Fri, 12 Feb 2010 15:27:36 -0800
Merge
duke@435 | 1 | /* |
xdono@1014 | 2 | * Copyright 2000-2009 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_chaitin.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | //============================================================================= |
duke@435 | 29 | |
duke@435 | 30 | #ifndef PRODUCT |
duke@435 | 31 | void LRG::dump( ) const { |
duke@435 | 32 | ttyLocker ttyl; |
duke@435 | 33 | tty->print("%d ",num_regs()); |
duke@435 | 34 | _mask.dump(); |
duke@435 | 35 | if( _msize_valid ) { |
duke@435 | 36 | if( mask_size() == compute_mask_size() ) tty->print(", #%d ",_mask_size); |
duke@435 | 37 | else tty->print(", #!!!_%d_vs_%d ",_mask_size,_mask.Size()); |
duke@435 | 38 | } else { |
duke@435 | 39 | tty->print(", #?(%d) ",_mask.Size()); |
duke@435 | 40 | } |
duke@435 | 41 | |
duke@435 | 42 | tty->print("EffDeg: "); |
duke@435 | 43 | if( _degree_valid ) tty->print( "%d ", _eff_degree ); |
duke@435 | 44 | else tty->print("? "); |
duke@435 | 45 | |
never@730 | 46 | if( is_multidef() ) { |
duke@435 | 47 | tty->print("MultiDef "); |
duke@435 | 48 | if (_defs != NULL) { |
duke@435 | 49 | tty->print("("); |
duke@435 | 50 | for (int i = 0; i < _defs->length(); i++) { |
duke@435 | 51 | tty->print("N%d ", _defs->at(i)->_idx); |
duke@435 | 52 | } |
duke@435 | 53 | tty->print(") "); |
duke@435 | 54 | } |
duke@435 | 55 | } |
duke@435 | 56 | else if( _def == 0 ) tty->print("Dead "); |
duke@435 | 57 | else tty->print("Def: N%d ",_def->_idx); |
duke@435 | 58 | |
duke@435 | 59 | tty->print("Cost:%4.2g Area:%4.2g Score:%4.2g ",_cost,_area, score()); |
duke@435 | 60 | // Flags |
duke@435 | 61 | if( _is_oop ) tty->print("Oop "); |
duke@435 | 62 | if( _is_float ) tty->print("Float "); |
duke@435 | 63 | if( _was_spilled1 ) tty->print("Spilled "); |
duke@435 | 64 | if( _was_spilled2 ) tty->print("Spilled2 "); |
duke@435 | 65 | if( _direct_conflict ) tty->print("Direct_conflict "); |
duke@435 | 66 | if( _fat_proj ) tty->print("Fat "); |
duke@435 | 67 | if( _was_lo ) tty->print("Lo "); |
duke@435 | 68 | if( _has_copy ) tty->print("Copy "); |
duke@435 | 69 | if( _at_risk ) tty->print("Risk "); |
duke@435 | 70 | |
duke@435 | 71 | if( _must_spill ) tty->print("Must_spill "); |
duke@435 | 72 | if( _is_bound ) tty->print("Bound "); |
duke@435 | 73 | if( _msize_valid ) { |
duke@435 | 74 | if( _degree_valid && lo_degree() ) tty->print("Trivial "); |
duke@435 | 75 | } |
duke@435 | 76 | |
duke@435 | 77 | tty->cr(); |
duke@435 | 78 | } |
duke@435 | 79 | #endif |
duke@435 | 80 | |
duke@435 | 81 | //------------------------------score------------------------------------------ |
duke@435 | 82 | // Compute score from cost and area. Low score is best to spill. |
duke@435 | 83 | static double raw_score( double cost, double area ) { |
duke@435 | 84 | return cost - (area*RegisterCostAreaRatio) * 1.52588e-5; |
duke@435 | 85 | } |
duke@435 | 86 | |
duke@435 | 87 | double LRG::score() const { |
duke@435 | 88 | // Scale _area by RegisterCostAreaRatio/64K then subtract from cost. |
duke@435 | 89 | // Bigger area lowers score, encourages spilling this live range. |
duke@435 | 90 | // Bigger cost raise score, prevents spilling this live range. |
duke@435 | 91 | // (Note: 1/65536 is the magic constant below; I dont trust the C optimizer |
duke@435 | 92 | // to turn a divide by a constant into a multiply by the reciprical). |
duke@435 | 93 | double score = raw_score( _cost, _area); |
duke@435 | 94 | |
duke@435 | 95 | // Account for area. Basically, LRGs covering large areas are better |
duke@435 | 96 | // to spill because more other LRGs get freed up. |
duke@435 | 97 | if( _area == 0.0 ) // No area? Then no progress to spill |
duke@435 | 98 | return 1e35; |
duke@435 | 99 | |
duke@435 | 100 | if( _was_spilled2 ) // If spilled once before, we are unlikely |
duke@435 | 101 | return score + 1e30; // to make progress again. |
duke@435 | 102 | |
duke@435 | 103 | if( _cost >= _area*3.0 ) // Tiny area relative to cost |
duke@435 | 104 | return score + 1e17; // Probably no progress to spill |
duke@435 | 105 | |
duke@435 | 106 | if( (_cost+_cost) >= _area*3.0 ) // Small area relative to cost |
duke@435 | 107 | return score + 1e10; // Likely no progress to spill |
duke@435 | 108 | |
duke@435 | 109 | return score; |
duke@435 | 110 | } |
duke@435 | 111 | |
duke@435 | 112 | //------------------------------LRG_List--------------------------------------- |
duke@435 | 113 | LRG_List::LRG_List( uint max ) : _cnt(max), _max(max), _lidxs(NEW_RESOURCE_ARRAY(uint,max)) { |
duke@435 | 114 | memset( _lidxs, 0, sizeof(uint)*max ); |
duke@435 | 115 | } |
duke@435 | 116 | |
duke@435 | 117 | void LRG_List::extend( uint nidx, uint lidx ) { |
duke@435 | 118 | _nesting.check(); |
duke@435 | 119 | if( nidx >= _max ) { |
duke@435 | 120 | uint size = 16; |
duke@435 | 121 | while( size <= nidx ) size <<=1; |
duke@435 | 122 | _lidxs = REALLOC_RESOURCE_ARRAY( uint, _lidxs, _max, size ); |
duke@435 | 123 | _max = size; |
duke@435 | 124 | } |
duke@435 | 125 | while( _cnt <= nidx ) |
duke@435 | 126 | _lidxs[_cnt++] = 0; |
duke@435 | 127 | _lidxs[nidx] = lidx; |
duke@435 | 128 | } |
duke@435 | 129 | |
duke@435 | 130 | #define NUMBUCKS 3 |
duke@435 | 131 | |
duke@435 | 132 | //------------------------------Chaitin---------------------------------------- |
duke@435 | 133 | PhaseChaitin::PhaseChaitin(uint unique, PhaseCFG &cfg, Matcher &matcher) |
duke@435 | 134 | : PhaseRegAlloc(unique, cfg, matcher, |
duke@435 | 135 | #ifndef PRODUCT |
duke@435 | 136 | print_chaitin_statistics |
duke@435 | 137 | #else |
duke@435 | 138 | NULL |
duke@435 | 139 | #endif |
duke@435 | 140 | ), |
duke@435 | 141 | _names(unique), _uf_map(unique), |
duke@435 | 142 | _maxlrg(0), _live(0), |
duke@435 | 143 | _spilled_once(Thread::current()->resource_area()), |
duke@435 | 144 | _spilled_twice(Thread::current()->resource_area()), |
duke@435 | 145 | _lo_degree(0), _lo_stk_degree(0), _hi_degree(0), _simplified(0), |
duke@435 | 146 | _oldphi(unique) |
duke@435 | 147 | #ifndef PRODUCT |
duke@435 | 148 | , _trace_spilling(TraceSpilling || C->method_has_option("TraceSpilling")) |
duke@435 | 149 | #endif |
duke@435 | 150 | { |
duke@435 | 151 | NOT_PRODUCT( Compile::TracePhase t3("ctorChaitin", &_t_ctorChaitin, TimeCompiler); ) |
kvn@1108 | 152 | |
kvn@1108 | 153 | _high_frequency_lrg = MIN2(float(OPTO_LRG_HIGH_FREQ), _cfg._outer_loop_freq); |
kvn@1108 | 154 | |
duke@435 | 155 | uint i,j; |
duke@435 | 156 | // Build a list of basic blocks, sorted by frequency |
duke@435 | 157 | _blks = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks ); |
duke@435 | 158 | // Experiment with sorting strategies to speed compilation |
duke@435 | 159 | double cutoff = BLOCK_FREQUENCY(1.0); // Cutoff for high frequency bucket |
duke@435 | 160 | Block **buckets[NUMBUCKS]; // Array of buckets |
duke@435 | 161 | uint buckcnt[NUMBUCKS]; // Array of bucket counters |
duke@435 | 162 | double buckval[NUMBUCKS]; // Array of bucket value cutoffs |
duke@435 | 163 | for( i = 0; i < NUMBUCKS; i++ ) { |
duke@435 | 164 | buckets[i] = NEW_RESOURCE_ARRAY( Block *, _cfg._num_blocks ); |
duke@435 | 165 | buckcnt[i] = 0; |
duke@435 | 166 | // Bump by three orders of magnitude each time |
duke@435 | 167 | cutoff *= 0.001; |
duke@435 | 168 | buckval[i] = cutoff; |
duke@435 | 169 | for( j = 0; j < _cfg._num_blocks; j++ ) { |
duke@435 | 170 | buckets[i][j] = NULL; |
duke@435 | 171 | } |
duke@435 | 172 | } |
duke@435 | 173 | // Sort blocks into buckets |
duke@435 | 174 | for( i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 175 | for( j = 0; j < NUMBUCKS; j++ ) { |
duke@435 | 176 | if( (j == NUMBUCKS-1) || (_cfg._blocks[i]->_freq > buckval[j]) ) { |
duke@435 | 177 | // Assign block to end of list for appropriate bucket |
duke@435 | 178 | buckets[j][buckcnt[j]++] = _cfg._blocks[i]; |
duke@435 | 179 | break; // kick out of inner loop |
duke@435 | 180 | } |
duke@435 | 181 | } |
duke@435 | 182 | } |
duke@435 | 183 | // Dump buckets into final block array |
duke@435 | 184 | uint blkcnt = 0; |
duke@435 | 185 | for( i = 0; i < NUMBUCKS; i++ ) { |
duke@435 | 186 | for( j = 0; j < buckcnt[i]; j++ ) { |
duke@435 | 187 | _blks[blkcnt++] = buckets[i][j]; |
duke@435 | 188 | } |
duke@435 | 189 | } |
duke@435 | 190 | |
duke@435 | 191 | assert(blkcnt == _cfg._num_blocks, "Block array not totally filled"); |
duke@435 | 192 | } |
duke@435 | 193 | |
duke@435 | 194 | void PhaseChaitin::Register_Allocate() { |
duke@435 | 195 | |
duke@435 | 196 | // Above the OLD FP (and in registers) are the incoming arguments. Stack |
duke@435 | 197 | // slots in this area are called "arg_slots". Above the NEW FP (and in |
duke@435 | 198 | // registers) is the outgoing argument area; above that is the spill/temp |
duke@435 | 199 | // area. These are all "frame_slots". Arg_slots start at the zero |
duke@435 | 200 | // stack_slots and count up to the known arg_size. Frame_slots start at |
duke@435 | 201 | // the stack_slot #arg_size and go up. After allocation I map stack |
duke@435 | 202 | // slots to actual offsets. Stack-slots in the arg_slot area are biased |
duke@435 | 203 | // by the frame_size; stack-slots in the frame_slot area are biased by 0. |
duke@435 | 204 | |
duke@435 | 205 | _trip_cnt = 0; |
duke@435 | 206 | _alternate = 0; |
duke@435 | 207 | _matcher._allocation_started = true; |
duke@435 | 208 | |
duke@435 | 209 | ResourceArea live_arena; // Arena for liveness & IFG info |
duke@435 | 210 | ResourceMark rm(&live_arena); |
duke@435 | 211 | |
duke@435 | 212 | // Need live-ness for the IFG; need the IFG for coalescing. If the |
duke@435 | 213 | // liveness is JUST for coalescing, then I can get some mileage by renaming |
duke@435 | 214 | // all copy-related live ranges low and then using the max copy-related |
duke@435 | 215 | // live range as a cut-off for LIVE and the IFG. In other words, I can |
duke@435 | 216 | // build a subset of LIVE and IFG just for copies. |
duke@435 | 217 | PhaseLive live(_cfg,_names,&live_arena); |
duke@435 | 218 | |
duke@435 | 219 | // Need IFG for coalescing and coloring |
duke@435 | 220 | PhaseIFG ifg( &live_arena ); |
duke@435 | 221 | _ifg = &ifg; |
duke@435 | 222 | |
duke@435 | 223 | if (C->unique() > _names.Size()) _names.extend(C->unique()-1, 0); |
duke@435 | 224 | |
duke@435 | 225 | // Come out of SSA world to the Named world. Assign (virtual) registers to |
duke@435 | 226 | // Nodes. Use the same register for all inputs and the output of PhiNodes |
duke@435 | 227 | // - effectively ending SSA form. This requires either coalescing live |
