1 /* 2 * Copyright (c) 2000, 2016, 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 45 #ifndef PRODUCT 46 void LRG::dump() const { 47 ttyLocker ttyl; 48 tty->print("%d ",num_regs()); 49 _mask.dump(); 50 if( _msize_valid ) { 51 if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size); 52 else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size()); 53 } else { 54 tty->print(", #?(%d) ",_mask.Size()); 55 } 56 57 tty->print("EffDeg: "); 58 if( _degree_valid ) tty->print( "%d ", _eff_degree ); 59 else tty->print("? "); 60 61 if( is_multidef() ) { 62 tty->print("MultiDef "); 63 if (_defs != NULL) { 64 tty->print("("); 65 for (int i = 0; i < _defs->length(); i++) { 66 tty->print("N%d ", _defs->at(i)->_idx); 67 } 68 tty->print(") "); 69 } 70 } 71 else if( _def == 0 ) tty->print("Dead "); 72 else tty->print("Def: N%d ",_def->_idx); 73 74 tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score()); 75 // Flags 76 if( _is_oop ) tty->print("Oop "); 77 if( _is_float ) tty->print("Float "); 78 if( _is_vector ) tty->print("Vector "); 79 if( _was_spilled1 ) tty->print("Spilled "); 80 if( _was_spilled2 ) tty->print("Spilled2 "); 81 if( _direct_conflict ) tty->print("Direct_conflict "); 82 if( _fat_proj ) tty->print("Fat "); 83 if( _was_lo ) tty->print("Lo "); 84 if( _has_copy ) tty->print("Copy "); 85 if( _at_risk ) tty->print("Risk "); 86 87 if( _must_spill ) tty->print("Must_spill "); 88 if( _is_bound ) tty->print("Bound "); 89 if( _msize_valid ) { 90 if( _degree_valid && lo_degree() ) tty->print("Trivial "); 91 } 92 93 tty->cr(); 94 } 95 #endif 96 97 // Compute score from cost and area. Low score is best to spill. 98 static double raw_score( double cost, double area ) { 99 return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; 100 } 101 102 double LRG::score() const { 103 // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. 104 // Bigger area lowers score, encourages spilling this live range. 105 // Bigger cost raise score, prevents spilling this live range. 106 // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer 107 // to turn a divide by a constant into a multiply by the reciprical). 108 double score = raw_score( _cost, _area); 109 110 // Account for area. Basically, LRGs covering large areas are better 111 // to spill because more other LRGs get freed up. 112 if( _area == 0.0 ) // No area? Then no progress to spill 113 return 1e35; 114 115 if( _was_spilled2 ) // If spilled once before, we are unlikely 116 return score + 1e30; // to make progress again. 117 118 if( _cost >= _area*3.0 ) // Tiny area relative to cost 119 return score + 1e17; // Probably no progress to spill 120 121 if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost 122 return score + 1e10; // Likely no progress to spill 123 124 return score; 125 } 126 127 #define NUMBUCKS 3 128 129 // Straight out of Tarjan's union-find algorithm 130 uint LiveRangeMap::find_compress(uint lrg) { 131 uint cur = lrg; 132 uint next = _uf_map.at(cur); 133 while (next != cur) { // Scan chain of equivalences 134 assert( next < cur, "always union smaller"); 135 cur = next; // until find a fixed-point 136 next = _uf_map.at(cur); 137 } 138 139 // Core of union-find algorithm: update chain of 140 // equivalences to be equal to the root. 141 while (lrg != next) { 142 uint tmp = _uf_map.at(lrg); 143 _uf_map.at_put(lrg, next); 144 lrg = tmp; 145 } 146 return lrg; 147 } 148 149 // Reset the Union-Find map to identity 150 void LiveRangeMap::reset_uf_map(uint max_lrg_id) { 151 _max_lrg_id= max_lrg_id; 152 // Force the Union-Find mapping to be at least this large 153 _uf_map.at_put_grow(_max_lrg_id, 0); 154 // Initialize it to be the ID mapping. 155 for (uint i = 0; i < _max_lrg_id; ++i) { 156 _uf_map.at_put(i, i); 157 } 158 } 159 160 // Make all Nodes map directly to their final live range; no need for 161 // the Union-Find mapping after this call. 162 void LiveRangeMap::compress_uf_map_for_nodes() { 163 // For all Nodes, compress mapping 164 uint unique = _names.length(); 165 for (uint i = 0; i < unique; ++i) { 166 uint lrg = _names.at(i); 167 uint compressed_lrg = find(lrg); 168 if (lrg != compressed_lrg) { 169 _names.at_put(i, compressed_lrg); 170 } 171 } 172 } 173 174 // Like Find above, but no path compress, so bad asymptotic behavior 175 uint LiveRangeMap::find_const(uint lrg) const { 176 if (!lrg) { 177 return lrg; // Ignore the zero LRG 178 } 179 180 // Off the end? This happens during debugging dumps when you got 181 // brand new live ranges but have not told the allocator yet. 182 if (lrg >= _max_lrg_id) { 183 return lrg; 184 } 185 186 uint next = _uf_map.at(lrg); 187 while (next != lrg) { // Scan chain of equivalences 188 assert(next < lrg, "always union smaller"); 189 lrg = next; // until find a fixed-point 190 next = _uf_map.at(lrg); 191 } 192 return next; 193 } 194 195 PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher, bool scheduling_info_generated) 196 : PhaseRegAlloc(unique, cfg, matcher, 197 #ifndef PRODUCT 198 print_chaitin_statistics 199 #else 200 NULL 201 #endif 202 ) 203 , _lrg_map(Thread::current()->resource_area(), unique) 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 , _scheduling_info_generated(scheduling_info_generated) 210 , _sched_int_pressure(0, INTPRESSURE) 211 , _sched_float_pressure(0, FLOATPRESSURE) 212 , _scratch_int_pressure(0, INTPRESSURE) 213 , _scratch_float_pressure(0, FLOATPRESSURE) 214 #ifndef PRODUCT 215 , _trace_spilling(C->directive()->TraceSpillingOption) 216 #endif 217 { 218 Compile::TracePhase tp("ctorChaitin", &timers[_t_ctorChaitin]); 219 220 _high_frequency_lrg = MIN2(double(OPTO_LRG_HIGH_FREQ), _cfg.get_outer_loop_frequency()); 221 222 // Build a list of basic blocks, sorted by frequency 223 _blks = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 224 // Experiment with sorting strategies to speed compilation 225 double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket 226 Block **buckets[NUMBUCKS]; // Array of buckets 227 uint buckcnt[NUMBUCKS]; // Array of bucket counters 228 double buckval[NUMBUCKS]; // Array of bucket value cutoffs 229 for (uint i = 0; i < NUMBUCKS; i++) { 230 buckets[i] = NEW_RESOURCE_ARRAY(Block *, _cfg.number_of_blocks()); 231 buckcnt[i] = 0; 232 // Bump by three orders of magnitude each time 233 cutoff *= 0.001; 234 buckval[i] = cutoff; 235 for (uint j = 0; j < _cfg.number_of_blocks(); j++) { 236 buckets[i][j] = NULL; 237 } 238 } 239 // Sort blocks into buckets 240 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 241 for (uint j = 0; j < NUMBUCKS; j++) { 242 if ((j == NUMBUCKS - 1) || (_cfg.get_block(i)->_freq > buckval[j])) { 243 // Assign block to end of list for appropriate bucket 244 buckets[j][buckcnt[j]++] = _cfg.get_block(i); 245 break; // kick out of inner loop 246 } 247 } 248 } 249 // Dump buckets into final block array 250 uint blkcnt = 0; 251 for (uint i = 0; i < NUMBUCKS; i++) { 252 for (uint j = 0; j < buckcnt[i]; j++) { 253 _blks[blkcnt++] = buckets[i][j]; 254 } 255 } 256 257 assert(blkcnt == _cfg.number_of_blocks(), "Block array not totally filled"); 258 } 259 260 // union 2 sets together. 261 void PhaseChaitin::Union( const Node *src_n, const Node *dst_n ) { 262 uint src = _lrg_map.find(src_n); 263 uint dst = _lrg_map.find(dst_n); 264 assert(src, ""); 265 assert(dst, ""); 266 assert(src < _lrg_map.max_lrg_id(), "oob"); 267 assert(dst < _lrg_map.max_lrg_id(), "oob"); 268 assert(src < dst, "always union smaller"); 269 _lrg_map.uf_map(dst, src); 270 } 271 272 void PhaseChaitin::new_lrg(const Node *x, uint lrg) { 273 // Make the Node->LRG mapping 274 _lrg_map.extend(x->_idx,lrg); 275 // Make the Union-Find mapping an identity function 276 _lrg_map.uf_extend(lrg, lrg); 277 } 278 279 280 int PhaseChaitin::clone_projs(Block* b, uint idx, Node* orig, Node* copy, uint& max_lrg_id) { 281 assert(b->find_node(copy) == (idx - 1), "incorrect insert index for copy kill projections"); 282 DEBUG_ONLY( Block* borig = _cfg.get_block_for_node(orig); ) 283 int found_projs = 0; 284 uint cnt = orig->outcnt(); 285 for (uint i = 0; i < cnt; i++) { 286 Node* proj = orig->raw_out(i); 287 if (proj->is_MachProj()) { 288 assert(proj->outcnt() == 0, "only kill projections are expected here"); 289 assert(_cfg.get_block_for_node(proj) == borig, "incorrect block for kill projections"); 290 found_projs++; 291 // Copy kill projections after the cloned node 292 Node* kills = proj->clone(); 293 kills->set_req(0, copy); 294 b->insert_node(kills, idx++); 295 _cfg.map_node_to_block(kills, b); 296 new_lrg(kills, max_lrg_id++); 297 } 298 } 299 return found_projs; 300 } 301 302 // Renumber the live ranges to compact them. Makes the IFG smaller. 303 void PhaseChaitin::compact() { 304 Compile::TracePhase tp("chaitinCompact", &timers[_t_chaitinCompact]); 305 306 // Current the _uf_map contains a series of short chains which are headed 307 // by a self-cycle. All the chains run from big numbers to little numbers. 308 // The Find() call chases the chains & shortens them for the next Find call. 309 // We are going to change this structure slightly. Numbers above a moving 310 // wave 'i' are unchanged. Numbers below 'j' point directly to their 311 // compacted live range with no further chaining. There are no chains or 312 // cycles below 'i', so the Find call no longer works. 313 uint j=1; 314 uint i; 315 for (i = 1; i < _lrg_map.max_lrg_id(); i++) { 316 uint lr = _lrg_map.uf_live_range_id(i); 317 // Ignore unallocated live ranges 318 if (!lr) { 319 continue; 320 } 321 assert(lr <= i, ""); 322 _lrg_map.uf_map(i, ( lr == i ) ? j++ : _lrg_map.uf_live_range_id(lr)); 323 } 324 // Now change the Node->LR mapping to reflect the compacted names 325 uint unique = _lrg_map.size(); 326 for (i = 0; i < unique; i++) { 327 uint lrg_id = _lrg_map.live_range_id(i); 328 _lrg_map.map(i, _lrg_map.uf_live_range_id(lrg_id)); 329 } 330 331 // Reset the Union-Find mapping 332 _lrg_map.reset_uf_map(j); 333 } 334 335 void PhaseChaitin::Register_Allocate() { 336 337 // Above the OLD FP (and in registers) are the incoming arguments. Stack 338 // slots in this area are called "arg_slots". Above the NEW FP (and in 339 // registers) is the outgoing argument area; above that is the spill/temp 340 // area. These are all "frame_slots". Arg_slots start at the zero 341 // stack_slots and count up to the known arg_size. Frame_slots start at 342 // the stack_slot #arg_size and go up. After allocation I map stack 343 // slots to actual offsets. Stack-slots in the arg_slot area are biased 344 // by the frame_size; stack-slots in the frame_slot area are biased by 0. 