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 , _live(0) 205 , _spilled_once(Thread::current()->resource_area()) 206 , _spilled_twice(Thread::current()->resource_area()) 207 , _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0) 208 , _oldphi(unique) 209 #ifndef PRODUCT 210 , _trace_spilling(C->directive()->TraceSpillingOption) 211 #endif 212 , _lrg_map(Thread::current()->resource_area(), unique) 213 , _scheduling_info_generated(scheduling_info_generated) 214 , _sched_int_pressure(0, INTPRESSURE) 215 , _sched_float_pressure(0, FLOATPRESSURE) 216 , _scratch_int_pressure(0, INTPRESSURE) 217 , _scratch_float_pressure(0, FLOATPRESSURE) 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(Matcher::vector_size_supported(T_BYTE,4), "sanity"); 917 assert(RegMask::num_registers(Op_VecS) == RegMask::SlotsPerVecS, "sanity"); 918 lrg.set_num_regs(RegMask::SlotsPerVecS); 919 lrg.set_reg_pressure(1); 920 break; 921 case Op_VecD: 922 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity"); 923 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity"); 924 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned"); 925 lrg.set_num_regs(RegMask::SlotsPerVecD); 926 lrg.set_reg_pressure(1); 927 break; 928 case Op_VecX: 929 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity"); 930 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity"); 931 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned"); 932 lrg.set_num_regs(RegMask::SlotsPerVecX); 933 lrg.set_reg_pressure(1); 934 break; 935 case Op_VecY: 936 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity"); 937 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity"); 938 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned"); 939 lrg.set_num_regs(RegMask::SlotsPerVecY); 940 lrg.set_reg_pressure(1); 941 break; 942 case Op_VecZ: 943 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity"); 944 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity"); 945 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned"); 946 lrg.set_num_regs(RegMask::SlotsPerVecZ); 947 lrg.set_reg_pressure(1); 948 break; 949 default: 950 ShouldNotReachHere(); 951 } 952 } 953 954 // Now do the same for inputs 955 uint cnt = n->req(); 956 // Setup for CISC SPILLING 957 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 958 if( UseCISCSpill && after_aggressive ) { 959 inp = n->cisc_operand(); 960 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 961 // Convert operand number to edge index number 962 inp = n->as_Mach()->operand_index(inp); 963 } 964 965 // Prepare register mask for each input 966 for( uint k = input_edge_start; k < cnt; k++ ) { 967 uint vreg = _lrg_map.live_range_id(n->in(k)); 968 if (!vreg) { 969 continue; 970 } 971 972 // If this instruction is CISC Spillable, add the flags 973 // bit to its appropriate input 974 if( UseCISCSpill && after_aggressive && inp == k ) { 975 #ifndef PRODUCT 976 if( TraceCISCSpill ) { 977 tty->print(" use_cisc_RegMask: "); 978 n->dump(); 979 } 980 #endif 981 n->as_Mach()->use_cisc_RegMask(); 982 } 983 984 if (is_machine_node && _scheduling_info_generated) { 985 MachNode* cur_node = n->as_Mach(); 986 // this is cleaned up by register allocation 987 if (k >= cur_node->num_opnds()) continue; 988 } 989 990 LRG &lrg = lrgs(vreg); 991 // // Testing for floating point code shape 992 // Node *test = n->in(k); 993 // if( test->is_Mach() ) { 994 // MachNode *m = test->as_Mach(); 995 // int op = m->ideal_Opcode(); 996 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 997 // int zzz = 1; 998 // } 999 // } 1000 1001 // Limit result register mask to acceptable registers. 1002 // Do not limit registers from uncommon uses before 1003 // AggressiveCoalesce. This effectively pre-virtual-splits 1004 // around uncommon uses of common defs. 1005 const RegMask &rm = n->in_RegMask(k); 1006 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { 1007 // Since we are BEFORE aggressive coalesce, leave the register 1008 // mask untrimmed by the call. This encourages more coalescing. 1009 // Later, AFTER aggressive, this live range will have to spill 1010 // but the spiller handles slow-path calls very nicely. 1011 } else { 1012 lrg.AND( rm ); 1013 } 1014 1015 // Check for bound register masks 1016 const RegMask &lrgmask = lrg.mask(); 1017 uint kreg = n->in(k)->ideal_reg(); 1018 bool is_vect = RegMask::is_vector(kreg); 1019 assert(n->in(k)->bottom_type()->isa_vect() == NULL || 1020 is_vect || kreg == Op_RegD || kreg == Op_RegL, 1021 "vector must be in vector registers"); 1022 if (lrgmask.is_bound(kreg)) 1023 lrg._is_bound = 1; 1024 1025 // If this use of a double forces a mis-aligned double, 1026 // flag as '_fat_proj' - really flag as allowing misalignment 1027 // AND changes how we count interferences. A mis-aligned 1028 // double can interfere with TWO aligned pairs, or effectively 1029 // FOUR registers! 1030 #ifdef ASSERT 1031 if (is_vect && !_scheduling_info_generated) { 1032 if (lrg.num_regs() != 0) { 1033 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned"); 1034 assert(!lrg._fat_proj, "sanity"); 1035 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity"); 1036 } else { 1037 assert(n->is_Phi(), "not all inputs processed only if Phi"); 1038 } 1039 } 1040 #endif 1041 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { 1042 lrg._fat_proj = 1; 1043 lrg._is_bound = 1; 1044 } 1045 // if the LRG is an unaligned pair, we will have to spill 1046 // so clear the LRG's register mask if it is not already spilled 1047 if (!is_vect && !n->is_SpillCopy() && 1048 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 1049 lrgmask.is_misaligned_pair()) { 1050 lrg.Clear(); 1051 } 1052 1053 // Check for maximum frequency value 1054 if (lrg._maxfreq < block->_freq) { 1055 lrg._maxfreq = block->_freq; 1056 } 1057 1058 } // End for all allocated inputs 1059 } // end for all instructions 1060 } // end for all blocks 1061 1062 // Final per-liverange setup 1063 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { 1064 LRG &lrg = lrgs(i2); 1065 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); 1066 if (lrg.num_regs() > 1 && !lrg._fat_proj) { 1067 lrg.clear_to_sets(); 1068 } 1069 lrg.compute_set_mask_size(); 1070 if (lrg.not_free()) { // Handle case where we lose from the start 1071 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 1072 lrg._direct_conflict = 1; 1073 } 1074 lrg.set_degree(0); // no neighbors in IFG yet 1075 } 1076 } 1077 1078 // Set the was-lo-degree bit. Conservative coalescing should not change the 1079 // colorability of the graph. If any live range was of low-degree before 1080 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 1081 // The bit is checked in Simplify. 