duke@435 | 228 | // ranges or inserting copies. For the moment, we insert "virtual copies" |
duke@435 | 229 | // - we pretend there is a copy prior to each Phi in predecessor blocks. |
duke@435 | 230 | // We will attempt to coalesce such "virtual copies" before we manifest |
duke@435 | 231 | // them for real. |
duke@435 | 232 | de_ssa(); |
duke@435 | 233 | |
kvn@1001 | 234 | #ifdef ASSERT |
kvn@1001 | 235 | // Veify the graph before RA. |
kvn@1001 | 236 | verify(&live_arena); |
kvn@1001 | 237 | #endif |
kvn@1001 | 238 | |
duke@435 | 239 | { |
duke@435 | 240 | NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) |
duke@435 | 241 | _live = NULL; // Mark live as being not available |
duke@435 | 242 | rm.reset_to_mark(); // Reclaim working storage |
duke@435 | 243 | IndexSet::reset_memory(C, &live_arena); |
duke@435 | 244 | ifg.init(_maxlrg); // Empty IFG |
duke@435 | 245 | gather_lrg_masks( false ); // Collect LRG masks |
duke@435 | 246 | live.compute( _maxlrg ); // Compute liveness |
duke@435 | 247 | _live = &live; // Mark LIVE as being available |
duke@435 | 248 | } |
duke@435 | 249 | |
duke@435 | 250 | // Base pointers are currently "used" by instructions which define new |
duke@435 | 251 | // derived pointers. This makes base pointers live up to the where the |
duke@435 | 252 | // derived pointer is made, but not beyond. Really, they need to be live |
duke@435 | 253 | // across any GC point where the derived value is live. So this code looks |
duke@435 | 254 | // at all the GC points, and "stretches" the live range of any base pointer |
duke@435 | 255 | // to the GC point. |
duke@435 | 256 | if( stretch_base_pointer_live_ranges(&live_arena) ) { |
duke@435 | 257 | NOT_PRODUCT( Compile::TracePhase t3("computeLive (sbplr)", &_t_computeLive, TimeCompiler); ) |
duke@435 | 258 | // Since some live range stretched, I need to recompute live |
duke@435 | 259 | _live = NULL; |
duke@435 | 260 | rm.reset_to_mark(); // Reclaim working storage |
duke@435 | 261 | IndexSet::reset_memory(C, &live_arena); |
duke@435 | 262 | ifg.init(_maxlrg); |
duke@435 | 263 | gather_lrg_masks( false ); |
duke@435 | 264 | live.compute( _maxlrg ); |
duke@435 | 265 | _live = &live; |
duke@435 | 266 | } |
duke@435 | 267 | // Create the interference graph using virtual copies |
duke@435 | 268 | build_ifg_virtual( ); // Include stack slots this time |
duke@435 | 269 | |
duke@435 | 270 | // Aggressive (but pessimistic) copy coalescing. |
duke@435 | 271 | // This pass works on virtual copies. Any virtual copies which are not |
duke@435 | 272 | // coalesced get manifested as actual copies |
duke@435 | 273 | { |
duke@435 | 274 | // The IFG is/was triangular. I am 'squaring it up' so Union can run |
duke@435 | 275 | // faster. Union requires a 'for all' operation which is slow on the |
duke@435 | 276 | // triangular adjacency matrix (quick reminder: the IFG is 'sparse' - |
duke@435 | 277 | // meaning I can visit all the Nodes neighbors less than a Node in time |
duke@435 | 278 | // O(# of neighbors), but I have to visit all the Nodes greater than a |
duke@435 | 279 | // given Node and search them for an instance, i.e., time O(#MaxLRG)). |
duke@435 | 280 | _ifg->SquareUp(); |
duke@435 | 281 | |
duke@435 | 282 | PhaseAggressiveCoalesce coalesce( *this ); |
duke@435 | 283 | coalesce.coalesce_driver( ); |
duke@435 | 284 | // Insert un-coalesced copies. Visit all Phis. Where inputs to a Phi do |
duke@435 | 285 | // not match the Phi itself, insert a copy. |
duke@435 | 286 | coalesce.insert_copies(_matcher); |
duke@435 | 287 | } |
duke@435 | 288 | |
duke@435 | 289 | // After aggressive coalesce, attempt a first cut at coloring. |
duke@435 | 290 | // To color, we need the IFG and for that we need LIVE. |
duke@435 | 291 | { |
duke@435 | 292 | NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) |
duke@435 | 293 | _live = NULL; |
duke@435 | 294 | rm.reset_to_mark(); // Reclaim working storage |
duke@435 | 295 | IndexSet::reset_memory(C, &live_arena); |
duke@435 | 296 | ifg.init(_maxlrg); |
duke@435 | 297 | gather_lrg_masks( true ); |
duke@435 | 298 | live.compute( _maxlrg ); |
duke@435 | 299 | _live = &live; |
duke@435 | 300 | } |
duke@435 | 301 | |
duke@435 | 302 | // Build physical interference graph |
duke@435 | 303 | uint must_spill = 0; |
duke@435 | 304 | must_spill = build_ifg_physical( &live_arena ); |
duke@435 | 305 | // If we have a guaranteed spill, might as well spill now |
duke@435 | 306 | if( must_spill ) { |
duke@435 | 307 | if( !_maxlrg ) return; |
duke@435 | 308 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit) |
duke@435 | 309 | C->check_node_count(10*must_spill, "out of nodes before split"); |
duke@435 | 310 | if (C->failing()) return; |
duke@435 | 311 | _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere |
duke@435 | 312 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) |
duke@435 | 313 | // or we failed to split |
duke@435 | 314 | C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after physical split"); |
duke@435 | 315 | if (C->failing()) return; |
duke@435 | 316 | |
duke@435 | 317 | NOT_PRODUCT( C->verify_graph_edges(); ) |
duke@435 | 318 | |
duke@435 | 319 | compact(); // Compact LRGs; return new lower max lrg |
duke@435 | 320 | |
duke@435 | 321 | { |
duke@435 | 322 | NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) |
duke@435 | 323 | _live = NULL; |
duke@435 | 324 | rm.reset_to_mark(); // Reclaim working storage |
duke@435 | 325 | IndexSet::reset_memory(C, &live_arena); |
duke@435 | 326 | ifg.init(_maxlrg); // Build a new interference graph |
duke@435 | 327 | gather_lrg_masks( true ); // Collect intersect mask |
duke@435 | 328 | live.compute( _maxlrg ); // Compute LIVE |
duke@435 | 329 | _live = &live; |
duke@435 | 330 | } |
duke@435 | 331 | build_ifg_physical( &live_arena ); |
duke@435 | 332 | _ifg->SquareUp(); |
duke@435 | 333 | _ifg->Compute_Effective_Degree(); |
duke@435 | 334 | // Only do conservative coalescing if requested |
duke@435 | 335 | if( OptoCoalesce ) { |
duke@435 | 336 | // Conservative (and pessimistic) copy coalescing of those spills |
duke@435 | 337 | PhaseConservativeCoalesce coalesce( *this ); |
duke@435 | 338 | // If max live ranges greater than cutoff, don't color the stack. |
duke@435 | 339 | // This cutoff can be larger than below since it is only done once. |
duke@435 | 340 | coalesce.coalesce_driver( ); |
duke@435 | 341 | } |
duke@435 | 342 | compress_uf_map_for_nodes(); |
duke@435 | 343 | |
duke@435 | 344 | #ifdef ASSERT |
kvn@1001 | 345 | verify(&live_arena, true); |
duke@435 | 346 | #endif |
duke@435 | 347 | } else { |
duke@435 | 348 | ifg.SquareUp(); |
duke@435 | 349 | ifg.Compute_Effective_Degree(); |
duke@435 | 350 | #ifdef ASSERT |
duke@435 | 351 | set_was_low(); |
duke@435 | 352 | #endif |
duke@435 | 353 | } |
duke@435 | 354 | |
duke@435 | 355 | // Prepare for Simplify & Select |
duke@435 | 356 | cache_lrg_info(); // Count degree of LRGs |
duke@435 | 357 | |
duke@435 | 358 | // Simplify the InterFerence Graph by removing LRGs of low degree. |
duke@435 | 359 | // LRGs of low degree are trivially colorable. |
duke@435 | 360 | Simplify(); |
duke@435 | 361 | |
duke@435 | 362 | // Select colors by re-inserting LRGs back into the IFG in reverse order. |
duke@435 | 363 | // Return whether or not something spills. |
duke@435 | 364 | uint spills = Select( ); |
duke@435 | 365 | |
duke@435 | 366 | // If we spill, split and recycle the entire thing |
duke@435 | 367 | while( spills ) { |
duke@435 | 368 | if( _trip_cnt++ > 24 ) { |
duke@435 | 369 | DEBUG_ONLY( dump_for_spill_split_recycle(); ) |
duke@435 | 370 | if( _trip_cnt > 27 ) { |
duke@435 | 371 | C->record_method_not_compilable("failed spill-split-recycle sanity check"); |
duke@435 | 372 | return; |
duke@435 | 373 | } |
duke@435 | 374 | } |
duke@435 | 375 | |
duke@435 | 376 | if( !_maxlrg ) return; |
duke@435 | 377 | _maxlrg = Split( _maxlrg ); // Split spilling LRG everywhere |
duke@435 | 378 | // Bail out if unique gets too large (ie - unique > MaxNodeLimit - 2*NodeLimitFudgeFactor) |
duke@435 | 379 | C->check_node_count(2*NodeLimitFudgeFactor, "out of nodes after split"); |
duke@435 | 380 | if (C->failing()) return; |
duke@435 | 381 | |
duke@435 | 382 | compact(); // Compact LRGs; return new lower max lrg |
duke@435 | 383 | |
duke@435 | 384 | // Nuke the live-ness and interference graph and LiveRanGe info |
duke@435 | 385 | { |
duke@435 | 386 | NOT_PRODUCT( Compile::TracePhase t3("computeLive", &_t_computeLive, TimeCompiler); ) |
duke@435 | 387 | _live = NULL; |
duke@435 | 388 | rm.reset_to_mark(); // Reclaim working storage |
duke@435 | 389 | IndexSet::reset_memory(C, &live_arena); |
duke@435 | 390 | ifg.init(_maxlrg); |
duke@435 | 391 | |
duke@435 | 392 | // Create LiveRanGe array. |
duke@435 | 393 | // Intersect register masks for all USEs and DEFs |
duke@435 | 394 | gather_lrg_masks( true ); |
duke@435 | 395 | live.compute( _maxlrg ); |
duke@435 | 396 | _live = &live; |
duke@435 | 397 | } |
duke@435 | 398 | must_spill = build_ifg_physical( &live_arena ); |
duke@435 | 399 | _ifg->SquareUp(); |
duke@435 | 400 | _ifg->Compute_Effective_Degree(); |
duke@435 | 401 | |
duke@435 | 402 | // Only do conservative coalescing if requested |
duke@435 | 403 | if( OptoCoalesce ) { |
duke@435 | 404 | // Conservative (and pessimistic) copy coalescing |
duke@435 | 405 | PhaseConservativeCoalesce coalesce( *this ); |
duke@435 | 406 | // Check for few live ranges determines how aggressive coalesce is. |
duke@435 | 407 | coalesce.coalesce_driver( ); |
duke@435 | 408 | } |
duke@435 | 409 | compress_uf_map_for_nodes(); |
duke@435 | 410 | #ifdef ASSERT |
kvn@1001 | 411 | verify(&live_arena, true); |
duke@435 | 412 | #endif |
duke@435 | 413 | cache_lrg_info(); // Count degree of LRGs |
duke@435 | 414 | |
duke@435 | 415 | // Simplify the InterFerence Graph by removing LRGs of low degree. |
duke@435 | 416 | // LRGs of low degree are trivially colorable. |
duke@435 | 417 | Simplify(); |
duke@435 | 418 | |
duke@435 | 419 | // Select colors by re-inserting LRGs back into the IFG in reverse order. |
duke@435 | 420 | // Return whether or not something spills. |
duke@435 | 421 | spills = Select( ); |
duke@435 | 422 | } |
duke@435 | 423 | |
duke@435 | 424 | // Count number of Simplify-Select trips per coloring success. |
duke@435 | 425 | _allocator_attempts += _trip_cnt + 1; |
duke@435 | 426 | _allocator_successes += 1; |
duke@435 | 427 | |
duke@435 | 428 | // Peephole remove copies |
duke@435 | 429 | post_allocate_copy_removal(); |
duke@435 | 430 | |
kvn@1001 | 431 | #ifdef ASSERT |
kvn@1001 | 432 | // Veify the graph after RA. |