345 346 _trip_cnt = 0; 347 _alternate = 0; 348 _matcher._allocation_started = true; 349 350 ResourceArea split_arena; // Arena for Split local resources 351 ResourceArea live_arena; // Arena for liveness & IFG info 352 ResourceMark rm(&live_arena); 353 354 // Need live-ness for the IFG; need the IFG for coalescing. If the 355 // liveness is JUST for coalescing, then I can get some mileage by renaming 356 // all copy-related live ranges low and then using the max copy-related 357 // live range as a cut-off for LIVE and the IFG. In other words, I can 358 // build a subset of LIVE and IFG just for copies. 359 PhaseLive live(_cfg, _lrg_map.names(), &live_arena, false); 360 361 // Need IFG for coalescing and coloring 362 PhaseIFG ifg(&live_arena); 363 _ifg = &ifg; 364 365 // Come out of SSA world to the Named world. Assign (virtual) registers to 366 // Nodes. Use the same register for all inputs and the output of PhiNodes 367 // - effectively ending SSA form. This requires either coalescing live 368 // ranges or inserting copies. For the moment, we insert "virtual copies" 369 // - we pretend there is a copy prior to each Phi in predecessor blocks. 370 // We will attempt to coalesce such "virtual copies" before we manifest 371 // them for real. 372 de_ssa(); 373 374 #ifdef ASSERT 375 // Veify the graph before RA. 376 verify(&live_arena); 377 #endif 378 379 { 380 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 381 _live = NULL; // Mark live as being not available 382 rm.reset_to_mark(); // Reclaim working storage 383 IndexSet::reset_memory(C, &live_arena); 384 ifg.init(_lrg_map.max_lrg_id()); // Empty IFG 385 gather_lrg_masks( false ); // Collect LRG masks 386 live.compute(_lrg_map.max_lrg_id()); // Compute liveness 387 _live = &live; // Mark LIVE as being available 388 } 389 390 // Base pointers are currently "used" by instructions which define new 391 // derived pointers. This makes base pointers live up to the where the 392 // derived pointer is made, but not beyond. Really, they need to be live 393 // across any GC point where the derived value is live. So this code looks 394 // at all the GC points, and "stretches" the live range of any base pointer 395 // to the GC point. 396 if (stretch_base_pointer_live_ranges(&live_arena)) { 397 Compile::TracePhase tp("computeLive (sbplr)", &timers[_t_computeLive]); 398 // Since some live range stretched, I need to recompute live 399 _live = NULL; 400 rm.reset_to_mark(); // Reclaim working storage 401 IndexSet::reset_memory(C, &live_arena); 402 ifg.init(_lrg_map.max_lrg_id()); 403 gather_lrg_masks(false); 404 live.compute(_lrg_map.max_lrg_id()); 405 _live = &live; 406 } 407 // Create the interference graph using virtual copies 408 build_ifg_virtual(); // Include stack slots this time 409 410 // The IFG is/was triangular. I am 'squaring it up' so Union can run 411 // faster. Union requires a 'for all' operation which is slow on the 412 // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - 413 // meaning I can visit all the Nodes neighbors less than a Node in time 414 // O(# of neighbors), but I have to visit all the Nodes greater than a 415 // given Node and search them for an instance, i.e., time O(#MaxLRG)). 416 _ifg->SquareUp(); 417 418 // Aggressive (but pessimistic) copy coalescing. 419 // This pass works on virtual copies. Any virtual copies which are not 420 // coalesced get manifested as actual copies 421 { 422 Compile::TracePhase tp("chaitinCoalesce1", &timers[_t_chaitinCoalesce1]); 423 424 PhaseAggressiveCoalesce coalesce(*this); 425 coalesce.coalesce_driver(); 426 // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do 427 // not match the Phi itself, insert a copy. 428 coalesce.insert_copies(_matcher); 429 if (C->failing()) { 430 return; 431 } 432 } 433 434 // After aggressive coalesce, attempt a first cut at coloring. 435 // To color, we need the IFG and for that we need LIVE. 436 { 437 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 438 _live = NULL; 439 rm.reset_to_mark(); // Reclaim working storage 440 IndexSet::reset_memory(C, &live_arena); 441 ifg.init(_lrg_map.max_lrg_id()); 442 gather_lrg_masks( true ); 443 live.compute(_lrg_map.max_lrg_id()); 444 _live = &live; 445 } 446 447 // Build physical interference graph 448 uint must_spill = 0; 449 must_spill = build_ifg_physical(&live_arena); 450 // If we have a guaranteed spill, might as well spill now 451 if (must_spill) { 452 if(!_lrg_map.max_lrg_id()) { 453 return; 454 } 455 // Bail out if unique gets too large (ie - unique > MaxNodeLimit) 456 C->check_node_count(10*must_spill, "out of nodes before split"); 457 if (C->failing()) { 458 return; 459 } 460 461 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 462 _lrg_map.set_max_lrg_id(new_max_lrg_id); 463 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 464 // or we failed to split 465 C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); 466 if (C->failing()) { 467 return; 468 } 469 470 NOT_PRODUCT(C->verify_graph_edges();) 471 472 compact(); // Compact LRGs; return new lower max lrg 473 474 { 475 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 476 _live = NULL; 477 rm.reset_to_mark(); // Reclaim working storage 478 IndexSet::reset_memory(C, &live_arena); 479 ifg.init(_lrg_map.max_lrg_id()); // Build a new interference graph 480 gather_lrg_masks( true ); // Collect intersect mask 481 live.compute(_lrg_map.max_lrg_id()); // Compute LIVE 482 _live = &live; 483 } 484 build_ifg_physical(&live_arena); 485 _ifg->SquareUp(); 486 _ifg->Compute_Effective_Degree(); 487 // Only do conservative coalescing if requested 488 if (OptoCoalesce) { 489 Compile::TracePhase tp("chaitinCoalesce2", &timers[_t_chaitinCoalesce2]); 490 // Conservative (and pessimistic) copy coalescing of those spills 491 PhaseConservativeCoalesce coalesce(*this); 492 // If max live ranges greater than cutoff, don't color the stack. 493 // This cutoff can be larger than below since it is only done once. 494 coalesce.coalesce_driver(); 495 } 496 _lrg_map.compress_uf_map_for_nodes(); 497 498 #ifdef ASSERT 499 verify(&live_arena, true); 500 #endif 501 } else { 502 ifg.SquareUp(); 503 ifg.Compute_Effective_Degree(); 504 #ifdef ASSERT 505 set_was_low(); 506 #endif 507 } 508 509 // Prepare for Simplify & Select 510 cache_lrg_info(); // Count degree of LRGs 511 512 // Simplify the InterFerence Graph by removing LRGs of low degree. 513 // LRGs of low degree are trivially colorable. 514 Simplify(); 515 516 // Select colors by re-inserting LRGs back into the IFG in reverse order. 517 // Return whether or not something spills. 518 uint spills = Select( ); 519 520 // If we spill, split and recycle the entire thing 521 while( spills ) { 522 if( _trip_cnt++ > 24 ) { 523 DEBUG_ONLY( dump_for_spill_split_recycle(); ) 524 if( _trip_cnt > 27 ) { 525 C->record_method_not_compilable("failed spill-split-recycle sanity check"); 526 return; 527 } 528 } 529 530 if (!_lrg_map.max_lrg_id()) { 531 return; 532 } 533 uint new_max_lrg_id = Split(_lrg_map.max_lrg_id(), &split_arena); // Split spilling LRG everywhere 534 _lrg_map.set_max_lrg_id(new_max_lrg_id); 535 // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) 536 C->check_node_count(2 * NodeLimitFudgeFactor, "out of nodes after split"); 537 if (C->failing()) { 538 return; 539 } 540 541 compact(); // Compact LRGs; return new lower max lrg 542 543 // Nuke the live-ness and interference graph and LiveRanGe info 544 { 545 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 546 _live = NULL; 547 rm.reset_to_mark(); // Reclaim working storage 548 IndexSet::reset_memory(C, &live_arena); 549 ifg.init(_lrg_map.max_lrg_id()); 550 551 // Create LiveRanGe array. 552 // Intersect register masks for all USEs and DEFs 553 gather_lrg_masks(true); 554 live.compute(_lrg_map.max_lrg_id()); 555 _live = &live; 556 } 557 must_spill = build_ifg_physical(&live_arena); 558 _ifg->SquareUp(); 559 _ifg->Compute_Effective_Degree(); 560 561 // Only do conservative coalescing if requested 562 if (OptoCoalesce) { 563 Compile::TracePhase tp("chaitinCoalesce3", &timers[_t_chaitinCoalesce3]); 564 // Conservative (and pessimistic) copy coalescing 565 PhaseConservativeCoalesce coalesce(*this); 566 // Check for few live ranges determines how aggressive coalesce is. 567 coalesce.coalesce_driver(); 568 } 569 _lrg_map.compress_uf_map_for_nodes(); 570 #ifdef ASSERT 571 verify(&live_arena, true); 572 #endif 573 cache_lrg_info(); // Count degree of LRGs 574 575 // Simplify the InterFerence Graph by removing LRGs of low degree. 576 // LRGs of low degree are trivially colorable. 577 Simplify(); 578 579 // Select colors by re-inserting LRGs back into the IFG in reverse order. 580 // Return whether or not something spills. 581 spills = Select(); 582 } 583 584 // Count number of Simplify-Select trips per coloring success. 585 _allocator_attempts += _trip_cnt + 1; 586 _allocator_successes += 1; 587 588 // Peephole remove copies 589 post_allocate_copy_removal(); 590 591 // Merge multidefs if multiple defs representing the same value are used in a single block. 592 merge_multidefs(); 593 594 #ifdef ASSERT 595 // Veify the graph after RA. 596 verify(&live_arena); 597 #endif 598 599 // max_reg is past the largest *register* used. 600 // Convert that to a frame_slot number. 601 if (_max_reg <= _matcher._new_SP) { 602 _framesize = C->out_preserve_stack_slots(); 603 } 604 else { 605 _framesize = _max_reg -_matcher._new_SP; 606 } 607 assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); 608 609 // This frame must preserve the required fp alignment 610 _framesize = round_to(_framesize, Matcher::stack_alignment_in_slots()); 611 assert(_framesize <= 1000000, "sanity check"); 612 #ifndef PRODUCT 613 _total_framesize += _framesize; 614 if ((int)_framesize > _max_framesize) { 615 _max_framesize = _framesize; 616 } 617 #endif 618 619 // Convert CISC spills 620 fixup_spills(); 621 622 // Log regalloc results 623 CompileLog* log = Compile::current()->log(); 624 if (log != NULL) { 625 log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); 626 } 627 628 if (C->failing()) { 629 return; 630 } 631 632 NOT_PRODUCT(C->verify_graph_edges();) 633 634 // Move important info out of the live_arena to longer lasting storage. 635 alloc_node_regs(_lrg_map.size()); 636 for (uint i=0; i < _lrg_map.size(); i++) { 637 if (_lrg_map.live_range_id(i)) { // Live range associated with Node? 638 LRG &lrg = lrgs(_lrg_map.live_range_id(i)); 639 if (!lrg.alive()) { 640 set_bad(i); 641 } else if (lrg.num_regs() == 1) { 642 set1(i, lrg.reg()); 643 } else { // Must be a register-set 644 if (!lrg._fat_proj) { // Must be aligned adjacent register set 645 // Live ranges record the highest register in their mask. 646 // We want the low register for the AD file writer's convenience. 