1082 void PhaseChaitin::set_was_low() { 1083 #ifdef ASSERT 1084 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1085 int size = lrgs(i).num_regs(); 1086 uint old_was_lo = lrgs(i)._was_lo; 1087 lrgs(i)._was_lo = 0; 1088 if( lrgs(i).lo_degree() ) { 1089 lrgs(i)._was_lo = 1; // Trivially of low degree 1090 } else { // Else check the Brigg's assertion 1091 // Brigg's observation is that the lo-degree neighbors of a 1092 // hi-degree live range will not interfere with the color choices 1093 // of said hi-degree live range. The Simplify reverse-stack-coloring 1094 // order takes care of the details. Hence you do not have to count 1095 // low-degree neighbors when determining if this guy colors. 1096 int briggs_degree = 0; 1097 IndexSet *s = _ifg->neighbors(i); 1098 IndexSetIterator elements(s); 1099 uint lidx; 1100 while((lidx = elements.next()) != 0) { 1101 if( !lrgs(lidx).lo_degree() ) 1102 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 1103 } 1104 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 1105 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 1106 } 1107 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 1108 } 1109 #endif 1110 } 1111 1112 #define REGISTER_CONSTRAINED 16 1113 1114 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 1115 void PhaseChaitin::cache_lrg_info( ) { 1116 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]); 1117 1118 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1119 LRG &lrg = lrgs(i); 1120 1121 // Check for being of low degree: means we can be trivially colored. 1122 // Low degree, dead or must-spill guys just get to simplify right away 1123 if( lrg.lo_degree() || 1124 !lrg.alive() || 1125 lrg._must_spill ) { 1126 // Split low degree list into those guys that must get a 1127 // register and those that can go to register or stack. 1128 // The idea is LRGs that can go register or stack color first when 1129 // they have a good chance of getting a register. The register-only 1130 // lo-degree live ranges always get a register. 1131 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 1132 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 1133 lrg._next = _lo_stk_degree; 1134 _lo_stk_degree = i; 1135 } else { 1136 lrg._next = _lo_degree; 1137 _lo_degree = i; 1138 } 1139 } else { // Else high degree 1140 lrgs(_hi_degree)._prev = i; 1141 lrg._next = _hi_degree; 1142 lrg._prev = 0; 1143 _hi_degree = i; 1144 } 1145 } 1146 } 1147 1148 // Simplify the IFG by removing LRGs of low degree that have NO copies 1149 void PhaseChaitin::Pre_Simplify( ) { 1150 1151 // Warm up the lo-degree no-copy list 1152 int lo_no_copy = 0; 1153 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1154 if ((lrgs(i).lo_degree() && !lrgs(i)._has_copy) || 1155 !lrgs(i).alive() || 1156 lrgs(i)._must_spill) { 1157 lrgs(i)._next = lo_no_copy; 1158 lo_no_copy = i; 1159 } 1160 } 1161 1162 while( lo_no_copy ) { 1163 uint lo = lo_no_copy; 1164 lo_no_copy = lrgs(lo)._next; 1165 int size = lrgs(lo).num_regs(); 1166 1167 // Put the simplified guy on the simplified list. 1168 lrgs(lo)._next = _simplified; 1169 _simplified = lo; 1170 1171 // Yank this guy from the IFG. 1172 IndexSet *adj = _ifg->remove_node( lo ); 1173 1174 // If any neighbors' degrees fall below their number of 1175 // allowed registers, then put that neighbor on the low degree 1176 // list. Note that 'degree' can only fall and 'numregs' is 1177 // unchanged by this action. Thus the two are equal at most once, 1178 // so LRGs hit the lo-degree worklists at most once. 1179 IndexSetIterator elements(adj); 1180 uint neighbor; 1181 while ((neighbor = elements.next()) != 0) { 1182 LRG *n = &lrgs(neighbor); 1183 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1184 1185 // Check for just becoming of-low-degree 1186 if( n->just_lo_degree() && !n->_has_copy ) { 1187 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1188 // Put on lo-degree list 1189 n->_next = lo_no_copy; 1190 lo_no_copy = neighbor; 1191 } 1192 } 1193 } // End of while lo-degree no_copy worklist not empty 1194 1195 // No more lo-degree no-copy live ranges to simplify 1196 } 1197 1198 // Simplify the IFG by removing LRGs of low degree. 1199 void PhaseChaitin::Simplify( ) { 1200 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]); 1201 1202 while( 1 ) { // Repeat till simplified it all 1203 // May want to explore simplifying lo_degree before _lo_stk_degree. 1204 // This might result in more spills coloring into registers during 1205 // Select(). 1206 while( _lo_degree || _lo_stk_degree ) { 1207 // If possible, pull from lo_stk first 1208 uint lo; 1209 if( _lo_degree ) { 1210 lo = _lo_degree; 1211 _lo_degree = lrgs(lo)._next; 1212 } else { 1213 lo = _lo_stk_degree; 1214 _lo_stk_degree = lrgs(lo)._next; 1215 } 1216 1217 // Put the simplified guy on the simplified list. 1218 lrgs(lo)._next = _simplified; 1219 _simplified = lo; 1220 // If this guy is "at risk" then mark his current neighbors 1221 if( lrgs(lo)._at_risk ) { 1222 IndexSetIterator elements(_ifg->neighbors(lo)); 1223 uint datum; 1224 while ((datum = elements.next()) != 0) { 1225 lrgs(datum)._risk_bias = lo; 1226 } 1227 } 1228 1229 // Yank this guy from the IFG. 1230 IndexSet *adj = _ifg->remove_node( lo ); 1231 1232 // If any neighbors' degrees fall below their number of 1233 // allowed registers, then put that neighbor on the low degree 1234 // list. Note that 'degree' can only fall and 'numregs' is 1235 // unchanged by this action. Thus the two are equal at most once, 1236 // so LRGs hit the lo-degree worklist at most once. 1237 IndexSetIterator elements(adj); 1238 uint neighbor; 1239 while ((neighbor = elements.next()) != 0) { 1240 LRG *n = &lrgs(neighbor); 1241 #ifdef ASSERT 1242 if( VerifyOpto || VerifyRegisterAllocator ) { 1243 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1244 } 1245 #endif 1246 1247 // Check for just becoming of-low-degree just counting registers. 1248 // _must_spill live ranges are already on the low degree list. 1249 if( n->just_lo_degree() && !n->_must_spill ) { 1250 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1251 // Pull from hi-degree list 1252 uint prev = n->_prev; 1253 uint next = n->_next; 1254 if( prev ) lrgs(prev)._next = next; 1255 else _hi_degree = next; 1256 lrgs(next)._prev = prev; 1257 n->_next = _lo_degree; 1258 _lo_degree = neighbor; 1259 } 1260 } 1261 } // End of while lo-degree/lo_stk_degree worklist not empty 1262 1263 // Check for got everything: is hi-degree list empty? 1264 if( !_hi_degree ) break; 1265 1266 // Time to pick a potential spill guy 1267 uint lo_score = _hi_degree; 1268 double score = lrgs(lo_score).score(); 1269 double area = lrgs(lo_score)._area; 1270 double cost = lrgs(lo_score)._cost; 1271 bool bound = lrgs(lo_score)._is_bound; 1272 1273 // Find cheapest guy 1274 debug_only( int lo_no_simplify=0; ); 1275 for( uint i = _hi_degree; i; i = lrgs(i)._next ) { 1276 assert( !