kvn@1001 | 433 | verify(&live_arena); |
kvn@1001 | 434 | #endif |
kvn@1001 | 435 | |
duke@435 | 436 | // max_reg is past the largest *register* used. |
duke@435 | 437 | // Convert that to a frame_slot number. |
duke@435 | 438 | if( _max_reg <= _matcher._new_SP ) |
duke@435 | 439 | _framesize = C->out_preserve_stack_slots(); |
duke@435 | 440 | else _framesize = _max_reg -_matcher._new_SP; |
duke@435 | 441 | assert((int)(_matcher._new_SP+_framesize) >= (int)_matcher._out_arg_limit, "framesize must be large enough"); |
duke@435 | 442 | |
duke@435 | 443 | // This frame must preserve the required fp alignment |
never@854 | 444 | _framesize = round_to(_framesize, Matcher::stack_alignment_in_slots()); |
duke@435 | 445 | assert( _framesize >= 0 && _framesize <= 1000000, "sanity check" ); |
duke@435 | 446 | #ifndef PRODUCT |
duke@435 | 447 | _total_framesize += _framesize; |
duke@435 | 448 | if( (int)_framesize > _max_framesize ) |
duke@435 | 449 | _max_framesize = _framesize; |
duke@435 | 450 | #endif |
duke@435 | 451 | |
duke@435 | 452 | // Convert CISC spills |
duke@435 | 453 | fixup_spills(); |
duke@435 | 454 | |
duke@435 | 455 | // Log regalloc results |
duke@435 | 456 | CompileLog* log = Compile::current()->log(); |
duke@435 | 457 | if (log != NULL) { |
duke@435 | 458 | log->elem("regalloc attempts='%d' success='%d'", _trip_cnt, !C->failing()); |
duke@435 | 459 | } |
duke@435 | 460 | |
duke@435 | 461 | if (C->failing()) return; |
duke@435 | 462 | |
duke@435 | 463 | NOT_PRODUCT( C->verify_graph_edges(); ) |
duke@435 | 464 | |
duke@435 | 465 | // Move important info out of the live_arena to longer lasting storage. |
duke@435 | 466 | alloc_node_regs(_names.Size()); |
duke@435 | 467 | for( uint i=0; i < _names.Size(); i++ ) { |
duke@435 | 468 | if( _names[i] ) { // Live range associated with Node? |
duke@435 | 469 | LRG &lrg = lrgs( _names[i] ); |
duke@435 | 470 | if( lrg.num_regs() == 1 ) { |
duke@435 | 471 | _node_regs[i].set1( lrg.reg() ); |
duke@435 | 472 | } else { // Must be a register-pair |
duke@435 | 473 | if( !lrg._fat_proj ) { // Must be aligned adjacent register pair |
duke@435 | 474 | // Live ranges record the highest register in their mask. |
duke@435 | 475 | // We want the low register for the AD file writer's convenience. |
duke@435 | 476 | _node_regs[i].set2( OptoReg::add(lrg.reg(),-1) ); |
duke@435 | 477 | } else { // Misaligned; extract 2 bits |
duke@435 | 478 | OptoReg::Name hi = lrg.reg(); // Get hi register |
duke@435 | 479 | lrg.Remove(hi); // Yank from mask |
duke@435 | 480 | int lo = lrg.mask().find_first_elem(); // Find lo |
duke@435 | 481 | _node_regs[i].set_pair( hi, lo ); |
duke@435 | 482 | } |
duke@435 | 483 | } |
duke@435 | 484 | if( lrg._is_oop ) _node_oops.set(i); |
duke@435 | 485 | } else { |
duke@435 | 486 | _node_regs[i].set_bad(); |
duke@435 | 487 | } |
duke@435 | 488 | } |
duke@435 | 489 | |
duke@435 | 490 | // Done! |
duke@435 | 491 | _live = NULL; |
duke@435 | 492 | _ifg = NULL; |
duke@435 | 493 | C->set_indexSet_arena(NULL); // ResourceArea is at end of scope |
duke@435 | 494 | } |
duke@435 | 495 | |
duke@435 | 496 | //------------------------------de_ssa----------------------------------------- |
duke@435 | 497 | void PhaseChaitin::de_ssa() { |
duke@435 | 498 | // Set initial Names for all Nodes. Most Nodes get the virtual register |
duke@435 | 499 | // number. A few get the ZERO live range number. These do not |
duke@435 | 500 | // get allocated, but instead rely on correct scheduling to ensure that |
duke@435 | 501 | // only one instance is simultaneously live at a time. |
duke@435 | 502 | uint lr_counter = 1; |
duke@435 | 503 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 504 | Block *b = _cfg._blocks[i]; |
duke@435 | 505 | uint cnt = b->_nodes.size(); |
duke@435 | 506 | |
duke@435 | 507 | // Handle all the normal Nodes in the block |
duke@435 | 508 | for( uint j = 0; j < cnt; j++ ) { |
duke@435 | 509 | Node *n = b->_nodes[j]; |
duke@435 | 510 | // Pre-color to the zero live range, or pick virtual register |
duke@435 | 511 | const RegMask &rm = n->out_RegMask(); |
duke@435 | 512 | _names.map( n->_idx, rm.is_NotEmpty() ? lr_counter++ : 0 ); |
duke@435 | 513 | } |
duke@435 | 514 | } |
duke@435 | 515 | // Reset the Union-Find mapping to be identity |
duke@435 | 516 | reset_uf_map(lr_counter); |
duke@435 | 517 | } |
duke@435 | 518 | |
duke@435 | 519 | |
duke@435 | 520 | //------------------------------gather_lrg_masks------------------------------- |
duke@435 | 521 | // Gather LiveRanGe information, including register masks. Modification of |
duke@435 | 522 | // cisc spillable in_RegMasks should not be done before AggressiveCoalesce. |
duke@435 | 523 | void PhaseChaitin::gather_lrg_masks( bool after_aggressive ) { |
duke@435 | 524 | |
duke@435 | 525 | // Nail down the frame pointer live range |
duke@435 | 526 | uint fp_lrg = n2lidx(_cfg._root->in(1)->in(TypeFunc::FramePtr)); |
duke@435 | 527 | lrgs(fp_lrg)._cost += 1e12; // Cost is infinite |
duke@435 | 528 | |
duke@435 | 529 | // For all blocks |
duke@435 | 530 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 531 | Block *b = _cfg._blocks[i]; |
duke@435 | 532 | |
duke@435 | 533 | // For all instructions |
duke@435 | 534 | for( uint j = 1; j < b->_nodes.size(); j++ ) { |
duke@435 | 535 | Node *n = b->_nodes[j]; |
duke@435 | 536 | uint input_edge_start =1; // Skip control most nodes |
duke@435 | 537 | if( n->is_Mach() ) input_edge_start = n->as_Mach()->oper_input_base(); |
duke@435 | 538 | uint idx = n->is_Copy(); |
duke@435 | 539 | |
duke@435 | 540 | // Get virtual register number, same as LiveRanGe index |
duke@435 | 541 | uint vreg = n2lidx(n); |
duke@435 | 542 | LRG &lrg = lrgs(vreg); |
duke@435 | 543 | if( vreg ) { // No vreg means un-allocable (e.g. memory) |
duke@435 | 544 | |
duke@435 | 545 | // Collect has-copy bit |
duke@435 | 546 | if( idx ) { |
duke@435 | 547 | lrg._has_copy = 1; |
duke@435 | 548 | uint clidx = n2lidx(n->in(idx)); |
duke@435 | 549 | LRG ©_src = lrgs(clidx); |
duke@435 | 550 | copy_src._has_copy = 1; |
duke@435 | 551 | } |
duke@435 | 552 | |
duke@435 | 553 | // Check for float-vs-int live range (used in register-pressure |
duke@435 | 554 | // calculations) |
duke@435 | 555 | const Type *n_type = n->bottom_type(); |
duke@435 | 556 | if( n_type->is_floatingpoint() ) |
duke@435 | 557 | lrg._is_float = 1; |
duke@435 | 558 | |
duke@435 | 559 | // Check for twice prior spilling. Once prior spilling might have |
duke@435 | 560 | // spilled 'soft', 2nd prior spill should have spilled 'hard' and |
duke@435 | 561 | // further spilling is unlikely to make progress. |
duke@435 | 562 | if( _spilled_once.test(n->_idx) ) { |
duke@435 | 563 | lrg._was_spilled1 = 1; |
duke@435 | 564 | if( _spilled_twice.test(n->_idx) ) |
duke@435 | 565 | lrg._was_spilled2 = 1; |
duke@435 | 566 | } |
duke@435 | 567 | |
duke@435 | 568 | #ifndef PRODUCT |
duke@435 | 569 | if (trace_spilling() && lrg._def != NULL) { |
duke@435 | 570 | // collect defs for MultiDef printing |
duke@435 | 571 | if (lrg._defs == NULL) { |
duke@435 | 572 | lrg._defs = new (_ifg->_arena) GrowableArray<Node*>(); |
duke@435 | 573 | lrg._defs->append(lrg._def); |
duke@435 | 574 | } |
duke@435 | 575 | lrg._defs->append(n); |
duke@435 | 576 | } |
duke@435 | 577 | #endif |
duke@435 | 578 | |
duke@435 | 579 | // Check for a single def LRG; these can spill nicely |
duke@435 | 580 | // via rematerialization. Flag as NULL for no def found |
duke@435 | 581 | // yet, or 'n' for single def or -1 for many defs. |
duke@435 | 582 | lrg._def = lrg._def ? NodeSentinel : n; |
duke@435 | 583 | |
duke@435 | 584 | // Limit result register mask to acceptable registers |
duke@435 | 585 | const RegMask &rm = n->out_RegMask(); |
duke@435 | 586 | lrg.AND( rm ); |
duke@435 | 587 | // Check for bound register masks |
duke@435 | 588 | const RegMask &lrgmask = lrg.mask(); |
duke@435 | 589 | if( lrgmask.is_bound1() || lrgmask.is_bound2() ) |
duke@435 | 590 | lrg._is_bound = 1; |
duke@435 | 591 | |
duke@435 | 592 | // Check for maximum frequency value |
duke@435 | 593 | if( lrg._maxfreq < b->_freq ) |
duke@435 | 594 | lrg._maxfreq = b->_freq; |
duke@435 | 595 | |
duke@435 | 596 | int ireg = n->ideal_reg(); |
duke@435 | 597 | assert( !n->bottom_type()->isa_oop_ptr() || ireg == Op_RegP, |
duke@435 | 598 | "oops must be in Op_RegP's" ); |
duke@435 | 599 | // Check for oop-iness, or long/double |
duke@435 | 600 | // Check for multi-kill projection |
duke@435 | 601 | switch( ireg ) { |
duke@435 | 602 | case MachProjNode::fat_proj: |
duke@435 | 603 | // Fat projections have size equal to number of registers killed |
duke@435 | 604 | lrg.set_num_regs(rm.Size()); |
duke@435 | 605 | lrg.set_reg_pressure(lrg.num_regs()); |
duke@435 | 606 | lrg._fat_proj = 1; |
duke@435 | 607 | lrg._is_bound = 1; |
duke@435 | 608 | break; |
duke@435 | 609 | case Op_RegP: |
duke@435 | 610 | #ifdef _LP64 |
duke@435 | 611 | lrg.set_num_regs(2); // Size is 2 stack words |
duke@435 | 612 | #else |
duke@435 | 613 | lrg.set_num_regs(1); // Size is 1 stack word |
duke@435 | 614 | #endif |
duke@435 | 615 | // Register pressure is tracked relative to the maximum values |
duke@435 | 616 | // suggested for that platform, INTPRESSURE and FLOATPRESSURE, |
duke@435 | 617 | // and relative to other types which compete for the same regs. |
duke@435 | 618 | // |
duke@435 | 619 | // The following table contains suggested values based on the |
duke@435 | 620 | // architectures as defined in each .ad file. |
duke@435 | 621 | // INTPRESSURE and FLOATPRESSURE may be tuned differently for |
duke@435 | 622 | // compile-speed or performance. |
duke@435 | 623 | // Note1: |
duke@435 | 624 | // SPARC and SPARCV9 reg_pressures are at 2 instead of 1 |
duke@435 | 625 | // since .ad registers are defined as high and low halves. |
duke@435 | 626 | // These reg_pressure values remain compatible with the code |
duke@435 | 627 | // in is_high_pressure() which relates get_invalid_mask_size(), |
duke@435 | 628 | // Block::_reg_pressure and INTPRESSURE, FLOATPRESSURE. |
duke@435 | 629 | // Note2: |
duke@435 | 630 | // SPARC -d32 has 24 registers available for integral values, |
duke@435 | 631 | // but only 10 of these are safe for 64-bit longs. |
duke@435 | 632 | // Using set_reg_pressure(2) for both int and long means |
duke@435 | 633 | // the allocator will believe it can fit 26 longs into |
duke@435 | 634 | // registers. Using 2 for longs and 1 for ints means the |
duke@435 | 635 | // allocator will attempt to put 52 integers into registers. |