647 OptoReg::Name hi = lrg.reg(); // Get hi register 648 OptoReg::Name lo = OptoReg::add(hi, (1-lrg.num_regs())); // Find lo 649 // We have to use pair [lo,lo+1] even for wide vectors because 650 // the rest of code generation works only with pairs. It is safe 651 // since for registers encoding only 'lo' is used. 652 // Second reg from pair is used in ScheduleAndBundle on SPARC where 653 // vector max size is 8 which corresponds to registers pair. 654 // It is also used in BuildOopMaps but oop operations are not 655 // vectorized. 656 set2(i, lo); 657 } else { // Misaligned; extract 2 bits 658 OptoReg::Name hi = lrg.reg(); // Get hi register 659 lrg.Remove(hi); // Yank from mask 660 int lo = lrg.mask().find_first_elem(); // Find lo 661 set_pair(i, hi, lo); 662 } 663 } 664 if( lrg._is_oop ) _node_oops.set(i); 665 } else { 666 set_bad(i); 667 } 668 } 669 670 // Done! 671 _live = NULL; 672 _ifg = NULL; 673 C->set_indexSet_arena(NULL); // ResourceArea is at end of scope 674 } 675 676 void PhaseChaitin::de_ssa() { 677 // Set initial Names for all Nodes. Most Nodes get the virtual register 678 // number. A few get the ZERO live range number. These do not 679 // get allocated, but instead rely on correct scheduling to ensure that 680 // only one instance is simultaneously live at a time. 681 uint lr_counter = 1; 682 for( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 683 Block* block = _cfg.get_block(i); 684 uint cnt = block->number_of_nodes(); 685 686 // Handle all the normal Nodes in the block 687 for( uint j = 0; j < cnt; j++ ) { 688 Node *n = block->get_node(j); 689 // Pre-color to the zero live range, or pick virtual register 690 const RegMask &rm = n->out_RegMask(); 691 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0); 692 } 693 } 694 695 // Reset the Union-Find mapping to be identity 696 _lrg_map.reset_uf_map(lr_counter); 697 } 698 699 void PhaseChaitin::mark_ssa() { 700 // Use ssa names to populate the live range maps or if no mask 701 // is available, use the 0 entry. 702 uint max_idx = 0; 703 for ( uint i = 0; i < _cfg.number_of_blocks(); i++ ) { 704 Block* block = _cfg.get_block(i); 705 uint cnt = block->number_of_nodes(); 706 707 // Handle all the normal Nodes in the block 708 for ( uint j = 0; j < cnt; j++ ) { 709 Node *n = block->get_node(j); 710 // Pre-color to the zero live range, or pick virtual register 711 const RegMask &rm = n->out_RegMask(); 712 _lrg_map.map(n->_idx, rm.is_NotEmpty() ? n->_idx : 0); 713 max_idx = (n->_idx > max_idx) ? n->_idx : max_idx; 714 } 715 } 716 _lrg_map.set_max_lrg_id(max_idx+1); 717 718 // Reset the Union-Find mapping to be identity 719 _lrg_map.reset_uf_map(max_idx+1); 720 } 721 722 723 // Gather LiveRanGe information, including register masks. Modification of 724 // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. 725 void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { 726 727 // Nail down the frame pointer live range 728 uint fp_lrg = _lrg_map.live_range_id(_cfg.get_root_node()->in(1)->in(TypeFunc::FramePtr)); 729 lrgs(fp_lrg)._cost += 1e12; // Cost is infinite 730 731 // For all blocks 732 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 733 Block* block = _cfg.get_block(i); 734 735 // For all instructions 736 for (uint j = 1; j < block->number_of_nodes(); j++) { 737 Node* n = block->get_node(j); 738 uint input_edge_start =1; // Skip control most nodes 739 bool is_machine_node = false; 740 if (n->is_Mach()) { 741 is_machine_node = true; 742 input_edge_start = n->as_Mach()->oper_input_base(); 743 } 744 uint idx = n->is_Copy(); 745 746 // Get virtual register number, same as LiveRanGe index 747 uint vreg = _lrg_map.live_range_id(n); 748 LRG& lrg = lrgs(vreg); 749 if (vreg) { // No vreg means un-allocable (e.g. memory) 750 751 // Collect has-copy bit 752 if (idx) { 753 lrg._has_copy = 1; 754 uint clidx = _lrg_map.live_range_id(n->in(idx)); 755 LRG& copy_src = lrgs(clidx); 756 copy_src._has_copy = 1; 757 } 758 759 // Check for float-vs-int live range (used in register-pressure 760 // calculations) 761 const Type *n_type = n->bottom_type(); 762 if (n_type->is_floatingpoint()) { 763 lrg._is_float = 1; 764 } 765 766 // Check for twice prior spilling. Once prior spilling might have 767 // spilled 'soft', 2nd prior spill should have spilled 'hard' and 768 // further spilling is unlikely to make progress. 769 if (_spilled_once.test(n->_idx)) { 770 lrg._was_spilled1 = 1; 771 if (_spilled_twice.test(n->_idx)) { 772 lrg._was_spilled2 = 1; 773 } 774 } 775 776 #ifndef PRODUCT 777 if (trace_spilling() && lrg._def != NULL) { 778 // collect defs for MultiDef printing 779 if (lrg._defs == NULL) { 780 lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(_ifg->_arena, 2, 0, NULL); 781 lrg._defs->append(lrg._def); 782 } 783 lrg._defs->append(n); 784 } 785 #endif 786 787 // Check for a single def LRG; these can spill nicely 788 // via rematerialization. Flag as NULL for no def found 789 // yet, or 'n' for single def or -1 for many defs. 790 lrg._def = lrg._def ? NodeSentinel : n; 791 792 // Limit result register mask to acceptable registers 793 const RegMask &rm = n->out_RegMask(); 794 lrg.AND( rm ); 795 796 int ireg = n->ideal_reg(); 797 assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, 798 "oops must be in Op_RegP's" ); 799 800 // Check for vector live range (only if vector register is used). 801 // On SPARC vector uses RegD which could be misaligned so it is not 802 // processes as vector in RA. 803 if (RegMask::is_vector(ireg)) 804 lrg._is_vector = 1; 805 assert(n_type->isa_vect() == NULL || lrg._is_vector || ireg == Op_RegD || ireg == Op_RegL, 806 "vector must be in vector registers"); 807 808 // Check for bound register masks 809 const RegMask &lrgmask = lrg.mask(); 810 if (lrgmask.is_bound(ireg)) { 811 lrg._is_bound = 1; 812 } 813 814 // Check for maximum frequency value 815 if (lrg._maxfreq < block->_freq) { 816 lrg._maxfreq = block->_freq; 817 } 818 819 // Check for oop-iness, or long/double 820 // Check for multi-kill projection 821 switch (ireg) { 822 case MachProjNode::fat_proj: 823 // Fat projections have size equal to number of registers killed 824 lrg.set_num_regs(rm.Size()); 825 lrg.set_reg_pressure(lrg.num_regs()); 826 lrg._fat_proj = 1; 827 lrg._is_bound = 1; 828 break; 829 case Op_RegP: 830 #ifdef _LP64 831 lrg.set_num_regs(2); // Size is 2 stack words 832 #else 833 lrg.set_num_regs(1); // Size is 1 stack word 834 #endif 835 // Register pressure is tracked relative to the maximum values 836 // suggested for that platform, INTPRESSURE and FLOATPRESSURE, 837 // and relative to other types which compete for the same regs. 838 // 839 // The following table contains suggested values based on the 840 // architectures as defined in each .ad file. 841 // INTPRESSURE and FLOATPRESSURE may be tuned differently for 842 // compile-speed or performance. 843 // Note1: 844 // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 845 // since .ad registers are defined as high and low halves. 846 // These reg_pressure values remain compatible with the code 847 // in is_high_pressure() which relates get_invalid_mask_size(), 848 // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. 849 // Note2: 850 // SPARC -d32 has 24 registers available for integral values, 851 // but only 10 of these are safe for 64-bit longs. 852 // Using set_reg_pressure(2) for both int and long means 853 // the allocator will believe it can fit 26 longs into 854 // registers. Using 2 for longs and 1 for ints means the 855 // allocator will attempt to put 52 integers into registers. 856 // The settings below limit this problem to methods with 857 // many long values which are being run on 32-bit SPARC. 858 // 859 // ------------------- reg_pressure -------------------- 860 // Each entry is reg_pressure_per_value,number_of_regs 861 // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE 862 // IA32 2 1 1 1 1 6 6 863 // IA64 1 1 1 1 1 50 41 864 // SPARC 2 2 2 2 2 48 (24) 52 (26) 865 // SPARCV9 2 2 2 2 2 48 (24) 52 (26) 866 // AMD64 1 1 1 1 1 14 15 867 // ----------------------------------------------------- 868 #if defined(SPARC) 869 lrg.set_reg_pressure(2); // use for v9 as well 870 #else 871 lrg.set_reg_pressure(1); // normally one value per register 872 #endif 873 if( n_type->isa_oop_ptr() ) { 874 lrg._is_oop = 1; 875 } 876 break; 877 case Op_RegL: // Check for long or double 878 case Op_RegD: 879 lrg.set_num_regs(2); 880 // Define platform specific register pressure 881 #if defined(SPARC) || defined(ARM32) 882 lrg.set_reg_pressure(2); 883 #elif defined(IA32) 884 if( ireg == Op_RegL ) { 885 lrg.set_reg_pressure(2); 886 } else { 887 lrg.set_reg_pressure(1); 888 } 889 #else 890 lrg.set_reg_pressure(1); // normally one value per register 891 #endif 892 // If this def of a double forces a mis-aligned double, 893 // flag as '_fat_proj' - really flag as allowing misalignment 894 // AND changes how we count interferences. A mis-aligned 895 // double can interfere with TWO aligned pairs, or effectively 896 // FOUR registers! 