(*_ifg->_yanked)[i], "" ); 1277 // It's just vaguely possible to move hi-degree to lo-degree without 1278 // going through a just-lo-degree stage: If you remove a double from 1279 // a float live range it's degree will drop by 2 and you can skip the 1280 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 1281 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1282 if( lrgs(i).lo_degree() ) { 1283 lo_score = i; 1284 break; 1285 } 1286 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1287 double iscore = lrgs(i).score(); 1288 double iarea = lrgs(i)._area; 1289 double icost = lrgs(i)._cost; 1290 bool ibound = lrgs(i)._is_bound; 1291 1292 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1293 // wins. Ties happen because all live ranges in question have spilled 1294 // a few times before and the spill-score adds a huge number which 1295 // washes out the low order bits. We are choosing the lesser of 2 1296 // evils; in this case pick largest area to spill. 1297 // Ties also happen when live ranges are defined and used only inside 1298 // one block. In which case their area is 0 and score set to max. 1299 // In such case choose bound live range over unbound to free registers 1300 // or with smaller cost to spill. 1301 if( iscore < score || 1302 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || 1303 (iscore == score && iarea == area && 1304 ( (ibound && !bound) || (ibound == bound && (icost < cost)) )) ) { 1305 lo_score = i; 1306 score = iscore; 1307 area = iarea; 1308 cost = icost; 1309 bound = ibound; 1310 } 1311 } 1312 LRG *lo_lrg = &lrgs(lo_score); 1313 // The live range we choose for spilling is either hi-degree, or very 1314 // rarely it can be low-degree. If we choose a hi-degree live range 1315 // there better not be any lo-degree choices. 1316 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1317 1318 // Pull from hi-degree list 1319 uint prev = lo_lrg->_prev; 1320 uint next = lo_lrg->_next; 1321 if( prev ) lrgs(prev)._next = next; 1322 else _hi_degree = next; 1323 lrgs(next)._prev = prev; 1324 // Jam him on the lo-degree list, despite his high degree. 1325 // Maybe he'll get a color, and maybe he'll spill. 1326 // Only Select() will know. 1327 lrgs(lo_score)._at_risk = true; 1328 _lo_degree = lo_score; 1329 lo_lrg->_next = 0; 1330 1331 } // End of while not simplified everything 1332 1333 } 1334 1335 // Is 'reg' register legal for 'lrg'? 1336 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { 1337 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && 1338 lrg.mask().Member(OptoReg::add(reg,-chunk))) { 1339 // RA uses OptoReg which represent the highest element of a registers set. 1340 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set 1341 // in which XMMd is used by RA to represent such vectors. A double value 1342 // uses [XMM,XMMb] pairs and XMMb is used by RA for it. 1343 // The register mask uses largest bits set of overlapping register sets. 1344 // On x86 with AVX it uses 8 bits for each XMM registers set. 1345 // 1346 // The 'lrg' already has cleared-to-set register mask (done in Select() 1347 // before calling choose_color()). Passing mask.Member(reg) check above 1348 // indicates that the size (num_regs) of 'reg' set is less or equal to 1349 // 'lrg' set size. 1350 // For set size 1 any register which is member of 'lrg' mask is legal. 1351 if (lrg.num_regs()==1) 1352 return true; 1353 // For larger sets only an aligned register with the same set size is legal. 1354 int mask = lrg.num_regs()-1; 1355 if ((reg&mask) == mask) 1356 return true; 1357 } 1358 return false; 1359 } 1360 1361 // Choose a color using the biasing heuristic 1362 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1363 1364 // Check for "at_risk" LRG's 1365 uint risk_lrg = _lrg_map.find(lrg._risk_bias); 1366 if( risk_lrg != 0 ) { 1367 // Walk the colored neighbors of the "at_risk" candidate 1368 // Choose a color which is both legal and already taken by a neighbor 1369 // of the "at_risk" candidate in order to improve the chances of the 1370 // "at_risk" candidate of coloring 1371 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1372 uint datum; 1373 while ((datum = elements.next()) != 0) { 1374 OptoReg::Name reg = lrgs(datum).reg(); 1375 // If this LRG's register is legal for us, choose it 1376 if (is_legal_reg(lrg, reg, chunk)) 1377 return reg; 1378 } 1379 } 1380 1381 uint copy_lrg = _lrg_map.find(lrg._copy_bias); 1382 if( copy_lrg != 0 ) { 1383 // If he has a color, 1384 if( !(*(_ifg->_yanked))[copy_lrg] ) { 1385 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1386 // And it is legal for you, 1387 if (is_legal_reg(lrg, reg, chunk)) 1388 return reg; 1389 } else if( chunk == 0 ) { 1390 // Choose a color which is legal for him 1391 RegMask tempmask = lrg.mask(); 1392 tempmask.AND(lrgs(copy_lrg).mask()); 1393 tempmask.clear_to_sets(lrg.num_regs()); 1394 OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); 1395 if (OptoReg::is_valid(reg)) 1396 return reg; 1397 } 1398 } 1399 1400 // If no bias info exists, just go with the register selection ordering 1401 if (lrg._is_vector || lrg.num_regs() == 2) { 1402 // Find an aligned set 1403 return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); 1404 } 1405 1406 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1407 // copy removal to remove many more copies, by preventing a just-assigned 1408 // register from being repeatedly assigned. 1409 OptoReg::Name reg = lrg.mask().find_first_elem(); 1410 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1411 // This 'Remove; find; Insert' idiom is an expensive way to find the 1412 // SECOND element in the mask. 1413 lrg.Remove(reg); 1414 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1415 lrg.Insert(reg); 1416 if( OptoReg::is_reg(reg2)) 1417 reg = reg2; 1418 } 1419 return OptoReg::add( reg, chunk ); 1420 } 1421 1422 // Choose a color in the current chunk 1423 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1424 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)"); 1425 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)"); 1426 1427 if( lrg.num_regs() == 1 || // Common Case 1428 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1429 // Use a heuristic to "bias" the color choice 1430 return bias_color(lrg, chunk); 1431 1432 assert(!lrg._is_vector, "should be not vector here" ); 1433 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1434 1435 // Fat-proj case or misaligned double argument. 