1 /* 2 * Copyright (c) 2000, 2020, 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 "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/block.hpp" 32 #include "opto/callnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/chaitin.hpp" 35 #include "opto/coalesce.hpp" 36 #include "opto/connode.hpp" 37 #include "opto/idealGraphPrinter.hpp" 38 #include "opto/indexSet.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/movenode.hpp" 42 #include "opto/opcodes.hpp" 43 #include "opto/rootnode.hpp" 44 #include "utilities/align.hpp" 45 46 #ifndef PRODUCT 47 void LRG::dump() const { 48 ttyLocker ttyl; 49 tty->print("%d ",num_regs()); 50 _mask.dump(); 51 if( _msize_valid ) { 52 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size); 53 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size()); 54 } else { 55 tty->print(", #?(%d) ",_mask.Size()); 56 } 57 58 tty->print("EffDeg: "); 59 if( _degree_valid ) tty->print( "%d ", _eff_degree ); 60 else tty->print("? "); 61 62 if( is_multidef() ) { 63 tty->print("MultiDef "); 64 if (_defs != NULL) { 65 tty->print("("); 66 for (int i = 0; i < _defs->length(); i++) { 67 tty->print("N%d ", _defs->at(i)->_idx); 68 } 69 tty->print(") "); 70 } 71 } 72 else if( _def == 0 ) tty->print("Dead "); 73 else tty->print("Def: N%d ",_def->_idx); 74 75 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score()); 76 // Flags 77 if( _is_oop ) tty->print("Oop "); 78 if( _is_float ) tty->print("Float "); 79 if( _is_vector ) tty->print("Vector "); 80 if( _was_spilled1 ) tty->print("Spilled "); 81 if( _was_spilled2 ) tty->print("Spilled2 "); 82 if( _direct_conflict ) tty->print("Direct_conflict "); 83 if( _fat_proj ) tty->print("Fat "); 84 if( _was_lo ) tty->print("Lo "); 85 if( _has_copy ) tty->print("Copy "); 86 if( _at_risk ) tty->print("Risk "); 87 88 if( _must_spill ) tty->print("Must_spill "); 89 if( _is_bound ) tty->print("Bound "); 90 if( _msize_valid ) { 91 if( _degree_valid && lo_degree() ) tty->print("Trivial "); 92 } 93 94 tty->cr(); 95 } 96 #endif 97 98 // Compute score from cost and area. Low score is best to spill. 99 static double raw_score( double cost, double area ) { 100 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; 101 } 102 103 double LRG::score() const { 104 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. 105 // Bigger area lowers score, encourages spilling this live range. 106 // Bigger cost raise score, prevents spilling this live range. 107 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer 108 // to turn a divide by a constant into a multiply by the reciprical). 109 double score = raw_score( _cost, _area); 110 111 // Account for area. Basically, LRGs covering large areas are better 112 // to spill because more other LRGs get freed up. 113 if( _area == 0.0 ) // No area? Then no progress to spill 114 return 1e35; 115 116 if( _was_spilled2 ) // If spilled once before, we are unlikely 117 return score + 1e30; // to make progress again. 118 119 if( _cost >= _area*3.0 ) // Tiny area relative to cost 120 return score + 1e17; // Probably no progress to spill 121 122 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost 123 return score + 1e10; // Likely no progress to spill 124 125 return score; 126 } 127 128 #define NUMBUCKS 3 129 130 // Straight out of Tarjan's union-find algorithm 131 uint LiveRangeMap::find_compress(uint lrg) { 132 uint cur = lrg; 133 uint next = _uf_map.at(cur); 134 while (next != cur) { // Scan chain of equivalences 135 assert( next < cur, "always union smaller"); 136 cur = next; // until find a fixed-point 137 next = _uf_map.at(cur); 138 } 139 140 // Core of union-find algorithm: update chain of 141 // equivalences to be equal to the root. 142 while (lrg != next) { 143 uint tmp = _uf_map.at(lrg); 144 _uf_map.at_put(lrg, next); 145 lrg = tmp; 146 } 147 return lrg; 148 } 149 150 // Reset the Union-Find map to identity 151 void LiveRangeMap::reset_uf_map(uint max_lrg_id) { 152 _max_lrg_id= max_lrg_id; 153 // Force the Union-Find mapping to be at least this large 154 _uf_map.at_put_grow(_max_lrg_id, 0); 155 // Initialize it to be the ID mapping. 156 for (uint i = 0; i < _max_lrg_id; ++i) { 157 _uf_map.at_put(i, i); 158 } 159 } 160 161 // Make all Nodes map directly to their final live range; no need for 162 // the Union-Find mapping after this call. 163 void LiveRangeMap::compress_uf_map_for_nodes() { 164 // For all Nodes, compress mapping 165 uint unique = _names.length(); 166 for (uint i = 0; i < unique; ++i) { 167 uint lrg = _names.at(i); 168 uint compressed_lrg = find(lrg); 169 if (lrg != compressed_lrg) { 170 _names.at_put(i, compressed_lrg); 171 } 172 } 173 } 174 175 // Like Find above, but no path compress, so bad asymptotic behavior 176 uint LiveRangeMap::find_const(uint lrg) const { 177 if (!lrg) { 178 return lrg; // Ignore the zero LRG 179 } 180 181 // Off the end? This happens during debugging dumps when you got 182 // brand new live ranges but have not told the allocator yet. 183 if (lrg >= _max_lrg_id) { 184 return lrg; 185 } 186 187 uint next = _uf_map.at(lrg); 188 while (next != lrg) { // Scan chain of equivalences 189 assert(next < lrg, "always union smaller"); 190 lrg = next; // until find a fixed-point 191 next = _uf_map.at(lrg); 192 } 193 return next; 194 } 195 196 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated) 197 : PhaseRegAlloc(unique, cfg, matcher, 198 #ifndef PRODUCT 199 print_chaitin_statistics 200 #else 201 NULL 202 #endif 203 ) 204 , _live(0) 205 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0) 206 , _oldphi(unique) 207 #ifndef PRODUCT 208 , _trace_spilling(C->directive()->TraceSpillingOption) 209 #endif 210 , _lrg_map(Thread::current()->resource_area(), unique) 211 , _scheduling_info_generated(scheduling_info_generated) 212 , _sched_int_pressure(0, INTPRESSURE) 213 , _sched_float_pressure(0, FLOATPRESSURE) 214 , _scratch_int_pressure(0, INTPRESSURE) 215 , _scratch_float_pressure(0, FLOATPRESSURE) 216 { 217 Compile::TracePhase tp("ctorChaitin", &timers[_t_ctorChaitin]); 218 219 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency()); 220 221 // Build a list of basic blocks, sorted by frequency 222 _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 223 // Experiment with sorting strategies to speed compilation 224 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket 225 Block **buckets[NUMBUCKS]; // Array of buckets 226 uint buckcnt[NUMBUCKS]; // Array of bucket counters 227 double buckval[NUMBUCKS]; // Array of bucket value cutoffs 228 for (uint i = 0; i < NUMBUCKS; i++) { 229 buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 230 buckcnt[i] = 0; 231 // Bump by three orders of magnitude each time 232 cutoff *= 0.001; 233 buckval[i] = cutoff; 234 for (uint j = 0; j < _cfg.number_of_blocks(); j++) { 235 buckets[i][j] = NULL; 236 } 237 } 238 // Sort blocks into buckets 239 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 240 for (uint j = 0; j < NUMBUCKS; j++) { 241 if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) { 242 // Assign block to end of list for appropriate bucket 243 buckets[j][buckcnt[j]++] = _cfg.get_block(i); 244 break; // kick out of inner loop 245 } 246 } 247 } 248 // Dump buckets into final block array 249 uint blkcnt = 0; 250 for (uint i = 0; i < NUMBUCKS; i++) { 251 for (uint j = 0; j < buckcnt[i]; j++) { 252 _blks[blkcnt++] = buckets[i][j]; 253 } 254 } 255 256 assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled"); 257 } 258 259 // union 2 sets together. 260 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) { 261 uint src = _lrg_map.find(src_n); 262 uint dst = _lrg_map.find(dst_n); 263 assert(src, ""); 264 assert(dst, ""); 265 assert(src < _lrg_map.max_lrg_id(), "oob"); 266 assert(dst < _lrg_map.max_lrg_id(), "oob"); 267 assert(src < dst, "always union smaller"); 268 _lrg_map.uf_map(dst, src); 269 } 270 271 void PhaseChaitin::new_lrg(const Node *x, uint lrg) { 272 // Make the Node->LRG mapping 273 _lrg_map.extend(x->_idx,lrg); 274 // Make the Union-Find mapping an identity function 275 _lrg_map.uf_extend(lrg, lrg); 276 } 277 278 279 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) { 280 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections"); 281 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); ) 282 int found_projs = 0; 283 uint cnt = orig->outcnt(); 284 for (uint i = 0; i < cnt; i++) { 285 Node* proj = orig->raw_out(i); 286 if (proj->is_MachProj()) { 287 assert(proj->outcnt() == 0, "only kill projections are expected here"); 288 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections"); 289 found_projs++; 290 // Copy kill projections after the cloned node 291 Node* kills = proj->clone(); 292 kills->set_req(0, copy); 293 b->insert_node(kills, idx++); 294 _cfg.map_node_to_block(kills, b); 295 new_lrg(kills, max_lrg_id++); 296 } 297 } 298 return found_projs; 299 } 300 301 // Renumber the live ranges to compact them. Makes the IFG smaller. 302 void PhaseChaitin::compact() { 303 Compile::TracePhase tp("chaitinCompact", &timers[_t_chaitinCompact]); 304 305 // Current the _uf_map contains a series of short chains which are headed 306 // by a self-cycle. All the chains run from big numbers to little numbers. 307 // The Find() call chases the chains & shortens them for the next Find call. 308 // We are going to change this structure slightly. Numbers above a moving 309 // wave 'i' are unchanged. Numbers below 'j' point directly to their 310 // compacted live range with no further chaining. There are no chains or 311 // cycles below 'i', so the Find call no longer works. 312 uint j=1; 313 uint i; 314 for (i = 1; i < _lrg_map.max_lrg_id(); i++) { 315 uint lr = _lrg_map.uf_live_range_id(i); 316 // Ignore unallocated live ranges 317 if (!lr) { 318 continue; 319 } 320 assert(lr <= i, ""); 321 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr)); 322 } 323 // Now change the Node->LR mapping to reflect the compacted names 324 uint unique = _lrg_map.size(); 325 for (i = 0; i < unique; i++) { 326 uint lrg_id = _lrg_map.live_range_id(i); 327 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id)); 328 } 329 330 // Reset the Union-Find mapping 331 _lrg_map.reset_uf_map(j); 332 } 333 334 void PhaseChaitin::Register_Allocate() { 335 336 // Above the OLD FP (and in registers) are the incoming arguments. Stack 337 // slots in this area are called "arg_slots". Above the NEW FP (and in 338 // registers) is the outgoing argument area; above that is the spill/temp 339 // area. These are all "frame_slots". Arg_slots start at the zero 340 // stack_slots and count up to the known arg_size. Frame_slots start at 341 // the stack_slot #arg_size and go up. After allocation I map stack 342 // slots to actual offsets. Stack-slots in the arg_slot area are biased 343 // by the frame_size; stack-slots in the frame_slot area are biased by 0. 