897 if (rm.is_misaligned_pair()) { 898 lrg._fat_proj = 1; 899 lrg._is_bound = 1; 900 } 901 break; 902 case Op_RegF: 903 case Op_RegI: 904 case Op_RegN: 905 case Op_RegFlags: 906 case 0: // not an ideal register 907 lrg.set_num_regs(1); 908 #ifdef SPARC 909 lrg.set_reg_pressure(2); 910 #else 911 lrg.set_reg_pressure(1); 912 #endif 913 break; 914 case Op_VecS: 915 assert(Matcher::vector_size_supported(T_BYTE,4), "sanity"); 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(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecD), "sanity"); 922 assert(RegMask::num_registers(Op_VecD) == RegMask::SlotsPerVecD, "sanity"); 923 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecD), "vector should be aligned"); 924 lrg.set_num_regs(RegMask::SlotsPerVecD); 925 lrg.set_reg_pressure(1); 926 break; 927 case Op_VecX: 928 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecX), "sanity"); 929 assert(RegMask::num_registers(Op_VecX) == RegMask::SlotsPerVecX, "sanity"); 930 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecX), "vector should be aligned"); 931 lrg.set_num_regs(RegMask::SlotsPerVecX); 932 lrg.set_reg_pressure(1); 933 break; 934 case Op_VecY: 935 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecY), "sanity"); 936 assert(RegMask::num_registers(Op_VecY) == RegMask::SlotsPerVecY, "sanity"); 937 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecY), "vector should be aligned"); 938 lrg.set_num_regs(RegMask::SlotsPerVecY); 939 lrg.set_reg_pressure(1); 940 break; 941 case Op_VecZ: 942 assert(Matcher::vector_size_supported(T_FLOAT,RegMask::SlotsPerVecZ), "sanity"); 943 assert(RegMask::num_registers(Op_VecZ) == RegMask::SlotsPerVecZ, "sanity"); 944 assert(lrgmask.is_aligned_sets(RegMask::SlotsPerVecZ), "vector should be aligned"); 945 lrg.set_num_regs(RegMask::SlotsPerVecZ); 946 lrg.set_reg_pressure(1); 947 break; 948 default: 949 ShouldNotReachHere(); 950 } 951 } 952 953 // Now do the same for inputs 954 uint cnt = n->req(); 955 // Setup for CISC SPILLING 956 uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; 957 if( UseCISCSpill && after_aggressive ) { 958 inp = n->cisc_operand(); 959 if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) 960 // Convert operand number to edge index number 961 inp = n->as_Mach()->operand_index(inp); 962 } 963 964 // Prepare register mask for each input 965 for( uint k = input_edge_start; k < cnt; k++ ) { 966 uint vreg = _lrg_map.live_range_id(n->in(k)); 967 if (!vreg) { 968 continue; 969 } 970 971 // If this instruction is CISC Spillable, add the flags 972 // bit to its appropriate input 973 if( UseCISCSpill && after_aggressive && inp == k ) { 974 #ifndef PRODUCT 975 if( TraceCISCSpill ) { 976 tty->print(" use_cisc_RegMask: "); 977 n->dump(); 978 } 979 #endif 980 n->as_Mach()->use_cisc_RegMask(); 981 } 982 983 if (is_machine_node && _scheduling_info_generated) { 984 MachNode* cur_node = n->as_Mach(); 985 // this is cleaned up by register allocation 986 if (k >= cur_node->num_opnds()) continue; 987 } 988 989 LRG &lrg = lrgs(vreg); 990 // // Testing for floating point code shape 991 // Node *test = n->in(k); 992 // if( test->is_Mach() ) { 993 // MachNode *m = test->as_Mach(); 994 // int op = m->ideal_Opcode(); 995 // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { 996 // int zzz = 1; 997 // } 998 // } 999 1000 // Limit result register mask to acceptable registers. 1001 // Do not limit registers from uncommon uses before 1002 // AggressiveCoalesce. This effectively pre-virtual-splits 1003 // around uncommon uses of common defs. 1004 const RegMask &rm = n->in_RegMask(k); 1005 if (!after_aggressive && _cfg.get_block_for_node(n->in(k))->_freq > 1000 * block->_freq) { 1006 // Since we are BEFORE aggressive coalesce, leave the register 1007 // mask untrimmed by the call. This encourages more coalescing. 1008 // Later, AFTER aggressive, this live range will have to spill 1009 // but the spiller handles slow-path calls very nicely. 1010 } else { 1011 lrg.AND( rm ); 1012 } 1013 1014 // Check for bound register masks 1015 const RegMask &lrgmask = lrg.mask(); 1016 int kreg = n->in(k)->ideal_reg(); 1017 bool is_vect = RegMask::is_vector(kreg); 1018 assert(n->in(k)->bottom_type()->isa_vect() == NULL || 1019 is_vect || kreg == Op_RegD || kreg == Op_RegL, 1020 "vector must be in vector registers"); 1021 if (lrgmask.is_bound(kreg)) 1022 lrg._is_bound = 1; 1023 1024 // If this use of a double forces a mis-aligned double, 1025 // flag as '_fat_proj' - really flag as allowing misalignment 1026 // AND changes how we count interferences. A mis-aligned 1027 // double can interfere with TWO aligned pairs, or effectively 1028 // FOUR registers! 1029 #ifdef ASSERT 1030 if (is_vect && !_scheduling_info_generated) { 1031 if (lrg.num_regs() != 0) { 1032 assert(lrgmask.is_aligned_sets(lrg.num_regs()), "vector should be aligned"); 1033 assert(!lrg._fat_proj, "sanity"); 1034 assert(RegMask::num_registers(kreg) == lrg.num_regs(), "sanity"); 1035 } else { 1036 assert(n->is_Phi(), "not all inputs processed only if Phi"); 1037 } 1038 } 1039 #endif 1040 if (!is_vect && lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_pair()) { 1041 lrg._fat_proj = 1; 1042 lrg._is_bound = 1; 1043 } 1044 // if the LRG is an unaligned pair, we will have to spill 1045 // so clear the LRG's register mask if it is not already spilled 1046 if (!is_vect && !n->is_SpillCopy() && 1047 (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && 1048 lrgmask.is_misaligned_pair()) { 1049 lrg.Clear(); 1050 } 1051 1052 // Check for maximum frequency value 1053 if (lrg._maxfreq < block->_freq) { 1054 lrg._maxfreq = block->_freq; 1055 } 1056 1057 } // End for all allocated inputs 1058 } // end for all instructions 1059 } // end for all blocks 1060 1061 // Final per-liverange setup 1062 for (uint i2 = 0; i2 < _lrg_map.max_lrg_id(); i2++) { 1063 LRG &lrg = lrgs(i2); 1064 assert(!lrg._is_vector || !lrg._fat_proj, "sanity"); 1065 if (lrg.num_regs() > 1 && !lrg._fat_proj) { 1066 lrg.clear_to_sets(); 1067 } 1068 lrg.compute_set_mask_size(); 1069 if (lrg.not_free()) { // Handle case where we lose from the start 1070 lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); 1071 lrg._direct_conflict = 1; 1072 } 1073 lrg.set_degree(0); // no neighbors in IFG yet 1074 } 1075 } 1076 1077 // Set the was-lo-degree bit. Conservative coalescing should not change the 1078 // colorability of the graph. If any live range was of low-degree before 1079 // coalescing, it should Simplify. This call sets the was-lo-degree bit. 1080 // The bit is checked in Simplify. 1081 void PhaseChaitin::set_was_low() { 1082 #ifdef ASSERT 1083 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1084 int size = lrgs(i).num_regs(); 1085 uint old_was_lo = lrgs(i)._was_lo; 1086 lrgs(i)._was_lo = 0; 1087 if( lrgs(i).lo_degree() ) { 1088 lrgs(i)._was_lo = 1; // Trivially of low degree 1089 } else { // Else check the Brigg's assertion 1090 // Brigg's observation is that the lo-degree neighbors of a 1091 // hi-degree live range will not interfere with the color choices 1092 // of said hi-degree live range. The Simplify reverse-stack-coloring 1093 // order takes care of the details. Hence you do not have to count 1094 // low-degree neighbors when determining if this guy colors. 1095 int briggs_degree = 0; 1096 IndexSet *s = _ifg->neighbors(i); 1097 IndexSetIterator elements(s); 1098 uint lidx; 1099 while((lidx = elements.next()) != 0) { 1100 if( !lrgs(lidx).lo_degree() ) 1101 briggs_degree += MAX2(size,lrgs(lidx).num_regs()); 1102 } 1103 if( briggs_degree < lrgs(i).degrees_of_freedom() ) 1104 lrgs(i)._was_lo = 1; // Low degree via the briggs assertion 1105 } 1106 assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); 1107 } 1108 #endif 1109 } 1110 1111 #define REGISTER_CONSTRAINED 16 1112 1113 // Compute cost/area ratio, in case we spill. Build the lo-degree list. 1114 void PhaseChaitin::cache_lrg_info( ) { 1115 Compile::TracePhase tp("chaitinCacheLRG", &timers[_t_chaitinCacheLRG]); 1116 1117 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1118 LRG &lrg = lrgs(i); 1119 1120 // Check for being of low degree: means we can be trivially colored. 1121 // Low degree, dead or must-spill guys just get to simplify right away 1122 if( lrg.lo_degree() || 1123 !lrg.alive() || 1124 lrg._must_spill ) { 1125 // Split low degree list into those guys that must get a 1126 // register and those that can go to register or stack. 1127 // The idea is LRGs that can go register or stack color first when 1128 // they have a good chance of getting a register. The register-only 1129 // lo-degree live ranges always get a register. 1130 OptoReg::Name hi_reg = lrg.mask().find_last_elem(); 1131 if( OptoReg::is_stack(hi_reg)) { // Can go to stack? 1132 lrg._next = _lo_stk_degree; 1133 _lo_stk_degree = i; 1134 } else { 1135 lrg._next = _lo_degree; 1136 _lo_degree = i; 1137 } 1138 } else { // Else high degree 1139 lrgs(_hi_degree)._prev = i; 1140 lrg._next = _hi_degree; 1141 lrg._prev = 0; 1142 _hi_degree = i; 1143 } 1144 } 1145 } 1146 1147 // Simplify the IFG by removing LRGs of low degree that have NO copies 1148 void PhaseChaitin::Pre_Simplify( ) { 1149 1150 // Warm up the lo-degree no-copy list 1151 int lo_no_copy = 0; 1152 for (uint i = 1; i < _lrg_map.max_lrg_id(); i++) { 1153 if ((lrgs(i).lo_degree() && !lrgs(i)._has_copy) || 1154 !lrgs(i).alive() || 1155 lrgs(i)._must_spill) { 1156 lrgs(i)._next = lo_no_copy; 1157 lo_no_copy = i; 1158 } 1159 } 1160 1161 while( lo_no_copy ) { 1162 uint lo = lo_no_copy; 1163 lo_no_copy = lrgs(lo)._next; 1164 int size = lrgs(lo).num_regs(); 1165 1166 // Put the simplified guy on the simplified list. 1167 lrgs(lo)._next = _simplified; 1168 _simplified = lo; 1169 1170 // Yank this guy from the IFG. 1171 IndexSet *adj = _ifg->remove_node( lo ); 1172 1173 // If any neighbors' degrees fall below their number of 1174 // allowed registers, then put that neighbor on the low degree 1175 // list. Note that 'degree' can only fall and 'numregs' is 1176 // unchanged by this action. Thus the two are equal at most once, 1177 // so LRGs hit the lo-degree worklists at most once. 1178 IndexSetIterator elements(adj); 1179 uint neighbor; 1180 while ((neighbor = elements.next()) != 0) { 1181 LRG *n = &lrgs(neighbor); 1182 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1183 1184 // Check for just becoming of-low-degree 1185 if( n->just_lo_degree() && !n->_has_copy ) { 1186 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1187 // Put on lo-degree list 1188 n->_next = lo_no_copy; 1189 lo_no_copy = neighbor; 1190 } 1191 } 1192 } // End of while lo-degree no_copy worklist not empty 1193 1194 // No more lo-degree no-copy live ranges to simplify 1195 } 1196 1197 // Simplify the IFG by removing LRGs of low degree. 1198 void PhaseChaitin::Simplify( ) { 1199 Compile::TracePhase tp("chaitinSimplify", &timers[_t_chaitinSimplify]); 1200 1201 while( 1 ) { // Repeat till simplified it all 1202 // May want to explore simplifying lo_degree before _lo_stk_degree. 1203 // This might result in more spills coloring into registers during 1204 // Select(). 1205 while( _lo_degree || _lo_stk_degree ) { 1206 // If possible, pull from lo_stk first 1207 uint lo; 1208 if( _lo_degree ) { 1209 lo = _lo_degree; 1210 _lo_degree = lrgs(lo)._next; 1211 } else { 1212 lo = _lo_stk_degree; 1213 _lo_stk_degree = lrgs(lo)._next; 1214 } 1215 1216 // Put the simplified guy on the simplified list. 1217 lrgs(lo)._next = _simplified; 1218 _simplified = lo; 1219 // If this guy is "at risk" then mark his current neighbors 1220 if( lrgs(lo)._at_risk ) { 1221 IndexSetIterator elements(_ifg->neighbors(lo)); 1222 uint datum; 1223 while ((datum = elements.next()) != 0) { 1224 lrgs(datum)._risk_bias = lo; 1225 } 1226 } 1227 1228 // Yank this guy from the IFG. 1229 IndexSet *adj = _ifg->remove_node( lo ); 1230 1231 // If any neighbors' degrees fall below their number of 1232 // allowed registers, then put that neighbor on the low degree 1233 // list. Note that 'degree' can only fall and 'numregs' is 1234 // unchanged by this action. Thus the two are equal at most once, 1235 // so LRGs hit the lo-degree worklist at most once. 1236 IndexSetIterator elements(adj); 1237 uint neighbor; 1238 while ((neighbor = elements.next()) != 0) { 1239 LRG *n = &lrgs(neighbor); 1240 #ifdef ASSERT 1241 if( VerifyOpto || VerifyRegisterAllocator ) { 1242 assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); 1243 } 1244 #endif 1245 1246 // Check for just becoming of-low-degree just counting registers. 