 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { 1773 int must_recompute_live = false; 1774 uint maxlrg = _lrg_map.max_lrg_id(); 1775 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1776 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1777 1778 // For all blocks in RPO do... 1779 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1780 Block* block = _cfg.get_block(i); 1781 // Note use of deep-copy constructor. I cannot hammer the original 1782 // liveout bits, because they are needed by the following coalesce pass. 1783 IndexSet liveout(_live->live(block)); 1784 1785 for (uint j = block->end_idx() + 1; j > 1; j--) { 1786 Node* n = block->get_node(j - 1); 1787 1788 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1789 // like to see in the same register. Compare uses the loop-phi and so 1790 // extends its live range BUT cannot be part of the cycle. If this 1791 // extended live range overlaps with the update of the loop-phi value 1792 // we need both alive at the same time -- which requires at least 1 1793 // copy. But because Intel has only 2-address registers we end up with 1794 // at least 2 copies, one before the loop-phi update instruction and 1795 // one after. Instead we split the input to the compare just after the 1796 // phi. 1797 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1798 Node *phi = n->in(1); 1799 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1800 Block *phi_block = _cfg.get_block_for_node(phi); 1801 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { 1802 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1803 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); 1804 insert_proj( phi_block, 1, spill, maxlrg++ ); 1805 n->set_req(1,spill); 1806 must_recompute_live = true; 1807 } 1808 } 1809 } 1810 1811 // Get value being defined 1812 uint lidx = _lrg_map.live_range_id(n); 1813 // Ignore the occasional brand-new live range 1814 if (lidx && lidx < _lrg_map.max_lrg_id()) { 1815 // Remove from live-out set 1816 liveout.remove(lidx); 1817 1818 // Copies do not define a new value and so do not interfere. 1819 // Remove the copies source from the liveout set before interfering. 1820 uint idx = n->is_Copy(); 1821 if (idx) { 1822 liveout.remove(_lrg_map.live_range_id(n->in(idx))); 1823 } 1824 } 1825 1826 // Found a safepoint? 1827 JVMState *jvms = n->jvms(); 1828 if( jvms ) { 1829 // Now scan for a live derived pointer 1830 IndexSetIterator elements(&liveout); 1831 uint neighbor; 1832 while ((neighbor = elements.next()) != 0) { 1833 // Find reaching DEF for base and derived values 1834 // This works because we are still in SSA during this call. 1835 Node *derived = lrgs(neighbor)._def; 1836 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1837 assert(!derived->bottom_type()->isa_narrowoop() || 1838 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1839 // If its an OOP with a non-zero offset, then it is derived. 1840 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { 1841 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); 1842 assert(base->_idx < _lrg_map.size(), ""); 1843 // Add reaching DEFs of derived pointer and base pointer as a 1844 // pair of inputs 1845 n->add_req(derived); 1846 n->add_req(base); 1847 1848 // See if the base pointer is already live to this point. 1849 // Since I'm working on the SSA form, live-ness amounts to 1850 // reaching def's. So if I find the base's live range then 1851 // I know the base's def reaches here. 1852 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or 1853 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND 1854 (_lrg_map.live_range_id(base) > 0) && // not a constant 1855 _cfg.get_block_for_node(base) != block) { // base not def'd in blk) 1856 // Base pointer is not currently live. Since I stretched 1857 // the base pointer to here and it crosses basic-block 1858 // boundaries, the global live info is now incorrect. 