duke@435 | 636 | // The settings below limit this problem to methods with |
duke@435 | 637 | // many long values which are being run on 32-bit SPARC. |
duke@435 | 638 | // |
duke@435 | 639 | // ------------------- reg_pressure -------------------- |
duke@435 | 640 | // Each entry is reg_pressure_per_value,number_of_regs |
duke@435 | 641 | // RegL RegI RegFlags RegF RegD INTPRESSURE FLOATPRESSURE |
duke@435 | 642 | // IA32 2 1 1 1 1 6 6 |
duke@435 | 643 | // IA64 1 1 1 1 1 50 41 |
duke@435 | 644 | // SPARC 2 2 2 2 2 48 (24) 52 (26) |
duke@435 | 645 | // SPARCV9 2 2 2 2 2 48 (24) 52 (26) |
duke@435 | 646 | // AMD64 1 1 1 1 1 14 15 |
duke@435 | 647 | // ----------------------------------------------------- |
duke@435 | 648 | #if defined(SPARC) |
duke@435 | 649 | lrg.set_reg_pressure(2); // use for v9 as well |
duke@435 | 650 | #else |
duke@435 | 651 | lrg.set_reg_pressure(1); // normally one value per register |
duke@435 | 652 | #endif |
duke@435 | 653 | if( n_type->isa_oop_ptr() ) { |
duke@435 | 654 | lrg._is_oop = 1; |
duke@435 | 655 | } |
duke@435 | 656 | break; |
duke@435 | 657 | case Op_RegL: // Check for long or double |
duke@435 | 658 | case Op_RegD: |
duke@435 | 659 | lrg.set_num_regs(2); |
duke@435 | 660 | // Define platform specific register pressure |
duke@435 | 661 | #ifdef SPARC |
duke@435 | 662 | lrg.set_reg_pressure(2); |
duke@435 | 663 | #elif defined(IA32) |
duke@435 | 664 | if( ireg == Op_RegL ) { |
duke@435 | 665 | lrg.set_reg_pressure(2); |
duke@435 | 666 | } else { |
duke@435 | 667 | lrg.set_reg_pressure(1); |
duke@435 | 668 | } |
duke@435 | 669 | #else |
duke@435 | 670 | lrg.set_reg_pressure(1); // normally one value per register |
duke@435 | 671 | #endif |
duke@435 | 672 | // If this def of a double forces a mis-aligned double, |
duke@435 | 673 | // flag as '_fat_proj' - really flag as allowing misalignment |
duke@435 | 674 | // AND changes how we count interferences. A mis-aligned |
duke@435 | 675 | // double can interfere with TWO aligned pairs, or effectively |
duke@435 | 676 | // FOUR registers! |
duke@435 | 677 | if( rm.is_misaligned_Pair() ) { |
duke@435 | 678 | lrg._fat_proj = 1; |
duke@435 | 679 | lrg._is_bound = 1; |
duke@435 | 680 | } |
duke@435 | 681 | break; |
duke@435 | 682 | case Op_RegF: |
duke@435 | 683 | case Op_RegI: |
coleenp@548 | 684 | case Op_RegN: |
duke@435 | 685 | case Op_RegFlags: |
duke@435 | 686 | case 0: // not an ideal register |
duke@435 | 687 | lrg.set_num_regs(1); |
duke@435 | 688 | #ifdef SPARC |
duke@435 | 689 | lrg.set_reg_pressure(2); |
duke@435 | 690 | #else |
duke@435 | 691 | lrg.set_reg_pressure(1); |
duke@435 | 692 | #endif |
duke@435 | 693 | break; |
duke@435 | 694 | default: |
duke@435 | 695 | ShouldNotReachHere(); |
duke@435 | 696 | } |
duke@435 | 697 | } |
duke@435 | 698 | |
duke@435 | 699 | // Now do the same for inputs |
duke@435 | 700 | uint cnt = n->req(); |
duke@435 | 701 | // Setup for CISC SPILLING |
duke@435 | 702 | uint inp = (uint)AdlcVMDeps::Not_cisc_spillable; |
duke@435 | 703 | if( UseCISCSpill && after_aggressive ) { |
duke@435 | 704 | inp = n->cisc_operand(); |
duke@435 | 705 | if( inp != (uint)AdlcVMDeps::Not_cisc_spillable ) |
duke@435 | 706 | // Convert operand number to edge index number |
duke@435 | 707 | inp = n->as_Mach()->operand_index(inp); |
duke@435 | 708 | } |
duke@435 | 709 | // Prepare register mask for each input |
duke@435 | 710 | for( uint k = input_edge_start; k < cnt; k++ ) { |
duke@435 | 711 | uint vreg = n2lidx(n->in(k)); |
duke@435 | 712 | if( !vreg ) continue; |
duke@435 | 713 | |
duke@435 | 714 | // If this instruction is CISC Spillable, add the flags |
duke@435 | 715 | // bit to its appropriate input |
duke@435 | 716 | if( UseCISCSpill && after_aggressive && inp == k ) { |
duke@435 | 717 | #ifndef PRODUCT |
duke@435 | 718 | if( TraceCISCSpill ) { |
duke@435 | 719 | tty->print(" use_cisc_RegMask: "); |
duke@435 | 720 | n->dump(); |
duke@435 | 721 | } |
duke@435 | 722 | #endif |
duke@435 | 723 | n->as_Mach()->use_cisc_RegMask(); |
duke@435 | 724 | } |
duke@435 | 725 | |
duke@435 | 726 | LRG &lrg = lrgs(vreg); |
duke@435 | 727 | // // Testing for floating point code shape |
duke@435 | 728 | // Node *test = n->in(k); |
duke@435 | 729 | // if( test->is_Mach() ) { |
duke@435 | 730 | // MachNode *m = test->as_Mach(); |
duke@435 | 731 | // int op = m->ideal_Opcode(); |
duke@435 | 732 | // if (n->is_Call() && (op == Op_AddF || op == Op_MulF) ) { |
duke@435 | 733 | // int zzz = 1; |
duke@435 | 734 | // } |
duke@435 | 735 | // } |
duke@435 | 736 | |
duke@435 | 737 | // Limit result register mask to acceptable registers. |
duke@435 | 738 | // Do not limit registers from uncommon uses before |
duke@435 | 739 | // AggressiveCoalesce. This effectively pre-virtual-splits |
duke@435 | 740 | // around uncommon uses of common defs. |
duke@435 | 741 | const RegMask &rm = n->in_RegMask(k); |
duke@435 | 742 | if( !after_aggressive && |
duke@435 | 743 | _cfg._bbs[n->in(k)->_idx]->_freq > 1000*b->_freq ) { |
duke@435 | 744 | // Since we are BEFORE aggressive coalesce, leave the register |
duke@435 | 745 | // mask untrimmed by the call. This encourages more coalescing. |
duke@435 | 746 | // Later, AFTER aggressive, this live range will have to spill |
duke@435 | 747 | // but the spiller handles slow-path calls very nicely. |
duke@435 | 748 | } else { |
duke@435 | 749 | lrg.AND( rm ); |
duke@435 | 750 | } |
duke@435 | 751 | // Check for bound register masks |
duke@435 | 752 | const RegMask &lrgmask = lrg.mask(); |
duke@435 | 753 | if( lrgmask.is_bound1() || lrgmask.is_bound2() ) |
duke@435 | 754 | lrg._is_bound = 1; |
duke@435 | 755 | // If this use of a double forces a mis-aligned double, |
duke@435 | 756 | // flag as '_fat_proj' - really flag as allowing misalignment |
duke@435 | 757 | // AND changes how we count interferences. A mis-aligned |
duke@435 | 758 | // double can interfere with TWO aligned pairs, or effectively |
duke@435 | 759 | // FOUR registers! |
duke@435 | 760 | if( lrg.num_regs() == 2 && !lrg._fat_proj && rm.is_misaligned_Pair() ) { |
duke@435 | 761 | lrg._fat_proj = 1; |
duke@435 | 762 | lrg._is_bound = 1; |
duke@435 | 763 | } |
duke@435 | 764 | // if the LRG is an unaligned pair, we will have to spill |
duke@435 | 765 | // so clear the LRG's register mask if it is not already spilled |
duke@435 | 766 | if ( !n->is_SpillCopy() && |
never@730 | 767 | (lrg._def == NULL || lrg.is_multidef() || !lrg._def->is_SpillCopy()) && |
duke@435 | 768 | lrgmask.is_misaligned_Pair()) { |
duke@435 | 769 | lrg.Clear(); |
duke@435 | 770 | } |
duke@435 | 771 | |
duke@435 | 772 | // Check for maximum frequency value |
duke@435 | 773 | if( lrg._maxfreq < b->_freq ) |
duke@435 | 774 | lrg._maxfreq = b->_freq; |
duke@435 | 775 | |
duke@435 | 776 | } // End for all allocated inputs |
duke@435 | 777 | } // end for all instructions |
duke@435 | 778 | } // end for all blocks |
duke@435 | 779 | |
duke@435 | 780 | // Final per-liverange setup |
duke@435 | 781 | for( uint i2=0; i2<_maxlrg; i2++ ) { |
duke@435 | 782 | LRG &lrg = lrgs(i2); |
duke@435 | 783 | if( lrg.num_regs() == 2 && !lrg._fat_proj ) |
duke@435 | 784 | lrg.ClearToPairs(); |
duke@435 | 785 | lrg.compute_set_mask_size(); |
duke@435 | 786 | if( lrg.not_free() ) { // Handle case where we lose from the start |
duke@435 | 787 | lrg.set_reg(OptoReg::Name(LRG::SPILL_REG)); |
duke@435 | 788 | lrg._direct_conflict = 1; |
duke@435 | 789 | } |
duke@435 | 790 | lrg.set_degree(0); // no neighbors in IFG yet |
duke@435 | 791 | } |
duke@435 | 792 | } |
duke@435 | 793 | |
duke@435 | 794 | //------------------------------set_was_low------------------------------------ |
duke@435 | 795 | // Set the was-lo-degree bit. Conservative coalescing should not change the |
duke@435 | 796 | // colorability of the graph. If any live range was of low-degree before |
duke@435 | 797 | // coalescing, it should Simplify. This call sets the was-lo-degree bit. |
duke@435 | 798 | // The bit is checked in Simplify. |
duke@435 | 799 | void PhaseChaitin::set_was_low() { |
duke@435 | 800 | #ifdef ASSERT |
duke@435 | 801 | for( uint i = 1; i < _maxlrg; i++ ) { |
duke@435 | 802 | int size = lrgs(i).num_regs(); |
duke@435 | 803 | uint old_was_lo = lrgs(i)._was_lo; |
duke@435 | 804 | lrgs(i)._was_lo = 0; |
duke@435 | 805 | if( lrgs(i).lo_degree() ) { |
duke@435 | 806 | lrgs(i)._was_lo = 1; // Trivially of low degree |
duke@435 | 807 | } else { // Else check the Brigg's assertion |
duke@435 | 808 | // Brigg's observation is that the lo-degree neighbors of a |
duke@435 | 809 | // hi-degree live range will not interfere with the color choices |
duke@435 | 810 | // of said hi-degree live range. The Simplify reverse-stack-coloring |
duke@435 | 811 | // order takes care of the details. Hence you do not have to count |
duke@435 | 812 | // low-degree neighbors when determining if this guy colors. |
duke@435 | 813 | int briggs_degree = 0; |
duke@435 | 814 | IndexSet *s = _ifg->neighbors(i); |
duke@435 | 815 | IndexSetIterator elements(s); |
duke@435 | 816 | uint lidx; |
duke@435 | 817 | while((lidx = elements.next()) != 0) { |
duke@435 | 818 | if( !lrgs(lidx).lo_degree() ) |
duke@435 | 819 | briggs_degree += MAX2(size,lrgs(lidx).num_regs()); |
duke@435 | 820 | } |
duke@435 | 821 | if( briggs_degree < lrgs(i).degrees_of_freedom() ) |
duke@435 | 822 | lrgs(i)._was_lo = 1; // Low degree via the briggs assertion |
duke@435 | 823 | } |
duke@435 | 824 | assert(old_was_lo <= lrgs(i)._was_lo, "_was_lo may not decrease"); |
duke@435 | 825 | } |
duke@435 | 826 | #endif |
duke@435 | 827 | } |
duke@435 | 828 | |
duke@435 | 829 | #define REGISTER_CONSTRAINED 16 |
duke@435 | 830 | |
duke@435 | 831 | //------------------------------cache_lrg_info--------------------------------- |
duke@435 | 832 | // Compute cost/area ratio, in case we spill. Build the lo-degree list. |
duke@435 | 833 | void PhaseChaitin::cache_lrg_info( ) { |
duke@435 | 834 | |
duke@435 | 835 | for( uint i = 1; i < _maxlrg; i++ ) { |
duke@435 | 836 | LRG &lrg = lrgs(i); |
duke@435 | 837 | |
duke@435 | 838 | // Check for being of low degree: means we can be trivially colored. |
duke@435 | 839 | // Low degree, dead or must-spill guys just get to simplify right away |
duke@435 | 840 | if( lrg.lo_degree() || |
duke@435 | 841 | !lrg.alive() || |
duke@435 | 842 | lrg._must_spill ) { |
duke@435 | 843 | // Split low degree list into those guys that must get a |
duke@435 | 844 | // register and those that can go to register or stack. |
duke@435 | 845 | // The idea is LRGs that can go register or stack color first when |