1247 // _must_spill live ranges are already on the low degree list. 1248 if( n->just_lo_degree() && !n->_must_spill ) { 1249 assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); 1250 // Pull from hi-degree list 1251 uint prev = n->_prev; 1252 uint next = n->_next; 1253 if( prev ) lrgs(prev)._next = next; 1254 else _hi_degree = next; 1255 lrgs(next)._prev = prev; 1256 n->_next = _lo_degree; 1257 _lo_degree = neighbor; 1258 } 1259 } 1260 } // End of while lo-degree/lo_stk_degree worklist not empty 1261 1262 // Check for got everything: is hi-degree list empty? 1263 if( !_hi_degree ) break; 1264 1265 // Time to pick a potential spill guy 1266 uint lo_score = _hi_degree; 1267 double score = lrgs(lo_score).score(); 1268 double area = lrgs(lo_score)._area; 1269 double cost = lrgs(lo_score)._cost; 1270 bool bound = lrgs(lo_score)._is_bound; 1271 1272 // Find cheapest guy 1273 debug_only( int lo_no_simplify=0; ); 1274 for( uint i = _hi_degree; i; i = lrgs(i)._next ) { 1275 assert( !(*_ifg->_yanked)[i], "" ); 1276 // It's just vaguely possible to move hi-degree to lo-degree without 1277 // going through a just-lo-degree stage: If you remove a double from 1278 // a float live range it's degree will drop by 2 and you can skip the 1279 // just-lo-degree stage. It's very rare (shows up after 5000+ methods 1280 // in -Xcomp of Java2Demo). So just choose this guy to simplify next. 1281 if( lrgs(i).lo_degree() ) { 1282 lo_score = i; 1283 break; 1284 } 1285 debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); 1286 double iscore = lrgs(i).score(); 1287 double iarea = lrgs(i)._area; 1288 double icost = lrgs(i)._cost; 1289 bool ibound = lrgs(i)._is_bound; 1290 1291 // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area 1292 // wins. Ties happen because all live ranges in question have spilled 1293 // a few times before and the spill-score adds a huge number which 1294 // washes out the low order bits. We are choosing the lesser of 2 1295 // evils; in this case pick largest area to spill. 1296 // Ties also happen when live ranges are defined and used only inside 1297 // one block. In which case their area is 0 and score set to max. 1298 // In such case choose bound live range over unbound to free registers 1299 // or with smaller cost to spill. 1300 if( iscore < score || 1301 (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || 1302 (iscore == score && iarea == area && 1303 ( (ibound && !bound) || ibound == bound && (icost < cost) )) ) { 1304 lo_score = i; 1305 score = iscore; 1306 area = iarea; 1307 cost = icost; 1308 bound = ibound; 1309 } 1310 } 1311 LRG *lo_lrg = &lrgs(lo_score); 1312 // The live range we choose for spilling is either hi-degree, or very 1313 // rarely it can be low-degree. If we choose a hi-degree live range 1314 // there better not be any lo-degree choices. 1315 assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); 1316 1317 // Pull from hi-degree list 1318 uint prev = lo_lrg->_prev; 1319 uint next = lo_lrg->_next; 1320 if( prev ) lrgs(prev)._next = next; 1321 else _hi_degree = next; 1322 lrgs(next)._prev = prev; 1323 // Jam him on the lo-degree list, despite his high degree. 1324 // Maybe he'll get a color, and maybe he'll spill. 1325 // Only Select() will know. 1326 lrgs(lo_score)._at_risk = true; 1327 _lo_degree = lo_score; 1328 lo_lrg->_next = 0; 1329 1330 } // End of while not simplified everything 1331 1332 } 1333 1334 // Is 'reg' register legal for 'lrg'? 1335 static bool is_legal_reg(LRG &lrg, OptoReg::Name reg, int chunk) { 1336 if (reg >= chunk && reg < (chunk + RegMask::CHUNK_SIZE) && 1337 lrg.mask().Member(OptoReg::add(reg,-chunk))) { 1338 // RA uses OptoReg which represent the highest element of a registers set. 1339 // For example, vectorX (128bit) on x86 uses [XMM,XMMb,XMMc,XMMd] set 1340 // in which XMMd is used by RA to represent such vectors. A double value 1341 // uses [XMM,XMMb] pairs and XMMb is used by RA for it. 1342 // The register mask uses largest bits set of overlapping register sets. 1343 // On x86 with AVX it uses 8 bits for each XMM registers set. 1344 // 1345 // The 'lrg' already has cleared-to-set register mask (done in Select() 1346 // before calling choose_color()). Passing mask.Member(reg) check above 1347 // indicates that the size (num_regs) of 'reg' set is less or equal to 1348 // 'lrg' set size. 1349 // For set size 1 any register which is member of 'lrg' mask is legal. 1350 if (lrg.num_regs()==1) 1351 return true; 1352 // For larger sets only an aligned register with the same set size is legal. 1353 int mask = lrg.num_regs()-1; 1354 if ((reg&mask) == mask) 1355 return true; 1356 } 1357 return false; 1358 } 1359 1360 // Choose a color using the biasing heuristic 1361 OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { 1362 1363 // Check for "at_risk" LRG's 1364 uint risk_lrg = _lrg_map.find(lrg._risk_bias); 1365 if( risk_lrg != 0 ) { 1366 // Walk the colored neighbors of the "at_risk" candidate 1367 // Choose a color which is both legal and already taken by a neighbor 1368 // of the "at_risk" candidate in order to improve the chances of the 1369 // "at_risk" candidate of coloring 1370 IndexSetIterator elements(_ifg->neighbors(risk_lrg)); 1371 uint datum; 1372 while ((datum = elements.next()) != 0) { 1373 OptoReg::Name reg = lrgs(datum).reg(); 1374 // If this LRG's register is legal for us, choose it 1375 if (is_legal_reg(lrg, reg, chunk)) 1376 return reg; 1377 } 1378 } 1379 1380 uint copy_lrg = _lrg_map.find(lrg._copy_bias); 1381 if( copy_lrg != 0 ) { 1382 // If he has a color, 1383 if( !(*(_ifg->_yanked))[copy_lrg] ) { 1384 OptoReg::Name reg = lrgs(copy_lrg).reg(); 1385 // And it is legal for you, 1386 if (is_legal_reg(lrg, reg, chunk)) 1387 return reg; 1388 } else if( chunk == 0 ) { 1389 // Choose a color which is legal for him 1390 RegMask tempmask = lrg.mask(); 1391 tempmask.AND(lrgs(copy_lrg).mask()); 1392 tempmask.clear_to_sets(lrg.num_regs()); 1393 OptoReg::Name reg = tempmask.find_first_set(lrg.num_regs()); 1394 if (OptoReg::is_valid(reg)) 1395 return reg; 1396 } 1397 } 1398 1399 // If no bias info exists, just go with the register selection ordering 1400 if (lrg._is_vector || lrg.num_regs() == 2) { 1401 // Find an aligned set 1402 return OptoReg::add(lrg.mask().find_first_set(lrg.num_regs()),chunk); 1403 } 1404 1405 // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate 1406 // copy removal to remove many more copies, by preventing a just-assigned 1407 // register from being repeatedly assigned. 1408 OptoReg::Name reg = lrg.mask().find_first_elem(); 1409 if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { 1410 // This 'Remove; find; Insert' idiom is an expensive way to find the 1411 // SECOND element in the mask. 1412 lrg.Remove(reg); 1413 OptoReg::Name reg2 = lrg.mask().find_first_elem(); 1414 lrg.Insert(reg); 1415 if( OptoReg::is_reg(reg2)) 1416 reg = reg2; 1417 } 1418 return OptoReg::add( reg, chunk ); 1419 } 1420 1421 // Choose a color in the current chunk 1422 OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { 1423 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)"); 1424 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)"); 1425 1426 if( lrg.num_regs() == 1 || // Common Case 1427 !lrg._fat_proj ) // Aligned+adjacent pairs ok 1428 // Use a heuristic to "bias" the color choice 1429 return bias_color(lrg, chunk); 1430 1431 assert(!lrg._is_vector, "should be not vector here" ); 1432 assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); 1433 1434 // Fat-proj case or misaligned double argument. 1435 assert(lrg.compute_mask_size() == lrg.num_regs() || 1436 lrg.num_regs() == 2,"fat projs exactly color" ); 1437 assert( !chunk, "always color in 1st chunk" ); 1438 // Return the highest element in the set. 1439 return lrg.mask().find_last_elem(); 1440 } 1441 1442 // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted 1443 // in reverse order of removal. As long as nothing of hi-degree was yanked, 1444 // everything going back is guaranteed a color. Select that color. If some 1445 // hi-degree LRG cannot get a color then we record that we must spill. 1446 uint PhaseChaitin::Select( ) { 1447 Compile::TracePhase tp("chaitinSelect", &timers[_t_chaitinSelect]); 1448 1449 uint spill_reg = LRG::SPILL_REG; 1450 _max_reg = OptoReg::Name(0); // Past max register used 1451 while( _simplified ) { 1452 // Pull next LRG from the simplified list - in reverse order of removal 1453 uint lidx = _simplified; 1454 LRG *lrg = &lrgs(lidx); 1455 _simplified = lrg->_next; 1456 1457 1458 #ifndef PRODUCT 1459 if (trace_spilling()) { 1460 ttyLocker ttyl; 1461 tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), 1462 lrg->degrees_of_freedom()); 1463 lrg->dump(); 1464 } 1465 #endif 1466 1467 // Re-insert into the IFG 1468 _ifg->re_insert(lidx); 1469 if( !lrg->alive() ) continue; 1470 // capture allstackedness flag before mask is hacked 1471 const int is_allstack = lrg->mask().is_AllStack(); 1472 1473 // Yeah, yeah, yeah, I know, I know. I can refactor this 1474 // to avoid the GOTO, although the refactored code will not 1475 // be much clearer. We arrive here IFF we have a stack-based 1476 // live range that cannot color in the current chunk, and it 1477 // has to move into the next free stack chunk. 1478 int chunk = 0; // Current chunk is first chunk 1479 retry_next_chunk: 1480 1481 // Remove neighbor colors 1482 IndexSet *s = _ifg->neighbors(lidx); 1483 1484 debug_only(RegMask orig_mask = lrg->mask();) 1485 IndexSetIterator elements(s); 1486 uint neighbor; 1487 while ((neighbor = elements.next()) != 0) { 1488 // Note that neighbor might be a spill_reg. In this case, exclusion 1489 // of its color will be a no-op, since the spill_reg chunk is in outer 1490 // space. Also, if neighbor is in a different chunk, this exclusion 1491 // will be a no-op. (Later on, if lrg runs out of possible colors in 1492 // its chunk, a new chunk of color may be tried, in which case 1493 // examination of neighbors is started again, at retry_next_chunk.) 