1859 // Recompute live. 1860 must_recompute_live = true; 1861 } // End of if base pointer is not live to debug info 1862 } 1863 } // End of scan all live data for derived ptrs crossing GC point 1864 } // End of if found a GC point 1865 1866 // Make all inputs live 1867 if (!n->is_Phi()) { // Phi function uses come from prior block 1868 for (uint k = 1; k < n->req(); k++) { 1869 uint lidx = _lrg_map.live_range_id(n->in(k)); 1870 if (lidx < _lrg_map.max_lrg_id()) { 1871 liveout.insert(lidx); 1872 } 1873 } 1874 } 1875 1876 } // End of forall instructions in block 1877 liveout.clear(); // Free the memory used by liveout. 1878 1879 } // End of forall blocks 1880 _lrg_map.set_max_lrg_id(maxlrg); 1881 1882 // If I created a new live range I need to recompute live 1883 if (maxlrg != _ifg->_maxlrg) { 1884 must_recompute_live = true; 1885 } 1886 1887 return must_recompute_live != 0; 1888 } 1889 1890 // Extend the node to LRG mapping 1891 1892 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { 1893 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); 1894 } 1895 1896 #ifndef PRODUCT 1897 void PhaseChaitin::dump(const Node *n) const { 1898 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; 1899 tty->print("L%d",r); 1900 if (r && n->Opcode() != Op_Phi) { 1901 if( _node_regs ) { // Got a post-allocation copy of allocation? 1902 tty->print("["); 1903 OptoReg::Name second = get_reg_second(n); 1904 if( OptoReg::is_valid(second) ) { 1905 if( OptoReg::is_reg(second) ) 1906 tty->print("%s:",Matcher::regName[second]); 1907 else 1908 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1909 } 1910 OptoReg::Name first = get_reg_first(n); 1911 if( OptoReg::is_reg(first) ) 1912 tty->print("%s]",Matcher::regName[first]); 1913 else 1914 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1915 } else 1916 n->out_RegMask().dump(); 1917 } 1918 tty->print("/N%d\t",n->_idx); 1919 tty->print("%s === ", n->Name()); 1920 uint k; 1921 for (k = 0; k < n->req(); k++) { 1922 Node *m = n->in(k); 1923 if (!m) { 1924 tty->print("_ "); 1925 } 1926 else { 1927 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1928 tty->print("L%d",r); 1929 // Data MultiNode's can have projections with no real registers. 1930 // Don't die while dumping them. 1931 int op = n->Opcode(); 1932 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1933 if( _node_regs ) { 1934 tty->print("["); 1935 OptoReg::Name second = get_reg_second(n->in(k)); 1936 if( OptoReg::is_valid(second) ) { 1937 if( OptoReg::is_reg(second) ) 1938 tty->print("%s:",Matcher::regName[second]); 1939 else 1940 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 1941 reg2offset_unchecked(second)); 1942 } 1943 OptoReg::Name first = get_reg_first(n->in(k)); 1944 if( OptoReg::is_reg(first) ) 1945 tty->print("%s]",Matcher::regName[first]); 1946 else 1947 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 1948 reg2offset_unchecked(first)); 1949 } else 1950 n->in_RegMask(k).dump(); 1951 } 1952 tty->print("/N%d ",m->_idx); 1953 } 1954 } 1955 if( k < n->len() && n->in(k) ) tty->print("| "); 1956 for( ; k < n->len(); k++ ) { 1957 Node *m = n->in(k); 1958 if(!m) { 1959 break; 1960 } 1961 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1962 tty->print("L%d",r); 1963 tty->print("/N%d ",m->_idx); 1964 } 1965 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 1966 else n->dump_spec(tty); 1967 if( _spilled_once.test(n->_idx ) ) { 1968 tty->print(" Spill_1"); 1969 if( _spilled_twice.test(n->_idx ) ) 1970 tty->print(" Spill_2"); 1971 } 1972 tty->print("\n"); 1973 } 1974 1975 void PhaseChaitin::dump(const Block *b) const { 1976 b->dump_head(&_cfg); 1977 1978 // For all instructions 1979 for( uint j = 0; j < b->number_of_nodes(); j++ ) 1980 dump(b->get_node(j)); 1981 // Print live-out info at end of block 1982 if( _live ) { 1983 tty->print("Liveout: "); 1984 IndexSet *live = _live->live(b); 1985 IndexSetIterator elements(live); 1986 tty->print("{"); 1987 uint i; 1988 while ((i = elements.next()) != 0) { 1989 tty->print("L%d ", _lrg_map.find_const(i)); 1990 } 1991 tty->print_cr("}"); 1992 } 1993 tty->print("\n"); 1994 } 1995 1996 void PhaseChaitin::dump() const { 1997 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 1998 _matcher._new_SP, _framesize ); 1999 2000 // For all blocks 2001 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2002 dump(_cfg.get_block(i)); 2003 } 2004 // End of per-block dump 2005 tty->print("\n"); 2006 2007 if (!