1436 assert(lrg.compute_mask_size() == lrg.num_regs() || 1437 lrg.num_regs() == 2,"fat projs exactly color" ); 1438 assert( !chunk, "always color in 1st chunk" ); 1439 // Return the highest element in the set. 1440 return lrg.mask().find_last_elem(); 1441 } 1442 1443 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1444 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1445 // everything going back is guaranteed a color. Select that color. If some 1446 // hi-degree LRG cannot get a color then we record that we must spill. 1447 uint PhaseChaitin::Select( ) { 1448 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]); 1449 1450 uint spill_reg = LRG::SPILL_REG; 1451 _max_reg = OptoReg::Name(0); // Past max register used 1452 while( _simplified ) { 1453 // Pull next LRG from the simplified list - in reverse order of removal 1454 uint lidx = _simplified; 1455 LRG *lrg = &lrgs(lidx); 1456 _simplified = lrg->_next; 1457 1458 1459 #ifndef PRODUCT 1460 if (trace_spilling()) { 1461 ttyLocker ttyl; 1462 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1463 lrg->degrees_of_freedom()); 1464 lrg->dump(); 1465 } 1466 #endif 1467 1468 // Re-insert into the IFG 1469 _ifg->re_insert(lidx); 1470 if( !lrg->alive() ) continue; 1471 // capture allstackedness flag before mask is hacked 1472 const int is_allstack = lrg->mask().is_AllStack(); 1473 1474 // Yeah, yeah, yeah, I know, I know. I can refactor this 1475 // to avoid the GOTO, although the refactored code will not 1476 // be much clearer. We arrive here IFF we have a stack-based 1477 // live range that cannot color in the current chunk, and it 1478 // has to move into the next free stack chunk. 1479 int chunk = 0; // Current chunk is first chunk 1480 retry_next_chunk: 1481 1482 // Remove neighbor colors 1483 IndexSet *s = _ifg->neighbors(lidx); 1484 1485 debug_only(RegMask orig_mask = lrg->mask();) 1486 IndexSetIterator elements(s); 1487 uint neighbor; 1488 while ((neighbor = elements.next()) != 0) { 1489 // Note that neighbor might be a spill_reg. In this case, exclusion 1490 // of its color will be a no-op, since the spill_reg chunk is in outer 1491 // space. Also, if neighbor is in a different chunk, this exclusion 1492 // will be a no-op. (Later on, if lrg runs out of possible colors in 1493 // its chunk, a new chunk of color may be tried, in which case 1494 // examination of neighbors is started again, at retry_next_chunk.) 1495 LRG &nlrg = lrgs(neighbor); 1496 OptoReg::Name nreg = nlrg.reg(); 1497 // Only subtract masks in the same chunk 1498 if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { 1499 #ifndef PRODUCT 1500 uint size = lrg->mask().Size(); 1501 RegMask rm = lrg->mask(); 1502 #endif 1503 lrg->SUBTRACT(nlrg.mask()); 1504 #ifndef PRODUCT 1505 if (trace_spilling() && lrg->mask().Size() != size) { 1506 ttyLocker ttyl; 1507 tty->print("L%d ", lidx); 1508 rm.dump(); 1509 tty->print(" intersected L%d ", neighbor); 1510 nlrg.mask().dump(); 1511 tty->print(" removed "); 1512 rm.SUBTRACT(lrg->mask()); 1513 rm.dump(); 1514 tty->print(" leaving "); 1515 lrg->mask().dump(); 1516 tty->cr(); 1517 } 1518 #endif 1519 } 1520 } 1521 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1522 // Aligned pairs need aligned masks 1523 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1524 if (lrg->num_regs() > 1 && !lrg->_fat_proj) { 1525 lrg->clear_to_sets(); 1526 } 1527 1528 // Check if a color is available and if so pick the color 1529 OptoReg::Name reg = choose_color( *lrg, chunk ); 1530 #ifdef SPARC 1531 debug_only(lrg->compute_set_mask_size()); 1532 assert(lrg->num_regs() < 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); 1533 #endif 1534 1535 //--------------- 1536 // If we fail to color and the AllStack flag is set, trigger 1537 // a chunk-rollover event 1538 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1539 // Bump register mask up to next stack chunk 1540 chunk += RegMask::CHUNK_SIZE; 1541 lrg->Set_All(); 1542 1543 goto retry_next_chunk; 1544 } 1545 1546 //--------------- 1547 // Did we get a color? 1548 else if( OptoReg::is_valid(reg)) { 1549 #ifndef PRODUCT 1550 RegMask avail_rm = lrg->mask(); 1551 #endif 1552 1553 // Record selected register 1554 lrg->set_reg(reg); 1555 1556 if( reg >= _max_reg ) // Compute max register limit 1557 _max_reg = OptoReg::add(reg,1); 1558 // Fold reg back into normal space 1559 reg = OptoReg::add(reg,-chunk); 1560 1561 // If the live range is not bound, then we actually had some choices 1562 // to make. In this case, the mask has more bits in it than the colors 1563 // chosen. Restrict the mask to just what was picked. 1564 int n_regs = lrg->num_regs(); 1565 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1566 if (n_regs == 1 || !lrg->_fat_proj) { 1567 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity"); 1568 lrg->Clear(); // Clear the mask 1569 lrg->Insert(reg); // Set regmask to match selected reg 1570 // For vectors and pairs, also insert the low bit of the pair 1571 for (int i = 1; i < n_regs; i++) 1572 lrg->Insert(OptoReg::add(reg,-i)); 1573 lrg->set_mask_size(n_regs); 1574 } else { // Else fatproj 1575 // mask must be equal to fatproj bits, by definition 1576 } 1577 #ifndef PRODUCT 1578 if (trace_spilling()) { 1579 ttyLocker ttyl; 1580 tty->print("L%d selected ", lidx); 1581 lrg->mask().dump(); 1582 tty->print(" from "); 1583 avail_rm.dump(); 1584 tty->cr(); 1585 } 1586 #endif 1587 // Note that reg is the highest-numbered register in the newly-bound mask. 1588 } // end color available case 1589 1590 //--------------- 1591 // Live range is live and no colors available 1592 else { 1593 assert( lrg->alive(), "" ); 1594 assert( !lrg->_fat_proj || lrg->is_multidef() || 1595 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1596 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1597 1598 // Assign the special spillreg register 1599 lrg->set_reg(OptoReg::Name(spill_reg++)); 1600 // Do not empty the regmask; leave mask_size lying around 1601 // for use during Spilling 1602 #ifndef PRODUCT 1603 if( trace_spilling() ) { 1604 ttyLocker ttyl; 1605 tty->print("L%d spilling with neighbors: ", lidx); 1606 s->dump(); 1607 debug_only(tty->print(" original mask: ")); 1608 debug_only(orig_mask.dump()); 1609 dump_lrg(lidx); 1610 } 1611 #endif 1612 } // end spill case 1613 1614 } 1615 1616 return spill_reg-LRG::SPILL_REG; // Return number of spills 1617 } 1618 1619 // Copy 'was_spilled'-edness from the source Node to the dst Node. 1620 void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { 1621 if( _spilled_once.test(src->_idx) ) { 1622 _spilled_once.set(dst->_idx); 1623 lrgs(_lrg_map.find(dst))._was_spilled1 = 1; 1624 if( _spilled_twice.test(src->_idx) ) { 1625 _spilled_twice.set(dst->_idx); 1626 lrgs(_lrg_map.find(dst))._was_spilled2 = 1; 1627 } 1628 } 1629 } 1630 1631 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1632 void PhaseChaitin::set_was_spilled( Node *n ) { 1633 if( _spilled_once.test_set(n->_idx) ) 1634 _spilled_twice.set(n->_idx); 1635 } 1636 1637 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1638 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1639 void PhaseChaitin::fixup_spills() { 1640 // This function does only cisc spill work. 1641 if( !UseCISCSpill ) return; 1642 1643 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]); 1644 1645 // Grab the Frame Pointer 1646 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); 1647 1648 // For all blocks 1649 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1650 Block* block = _cfg.get_block(i); 1651 1652 // For all instructions in block 1653 uint last_inst = block->end_idx(); 1654 for (uint j = 1; j <= last_inst; j++) { 1655 Node* n = block->get_node(j); 1656 1657 // Dead instruction??? 