1494 LRG &nlrg = lrgs(neighbor); 1495 OptoReg::Name nreg = nlrg.reg(); 1496 // Only subtract masks in the same chunk 1497 if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { 1498 #ifndef PRODUCT 1499 uint size = lrg->mask().Size(); 1500 RegMask rm = lrg->mask(); 1501 #endif 1502 lrg->SUBTRACT(nlrg.mask()); 1503 #ifndef PRODUCT 1504 if (trace_spilling() && lrg->mask().Size() != size) { 1505 ttyLocker ttyl; 1506 tty->print("L%d ", lidx); 1507 rm.dump(); 1508 tty->print(" intersected L%d ", neighbor); 1509 nlrg.mask().dump(); 1510 tty->print(" removed "); 1511 rm.SUBTRACT(lrg->mask()); 1512 rm.dump(); 1513 tty->print(" leaving "); 1514 lrg->mask().dump(); 1515 tty->cr(); 1516 } 1517 #endif 1518 } 1519 } 1520 //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); 1521 // Aligned pairs need aligned masks 1522 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1523 if (lrg->num_regs() > 1 && !lrg->_fat_proj) { 1524 lrg->clear_to_sets(); 1525 } 1526 1527 // Check if a color is available and if so pick the color 1528 OptoReg::Name reg = choose_color( *lrg, chunk ); 1529 #ifdef SPARC 1530 debug_only(lrg->compute_set_mask_size()); 1531 assert(lrg->num_regs() < 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); 1532 #endif 1533 1534 //--------------- 1535 // If we fail to color and the AllStack flag is set, trigger 1536 // a chunk-rollover event 1537 if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { 1538 // Bump register mask up to next stack chunk 1539 chunk += RegMask::CHUNK_SIZE; 1540 lrg->Set_All(); 1541 1542 goto retry_next_chunk; 1543 } 1544 1545 //--------------- 1546 // Did we get a color? 1547 else if( OptoReg::is_valid(reg)) { 1548 #ifndef PRODUCT 1549 RegMask avail_rm = lrg->mask(); 1550 #endif 1551 1552 // Record selected register 1553 lrg->set_reg(reg); 1554 1555 if( reg >= _max_reg ) // Compute max register limit 1556 _max_reg = OptoReg::add(reg,1); 1557 // Fold reg back into normal space 1558 reg = OptoReg::add(reg,-chunk); 1559 1560 // If the live range is not bound, then we actually had some choices 1561 // to make. In this case, the mask has more bits in it than the colors 1562 // chosen. Restrict the mask to just what was picked. 1563 int n_regs = lrg->num_regs(); 1564 assert(!lrg->_is_vector || !lrg->_fat_proj, "sanity"); 1565 if (n_regs == 1 || !lrg->_fat_proj) { 1566 assert(!lrg->_is_vector || n_regs <= RegMask::SlotsPerVecZ, "sanity"); 1567 lrg->Clear(); // Clear the mask 1568 lrg->Insert(reg); // Set regmask to match selected reg 1569 // For vectors and pairs, also insert the low bit of the pair 1570 for (int i = 1; i < n_regs; i++) 1571 lrg->Insert(OptoReg::add(reg,-i)); 1572 lrg->set_mask_size(n_regs); 1573 } else { // Else fatproj 1574 // mask must be equal to fatproj bits, by definition 1575 } 1576 #ifndef PRODUCT 1577 if (trace_spilling()) { 1578 ttyLocker ttyl; 1579 tty->print("L%d selected ", lidx); 1580 lrg->mask().dump(); 1581 tty->print(" from "); 1582 avail_rm.dump(); 1583 tty->cr(); 1584 } 1585 #endif 1586 // Note that reg is the highest-numbered register in the newly-bound mask. 1587 } // end color available case 1588 1589 //--------------- 1590 // Live range is live and no colors available 1591 else { 1592 assert( lrg->alive(), "" ); 1593 assert( !lrg->_fat_proj || lrg->is_multidef() || 1594 lrg->_def->outcnt() > 0, "fat_proj cannot spill"); 1595 assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); 1596 1597 // Assign the special spillreg register 1598 lrg->set_reg(OptoReg::Name(spill_reg++)); 1599 // Do not empty the regmask; leave mask_size lying around 1600 // for use during Spilling 1601 #ifndef PRODUCT 1602 if( trace_spilling() ) { 1603 ttyLocker ttyl; 1604 tty->print("L%d spilling with neighbors: ", lidx); 1605 s->dump(); 1606 debug_only(tty->print(" original mask: ")); 1607 debug_only(orig_mask.dump()); 1608 dump_lrg(lidx); 1609 } 1610 #endif 1611 } // end spill case 1612 1613 } 1614 1615 return spill_reg-LRG::SPILL_REG; // Return number of spills 1616 } 1617 1618 // Copy 'was_spilled'-edness from the source Node to the dst Node. 1619 void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { 1620 if( _spilled_once.test(src->_idx) ) { 1621 _spilled_once.set(dst->_idx); 1622 lrgs(_lrg_map.find(dst))._was_spilled1 = 1; 1623 if( _spilled_twice.test(src->_idx) ) { 1624 _spilled_twice.set(dst->_idx); 1625 lrgs(_lrg_map.find(dst))._was_spilled2 = 1; 1626 } 1627 } 1628 } 1629 1630 // Set the 'spilled_once' or 'spilled_twice' flag on a node. 1631 void PhaseChaitin::set_was_spilled( Node *n ) { 1632 if( _spilled_once.test_set(n->_idx) ) 1633 _spilled_twice.set(n->_idx); 1634 } 1635 1636 // Convert Ideal spill instructions into proper FramePtr + offset Loads and 1637 // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. 1638 void PhaseChaitin::fixup_spills() { 1639 // This function does only cisc spill work. 1640 if( !UseCISCSpill ) return; 1641 1642 Compile::TracePhase tp("fixupSpills", &timers[_t_fixupSpills]); 1643 1644 // Grab the Frame Pointer 1645 Node *fp = _cfg.get_root_block()->head()->in(1)->in(TypeFunc::FramePtr); 1646 1647 // For all blocks 1648 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1649 Block* block = _cfg.get_block(i); 1650 1651 // For all instructions in block 1652 uint last_inst = block->end_idx(); 1653 for (uint j = 1; j <= last_inst; j++) { 1654 Node* n = block->get_node(j); 1655 1656 // Dead instruction??? 1657 assert( n->outcnt() != 0 ||// Nothing dead after post alloc 1658 C->top() == n || // Or the random TOP node 1659 n->is_Proj(), // Or a fat-proj kill node 1660 "No dead instructions after post-alloc" ); 1661 1662 int inp = n->cisc_operand(); 1663 if( inp != AdlcVMDeps::Not_cisc_spillable ) { 1664 // Convert operand number to edge index number 1665 MachNode *mach = n->as_Mach(); 1666 inp = mach->operand_index(inp); 1667 Node *src = n->in(inp); // Value to load or store 1668 LRG &lrg_cisc = lrgs(_lrg_map.find_const(src)); 1669 OptoReg::Name src_reg = lrg_cisc.reg(); 1670 // Doubles record the HIGH register of an adjacent pair. 1671 src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); 1672 if( OptoReg::is_stack(src_reg) ) { // If input is on stack 1673 // This is a CISC Spill, get stack offset and construct new node 1674 #ifndef PRODUCT 1675 if( TraceCISCSpill ) { 1676 tty->print(" reg-instr: "); 1677 n->dump(); 1678 } 1679 #endif 1680 int stk_offset = reg2offset(src_reg); 1681 // Bailout if we might exceed node limit when spilling this instruction 1682 C->check_node_count(0, "out of nodes fixing spills"); 1683 if (C->failing()) return; 1684 // Transform node 1685 MachNode *cisc = mach->cisc_version(stk_offset)->as_Mach(); 1686 cisc->set_req(inp,fp); // Base register is frame pointer 1687 if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { 1688 assert( cisc->oper_input_base() == 2, "Only adding one edge"); 1689 cisc->ins_req(1,src); // Requires a memory edge 1690 } 1691 block->map_node(cisc, j); // Insert into basic block 1692 n->subsume_by(cisc, C); // Correct graph 1693 // 1694 ++_used_cisc_instructions; 1695 #ifndef PRODUCT 1696 if( TraceCISCSpill ) { 1697 tty->print(" cisc-instr: "); 1698 cisc->dump(); 1699 } 1700 #endif 1701 } else { 1702 #ifndef PRODUCT 1703 if( TraceCISCSpill ) { 1704 tty->print(" using reg-instr: "); 1705 n->dump(); 1706 } 1707 #endif 1708 ++_unused_cisc_instructions; // input can be on stack 1709 } 1710 } 1711 1712 } // End of for all instructions 1713 1714 } // End of for all blocks 1715 } 1716 1717 // Helper to stretch above; recursively discover the base Node for a 1718 // given derived Node. Easy for AddP-related machine nodes, but needs 1719 // to be recursive for derived Phis. 1720 Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { 1721 // See if already computed; if so return it 1722 if( derived_base_map[derived->_idx] ) 1723 return derived_base_map[derived->_idx]; 1724 1725 // See if this happens to be a base. 1726 // NOTE: we use TypePtr instead of TypeOopPtr because we can have 1727 // pointers derived from NULL! These are always along paths that 1728 // can't happen at run-time but the optimizer cannot deduce it so 1729 // we have to handle it gracefully. 1730 assert(!derived->bottom_type()->isa_narrowoop() || 1731 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1732 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1733 // If its an OOP with a non-zero offset, then it is derived. 1734 if( tj == NULL || tj->_offset == 0 ) { 1735 derived_base_map[derived->_idx] = derived; 1736 return derived; 1737 } 1738 // Derived is NULL+offset? Base is NULL! 1739 if( derived->is_Con() ) { 1740 Node *base = _matcher.mach_null(); 1741 assert(base != NULL, "sanity"); 1742 if (base->in(0) == NULL) { 1743 // Initialize it once and make it shared: 1744 // set control to _root and place it into Start block 1745 // (where top() node is placed). 1746 base->init_req(0, _cfg.get_root_node()); 1747 Block *startb = _cfg.get_block_for_node(C->top()); 1748 uint node_pos = startb->find_node(C->top()); 1749 startb->insert_node(base, node_pos); 1750 _cfg.map_node_to_block(base, startb); 1751 assert(_lrg_map.live_range_id(base) == 0, "should not have LRG yet"); 1752 1753 // The loadConP0 might have projection nodes depending on architecture 1754 // Add the projection nodes to the CFG 1755 for (DUIterator_Fast imax, i = base->fast_outs(imax); i < imax; i++) { 1756 Node* use = base->fast_out(i); 1757 if (use->is_MachProj()) { 1758 startb->insert_node(use, ++node_pos); 1759 _cfg.map_node_to_block(use, startb); 1760 new_lrg(use, maxlrg++); 1761 } 1762 } 1763 } 1764 if (_lrg_map.live_range_id(base) == 0) { 1765 new_lrg(base, maxlrg++); 1766 } 1767 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"); 1768 derived_base_map[derived->_idx] = base; 1769 return base; 1770 } 1771 1772 // Check for AddP-related opcodes 1773 if (!derived->is_Phi()) { 1774 assert(derived->as_Mach()->ideal_Opcode() == Op_AddP, "but is: %s", derived->Name()); 1775 Node *base = derived->in(AddPNode::Base); 1776 derived_base_map[derived->_idx] = base; 1777 return base; 1778 } 1779 1780 // Recursively find bases for Phis. 1781 // First check to see if we can avoid a base Phi here. 1782 Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); 1783 uint i; 1784 for( i = 2; i < derived->req(); i++ ) 1785 if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) 1786 break; 1787 // Went to the end without finding any different bases? 1788 if( i == derived->req() ) { // No need for a base Phi here 1789 derived_base_map[derived->_idx] = base; 1790 return base; 1791 } 1792 1793 // Now we see we need a base-Phi here to merge the bases 1794 const Type *t = base->bottom_type(); 1795 base = new PhiNode( derived->in(0), t ); 1796 for( i = 1; i < derived->req(); i++ ) { 1797 base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); 1798 t = t->meet(base->in(i)->bottom_type()); 1799 } 1800 base->as_Phi()->set_type(t); 1801 1802 // Search the current block for an existing base-Phi 1803 Block *b = _cfg.get_block_for_node(derived); 1804 for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi 1805 Node *phi = b->get_node(i); 1806 if( !phi->is_Phi() ) { // Found end of Phis with no match? 