_ifg) { 2008 tty->print("(No IFG.)\n"); 2009 return; 2010 } 2011 2012 // Dump LRG array 2013 tty->print("--- Live RanGe Array ---\n"); 2014 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { 2015 tty->print("L%d: ",i2); 2016 if (i2 < _ifg->_maxlrg) { 2017 lrgs(i2).dump(); 2018 } 2019 else { 2020 tty->print_cr("new LRG"); 2021 } 2022 } 2023 tty->cr(); 2024 2025 // Dump lo-degree list 2026 tty->print("Lo degree: "); 2027 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 2028 tty->print("L%d ",i3); 2029 tty->cr(); 2030 2031 // Dump lo-stk-degree list 2032 tty->print("Lo stk degree: "); 2033 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 2034 tty->print("L%d ",i4); 2035 tty->cr(); 2036 2037 // Dump lo-degree list 2038 tty->print("Hi degree: "); 2039 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 2040 tty->print("L%d ",i5); 2041 tty->cr(); 2042 } 2043 2044 void PhaseChaitin::dump_degree_lists() const { 2045 // Dump lo-degree list 2046 tty->print("Lo degree: "); 2047 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 2048 tty->print("L%d ",i); 2049 tty->cr(); 2050 2051 // Dump lo-stk-degree list 2052 tty->print("Lo stk degree: "); 2053 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 2054 tty->print("L%d ",i2); 2055 tty->cr(); 2056 2057 // Dump lo-degree list 2058 tty->print("Hi degree: "); 2059 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 2060 tty->print("L%d ",i3); 2061 tty->cr(); 2062 } 2063 2064 void PhaseChaitin::dump_simplified() const { 2065 tty->print("Simplified: "); 2066 for( uint i = _simplified; i; i = lrgs(i)._next ) 2067 tty->print("L%d ",i); 2068 tty->cr(); 2069 } 2070 2071 static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { 2072 if ((int)reg < 0) 2073 sprintf(buf, "<OptoReg::%d>", (int)reg); 2074 else if (OptoReg::is_reg(reg)) 2075 strcpy(buf, Matcher::regName[reg]); 2076 else 2077 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 2078 pc->reg2offset(reg)); 2079 return buf+strlen(buf); 2080 } 2081 2082 // Dump a register name into a buffer. Be intelligent if we get called 2083 // before allocation is complete. 2084 char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { 2085 if( !this ) { // Not got anything? 2086 sprintf(buf,"N%d",n->_idx); // Then use Node index 2087 } else if( _node_regs ) { 2088 // Post allocation, use direct mappings, no LRG info available 2089 print_reg( get_reg_first(n), this, buf ); 2090 } else { 2091 uint lidx = _lrg_map.find_const(n); // Grab LRG number 2092 if( !_ifg ) { 2093 sprintf(buf,"L%d",lidx); // No register binding yet 2094 } else if( !lidx ) { // Special, not allocated value 2095 strcpy(buf,"Special"); 2096 } else { 2097 if (lrgs(lidx)._is_vector) { 2098 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) 2099 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register 2100 else 2101 sprintf(buf,"L%d",lidx); // No register binding yet 2102 } else if( (lrgs(lidx).num_regs() == 1) 2103 ? lrgs(lidx).mask().is_bound1() 2104 : lrgs(lidx).mask().is_bound_pair() ) { 2105 // Hah! We have a bound machine register 2106 print_reg( lrgs(lidx).reg(), this, buf ); 2107 } else { 2108 sprintf(buf,"L%d",lidx); // No register binding yet 2109 } 2110 } 2111 } 2112 return buf+strlen(buf); 2113 } 2114 2115 void PhaseChaitin::dump_for_spill_split_recycle() const { 2116 if( WizardMode && (PrintCompilation || PrintOpto) ) { 2117 // Display which live ranges