344 345 _trip_cnt = 0; 346 _alternate = 0; 347 _matcher._allocation_started = true; 348 349 ResourceArea split_arena(mtCompiler); // Arena for Split local resources 350 ResourceArea live_arena(mtCompiler); // Arena for liveness & IFG info 351 ResourceMark rm(&live_arena); 352 353 // Need live-ness for the IFG; need the IFG for coalescing. If the 354 // liveness is JUST for coalescing, then I can get some mileage by renaming 355 // all copy-related live ranges low and then using the max copy-related 356 // live range as a cut-off for LIVE and the IFG. In other words, I can 357 // build a subset of LIVE and IFG just for copies. 358 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false); 359 360 // Need IFG for coalescing and coloring 361 PhaseIFG ifg(&live_arena); 362 _ifg = &ifg; 363 364 // Come out of SSA world to the Named world. Assign (virtual) registers to 365 // Nodes. Use the same register for all inputs and the output of PhiNodes 366 // - effectively ending SSA form. This requires either coalescing live 367 // ranges or inserting copies. For the moment, we insert "virtual copies" 368 // - we pretend there is a copy prior to each Phi in predecessor blocks. 369 // We will attempt to coalesce such "virtual copies" before we manifest 370 // them for real. 371 de_ssa(); 372 373 #ifdef ASSERT 374 // Veify the graph before RA. 375 verify(&live_arena); 376 #endif 377 378 { 379 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 380 _live = NULL; // Mark live as being not available 381 rm.reset_to_mark(); // Reclaim working storage 382 IndexSet::reset_memory(C, &live_arena); 383 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG 384 gather_lrg_masks( false ); // Collect LRG masks 385 live.compute(_lrg_map.max_lrg_id()); // Compute liveness 386 _live = &live; // Mark LIVE as being available 387 } 388 389 // Base pointers are currently "used" by instructions which define new 390 // derived pointers. This makes base pointers live up to the where the 391 // derived pointer is made, but not beyond. Really, they need to be live 392 // across any GC point where the derived value is live. So this code looks 393 // at all the GC points, and "stretches" the live range of any base pointer 394 // to the GC point. 395 if (stretch_base_pointer_live_ranges(&live_arena)) { 396 Compile::TracePhase tp("computeLive (sbplr)", &timers[_t_computeLive]); 397 // Since some live range stretched, I need to recompute live 398 _live = NULL; 399 rm.reset_to_mark(); // Reclaim working storage 400 IndexSet::reset_memory(C, &live_arena); 401 ifg.init(_lrg_map.max_lrg_id()); 402 gather_lrg_masks(false); 403 live.compute(_lrg_map.max_lrg_id()); 404 _live = &live; 405 } 406 // Create the interference graph using virtual copies 407 build_ifg_virtual(); // Include stack slots this time 408 409 // The IFG is/was triangular. I am 'squaring it up' so Union can run 410 // faster. Union requires a 'for all' operation which is slow on the 411 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - 412 // meaning I can visit all the Nodes neighbors less than a Node in time 413 // O(# of neighbors), but I have to visit all the Nodes greater than a 414 // given Node and search them for an instance, i.e., time O(#MaxLRG)). 415 _ifg->SquareUp(); 416 417 // Aggressive (but pessimistic) copy coalescing. 418 // This pass works on virtual copies. Any virtual copies which are not 419 // coalesced get manifested as actual copies 420 { 421 Compile::TracePhase tp("chaitinCoalesce1", &timers[_t_chaitinCoalesce1]); 422 423 PhaseAggressiveCoalesce coalesce(*this); 424 coalesce.coalesce_driver(); 425 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do 426 // not match the Phi itself, insert a copy. 427 coalesce.insert_copies(_matcher); 428 if (C->failing()) { 429 return; 430 } 431 } 432 433 // After aggressive coalesce, attempt a first cut at coloring. 434 // To color, we need the IFG and for that we need LIVE. 435 { 436 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 437 _live = NULL; 438 rm.reset_to_mark(); // Reclaim working storage 439 IndexSet::reset_memory(C, &live_arena); 440 ifg.init(_lrg_map.max_lrg_id()); 441 gather_lrg_masks( true ); 442 live.compute(_lrg_map.max_lrg_id()); 443 _live = &live; 444 } 445 446 // Build physical interference graph 447 uint must_spill = 0; 448 must_spill = build_ifg_physical(&live_arena); 449 // If we have a guaranteed spill, might as well spill now 450 if (must_spill) { 451 if(!_lrg_map.max_lrg_id()) { 452 return; 453 } 454 // Bail out if unique gets too large (ie - unique > MaxNodeLimit) 455 C->check_node_count(10*must_spill, "out of nodes before split"); 456 if (C->failing()) { 457 return; 458 } 459 460 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 461 _lrg_map.set_max_lrg_id(new_max_lrg_id); 462 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 463 // or we failed to split 464 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); 465 if (C->failing()) { 466 return; 467 } 468 469 NOT_PRODUCT(C->verify_graph_edges();) 470 471 compact(); // Compact LRGs; return new lower max lrg 472 473 { 474 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 475 _live = NULL; 476 rm.reset_to_mark(); // Reclaim working storage 477 IndexSet::reset_memory(C, &live_arena); 478 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph 479 gather_lrg_masks( true ); // Collect intersect mask 480 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE 481 _live = &live; 482 } 483 build_ifg_physical(&live_arena); 484 _ifg->SquareUp(); 485 _ifg->Compute_Effective_Degree(); 486 // Only do conservative coalescing if requested 487 if (OptoCoalesce) { 488 Compile::TracePhase tp("chaitinCoalesce2", &timers[_t_chaitinCoalesce2]); 489 // Conservative (and pessimistic) copy coalescing of those spills 490 PhaseConservativeCoalesce coalesce(*this); 491 // If max live ranges greater than cutoff, don't color the stack. 492 // This cutoff can be larger than below since it is only done once. 493 coalesce.coalesce_driver(); 494 } 495 _lrg_map.compress_uf_map_for_nodes(); 496 497 #ifdef ASSERT 498 verify(&live_arena, true); 499 #endif 500 } else { 501 ifg.SquareUp(); 502 ifg.Compute_Effective_Degree(); 503 #ifdef ASSERT 504 set_was_low(); 505 #endif 506 } 507 508 // Prepare for Simplify & Select 509 cache_lrg_info(); // Count degree of LRGs 510 511 // Simplify the InterFerence Graph by removing LRGs of low degree. 512 // LRGs of low degree are trivially colorable. 513 Simplify(); 514 515 // Select colors by re-inserting LRGs back into the IFG in reverse order. 516 // Return whether or not something spills. 517 uint spills = Select( ); 518 519 // If we spill, split and recycle the entire thing 520 while( spills ) { 521 if( _trip_cnt++ > 24 ) { 522 DEBUG_ONLY( dump_for_spill_split_recycle(); ) 523 if( _trip_cnt > 27 ) { 524 C->record_method_not_compilable("failed spill-split-recycle sanity check"); 525 return; 526 } 527 } 528 529 if (!_lrg_map.max_lrg_id()) { 530 return; 531 } 532 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 533 _lrg_map.set_max_lrg_id(new_max_lrg_id); 534 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 535 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split"); 536 if (C->failing()) { 537 return; 538 } 539 540 compact(); // Compact LRGs; return new lower max lrg 541 542 // Nuke the live-ness and interference graph and LiveRanGe info 543 { 544 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 545 _live = NULL; 546 rm.reset_to_mark(); // Reclaim working storage 547 IndexSet::reset_memory(C, &live_arena); 548 ifg.init(_lrg_map.max_lrg_id()); 549 550 // Create LiveRanGe array. 551 // Intersect register masks for all USEs and DEFs 552 gather_lrg_masks(true); 553 live.compute(_lrg_map.max_lrg_id()); 554 _live = &live; 555 } 556 must_spill = build_ifg_physical(&live_arena); 557 _ifg->SquareUp(); 558 _ifg->Compute_Effective_Degree(); 559 560 // Only do conservative coalescing if requested 561 if (OptoCoalesce) { 562 Compile::TracePhase tp("chaitinCoalesce3", &timers[_t_chaitinCoalesce3]); 563 // Conservative (and pessimistic) copy coalescing 564 PhaseConservativeCoalesce coalesce(*this); 565 // Check for few live ranges determines how aggressive coalesce is. 566 coalesce.coalesce_driver(); 567 } 568 _lrg_map.compress_uf_map_for_nodes(); 569 #ifdef ASSERT 570 verify(&live_arena, true); 571 #endif 572 cache_lrg_info(); // Count degree of LRGs 573 574 // Simplify the InterFerence Graph by removing LRGs of low degree. 575 // LRGs of low degree are trivially colorable. 576 Simplify(); 577 578 // Select colors by re-inserting LRGs back into the IFG in reverse order. 579 // Return whether or not something spills. 580 spills = Select(); 581 } 582 583 // Count number of Simplify-Select trips per coloring success. 584 _allocator_attempts += _trip_cnt + 1; 585 _allocator_successes += 1; 586 587 // Peephole remove copies 588 post_allocate_copy_removal(); 589 590 // Merge multidefs if multiple defs representing the same value are used in a single block. 591 merge_multidefs(); 592 593 #ifdef ASSERT 594 // Veify the graph after RA. 595 verify(&live_arena); 596 #endif 597 598 // max_reg is past the largest *register* used. 599 // Convert that to a frame_slot number. 600 if (_max_reg <= _matcher._new_SP) { 601 _framesize = C->out_preserve_stack_slots(); 602 } 603 else { 604 _framesize = _max_reg -_matcher._new_SP; 605 } 606 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); 607 608 // This frame must preserve the required fp alignment 609 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots()); 610 assert(_framesize <= 1000000, "sanity check"); 611 #ifndef PRODUCT 612 _total_framesize += _framesize; 613 if ((int)_framesize > _max_framesize) { 614 _max_framesize = _framesize; 615 } 616 #endif 617 618 // Convert CISC spills 619 fixup_spills(); 620 621 // Log regalloc results 622 CompileLog* log = Compile::current()->log(); 623 if (log != NULL) { 624 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); 625 } 626 627 if (C->failing()) { 628 return; 629 } 630 631 NOT_PRODUCT(C->verify_graph_edges();) 632 633 // Move important info out of the live_arena to longer lasting storage. 634 alloc_node_regs(_lrg_map.size()); 635 for (uint i=0; i < _lrg_map.size(); i++) { 636 if (_lrg_map.live_range_id(i)) { // Live range associated with Node? 637 LRG &lrg = lrgs(_lrg_map.live_range_id(i)); 638 if (!lrg.alive()) { 639 set_bad(i); 640 } else if (lrg.num_regs() == 1) { 641 set1(i, lrg.reg()); 642 } else { // Must be a register-set 643 if (!lrg._fat_proj) { // Must be aligned adjacent register set 644 // Live ranges record the highest register in their mask. 645 // We want the low register for the AD file writer's convenience. 