1807 b->insert_node(base, i); // Must insert created Phi here as base 1808 _cfg.map_node_to_block(base, b); 1809 new_lrg(base,maxlrg++); 1810 break; 1811 } 1812 // See if Phi matches. 1813 uint j; 1814 for( j = 1; j < base->req(); j++ ) 1815 if( phi->in(j) != base->in(j) && 1816 !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs 1817 break; 1818 if( j == base->req() ) { // All inputs match? 1819 base = phi; // Then use existing 'phi' and drop 'base' 1820 break; 1821 } 1822 } 1823 1824 1825 // Cache info for later passes 1826 derived_base_map[derived->_idx] = base; 1827 return base; 1828 } 1829 1830 // At each Safepoint, insert extra debug edges for each pair of derived value/ 1831 // base pointer that is live across the Safepoint for oopmap building. The 1832 // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the 1833 // required edge set. 1834 bool PhaseChaitin::stretch_base_pointer_live_ranges(ResourceArea *a) { 1835 int must_recompute_live = false; 1836 uint maxlrg = _lrg_map.max_lrg_id(); 1837 Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); 1838 memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); 1839 1840 // For all blocks in RPO do... 1841 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 1842 Block* block = _cfg.get_block(i); 1843 // Note use of deep-copy constructor. I cannot hammer the original 1844 // liveout bits, because they are needed by the following coalesce pass. 1845 IndexSet liveout(_live->live(block)); 1846 1847 for (uint j = block->end_idx() + 1; j > 1; j--) { 1848 Node* n = block->get_node(j - 1); 1849 1850 // Pre-split compares of loop-phis. Loop-phis form a cycle we would 1851 // like to see in the same register. Compare uses the loop-phi and so 1852 // extends its live range BUT cannot be part of the cycle. If this 1853 // extended live range overlaps with the update of the loop-phi value 1854 // we need both alive at the same time -- which requires at least 1 1855 // copy. But because Intel has only 2-address registers we end up with 1856 // at least 2 copies, one before the loop-phi update instruction and 1857 // one after. Instead we split the input to the compare just after the 1858 // phi. 1859 if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { 1860 Node *phi = n->in(1); 1861 if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { 1862 Block *phi_block = _cfg.get_block_for_node(phi); 1863 if (_cfg.get_block_for_node(phi_block->pred(2)) == block) { 1864 const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; 1865 Node *spill = new MachSpillCopyNode(MachSpillCopyNode::LoopPhiInput, phi, *mask, *mask); 1866 insert_proj( phi_block, 1, spill, maxlrg++ ); 1867 n->set_req(1,spill); 1868 must_recompute_live = true; 1869 } 1870 } 1871 } 1872 1873 // Get value being defined 1874 uint lidx = _lrg_map.live_range_id(n); 1875 // Ignore the occasional brand-new live range 1876 if (lidx && lidx < _lrg_map.max_lrg_id()) { 1877 // Remove from live-out set 1878 liveout.remove(lidx); 1879 1880 // Copies do not define a new value and so do not interfere. 1881 // Remove the copies source from the liveout set before interfering. 1882 uint idx = n->is_Copy(); 1883 if (idx) { 1884 liveout.remove(_lrg_map.live_range_id(n->in(idx))); 1885 } 1886 } 1887 1888 // Found a safepoint? 1889 JVMState *jvms = n->jvms(); 1890 if( jvms ) { 1891 // Now scan for a live derived pointer 1892 IndexSetIterator elements(&liveout); 1893 uint neighbor; 1894 while ((neighbor = elements.next()) != 0) { 1895 // Find reaching DEF for base and derived values 1896 // This works because we are still in SSA during this call. 1897 Node *derived = lrgs(neighbor)._def; 1898 const TypePtr *tj = derived->bottom_type()->isa_ptr(); 1899 assert(!derived->bottom_type()->isa_narrowoop() || 1900 derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); 1901 // If its an OOP with a non-zero offset, then it is derived. 1902 if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { 1903 Node *base = find_base_for_derived(derived_base_map, derived, maxlrg); 1904 assert(base->_idx < _lrg_map.size(), ""); 1905 // Add reaching DEFs of derived pointer and base pointer as a 1906 // pair of inputs 1907 n->add_req(derived); 1908 n->add_req(base); 1909 1910 // See if the base pointer is already live to this point. 1911 // Since I'm working on the SSA form, live-ness amounts to 1912 // reaching def's. So if I find the base's live range then 1913 // I know the base's def reaches here. 1914 if ((_lrg_map.live_range_id(base) >= _lrg_map.max_lrg_id() || // (Brand new base (hence not live) or 1915 !liveout.member(_lrg_map.live_range_id(base))) && // not live) AND 1916 (_lrg_map.live_range_id(base) > 0) && // not a constant 1917 _cfg.get_block_for_node(base) != block) { // base not def'd in blk) 1918 // Base pointer is not currently live. Since I stretched 1919 // the base pointer to here and it crosses basic-block 1920 // boundaries, the global live info is now incorrect. 1921 // Recompute live. 1922 must_recompute_live = true; 1923 } // End of if base pointer is not live to debug info 1924 } 1925 } // End of scan all live data for derived ptrs crossing GC point 1926 } // End of if found a GC point 1927 1928 // Make all inputs live 1929 if (!n->is_Phi()) { // Phi function uses come from prior block 1930 for (uint k = 1; k < n->req(); k++) { 1931 uint lidx = _lrg_map.live_range_id(n->in(k)); 1932 if (lidx < _lrg_map.max_lrg_id()) { 1933 liveout.insert(lidx); 1934 } 1935 } 1936 } 1937 1938 } // End of forall instructions in block 1939 liveout.clear(); // Free the memory used by liveout. 1940 1941 } // End of forall blocks 1942 _lrg_map.set_max_lrg_id(maxlrg); 1943 1944 // If I created a new live range I need to recompute live 1945 if (maxlrg != _ifg->_maxlrg) { 1946 must_recompute_live = true; 1947 } 1948 1949 return must_recompute_live != 0; 1950 } 1951 1952 // Extend the node to LRG mapping 1953 1954 void PhaseChaitin::add_reference(const Node *node, const Node *old_node) { 1955 _lrg_map.extend(node->_idx, _lrg_map.live_range_id(old_node)); 1956 } 1957 1958 #ifndef PRODUCT 1959 void PhaseChaitin::dump(const Node *n) const { 1960 uint r = (n->_idx < _lrg_map.size()) ? _lrg_map.find_const(n) : 0; 1961 tty->print("L%d",r); 1962 if (r && n->Opcode() != Op_Phi) { 1963 if( _node_regs ) { // Got a post-allocation copy of allocation? 1964 tty->print("["); 1965 OptoReg::Name second = get_reg_second(n); 1966 if( OptoReg::is_valid(second) ) { 1967 if( OptoReg::is_reg(second) ) 1968 tty->print("%s:",Matcher::regName[second]); 1969 else 1970 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); 1971 } 1972 OptoReg::Name first = get_reg_first(n); 1973 if( OptoReg::is_reg(first) ) 1974 tty->print("%s]",Matcher::regName[first]); 1975 else 1976 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); 1977 } else 1978 n->out_RegMask().dump(); 1979 } 1980 tty->print("/N%d\t",n->_idx); 1981 tty->print("%s === ", n->Name()); 1982 uint k; 1983 for (k = 0; k < n->req(); k++) { 1984 Node *m = n->in(k); 1985 if (!m) { 1986 tty->print("_ "); 1987 } 1988 else { 1989 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 1990 tty->print("L%d",r); 1991 // Data MultiNode's can have projections with no real registers. 1992 // Don't die while dumping them. 1993 int op = n->Opcode(); 1994 if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { 1995 if( _node_regs ) { 1996 tty->print("["); 1997 OptoReg::Name second = get_reg_second(n->in(k)); 1998 if( OptoReg::is_valid(second) ) { 1999 if( OptoReg::is_reg(second) ) 2000 tty->print("%s:",Matcher::regName[second]); 2001 else 2002 tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), 2003 reg2offset_unchecked(second)); 2004 } 2005 OptoReg::Name first = get_reg_first(n->in(k)); 2006 if( OptoReg::is_reg(first) ) 2007 tty->print("%s]",Matcher::regName[first]); 2008 else 2009 tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), 2010 reg2offset_unchecked(first)); 2011 } else 2012 n->in_RegMask(k).dump(); 2013 } 2014 tty->print("/N%d ",m->_idx); 2015 } 2016 } 2017 if( k < n->len() && n->in(k) ) tty->print("| "); 2018 for( ; k < n->len(); k++ ) { 2019 Node *m = n->in(k); 2020 if(!m) { 2021 break; 2022 } 2023 uint r = (m->_idx < _lrg_map.size()) ? _lrg_map.find_const(m) : 0; 2024 tty->print("L%d",r); 2025 tty->print("/N%d ",m->_idx); 2026 } 2027 if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); 2028 else n->dump_spec(tty); 2029 if( _spilled_once.test(n->_idx ) ) { 2030 tty->print(" Spill_1"); 2031 if( _spilled_twice.test(n->_idx ) ) 2032 tty->print(" Spill_2"); 2033 } 2034 tty->print("\n"); 2035 } 2036 2037 void PhaseChaitin::dump(const Block *b) const { 2038 b->dump_head(&_cfg); 2039 2040 // For all instructions 2041 for( uint j = 0; j < b->number_of_nodes(); j++ ) 2042 dump(b->get_node(j)); 2043 // Print live-out info at end of block 2044 if( _live ) { 2045 tty->print("Liveout: "); 2046 IndexSet *live = _live->live(b); 2047 IndexSetIterator elements(live); 2048 tty->print("{"); 2049 uint i; 2050 while ((i = elements.next()) != 0) { 2051 tty->print("L%d ", _lrg_map.find_const(i)); 2052 } 2053 tty->print_cr("}"); 2054 } 2055 tty->print("\n"); 2056 } 2057 2058 void PhaseChaitin::dump() const { 2059 tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", 2060 _matcher._new_SP, _framesize ); 2061 2062 // For all blocks 2063 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2064 dump(_cfg.get_block(i)); 2065 } 2066 // End of per-block dump 2067 tty->print("\n"); 2068 2069 if (!_ifg) { 2070 tty->print("(No IFG.)