1 /* 2 * Copyright (c) 2000, 2017, 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 , _lrg_map(Thread::current()->resource_area(), unique) 205 , _live(0) 206 , _spilled_once(Thread::current()->resource_area()) 207 , _spilled_twice(Thread::current()->resource_area()) 208 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0) 209 , _oldphi(unique) 210 , _scheduling_info_generated(scheduling_info_generated) 211 , _sched_int_pressure(0, INTPRESSURE) 212 , _sched_float_pressure(0, FLOATPRESSURE) 213 , _scratch_int_pressure(0, INTPRESSURE) 214 , _scratch_float_pressure(0, FLOATPRESSURE) 215 #ifndef PRODUCT 216 , _trace_spilling(C->directive()->TraceSpillingOption) 217 #endif 218 { 219 Compile::TracePhase tp("ctorChaitin", &timers[_t_ctorChaitin]); 220 221 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency()); 222 223 // Build a list of basic blocks, sorted by frequency 224 _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 225 // Experiment with sorting strategies to speed compilation 226 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket 227 Block **buckets[NUMBUCKS]; // Array of buckets 228 uint buckcnt[NUMBUCKS]; // Array of bucket counters 229 double buckval[NUMBUCKS]; // Array of bucket value cutoffs 230 for (uint i = 0; i < NUMBUCKS; i++) { 231 buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 232 buckcnt[i] = 0; 233 // Bump by three orders of magnitude each time 234 cutoff *= 0.001; 235 buckval[i] = cutoff; 236 for (uint j = 0; j < _cfg.number_of_blocks(); j++) { 237 buckets[i][j] = NULL; 238 } 239 } 240 // Sort blocks into buckets 241 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 242 for (uint j = 0; j < NUMBUCKS; j++) { 243 if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) { 244 // Assign block to end of list for appropriate bucket 245 buckets[j][buckcnt[j]++] = _cfg.get_block(i); 246 break; // kick out of inner loop 247 } 248 } 249 } 250 // Dump buckets into final block array 251 uint blkcnt = 0; 252 for (uint i = 0; i < NUMBUCKS; i++) { 253 for (uint j = 0; j < buckcnt[i]; j++) { 254 _blks[blkcnt++] = buckets[i][j]; 255 } 256 } 257 258 assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled"); 259 } 260 261 // union 2 sets together. 262 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) { 263 uint src = _lrg_map.find(src_n); 264 uint dst = _lrg_map.find(dst_n); 265 assert(src, ""); 266 assert(dst, ""); 267 assert(src < _lrg_map.max_lrg_id(), "oob"); 268 assert(dst < _lrg_map.max_lrg_id(), "oob"); 269 assert(src < dst, "always union smaller"); 270 _lrg_map.uf_map(dst, src); 271 } 272 273 void PhaseChaitin::new_lrg(const Node *x, uint lrg) { 274 // Make the Node->LRG mapping 275 _lrg_map.extend(x->_idx,lrg); 276 // Make the Union-Find mapping an identity function 277 _lrg_map.uf_extend(lrg, lrg); 278 } 279 280 281 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) { 282 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections"); 283 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); ) 284 int found_projs = 0; 285 uint cnt = orig->outcnt(); 286 for (uint i = 0; i < cnt; i++) { 287 Node* proj = orig->raw_out(i); 288 if (proj->is_MachProj()) { 289 assert(proj->outcnt() == 0, "only kill projections are expected here"); 290 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections"); 291 found_projs++; 292 // Copy kill projections after the cloned node 293 Node* kills = proj->clone(); 294 kills->set_req(0, copy); 295 b->insert_node(kills, idx++); 296 _cfg.map_node_to_block(kills, b); 297 new_lrg(kills, max_lrg_id++); 298 } 299 } 300 return found_projs; 301 } 302 303 // Renumber the live ranges to compact them. Makes the IFG smaller. 304 void PhaseChaitin::compact() { 305 Compile::TracePhase tp("chaitinCompact", &timers[_t_chaitinCompact]); 306 307 // Current the _uf_map contains a series of short chains which are headed 308 // by a self-cycle. All the chains run from big numbers to little numbers. 309 // The Find() call chases the chains & shortens them for the next Find call. 310 // We are going to change this structure slightly. Numbers above a moving 311 // wave 'i' are unchanged. Numbers below 'j' point directly to their 312 // compacted live range with no further chaining. There are no chains or 313 // cycles below 'i', so the Find call no longer works. 314 uint j=1; 315 uint i; 316 for (i = 1; i < _lrg_map.max_lrg_id(); i++) { 317 uint lr = _lrg_map.uf_live_range_id(i); 318 // Ignore unallocated live ranges 319 if (!lr) { 320 continue; 321 } 322 assert(lr <= i, ""); 323 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr)); 324 } 325 // Now change the Node->LR mapping to reflect the compacted names 326 uint unique = _lrg_map.size(); 327 for (i = 0; i < unique; i++) { 328 uint lrg_id = _lrg_map.live_range_id(i); 329 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id)); 330 } 331 332 // Reset the Union-Find mapping 333 _lrg_map.reset_uf_map(j); 334 } 335 336 void PhaseChaitin::Register_Allocate() { 337 338 // Above the OLD FP (and in registers) are the incoming arguments. Stack 339 // slots in this area are called "arg_slots". Above the NEW FP (and in 340 // registers) is the outgoing argument area; above that is the spill/temp 341 // area. These are all "frame_slots". Arg_slots start at the zero 342 // stack_slots and count up to the known arg_size. Frame_slots start at 343 // the stack_slot #arg_size and go up. After allocation I map stack 344 // slots to actual offsets. Stack-slots in the arg_slot area are biased 345 // by the frame_size; stack-slots in the frame_slot area are biased by 0. 346 347 _trip_cnt = 0; 348 _alternate = 0; 349 _matcher._allocation_started = true; 350 351 ResourceArea split_arena(mtCompiler); // Arena for Split local resources 352 ResourceArea live_arena(mtCompiler); // Arena for liveness & IFG info 353 ResourceMark rm(&live_arena); 354 355 // Need live-ness for the IFG; need the IFG for coalescing. If the 356 // liveness is JUST for coalescing, then I can get some mileage by renaming 357 // all copy-related live ranges low and then using the max copy-related 358 // live range as a cut-off for LIVE and the IFG. In other words, I can 359 // build a subset of LIVE and IFG just for copies. 360 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false); 361 362 // Need IFG for coalescing and coloring 363 PhaseIFG ifg(&live_arena); 364 _ifg = &ifg; 365 366 // Come out of SSA world to the Named world. Assign (virtual) registers to 367 // Nodes. Use the same register for all inputs and the output of PhiNodes 368 // - effectively ending SSA form. This requires either coalescing live 369 // ranges or inserting copies. For the moment, we insert "virtual copies" 370 // - we pretend there is a copy prior to each Phi in predecessor blocks. 371 // We will attempt to coalesce such "virtual copies" before we manifest 372 // them for real. 373 de_ssa(); 374 375 #ifdef ASSERT 376 // Veify the graph before RA. 377 verify(&live_arena); 378 #endif 379 380 { 381 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 382 _live = NULL; // Mark live as being not available 383 rm.reset_to_mark(); // Reclaim working storage 384 IndexSet::reset_memory(C, &live_arena); 385 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG 386 gather_lrg_masks( false ); // Collect LRG masks 387 live.compute(_lrg_map.max_lrg_id()); // Compute liveness 388 _live = &live; // Mark LIVE as being available 389 } 390 391 // Base pointers are currently "used" by instructions which define new 392 // derived pointers. This makes base pointers live up to the where the 393 // derived pointer is made, but not beyond. Really, they need to be live 394 // across any GC point where the derived value is live. So this code looks 395 // at all the GC points, and "stretches" the live range of any base pointer 396 // to the GC point. 397 if (stretch_base_pointer_live_ranges(&live_arena)) { 398 Compile::TracePhase tp("computeLive (sbplr)", &timers[_t_computeLive]); 399 // Since some live range stretched, I need to recompute live 400 _live = NULL; 401 rm.reset_to_mark(); // Reclaim working storage 402 IndexSet::reset_memory(C, &live_arena); 403 ifg.init(_lrg_map.max_lrg_id()); 404 gather_lrg_masks(false); 405 live.compute(_lrg_map.max_lrg_id()); 406 _live = &live; 407 } 408 // Create the interference graph using virtual copies 409 build_ifg_virtual(); // Include stack slots this time 410 411 // The IFG is/was triangular. I am 'squaring it up' so Union can run 412 // faster. Union requires a 'for all' operation which is slow on the 413 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - 414 // meaning I can visit all the Nodes neighbors less than a Node in time 415 // O(# of neighbors), but I have to visit all the Nodes greater than a 416 // given Node and search them for an instance, i.e., time O(#MaxLRG)). 417 _ifg->SquareUp(); 418 419 // Aggressive (but pessimistic) copy coalescing. 420 // This pass works on virtual copies. Any virtual copies which are not 421 // coalesced get manifested as actual copies 422 { 423 Compile::TracePhase tp("chaitinCoalesce1", &timers[_t_chaitinCoalesce1]); 424 425 PhaseAggressiveCoalesce coalesce(*this); 426 coalesce.coalesce_driver(); 427 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do 428 // not match the Phi itself, insert a copy. 429 coalesce.insert_copies(_matcher); 430 if (C->failing()) { 431 return; 432 } 433 } 434 435 // After aggressive coalesce, attempt a first cut at coloring. 436 // To color, we need the IFG and for that we need LIVE. 437 { 438 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 439 _live = NULL; 440 rm.reset_to_mark(); // Reclaim working storage 441 IndexSet::reset_memory(C, &live_arena); 442 ifg.init(_lrg_map.max_lrg_id()); 443 gather_lrg_masks( true ); 444 live.compute(_lrg_map.max_lrg_id()); 445 _live = &live; 446 } 447 448 // Build physical interference graph 449 uint must_spill = 0; 450 must_spill = build_ifg_physical(&live_arena); 451 // If we have a guaranteed spill, might as well spill now 452 if (must_spill) { 453 if(!