need to be split and the allocator's state 2118 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 2119 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { 2120 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 2121 tty->print("L%d: ", bidx); 2122 lrgs(bidx).dump(); 2123 } 2124 } 2125 tty->cr(); 2126 dump(); 2127 } 2128 } 2129 2130 void PhaseChaitin::dump_frame() const { 2131 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 2132 const TypeTuple *domain = C->tf()->domain(); 2133 const int argcnt = domain->cnt() - TypeFunc::Parms; 2134 2135 // Incoming arguments in registers dump 2136 for( int k = 0; k < argcnt; k++ ) { 2137 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 2138 if( OptoReg::is_reg(parmreg)) { 2139 const char *reg_name = OptoReg::regname(parmreg); 2140 tty->print("#r%3.3d %s", parmreg, reg_name); 2141 parmreg = _matcher._parm_regs[k].second(); 2142 if( OptoReg::is_reg(parmreg)) { 2143 tty->print(":%s", OptoReg::regname(parmreg)); 2144 } 2145 tty->print(" : parm %d: ", k); 2146 domain->field_at(k + TypeFunc::Parms)->dump(); 2147 tty->cr(); 2148 } 2149 } 2150 2151 // Check for un-owned padding above incoming args 2152 OptoReg::Name reg = _matcher._new_SP; 2153 if( reg > _matcher._in_arg_limit ) { 2154 reg = OptoReg::add(reg, -1); 2155 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 2156 } 2157 2158 // Incoming argument area dump 2159 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 2160 while( reg > begin_in_arg ) { 2161 reg = OptoReg::add(reg, -1); 2162 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2163 int j; 2164 for( j = 0; j < argcnt; j++) { 2165 if( _matcher._parm_regs[j].first() == reg || 2166 _matcher._parm_regs[j].second() == reg ) { 2167 tty->print("parm %d: ",j); 2168 domain->field_at(j + TypeFunc::Parms)->dump(); 2169 tty->cr(); 2170 break; 2171 } 2172 } 2173 if( j >= argcnt ) 2174 tty->print_cr("HOLE, owned by SELF"); 2175 } 2176 2177 // Old outgoing preserve area 2178 while( reg > _matcher._old_SP ) { 2179 reg = OptoReg::add(reg, -1); 2180 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 2181 } 2182 2183 // Old SP 2184 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 2185 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 2186 2187 // Preserve area dump 2188 int fixed_slots = C->fixed_slots(); 2189 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); 2190 OptoReg::Name return_addr = _matcher.return_addr(); 2191 2192 reg = OptoReg::add(reg, -1); 2193 while (OptoReg::is_stack(reg)) { 2194 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2195 if (return_addr == reg) { 2196 tty->print_cr("return address"); 2197 } else if (reg >= begin_in_preserve) { 2198 // Preserved slots are present on x86 2199 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) 2200 tty->print_cr("saved fp register"); 2201 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && 2202 VerifyStackAtCalls) 2203 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 2204 else 2205 tty->print_cr("in_preserve"); 2206 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { 2207 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 2208 } else { 2209 tty->print_cr("pad2, stack alignment"); 2210 } 2211 reg = OptoReg::add(reg, -1); 2212 } 2213 2214 // Spill area dump 2215 reg = OptoReg::add(_matcher._new_SP, _framesize ); 2216 while( reg > _matcher._out_arg_limit ) { 2217 reg = OptoReg::add(reg, -1); 2218 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 2219 } 2220 2221 // Outgoing argument area dump 2222 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 2223 reg = OptoReg::add(reg, -1); 2224 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 2225 } 2226 2227 // Outgoing new preserve area 2228 while( reg > _matcher._new_SP ) { 2229 reg = OptoReg::add(reg, -1); 2230 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 2231 } 2232 tty->print_cr("#"); 2233 } 2234 2235 void PhaseChaitin::dump_bb( uint pre_order ) const { 2236 tty->print_cr("---dump of B%d---",pre_order); 2237 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2238 Block* block = _cfg.get_block(i); 2239 if (block->_pre_order == pre_order) { 2240 dump(block); 2241 } 2242 } 2243 } 2244 2245 void PhaseChaitin::dump_lrg( uint lidx, bool defs_only ) const { 2246 tty->print_cr("---dump of L%d---",lidx); 2247 2248 if (_ifg) { 2249 if (lidx >= _lrg_map.max_lrg_id()) { 2250 tty->print("Attempt to print live range index beyond max live range.