\n"); 2071 return; 2072 } 2073 2074 // Dump LRG array 2075 tty->print("--- Live RanGe Array ---\n"); 2076 for (uint i2 = 1; i2 < _lrg_map.max_lrg_id(); i2++) { 2077 tty->print("L%d: ",i2); 2078 if (i2 < _ifg->_maxlrg) { 2079 lrgs(i2).dump(); 2080 } 2081 else { 2082 tty->print_cr("new LRG"); 2083 } 2084 } 2085 tty->cr(); 2086 2087 // Dump lo-degree list 2088 tty->print("Lo degree: "); 2089 for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) 2090 tty->print("L%d ",i3); 2091 tty->cr(); 2092 2093 // Dump lo-stk-degree list 2094 tty->print("Lo stk degree: "); 2095 for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) 2096 tty->print("L%d ",i4); 2097 tty->cr(); 2098 2099 // Dump lo-degree list 2100 tty->print("Hi degree: "); 2101 for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) 2102 tty->print("L%d ",i5); 2103 tty->cr(); 2104 } 2105 2106 void PhaseChaitin::dump_degree_lists() const { 2107 // Dump lo-degree list 2108 tty->print("Lo degree: "); 2109 for( uint i = _lo_degree; i; i = lrgs(i)._next ) 2110 tty->print("L%d ",i); 2111 tty->cr(); 2112 2113 // Dump lo-stk-degree list 2114 tty->print("Lo stk degree: "); 2115 for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) 2116 tty->print("L%d ",i2); 2117 tty->cr(); 2118 2119 // Dump lo-degree list 2120 tty->print("Hi degree: "); 2121 for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) 2122 tty->print("L%d ",i3); 2123 tty->cr(); 2124 } 2125 2126 void PhaseChaitin::dump_simplified() const { 2127 tty->print("Simplified: "); 2128 for( uint i = _simplified; i; i = lrgs(i)._next ) 2129 tty->print("L%d ",i); 2130 tty->cr(); 2131 } 2132 2133 static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { 2134 if ((int)reg < 0) 2135 sprintf(buf, "<OptoReg::%d>", (int)reg); 2136 else if (OptoReg::is_reg(reg)) 2137 strcpy(buf, Matcher::regName[reg]); 2138 else 2139 sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), 2140 pc->reg2offset(reg)); 2141 return buf+strlen(buf); 2142 } 2143 2144 // Dump a register name into a buffer. Be intelligent if we get called 2145 // before allocation is complete. 2146 char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { 2147 if( this == NULL ) { // Not got anything? 2148 sprintf(buf,"N%d",n->_idx); // Then use Node index 2149 } else if( _node_regs ) { 2150 // Post allocation, use direct mappings, no LRG info available 2151 print_reg( get_reg_first(n), this, buf ); 2152 } else { 2153 uint lidx = _lrg_map.find_const(n); // Grab LRG number 2154 if( !_ifg ) { 2155 sprintf(buf,"L%d",lidx); // No register binding yet 2156 } else if( !lidx ) { // Special, not allocated value 2157 strcpy(buf,"Special"); 2158 } else { 2159 if (lrgs(lidx)._is_vector) { 2160 if (lrgs(lidx).mask().is_bound_set(lrgs(lidx).num_regs())) 2161 print_reg( lrgs(lidx).reg(), this, buf ); // a bound machine register 2162 else 2163 sprintf(buf,"L%d",lidx); // No register binding yet 2164 } else if( (lrgs(lidx).num_regs() == 1) 2165 ? lrgs(lidx).mask().is_bound1() 2166 : lrgs(lidx).mask().is_bound_pair() ) { 2167 // Hah! We have a bound machine register 2168 print_reg( lrgs(lidx).reg(), this, buf ); 2169 } else { 2170 sprintf(buf,"L%d",lidx); // No register binding yet 2171 } 2172 } 2173 } 2174 return buf+strlen(buf); 2175 } 2176 2177 void PhaseChaitin::dump_for_spill_split_recycle() const { 2178 if( WizardMode && (PrintCompilation || PrintOpto) ) { 2179 // Display which live ranges need to be split and the allocator's state 2180 tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); 2181 for (uint bidx = 1; bidx < _lrg_map.max_lrg_id(); bidx++) { 2182 if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { 2183 tty->print("L%d: ", bidx); 2184 lrgs(bidx).dump(); 2185 } 2186 } 2187 tty->cr(); 2188 dump(); 2189 } 2190 } 2191 2192 void PhaseChaitin::dump_frame() const { 2193 const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); 2194 const TypeTuple *domain = C->tf()->domain(); 2195 const int argcnt = domain->cnt() - TypeFunc::Parms; 2196 2197 // Incoming arguments in registers dump 2198 for( int k = 0; k < argcnt; k++ ) { 2199 OptoReg::Name parmreg = _matcher._parm_regs[k].first(); 2200 if( OptoReg::is_reg(parmreg)) { 2201 const char *reg_name = OptoReg::regname(parmreg); 2202 tty->print("#r%3.3d %s", parmreg, reg_name); 2203 parmreg = _matcher._parm_regs[k].second(); 2204 if( OptoReg::is_reg(parmreg)) { 2205 tty->print(":%s", OptoReg::regname(parmreg)); 2206 } 2207 tty->print(" : parm %d: ", k); 2208 domain->field_at(k + TypeFunc::Parms)->dump(); 2209 tty->cr(); 2210 } 2211 } 2212 2213 // Check for un-owned padding above incoming args 2214 OptoReg::Name reg = _matcher._new_SP; 2215 if( reg > _matcher._in_arg_limit ) { 2216 reg = OptoReg::add(reg, -1); 2217 tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); 2218 } 2219 2220 // Incoming argument area dump 2221 OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); 2222 while( reg > begin_in_arg ) { 2223 reg = OptoReg::add(reg, -1); 2224 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2225 int j; 2226 for( j = 0; j < argcnt; j++) { 2227 if( _matcher._parm_regs[j].first() == reg || 2228 _matcher._parm_regs[j].second() == reg ) { 2229 tty->print("parm %d: ",j); 2230 domain->field_at(j + TypeFunc::Parms)->dump(); 2231 tty->cr(); 2232 break; 2233 } 2234 } 2235 if( j >= argcnt ) 2236 tty->print_cr("HOLE, owned by SELF"); 2237 } 2238 2239 // Old outgoing preserve area 2240 while( reg > _matcher._old_SP ) { 2241 reg = OptoReg::add(reg, -1); 2242 tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); 2243 } 2244 2245 // Old SP 2246 tty->print_cr("# -- Old %s -- Framesize: %d --",fp, 2247 reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); 2248 2249 // Preserve area dump 2250 int fixed_slots = C->fixed_slots(); 2251 OptoReg::Name begin_in_preserve = OptoReg::add(_matcher._old_SP, -(int)C->in_preserve_stack_slots()); 2252 OptoReg::Name return_addr = _matcher.return_addr(); 2253 2254 reg = OptoReg::add(reg, -1); 2255 while (OptoReg::is_stack(reg)) { 2256 tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); 2257 if (return_addr == reg) { 2258 tty->print_cr("return address"); 2259 } else if (reg >= begin_in_preserve) { 2260 // Preserved slots are present on x86 2261 if (return_addr == OptoReg::add(reg, VMRegImpl::slots_per_word)) 2262 tty->print_cr("saved fp register"); 2263 else if (return_addr == OptoReg::add(reg, 2*VMRegImpl::slots_per_word) && 2264 VerifyStackAtCalls) 2265 tty->print_cr("0xBADB100D +VerifyStackAtCalls"); 2266 else 2267 tty->print_cr("in_preserve"); 2268 } else if ((int)OptoReg::reg2stack(reg) < fixed_slots) { 2269 tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); 2270 } else { 2271 tty->print_cr("pad2, stack alignment"); 2272 } 2273 reg = OptoReg::add(reg, -1); 2274 } 2275 2276 // Spill area dump 2277 reg = OptoReg::add(_matcher._new_SP, _framesize ); 2278 while( reg > _matcher._out_arg_limit ) { 2279 reg = OptoReg::add(reg, -1); 2280 tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); 2281 } 2282 2283 // Outgoing argument area dump 2284 while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { 2285 reg = OptoReg::add(reg, -1); 2286 tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); 2287 } 2288 2289 // Outgoing new preserve area 2290 while( reg > _matcher._new_SP ) { 2291 reg = OptoReg::add(reg, -1); 2292 tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); 2293 } 2294 tty->print_cr("#"); 2295 } 2296 2297 void PhaseChaitin::dump_bb( uint pre_order ) const { 2298 tty->print_cr("---dump of B%d---",pre_order); 2299 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2300 Block* block = _cfg.get_block(i); 2301 if (block->_pre_order == pre_order) { 2302 dump(block); 2303 } 2304 } 2305 } 2306 2307 void PhaseChaitin::dump_lrg( uint lidx, bool defs_only ) const { 2308 tty->print_cr("---dump of L%d---",lidx); 2309 2310 if (_ifg) { 2311 if (lidx >= _lrg_map.max_lrg_id()) { 2312 tty->print("Attempt to print live range index beyond max live range.\n"); 2313 return; 2314 } 2315 tty->print("L%d: ",lidx); 2316 if (lidx < _ifg->_maxlrg) { 2317 lrgs(lidx).dump(); 2318 } else { 2319 tty->print_cr("new LRG"); 2320 } 2321 } 2322 if( _ifg && lidx < _ifg->_maxlrg) { 2323 tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); 2324 _ifg->neighbors(lidx)->dump(); 2325 tty->cr(); 2326 } 2327 // For all blocks 2328 for (uint i = 0; i < _cfg.number_of_blocks(); i++) { 2329 Block* block = _cfg.get_block(i); 2330 int dump_once = 0; 2331 2332 // For all instructions 2333 for( uint j = 0; j < block->number_of_nodes(); j++ ) { 2334 Node *n = block->get_node(j); 2335 if (_lrg_map.find_const(n) == lidx) { 2336 if (!dump_once++) { 2337 tty->cr(); 2338 block->dump_head(&_cfg); 2339 } 2340 dump(n); 2341 continue; 2342 } 2343 if (!defs_only) { 2344 uint cnt = n->req(); 2345 for( uint k = 1; k < cnt; k++ ) { 2346 Node *m = n->in(k); 2347 if (!m) { 2348 continue; // be robust in the dumper 2349 } 2350 if (_lrg_map.find_const(m) == lidx) { 2351 if (!dump_once++) { 2352 tty->cr(); 2353 block->dump_head(&_cfg); 2354 } 2355 dump(n); 2356 } 2357 } 2358 } 2359 } 2360 } // End of per-block dump 2361 tty->cr(); 2362 } 2363 #endif // not PRODUCT 2364 2365 int PhaseChaitin::_final_loads = 0; 2366 int PhaseChaitin::_final_stores = 0; 2367 int PhaseChaitin::_final_memoves= 0; 2368 int PhaseChaitin::_final_copies = 0; 2369 double PhaseChaitin::_final_load_cost = 0; 2370 double PhaseChaitin::_final_store_cost = 0; 2371 double PhaseChaitin::_final_memove_cost= 0; 2372 double PhaseChaitin::_final_copy_cost = 0; 2373 int PhaseChaitin::_conserv_coalesce = 0; 2374 int PhaseChaitin::_conserv_coalesce_pair = 0; 2375 int PhaseChaitin::_conserv_coalesce_trie = 0; 2376 int PhaseChaitin::_conserv_coalesce_quad = 0; 2377 int PhaseChaitin::_post_alloc = 0; 2378 int PhaseChaitin::_lost_opp_pp_coalesce = 0; 2379 int PhaseChaitin::_lost_opp_cflow_coalesce = 0; 2380 int PhaseChaitin::_used_cisc_instructions = 0; 2381 int PhaseChaitin::_unused_cisc_instructions = 0; 2382 int PhaseChaitin::_allocator_attempts = 0; 2383 int PhaseChaitin::_allocator_successes = 0; 2384 2385 #ifndef PRODUCT 2386 uint PhaseChaitin::_high_pressure = 0; 2387 uint PhaseChaitin::_low_pressure = 0; 2388 2389 void PhaseChaitin::print_chaitin_statistics() { 2390 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); 2391 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); 2392 tty->print_cr("Adjusted spill cost = %7.0f.", 2393 _final_load_cost*4.0 + _final_store_cost * 2.0 + 2394 _final_copy_cost*1.0 + _final_memove_cost*12.0); 2395 tty->print("Conservatively coalesced %d copies, %d pairs", 2396 _conserv_coalesce, _conserv_coalesce_pair); 2397 if( _conserv_coalesce_trie || _conserv_coalesce_quad ) 2398 tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); 2399 tty->print_cr(", %d post alloc.", _post_alloc); 2400 if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) 2401 tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", 2402 _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); 2403 if( _used_cisc_instructions || _unused_cisc_instructions ) 2404 tty->print_cr("Used cisc instruction %d, remained in register %d", 2405 _used_cisc_instructions, _unused_cisc_instructions); 2406 if( _allocator_successes != 0 ) 2407 tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); 2408 tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); 2409 } 2410 #endif // not PRODUCT