_lrg_map.max_lrg_id()) { 454 return; 455 } 456 // Bail out if unique gets too large (ie - unique > MaxNodeLimit) 457 C->check_node_count(10*must_spill, "out of nodes before split"); 458 if (C->failing()) { 459 return; 460 } 461 462 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 463 _lrg_map.set_max_lrg_id(new_max_lrg_id); 464 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 465 // or we failed to split 466 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); 467 if (C->failing()) { 468 return; 469 } 470 471 NOT_PRODUCT(C->verify_graph_edges();) 472 473 compact(); // Compact LRGs; return new lower max lrg 474 475 { 476 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 477 _live = NULL; 478 rm.reset_to_mark(); // Reclaim working storage 479 IndexSet::reset_memory(C, &live_arena); 480 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph 481 gather_lrg_masks( true ); // Collect intersect mask 482 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE 483 _live = &live; 484 } 485 build_ifg_physical(&live_arena); 486 _ifg->SquareUp(); 487 _ifg->Compute_Effective_Degree(); 488 // Only do conservative coalescing if requested 489 if (OptoCoalesce) { 490 Compile::TracePhase tp("chaitinCoalesce2", &timers[_t_chaitinCoalesce2]); 491 // Conservative (and pessimistic) copy coalescing of those spills 492 PhaseConservativeCoalesce coalesce(*this); 493 // If max live ranges greater than cutoff, don't color the stack. 494 // This cutoff can be larger than below since it is only done once. 495 coalesce.coalesce_driver(); 496 } 497 _lrg_map.compress_uf_map_for_nodes(); 498 499 #ifdef ASSERT 500 verify(&live_arena, true); 501 #endif 502 } else { 503 ifg.SquareUp(); 504 ifg.Compute_Effective_Degree(); 505 #ifdef ASSERT 506 set_was_low(); 507 #endif 508 } 509 510 // Prepare for Simplify & Select 511 cache_lrg_info(); // Count degree of LRGs 512 513 // Simplify the InterFerence Graph by removing LRGs of low degree. 514 // LRGs of low degree are trivially colorable. 515 Simplify(); 516 517 // Select colors by re-inserting LRGs back into the IFG in reverse order. 518 // Return whether or not something spills. 519 uint spills = Select( ); 520 521 // If we spill, split and recycle the entire thing 522 while( spills ) { 523 if( _trip_cnt++ > 24 ) { 524 DEBUG_ONLY( dump_for_spill_split_recycle(); ) 525 if( _trip_cnt > 27 ) { 526 C->record_method_not_compilable("failed spill-split-recycle sanity check"); 527 return; 528 } 529 } 530 531 if (!_lrg_map.max_lrg_id()) { 532 return; 533 } 534 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 535 _lrg_map.set_max_lrg_id(new_max_lrg_id); 536 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 537 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split"); 538 if (C->failing()) { 539 return; 540 } 541 542 compact(); // Compact LRGs; return new lower max lrg 543 544 // Nuke the live-ness and interference graph and LiveRanGe info 545 { 546 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 547 _live = NULL; 548 rm.reset_to_mark(); // Reclaim working storage 549 IndexSet::reset_memory(C, &live_arena); 550 ifg.init(_lrg_map.max_lrg_id()); 551 552 // Create LiveRanGe array. 553 // Intersect register masks for all USEs and DEFs 554 gather_lrg_masks(true); 555 live.compute(_lrg_map.max_lrg_id()); 556 _live = &live; 557 } 558 must_spill = build_ifg_physical(&live_arena); 559 _ifg->SquareUp(); 560 _ifg->Compute_Effective_Degree(); 561 562 // Only do conservative coalescing if requested 563 if (OptoCoalesce) { 564 Compile::TracePhase tp("chaitinCoalesce3", &timers[_t_chaitinCoalesce3]); 565 // Conservative (and pessimistic) copy coalescing 566 PhaseConservativeCoalesce coalesce(*this); 567 // Check for few live ranges determines how aggressive coalesce is. 568 coalesce.coalesce_driver(); 569 } 570 _lrg_map.compress_uf_map_for_nodes(); 571 #ifdef ASSERT 572 verify(&live_arena, true); 573 #endif 574 cache_lrg_info(); // Count degree of LRGs 575 576 // Simplify the InterFerence Graph by removing LRGs of low degree. 577 // LRGs of low degree are trivially colorable. 578 Simplify(); 579 580 // Select colors by re-inserting LRGs back into the IFG in reverse order. 581 // Return whether or not something spills. 582 spills = Select(); 583 } 584 585 // Count number of Simplify-Select trips per coloring success. 586 _allocator_attempts += _trip_cnt + 1; 587 _allocator_successes += 1; 588 589 // Peephole remove copies 590 post_allocate_copy_removal(); 591 592 // Merge multidefs if multiple defs representing the same value are used in a single block. 593 merge_multidefs(); 594 595 #ifdef ASSERT 596 // Veify the graph after RA. 597 verify(&live_arena); 598 #endif 599 600 // max_reg is past the largest *register* used. 601 // Convert that to a frame_slot number. 602 if (_max_reg <= _matcher._new_SP) { 603 _framesize = C->out_preserve_stack_slots(); 604 } 605 else { 606 _framesize = _max_reg -_matcher._new_SP; 607 } 608 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); 609 610 // This frame must preserve the required fp alignment 611 _framesize = align_up(_framesize, Matcher::stack_alignment_in_slots()); 612 assert(_framesize <= 1000000, "sanity check"); 613 #ifndef PRODUCT 614 _total_framesize += _framesize; 615 if ((int)_framesize > _max_framesize) { 616 _max_framesize = _framesize; 617 } 618 #endif 619 620 // Convert CISC spills 621 fixup_spills(); 622 623 // Log regalloc results 624 CompileLog* log = Compile::current()->log(); 625 if (log != NULL) { 626 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); 627 } 628 629 if (C->failing()) { 630 return; 631 } 632 633 NOT_PRODUCT(C->verify_graph_edges();) 634 635 // Move important info out of the live_arena to longer lasting storage. 636 alloc_node_regs(_lrg_map.size()); 637 for (uint i=0; i < _lrg_map.size(); i++) { 638 if (_lrg_map.live_range_id(i)) { // Live range associated with Node? 639 LRG &lrg = lrgs(_lrg_map.live_range_id(i)); 640 if (!lrg.alive()) { 641 set_bad(i); 642 } else if (lrg.num_regs() == 1) { 643 set1(i, lrg.reg()); 644 } else { // Must be a register-set 645 if (!lrg._fat_proj) { // Must be aligned adjacent register set 646 // Live ranges record the highest register in their mask. 647 // We want the low register for the AD file writer's convenience. 648 OptoReg::Name hi = lrg.reg(); // Get hi register 649 OptoReg::Name lo = OptoReg::add(hi, (1-lrg.num_regs())); // Find lo 650 // We have to use pair [lo,lo+1] even for wide vectors because 651 // the rest of code generation works only with pairs. It is safe 652 // since for registers encoding only 'lo' is used. 653 // Second reg from pair is used in ScheduleAndBundle on SPARC where 654 // vector max size is 8 which corresponds to registers pair. 655 // It is also used in BuildOopMaps but oop operations are not 656 // vectorized. 657 set2(i, lo); 658 } else { // Misaligned; extract 2 bits 659 OptoReg::Name hi = lrg.reg(); // Get hi register 660 lrg.Remove(hi); // Yank from mask 661 int lo = lrg.mask().find_first_elem(); // Find lo 662 set_pair(i, hi, lo); 663 } 664 } 665 if( lrg._is_oop ) _node_oops.set(i); 666 } else { 667 set_bad(i); 668 } 669 } 670 671 // Done! 672 _live = NULL; 673 _ifg = NULL; 674 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope 675 } 676 677 void PhaseChaitin::de_ssa() { 678 // Set initial Names for all Nodes. Most Nodes get the virtual register 679 // number. A few get the ZERO live range number. These do not 680 // get allocated, but instead rely on correct scheduling to ensure that 681 // only one instance is simultaneously live at a time. 682 uint lr_counter = 1; 683 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 684 Block* block = _cfg.get_block(i); 685 uint cnt = block->number_of_nodes(); 686 687 // Handle all the normal Nodes in the block 688 for( uint j = 0; j < cnt; j++ ) { 689 Node *n = block->get_node(j); 690 // Pre-color to the zero live range, or pick virtual register 691 const RegMask &rm = n->out_RegMask(); 692 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0); 693 } 694 } 695 696 // Reset the Union-Find mapping to be identity 697 _lrg_map.reset_uf_map(lr_counter); 698 } 699 700 void PhaseChaitin::mark_ssa() { 701 // Use ssa names to populate the live range maps or if no mask 702 // is available, use the 0 entry. 703 uint max_idx = 0; 704 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 705 Block* block = _cfg.get_block(i); 706 uint cnt = block->number_of_nodes(); 707 708 // Handle all the normal Nodes in the block 709 for ( uint j = 0; j < cnt; j++ ) { 710 Node *n = block->get_node(j); 711 // Pre-color to the zero live range, or pick virtual register 712 const RegMask &rm = n->out_RegMask(); 713 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0); 714 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx; 715 } 716 } 717 _lrg_map.set_max_lrg_id(max_idx+1); 718 719 // Reset the Union-Find mapping to be identity 720 _lrg_map.reset_uf_map(max_idx+1); 721 } 722 723 724 // Gather LiveRanGe information, including register masks. Modification of 725 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. 