646 OptoReg::Name hi = lrg.reg(); // Get hi register 647 OptoReg::Name lo = OptoReg::add(hi, (1-lrg.num_regs())); // Find lo 648 // We have to use pair [lo,lo+1] even for wide vectors because 649 // the rest of code generation works only with pairs. It is safe 650 // since for registers encoding only 'lo' is used. 651 // Second reg from pair is used in ScheduleAndBundle on SPARC where 652 // vector max size is 8 which corresponds to registers pair. 653 // It is also used in BuildOopMaps but oop operations are not 654 // vectorized. 655 set2(i, lo); 656 } else { // Misaligned; extract 2 bits 657 OptoReg::Name hi = lrg.reg(); // Get hi register 658 lrg.Remove(hi); // Yank from mask 659 int lo = lrg.mask().find_first_elem(); // Find lo 660 set_pair(i, hi, lo); 661 } 662 } 663 if( lrg._is_oop ) _node_oops.set(i); 664 } else { 665 set_bad(i); 666 } 667 } 668 669 // Done! 670 _live = NULL; 671 _ifg = NULL; 672 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope 673 } 674 675 void PhaseChaitin::de_ssa() { 676 // Set initial Names for all Nodes. Most Nodes get the virtual register 677 // number. A few get the ZERO live range number. These do not 678 // get allocated, but instead rely on correct scheduling to ensure that 679 // only one instance is simultaneously live at a time. 680 uint lr_counter = 1; 681 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 682 Block* block = _cfg.get_block(i); 683 uint cnt = block->number_of_nodes(); 684 685 // Handle all the normal Nodes in the block 686 for( uint j = 0; j < cnt; j++ ) { 687 Node *n = block->get_node(j); 688 // Pre-color to the zero live range, or pick virtual register 689 const RegMask &rm = n->out_RegMask(); 690 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0); 691 } 692 } 693 694 // Reset the Union-Find mapping to be identity 695 _lrg_map.reset_uf_map(lr_counter); 696 } 697 698 void PhaseChaitin::mark_ssa() { 699 // Use ssa names to populate the live range maps or if no mask 700 // is available, use the 0 entry. 701 uint max_idx = 0; 702 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 703 Block* block = _cfg.get_block(i); 704 uint cnt = block->number_of_nodes(); 705 706 // Handle all the normal Nodes in the block 707 for ( uint j = 0; j < cnt; j++ ) { 708 Node *n = block->get_node(j); 709 // Pre-color to the zero live range, or pick virtual register 710 const RegMask &rm = n->out_RegMask(); 711 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0); 712 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx; 713 } 714 } 715 _lrg_map.set_max_lrg_id(max_idx+1); 716 717 // Reset the Union-Find mapping to be identity 718 _lrg_map.reset_uf_map(max_idx+1); 719 } 720 721 722 // Gather LiveRanGe information, including register masks. Modification of 723 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. 724 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { 725 726 // Nail down the frame pointer live range 727 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr)); 728 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite 729 730 // For all blocks 731 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 732 Block* block = _cfg.get_block(i); 733 734 // For all instructions 735 for (uint j = 1; j < block->number_of_nodes(); j++) { 736 Node* n = block->get_node(j); 737 uint input_edge_start =1; // Skip control most nodes 738 bool is_machine_node = false; 739 if (n->is_Mach()) { 740 is_machine_node = true; 741 input_edge_start = n->as_Mach()->oper_input_base(); 742 } 743 uint idx = n->is_Copy(); 744 745 // Get virtual register number, same as LiveRanGe index 746 uint vreg = _lrg_map.live_range_id(n); 747 LRG& lrg = lrgs(vreg); 748 if (vreg) { // No vreg means un-allocable (e.g. memory) 749 750 // Check for float-vs-int live range (used in register-pressure 751 // calculations) 752 const Type *n_type = n->bottom_type(); 753 if (n_type->is_floatingpoint()) { 754 lrg._is_float = 1; 755 } 756 757 // Check for twice prior spilling. Once prior spilling might have 758 // spilled 'soft', 2nd prior spill should have spilled 'hard' and 759 // further spilling is unlikely to make progress. 760 if (_spilled_once.test(n->_idx)) { 761 lrg._was_spilled1 = 1; 762 if (_spilled_twice.test(n->_idx)) { 763 lrg._was_spilled2 = 1; 764 } 765 } 766 767 #ifndef PRODUCT 768 // Collect bits not used by product code, but which may be useful for 769 // debugging. 770 771 // Collect has-copy bit 772 if (idx) { 773 lrg._has_copy = 1; 774 uint clidx = _lrg_map.live_range_id(n->in(idx)); 775 LRG& copy_src = lrgs(clidx); 776 copy_src._has_copy = 1; 777 } 778 779 if (trace_spilling() && lrg._def != NULL) { 780 // collect defs for MultiDef printing 781 if (lrg._defs == NULL) { 782 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL); 783 lrg._defs->append(lrg._def); 784 } 785 lrg._defs->append(n); 786 } 787 #endif 788 789 // Check for a single def LRG; these can spill nicely 790 // via rematerialization. Flag as NULL for no def found 791 // yet, or 'n' for single def or -1 for many defs. 792 lrg._def = lrg._def ? NodeSentinel : n; 793 794 // Limit result register mask to acceptable registers 795 const RegMask &rm = n->out_RegMask(); 796 lrg.AND( rm ); 797 798 uint ireg = n->ideal_reg(); 799 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, 800 "oops must be in Op_RegP's" ); 801 802 // Check for vector live range (only if vector register is used). 803 // On SPARC vector uses RegD which could be misaligned so it is not 804 // processes as vector in RA. 805 if (RegMask::is_vector(ireg)) 806 lrg._is_vector = 1; 807 assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL, 808 "vector must be in vector registers"); 809 810 // Check for bound register masks 811 const RegMask &lrgmask = lrg.mask(); 812 if (lrgmask.is_bound(ireg)) { 813 lrg._is_bound = 1; 814 } 815 816 // Check for maximum frequency value 817 if (lrg._maxfreq < block->_freq) { 818 lrg._maxfreq = block->_freq; 819 } 820 821 // Check for oop-iness, or long/double 822 // Check for multi-kill projection 823 switch (ireg) { 824 case MachProjNode::fat_proj: 825 // Fat projections have size equal to number of registers killed 826 lrg.set_num_regs(rm.Size()); 827 lrg.set_reg_pressure(lrg.num_regs()); 828 lrg._fat_proj = 1; 829 lrg._is_bound = 1; 830 break; 831 case Op_RegP: 832 #ifdef _LP64 833 lrg.set_num_regs(2); // Size is 2 stack words 834 #else 835 lrg.set_num_regs(1); // Size is 1 stack word 836 #endif 837 // Register pressure is tracked relative to the maximum values 838 // suggested for that platform, INTPRESSURE and FLOATPRESSURE, 839 // and relative to other types which compete for the same regs. 840 // 841 // The following table contains suggested values based on the 842 // architectures as defined in each .ad file. 843 // INTPRESSURE and FLOATPRESSURE may be tuned differently for 844 // compile-speed or performance. 845 // Note1: 846 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 847 // since .ad registers are defined as high and low halves. 848 // These reg_pressure values remain compatible with the code 849 // in is_high_pressure() which relates get_invalid_mask_size(), 850 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. 851 // Note2: 852 // SPARC -d32 has 24 registers available for integral values, 853 // but only 10 of these are safe for 64-bit longs. 854 // Using set_reg_pressure(2) for both int and long means 855 // the allocator will believe it can fit 26 longs into 856 // registers. Using 2 for longs and 1 for ints means the 857 // allocator will attempt to put 52 integers into registers. 858 // The settings below limit this problem to methods with 859 // many long values which are being run on 32-bit SPARC. 860 // 861 // ------------------- reg_pressure -------------------- 862 // Each entry is reg_pressure_per_value,number_of_regs 863 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE 864 // IA32 2 1 1 1 1 6 6 865 // IA64 1 1 1 1 1 50 41 866 // SPARC 2 2 2 2 2 48 (24) 52 (26) 867 // SPARCV9 2 2 2 2 2 48 (24) 52 (26) 868 // AMD64 1 1 1 1 1 14 15 869 // ----------------------------------------------------- 870 lrg.set_reg_pressure(1); // normally one value per register 871 if( n_type->isa_oop_ptr() ) { 872 lrg._is_oop = 1; 873 } 874 break; 875 case Op_RegL: // Check for long or double 876 case Op_RegD: 877 lrg.set_num_regs(2); 878 // Define platform specific register pressure 879 #if defined(ARM32) 880 lrg.set_reg_pressure(2); 881 #elif defined(IA32) 882 if( ireg == Op_RegL ) { 883 lrg.set_reg_pressure(2); 884 } else { 885 lrg.set_reg_pressure(1); 886 } 887 #else 888 lrg.set_reg_pressure(1); // normally one value per register 889 #endif 890 // If this def of a double forces a mis-aligned double, 891 // flag as '_fat_proj' - really flag as allowing misalignment 892 // AND changes how we count interferences. A mis-aligned 893 // double can interfere with TWO aligned pairs, or effectively 894 // FOUR registers! 