\n"); 2251 return; 2252 } 2253 tty->print("L%d: ",lidx); 2254 if (lidx < _ifg->_maxlrg) { 2255 lrgs(lidx).dump(); 2256 } else { 2257 tty->print_cr("new LRG"); 2258 } 2259 } 2260 if( _ifg && lidx < _ifg->_maxlrg) { 2261 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 2262 _ifg->neighbors(lidx)->dump(); 2263 tty->cr(); 2264 } 2265 // For all blocks 2266 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2267 Block* block = _cfg.get_block(i); 2268 int dump_once = 0; 2269 2270 // For all instructions 2271 for( uint j = 0; j < block->number_of_nodes(); j++ ) { 2272 Node *n = block->get_node(j); 2273 if (_lrg_map.find_const(n) == lidx) { 2274 if (!dump_once++) { 2275 tty->cr(); 2276 block->dump_head(&_cfg); 2277 } 2278 dump(n); 2279 continue; 2280 } 2281 if (!defs_only) { 2282 uint cnt = n->req(); 2283 for( uint k = 1; k < cnt; k++ ) { 2284 Node *m = n->in(k); 2285 if (!m) { 2286 continue; // be robust in the dumper 2287 } 2288 if (_lrg_map.find_const(m) == lidx) { 2289 if (!dump_once++) { 2290 tty->cr(); 2291 block->dump_head(&_cfg); 2292 } 2293 dump(n); 2294 } 2295 } 2296 } 2297 } 2298 } // End of per-block dump 2299 tty->cr(); 2300 } 2301 #endif // not PRODUCT 2302 2303 int PhaseChaitin::_final_loads = 0; 2304 int PhaseChaitin::_final_stores = 0; 2305 int PhaseChaitin::_final_memoves= 0; 2306 int PhaseChaitin::_final_copies = 0; 2307 double PhaseChaitin::_final_load_cost = 0; 2308 double PhaseChaitin::_final_store_cost = 0; 2309 double PhaseChaitin::_final_memove_cost= 0; 2310 double PhaseChaitin::_final_copy_cost = 0; 2311 int PhaseChaitin::_conserv_coalesce = 0; 2312 int PhaseChaitin::_conserv_coalesce_pair = 0; 2313 int PhaseChaitin::_conserv_coalesce_trie = 0; 2314 int PhaseChaitin::_conserv_coalesce_quad = 0; 2315 int PhaseChaitin::_post_alloc = 0; 2316 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2317 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2318 int PhaseChaitin::_used_cisc_instructions = 0; 2319 int PhaseChaitin::_unused_cisc_instructions = 0; 2320 int PhaseChaitin::_allocator_attempts = 0; 2321 int PhaseChaitin::_allocator_successes = 0; 2322 2323 #ifndef PRODUCT 2324 uint PhaseChaitin::_high_pressure = 0; 2325 uint PhaseChaitin::_low_pressure = 0; 2326 2327 void PhaseChaitin::print_chaitin_statistics() { 2328 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); 2329 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); 2330 tty->print_cr("Adjusted spill cost = %7.0f.", 2331 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2332 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2333 tty->print("Conservatively coalesced %d copies, %d pairs", 2334 _conserv_coalesce, _conserv_coalesce_pair); 2335 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2336 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2337 tty->print_cr(", %d post alloc.", _post_alloc); 2338 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2339 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2340 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2341 if( _used_cisc_instructions || _unused_cisc_instructions ) 2342 tty->print_cr("Used cisc instruction %d, remained in register %d", 2343 _used_cisc_instructions, _unused_cisc_instructions); 2344 if( _allocator_successes != 0 ) 2345 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2346 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2347 } 2348 #endif // not PRODUCT