duke@435 | 846 | // they have a good chance of getting a register. The register-only |
duke@435 | 847 | // lo-degree live ranges always get a register. |
duke@435 | 848 | OptoReg::Name hi_reg = lrg.mask().find_last_elem(); |
duke@435 | 849 | if( OptoReg::is_stack(hi_reg)) { // Can go to stack? |
duke@435 | 850 | lrg._next = _lo_stk_degree; |
duke@435 | 851 | _lo_stk_degree = i; |
duke@435 | 852 | } else { |
duke@435 | 853 | lrg._next = _lo_degree; |
duke@435 | 854 | _lo_degree = i; |
duke@435 | 855 | } |
duke@435 | 856 | } else { // Else high degree |
duke@435 | 857 | lrgs(_hi_degree)._prev = i; |
duke@435 | 858 | lrg._next = _hi_degree; |
duke@435 | 859 | lrg._prev = 0; |
duke@435 | 860 | _hi_degree = i; |
duke@435 | 861 | } |
duke@435 | 862 | } |
duke@435 | 863 | } |
duke@435 | 864 | |
duke@435 | 865 | //------------------------------Pre-Simplify----------------------------------- |
duke@435 | 866 | // Simplify the IFG by removing LRGs of low degree that have NO copies |
duke@435 | 867 | void PhaseChaitin::Pre_Simplify( ) { |
duke@435 | 868 | |
duke@435 | 869 | // Warm up the lo-degree no-copy list |
duke@435 | 870 | int lo_no_copy = 0; |
duke@435 | 871 | for( uint i = 1; i < _maxlrg; i++ ) { |
duke@435 | 872 | if( (lrgs(i).lo_degree() && !lrgs(i)._has_copy) || |
duke@435 | 873 | !lrgs(i).alive() || |
duke@435 | 874 | lrgs(i)._must_spill ) { |
duke@435 | 875 | lrgs(i)._next = lo_no_copy; |
duke@435 | 876 | lo_no_copy = i; |
duke@435 | 877 | } |
duke@435 | 878 | } |
duke@435 | 879 | |
duke@435 | 880 | while( lo_no_copy ) { |
duke@435 | 881 | uint lo = lo_no_copy; |
duke@435 | 882 | lo_no_copy = lrgs(lo)._next; |
duke@435 | 883 | int size = lrgs(lo).num_regs(); |
duke@435 | 884 | |
duke@435 | 885 | // Put the simplified guy on the simplified list. |
duke@435 | 886 | lrgs(lo)._next = _simplified; |
duke@435 | 887 | _simplified = lo; |
duke@435 | 888 | |
duke@435 | 889 | // Yank this guy from the IFG. |
duke@435 | 890 | IndexSet *adj = _ifg->remove_node( lo ); |
duke@435 | 891 | |
duke@435 | 892 | // If any neighbors' degrees fall below their number of |
duke@435 | 893 | // allowed registers, then put that neighbor on the low degree |
duke@435 | 894 | // list. Note that 'degree' can only fall and 'numregs' is |
duke@435 | 895 | // unchanged by this action. Thus the two are equal at most once, |
duke@435 | 896 | // so LRGs hit the lo-degree worklists at most once. |
duke@435 | 897 | IndexSetIterator elements(adj); |
duke@435 | 898 | uint neighbor; |
duke@435 | 899 | while ((neighbor = elements.next()) != 0) { |
duke@435 | 900 | LRG *n = &lrgs(neighbor); |
duke@435 | 901 | assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); |
duke@435 | 902 | |
duke@435 | 903 | // Check for just becoming of-low-degree |
duke@435 | 904 | if( n->just_lo_degree() && !n->_has_copy ) { |
duke@435 | 905 | assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); |
duke@435 | 906 | // Put on lo-degree list |
duke@435 | 907 | n->_next = lo_no_copy; |
duke@435 | 908 | lo_no_copy = neighbor; |
duke@435 | 909 | } |
duke@435 | 910 | } |
duke@435 | 911 | } // End of while lo-degree no_copy worklist not empty |
duke@435 | 912 | |
duke@435 | 913 | // No more lo-degree no-copy live ranges to simplify |
duke@435 | 914 | } |
duke@435 | 915 | |
duke@435 | 916 | //------------------------------Simplify--------------------------------------- |
duke@435 | 917 | // Simplify the IFG by removing LRGs of low degree. |
duke@435 | 918 | void PhaseChaitin::Simplify( ) { |
duke@435 | 919 | |
duke@435 | 920 | while( 1 ) { // Repeat till simplified it all |
duke@435 | 921 | // May want to explore simplifying lo_degree before _lo_stk_degree. |
duke@435 | 922 | // This might result in more spills coloring into registers during |
duke@435 | 923 | // Select(). |
duke@435 | 924 | while( _lo_degree || _lo_stk_degree ) { |
duke@435 | 925 | // If possible, pull from lo_stk first |
duke@435 | 926 | uint lo; |
duke@435 | 927 | if( _lo_degree ) { |
duke@435 | 928 | lo = _lo_degree; |
duke@435 | 929 | _lo_degree = lrgs(lo)._next; |
duke@435 | 930 | } else { |
duke@435 | 931 | lo = _lo_stk_degree; |
duke@435 | 932 | _lo_stk_degree = lrgs(lo)._next; |
duke@435 | 933 | } |
duke@435 | 934 | |
duke@435 | 935 | // Put the simplified guy on the simplified list. |
duke@435 | 936 | lrgs(lo)._next = _simplified; |
duke@435 | 937 | _simplified = lo; |
duke@435 | 938 | // If this guy is "at risk" then mark his current neighbors |
duke@435 | 939 | if( lrgs(lo)._at_risk ) { |
duke@435 | 940 | IndexSetIterator elements(_ifg->neighbors(lo)); |
duke@435 | 941 | uint datum; |
duke@435 | 942 | while ((datum = elements.next()) != 0) { |
duke@435 | 943 | lrgs(datum)._risk_bias = lo; |
duke@435 | 944 | } |
duke@435 | 945 | } |
duke@435 | 946 | |
duke@435 | 947 | // Yank this guy from the IFG. |
duke@435 | 948 | IndexSet *adj = _ifg->remove_node( lo ); |
duke@435 | 949 | |
duke@435 | 950 | // If any neighbors' degrees fall below their number of |
duke@435 | 951 | // allowed registers, then put that neighbor on the low degree |
duke@435 | 952 | // list. Note that 'degree' can only fall and 'numregs' is |
duke@435 | 953 | // unchanged by this action. Thus the two are equal at most once, |
duke@435 | 954 | // so LRGs hit the lo-degree worklist at most once. |
duke@435 | 955 | IndexSetIterator elements(adj); |
duke@435 | 956 | uint neighbor; |
duke@435 | 957 | while ((neighbor = elements.next()) != 0) { |
duke@435 | 958 | LRG *n = &lrgs(neighbor); |
duke@435 | 959 | #ifdef ASSERT |
kvn@985 | 960 | if( VerifyOpto || VerifyRegisterAllocator ) { |
duke@435 | 961 | assert( _ifg->effective_degree(neighbor) == n->degree(), "" ); |
duke@435 | 962 | } |
duke@435 | 963 | #endif |
duke@435 | 964 | |
duke@435 | 965 | // Check for just becoming of-low-degree just counting registers. |
duke@435 | 966 | // _must_spill live ranges are already on the low degree list. |
duke@435 | 967 | if( n->just_lo_degree() && !n->_must_spill ) { |
duke@435 | 968 | assert(!(*_ifg->_yanked)[neighbor],"Cannot move to lo degree twice"); |
duke@435 | 969 | // Pull from hi-degree list |
duke@435 | 970 | uint prev = n->_prev; |
duke@435 | 971 | uint next = n->_next; |
duke@435 | 972 | if( prev ) lrgs(prev)._next = next; |
duke@435 | 973 | else _hi_degree = next; |
duke@435 | 974 | lrgs(next)._prev = prev; |
duke@435 | 975 | n->_next = _lo_degree; |
duke@435 | 976 | _lo_degree = neighbor; |
duke@435 | 977 | } |
duke@435 | 978 | } |
duke@435 | 979 | } // End of while lo-degree/lo_stk_degree worklist not empty |
duke@435 | 980 | |
duke@435 | 981 | // Check for got everything: is hi-degree list empty? |
duke@435 | 982 | if( !_hi_degree ) break; |
duke@435 | 983 | |
duke@435 | 984 | // Time to pick a potential spill guy |
duke@435 | 985 | uint lo_score = _hi_degree; |
duke@435 | 986 | double score = lrgs(lo_score).score(); |
duke@435 | 987 | double area = lrgs(lo_score)._area; |
kvn@1443 | 988 | double cost = lrgs(lo_score)._cost; |
kvn@1443 | 989 | bool bound = lrgs(lo_score)._is_bound; |
duke@435 | 990 | |
duke@435 | 991 | // Find cheapest guy |
duke@435 | 992 | debug_only( int lo_no_simplify=0; ); |
kvn@1447 | 993 | for( uint i = _hi_degree; i; i = lrgs(i)._next ) { |
duke@435 | 994 | assert( !(*_ifg->_yanked)[i], "" ); |
duke@435 | 995 | // It's just vaguely possible to move hi-degree to lo-degree without |
duke@435 | 996 | // going through a just-lo-degree stage: If you remove a double from |
duke@435 | 997 | // a float live range it's degree will drop by 2 and you can skip the |
duke@435 | 998 | // just-lo-degree stage. It's very rare (shows up after 5000+ methods |
duke@435 | 999 | // in -Xcomp of Java2Demo). So just choose this guy to simplify next. |
duke@435 | 1000 | if( lrgs(i).lo_degree() ) { |
duke@435 | 1001 | lo_score = i; |
duke@435 | 1002 | break; |
duke@435 | 1003 | } |
duke@435 | 1004 | debug_only( if( lrgs(i)._was_lo ) lo_no_simplify=i; ); |
duke@435 | 1005 | double iscore = lrgs(i).score(); |
duke@435 | 1006 | double iarea = lrgs(i)._area; |
kvn@1443 | 1007 | double icost = lrgs(i)._cost; |
kvn@1443 | 1008 | bool ibound = lrgs(i)._is_bound; |
duke@435 | 1009 | |
duke@435 | 1010 | // Compare cost/area of i vs cost/area of lo_score. Smaller cost/area |
duke@435 | 1011 | // wins. Ties happen because all live ranges in question have spilled |
duke@435 | 1012 | // a few times before and the spill-score adds a huge number which |
duke@435 | 1013 | // washes out the low order bits. We are choosing the lesser of 2 |
duke@435 | 1014 | // evils; in this case pick largest area to spill. |
kvn@1443 | 1015 | // Ties also happen when live ranges are defined and used only inside |
kvn@1443 | 1016 | // one block. In which case their area is 0 and score set to max. |
kvn@1443 | 1017 | // In such case choose bound live range over unbound to free registers |
kvn@1443 | 1018 | // or with smaller cost to spill. |
duke@435 | 1019 | if( iscore < score || |
kvn@1443 | 1020 | (iscore == score && iarea > area && lrgs(lo_score)._was_spilled2) || |
kvn@1443 | 1021 | (iscore == score && iarea == area && |
kvn@1443 | 1022 | ( (ibound && !bound) || ibound == bound && (icost < cost) )) ) { |
duke@435 | 1023 | lo_score = i; |
duke@435 | 1024 | score = iscore; |
duke@435 | 1025 | area = iarea; |
kvn@1443 | 1026 | cost = icost; |
kvn@1443 | 1027 | bound = ibound; |
duke@435 | 1028 | } |
duke@435 | 1029 | } |
duke@435 | 1030 | LRG *lo_lrg = &lrgs(lo_score); |
duke@435 | 1031 | // The live range we choose for spilling is either hi-degree, or very |
duke@435 | 1032 | // rarely it can be low-degree. If we choose a hi-degree live range |
duke@435 | 1033 | // there better not be any lo-degree choices. |
duke@435 | 1034 | assert( lo_lrg->lo_degree() || !lo_no_simplify, "Live range was lo-degree before coalesce; should simplify" ); |
duke@435 | 1035 | |
duke@435 | 1036 | // Pull from hi-degree list |
duke@435 | 1037 | uint prev = lo_lrg->_prev; |
duke@435 | 1038 | uint next = lo_lrg->_next; |
duke@435 | 1039 | if( prev ) lrgs(prev)._next = next; |
duke@435 | 1040 | else _hi_degree = next; |
duke@435 | 1041 | lrgs(next)._prev = prev; |
duke@435 | 1042 | // Jam him on the lo-degree list, despite his high degree. |
duke@435 | 1043 | // Maybe he'll get a color, and maybe he'll spill. |
duke@435 | 1044 | // Only Select() will know. |
duke@435 | 1045 | lrgs(lo_score)._at_risk = true; |
duke@435 | 1046 | _lo_degree = lo_score; |
duke@435 | 1047 | lo_lrg->_next = 0; |
duke@435 | 1048 | |
duke@435 | 1049 | } // End of while not simplified everything |
duke@435 | 1050 | |
duke@435 | 1051 | } |
duke@435 | 1052 | |