726 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { 727 728 // Nail down the frame pointer live range 729 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr)); 730 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite 731 732 // For all blocks 733 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 734 Block* block = _cfg.get_block(i); 735 736 // For all instructions 737 for (uint j = 1; j < block->number_of_nodes(); j++) { 738 Node* n = block->get_node(j); 739 uint input_edge_start =1; // Skip control most nodes 740 bool is_machine_node = false; 741 if (n->is_Mach()) { 742 is_machine_node = true; 743 input_edge_start = n->as_Mach()->oper_input_base(); 744 } 745 uint idx = n->is_Copy(); 746 747 // Get virtual register number, same as LiveRanGe index 748 uint vreg = _lrg_map.live_range_id(n); 749 LRG& lrg = lrgs(vreg); 750 if (vreg) { // No vreg means un-allocable (e.g. memory) 751 752 // Collect has-copy bit 753 if (idx) { 754 lrg._has_copy = 1; 755 uint clidx = _lrg_map.live_range_id(n->in(idx)); 756 LRG& copy_src = lrgs(clidx); 757 copy_src._has_copy = 1; 758 } 759 760 // Check for float-vs-int live range (used in register-pressure 761 // calculations) 762 const Type *n_type = n->bottom_type(); 763 if (n_type->is_floatingpoint()) { 764 lrg._is_float = 1; 765 } 766 767 // Check for twice prior spilling. Once prior spilling might have 768 // spilled 'soft', 2nd prior spill should have spilled 'hard' and 769 // further spilling is unlikely to make progress. 770 if (_spilled_once.test(n->_idx)) { 771 lrg._was_spilled1 = 1; 772 if (_spilled_twice.test(n->_idx)) { 773 lrg._was_spilled2 = 1; 774 } 775 } 776 777 #ifndef PRODUCT 778 if (trace_spilling() && lrg._def != NULL) { 779 // collect defs for MultiDef printing 780 if (lrg._defs == NULL) { 781 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL); 782 lrg._defs->append(lrg._def); 783 } 784 lrg._defs->append(n); 785 } 786 #endif 787 788 // Check for a single def LRG; these can spill nicely 789 // via rematerialization. Flag as NULL for no def found 790 // yet, or 'n' for single def or -1 for many defs. 791 lrg._def = lrg._def ? NodeSentinel : n; 792 793 // Limit result register mask to acceptable registers 794 const RegMask &rm = n->out_RegMask(); 795 lrg.AND( rm ); 796 797 uint ireg = n->ideal_reg(); 798 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, 799 "oops must be in Op_RegP's" ); 800 801 // Check for vector live range (only if vector register is used). 802 // On SPARC vector uses RegD which could be misaligned so it is not 803 // processes as vector in RA. 804 if (RegMask::is_vector(ireg)) 805 lrg._is_vector = 1; 806 assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL, 807 "vector must be in vector registers"); 808 809 // Check for bound register masks 810 const RegMask &lrgmask = lrg.mask(); 811 if (lrgmask.is_bound(ireg)) { 812 lrg._is_bound = 1; 813 } 814 815 // Check for maximum frequency value 816 if (lrg._maxfreq < block->_freq) { 817 lrg._maxfreq = block->_freq; 818 } 819 820 // Check for oop-iness, or long/double 821 // Check for multi-kill projection 822 switch (ireg) { 823 case MachProjNode::fat_proj: 824 // Fat projections have size equal to number of registers killed 825 lrg.set_num_regs(rm.Size()); 826 lrg.set_reg_pressure(lrg.num_regs()); 827 lrg._fat_proj = 1; 828 lrg._is_bound = 1; 829 break; 830 case Op_RegP: 831 #ifdef _LP64 832 lrg.set_num_regs(2); // Size is 2 stack words 833 #else 834 lrg.set_num_regs(1); // Size is 1 stack word 835 #endif 836 // Register pressure is tracked relative to the maximum values 837 // suggested for that platform, INTPRESSURE and FLOATPRESSURE, 838 // and relative to other types which compete for the same regs. 839 // 840 // The following table contains suggested values based on the 841 // architectures as defined in each .ad file. 842 // INTPRESSURE and FLOATPRESSURE may be tuned differently for 843 // compile-speed or performance. 844 // Note1: 845 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 846 // since .ad registers are defined as high and low halves. 847 // These reg_pressure values remain compatible with the code 848 // in is_high_pressure() which relates get_invalid_mask_size(), 849 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. 850 // Note2: 851 // SPARC -d32 has 24 registers available for integral values, 852 // but only 10 of these are safe for 64-bit longs. 853 // Using set_reg_pressure(2) for both int and long means 854 // the allocator will believe it can fit 26 longs into 855 // registers. Using 2 for longs and 1 for ints means the 856 // allocator will attempt to put 52 integers into registers. 857 // The settings below limit this problem to methods with 858 // many long values which are being run on 32-bit SPARC. 859 // 860 // ------------------- reg_pressure -------------------- 861 // Each entry is reg_pressure_per_value,number_of_regs 862 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE 863 // IA32 2 1 1 1 1 6 6 864 // IA64 1 1 1 1 1 50 41 865 // SPARC 2 2 2 2 2 48 (24) 52 (26) 866 // SPARCV9 2 2 2 2 2 48 (24) 52 (26) 867 // AMD64 1 1 1 1 1 14 15 868 // ----------------------------------------------------- 869 #if defined(SPARC) 870 lrg.set_reg_pressure(2); // use for v9 as well 871 #else 872 lrg.set_reg_pressure(1); // normally one value per register 873 #endif 874 if( n_type->isa_oop_ptr() ) { 875 lrg._is_oop = 1; 876 } 877 break; 878 case Op_RegL: // Check for long or double 879 case Op_RegD: 880 lrg.set_num_regs(2); 881 // Define platform specific register pressure 882 #if defined(SPARC) || defined(ARM32) 883 lrg.set_reg_pressure(2); 884 #elif defined(IA32) 885 if( ireg == Op_RegL ) { 886 lrg.set_reg_pressure(2); 887 } else { 888 lrg.set_reg_pressure(1); 889 } 890 #else 891 lrg.set_reg_pressure(1); // normally one value per register 892 #endif 893 // If this def of a double forces a mis-aligned double, 894 // flag as '_fat_proj' - really flag as allowing misalignment 895 // AND changes how we count interferences. A mis-aligned 896 // double can interfere with TWO aligned pairs, or effectively 897 // FOUR registers! 898 if (rm.is_misaligned_pair()) { 899 lrg._fat_proj = 1; 900 lrg._is_bound = 1; 901 } 902 break; 903 case Op_RegF: 904 case Op_RegI: 905 case Op_RegN: 906 case Op_RegFlags: 907 case 0: // not an ideal register 908 lrg.set_num_regs(1); 909 #ifdef SPARC 910 lrg.set_reg_pressure(2); 911 #else 912 lrg.set_reg_pressure(1); 913 #endif 914 break; 915 case Op_VecS: 916 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity"); 917 lrg.set_num_regs(RegMask::SlotsPerVecS); 918 lrg.set_reg_pressure(1); 919 break; 920 case Op_VecD: 921 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity"); 922 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned"); 923 lrg.set_num_regs(RegMask::SlotsPerVecD); 924 lrg.set_reg_pressure(1); 925 break; 926 case Op_VecX: 927 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity"); 928 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned"); 929 lrg.set_num_regs(RegMask::SlotsPerVecX); 930 lrg.set_reg_pressure(1); 931 break; 932 case Op_VecY: 933 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity"); 934 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned"); 935 lrg.set_num_regs(RegMask::SlotsPerVecY); 936 lrg.set_reg_pressure(1); 937 break; 938 case Op_VecZ: 939 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity"); 940 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned"); 941 lrg.set_num_regs(RegMask::SlotsPerVecZ); 942 lrg.set_reg_pressure(1); 943 break; 944 default: 945 ShouldNotReachHere(); 946 } 947 } 948 949 // Now do the same for inputs 950 uint cnt = n->req(); 951 // Setup for CISC SPILLING 952 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 953 if( UseCISCSpill && after_aggressive ) { 954 inp = n->cisc_operand(); 955 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 956 // Convert operand number to edge index number 957 inp = n->as_Mach()->operand_index(inp); 958 } 959 960 // Prepare register mask for each input 961 for( uint k = input_edge_start; k < cnt; k++ ) { 962 uint vreg = _lrg_map.live_range_id(n->in(k)); 963 if (!vreg) { 964 continue; 965 } 966 967 // If this instruction is CISC Spillable, add the flags 968 // bit to its appropriate input 969 if( UseCISCSpill && after_aggressive && inp == k ) { 970 #ifndef PRODUCT 971 if( TraceCISCSpill ) { 972 tty->print(" use_cisc_RegMask: "); 973 n->dump(); 974 } 975 #endif 976 n->as_Mach()->use_cisc_RegMask(); 977 } 978 979 if (is_machine_node && _scheduling_info_generated) { 980 MachNode* cur_node = n->as_Mach(); 981 // this is cleaned up by register allocation 982 if (k >= cur_node->num_opnds()) continue; 983 } 984 985 LRG &lrg = lrgs(vreg); 986 // // Testing for floating point code shape 987 // Node *test = n->in(k); 988 // if( test->is_Mach() ) { 989 // MachNode *m = test->as_Mach(); 990 // int op = m->ideal_Opcode(); 991 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 992 // int zzz = 1; 993 // } 994 // } 995 996 // Limit result register mask to acceptable registers. 997 // Do not limit registers from uncommon uses before 998 // AggressiveCoalesce. This effectively pre-virtual-splits 999 // around uncommon uses of common defs. 1000 const RegMask &rm = n->in_RegMask(k); 1001 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { 1002 // Since we are BEFORE aggressive coalesce, leave the register 1003 // mask untrimmed by the call. This encourages more coalescing. 1004 // Later, AFTER aggressive, this live range will have to spill 1005 // but the spiller handles slow-path calls very nicely. 1006 } else { 1007 lrg.AND( rm ); 1008 } 1009 1010 // Check for bound register masks 1011 const RegMask &lrgmask = lrg.mask(); 1012 uint kreg = n->in(k)->ideal_reg(); 1013 bool is_vect = RegMask::is_vector(kreg); 1014 assert(n->in(k)->bottom_type()->isa_vect() == NULL || 1015 is_vect || kreg == Op_RegD || kreg == Op_RegL, 1016 "vector must be in vector registers"); 1017 if (lrgmask.is_bound(kreg)) 1018 lrg._is_bound = 1; 1019 1020 // If this use of a double forces a mis-aligned double, 1021 // flag as '_fat_proj' - really flag as allowing misalignment 1022 // AND changes how we count interferences. A mis-aligned 1023 // double can interfere with TWO aligned pairs, or effectively 1024 // FOUR registers! 