1658 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1659 C->top() == n || // Or the random TOP node 1660 n->is_Proj(), // Or a fat-proj kill node 1661 "No dead instructions after post-alloc" ); 1662 1663 int inp = n->cisc_operand(); 1664 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1665 // Convert operand number to edge index number 1666 MachNode *mach = n->as_Mach(); 1667 inp = mach->operand_index(inp); 1668 Node *src = n->in(inp); // Value to load or store 1669 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); 1670 OptoReg::Name src_reg = lrg_cisc.reg(); 1671 // Doubles record the HIGH register of an adjacent pair. 1672 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1673 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1674 // This is a CISC Spill, get stack offset and construct new node 1675 #ifndef PRODUCT 1676 if( TraceCISCSpill ) { 1677 tty->print(" reg-instr: "); 1678 n->dump(); 1679 } 1680 #endif 1681 int stk_offset = reg2offset(src_reg); 1682 // Bailout if we might exceed node limit when spilling this instruction 1683 C->check_node_count(0, "out of nodes fixing spills"); 1684 if (C->failing()) return; 1685 // Transform node 1686 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); 1687 cisc->set_req(inp,fp); // Base register is frame pointer 1688 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1689 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1690 cisc->ins_req(1,src); // Requires a memory edge 1691 } 1692 block->map_node(cisc, j); // Insert into basic block 1693 n->subsume_by(cisc, C); // Correct graph 1694 // 1695 ++_used_cisc_instructions; 1696 #ifndef PRODUCT 1697 if( TraceCISCSpill ) { 1698 tty->print(" cisc-instr: "); 1699 cisc->dump(); 1700 } 1701 #endif 1702 } else { 1703 #ifndef PRODUCT 1704 if( TraceCISCSpill ) { 1705 tty->print(" using reg-instr: "); 1706 n->dump(); 1707 } 1708 #endif 1709 ++_unused_cisc_instructions; // input can be on stack 1710 } 1711 } 1712 1713 } // End of for all instructions 1714 1715 } // End of for all blocks 1716 } 1717 1718 // Helper to stretch above; recursively discover the base Node for a 1719 // given derived Node. Easy for AddP-related machine nodes, but needs 1720 // to be recursive for derived Phis. 1721 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1722 // See if already computed; if so return it 1723 if( derived_base_map[derived->_idx] ) 1724 return derived_base_map[derived->_idx]; 1725 1726 // See if this happens to be a base. 1727 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1728 // pointers derived from NULL! These are always along paths that 1729 // can't happen at run-time but the optimizer cannot deduce it so 1730 // we have to handle it gracefully. 1731 assert(!derived->bottom_type()->isa_narrowoop() || 1732 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity"); 1733 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1734 // If its an OOP with a non-zero offset, then it is derived. 1735 if (tj == NULL || tj->offset() == 0) { 1736 derived_base_map[derived->_idx] = derived; 1737 return derived; 1738 } 1739 // Derived is NULL+offset? Base is NULL! 1740 if( derived->is_Con() ) { 1741 Node *base = _matcher.mach_null(); 1742 assert(base != NULL, "sanity"); 1743 if (base->in(0) == NULL) { 1744 // Initialize it once and make it shared: 1745 // set control to _root and place it into Start block 1746 // (where top() node is placed). 1747 base->init_req(0, _cfg.get_root_node()); 1748 Block *startb = _cfg.get_block_for_node(C->top()); 1749 uint node_pos = startb->find_node(C->top()); 1750 startb->insert_node(base, node_pos); 1751 _cfg.map_node_to_block(base, startb); 1752 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet"); 1753 1754 // The loadConP0 might have projection nodes depending on architecture 1755 // Add the projection nodes to the CFG 1756 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { 1757 Node* use = base->fast_out(i); 1758 if (use->is_MachProj()) { 1759 startb->insert_node(use, ++node_pos); 1760 _cfg.map_node_to_block(use, startb); 1761 new_lrg(use, maxlrg++); 1762 } 1763 } 1764 } 1765 if (_lrg_map.live_range_id(base) == 0) { 1766 new_lrg(base, maxlrg++); 1767 } 1768 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"); 1769 derived_base_map[derived->_idx] = base; 1770 return base; 1771 } 1772 1773 // Check for AddP-related opcodes 1774 if (!derived->is_Phi()) { 1775 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name()); 1776 Node *base = derived->in(AddPNode::Base); 1777 derived_base_map[derived->_idx] = base; 1778 return base; 1779 } 1780 1781 // Recursively find bases for Phis. 1782 // First check to see if we can avoid a base Phi here. 1783 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1784 uint i; 1785 for( i = 2; i < derived->req(); i++ ) 1786 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1787 break; 1788 // Went to the end without finding any different bases? 1789 if( i == derived->req() ) { // No need for a base Phi here 1790 derived_base_map[derived->_idx] = base; 1791 return base; 1792 } 1793 1794 // Now we see we need a base-Phi here to merge the bases 1795 const Type *t = base->bottom_type(); 1796 base = new PhiNode( derived->in(0), t ); 1797 for( i = 1; i < derived->req(); i++ ) { 1798 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1799 t = t->meet(base->in(i)->bottom_type()); 1800 } 1801 base->as_Phi()->set_type(t); 1802 1803 // Search the current block for an existing base-Phi 1804 Block *b = _cfg.get_block_for_node(derived); 1805 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1806 Node *phi = b->get_node(i); 1807 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1808 b->insert_node(base, i); // Must insert created Phi here as base 1809 _cfg.map_node_to_block(base, b); 1810 new_lrg(base,maxlrg++); 1811 break; 1812 } 1813 // See if Phi matches. 