895 if (rm.is_misaligned_pair()) { 896 lrg._fat_proj = 1; 897 lrg._is_bound = 1; 898 } 899 break; 900 case Op_RegF: 901 case Op_RegI: 902 case Op_RegN: 903 case Op_RegFlags: 904 case 0: // not an ideal register 905 lrg.set_num_regs(1); 906 lrg.set_reg_pressure(1); 907 break; 908 case Op_VecS: 909 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity"); 910 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity"); 911 lrg.set_num_regs(RegMask::SlotsPerVecS); 912 lrg.set_reg_pressure(1); 913 break; 914 case Op_VecD: 915 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity"); 916 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity"); 917 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned"); 918 lrg.set_num_regs(RegMask::SlotsPerVecD); 919 lrg.set_reg_pressure(1); 920 break; 921 case Op_VecX: 922 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity"); 923 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity"); 924 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned"); 925 lrg.set_num_regs(RegMask::SlotsPerVecX); 926 lrg.set_reg_pressure(1); 927 break; 928 case Op_VecY: 929 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity"); 930 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity"); 931 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned"); 932 lrg.set_num_regs(RegMask::SlotsPerVecY); 933 lrg.set_reg_pressure(1); 934 break; 935 case Op_VecZ: 936 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity"); 937 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity"); 938 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned"); 939 lrg.set_num_regs(RegMask::SlotsPerVecZ); 940 lrg.set_reg_pressure(1); 941 break; 942 default: 943 ShouldNotReachHere(); 944 } 945 } 946 947 // Now do the same for inputs 948 uint cnt = n->req(); 949 // Setup for CISC SPILLING 950 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 951 if( UseCISCSpill && after_aggressive ) { 952 inp = n->cisc_operand(); 953 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 954 // Convert operand number to edge index number 955 inp = n->as_Mach()->operand_index(inp); 956 } 957 958 // Prepare register mask for each input 959 for( uint k = input_edge_start; k < cnt; k++ ) { 960 uint vreg = _lrg_map.live_range_id(n->in(k)); 961 if (!vreg) { 962 continue; 963 } 964 965 // If this instruction is CISC Spillable, add the flags 966 // bit to its appropriate input 967 if( UseCISCSpill && after_aggressive && inp == k ) { 968 #ifndef PRODUCT 969 if( TraceCISCSpill ) { 970 tty->print(" use_cisc_RegMask: "); 971 n->dump(); 972 } 973 #endif 974 n->as_Mach()->use_cisc_RegMask(); 975 } 976 977 if (is_machine_node && _scheduling_info_generated) { 978 MachNode* cur_node = n->as_Mach(); 979 // this is cleaned up by register allocation 980 if (k >= cur_node->num_opnds()) continue; 981 } 982 983 LRG &lrg = lrgs(vreg); 984 // // Testing for floating point code shape 985 // Node *test = n->in(k); 986 // if( test->is_Mach() ) { 987 // MachNode *m = test->as_Mach(); 988 // int op = m->ideal_Opcode(); 989 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 990 // int zzz = 1; 991 // } 992 // } 993 994 // Limit result register mask to acceptable registers. 995 // Do not limit registers from uncommon uses before 996 // AggressiveCoalesce. This effectively pre-virtual-splits 997 // around uncommon uses of common defs. 998 const RegMask &rm = n->in_RegMask(k); 999 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { 1000 // Since we are BEFORE aggressive coalesce, leave the register 1001 // mask untrimmed by the call. This encourages more coalescing. 1002 // Later, AFTER aggressive, this live range will have to spill 1003 // but the spiller handles slow-path calls very nicely. 1004 } else { 1005 lrg.AND( rm ); 1006 } 1007 1008 // Check for bound register masks 1009 const RegMask &lrgmask = lrg.mask(); 1010 uint kreg = n->in(k)->ideal_reg(); 1011 bool is_vect = RegMask::is_vector(kreg); 1012 assert(n->in(k)->bottom_type()->isa_vect() == NULL || 1013 is_vect || kreg == Op_RegD || kreg == Op_RegL, 1014 "vector must be in vector registers"); 1015 if (lrgmask.is_bound(kreg)) 1016 lrg._is_bound = 1; 1017 1018 // If this use of a double forces a mis-aligned double, 1019 // flag as '_fat_proj' - really flag as allowing misalignment 1020 // AND changes how we count interferences. A mis-aligned 1021 // double can interfere with TWO aligned pairs, or effectively 1022 // FOUR registers! 1023 #ifdef ASSERT 1024 if (is_vect && !_scheduling_info_generated) { 1025 if (lrg.num_regs() != 0) { 1026 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned"); 1027 assert(!lrg._fat_proj, "sanity"); 1028 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity"); 1029 } else { 1030 assert(n->is_Phi(), "not all inputs processed only if Phi"); 1031 } 1032 } 1033 #endif 1034 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { 1035 lrg._fat_proj = 1; 1036 lrg._is_bound = 1; 1037 } 1038 // if the LRG is an unaligned pair, we will have to spill 1039 // so clear the LRG's register mask if it is not already spilled 1040 if (!is_vect && !n->is_SpillCopy() && 1041 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 1042 lrgmask.is_misaligned_pair()) { 1043 lrg.Clear(); 1044 } 1045 1046 // Check for maximum frequency value 1047 if (lrg._maxfreq < block->_freq) { 1048 lrg._maxfreq = block->_freq; 1049 } 1050 1051 } // End for all allocated inputs 1052 } // end for all instructions 1053 } // end for all blocks 1054 1055 // Final per-liverange setup 1056 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { 1057 LRG &lrg = lrgs(i2); 1058 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); 1059 if (lrg.num_regs() > 1 && !lrg._fat_proj) { 1060 lrg.clear_to_sets(); 1061 } 1062 lrg.compute_set_mask_size(); 1063 if (lrg.not_free()) { // Handle case where we lose from the start 1064 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 1065 lrg._direct_conflict = 1; 1066 } 1067 lrg.set_degree(0); // no neighbors in IFG yet 1068 } 1069 } 1070 1071 // Set the was-lo-degree bit. Conservative coalescing should not change the 1072 // colorability of the graph. If any live range was of low-degree before 1073 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 1074 // The bit is checked in Simplify. 1075 void PhaseChaitin::set_was_low() { 1076 #ifdef ASSERT 1077 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1078 int size = lrgs(i).num_regs(); 1079 uint old_was_lo = lrgs(i)._was_lo; 1080 lrgs(i)._was_lo = 0; 1081 if( lrgs(i).lo_degree() ) { 1082 lrgs(i)._was_lo = 1; // Trivially of low degree 1083 } else { // Else check the Brigg's assertion 1084 // Brigg's observation is that the lo-degree neighbors of a 1085 // hi-degree live range will not interfere with the color choices 1086 // of said hi-degree live range. The Simplify reverse-stack-coloring 1087 // order takes care of the details. Hence you do not have to count 1088 // low-degree neighbors when determining if this guy colors. 1089 int briggs_degree = 0; 1090 IndexSet *s = _ifg->neighbors(i); 1091 IndexSetIterator elements(s); 1092 uint lidx; 1093 while((lidx = elements.next()) != 0) { 1094 if( !lrgs(lidx).lo_degree() ) 1095 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 1096 } 1097 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 1098 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 1099 } 1100 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 1101 } 1102 #endif 1103 } 1104 1105 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 1106 void PhaseChaitin::cache_lrg_info( ) { 1107 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]); 1108 1109 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1110 LRG &lrg = lrgs(i); 1111 1112 // Check for being of low degree: means we can be trivially colored. 1113 // Low degree, dead or must-spill guys just get to simplify right away 1114 if( lrg.lo_degree() || 1115 !lrg.alive() || 1116 lrg._must_spill ) { 1117 // Split low degree list into those guys that must get a 1118 // register and those that can go to register or stack. 1119 // The idea is LRGs that can go register or stack color first when 1120 // they have a good chance of getting a register. The register-only 1121 // lo-degree live ranges always get a register. 1122 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 1123 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 1124 lrg._next = _lo_stk_degree; 1125 _lo_stk_degree = i; 1126 } else { 1127 lrg._next = _lo_degree; 1128 _lo_degree = i; 1129 } 1130 } else { // Else high degree 1131 lrgs(_hi_degree)._prev = i; 1132 lrg._next = _hi_degree; 1133 lrg._prev = 0; 1134 _hi_degree = i; 1135 } 1136 } 1137 } 1138 1139 // Simplify the IFG by removing LRGs of low degree. 1140 void PhaseChaitin::Simplify( ) { 1141 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]); 1142 1143 while( 1 ) { // Repeat till simplified it all 1144 // May want to explore simplifying lo_degree before _lo_stk_degree. 1145 // This might result in more spills coloring into registers during 1146 // Select(). 1147 while( _lo_degree || _lo_stk_degree ) { 1148 // If possible, pull from lo_stk first 1149 uint lo; 1150 if( _lo_degree ) { 1151 lo = _lo_degree; 1152 _lo_degree = lrgs(lo)._next; 1153 } else { 1154 lo = _lo_stk_degree; 1155 _lo_stk_degree = lrgs(lo)._next; 1156 } 1157 1158 // Put the simplified guy on the simplified list. 1159 lrgs(lo)._next = _simplified; 1160 _simplified = lo; 1161 // If this guy is "at risk" then mark his current neighbors 1162 if (lrgs(lo)._at_risk && !_ifg->neighbors(lo)->is_empty()) { 1163 IndexSetIterator elements(_ifg->neighbors(lo)); 1164 uint datum; 1165 while ((datum = elements.next()) != 0) { 1166 lrgs(datum)._risk_bias = lo; 1167 } 1168 } 1169 1170 // Yank this guy from the IFG. 1171 IndexSet *adj = _ifg->remove_node(lo); 1172 if (adj->is_empty()) { 1173 continue; 1174 } 1175 1176 // If any neighbors' degrees fall below their number of 1177 // allowed registers, then put that neighbor on the low degree 1178 // list. Note that 'degree' can only fall and 'numregs' is 1179 // unchanged by this action. Thus the two are equal at most once, 1180 // so LRGs hit the lo-degree worklist at most once. 1181 IndexSetIterator elements(adj); 1182 uint neighbor; 1183 while ((neighbor = elements.next()) != 0) { 1184 LRG *n = &lrgs(neighbor); 1185 #ifdef ASSERT 1186 if (VerifyRegisterAllocator) { 1187 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1188 } 1189 #endif 1190 1191 // Check for just becoming of-low-degree just counting registers. 1192 // _must_spill live ranges are already on the low degree list. 1193 if (n->just_lo_degree() && !n->_must_spill) { 1194 assert(!_ifg->_yanked->test(neighbor), "Cannot move to lo degree twice"); 1195 // Pull from hi-degree list 1196 uint prev = n->_prev; 1197 uint next = n->_next; 1198 if (prev) { 1199 lrgs(prev)._next = next; 1200 } else { 1201 _hi_degree = next; 1202 } 1203 lrgs(next)._prev = prev; 1204 n->_next = _lo_degree; 1205 _lo_degree = neighbor; 1206 } 1207 } 1208 } // End of while lo-degree/lo_stk_degree worklist not empty 1209 1210 // Check for got everything: is hi-degree list empty? 1211 if (!_hi_degree) break; 1212 1213 // Time to pick a potential spill guy 1214 uint lo_score = _hi_degree; 1215 double score = lrgs(lo_score).score(); 1216 double area = lrgs(lo_score)._area; 1217 double cost = lrgs(lo_score)._cost; 1218 bool bound = lrgs(lo_score)._is_bound; 1219 1220 // Find cheapest guy 1221 debug_only( int lo_no_simplify=0; ); 1222 for (uint i = _hi_degree; i; i = lrgs(i)._next) { 1223 assert(!_ifg->_yanked->test(i), ""); 1224 // It's just vaguely possible to move hi-degree to lo-degree without 1225 // going through a just-lo-degree stage: If you remove a double from 1226 // a float live range it's degree will drop by 2 and you can skip the 1227 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 1228 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1229 if( lrgs(i).lo_degree() ) { 1230 lo_score = i; 1231 break; 1232 } 1233 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1234 double iscore = lrgs(i).score(); 1235 double iarea = lrgs(i)._area; 1236 double icost = lrgs(i)._cost; 1237 bool ibound = lrgs(i)._is_bound; 1238 1239 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1240 // wins. Ties happen because all live ranges in question have spilled 1241 // a few times before and the spill-score adds a huge number which 1242 // washes out the low order bits. We are choosing the lesser of 2 1243 // evils; in this case pick largest area to spill. 