1025 #ifdef ASSERT 1026 if (is_vect && !_scheduling_info_generated) { 1027 if (lrg.num_regs() != 0) { 1028 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned"); 1029 assert(!lrg._fat_proj, "sanity"); 1030 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity"); 1031 } else { 1032 assert(n->is_Phi(), "not all inputs processed only if Phi"); 1033 } 1034 } 1035 #endif 1036 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { 1037 lrg._fat_proj = 1; 1038 lrg._is_bound = 1; 1039 } 1040 // if the LRG is an unaligned pair, we will have to spill 1041 // so clear the LRG's register mask if it is not already spilled 1042 if (!is_vect && !n->is_SpillCopy() && 1043 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 1044 lrgmask.is_misaligned_pair()) { 1045 lrg.Clear(); 1046 } 1047 1048 // Check for maximum frequency value 1049 if (lrg._maxfreq < block->_freq) { 1050 lrg._maxfreq = block->_freq; 1051 } 1052 1053 } // End for all allocated inputs 1054 } // end for all instructions 1055 } // end for all blocks 1056 1057 // Final per-liverange setup 1058 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { 1059 LRG &lrg = lrgs(i2); 1060 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); 1061 if (lrg.num_regs() > 1 && !lrg._fat_proj) { 1062 lrg.clear_to_sets(); 1063 } 1064 lrg.compute_set_mask_size(); 1065 if (lrg.not_free()) { // Handle case where we lose from the start 1066 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 1067 lrg._direct_conflict = 1; 1068 } 1069 lrg.set_degree(0); // no neighbors in IFG yet 1070 } 1071 } 1072 1073 // Set the was-lo-degree bit. Conservative coalescing should not change the 1074 // colorability of the graph. If any live range was of low-degree before 1075 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 1076 // The bit is checked in Simplify. 1077 void PhaseChaitin::set_was_low() { 1078 #ifdef ASSERT 1079 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1080 int size = lrgs(i).num_regs(); 1081 uint old_was_lo = lrgs(i)._was_lo; 1082 lrgs(i)._was_lo = 0; 1083 if( lrgs(i).lo_degree() ) { 1084 lrgs(i)._was_lo = 1; // Trivially of low degree 1085 } else { // Else check the Brigg's assertion 1086 // Brigg's observation is that the lo-degree neighbors of a 1087 // hi-degree live range will not interfere with the color choices 1088 // of said hi-degree live range. The Simplify reverse-stack-coloring 1089 // order takes care of the details. Hence you do not have to count 1090 // low-degree neighbors when determining if this guy colors. 1091 int briggs_degree = 0; 1092 IndexSet *s = _ifg->neighbors(i); 1093 IndexSetIterator elements(s); 1094 uint lidx; 1095 while((lidx = elements.next()) != 0) { 1096 if( !lrgs(lidx).lo_degree() ) 1097 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 1098 } 1099 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 1100 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 1101 } 1102 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 1103 } 1104 #endif 1105 } 1106 1107 #define REGISTER_CONSTRAINED 16 1108 1109 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 1110 void PhaseChaitin::cache_lrg_info( ) { 1111 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]); 1112 1113 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1114 LRG &lrg = lrgs(i); 1115 1116 // Check for being of low degree: means we can be trivially colored. 1117 // Low degree, dead or must-spill guys just get to simplify right away 1118 if( lrg.lo_degree() || 1119 !lrg.alive() || 1120 lrg._must_spill ) { 1121 // Split low degree list into those guys that must get a 1122 // register and those that can go to register or stack. 1123 // The idea is LRGs that can go register or stack color first when 1124 // they have a good chance of getting a register. The register-only 1125 // lo-degree live ranges always get a register. 1126 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 1127 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 1128 lrg._next = _lo_stk_degree; 1129 _lo_stk_degree = i; 1130 } else { 1131 lrg._next = _lo_degree; 1132 _lo_degree = i; 1133 } 1134 } else { // Else high degree 1135 lrgs(_hi_degree)._prev = i; 1136 lrg._next = _hi_degree; 1137 lrg._prev = 0; 1138 _hi_degree = i; 1139 } 1140 } 1141 } 1142 1143 // Simplify the IFG by removing LRGs of low degree that have NO copies 1144 void PhaseChaitin::Pre_Simplify( ) { 1145 1146 // Warm up the lo-degree no-copy list 1147 int lo_no_copy = 0; 1148 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1149 if ((lrgs(i).lo_degree() && !lrgs(i)._has_copy) || 1150 !lrgs(i).alive() || 1151 lrgs(i)._must_spill) { 1152 lrgs(i)._next = lo_no_copy; 1153 lo_no_copy = i; 1154 } 1155 } 1156 1157 while( lo_no_copy ) { 1158 uint lo = lo_no_copy; 1159 lo_no_copy = lrgs(lo)._next; 1160 int size = lrgs(lo).num_regs(); 1161 1162 // Put the simplified guy on the simplified list. 1163 lrgs(lo)._next = _simplified; 1164 _simplified = lo; 1165 1166 // Yank this guy from the IFG. 1167 IndexSet *adj = _ifg->remove_node( lo ); 1168 1169 // If any neighbors' degrees fall below their number of 1170 // allowed registers, then put that neighbor on the low degree 1171 // list. Note that 'degree' can only fall and 'numregs' is 1172 // unchanged by this action. Thus the two are equal at most once, 1173 // so LRGs hit the lo-degree worklists at most once. 1174 IndexSetIterator elements(adj); 1175 uint neighbor; 1176 while ((neighbor = elements.next()) != 0) { 1177 LRG *n = &lrgs(neighbor); 1178 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1179 1180 // Check for just becoming of-low-degree 1181 if( n->just_lo_degree() && !n->_has_copy ) { 1182 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1183 // Put on lo-degree list 1184 n->_next = lo_no_copy; 1185 lo_no_copy = neighbor; 1186 } 1187 } 1188 } // End of while lo-degree no_copy worklist not empty 1189 1190 // No more lo-degree no-copy live ranges to simplify 1191 } 1192 1193 // Simplify the IFG by removing LRGs of low degree. 1194 void PhaseChaitin::Simplify( ) { 1195 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]); 1196 1197 while( 1 ) { // Repeat till simplified it all 1198 // May want to explore simplifying lo_degree before _lo_stk_degree. 1199 // This might result in more spills coloring into registers during 1200 // Select(). 1201 while( _lo_degree || _lo_stk_degree ) { 1202 // If possible, pull from lo_stk first 1203 uint lo; 1204 if( _lo_degree ) { 1205 lo = _lo_degree; 1206 _lo_degree = lrgs(lo)._next; 1207 } else { 1208 lo = _lo_stk_degree; 1209 _lo_stk_degree = lrgs(lo)._next; 1210 } 1211 1212 // Put the simplified guy on the simplified list. 1213 lrgs(lo)._next = _simplified; 1214 _simplified = lo; 1215 // If this guy is "at risk" then mark his current neighbors 1216 if( lrgs(lo)._at_risk ) { 1217 IndexSetIterator elements(_ifg->neighbors(lo)); 1218 uint datum; 1219 while ((datum = elements.next()) != 0) { 1220 lrgs(datum)._risk_bias = lo; 1221 } 1222 } 1223 1224 // Yank this guy from the IFG. 1225 IndexSet *adj = _ifg->remove_node( lo ); 1226 1227 // If any neighbors' degrees fall below their number of 1228 // allowed registers, then put that neighbor on the low degree 1229 // list. Note that 'degree' can only fall and 'numregs' is 1230 // unchanged by this action. Thus the two are equal at most once, 1231 // so LRGs hit the lo-degree worklist at most once. 1232 IndexSetIterator elements(adj); 1233 uint neighbor; 1234 while ((neighbor = elements.next()) != 0) { 1235 LRG *n = &lrgs(neighbor); 1236 #ifdef ASSERT 1237 if( VerifyOpto || VerifyRegisterAllocator ) { 1238 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1239 } 1240 #endif 1241 1242 // Check for just becoming of-low-degree just counting registers. 1243 // _must_spill live ranges are already on the low degree list. 1244 if( n->just_lo_degree() && !n->_must_spill ) { 1245 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1246 // Pull from hi-degree list 1247 uint prev = n->_prev; 1248 uint next = n->_next; 1249 if( prev ) lrgs(prev)._next = next; 1250 else _hi_degree = next; 1251 lrgs(next)._prev = prev; 1252 n->_next = _lo_degree; 1253 _lo_degree = neighbor; 1254 } 1255 } 1256 } // End of while lo-degree/lo_stk_degree worklist not empty 1257 1258 // Check for got everything: is hi-degree list empty? 1259 if( !_hi_degree ) break; 1260 1261 // Time to pick a potential spill guy 1262 uint lo_score = _hi_degree; 1263 double score = lrgs(lo_score).score(); 1264 double area = lrgs(lo_score)._area; 1265 double cost = lrgs(lo_score)._cost; 1266 bool bound = lrgs(lo_score)._is_bound; 1267 1268 // Find cheapest guy 1269 debug_only( int lo_no_simplify=0; ); 1270 for( uint i = _hi_degree; i; i = lrgs(i)._next ) { 1271 assert( !(*_ifg->_yanked)[i], "" ); 1272 // It's just vaguely possible to move hi-degree to lo-degree without 1273 // going through a just-lo-degree stage: If you remove a double from 1274 // a float live range it's degree will drop by 2 and you can skip the 1275 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 1276 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1277 if( lrgs(i).lo_degree() ) { 1278 lo_score = i; 1279 break; 1280 } 1281 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1282 double iscore = lrgs(i).score(); 1283 double iarea = lrgs(i)._area; 1284 double icost = lrgs(i)._cost; 1285 bool ibound = lrgs(i)._is_bound; 1286 1287 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1288 // wins. Ties happen because all live ranges in question have spilled 1289 // a few times before and the spill-score adds a huge number which 1290 // washes out the low order bits. We are choosing the lesser of 2 1291 // evils; in this case pick largest area to spill. 