1814 uint j; 1815 for( j = 1; j < base->req(); j++ ) 1816 if( phi->in(j) != base->in(j) && 1817 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1818 break; 1819 if( j == base->req() ) { // All inputs match? 1820 base = phi; // Then use existing 'phi' and drop 'base' 1821 break; 1822 } 1823 } 1824 1825 1826 // Cache info for later passes 1827 derived_base_map[derived->_idx] = base; 1828 return base; 1829 } 1830 1831 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1832 // base pointer that is live across the Safepoint for oopmap building. The 1833 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1834 // required edge set. 1835 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { 1836 int must_recompute_live = false; 1837 uint maxlrg = _lrg_map.max_lrg_id(); 1838 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1839 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1840 1841 // For all blocks in RPO do... 1842 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1843 Block* block = _cfg.get_block(i); 1844 // Note use of deep-copy constructor. I cannot hammer the original 1845 // liveout bits, because they are needed by the following coalesce pass. 1846 IndexSet liveout(_live->live(block)); 1847 1848 for (uint j = block->end_idx() + 1; j > 1; j--) { 1849 Node* n = block->get_node(j - 1); 1850 1851 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1852 // like to see in the same register. Compare uses the loop-phi and so 1853 // extends its live range BUT cannot be part of the cycle. If this 1854 // extended live range overlaps with the update of the loop-phi value 1855 // we need both alive at the same time -- which requires at least 1 1856 // copy. But because Intel has only 2-address registers we end up with 1857 // at least 2 copies, one before the loop-phi update instruction and 1858 // one after. Instead we split the input to the compare just after the 1859 // phi. 1860 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1861 Node *phi = n->in(1); 1862 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1863 Block *phi_block = _cfg.get_block_for_node(phi); 1864 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { 1865 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1866 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); 1867 insert_proj( phi_block, 1, spill, maxlrg++ ); 1868 n->set_req(1,spill); 1869 must_recompute_live = true; 1870 } 1871 } 1872 } 1873 1874 // Get value being defined 1875 uint lidx = _lrg_map.live_range_id(n); 1876 // Ignore the occasional brand-new live range 1877 if (lidx && lidx < _lrg_map.max_lrg_id()) { 1878 // Remove from live-out set 1879 liveout.remove(lidx); 1880 1881 // Copies do not define a new value and so do not interfere. 1882 // Remove the copies source from the liveout set before interfering. 1883 uint idx = n->is_Copy(); 1884 if (idx) { 1885 liveout.remove(_lrg_map.live_range_id(n->in(idx))); 1886 } 1887 } 1888 1889 // Found a safepoint? 1890 JVMState *jvms = n->jvms(); 1891 if( jvms ) { 1892 // Now scan for a live derived pointer 1893 IndexSetIterator elements(&liveout); 1894 uint neighbor; 1895 while ((neighbor = elements.next()) != 0) { 1896 // Find reaching DEF for base and derived values 1897 // This works because we are still in SSA during this call. 1898 Node *derived = lrgs(neighbor)._def; 1899 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1900 assert(!derived->bottom_type()->isa_narrowoop() || 1901 derived->bottom_type()->make_ptr()->is_ptr()->offset() == 0, "sanity"); 1902 // If its an OOP with a non-zero offset, then it is derived. 1903 if (tj && tj->offset() != 0 && tj->isa_oop_ptr()) { 1904 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); 1905 assert(base->_idx < _lrg_map.size(), ""); 1906 // Add reaching DEFs of derived pointer and base pointer as a 1907 // pair of inputs 1908 n->add_req(derived); 1909 n->add_req(base); 1910 1911 // See if the base pointer is already live to this point. 1912 // Since I'm working on the SSA form, live-ness amounts to 1913 // reaching def's. So if I find the base's live range then 1914 // I know the base's def reaches here. 1915 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or 1916 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND 1917 (_lrg_map.live_range_id(base) > 0) && // not a constant 1918 _cfg.get_block_for_node(base) != block) { // base not def'd in blk) 1919 // Base pointer is not currently live. Since I stretched 1920 // the base pointer to here and it crosses basic-block 1921 // boundaries, the global live info is now incorrect. 1922 // Recompute live. 1923 must_recompute_live = true; 1924 } // End of if base pointer is not live to debug info 1925 } 1926 } // End of scan all live data for derived ptrs crossing GC point 1927 } // End of if found a GC point 1928 1929 // Make all inputs live 1930 if (!n->is_Phi()) { // Phi function uses come from prior block 1931 for (uint k = 1; k < n->req(); k++) { 1932 uint lidx = _lrg_map.live_range_id(n->in(k)); 1933 if (lidx < _lrg_map.max_lrg_id()) { 1934 liveout.insert(lidx); 1935 } 1936 } 1937 } 1938 1939 } // End of forall instructions in block 1940 liveout.clear(); // Free the memory used by liveout. 1941 1942 } // End of forall blocks 1943 _lrg_map.set_max_lrg_id(maxlrg); 1944 1945 // If I created a new live range I need to recompute live 1946 if (maxlrg != _ifg->_maxlrg) { 1947 must_recompute_live = true; 1948 } 1949 1950 return must_recompute_live != 0; 1951 } 1952 1953 // Extend the node to LRG mapping 1954 1955 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { 1956 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); 1957 } 1958 1959 #ifndef PRODUCT 1960 void PhaseChaitin::dump(const Node *n) const { 1961 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; 1962 tty->print("L%d",r); 1963 if (r && n->Opcode() != Op_Phi) { 1964 if( _node_regs ) { // Got a post-allocation copy of allocation? 1965 tty->print("["); 1966 OptoReg::Name second = get_reg_second(n); 1967 if( OptoReg::is_valid(second) ) { 1968 if( OptoReg::is_reg(second) ) 1969 tty->print("%s:",Matcher::regName[second]); 1970 else 1971 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1972 } 1973 OptoReg::Name first = get_reg_first(n); 1974 if( OptoReg::is_reg(first) ) 1975 tty->print("%s]",Matcher::regName[first]); 1976 else 1977 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1978 } else 1979 n->out_RegMask().dump(); 1980 } 1981 tty->print("/N%d\t",n->_idx); 1982 tty->print("%s === ", n->Name()); 1983 uint k; 1984 for (k = 0; k < n->req(); k++) { 1985 Node *m = n->in(k); 1986 if (!m) { 1987 tty->print("_ "); 1988 } 1989 else { 1990 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1991 tty->print("L%d",r); 1992 // Data MultiNode's can have projections with no real registers. 