1244 // Ties also happen when live ranges are defined and used only inside 1245 // one block. In which case their area is 0 and score set to max. 1246 // In such case choose bound live range over unbound to free registers 1247 // or with smaller cost to spill. 1248 if ( iscore < score || 1249 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || 1250 (iscore == score && iarea == area && 1251 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) { 1252 lo_score = i; 1253 score = iscore; 1254 area = iarea; 1255 cost = icost; 1256 bound = ibound; 1257 } 1258 } 1259 LRG *lo_lrg = &lrgs(lo_score); 1260 // The live range we choose for spilling is either hi-degree, or very 1261 // rarely it can be low-degree. If we choose a hi-degree live range 1262 // there better not be any lo-degree choices. 1263 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1264 1265 // Pull from hi-degree list 1266 uint prev = lo_lrg->_prev; 1267 uint next = lo_lrg->_next; 1268 if( prev ) lrgs(prev)._next = next; 1269 else _hi_degree = next; 1270 lrgs(next)._prev = prev; 1271 // Jam him on the lo-degree list, despite his high degree. 1272 // Maybe he'll get a color, and maybe he'll spill. 1273 // Only Select() will know. 1274 lrgs(lo_score)._at_risk = true; 1275 _lo_degree = lo_score; 1276 lo_lrg->_next = 0; 1277 1278 } // End of while not simplified everything 1279 1280 } 1281 1282 // Is 'reg' register legal for 'lrg'? 1283 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { 1284 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && 1285 lrg.mask().Member(OptoReg::add(reg,-chunk))) { 1286 // RA uses OptoReg which represent the highest element of a registers set. 1287 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set 1288 // in which XMMd is used by RA to represent such vectors. A double value 1289 // uses [XMM,XMMb] pairs and XMMb is used by RA for it. 1290 // The register mask uses largest bits set of overlapping register sets. 1291 // On x86 with AVX it uses 8 bits for each XMM registers set. 1292 // 1293 // The 'lrg' already has cleared-to-set register mask (done in Select() 1294 // before calling choose_color()). Passing mask.Member(reg) check above 1295 // indicates that the size (num_regs) of 'reg' set is less or equal to 1296 // 'lrg' set size. 1297 // For set size 1 any register which is member of 'lrg' mask is legal. 1298 if (lrg.num_regs()==1) 1299 return true; 1300 // For larger sets only an aligned register with the same set size is legal. 1301 int mask = lrg.num_regs()-1; 1302 if ((reg&mask) == mask) 1303 return true; 1304 } 1305 return false; 1306 } 1307 1308 // Choose a color using the biasing heuristic 1309 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1310 1311 // Check for "at_risk" LRG's 1312 uint risk_lrg = _lrg_map.find(lrg._risk_bias); 1313 if (risk_lrg != 0 && !_ifg->neighbors(risk_lrg)->is_empty()) { 1314 // Walk the colored neighbors of the "at_risk" candidate 1315 // Choose a color which is both legal and already taken by a neighbor 1316 // of the "at_risk" candidate in order to improve the chances of the 1317 // "at_risk" candidate of coloring 1318 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1319 uint datum; 1320 while ((datum = elements.next()) != 0) { 1321 OptoReg::Name reg = lrgs(datum).reg(); 1322 // If this LRG's register is legal for us, choose it 1323 if (is_legal_reg(lrg, reg, chunk)) 1324 return reg; 1325 } 1326 } 1327 1328 uint copy_lrg = _lrg_map.find(lrg._copy_bias); 1329 if (copy_lrg != 0) { 1330 // If he has a color, 1331 if(!_ifg->_yanked->test(copy_lrg)) { 1332 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1333 // And it is legal for you, 1334 if (is_legal_reg(lrg, reg, chunk)) 1335 return reg; 1336 } else if( chunk == 0 ) { 1337 // Choose a color which is legal for him 1338 RegMask tempmask = lrg.mask(); 1339 tempmask.AND(lrgs(copy_lrg).mask()); 1340 tempmask.clear_to_sets(lrg.num_regs()); 1341 OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); 1342 if (OptoReg::is_valid(reg)) 1343 return reg; 1344 } 1345 } 1346 1347 // If no bias info exists, just go with the register selection ordering 1348 if (lrg._is_vector || lrg.num_regs() == 2) { 1349 // Find an aligned set 1350 return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); 1351 } 1352 1353 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1354 // copy removal to remove many more copies, by preventing a just-assigned 1355 // register from being repeatedly assigned. 1356 OptoReg::Name reg = lrg.mask().find_first_elem(); 1357 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1358 // This 'Remove; find; Insert' idiom is an expensive way to find the 1359 // SECOND element in the mask. 1360 lrg.Remove(reg); 1361 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1362 lrg.Insert(reg); 1363 if( OptoReg::is_reg(reg2)) 1364 reg = reg2; 1365 } 1366 return OptoReg::add( reg, chunk ); 1367 } 1368 1369 // Choose a color in the current chunk 1370 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1371 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)"); 1372 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)"); 1373 1374 if( lrg.num_regs() == 1 || // Common Case 1375 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1376 // Use a heuristic to "bias" the color choice 1377 return bias_color(lrg, chunk); 1378 1379 assert(!lrg._is_vector, "should be not vector here" ); 1380 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1381 1382 // Fat-proj case or misaligned double argument. 1383 assert(lrg.compute_mask_size() == lrg.num_regs() || 1384 lrg.num_regs() == 2,"fat projs exactly color" ); 1385 assert( !chunk, "always color in 1st chunk" ); 1386 // Return the highest element in the set. 1387 return lrg.mask().find_last_elem(); 1388 } 1389 1390 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1391 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1392 // everything going back is guaranteed a color. Select that color. If some 1393 // hi-degree LRG cannot get a color then we record that we must spill. 1394 uint PhaseChaitin::Select( ) { 1395 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]); 1396 1397 uint spill_reg = LRG::SPILL_REG; 1398 _max_reg = OptoReg::Name(0); // Past max register used 1399 while( _simplified ) { 1400 // Pull next LRG from the simplified list - in reverse order of removal 1401 uint lidx = _simplified; 1402 LRG *lrg = &lrgs(lidx); 1403 _simplified = lrg->_next; 1404 1405 1406 #ifndef PRODUCT 1407 if (trace_spilling()) { 1408 ttyLocker ttyl; 1409 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1410 lrg->degrees_of_freedom()); 1411 lrg->dump(); 1412 } 1413 #endif 1414 1415 // Re-insert into the IFG 1416 _ifg->re_insert(lidx); 1417 if( !lrg->alive() ) continue; 1418 // capture allstackedness flag before mask is hacked 1419 const int is_allstack = lrg->mask().is_AllStack(); 1420 1421 // Yeah, yeah, yeah, I know, I know. I can refactor this 1422 // to avoid the GOTO, although the refactored code will not 1423 // be much clearer. We arrive here IFF we have a stack-based 1424 // live range that cannot color in the current chunk, and it 1425 // has to move into the next free stack chunk. 1426 int chunk = 0; // Current chunk is first chunk 1427 retry_next_chunk: 1428 1429 // Remove neighbor colors 1430 IndexSet *s = _ifg->neighbors(lidx); 1431 debug_only(RegMask orig_mask = lrg->mask();) 1432 1433 if (!s->is_empty()) { 1434 IndexSetIterator elements(s); 1435 uint neighbor; 1436 while ((neighbor = elements.next()) != 0) { 1437 // Note that neighbor might be a spill_reg. In this case, exclusion 1438 // of its color will be a no-op, since the spill_reg chunk is in outer 1439 // space. Also, if neighbor is in a different chunk, this exclusion 1440 // will be a no-op. (Later on, if lrg runs out of possible colors in 1441 // its chunk, a new chunk of color may be tried, in which case 1442 // examination of neighbors is started again, at retry_next_chunk.) 1443 LRG &nlrg = lrgs(neighbor); 1444 OptoReg::Name nreg = nlrg.reg(); 1445 // Only subtract masks in the same chunk 1446 if (nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE) { 1447 #ifndef PRODUCT 1448 uint size = lrg->mask().Size(); 1449 RegMask rm = lrg->mask(); 1450 #endif 1451 lrg->SUBTRACT(nlrg.mask()); 1452 #ifndef PRODUCT 1453 if (trace_spilling() && lrg->mask().Size() != size) { 1454 ttyLocker ttyl; 1455 tty->print("L%d ", lidx); 1456 rm.dump(); 1457 tty->print(" intersected L%d ", neighbor); 1458 nlrg.mask().dump(); 1459 tty->print(" removed "); 1460 rm.SUBTRACT(lrg->mask()); 1461 rm.dump(); 1462 tty->print(" leaving "); 1463 lrg->mask().dump(); 1464 tty->cr(); 1465 } 1466 #endif 1467 } 1468 } 1469 } 1470 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1471 // Aligned pairs need aligned masks 1472 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1473 if (lrg->num_regs() > 1 && !lrg->_fat_proj) { 1474 lrg->clear_to_sets(); 1475 } 1476 1477 // Check if a color is available and if so pick the color 1478 OptoReg::Name reg = choose_color( *lrg, chunk ); 1479 1480 //--------------- 1481 // If we fail to color and the AllStack flag is set, trigger 1482 // a chunk-rollover event 1483 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1484 // Bump register mask up to next stack chunk 1485 chunk += RegMask::CHUNK_SIZE; 1486 lrg->Set_All(); 1487 1488 goto retry_next_chunk; 1489 } 1490 1491 //--------------- 1492 // Did we get a color? 1493 else if( OptoReg::is_valid(reg)) { 1494 #ifndef PRODUCT 1495 RegMask avail_rm = lrg->mask(); 1496 #endif 1497 1498 // Record selected register 1499 lrg->set_reg(reg); 1500 1501 if( reg >= _max_reg ) // Compute max register limit 1502 _max_reg = OptoReg::add(reg,1); 1503 // Fold reg back into normal space 1504 reg = OptoReg::add(reg,-chunk); 1505 1506 // If the live range is not bound, then we actually had some choices 1507 // to make. In this case, the mask has more bits in it than the colors 1508 // chosen. Restrict the mask to just what was picked. 