1292 // Ties also happen when live ranges are defined and used only inside 1293 // one block. In which case their area is 0 and score set to max. 1294 // In such case choose bound live range over unbound to free registers 1295 // or with smaller cost to spill. 1296 if( iscore < score || 1297 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || 1298 (iscore == score && iarea == area && 1299 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) { 1300 lo_score = i; 1301 score = iscore; 1302 area = iarea; 1303 cost = icost; 1304 bound = ibound; 1305 } 1306 } 1307 LRG *lo_lrg = &lrgs(lo_score); 1308 // The live range we choose for spilling is either hi-degree, or very 1309 // rarely it can be low-degree. If we choose a hi-degree live range 1310 // there better not be any lo-degree choices. 1311 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1312 1313 // Pull from hi-degree list 1314 uint prev = lo_lrg->_prev; 1315 uint next = lo_lrg->_next; 1316 if( prev ) lrgs(prev)._next = next; 1317 else _hi_degree = next; 1318 lrgs(next)._prev = prev; 1319 // Jam him on the lo-degree list, despite his high degree. 1320 // Maybe he'll get a color, and maybe he'll spill. 1321 // Only Select() will know. 1322 lrgs(lo_score)._at_risk = true; 1323 _lo_degree = lo_score; 1324 lo_lrg->_next = 0; 1325 1326 } // End of while not simplified everything 1327 1328 } 1329 1330 // Is 'reg' register legal for 'lrg'? 1331 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { 1332 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && 1333 lrg.mask().Member(OptoReg::add(reg,-chunk))) { 1334 // RA uses OptoReg which represent the highest element of a registers set. 1335 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set 1336 // in which XMMd is used by RA to represent such vectors. A double value 1337 // uses [XMM,XMMb] pairs and XMMb is used by RA for it. 1338 // The register mask uses largest bits set of overlapping register sets. 1339 // On x86 with AVX it uses 8 bits for each XMM registers set. 1340 // 1341 // The 'lrg' already has cleared-to-set register mask (done in Select() 1342 // before calling choose_color()). Passing mask.Member(reg) check above 1343 // indicates that the size (num_regs) of 'reg' set is less or equal to 1344 // 'lrg' set size. 1345 // For set size 1 any register which is member of 'lrg' mask is legal. 1346 if (lrg.num_regs()==1) 1347 return true; 1348 // For larger sets only an aligned register with the same set size is legal. 1349 int mask = lrg.num_regs()-1; 1350 if ((reg&mask) == mask) 1351 return true; 1352 } 1353 return false; 1354 } 1355 1356 // Choose a color using the biasing heuristic 1357 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1358 1359 // Check for "at_risk" LRG's 1360 uint risk_lrg = _lrg_map.find(lrg._risk_bias); 1361 if( risk_lrg != 0 ) { 1362 // Walk the colored neighbors of the "at_risk" candidate 1363 // Choose a color which is both legal and already taken by a neighbor 1364 // of the "at_risk" candidate in order to improve the chances of the 1365 // "at_risk" candidate of coloring 1366 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1367 uint datum; 1368 while ((datum = elements.next()) != 0) { 1369 OptoReg::Name reg = lrgs(datum).reg(); 1370 // If this LRG's register is legal for us, choose it 1371 if (is_legal_reg(lrg, reg, chunk)) 1372 return reg; 1373 } 1374 } 1375 1376 uint copy_lrg = _lrg_map.find(lrg._copy_bias); 1377 if( copy_lrg != 0 ) { 1378 // If he has a color, 1379 if( !(*(_ifg->_yanked))[copy_lrg] ) { 1380 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1381 // And it is legal for you, 1382 if (is_legal_reg(lrg, reg, chunk)) 1383 return reg; 1384 } else if( chunk == 0 ) { 1385 // Choose a color which is legal for him 1386 RegMask tempmask = lrg.mask(); 1387 tempmask.AND(lrgs(copy_lrg).mask()); 1388 tempmask.clear_to_sets(lrg.num_regs()); 1389 OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); 1390 if (OptoReg::is_valid(reg)) 1391 return reg; 1392 } 1393 } 1394 1395 // If no bias info exists, just go with the register selection ordering 1396 if (lrg._is_vector || lrg.num_regs() == 2) { 1397 // Find an aligned set 1398 return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); 1399 } 1400 1401 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1402 // copy removal to remove many more copies, by preventing a just-assigned 1403 // register from being repeatedly assigned. 1404 OptoReg::Name reg = lrg.mask().find_first_elem(); 1405 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1406 // This 'Remove; find; Insert' idiom is an expensive way to find the 1407 // SECOND element in the mask. 1408 lrg.Remove(reg); 1409 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1410 lrg.Insert(reg); 1411 if( OptoReg::is_reg(reg2)) 1412 reg = reg2; 1413 } 1414 return OptoReg::add( reg, chunk ); 1415 } 1416 1417 // Choose a color in the current chunk 1418 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1419 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)"); 1420 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)"); 1421 1422 if( lrg.num_regs() == 1 || // Common Case 1423 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1424 // Use a heuristic to "bias" the color choice 1425 return bias_color(lrg, chunk); 1426 1427 assert(!lrg._is_vector, "should be not vector here" ); 1428 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1429 1430 // Fat-proj case or misaligned double argument. 1431 assert(lrg.compute_mask_size() == lrg.num_regs() || 1432 lrg.num_regs() == 2,"fat projs exactly color" ); 1433 assert( !chunk, "always color in 1st chunk" ); 1434 // Return the highest element in the set. 1435 return lrg.mask().find_last_elem(); 1436 } 1437 1438 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1439 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1440 // everything going back is guaranteed a color. Select that color. If some 1441 // hi-degree LRG cannot get a color then we record that we must spill. 1442 uint PhaseChaitin::Select( ) { 1443 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]); 1444 1445 uint spill_reg = LRG::SPILL_REG; 1446 _max_reg = OptoReg::Name(0); // Past max register used 1447 while( _simplified ) { 1448 // Pull next LRG from the simplified list - in reverse order of removal 1449 uint lidx = _simplified; 1450 LRG *lrg = &lrgs(lidx); 1451 _simplified = lrg->_next; 1452 1453 1454 #ifndef PRODUCT 1455 if (trace_spilling()) { 1456 ttyLocker ttyl; 1457 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1458 lrg->degrees_of_freedom()); 1459 lrg->dump(); 1460 } 1461 #endif 1462 1463 // Re-insert into the IFG 1464 _ifg->re_insert(lidx); 1465 if( !lrg->alive() ) continue; 1466 // capture allstackedness flag before mask is hacked 1467 const int is_allstack = lrg->mask().is_AllStack(); 1468 1469 // Yeah, yeah, yeah, I know, I know. I can refactor this 1470 // to avoid the GOTO, although the refactored code will not 1471 // be much clearer. We arrive here IFF we have a stack-based 1472 // live range that cannot color in the current chunk, and it 1473 // has to move into the next free stack chunk. 1474 int chunk = 0; // Current chunk is first chunk 1475 retry_next_chunk: 1476 1477 // Remove neighbor colors 1478 IndexSet *s = _ifg->neighbors(lidx); 1479 1480 debug_only(RegMask orig_mask = lrg->mask();) 1481 IndexSetIterator elements(s); 1482 uint neighbor; 1483 while ((neighbor = elements.next()) != 0) { 1484 // Note that neighbor might be a spill_reg. In this case, exclusion 1485 // of its color will be a no-op, since the spill_reg chunk is in outer 1486 // space. Also, if neighbor is in a different chunk, this exclusion 1487 // will be a no-op. (Later on, if lrg runs out of possible colors in 1488 // its chunk, a new chunk of color may be tried, in which case 1489 // examination of neighbors is started again, at retry_next_chunk.) 1490 LRG &nlrg = lrgs(neighbor); 1491 OptoReg::Name nreg = nlrg.reg(); 1492 // Only subtract masks in the same chunk 1493 if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { 1494 #ifndef PRODUCT 1495 uint size = lrg->mask().Size(); 1496 RegMask rm = lrg->mask(); 1497 #endif 1498 lrg->SUBTRACT(nlrg.mask()); 1499 #ifndef PRODUCT 1500 if (trace_spilling() && lrg->mask().Size() != size) { 1501 ttyLocker ttyl; 1502 tty->print("L%d ", lidx); 1503 rm.dump(); 1504 tty->print(" intersected L%d ", neighbor); 1505 nlrg.mask().dump(); 1506 tty->print(" removed "); 1507 rm.SUBTRACT(lrg->mask()); 1508 rm.dump(); 1509 tty->print(" leaving "); 1510 lrg->mask().dump(); 1511 tty->cr(); 1512 } 1513 #endif 1514 } 1515 } 1516 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1517 // Aligned pairs need aligned masks 1518 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1519 if (lrg->num_regs() > 1 && !lrg->_fat_proj) { 1520 lrg->clear_to_sets(); 1521 } 1522 1523 // Check if a color is available and if so pick the color 1524 OptoReg::Name reg = choose_color( *lrg, chunk ); 1525 #ifdef SPARC 1526 debug_only(lrg->compute_set_mask_size()); 1527 assert(lrg->num_regs() < 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); 1528 #endif 1529 1530 //--------------- 1531 // If we fail to color and the AllStack flag is set, trigger 1532 // a chunk-rollover event 1533 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1534 // Bump register mask up to next stack chunk 1535 chunk += RegMask::CHUNK_SIZE; 1536 lrg->Set_All(); 1537 1538 goto retry_next_chunk; 1539 } 1540 1541 //--------------- 1542 // Did we get a color? 1543 else if( OptoReg::is_valid(reg)) { 1544 #ifndef PRODUCT 1545 RegMask avail_rm = lrg->mask(); 1546 #endif 1547 1548 // Record selected register 1549 lrg->set_reg(reg); 1550 1551 if( reg >= _max_reg ) // Compute max register limit 1552 _max_reg = OptoReg::add(reg,1); 1553 // Fold reg back into normal space 1554 reg = OptoReg::add(reg,-chunk); 1555 1556 // If the live range is not bound, then we actually had some choices 1557 // to make. In this case, the mask has more bits in it than the colors 1558 // chosen. Restrict the mask to just what was picked. 