1993 // Don't die while dumping them. 1994 int op = n->Opcode(); 1995 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1996 if( _node_regs ) { 1997 tty->print("["); 1998 OptoReg::Name second = get_reg_second(n->in(k)); 1999 if( OptoReg::is_valid(second) ) { 2000 if( OptoReg::is_reg(second) ) 2001 tty->print("%s:",Matcher::regName[second]); 2002 else 2003 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 2004 reg2offset_unchecked(second)); 2005 } 2006 OptoReg::Name first = get_reg_first(n->in(k)); 2007 if( OptoReg::is_reg(first) ) 2008 tty->print("%s]",Matcher::regName[first]); 2009 else 2010 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 2011 reg2offset_unchecked(first)); 2012 } else 2013 n->in_RegMask(k).dump(); 2014 } 2015 tty->print("/N%d ",m->_idx); 2016 } 2017 } 2018 if( k < n->len() && n->in(k) ) tty->print("| "); 2019 for( ; k < n->len(); k++ ) { 2020 Node *m = n->in(k); 2021 if(!m) { 2022 break; 2023 } 2024 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 2025 tty->print("L%d",r); 2026 tty->print("/N%d ",m->_idx); 2027 } 2028 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 2029 else n->dump_spec(tty); 2030 if( _spilled_once.test(n->_idx ) ) { 2031 tty->print(" Spill_1"); 2032 if( _spilled_twice.test(n->_idx ) ) 2033 tty->print(" Spill_2"); 2034 } 2035 tty->print("\n"); 2036 } 2037 2038 void PhaseChaitin::dump(const Block *b) const { 2039 b->dump_head(&_cfg); 2040 2041 // For all instructions 2042 for( uint j = 0; j < b->number_of_nodes(); j++ ) 2043 dump(b->get_node(j)); 2044 // Print live-out info at end of block 2045 if( _live ) { 2046 tty->print("Liveout: "); 2047 IndexSet *live = _live->live(b); 2048 IndexSetIterator elements(live); 2049 tty->print("{"); 2050 uint i; 2051 while ((i = elements.next()) != 0) { 2052 tty->print("L%d ", _lrg_map.find_const(i)); 2053 } 2054 tty->print_cr("}"); 2055 } 2056 tty->print("\n"); 2057 } 2058 2059 void PhaseChaitin::dump() const { 2060 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 2061 _matcher._new_SP, _framesize ); 2062 2063 // For all blocks 2064 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2065 dump(_cfg.get_block(i)); 2066 } 2067 // End of per-block dump 2068 tty->print("\n"); 2069 2070 if (!_ifg) { 2071 tty->print("(No IFG.)\n"); 2072 return; 2073 } 2074 2075 // Dump LRG array 2076 tty->print("--- Live RanGe Array ---\n"); 2077 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { 2078 tty->print("L%d: ",i2); 2079 if (i2 < _ifg->_maxlrg) { 2080 lrgs(i2).dump(); 2081 } 2082 else { 2083 tty->print_cr("new LRG"); 2084 } 2085 } 2086 tty->cr(); 2087 2088 // Dump lo-degree list 2089 tty->print("Lo degree: "); 2090 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 2091 tty->print("L%d ",i3); 2092 tty->cr(); 2093 2094 // Dump lo-stk-degree list 2095 tty->print("Lo stk degree: "); 2096 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 2097 tty->print("L%d ",i4); 2098 tty->cr(); 2099 2100 // Dump lo-degree list 2101 tty->print("Hi degree: "); 2102 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 2103 tty->print("L%d ",i5); 2104 tty->cr(); 2105 } 2106 2107 void PhaseChaitin::dump_degree_lists() const { 2108 // Dump lo-degree list 2109 tty->print("Lo degree: "); 2110 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 2111 tty->print("L%d ",i); 2112 tty->cr(); 2113 2114 // Dump lo-stk-degree list 2115 tty->print("Lo stk degree: "); 2116 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 2117 tty->print("L%d ",i2); 2118 tty->cr(); 2119 2120 // Dump lo-degree list 2121 tty->print("Hi degree: "); 2122 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 2123 tty->print("L%d ",i3); 2124 tty->cr(); 2125 } 2126 2127 void PhaseChaitin::dump_simplified() const { 2128 tty->print("Simplified: "); 2129 for( uint i = _simplified; i; i = lrgs(i)._next ) 2130 tty->print("L%d ",i); 2131 tty->cr(); 2132 } 2133 2134 static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { 2135 if ((int)reg < 0) 2136 sprintf(buf, "<OptoReg::%d>", (int)reg); 2137 else if (OptoReg::is_reg(reg)) 2138 strcpy(buf, Matcher::regName[reg]); 2139 else 2140 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 2141 pc->reg2offset(reg)); 2142 return buf+strlen(buf); 2143 } 2144 2145 // Dump a register name into a buffer. Be intelligent if we get called 2146 // before allocation is complete. 2147 char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { 2148 if( this == NULL ) { // Not got anything? 2149 sprintf(buf,"N%d",n->_idx); // Then use Node index 2150 } else if( _node_regs ) { 2151 // Post allocation, use direct mappings, no LRG info available 2152 print_reg( get_reg_first(n), this, buf ); 2153 } else { 2154 uint lidx = _lrg_map.find_const(n); // Grab LRG number 2155 if( !_ifg ) { 2156 sprintf(buf,"L%d",lidx); // No register binding yet 2157 } else if( !lidx ) { // Special, not allocated value 2158 strcpy(buf,"Special"); 2159 } else { 2160 if (lrgs(lidx)._is_vector) { 2161 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) 2162 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register 2163 else 2164 sprintf(buf,"L%d",lidx); // No register binding yet 2165 } else if( (lrgs(lidx).num_regs() == 1) 2166 ? lrgs(lidx).mask().is_bound1() 2167 : lrgs(lidx).mask().is_bound_pair() ) { 2168 // Hah! We have a bound machine register 2169 print_reg( lrgs(lidx).reg(), this, buf ); 2170 } else { 2171 sprintf(buf,"L%d",lidx); // No register binding yet 2172 } 2173 } 2174 } 2175 return buf+strlen(buf); 2176 } 2177 2178 void PhaseChaitin::dump_for_spill_split_recycle() const { 2179 if( WizardMode && (PrintCompilation || PrintOpto) ) { 2180 // Display which live ranges need to be split and the allocator's state 2181 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 2182 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { 2183 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 2184 tty->print("L%d: ", bidx); 2185 lrgs(bidx).dump(); 2186 } 2187 } 2188 tty->cr(); 2189 dump(); 2190 } 2191 } 2192 2193 void PhaseChaitin::dump_frame() const { 2194 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 2195 const TypeTuple *domain = C->tf()->domain_cc(); 2196 const int argcnt = domain->cnt() - TypeFunc::Parms; 2197 2198 // Incoming arguments in registers dump 2199 for( int k = 0; k < argcnt; k++ ) { 2200 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 2201 if( OptoReg::is_reg(parmreg)) { 2202 const char *reg_name = OptoReg::regname(parmreg); 2203 tty->print("#r%3.3d %s", parmreg, reg_name); 2204 parmreg = _matcher._parm_regs[k].second(); 2205 if( OptoReg::is_reg(parmreg)) { 2206 tty->print(":%s", OptoReg::regname(parmreg)); 2207 } 2208 tty->print(" : parm %d: ", k); 2209 domain->field_at(k + TypeFunc::Parms)->dump(); 2210 tty->cr(); 2211 } 2212 } 2213 2214 // Check for un-owned padding above incoming args 2215 