1509 int n_regs = lrg->num_regs(); 1510 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1511 if (n_regs == 1 || !lrg->_fat_proj) { 1512 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity"); 1513 lrg->Clear(); // Clear the mask 1514 lrg->Insert(reg); // Set regmask to match selected reg 1515 // For vectors and pairs, also insert the low bit of the pair 1516 for (int i = 1; i < n_regs; i++) 1517 lrg->Insert(OptoReg::add(reg,-i)); 1518 lrg->set_mask_size(n_regs); 1519 } else { // Else fatproj 1520 // mask must be equal to fatproj bits, by definition 1521 } 1522 #ifndef PRODUCT 1523 if (trace_spilling()) { 1524 ttyLocker ttyl; 1525 tty->print("L%d selected ", lidx); 1526 lrg->mask().dump(); 1527 tty->print(" from "); 1528 avail_rm.dump(); 1529 tty->cr(); 1530 } 1531 #endif 1532 // Note that reg is the highest-numbered register in the newly-bound mask. 1533 } // end color available case 1534 1535 //--------------- 1536 // Live range is live and no colors available 1537 else { 1538 assert( lrg->alive(), "" ); 1539 assert( !lrg->_fat_proj || lrg->is_multidef() || 1540 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1541 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1542 1543 // Assign the special spillreg register 1544 lrg->set_reg(OptoReg::Name(spill_reg++)); 1545 // Do not empty the regmask; leave mask_size lying around 1546 // for use during Spilling 1547 #ifndef PRODUCT 1548 if( trace_spilling() ) { 1549 ttyLocker ttyl; 1550 tty->print("L%d spilling with neighbors: ", lidx); 1551 s->dump(); 1552 debug_only(tty->print(" original mask: ")); 1553 debug_only(orig_mask.dump()); 1554 dump_lrg(lidx); 1555 } 1556 #endif 1557 } // end spill case 1558 1559 } 1560 1561 return spill_reg-LRG::SPILL_REG; // Return number of spills 1562 } 1563 1564 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1565 void PhaseChaitin::set_was_spilled( Node *n ) { 1566 if( _spilled_once.test_set(n->_idx) ) 1567 _spilled_twice.set(n->_idx); 1568 } 1569 1570 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1571 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1572 void PhaseChaitin::fixup_spills() { 1573 // This function does only cisc spill work. 1574 if( !UseCISCSpill ) return; 1575 1576 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]); 1577 1578 // Grab the Frame Pointer 1579 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); 1580 1581 // For all blocks 1582 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1583 Block* block = _cfg.get_block(i); 1584 1585 // For all instructions in block 1586 uint last_inst = block->end_idx(); 1587 for (uint j = 1; j <= last_inst; j++) { 1588 Node* n = block->get_node(j); 1589 1590 // Dead instruction??? 1591 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1592 C->top() == n || // Or the random TOP node 1593 n->is_Proj(), // Or a fat-proj kill node 1594 "No dead instructions after post-alloc" ); 1595 1596 int inp = n->cisc_operand(); 1597 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1598 // Convert operand number to edge index number 1599 MachNode *mach = n->as_Mach(); 1600 inp = mach->operand_index(inp); 1601 Node *src = n->in(inp); // Value to load or store 1602 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); 1603 OptoReg::Name src_reg = lrg_cisc.reg(); 1604 // Doubles record the HIGH register of an adjacent pair. 1605 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1606 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1607 // This is a CISC Spill, get stack offset and construct new node 1608 #ifndef PRODUCT 1609 if( TraceCISCSpill ) { 1610 tty->print(" reg-instr: "); 1611 n->dump(); 1612 } 1613 #endif 1614 int stk_offset = reg2offset(src_reg); 1615 // Bailout if we might exceed node limit when spilling this instruction 1616 C->check_node_count(0, "out of nodes fixing spills"); 1617 if (C->failing()) return; 1618 // Transform node 1619 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); 1620 cisc->set_req(inp,fp); // Base register is frame pointer 1621 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1622 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1623 cisc->ins_req(1,src); // Requires a memory edge 1624 } 1625 block->map_node(cisc, j); // Insert into basic block 1626 n->subsume_by(cisc, C); // Correct graph 1627 // 1628 ++_used_cisc_instructions; 1629 #ifndef PRODUCT 1630 if( TraceCISCSpill ) { 1631 tty->print(" cisc-instr: "); 1632 cisc->dump(); 1633 } 1634 #endif 1635 } else { 1636 #ifndef PRODUCT 1637 if( TraceCISCSpill ) { 1638 tty->print(" using reg-instr: "); 1639 n->dump(); 1640 } 1641 #endif 1642 ++_unused_cisc_instructions; // input can be on stack 1643 } 1644 } 1645 1646 } // End of for all instructions 1647 1648 } // End of for all blocks 1649 } 1650 1651 // Helper to stretch above; recursively discover the base Node for a 1652 // given derived Node. Easy for AddP-related machine nodes, but needs 1653 // to be recursive for derived Phis. 1654 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1655 // See if already computed; if so return it 1656 if( derived_base_map[derived->_idx] ) 1657 return derived_base_map[derived->_idx]; 1658 1659 // See if this happens to be a base. 1660 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1661 // pointers derived from NULL! These are always along paths that 1662 // can't happen at run-time but the optimizer cannot deduce it so 1663 // we have to handle it gracefully. 1664 assert(!derived->bottom_type()->isa_narrowoop() || 1665 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1666 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1667 // If its an OOP with a non-zero offset, then it is derived. 1668 if( tj == NULL || tj->_offset == 0 ) { 1669 derived_base_map[derived->_idx] = derived; 1670 return derived; 1671 } 1672 // Derived is NULL+offset? Base is NULL! 1673 if( derived->is_Con() ) { 1674 Node *base = _matcher.mach_null(); 1675 assert(base != NULL, "sanity"); 1676 if (base->in(0) == NULL) { 1677 // Initialize it once and make it shared: 1678 // set control to _root and place it into Start block 1679 // (where top() node is placed). 1680 base->init_req(0, _cfg.get_root_node()); 1681 Block *startb = _cfg.get_block_for_node(C->top()); 1682 uint node_pos = startb->find_node(C->top()); 1683 startb->insert_node(base, node_pos); 1684 _cfg.map_node_to_block(base, startb); 1685 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet"); 1686 1687 // The loadConP0 might have projection nodes depending on architecture 1688 // Add the projection nodes to the CFG 1689 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { 1690 Node* use = base->fast_out(i); 1691 if (use->is_MachProj()) { 1692 startb->insert_node(use, ++node_pos); 1693 _cfg.map_node_to_block(use, startb); 1694 new_lrg(use, maxlrg++); 1695 } 1696 } 1697 } 1698 if (_lrg_map.live_range_id(base) == 0) { 1699 new_lrg(base, maxlrg++); 1700 } 1701 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"); 1702 derived_base_map[derived->_idx] = base; 1703 return base; 1704 } 1705 1706 // Check for AddP-related opcodes 1707 if (!derived->is_Phi()) { 1708 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name()); 1709 Node *base = derived->in(AddPNode::Base); 1710 derived_base_map[derived->_idx] = base; 1711 return base; 1712 } 1713 1714 // Recursively find bases for Phis. 1715 // First check to see if we can avoid a base Phi here. 1716 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1717 uint i; 1718 for( i = 2; i < derived->req(); i++ ) 1719 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1720 break; 1721 // Went to the end without finding any different bases? 1722 if( i == derived->req() ) { // No need for a base Phi here 1723 derived_base_map[derived->_idx] = base; 1724 return base; 1725 } 1726 1727 // Now we see we need a base-Phi here to merge the bases 1728 const Type *t = base->bottom_type(); 1729 base = new PhiNode( derived->in(0), t ); 1730 for( i = 1; i < derived->req(); i++ ) { 1731 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1732 t = t->meet(base->in(i)->bottom_type()); 1733 } 1734 base->as_Phi()->set_type(t); 1735 1736 // Search the current block for an existing base-Phi 1737 Block *b = _cfg.get_block_for_node(derived); 1738 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1739 Node *phi = b->get_node(i); 1740 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1741 b->insert_node(base, i); // Must insert created Phi here as base 1742 _cfg.map_node_to_block(base, b); 1743 new_lrg(base,maxlrg++); 1744 break; 1745 } 1746 // See if Phi matches. 1747 uint j; 1748 for( j = 1; j < base->req(); j++ ) 1749 if( phi->in(j) != base->in(j) && 1750 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1751 break; 1752 if( j == base->req() ) { // All inputs match? 1753 base = phi; // Then use existing 'phi' and drop 'base' 1754 break; 1755 } 1756 } 1757 1758 1759 // Cache info for later passes 1760 derived_base_map[derived->_idx] = base; 1761 return base; 1762 } 1763 1764 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1765 // base pointer that is live across the Safepoint for oopmap building. The 1766 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1767 // required edge set. 1768 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { 1769 int must_recompute_live = false; 1770 uint maxlrg = _lrg_map.max_lrg_id(); 1771 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1772 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1773 1774 // For all blocks in RPO do... 1775 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1776 Block* block = _cfg.get_block(i); 1777 // Note use of deep-copy constructor. I cannot hammer the original 1778 // liveout bits, because they are needed by the following coalesce pass. 1779 IndexSet liveout(_live->live(block)); 1780 1781 for (uint j = block->end_idx() + 1; j > 1; j--) { 1782 Node* n = block->get_node(j - 1); 1783 1784 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1785 // like to see in the same register. Compare uses the loop-phi and so 1786 // extends its live range BUT cannot be part of the cycle. If this 1787 // extended live range overlaps with the update of the loop-phi value 1788 // we need both alive at the same time -- which requires at least 1 1789 // copy. But because Intel has only 2-address registers we end up with 1790 // at least 2 copies, one before the loop-phi update instruction and 1791 // one after. Instead we split the input to the compare just after the 1792 // phi. 1793 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1794 Node *phi = n->in(1); 1795 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1796 Block *phi_block = _cfg.get_block_for_node(phi); 1797 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { 1798 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1799 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); 1800 insert_proj( phi_block, 1, spill, maxlrg++ ); 1801 n->set_req(1,spill); 1802 must_recompute_live = true; 1803 } 1804 } 1805 } 1806 1807 // Get value being defined 1808 uint lidx = _lrg_map.live_range_id(n); 1809 // Ignore the occasional brand-new live range 1810 if (lidx && lidx < _lrg_map.max_lrg_id()) { 1811 // Remove from live-out set 1812 liveout.remove(lidx); 1813 1814 // Copies do not define a new value and so do not interfere. 