1559 int n_regs = lrg->num_regs(); 1560 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1561 if (n_regs == 1 || !lrg->_fat_proj) { 1562 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity"); 1563 lrg->Clear(); // Clear the mask 1564 lrg->Insert(reg); // Set regmask to match selected reg 1565 // For vectors and pairs, also insert the low bit of the pair 1566 for (int i = 1; i < n_regs; i++) 1567 lrg->Insert(OptoReg::add(reg,-i)); 1568 lrg->set_mask_size(n_regs); 1569 } else { // Else fatproj 1570 // mask must be equal to fatproj bits, by definition 1571 } 1572 #ifndef PRODUCT 1573 if (trace_spilling()) { 1574 ttyLocker ttyl; 1575 tty->print("L%d selected ", lidx); 1576 lrg->mask().dump(); 1577 tty->print(" from "); 1578 avail_rm.dump(); 1579 tty->cr(); 1580 } 1581 #endif 1582 // Note that reg is the highest-numbered register in the newly-bound mask. 1583 } // end color available case 1584 1585 //--------------- 1586 // Live range is live and no colors available 1587 else { 1588 assert( lrg->alive(), "" ); 1589 assert( !lrg->_fat_proj || lrg->is_multidef() || 1590 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1591 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1592 1593 // Assign the special spillreg register 1594 lrg->set_reg(OptoReg::Name(spill_reg++)); 1595 // Do not empty the regmask; leave mask_size lying around 1596 // for use during Spilling 1597 #ifndef PRODUCT 1598 if( trace_spilling() ) { 1599 ttyLocker ttyl; 1600 tty->print("L%d spilling with neighbors: ", lidx); 1601 s->dump(); 1602 debug_only(tty->print(" original mask: ")); 1603 debug_only(orig_mask.dump()); 1604 dump_lrg(lidx); 1605 } 1606 #endif 1607 } // end spill case 1608 1609 } 1610 1611 return spill_reg-LRG::SPILL_REG; // Return number of spills 1612 } 1613 1614 // Copy 'was_spilled'-edness from the source Node to the dst Node. 1615 void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { 1616 if( _spilled_once.test(src->_idx) ) { 1617 _spilled_once.set(dst->_idx); 1618 lrgs(_lrg_map.find(dst))._was_spilled1 = 1; 1619 if( _spilled_twice.test(src->_idx) ) { 1620 _spilled_twice.set(dst->_idx); 1621 lrgs(_lrg_map.find(dst))._was_spilled2 = 1; 1622 } 1623 } 1624 } 1625 1626 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1627 void PhaseChaitin::set_was_spilled( Node *n ) { 1628 if( _spilled_once.test_set(n->_idx) ) 1629 _spilled_twice.set(n->_idx); 1630 } 1631 1632 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1633 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1634 void PhaseChaitin::fixup_spills() { 1635 // This function does only cisc spill work. 1636 if( !UseCISCSpill ) return; 1637 1638 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]); 1639 1640 // Grab the Frame Pointer 1641 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); 1642 1643 // For all blocks 1644 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1645 Block* block = _cfg.get_block(i); 1646 1647 // For all instructions in block 1648 uint last_inst = block->end_idx(); 1649 for (uint j = 1; j <= last_inst; j++) { 1650 Node* n = block->get_node(j); 1651 1652 // Dead instruction??? 1653 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1654 C->top() == n || // Or the random TOP node 1655 n->is_Proj(), // Or a fat-proj kill node 1656 "No dead instructions after post-alloc" ); 1657 1658 int inp = n->cisc_operand(); 1659 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1660 // Convert operand number to edge index number 1661 MachNode *mach = n->as_Mach(); 1662 inp = mach->operand_index(inp); 1663 Node *src = n->in(inp); // Value to load or store 1664 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); 1665 OptoReg::Name src_reg = lrg_cisc.reg(); 1666 // Doubles record the HIGH register of an adjacent pair. 1667 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1668 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1669 // This is a CISC Spill, get stack offset and construct new node 1670 #ifndef PRODUCT 1671 if( TraceCISCSpill ) { 1672 tty->print(" reg-instr: "); 1673 n->dump(); 1674 } 1675 #endif 1676 int stk_offset = reg2offset(src_reg); 1677 // Bailout if we might exceed node limit when spilling this instruction 1678 C->check_node_count(0, "out of nodes fixing spills"); 1679 if (C->failing()) return; 1680 // Transform node 1681 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); 1682 cisc->set_req(inp,fp); // Base register is frame pointer 1683 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1684 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1685 cisc->ins_req(1,src); // Requires a memory edge 1686 } 1687 block->map_node(cisc, j); // Insert into basic block 1688 n->subsume_by(cisc, C); // Correct graph 1689 // 1690 ++_used_cisc_instructions; 1691 #ifndef PRODUCT 1692 if( TraceCISCSpill ) { 1693 tty->print(" cisc-instr: "); 1694 cisc->dump(); 1695 } 1696 #endif 1697 } else { 1698 #ifndef PRODUCT 1699 if( TraceCISCSpill ) { 1700 tty->print(" using reg-instr: "); 1701 n->dump(); 1702 } 1703 #endif 1704 ++_unused_cisc_instructions; // input can be on stack 1705 } 1706 } 1707 1708 } // End of for all instructions 1709 1710 } // End of for all blocks 1711 } 1712 1713 // Helper to stretch above; recursively discover the base Node for a 1714 // given derived Node. Easy for AddP-related machine nodes, but needs 1715 // to be recursive for derived Phis. 1716 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1717 // See if already computed; if so return it 1718 if( derived_base_map[derived->_idx] ) 1719 return derived_base_map[derived->_idx]; 1720 1721 // See if this happens to be a base. 1722 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1723 // pointers derived from NULL! These are always along paths that 1724 // can't happen at run-time but the optimizer cannot deduce it so 1725 // we have to handle it gracefully. 1726 assert(!derived->bottom_type()->isa_narrowoop() || 1727 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1728 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1729 // If its an OOP with a non-zero offset, then it is derived. 1730 if( tj == NULL || tj->_offset == 0 ) { 1731 derived_base_map[derived->_idx] = derived; 1732 return derived; 1733 } 1734 // Derived is NULL+offset? Base is NULL! 1735 if( derived->is_Con() ) { 1736 Node *base = _matcher.mach_null(); 1737 assert(base != NULL, "sanity"); 1738 if (base->in(0) == NULL) { 1739 // Initialize it once and make it shared: 1740 // set control to _root and place it into Start block 1741 // (where top() node is placed). 1742 base->init_req(0, _cfg.get_root_node()); 1743 Block *startb = _cfg.get_block_for_node(C->top()); 1744 uint node_pos = startb->find_node(C->top()); 1745 startb->insert_node(base, node_pos); 1746 _cfg.map_node_to_block(base, startb); 1747 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet"); 1748 1749 // The loadConP0 might have projection nodes depending on architecture 1750 // Add the projection nodes to the CFG 1751 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { 1752 Node* use = base->fast_out(i); 1753 if (use->is_MachProj()) { 1754 startb->insert_node(use, ++node_pos); 1755 _cfg.map_node_to_block(use, startb); 1756 new_lrg(use, maxlrg++); 1757 } 1758 } 1759 } 1760 if (_lrg_map.live_range_id(base) == 0) { 1761 new_lrg(base, maxlrg++); 1762 } 1763 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"); 1764 derived_base_map[derived->_idx] = base; 1765 return base; 1766 } 1767 1768 // Check for AddP-related opcodes 1769 if (!derived->is_Phi()) { 1770 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name()); 1771 Node *base = derived->in(AddPNode::Base); 1772 derived_base_map[derived->_idx] = base; 1773 return base; 1774 } 1775 1776 // Recursively find bases for Phis. 1777 // First check to see if we can avoid a base Phi here. 1778 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1779 uint i; 1780 for( i = 2; i < derived->req(); i++ ) 1781 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1782 break; 1783 // Went to the end without finding any different bases? 1784 if( i == derived->req() ) { // No need for a base Phi here 1785 derived_base_map[derived->_idx] = base; 1786 return base; 1787 } 1788 1789 // Now we see we need a base-Phi here to merge the bases 1790 const Type *t = base->bottom_type(); 1791 base = new PhiNode( derived->in(0), t ); 1792 for( i = 1; i < derived->req(); i++ ) { 1793 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1794 t = t->meet(base->in(i)->bottom_type()); 1795 } 1796 base->as_Phi()->set_type(t); 1797 1798 // Search the current block for an existing base-Phi 1799 Block *b = _cfg.get_block_for_node(derived); 1800 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1801 Node *phi = b->get_node(i); 1802 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1803 b->insert_node(base, i); // Must insert created Phi here as base 1804 _cfg.map_node_to_block(base, b); 1805 new_lrg(base,maxlrg++); 1806 break; 1807 } 1808 // See if Phi matches. 1809 uint j; 1810 for( j = 1; j < base->req(); j++ ) 1811 if( phi->in(j) != base->in(j) && 1812 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1813 break; 1814 if( j == base->req() ) { // All inputs match? 