duke@435 | 1053 | //------------------------------bias_color------------------------------------- |
duke@435 | 1054 | // Choose a color using the biasing heuristic |
duke@435 | 1055 | OptoReg::Name PhaseChaitin::bias_color( LRG &lrg, int chunk ) { |
duke@435 | 1056 | |
duke@435 | 1057 | // Check for "at_risk" LRG's |
duke@435 | 1058 | uint risk_lrg = Find(lrg._risk_bias); |
duke@435 | 1059 | if( risk_lrg != 0 ) { |
duke@435 | 1060 | // Walk the colored neighbors of the "at_risk" candidate |
duke@435 | 1061 | // Choose a color which is both legal and already taken by a neighbor |
duke@435 | 1062 | // of the "at_risk" candidate in order to improve the chances of the |
duke@435 | 1063 | // "at_risk" candidate of coloring |
duke@435 | 1064 | IndexSetIterator elements(_ifg->neighbors(risk_lrg)); |
duke@435 | 1065 | uint datum; |
duke@435 | 1066 | while ((datum = elements.next()) != 0) { |
duke@435 | 1067 | OptoReg::Name reg = lrgs(datum).reg(); |
duke@435 | 1068 | // If this LRG's register is legal for us, choose it |
duke@435 | 1069 | if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE && |
duke@435 | 1070 | lrg.mask().Member(OptoReg::add(reg,-chunk)) && |
duke@435 | 1071 | (lrg.num_regs()==1 || // either size 1 |
duke@435 | 1072 | (reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs) |
duke@435 | 1073 | return reg; |
duke@435 | 1074 | } |
duke@435 | 1075 | } |
duke@435 | 1076 | |
duke@435 | 1077 | uint copy_lrg = Find(lrg._copy_bias); |
duke@435 | 1078 | if( copy_lrg != 0 ) { |
duke@435 | 1079 | // If he has a color, |
duke@435 | 1080 | if( !(*(_ifg->_yanked))[copy_lrg] ) { |
duke@435 | 1081 | OptoReg::Name reg = lrgs(copy_lrg).reg(); |
duke@435 | 1082 | // And it is legal for you, |
duke@435 | 1083 | if( reg >= chunk && reg < chunk + RegMask::CHUNK_SIZE && |
duke@435 | 1084 | lrg.mask().Member(OptoReg::add(reg,-chunk)) && |
duke@435 | 1085 | (lrg.num_regs()==1 || // either size 1 |
duke@435 | 1086 | (reg&1) == 1) ) // or aligned (adjacent reg is available since we already cleared-to-pairs) |
duke@435 | 1087 | return reg; |
duke@435 | 1088 | } else if( chunk == 0 ) { |
duke@435 | 1089 | // Choose a color which is legal for him |
duke@435 | 1090 | RegMask tempmask = lrg.mask(); |
duke@435 | 1091 | tempmask.AND(lrgs(copy_lrg).mask()); |
duke@435 | 1092 | OptoReg::Name reg; |
duke@435 | 1093 | if( lrg.num_regs() == 1 ) { |
duke@435 | 1094 | reg = tempmask.find_first_elem(); |
duke@435 | 1095 | } else { |
duke@435 | 1096 | tempmask.ClearToPairs(); |
duke@435 | 1097 | reg = tempmask.find_first_pair(); |
duke@435 | 1098 | } |
duke@435 | 1099 | if( OptoReg::is_valid(reg) ) |
duke@435 | 1100 | return reg; |
duke@435 | 1101 | } |
duke@435 | 1102 | } |
duke@435 | 1103 | |
duke@435 | 1104 | // If no bias info exists, just go with the register selection ordering |
duke@435 | 1105 | if( lrg.num_regs() == 2 ) { |
duke@435 | 1106 | // Find an aligned pair |
duke@435 | 1107 | return OptoReg::add(lrg.mask().find_first_pair(),chunk); |
duke@435 | 1108 | } |
duke@435 | 1109 | |
duke@435 | 1110 | // CNC - Fun hack. Alternate 1st and 2nd selection. Enables post-allocate |
duke@435 | 1111 | // copy removal to remove many more copies, by preventing a just-assigned |
duke@435 | 1112 | // register from being repeatedly assigned. |
duke@435 | 1113 | OptoReg::Name reg = lrg.mask().find_first_elem(); |
duke@435 | 1114 | if( (++_alternate & 1) && OptoReg::is_valid(reg) ) { |
duke@435 | 1115 | // This 'Remove; find; Insert' idiom is an expensive way to find the |
duke@435 | 1116 | // SECOND element in the mask. |
duke@435 | 1117 | lrg.Remove(reg); |
duke@435 | 1118 | OptoReg::Name reg2 = lrg.mask().find_first_elem(); |
duke@435 | 1119 | lrg.Insert(reg); |
duke@435 | 1120 | if( OptoReg::is_reg(reg2)) |
duke@435 | 1121 | reg = reg2; |
duke@435 | 1122 | } |
duke@435 | 1123 | return OptoReg::add( reg, chunk ); |
duke@435 | 1124 | } |
duke@435 | 1125 | |
duke@435 | 1126 | //------------------------------choose_color----------------------------------- |
duke@435 | 1127 | // Choose a color in the current chunk |
duke@435 | 1128 | OptoReg::Name PhaseChaitin::choose_color( LRG &lrg, int chunk ) { |
duke@435 | 1129 | 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)"); |
duke@435 | 1130 | 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)"); |
duke@435 | 1131 | |
duke@435 | 1132 | if( lrg.num_regs() == 1 || // Common Case |
duke@435 | 1133 | !lrg._fat_proj ) // Aligned+adjacent pairs ok |
duke@435 | 1134 | // Use a heuristic to "bias" the color choice |
duke@435 | 1135 | return bias_color(lrg, chunk); |
duke@435 | 1136 | |
duke@435 | 1137 | assert( lrg.num_regs() >= 2, "dead live ranges do not color" ); |
duke@435 | 1138 | |
duke@435 | 1139 | // Fat-proj case or misaligned double argument. |
duke@435 | 1140 | assert(lrg.compute_mask_size() == lrg.num_regs() || |
duke@435 | 1141 | lrg.num_regs() == 2,"fat projs exactly color" ); |
duke@435 | 1142 | assert( !chunk, "always color in 1st chunk" ); |
duke@435 | 1143 | // Return the highest element in the set. |
duke@435 | 1144 | return lrg.mask().find_last_elem(); |
duke@435 | 1145 | } |
duke@435 | 1146 | |
duke@435 | 1147 | //------------------------------Select----------------------------------------- |
duke@435 | 1148 | // Select colors by re-inserting LRGs back into the IFG. LRGs are re-inserted |
duke@435 | 1149 | // in reverse order of removal. As long as nothing of hi-degree was yanked, |
duke@435 | 1150 | // everything going back is guaranteed a color. Select that color. If some |
duke@435 | 1151 | // hi-degree LRG cannot get a color then we record that we must spill. |
duke@435 | 1152 | uint PhaseChaitin::Select( ) { |
duke@435 | 1153 | uint spill_reg = LRG::SPILL_REG; |
duke@435 | 1154 | _max_reg = OptoReg::Name(0); // Past max register used |
duke@435 | 1155 | while( _simplified ) { |
duke@435 | 1156 | // Pull next LRG from the simplified list - in reverse order of removal |
duke@435 | 1157 | uint lidx = _simplified; |
duke@435 | 1158 | LRG *lrg = &lrgs(lidx); |
duke@435 | 1159 | _simplified = lrg->_next; |
duke@435 | 1160 | |
duke@435 | 1161 | |
duke@435 | 1162 | #ifndef PRODUCT |
duke@435 | 1163 | if (trace_spilling()) { |
duke@435 | 1164 | ttyLocker ttyl; |
duke@435 | 1165 | tty->print_cr("L%d selecting degree %d degrees_of_freedom %d", lidx, lrg->degree(), |
duke@435 | 1166 | lrg->degrees_of_freedom()); |
duke@435 | 1167 | lrg->dump(); |
duke@435 | 1168 | } |
duke@435 | 1169 | #endif |
duke@435 | 1170 | |
duke@435 | 1171 | // Re-insert into the IFG |
duke@435 | 1172 | _ifg->re_insert(lidx); |
duke@435 | 1173 | if( !lrg->alive() ) continue; |
duke@435 | 1174 | // capture allstackedness flag before mask is hacked |
duke@435 | 1175 | const int is_allstack = lrg->mask().is_AllStack(); |
duke@435 | 1176 | |
duke@435 | 1177 | // Yeah, yeah, yeah, I know, I know. I can refactor this |
duke@435 | 1178 | // to avoid the GOTO, although the refactored code will not |
duke@435 | 1179 | // be much clearer. We arrive here IFF we have a stack-based |
duke@435 | 1180 | // live range that cannot color in the current chunk, and it |
duke@435 | 1181 | // has to move into the next free stack chunk. |
duke@435 | 1182 | int chunk = 0; // Current chunk is first chunk |
duke@435 | 1183 | retry_next_chunk: |
duke@435 | 1184 | |
duke@435 | 1185 | // Remove neighbor colors |
duke@435 | 1186 | IndexSet *s = _ifg->neighbors(lidx); |
duke@435 | 1187 | |
duke@435 | 1188 | debug_only(RegMask orig_mask = lrg->mask();) |
duke@435 | 1189 | IndexSetIterator elements(s); |
duke@435 | 1190 | uint neighbor; |
duke@435 | 1191 | while ((neighbor = elements.next()) != 0) { |
duke@435 | 1192 | // Note that neighbor might be a spill_reg. In this case, exclusion |
duke@435 | 1193 | // of its color will be a no-op, since the spill_reg chunk is in outer |
duke@435 | 1194 | // space. Also, if neighbor is in a different chunk, this exclusion |
duke@435 | 1195 | // will be a no-op. (Later on, if lrg runs out of possible colors in |
duke@435 | 1196 | // its chunk, a new chunk of color may be tried, in which case |
duke@435 | 1197 | // examination of neighbors is started again, at retry_next_chunk.) |
duke@435 | 1198 | LRG &nlrg = lrgs(neighbor); |
duke@435 | 1199 | OptoReg::Name nreg = nlrg.reg(); |
duke@435 | 1200 | // Only subtract masks in the same chunk |
duke@435 | 1201 | if( nreg >= chunk && nreg < chunk + RegMask::CHUNK_SIZE ) { |
duke@435 | 1202 | #ifndef PRODUCT |
duke@435 | 1203 | uint size = lrg->mask().Size(); |
duke@435 | 1204 | RegMask rm = lrg->mask(); |
duke@435 | 1205 | #endif |
duke@435 | 1206 | lrg->SUBTRACT(nlrg.mask()); |
duke@435 | 1207 | #ifndef PRODUCT |
duke@435 | 1208 | if (trace_spilling() && lrg->mask().Size() != size) { |
duke@435 | 1209 | ttyLocker ttyl; |
duke@435 | 1210 | tty->print("L%d ", lidx); |
duke@435 | 1211 | rm.dump(); |
duke@435 | 1212 | tty->print(" intersected L%d ", neighbor); |
duke@435 | 1213 | nlrg.mask().dump(); |
duke@435 | 1214 | tty->print(" removed "); |
duke@435 | 1215 | rm.SUBTRACT(lrg->mask()); |
duke@435 | 1216 | rm.dump(); |
duke@435 | 1217 | tty->print(" leaving "); |
duke@435 | 1218 | lrg->mask().dump(); |
duke@435 | 1219 | tty->cr(); |
duke@435 | 1220 | } |
duke@435 | 1221 | #endif |
duke@435 | 1222 | } |
duke@435 | 1223 | } |
duke@435 | 1224 | //assert(is_allstack == lrg->mask().is_AllStack(), "nbrs must not change AllStackedness"); |
duke@435 | 1225 | // Aligned pairs need aligned masks |
duke@435 | 1226 | if( lrg->num_regs() == 2 && !lrg->_fat_proj ) |
duke@435 | 1227 | lrg->ClearToPairs(); |
duke@435 | 1228 | |
duke@435 | 1229 | // Check if a color is available and if so pick the color |
duke@435 | 1230 | OptoReg::Name reg = choose_color( *lrg, chunk ); |
duke@435 | 1231 | #ifdef SPARC |
duke@435 | 1232 | debug_only(lrg->compute_set_mask_size()); |
duke@435 | 1233 | assert(lrg->num_regs() != 2 || lrg->is_bound() || is_even(reg-1), "allocate all doubles aligned"); |
duke@435 | 1234 | #endif |
duke@435 | 1235 | |
duke@435 | 1236 | //--------------- |
duke@435 | 1237 | // If we fail to color and the AllStack flag is set, trigger |
duke@435 | 1238 | // a chunk-rollover event |
duke@435 | 1239 | if(!OptoReg::is_valid(OptoReg::add(reg,-chunk)) && is_allstack) { |
duke@435 | 1240 | // Bump register mask up to next stack chunk |
duke@435 | 1241 | chunk += RegMask::CHUNK_SIZE; |
duke@435 | 1242 | lrg->Set_All(); |
duke@435 | 1243 | |
duke@435 | 1244 | goto retry_next_chunk; |
duke@435 | 1245 | } |
duke@435 | 1246 | |
duke@435 | 1247 | //--------------- |
duke@435 | 1248 | // Did we get a color? |
duke@435 | 1249 | else if( OptoReg::is_valid(reg)) { |