1815 // Remove the copies source from the liveout set before interfering. 1816 uint idx = n->is_Copy(); 1817 if (idx) { 1818 liveout.remove(_lrg_map.live_range_id(n->in(idx))); 1819 } 1820 } 1821 1822 // Found a safepoint? 1823 JVMState *jvms = n->jvms(); 1824 if (jvms && !liveout.is_empty()) { 1825 // Now scan for a live derived pointer 1826 IndexSetIterator elements(&liveout); 1827 uint neighbor; 1828 while ((neighbor = elements.next()) != 0) { 1829 // Find reaching DEF for base and derived values 1830 // This works because we are still in SSA during this call. 1831 Node *derived = lrgs(neighbor)._def; 1832 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1833 assert(!derived->bottom_type()->isa_narrowoop() || 1834 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1835 // If its an OOP with a non-zero offset, then it is derived. 1836 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { 1837 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); 1838 assert(base->_idx < _lrg_map.size(), ""); 1839 // Add reaching DEFs of derived pointer and base pointer as a 1840 // pair of inputs 1841 n->add_req(derived); 1842 n->add_req(base); 1843 1844 // See if the base pointer is already live to this point. 1845 // Since I'm working on the SSA form, live-ness amounts to 1846 // reaching def's. So if I find the base's live range then 1847 // I know the base's def reaches here. 1848 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or 1849 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND 1850 (_lrg_map.live_range_id(base) > 0) && // not a constant 1851 _cfg.get_block_for_node(base) != block) { // base not def'd in blk) 1852 // Base pointer is not currently live. Since I stretched 1853 // the base pointer to here and it crosses basic-block 1854 // boundaries, the global live info is now incorrect. 1855 // Recompute live. 1856 must_recompute_live = true; 1857 } // End of if base pointer is not live to debug info 1858 } 1859 } // End of scan all live data for derived ptrs crossing GC point 1860 } // End of if found a GC point 1861 1862 // Make all inputs live 1863 if (!n->is_Phi()) { // Phi function uses come from prior block 1864 for (uint k = 1; k < n->req(); k++) { 1865 uint lidx = _lrg_map.live_range_id(n->in(k)); 1866 if (lidx < _lrg_map.max_lrg_id()) { 1867 liveout.insert(lidx); 1868 } 1869 } 1870 } 1871 1872 } // End of forall instructions in block 1873 liveout.clear(); // Free the memory used by liveout. 1874 1875 } // End of forall blocks 1876 _lrg_map.set_max_lrg_id(maxlrg); 1877 1878 // If I created a new live range I need to recompute live 1879 if (maxlrg != _ifg->_maxlrg) { 1880 must_recompute_live = true; 1881 } 1882 1883 return must_recompute_live != 0; 1884 } 1885 1886 // Extend the node to LRG mapping 1887 1888 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { 1889 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); 1890 } 1891 1892 #ifndef PRODUCT 1893 void PhaseChaitin::dump(const Node* n) const { 1894 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; 1895 tty->print("L%d",r); 1896 if (r && n->Opcode() != Op_Phi) { 1897 if( _node_regs ) { // Got a post-allocation copy of allocation? 1898 tty->print("["); 1899 OptoReg::Name second = get_reg_second(n); 1900 if( OptoReg::is_valid(second) ) { 1901 if( OptoReg::is_reg(second) ) 1902 tty->print("%s:",Matcher::regName[second]); 1903 else 1904 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1905 } 1906 OptoReg::Name first = get_reg_first(n); 1907 if( OptoReg::is_reg(first) ) 1908 tty->print("%s]",Matcher::regName[first]); 1909 else 1910 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1911 } else 1912 n->out_RegMask().dump(); 1913 } 1914 tty->print("/N%d\t",n->_idx); 1915 tty->print("%s === ", n->Name()); 1916 uint k; 1917 for (k = 0; k < n->req(); k++) { 1918 Node *m = n->in(k); 1919 if (!m) { 1920 tty->print("_ "); 1921 } 1922 else { 1923 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1924 tty->print("L%d",r); 1925 // Data MultiNode's can have projections with no real registers. 1926 // Don't die while dumping them. 1927 int op = n->Opcode(); 1928 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1929 if( _node_regs ) { 1930 tty->print("["); 1931 OptoReg::Name second = get_reg_second(n->in(k)); 1932 if( OptoReg::is_valid(second) ) { 1933 if( OptoReg::is_reg(second) ) 1934 tty->print("%s:",Matcher::regName[second]); 1935 else 1936 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 1937 reg2offset_unchecked(second)); 1938 } 1939 OptoReg::Name first = get_reg_first(n->in(k)); 1940 if( OptoReg::is_reg(first) ) 1941 tty->print("%s]",Matcher::regName[first]); 1942 else 1943 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 1944 reg2offset_unchecked(first)); 1945 } else 1946 n->in_RegMask(k).dump(); 1947 } 1948 tty->print("/N%d ",m->_idx); 1949 } 1950 } 1951 if( k < n->len() && n->in(k) ) tty->print("| "); 1952 for( ; k < n->len(); k++ ) { 1953 Node *m = n->in(k); 1954 if(!m) { 1955 break; 1956 } 1957 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1958 tty->print("L%d",r); 1959 tty->print("/N%d ",m->_idx); 1960 } 1961 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 1962 else n->dump_spec(tty); 1963 if( _spilled_once.test(n->_idx ) ) { 1964 tty->print(" Spill_1"); 1965 if( _spilled_twice.test(n->_idx ) ) 1966 tty->print(" Spill_2"); 1967 } 1968 tty->print("\n"); 1969 } 1970 1971 void PhaseChaitin::dump(const Block* b) const { 1972 b->dump_head(&_cfg); 1973 1974 // For all instructions 1975 for( uint j = 0; j < b->number_of_nodes(); j++ ) 1976 dump(b->get_node(j)); 1977 // Print live-out info at end of block 1978 if( _live ) { 1979 tty->print("Liveout: "); 1980 IndexSet *live = _live->live(b); 1981 IndexSetIterator elements(live); 1982 tty->print("{"); 1983 uint i; 1984 while ((i = elements.next()) != 0) { 1985 tty->print("L%d ", _lrg_map.find_const(i)); 1986 } 1987 tty->print_cr("}"); 1988 } 1989 tty->print("\n"); 1990 } 1991 1992 void PhaseChaitin::dump() const { 1993 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 1994 _matcher._new_SP, _framesize ); 1995 1996 // For all blocks 1997 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1998 dump(_cfg.get_block(i)); 1999 } 2000 // End of per-block dump 2001 tty->print("\n"); 2002 2003 if (!_ifg) { 2004 tty->print("(No IFG.)\n"); 2005 return; 2006 } 2007 2008 // Dump LRG array 2009 tty->print("--- Live RanGe Array ---\n"); 2010 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { 2011 tty->print("L%d: ",i2); 2012 if (i2 < _ifg->_maxlrg) { 2013 lrgs(i2).dump(); 2014 } 2015 else { 2016 tty->print_cr("new LRG"); 2017 } 2018 } 2019 tty->cr(); 2020 2021 // Dump lo-degree list 2022 tty->print("Lo degree: "); 2023 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 2024 tty->print("L%d ",i3); 2025 tty->cr(); 2026 2027 // Dump lo-stk-degree list 2028 tty->print("Lo stk degree: "); 2029 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 2030 tty->print("L%d ",i4); 2031 tty->cr(); 2032 2033 // Dump lo-degree list 2034 tty->print("Hi degree: "); 2035 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 2036 tty->print("L%d ",i5); 2037 tty->cr(); 2038 } 2039 2040 void PhaseChaitin::dump_degree_lists() const { 2041 // Dump lo-degree list 2042 tty->print("Lo degree: "); 2043 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 2044 tty->print("L%d ",i); 2045 tty->cr(); 2046 2047 // Dump lo-stk-degree list 2048 tty->print("Lo stk degree: "); 2049 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 2050 tty->print("L%d ",i2); 2051 tty->cr(); 2052 2053 // Dump lo-degree list 2054 tty->print("Hi degree: "); 2055 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 2056 tty->print("L%d ",i3); 2057 tty->cr(); 2058 } 2059 2060 void PhaseChaitin::dump_simplified() const { 2061 tty->print("Simplified: "); 2062 for( uint i = _simplified; i; i = lrgs(i)._next ) 2063 tty->print("L%d ",i); 2064 tty->cr(); 2065 } 2066 2067 static char *print_reg(OptoReg::Name reg, const PhaseChaitin* pc, char* buf) { 2068 if ((int)reg < 0) 2069 sprintf(buf, "<OptoReg::%d>", (int)reg); 2070 else if (OptoReg::is_reg(reg)) 2071 strcpy(buf, Matcher::regName[reg]); 2072 else 2073 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 2074 pc->reg2offset(reg)); 2075 return buf+strlen(buf); 2076 } 2077 2078 // Dump a register name into a buffer. Be intelligent if we get called 2079 // before allocation is complete. 2080 char *PhaseChaitin::dump_register(const Node* n, char* buf) const { 2081 if( _node_regs ) { 2082 // Post allocation, use direct mappings, no LRG info available 2083 print_reg( get_reg_first(n), this, buf ); 2084 } else { 2085 uint lidx = _lrg_map.find_const(n); // Grab LRG number 2086 if( !_ifg ) { 2087 sprintf(buf,"L%d",lidx); // No register binding yet 2088 } else if( !lidx ) { // Special, not allocated value 2089 strcpy(buf,"Special"); 2090 } else { 2091 if (lrgs(lidx)._is_vector) { 2092 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) 2093 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register 2094 else 2095 sprintf(buf,"L%d",lidx); // No register binding yet 2096 } else if( (lrgs(lidx).num_regs() == 1) 2097 ? lrgs(lidx).mask().is_bound1() 2098 : lrgs(lidx).mask().is_bound_pair() ) { 2099 // Hah! We have a bound machine register 2100 print_reg( lrgs(lidx).reg(), this, buf ); 2101 } else { 2102 sprintf(buf,"L%d",lidx); // No register binding yet 2103 } 2104 } 2105 } 2106 return buf+strlen(buf); 2107 } 2108 2109 void PhaseChaitin::dump_for_spill_split_recycle() const { 2110 if( WizardMode && (PrintCompilation || PrintOpto) ) { 2111 // Display which live ranges need to be split and the allocator's state 2112 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 2113 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { 2114 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 2115 tty->print("L%d: ", bidx); 2116 lrgs(bidx).dump(); 2117 } 2118 } 2119 tty->cr(); 2120 dump(); 2121 } 2122 } 2123 2124 void PhaseChaitin::dump_frame() const { 2125 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 2126 const TypeTuple *domain = C->tf()->domain(); 2127 const int argcnt = domain->cnt() - TypeFunc::Parms; 2128 2129 // Incoming arguments in registers dump 2130 for( int k = 0; k < argcnt; k++ ) { 2131 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 2132 if( OptoReg::is_reg(parmreg)) { 2133 const char *reg_name = OptoReg::regname(parmreg); 