1815 base = phi; // Then use existing 'phi' and drop 'base' 1816 break; 1817 } 1818 } 1819 1820 1821 // Cache info for later passes 1822 derived_base_map[derived->_idx] = base; 1823 return base; 1824 } 1825 1826 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1827 // base pointer that is live across the Safepoint for oopmap building. The 1828 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1829 // required edge set. 1830 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { 1831 int must_recompute_live = false; 1832 uint maxlrg = _lrg_map.max_lrg_id(); 1833 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1834 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1835 1836 // For all blocks in RPO do... 1837 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1838 Block* block = _cfg.get_block(i); 1839 // Note use of deep-copy constructor. I cannot hammer the original 1840 // liveout bits, because they are needed by the following coalesce pass. 1841 IndexSet liveout(_live->live(block)); 1842 1843 for (uint j = block->end_idx() + 1; j > 1; j--) { 1844 Node* n = block->get_node(j - 1); 1845 1846 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1847 // like to see in the same register. Compare uses the loop-phi and so 1848 // extends its live range BUT cannot be part of the cycle. If this 1849 // extended live range overlaps with the update of the loop-phi value 1850 // we need both alive at the same time -- which requires at least 1 1851 // copy. But because Intel has only 2-address registers we end up with 1852 // at least 2 copies, one before the loop-phi update instruction and 1853 // one after. Instead we split the input to the compare just after the 1854 // phi. 1855 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1856 Node *phi = n->in(1); 1857 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1858 Block *phi_block = _cfg.get_block_for_node(phi); 1859 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { 1860 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1861 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); 1862 insert_proj( phi_block, 1, spill, maxlrg++ ); 1863 n->set_req(1,spill); 1864 must_recompute_live = true; 1865 } 1866 } 1867 } 1868 1869 // Get value being defined 1870 uint lidx = _lrg_map.live_range_id(n); 1871 // Ignore the occasional brand-new live range 1872 if (lidx && lidx < _lrg_map.max_lrg_id()) { 1873 // Remove from live-out set 1874 liveout.remove(lidx); 1875 1876 // Copies do not define a new value and so do not interfere. 1877 // Remove the copies source from the liveout set before interfering. 1878 uint idx = n->is_Copy(); 1879 if (idx) { 1880 liveout.remove(_lrg_map.live_range_id(n->in(idx))); 1881 } 1882 } 1883 1884 // Found a safepoint? 1885 JVMState *jvms = n->jvms(); 1886 if( jvms ) { 1887 // Now scan for a live derived pointer 1888 IndexSetIterator elements(&liveout); 1889 uint neighbor; 1890 while ((neighbor = elements.next()) != 0) { 1891 // Find reaching DEF for base and derived values 1892 // This works because we are still in SSA during this call. 1893 Node *derived = lrgs(neighbor)._def; 1894 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1895 assert(!derived->bottom_type()->isa_narrowoop() || 1896 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1897 // If its an OOP with a non-zero offset, then it is derived. 1898 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { 1899 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); 1900 assert(base->_idx < _lrg_map.size(), ""); 1901 // Add reaching DEFs of derived pointer and base pointer as a 1902 // pair of inputs 1903 n->add_req(derived); 1904 n->add_req(base); 1905 1906 // See if the base pointer is already live to this point. 1907 // Since I'm working on the SSA form, live-ness amounts to 1908 // reaching def's. So if I find the base's live range then 1909 // I know the base's def reaches here. 1910 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or 1911 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND 1912 (_lrg_map.live_range_id(base) > 0) && // not a constant 1913 _cfg.get_block_for_node(base) != block) { // base not def'd in blk) 1914 // Base pointer is not currently live. Since I stretched 1915 // the base pointer to here and it crosses basic-block 1916 // boundaries, the global live info is now incorrect. 1917 // Recompute live. 1918 must_recompute_live = true; 1919 } // End of if base pointer is not live to debug info 1920 } 1921 } // End of scan all live data for derived ptrs crossing GC point 1922 } // End of if found a GC point 1923 1924 // Make all inputs live 1925 if (!n->is_Phi()) { // Phi function uses come from prior block 1926 for (uint k = 1; k < n->req(); k++) { 1927 uint lidx = _lrg_map.live_range_id(n->in(k)); 1928 if (lidx < _lrg_map.max_lrg_id()) { 1929 liveout.insert(lidx); 1930 } 1931 } 1932 } 1933 1934 } // End of forall instructions in block 1935 liveout.clear(); // Free the memory used by liveout. 1936 1937 } // End of forall blocks 1938 _lrg_map.set_max_lrg_id(maxlrg); 1939 1940 // If I created a new live range I need to recompute live 1941 if (maxlrg != _ifg->_maxlrg) { 1942 must_recompute_live = true; 1943 } 1944 1945 return must_recompute_live != 0; 1946 } 1947 1948 // Extend the node to LRG mapping 1949 1950 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { 1951 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); 1952 } 1953 1954 #ifndef PRODUCT 1955 void PhaseChaitin::dump(const Node *n) const { 1956 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; 1957 tty->print("L%d",r); 1958 if (r && n->Opcode() != Op_Phi) { 1959 if( _node_regs ) { // Got a post-allocation copy of allocation? 1960 tty->print("["); 1961 OptoReg::Name second = get_reg_second(n); 1962 if( OptoReg::is_valid(second) ) { 1963 if( OptoReg::is_reg(second) ) 1964 tty->print("%s:",Matcher::regName[second]); 1965 else 1966 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1967 } 1968 OptoReg::Name first = get_reg_first(n); 1969 if( OptoReg::is_reg(first) ) 1970 tty->print("%s]",Matcher::regName[first]); 1971 else 1972 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1973 } else 1974 n->out_RegMask().dump(); 1975 } 1976 tty->print("/N%d\t",n->_idx); 1977 tty->print("%s === ", n->Name()); 1978 uint k; 1979 for (k = 0; k < n->req(); k++) { 1980 Node *m = n->in(k); 1981 if (!m) { 1982 tty->print("_ "); 1983 } 1984 else { 1985 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1986 tty->print("L%d",r); 1987 // Data MultiNode's can have projections with no real registers. 1988 // Don't die while dumping them. 1989 int op = n->Opcode(); 1990 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1991 if( _node_regs ) { 1992 tty->print("["); 1993 OptoReg::Name second = get_reg_second(n->in(k)); 1994 if( OptoReg::is_valid(second) ) { 1995 if( OptoReg::is_reg(second) ) 1996 tty->print("%s:",Matcher::regName[second]); 1997 else 1998 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 1999 reg2offset_unchecked(second)); 2000 } 2001 OptoReg::Name first = get_reg_first(n->in(k)); 2002 if( OptoReg::is_reg(first) ) 2003 tty->print("%s]",Matcher::regName[first]); 2004 else 2005 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 2006 reg2offset_unchecked(first)); 2007 } else 2008 n->in_RegMask(k).dump(); 2009 } 2010 tty->print("/N%d ",m->_idx); 2011 } 2012 } 2013 if( k < n->len() && n->in(k) ) tty->print("| "); 2014 for( ; k < n->len(); k++ ) { 2015 Node *m = n->in(k); 2016 if(!m) { 2017 break; 2018 } 2019 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 2020 tty->print("L%d",r); 2021 tty->print("/N%d ",m->_idx); 2022 } 2023 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 2024 else n->dump_spec(tty); 2025 if( _spilled_once.test(n->_idx ) ) { 2026 tty->print(" Spill_1"); 2027 if( _spilled_twice.test(n->_idx ) ) 2028 tty->print(" Spill_2"); 2029 } 2030 tty->print("\n"); 2031 } 2032 2033 void PhaseChaitin::dump(const Block *b) const { 2034 b->dump_head(&_cfg); 2035 2036 // For all instructions 2037 for( uint j = 0; j < b->number_of_nodes(); j++ ) 2038 dump(b->get_node(j)); 2039 // Print live-out info at end of block 2040 if( _live ) { 2041 tty->print("Liveout: "); 2042 IndexSet *live = _live->live(b); 2043 IndexSetIterator elements(live); 2044 tty->print("{"); 2045 uint i; 2046 while ((i = elements.next()) != 0) { 2047 tty->print("L%d ", _lrg_map.find_const(i)); 2048 } 2049 tty->print_cr("}"); 2050 } 2051 tty->print("\n"); 2052 } 2053 2054 void PhaseChaitin::dump() const { 2055 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 2056 _matcher._new_SP, _framesize ); 2057 2058 // For all blocks 2059 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2060 dump(_cfg.get_block(i)); 2061 } 2062 // End of per-block dump 2063 tty->print("\n"); 2064 2065 if (!_ifg) { 2066 tty->print("(No IFG.)\n"); 2067 return; 2068 } 2069 2070 // Dump LRG array 2071 tty->print("--- Live RanGe Array ---\n"); 2072 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { 2073 tty->print("L%d: ",i2); 2074 if (i2 < _ifg->_maxlrg) { 2075 lrgs(i2).dump(); 2076 } 2077 else { 2078 tty->print_cr("new LRG"); 2079 } 2080 } 2081 tty->cr(); 2082 2083 // Dump lo-degree list 2084 tty->print("Lo degree: "); 2085 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 2086 tty->print("L%d ",i3); 2087 tty->cr(); 2088 2089 // Dump lo-stk-degree list 2090 tty->print("Lo stk degree: "); 2091 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 2092 tty->print("L%d ",i4); 2093 tty->cr(); 2094 2095 // Dump lo-degree list 2096 tty->print("Hi degree: "); 2097 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 2098 tty->print("L%d ",i5); 2099 tty->cr(); 2100 } 2101 2102 void PhaseChaitin::dump_degree_lists() const { 2103 // Dump lo-degree list 2104 tty->print("Lo degree: "); 2105 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 2106 tty->print("L%d ",i); 2107 tty->cr(); 2108 2109 // Dump lo-stk-degree list 2110 tty->print("Lo stk degree: "); 2111 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 2112 tty->print("L%d ",i2); 2113 tty->cr(); 2114 2115 // Dump lo-degree list 2116 tty->print("Hi degree: "); 2117 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 2118 tty->print("L%d ",i3); 2119 tty->cr(); 2120 } 2121 2122 void PhaseChaitin::dump_simplified() const { 2123 tty->print("Simplified: "); 2124 for( uint i = _simplified; i; i = lrgs(i)._next ) 2125 tty->print("L%d ",i); 2126 tty->cr(); 2127 } 2128 2129 static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { 2130 if ((int)reg < 0) 2131 sprintf(buf, "<OptoReg::%d>", (int)reg); 2132 else if (OptoReg::is_reg(reg)) 2133 strcpy(buf, Matcher::regName[reg]); 2134 else 2135 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 2136 pc->reg2offset(reg)); 2137 return buf+strlen(buf); 2138 } 2139 2140 // Dump a register name into a buffer. Be intelligent if we get called 2141 // before allocation is complete. 