duke@435 | 1250 | #ifndef PRODUCT |
duke@435 | 1251 | RegMask avail_rm = lrg->mask(); |
duke@435 | 1252 | #endif |
duke@435 | 1253 | |
duke@435 | 1254 | // Record selected register |
duke@435 | 1255 | lrg->set_reg(reg); |
duke@435 | 1256 | |
duke@435 | 1257 | if( reg >= _max_reg ) // Compute max register limit |
duke@435 | 1258 | _max_reg = OptoReg::add(reg,1); |
duke@435 | 1259 | // Fold reg back into normal space |
duke@435 | 1260 | reg = OptoReg::add(reg,-chunk); |
duke@435 | 1261 | |
duke@435 | 1262 | // If the live range is not bound, then we actually had some choices |
duke@435 | 1263 | // to make. In this case, the mask has more bits in it than the colors |
twisti@1040 | 1264 | // chosen. Restrict the mask to just what was picked. |
duke@435 | 1265 | if( lrg->num_regs() == 1 ) { // Size 1 live range |
duke@435 | 1266 | lrg->Clear(); // Clear the mask |
duke@435 | 1267 | lrg->Insert(reg); // Set regmask to match selected reg |
duke@435 | 1268 | lrg->set_mask_size(1); |
duke@435 | 1269 | } else if( !lrg->_fat_proj ) { |
duke@435 | 1270 | // For pairs, also insert the low bit of the pair |
duke@435 | 1271 | assert( lrg->num_regs() == 2, "unbound fatproj???" ); |
duke@435 | 1272 | lrg->Clear(); // Clear the mask |
duke@435 | 1273 | lrg->Insert(reg); // Set regmask to match selected reg |
duke@435 | 1274 | lrg->Insert(OptoReg::add(reg,-1)); |
duke@435 | 1275 | lrg->set_mask_size(2); |
duke@435 | 1276 | } else { // Else fatproj |
duke@435 | 1277 | // mask must be equal to fatproj bits, by definition |
duke@435 | 1278 | } |
duke@435 | 1279 | #ifndef PRODUCT |
duke@435 | 1280 | if (trace_spilling()) { |
duke@435 | 1281 | ttyLocker ttyl; |
duke@435 | 1282 | tty->print("L%d selected ", lidx); |
duke@435 | 1283 | lrg->mask().dump(); |
duke@435 | 1284 | tty->print(" from "); |
duke@435 | 1285 | avail_rm.dump(); |
duke@435 | 1286 | tty->cr(); |
duke@435 | 1287 | } |
duke@435 | 1288 | #endif |
duke@435 | 1289 | // Note that reg is the highest-numbered register in the newly-bound mask. |
duke@435 | 1290 | } // end color available case |
duke@435 | 1291 | |
duke@435 | 1292 | //--------------- |
duke@435 | 1293 | // Live range is live and no colors available |
duke@435 | 1294 | else { |
duke@435 | 1295 | assert( lrg->alive(), "" ); |
never@730 | 1296 | assert( !lrg->_fat_proj || lrg->is_multidef() || |
duke@435 | 1297 | lrg->_def->outcnt() > 0, "fat_proj cannot spill"); |
duke@435 | 1298 | assert( !orig_mask.is_AllStack(), "All Stack does not spill" ); |
duke@435 | 1299 | |
duke@435 | 1300 | // Assign the special spillreg register |
duke@435 | 1301 | lrg->set_reg(OptoReg::Name(spill_reg++)); |
duke@435 | 1302 | // Do not empty the regmask; leave mask_size lying around |
duke@435 | 1303 | // for use during Spilling |
duke@435 | 1304 | #ifndef PRODUCT |
duke@435 | 1305 | if( trace_spilling() ) { |
duke@435 | 1306 | ttyLocker ttyl; |
duke@435 | 1307 | tty->print("L%d spilling with neighbors: ", lidx); |
duke@435 | 1308 | s->dump(); |
duke@435 | 1309 | debug_only(tty->print(" original mask: ")); |
duke@435 | 1310 | debug_only(orig_mask.dump()); |
duke@435 | 1311 | dump_lrg(lidx); |
duke@435 | 1312 | } |
duke@435 | 1313 | #endif |
duke@435 | 1314 | } // end spill case |
duke@435 | 1315 | |
duke@435 | 1316 | } |
duke@435 | 1317 | |
duke@435 | 1318 | return spill_reg-LRG::SPILL_REG; // Return number of spills |
duke@435 | 1319 | } |
duke@435 | 1320 | |
duke@435 | 1321 | |
duke@435 | 1322 | //------------------------------copy_was_spilled------------------------------- |
duke@435 | 1323 | // Copy 'was_spilled'-edness from the source Node to the dst Node. |
duke@435 | 1324 | void PhaseChaitin::copy_was_spilled( Node *src, Node *dst ) { |
duke@435 | 1325 | if( _spilled_once.test(src->_idx) ) { |
duke@435 | 1326 | _spilled_once.set(dst->_idx); |
duke@435 | 1327 | lrgs(Find(dst))._was_spilled1 = 1; |
duke@435 | 1328 | if( _spilled_twice.test(src->_idx) ) { |
duke@435 | 1329 | _spilled_twice.set(dst->_idx); |
duke@435 | 1330 | lrgs(Find(dst))._was_spilled2 = 1; |
duke@435 | 1331 | } |
duke@435 | 1332 | } |
duke@435 | 1333 | } |
duke@435 | 1334 | |
duke@435 | 1335 | //------------------------------set_was_spilled-------------------------------- |
duke@435 | 1336 | // Set the 'spilled_once' or 'spilled_twice' flag on a node. |
duke@435 | 1337 | void PhaseChaitin::set_was_spilled( Node *n ) { |
duke@435 | 1338 | if( _spilled_once.test_set(n->_idx) ) |
duke@435 | 1339 | _spilled_twice.set(n->_idx); |
duke@435 | 1340 | } |
duke@435 | 1341 | |
duke@435 | 1342 | //------------------------------fixup_spills----------------------------------- |
duke@435 | 1343 | // Convert Ideal spill instructions into proper FramePtr + offset Loads and |
duke@435 | 1344 | // Stores. Use-def chains are NOT preserved, but Node->LRG->reg maps are. |
duke@435 | 1345 | void PhaseChaitin::fixup_spills() { |
duke@435 | 1346 | // This function does only cisc spill work. |
duke@435 | 1347 | if( !UseCISCSpill ) return; |
duke@435 | 1348 | |
duke@435 | 1349 | NOT_PRODUCT( Compile::TracePhase t3("fixupSpills", &_t_fixupSpills, TimeCompiler); ) |
duke@435 | 1350 | |
duke@435 | 1351 | // Grab the Frame Pointer |
duke@435 | 1352 | Node *fp = _cfg._broot->head()->in(1)->in(TypeFunc::FramePtr); |
duke@435 | 1353 | |
duke@435 | 1354 | // For all blocks |
duke@435 | 1355 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 1356 | Block *b = _cfg._blocks[i]; |
duke@435 | 1357 | |
duke@435 | 1358 | // For all instructions in block |
duke@435 | 1359 | uint last_inst = b->end_idx(); |
duke@435 | 1360 | for( uint j = 1; j <= last_inst; j++ ) { |
duke@435 | 1361 | Node *n = b->_nodes[j]; |
duke@435 | 1362 | |
duke@435 | 1363 | // Dead instruction??? |
duke@435 | 1364 | assert( n->outcnt() != 0 ||// Nothing dead after post alloc |
duke@435 | 1365 | C->top() == n || // Or the random TOP node |
duke@435 | 1366 | n->is_Proj(), // Or a fat-proj kill node |
duke@435 | 1367 | "No dead instructions after post-alloc" ); |
duke@435 | 1368 | |
duke@435 | 1369 | int inp = n->cisc_operand(); |
duke@435 | 1370 | if( inp != AdlcVMDeps::Not_cisc_spillable ) { |
duke@435 | 1371 | // Convert operand number to edge index number |
duke@435 | 1372 | MachNode *mach = n->as_Mach(); |
duke@435 | 1373 | inp = mach->operand_index(inp); |
duke@435 | 1374 | Node *src = n->in(inp); // Value to load or store |
duke@435 | 1375 | LRG &lrg_cisc = lrgs( Find_const(src) ); |
duke@435 | 1376 | OptoReg::Name src_reg = lrg_cisc.reg(); |
duke@435 | 1377 | // Doubles record the HIGH register of an adjacent pair. |
duke@435 | 1378 | src_reg = OptoReg::add(src_reg,1-lrg_cisc.num_regs()); |
duke@435 | 1379 | if( OptoReg::is_stack(src_reg) ) { // If input is on stack |
duke@435 | 1380 | // This is a CISC Spill, get stack offset and construct new node |
duke@435 | 1381 | #ifndef PRODUCT |
duke@435 | 1382 | if( TraceCISCSpill ) { |
duke@435 | 1383 | tty->print(" reg-instr: "); |
duke@435 | 1384 | n->dump(); |
duke@435 | 1385 | } |
duke@435 | 1386 | #endif |
duke@435 | 1387 | int stk_offset = reg2offset(src_reg); |
duke@435 | 1388 | // Bailout if we might exceed node limit when spilling this instruction |
duke@435 | 1389 | C->check_node_count(0, "out of nodes fixing spills"); |
duke@435 | 1390 | if (C->failing()) return; |
duke@435 | 1391 | // Transform node |
duke@435 | 1392 | MachNode *cisc = mach->cisc_version(stk_offset, C)->as_Mach(); |
duke@435 | 1393 | cisc->set_req(inp,fp); // Base register is frame pointer |
duke@435 | 1394 | if( cisc->oper_input_base() > 1 && mach->oper_input_base() <= 1 ) { |
duke@435 | 1395 | assert( cisc->oper_input_base() == 2, "Only adding one edge"); |
duke@435 | 1396 | cisc->ins_req(1,src); // Requires a memory edge |
duke@435 | 1397 | } |
duke@435 | 1398 | b->_nodes.map(j,cisc); // Insert into basic block |
kvn@603 | 1399 | n->subsume_by(cisc); // Correct graph |
duke@435 | 1400 | // |
duke@435 | 1401 | ++_used_cisc_instructions; |
duke@435 | 1402 | #ifndef PRODUCT |
duke@435 | 1403 | if( TraceCISCSpill ) { |
duke@435 | 1404 | tty->print(" cisc-instr: "); |
duke@435 | 1405 | cisc->dump(); |
duke@435 | 1406 | } |
duke@435 | 1407 | #endif |
duke@435 | 1408 | } else { |
duke@435 | 1409 | #ifndef PRODUCT |
duke@435 | 1410 | if( TraceCISCSpill ) { |
duke@435 | 1411 | tty->print(" using reg-instr: "); |
duke@435 | 1412 | n->dump(); |
duke@435 | 1413 | } |
duke@435 | 1414 | #endif |
duke@435 | 1415 | ++_unused_cisc_instructions; // input can be on stack |
duke@435 | 1416 | } |
duke@435 | 1417 | } |
duke@435 | 1418 | |
duke@435 | 1419 | } // End of for all instructions |
duke@435 | 1420 | |
duke@435 | 1421 | } // End of for all blocks |
duke@435 | 1422 | } |
duke@435 | 1423 | |
duke@435 | 1424 | //------------------------------find_base_for_derived-------------------------- |
duke@435 | 1425 | // Helper to stretch above; recursively discover the base Node for a |
duke@435 | 1426 | // given derived Node. Easy for AddP-related machine nodes, but needs |
duke@435 | 1427 | // to be recursive for derived Phis. |
duke@435 | 1428 | Node *PhaseChaitin::find_base_for_derived( Node **derived_base_map, Node *derived, uint &maxlrg ) { |
duke@435 | 1429 | // See if already computed; if so return it |
duke@435 | 1430 | if( derived_base_map[derived->_idx] ) |
duke@435 | 1431 | return derived_base_map[derived->_idx]; |
duke@435 | 1432 | |
duke@435 | 1433 | // See if this happens to be a base. |
duke@435 | 1434 | // NOTE: we use TypePtr instead of TypeOopPtr because we can have |
duke@435 | 1435 | // pointers derived from NULL! These are always along paths that |
duke@435 | 1436 | // can't happen at run-time but the optimizer cannot deduce it so |
duke@435 | 1437 | // we have to handle it gracefully. |
kvn@1164 | 1438 | assert(!derived->bottom_type()->isa_narrowoop() || |
kvn@1164 | 1439 | derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); |
duke@435 | 1440 | const TypePtr *tj = derived->bottom_type()->isa_ptr(); |
duke@435 | 1441 | // If its an OOP with a non-zero offset, then it is derived. |
kvn@1164 | 1442 | if( tj == NULL || tj->_offset == 0 ) { |
duke@435 | 1443 | derived_base_map[derived->_idx] = derived; |
duke@435 | 1444 | return derived; |
duke@435 | 1445 | } |
duke@435 | 1446 | // Derived is NULL+offset? Base is NULL! |
duke@435 | 1447 | if( derived->is_Con() ) { |
kvn@1164 | 1448 | Node *base = _matcher.mach_null(); |
kvn@1164 | 1449 | assert(base != NULL, "sanity"); |
kvn@1164 | 1450 | if (base->in(0) == NULL) { |
kvn@1164 | 1451 | // Initialize it once and make it shared: |
kvn@1164 | 1452 | // set control to _root and place it into Start block |