2134 tty->print("#r%3.3d %s", parmreg, reg_name); 2135 parmreg = _matcher._parm_regs[k].second(); 2136 if( OptoReg::is_reg(parmreg)) { 2137 tty->print(":%s", OptoReg::regname(parmreg)); 2138 } 2139 tty->print(" : parm %d: ", k); 2140 domain->field_at(k + TypeFunc::Parms)->dump(); 2141 tty->cr(); 2142 } 2143 } 2144 2145 // Check for un-owned padding above incoming args 2146 OptoReg::Name reg = _matcher._new_SP; 2147 if( reg > _matcher._in_arg_limit ) { 2148 reg = OptoReg::add(reg, -1); 2149 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 2150 } 2151 2152 // Incoming argument area dump 2153 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 2154 while( reg > begin_in_arg ) { 2155 reg = OptoReg::add(reg, -1); 2156 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2157 int j; 2158 for( j = 0; j < argcnt; j++) { 2159 if( _matcher._parm_regs[j].first() == reg || 2160 _matcher._parm_regs[j].second() == reg ) { 2161 tty->print("parm %d: ",j); 2162 domain->field_at(j + TypeFunc::Parms)->dump(); 2163 tty->cr(); 2164 break; 2165 } 2166 } 2167 if( j >= argcnt ) 2168 tty->print_cr("HOLE, owned by SELF"); 2169 } 2170 2171 // Old outgoing preserve area 2172 while( reg > _matcher._old_SP ) { 2173 reg = OptoReg::add(reg, -1); 2174 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 2175 } 2176 2177 // Old SP 2178 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 2179 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 2180 2181 // Preserve area dump 2182 int fixed_slots = C->fixed_slots(); 2183 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); 2184 OptoReg::Name return_addr = _matcher.return_addr(); 2185 2186 reg = OptoReg::add(reg, -1); 2187 while (OptoReg::is_stack(reg)) { 2188 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2189 if (return_addr == reg) { 2190 tty->print_cr("return address"); 2191 } else if (reg >= begin_in_preserve) { 2192 // Preserved slots are present on x86 2193 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) 2194 tty->print_cr("saved fp register"); 2195 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && 2196 VerifyStackAtCalls) 2197 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 2198 else 2199 tty->print_cr("in_preserve"); 2200 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { 2201 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 2202 } else { 2203 tty->print_cr("pad2, stack alignment"); 2204 } 2205 reg = OptoReg::add(reg, -1); 2206 } 2207 2208 // Spill area dump 2209 reg = OptoReg::add(_matcher._new_SP, _framesize ); 2210 while( reg > _matcher._out_arg_limit ) { 2211 reg = OptoReg::add(reg, -1); 2212 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 2213 } 2214 2215 // Outgoing argument area dump 2216 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 2217 reg = OptoReg::add(reg, -1); 2218 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 2219 } 2220 2221 // Outgoing new preserve area 2222 while( reg > _matcher._new_SP ) { 2223 reg = OptoReg::add(reg, -1); 2224 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 2225 } 2226 tty->print_cr("#"); 2227 } 2228 2229 void PhaseChaitin::dump_bb(uint pre_order) const { 2230 tty->print_cr("---dump of B%d---",pre_order); 2231 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2232 Block* block = _cfg.get_block(i); 2233 if (block->_pre_order == pre_order) { 2234 dump(block); 2235 } 2236 } 2237 } 2238 2239 void PhaseChaitin::dump_lrg(uint lidx, bool defs_only) const { 2240 tty->print_cr("---dump of L%d---",lidx); 2241 2242 if (_ifg) { 2243 if (lidx >= _lrg_map.max_lrg_id()) { 2244 tty->print("Attempt to print live range index beyond max live range.\n"); 2245 return; 2246 } 2247 tty->print("L%d: ",lidx); 2248 if (lidx < _ifg->_maxlrg) { 2249 lrgs(lidx).dump(); 2250 } else { 2251 tty->print_cr("new LRG"); 2252 } 2253 } 2254 if( _ifg && lidx < _ifg->_maxlrg) { 2255 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 2256 _ifg->neighbors(lidx)->dump(); 2257 tty->cr(); 2258 } 2259 // For all blocks 2260 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2261 Block* block = _cfg.get_block(i); 2262 int dump_once = 0; 2263 2264 // For all instructions 2265 for( uint j = 0; j < block->number_of_nodes(); j++ ) { 2266 Node *n = block->get_node(j); 2267 if (_lrg_map.find_const(n) == lidx) { 2268 if (!dump_once++) { 2269 tty->cr(); 2270 block->dump_head(&_cfg); 2271 } 2272 dump(n); 2273 continue; 2274 } 2275 if (!defs_only) { 2276 uint cnt = n->req(); 2277 for( uint k = 1; k < cnt; k++ ) { 2278 Node *m = n->in(k); 2279 if (!m) { 2280 continue; // be robust in the dumper 2281 } 2282 if (_lrg_map.find_const(m) == lidx) { 2283 if (!dump_once++) { 2284 tty->cr(); 2285 block->dump_head(&_cfg); 2286 } 2287 dump(n); 2288 } 2289 } 2290 } 2291 } 2292 } // End of per-block dump 2293 tty->cr(); 2294 } 2295 #endif // not PRODUCT 2296 2297 #ifdef ASSERT 2298 // Verify that base pointers and derived pointers are still sane. 2299 void PhaseChaitin::verify_base_ptrs(ResourceArea* a) const { 2300 Unique_Node_List worklist(a); 2301 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2302 Block* block = _cfg.get_block(i); 2303 for (uint j = block->end_idx() + 1; j > 1; j--) { 2304 Node* n = block->get_node(j-1); 2305 if (n->is_Phi()) { 2306 break; 2307 } 2308 // Found a safepoint? 2309 if (n->is_MachSafePoint()) { 2310 MachSafePointNode* sfpt = n->as_MachSafePoint(); 2311 JVMState* jvms = sfpt->jvms(); 2312 if (jvms != NULL) { 2313 // Now scan for a live derived pointer 2314 if (jvms->oopoff() < sfpt->req()) { 2315 // Check each derived/base pair 2316 for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx++) { 2317 Node* check = sfpt->in(idx); 2318 bool is_derived = ((idx - jvms->oopoff()) & 1) == 0; 2319 // search upwards through spills and spill phis for AddP 2320 worklist.clear(); 2321 worklist.push(check); 2322 uint k = 0; 2323 while (k < worklist.size()) { 2324 check = worklist.at(k); 2325 assert(check, "Bad base or derived pointer"); 2326 // See PhaseChaitin::find_base_for_derived() for all cases. 2327 int isc = check->is_Copy(); 2328 if (isc) { 2329 worklist.push(check->in(isc)); 2330 } else if (check->is_Phi()) { 2331 for (uint m = 1; m < check->req(); m++) { 2332 worklist.push(check->in(m)); 2333 } 2334 } else if (check->is_Con()) { 2335 if (is_derived) { 2336 // Derived is NULL+offset 2337 assert(!is_derived || check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad derived pointer"); 2338 } else { 2339 assert(check->bottom_type()->is_ptr()->_offset == 0, "Bad base pointer"); 2340 // Base either ConP(NULL) or loadConP 2341 if (check->is_Mach()) { 2342 assert(check->as_Mach()->ideal_Opcode() == Op_ConP, "Bad base pointer"); 2343 } else { 2344 assert(check->Opcode() == Op_ConP && 2345 check->bottom_type()->is_ptr()->ptr() == TypePtr::Null, "Bad base pointer"); 2346 } 2347 } 2348 } else if (check->bottom_type()->is_ptr()->_offset == 0) { 2349 if (check->is_Proj() || (check->is_Mach() && 2350 (check->as_Mach()->ideal_Opcode() == Op_CreateEx || 2351 check->as_Mach()->ideal_Opcode() == Op_ThreadLocal || 2352 check->as_Mach()->ideal_Opcode() == Op_CMoveP || 2353 check->as_Mach()->ideal_Opcode() == Op_CheckCastPP || 2354 #ifdef _LP64 2355 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_CastPP) || 2356 (UseCompressedOops && check->as_Mach()->ideal_Opcode() == Op_DecodeN) || 2357 (UseCompressedClassPointers && check->as_Mach()->ideal_Opcode() == Op_DecodeNKlass) || 2358 #endif // _LP64 2359 check->as_Mach()->ideal_Opcode() == Op_LoadP || 2360 check->as_Mach()->ideal_Opcode() == Op_LoadKlass))) { 2361 // Valid nodes 2362 } else { 2363 check->dump(); 2364 assert(false, "Bad base or derived pointer"); 2365 } 2366 } else { 2367 assert(is_derived, "Bad base pointer"); 2368 assert(check->is_Mach() && check->as_Mach()->ideal_Opcode() == Op_AddP, "Bad derived pointer"); 2369 } 2370 k++; 2371 assert(k < 100000, "Derived pointer checking in infinite loop"); 2372 } // End while 2373 } 2374 } // End of check for derived pointers 2375 } // End of Kcheck for debug info 2376 } // End of if found a safepoint 2377 } // End of forall instructions in block 2378 } // End of forall blocks 2379 } 2380 2381 // Verify that graphs and base pointers are still sane. 2382 void PhaseChaitin::verify(ResourceArea* a, bool verify_ifg) const { 2383 if (VerifyRegisterAllocator) { 2384 _cfg.verify(); 2385 verify_base_ptrs(a); 2386 if (verify_ifg) { 2387 _ifg->verify(this); 2388 } 2389 } 2390 } 2391 #endif // ASSERT 2392 2393 int PhaseChaitin::_final_loads = 0; 2394 int PhaseChaitin::_final_stores = 0; 2395 int PhaseChaitin::_final_memoves= 0; 2396 int PhaseChaitin::_final_copies = 0; 2397 double PhaseChaitin::_final_load_cost = 0; 2398 double PhaseChaitin::_final_store_cost = 0; 2399 double PhaseChaitin::_final_memove_cost= 0; 2400 double PhaseChaitin::_final_copy_cost = 0; 2401 int PhaseChaitin::_conserv_coalesce = 0; 2402 int PhaseChaitin::_conserv_coalesce_pair = 0; 2403 int PhaseChaitin::_conserv_coalesce_trie = 0; 2404 int PhaseChaitin::_conserv_coalesce_quad = 0; 2405 int PhaseChaitin::_post_alloc = 0; 2406 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2407 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2408 int PhaseChaitin::_used_cisc_instructions = 0; 2409 int PhaseChaitin::_unused_cisc_instructions = 0; 2410 int PhaseChaitin::_allocator_attempts = 0; 2411 int PhaseChaitin::_allocator_successes = 0; 2412 2413 #ifndef PRODUCT 2414 uint PhaseChaitin::_high_pressure = 0; 2415 uint PhaseChaitin::_low_pressure = 0; 2416 2417 void PhaseChaitin::print_chaitin_statistics() { 2418 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); 2419 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); 2420 tty->print_cr("Adjusted spill cost = %7.0f.", 2421 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2422 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2423 tty->print("Conservatively coalesced %d copies, %d pairs", 2424 _conserv_coalesce, _conserv_coalesce_pair); 2425 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2426 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2427 tty->print_cr(", %d post alloc.", _post_alloc); 2428 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2429 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2430 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2431 if( _used_cisc_instructions || _unused_cisc_instructions ) 2432 tty->print_cr("Used cisc instruction %d, remained in register %d", 2433 _used_cisc_instructions, _unused_cisc_instructions); 2434 if( _allocator_successes != 0 ) 2435 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2436 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2437 } 2438 #endif // not PRODUCT