2142 char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { 2143 if( this == NULL ) { // Not got anything? 2144 sprintf(buf,"N%d",n->_idx); // Then use Node index 2145 } else if( _node_regs ) { 2146 // Post allocation, use direct mappings, no LRG info available 2147 print_reg( get_reg_first(n), this, buf ); 2148 } else { 2149 uint lidx = _lrg_map.find_const(n); // Grab LRG number 2150 if( !_ifg ) { 2151 sprintf(buf,"L%d",lidx); // No register binding yet 2152 } else if( !lidx ) { // Special, not allocated value 2153 strcpy(buf,"Special"); 2154 } else { 2155 if (lrgs(lidx)._is_vector) { 2156 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) 2157 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register 2158 else 2159 sprintf(buf,"L%d",lidx); // No register binding yet 2160 } else if( (lrgs(lidx).num_regs() == 1) 2161 ? lrgs(lidx).mask().is_bound1() 2162 : lrgs(lidx).mask().is_bound_pair() ) { 2163 // Hah! We have a bound machine register 2164 print_reg( lrgs(lidx).reg(), this, buf ); 2165 } else { 2166 sprintf(buf,"L%d",lidx); // No register binding yet 2167 } 2168 } 2169 } 2170 return buf+strlen(buf); 2171 } 2172 2173 void PhaseChaitin::dump_for_spill_split_recycle() const { 2174 if( WizardMode && (PrintCompilation || PrintOpto) ) { 2175 // Display which live ranges need to be split and the allocator's state 2176 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 2177 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { 2178 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 2179 tty->print("L%d: ", bidx); 2180 lrgs(bidx).dump(); 2181 } 2182 } 2183 tty->cr(); 2184 dump(); 2185 } 2186 } 2187 2188 void PhaseChaitin::dump_frame() const { 2189 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 2190 const TypeTuple *domain = C->tf()->domain(); 2191 const int argcnt = domain->cnt() - TypeFunc::Parms; 2192 2193 // Incoming arguments in registers dump 2194 for( int k = 0; k < argcnt; k++ ) { 2195 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 2196 if( OptoReg::is_reg(parmreg)) { 2197 const char *reg_name = OptoReg::regname(parmreg); 2198 tty->print("#r%3.3d %s", parmreg, reg_name); 2199 parmreg = _matcher._parm_regs[k].second(); 2200 if( OptoReg::is_reg(parmreg)) { 2201 tty->print(":%s", OptoReg::regname(parmreg)); 2202 } 2203 tty->print(" : parm %d: ", k); 2204 domain->field_at(k + TypeFunc::Parms)->dump(); 2205 tty->cr(); 2206 } 2207 } 2208 2209 // Check for un-owned padding above incoming args 2210 OptoReg::Name reg = _matcher._new_SP; 2211 if( reg > _matcher._in_arg_limit ) { 2212 reg = OptoReg::add(reg, -1); 2213 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 2214 } 2215 2216 // Incoming argument area dump 2217 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 2218 while( reg > begin_in_arg ) { 2219 reg = OptoReg::add(reg, -1); 2220 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2221 int j; 2222 for( j = 0; j < argcnt; j++) { 2223 if( _matcher._parm_regs[j].first() == reg || 2224 _matcher._parm_regs[j].second() == reg ) { 2225 tty->print("parm %d: ",j); 2226 domain->field_at(j + TypeFunc::Parms)->dump(); 2227 tty->cr(); 2228 break; 2229 } 2230 } 2231 if( j >= argcnt ) 2232 tty->print_cr("HOLE, owned by SELF"); 2233 } 2234 2235 // Old outgoing preserve area 2236 while( reg > _matcher._old_SP ) { 2237 reg = OptoReg::add(reg, -1); 2238 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 2239 } 2240 2241 // Old SP 2242 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 2243 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 2244 2245 // Preserve area dump 2246 int fixed_slots = C->fixed_slots(); 2247 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); 2248 OptoReg::Name return_addr = _matcher.return_addr(); 2249 2250 reg = OptoReg::add(reg, -1); 2251 while (OptoReg::is_stack(reg)) { 2252 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2253 if (return_addr == reg) { 2254 tty->print_cr("return address"); 2255 } else if (reg >= begin_in_preserve) { 2256 // Preserved slots are present on x86 2257 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) 2258 tty->print_cr("saved fp register"); 2259 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && 2260 VerifyStackAtCalls) 2261 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 2262 else 2263 tty->print_cr("in_preserve"); 2264 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { 2265 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 2266 } else { 2267 tty->print_cr("pad2, stack alignment"); 2268 } 2269 reg = OptoReg::add(reg, -1); 2270 } 2271 2272 // Spill area dump 2273 reg = OptoReg::add(_matcher._new_SP, _framesize ); 2274 while( reg > _matcher._out_arg_limit ) { 2275 reg = OptoReg::add(reg, -1); 2276 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 2277 } 2278 2279 // Outgoing argument area dump 2280 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 2281 reg = OptoReg::add(reg, -1); 2282 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 2283 } 2284 2285 // Outgoing new preserve area 2286 while( reg > _matcher._new_SP ) { 2287 reg = OptoReg::add(reg, -1); 2288 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 2289 } 2290 tty->print_cr("#"); 2291 } 2292 2293 void PhaseChaitin::dump_bb( uint pre_order ) const { 2294 tty->print_cr("---dump of B%d---",pre_order); 2295 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2296 Block* block = _cfg.get_block(i); 2297 if (block->_pre_order == pre_order) { 2298 dump(block); 2299 } 2300 } 2301 } 2302 2303 void PhaseChaitin::dump_lrg( uint lidx, bool defs_only ) const { 2304 tty->print_cr("---dump of L%d---",lidx); 2305 2306 if (_ifg) { 2307 if (lidx >= _lrg_map.max_lrg_id()) { 2308 tty->print("Attempt to print live range index beyond max live range.\n"); 2309 return; 2310 } 2311 tty->print("L%d: ",lidx); 2312 if (lidx < _ifg->_maxlrg) { 2313 lrgs(lidx).dump(); 2314 } else { 2315 tty->print_cr("new LRG"); 2316 } 2317 } 2318 if( _ifg && lidx < _ifg->_maxlrg) { 2319 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 2320 _ifg->neighbors(lidx)->dump(); 2321 tty->cr(); 2322 } 2323 // For all blocks 2324 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2325 Block* block = _cfg.get_block(i); 2326 int dump_once = 0; 2327 2328 // For all instructions 2329 for( uint j = 0; j < block->number_of_nodes(); j++ ) { 2330 Node *n = block->get_node(j); 2331 if (_lrg_map.find_const(n) == lidx) { 2332 if (!dump_once++) { 2333 tty->cr(); 2334 block->dump_head(&_cfg); 2335 } 2336 dump(n); 2337 continue; 2338 } 2339 if (!defs_only) { 2340 uint cnt = n->req(); 2341 for( uint k = 1; k < cnt; k++ ) { 2342 Node *m = n->in(k); 2343 if (!m) { 2344 continue; // be robust in the dumper 2345 } 2346 if (_lrg_map.find_const(m) == lidx) { 2347 if (!dump_once++) { 2348 tty->cr(); 2349 block->dump_head(&_cfg); 2350 } 2351 dump(n); 2352 } 2353 } 2354 } 2355 } 2356 } // End of per-block dump 2357 tty->cr(); 2358 } 2359 #endif // not PRODUCT 2360 2361 int PhaseChaitin::_final_loads = 0; 2362 int PhaseChaitin::_final_stores = 0; 2363 int PhaseChaitin::_final_memoves= 0; 2364 int PhaseChaitin::_final_copies = 0; 2365 double PhaseChaitin::_final_load_cost = 0; 2366 double PhaseChaitin::_final_store_cost = 0; 2367 double PhaseChaitin::_final_memove_cost= 0; 2368 double PhaseChaitin::_final_copy_cost = 0; 2369 int PhaseChaitin::_conserv_coalesce = 0; 2370 int PhaseChaitin::_conserv_coalesce_pair = 0; 2371 int PhaseChaitin::_conserv_coalesce_trie = 0; 2372 int PhaseChaitin::_conserv_coalesce_quad = 0; 2373 int PhaseChaitin::_post_alloc = 0; 2374 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2375 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2376 int PhaseChaitin::_used_cisc_instructions = 0; 2377 int PhaseChaitin::_unused_cisc_instructions = 0; 2378 int PhaseChaitin::_allocator_attempts = 0; 2379 int PhaseChaitin::_allocator_successes = 0; 2380 2381 #ifndef PRODUCT 2382 uint PhaseChaitin::_high_pressure = 0; 2383 uint PhaseChaitin::_low_pressure = 0; 2384 2385 void PhaseChaitin::print_chaitin_statistics() { 2386 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); 2387 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); 2388 tty->print_cr("Adjusted spill cost = %7.0f.", 2389 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2390 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2391 tty->print("Conservatively coalesced %d copies, %d pairs", 2392 _conserv_coalesce, _conserv_coalesce_pair); 2393 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2394 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2395 tty->print_cr(", %d post alloc.", _post_alloc); 2396 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2397 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2398 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2399 if( _used_cisc_instructions || _unused_cisc_instructions ) 2400 tty->print_cr("Used cisc instruction %d, remained in register %d", 2401 _used_cisc_instructions, _unused_cisc_instructions); 2402 if( _allocator_successes != 0 ) 2403 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2404 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2405 } 2406 #endif // not PRODUCT