OptoReg::Name reg = _matcher._new_SP; 2216 if( reg > _matcher._in_arg_limit ) { 2217 reg = OptoReg::add(reg, -1); 2218 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 2219 } 2220 2221 // Incoming argument area dump 2222 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 2223 while( reg > begin_in_arg ) { 2224 reg = OptoReg::add(reg, -1); 2225 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2226 int j; 2227 for( j = 0; j < argcnt; j++) { 2228 if( _matcher._parm_regs[j].first() == reg || 2229 _matcher._parm_regs[j].second() == reg ) { 2230 tty->print("parm %d: ",j); 2231 domain->field_at(j + TypeFunc::Parms)->dump(); 2232 if (j == C->get_res_entry()._offset) { 2233 tty->print(" [RESERVED] "); 2234 } 2235 tty->cr(); 2236 break; 2237 } 2238 } 2239 if( j >= argcnt ) 2240 tty->print_cr("HOLE, owned by SELF"); 2241 } 2242 2243 // Old outgoing preserve area 2244 while( reg > _matcher._old_SP ) { 2245 reg = OptoReg::add(reg, -1); 2246 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 2247 } 2248 2249 // Old SP 2250 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 2251 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 2252 2253 // Preserve area dump 2254 int fixed_slots = C->fixed_slots(); 2255 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); 2256 OptoReg::Name return_addr = _matcher.return_addr(); 2257 2258 reg = OptoReg::add(reg, -1); 2259 while (OptoReg::is_stack(reg)) { 2260 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2261 if (return_addr == reg) { 2262 tty->print_cr("return address"); 2263 } else if (reg >= begin_in_preserve) { 2264 // Preserved slots are present on x86 2265 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) 2266 tty->print_cr("saved fp register"); 2267 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && 2268 VerifyStackAtCalls) 2269 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 2270 else 2271 tty->print_cr("in_preserve"); 2272 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { 2273 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 2274 } else { 2275 tty->print_cr("pad2, stack alignment"); 2276 } 2277 reg = OptoReg::add(reg, -1); 2278 } 2279 2280 // Spill area dump 2281 reg = OptoReg::add(_matcher._new_SP, _framesize ); 2282 while( reg > _matcher._out_arg_limit ) { 2283 reg = OptoReg::add(reg, -1); 2284 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 2285 } 2286 2287 // Outgoing argument area dump 2288 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 2289 reg = OptoReg::add(reg, -1); 2290 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 2291 } 2292 2293 // Outgoing new preserve area 2294 while( reg > _matcher._new_SP ) { 2295 reg = OptoReg::add(reg, -1); 2296 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 2297 } 2298 tty->print_cr("#"); 2299 } 2300 2301 void PhaseChaitin::dump_bb( uint pre_order ) const { 2302 tty->print_cr("---dump of B%d---",pre_order); 2303 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2304 Block* block = _cfg.get_block(i); 2305 if (block->_pre_order == pre_order) { 2306 dump(block); 2307 } 2308 } 2309 } 2310 2311 void PhaseChaitin::dump_lrg( uint lidx, bool defs_only ) const { 2312 tty->print_cr("---dump of L%d---",lidx); 2313 2314 if (_ifg) { 2315 if (lidx >= _lrg_map.max_lrg_id()) { 2316 tty->print("Attempt to print live range index beyond max live range.\n"); 2317 return; 2318 } 2319 tty->print("L%d: ",lidx); 2320 if (lidx < _ifg->_maxlrg) { 2321 lrgs(lidx).dump(); 2322 } else { 2323 tty->print_cr("new LRG"); 2324 } 2325 } 2326 if( _ifg && lidx < _ifg->_maxlrg) { 2327 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 2328 _ifg->neighbors(lidx)->dump(); 2329 tty->cr(); 2330 } 2331 // For all blocks 2332 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2333 Block* block = _cfg.get_block(i); 2334 int dump_once = 0; 2335 2336 // For all instructions 2337 for( uint j = 0; j < block->number_of_nodes(); j++ ) { 2338 Node *n = block->get_node(j); 2339 if (_lrg_map.find_const(n) == lidx) { 2340 if (!dump_once++) { 2341 tty->cr(); 2342 block->dump_head(&_cfg); 2343 } 2344 dump(n); 2345 continue; 2346 } 2347 if (!defs_only) { 2348 uint cnt = n->req(); 2349 for( uint k = 1; k < cnt; k++ ) { 2350 Node *m = n->in(k); 2351 if (!m) { 2352 continue; // be robust in the dumper 2353 } 2354 if (_lrg_map.find_const(m) == lidx) { 2355 if (!dump_once++) { 2356 tty->cr(); 2357 block->dump_head(&_cfg); 2358 } 2359 dump(n); 2360 } 2361 } 2362 } 2363 } 2364 } // End of per-block dump 2365 tty->cr(); 2366 } 2367 #endif // not PRODUCT 2368 2369 int PhaseChaitin::_final_loads = 0; 2370 int PhaseChaitin::_final_stores = 0; 2371 int PhaseChaitin::_final_memoves= 0; 2372 int PhaseChaitin::_final_copies = 0; 2373 double PhaseChaitin::_final_load_cost = 0; 2374 double PhaseChaitin::_final_store_cost = 0; 2375 double PhaseChaitin::_final_memove_cost= 0; 2376 double PhaseChaitin::_final_copy_cost = 0; 2377 int PhaseChaitin::_conserv_coalesce = 0; 2378 int PhaseChaitin::_conserv_coalesce_pair = 0; 2379 int PhaseChaitin::_conserv_coalesce_trie = 0; 2380 int PhaseChaitin::_conserv_coalesce_quad = 0; 2381 int PhaseChaitin::_post_alloc = 0; 2382 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2383 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2384 int PhaseChaitin::_used_cisc_instructions = 0; 2385 int PhaseChaitin::_unused_cisc_instructions = 0; 2386 int PhaseChaitin::_allocator_attempts = 0; 2387 int PhaseChaitin::_allocator_successes = 0; 2388 2389 #ifndef PRODUCT 2390 uint PhaseChaitin::_high_pressure = 0; 2391 uint PhaseChaitin::_low_pressure = 0; 2392 2393 void PhaseChaitin::print_chaitin_statistics() { 2394 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); 2395 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); 2396 tty->print_cr("Adjusted spill cost = %7.0f.", 2397 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2398 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2399 tty->print("Conservatively coalesced %d copies, %d pairs", 2400 _conserv_coalesce, _conserv_coalesce_pair); 2401 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2402 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2403 tty->print_cr(", %d post alloc.", _post_alloc); 2404 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2405 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2406 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2407 if( _used_cisc_instructions || _unused_cisc_instructions ) 2408 tty->print_cr("Used cisc instruction %d, remained in register %d", 2409 _used_cisc_instructions, _unused_cisc_instructions); 2410 if( _allocator_successes != 0 ) 2411 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2412 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2413 } 2414 #endif // not PRODUCT