kvn@1164 | 1453 | // (where top() node is placed). |
kvn@1164 | 1454 | base->init_req(0, _cfg._root); |
kvn@1164 | 1455 | Block *startb = _cfg._bbs[C->top()->_idx]; |
kvn@1164 | 1456 | startb->_nodes.insert(startb->find_node(C->top()), base ); |
kvn@1164 | 1457 | _cfg._bbs.map( base->_idx, startb ); |
kvn@1164 | 1458 | assert (n2lidx(base) == 0, "should not have LRG yet"); |
kvn@1164 | 1459 | } |
kvn@1164 | 1460 | if (n2lidx(base) == 0) { |
kvn@1164 | 1461 | new_lrg(base, maxlrg++); |
kvn@1164 | 1462 | } |
kvn@1164 | 1463 | assert(base->in(0) == _cfg._root && |
kvn@1164 | 1464 | _cfg._bbs[base->_idx] == _cfg._bbs[C->top()->_idx], "base NULL should be shared"); |
duke@435 | 1465 | derived_base_map[derived->_idx] = base; |
duke@435 | 1466 | return base; |
duke@435 | 1467 | } |
duke@435 | 1468 | |
duke@435 | 1469 | // Check for AddP-related opcodes |
duke@435 | 1470 | if( !derived->is_Phi() ) { |
duke@435 | 1471 | assert( derived->as_Mach()->ideal_Opcode() == Op_AddP, "" ); |
duke@435 | 1472 | Node *base = derived->in(AddPNode::Base); |
duke@435 | 1473 | derived_base_map[derived->_idx] = base; |
duke@435 | 1474 | return base; |
duke@435 | 1475 | } |
duke@435 | 1476 | |
duke@435 | 1477 | // Recursively find bases for Phis. |
duke@435 | 1478 | // First check to see if we can avoid a base Phi here. |
duke@435 | 1479 | Node *base = find_base_for_derived( derived_base_map, derived->in(1),maxlrg); |
duke@435 | 1480 | uint i; |
duke@435 | 1481 | for( i = 2; i < derived->req(); i++ ) |
duke@435 | 1482 | if( base != find_base_for_derived( derived_base_map,derived->in(i),maxlrg)) |
duke@435 | 1483 | break; |
duke@435 | 1484 | // Went to the end without finding any different bases? |
duke@435 | 1485 | if( i == derived->req() ) { // No need for a base Phi here |
duke@435 | 1486 | derived_base_map[derived->_idx] = base; |
duke@435 | 1487 | return base; |
duke@435 | 1488 | } |
duke@435 | 1489 | |
duke@435 | 1490 | // Now we see we need a base-Phi here to merge the bases |
kvn@1164 | 1491 | const Type *t = base->bottom_type(); |
kvn@1164 | 1492 | base = new (C, derived->req()) PhiNode( derived->in(0), t ); |
kvn@1164 | 1493 | for( i = 1; i < derived->req(); i++ ) { |
duke@435 | 1494 | base->init_req(i, find_base_for_derived(derived_base_map, derived->in(i), maxlrg)); |
kvn@1164 | 1495 | t = t->meet(base->in(i)->bottom_type()); |
kvn@1164 | 1496 | } |
kvn@1164 | 1497 | base->as_Phi()->set_type(t); |
duke@435 | 1498 | |
duke@435 | 1499 | // Search the current block for an existing base-Phi |
duke@435 | 1500 | Block *b = _cfg._bbs[derived->_idx]; |
duke@435 | 1501 | for( i = 1; i <= b->end_idx(); i++ ) {// Search for matching Phi |
duke@435 | 1502 | Node *phi = b->_nodes[i]; |
duke@435 | 1503 | if( !phi->is_Phi() ) { // Found end of Phis with no match? |
duke@435 | 1504 | b->_nodes.insert( i, base ); // Must insert created Phi here as base |
duke@435 | 1505 | _cfg._bbs.map( base->_idx, b ); |
duke@435 | 1506 | new_lrg(base,maxlrg++); |
duke@435 | 1507 | break; |
duke@435 | 1508 | } |
duke@435 | 1509 | // See if Phi matches. |
duke@435 | 1510 | uint j; |
duke@435 | 1511 | for( j = 1; j < base->req(); j++ ) |
duke@435 | 1512 | if( phi->in(j) != base->in(j) && |
duke@435 | 1513 | !(phi->in(j)->is_Con() && base->in(j)->is_Con()) ) // allow different NULLs |
duke@435 | 1514 | break; |
duke@435 | 1515 | if( j == base->req() ) { // All inputs match? |
duke@435 | 1516 | base = phi; // Then use existing 'phi' and drop 'base' |
duke@435 | 1517 | break; |
duke@435 | 1518 | } |
duke@435 | 1519 | } |
duke@435 | 1520 | |
duke@435 | 1521 | |
duke@435 | 1522 | // Cache info for later passes |
duke@435 | 1523 | derived_base_map[derived->_idx] = base; |
duke@435 | 1524 | return base; |
duke@435 | 1525 | } |
duke@435 | 1526 | |
duke@435 | 1527 | |
duke@435 | 1528 | //------------------------------stretch_base_pointer_live_ranges--------------- |
duke@435 | 1529 | // At each Safepoint, insert extra debug edges for each pair of derived value/ |
duke@435 | 1530 | // base pointer that is live across the Safepoint for oopmap building. The |
duke@435 | 1531 | // edge pairs get added in after sfpt->jvmtail()->oopoff(), but are in the |
duke@435 | 1532 | // required edge set. |
duke@435 | 1533 | bool PhaseChaitin::stretch_base_pointer_live_ranges( ResourceArea *a ) { |
duke@435 | 1534 | int must_recompute_live = false; |
duke@435 | 1535 | uint maxlrg = _maxlrg; |
duke@435 | 1536 | Node **derived_base_map = (Node**)a->Amalloc(sizeof(Node*)*C->unique()); |
duke@435 | 1537 | memset( derived_base_map, 0, sizeof(Node*)*C->unique() ); |
duke@435 | 1538 | |
duke@435 | 1539 | // For all blocks in RPO do... |
duke@435 | 1540 | for( uint i=0; i<_cfg._num_blocks; i++ ) { |
duke@435 | 1541 | Block *b = _cfg._blocks[i]; |
duke@435 | 1542 | // Note use of deep-copy constructor. I cannot hammer the original |
duke@435 | 1543 | // liveout bits, because they are needed by the following coalesce pass. |
duke@435 | 1544 | IndexSet liveout(_live->live(b)); |
duke@435 | 1545 | |
duke@435 | 1546 | for( uint j = b->end_idx() + 1; j > 1; j-- ) { |
duke@435 | 1547 | Node *n = b->_nodes[j-1]; |
duke@435 | 1548 | |
duke@435 | 1549 | // Pre-split compares of loop-phis. Loop-phis form a cycle we would |
duke@435 | 1550 | // like to see in the same register. Compare uses the loop-phi and so |
duke@435 | 1551 | // extends its live range BUT cannot be part of the cycle. If this |
duke@435 | 1552 | // extended live range overlaps with the update of the loop-phi value |
duke@435 | 1553 | // we need both alive at the same time -- which requires at least 1 |
duke@435 | 1554 | // copy. But because Intel has only 2-address registers we end up with |
duke@435 | 1555 | // at least 2 copies, one before the loop-phi update instruction and |
duke@435 | 1556 | // one after. Instead we split the input to the compare just after the |
duke@435 | 1557 | // phi. |
duke@435 | 1558 | if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CmpI ) { |
duke@435 | 1559 | Node *phi = n->in(1); |
duke@435 | 1560 | if( phi->is_Phi() && phi->as_Phi()->region()->is_Loop() ) { |
duke@435 | 1561 | Block *phi_block = _cfg._bbs[phi->_idx]; |
duke@435 | 1562 | if( _cfg._bbs[phi_block->pred(2)->_idx] == b ) { |
duke@435 | 1563 | const RegMask *mask = C->matcher()->idealreg2spillmask[Op_RegI]; |
duke@435 | 1564 | Node *spill = new (C) MachSpillCopyNode( phi, *mask, *mask ); |
duke@435 | 1565 | insert_proj( phi_block, 1, spill, maxlrg++ ); |
duke@435 | 1566 | n->set_req(1,spill); |
duke@435 | 1567 | must_recompute_live = true; |
duke@435 | 1568 | } |
duke@435 | 1569 | } |
duke@435 | 1570 | } |
duke@435 | 1571 | |
duke@435 | 1572 | // Get value being defined |
duke@435 | 1573 | uint lidx = n2lidx(n); |
duke@435 | 1574 | if( lidx && lidx < _maxlrg /* Ignore the occasional brand-new live range */) { |
duke@435 | 1575 | // Remove from live-out set |
duke@435 | 1576 | liveout.remove(lidx); |
duke@435 | 1577 | |
duke@435 | 1578 | // Copies do not define a new value and so do not interfere. |
duke@435 | 1579 | // Remove the copies source from the liveout set before interfering. |
duke@435 | 1580 | uint idx = n->is_Copy(); |
duke@435 | 1581 | if( idx ) liveout.remove( n2lidx(n->in(idx)) ); |
duke@435 | 1582 | } |
duke@435 | 1583 | |
duke@435 | 1584 | // Found a safepoint? |
duke@435 | 1585 | JVMState *jvms = n->jvms(); |
duke@435 | 1586 | if( jvms ) { |
duke@435 | 1587 | // Now scan for a live derived pointer |
duke@435 | 1588 | IndexSetIterator elements(&liveout); |
duke@435 | 1589 | uint neighbor; |
duke@435 | 1590 | while ((neighbor = elements.next()) != 0) { |
duke@435 | 1591 | // Find reaching DEF for base and derived values |
duke@435 | 1592 | // This works because we are still in SSA during this call. |
duke@435 | 1593 | Node *derived = lrgs(neighbor)._def; |
duke@435 | 1594 | const TypePtr *tj = derived->bottom_type()->isa_ptr(); |
kvn@1164 | 1595 | assert(!derived->bottom_type()->isa_narrowoop() || |
kvn@1164 | 1596 | derived->bottom_type()->make_ptr()->is_ptr()->_offset == 0, "sanity"); |
duke@435 | 1597 | // If its an OOP with a non-zero offset, then it is derived. |
duke@435 | 1598 | if( tj && tj->_offset != 0 && tj->isa_oop_ptr() ) { |
duke@435 | 1599 | Node *base = find_base_for_derived( derived_base_map, derived, maxlrg ); |
duke@435 | 1600 | assert( base->_idx < _names.Size(), "" ); |
duke@435 | 1601 | // Add reaching DEFs of derived pointer and base pointer as a |
duke@435 | 1602 | // pair of inputs |
duke@435 | 1603 | n->add_req( derived ); |
duke@435 | 1604 | n->add_req( base ); |
duke@435 | 1605 | |
duke@435 | 1606 | // See if the base pointer is already live to this point. |
duke@435 | 1607 | // Since I'm working on the SSA form, live-ness amounts to |
duke@435 | 1608 | // reaching def's. So if I find the base's live range then |
duke@435 | 1609 | // I know the base's def reaches here. |
duke@435 | 1610 | if( (n2lidx(base) >= _maxlrg ||// (Brand new base (hence not live) or |
duke@435 | 1611 | !liveout.member( n2lidx(base) ) ) && // not live) AND |
duke@435 | 1612 | (n2lidx(base) > 0) && // not a constant |
duke@435 | 1613 | _cfg._bbs[base->_idx] != b ) { // base not def'd in blk) |
duke@435 | 1614 | // Base pointer is not currently live. Since I stretched |
duke@435 | 1615 | // the base pointer to here and it crosses basic-block |
duke@435 | 1616 | // boundaries, the global live info is now incorrect. |
duke@435 | 1617 | // Recompute live. |
duke@435 | 1618 | must_recompute_live = true; |
duke@435 | 1619 | } // End of if base pointer is not live to debug info |
duke@435 | 1620 | } |
duke@435 | 1621 | } // End of scan all live data for derived ptrs crossing GC point |
duke@435 | 1622 | } // End of if found a GC point |
duke@435 | 1623 | |
duke@435 | 1624 | // Make all inputs live |
duke@435 | 1625 | if( !n->is_Phi() ) { // Phi function uses come from prior block |
duke@435 | 1626 | for( uint k = 1; k < n->req(); k++ ) { |
duke@435 | 1627 | uint lidx = n2lidx(n->in(k)); |
duke@435 | 1628 | if( lidx < _maxlrg ) |
duke@435 | 1629 | liveout.insert( lidx ); |
duke@435 | 1630 | } |
duke@435 | 1631 | } |
duke@435 | 1632 | |
duke@435 | 1633 | } // End of forall instructions in block |
duke@435 | 1634 | liveout.clear(); // Free the memory used by liveout. |
duke@435 | 1635 | |
duke@435 | 1636 | } // End of forall blocks |
duke@435 | 1637 | _maxlrg = maxlrg; |
duke@435 | 1638 | |
duke@435 | 1639 | // If I created a new live range I need to recompute live |
duke@435 | 1640 | if( maxlrg != _ifg->_maxlrg ) |
duke@435 | 1641 | must_recompute_live = true; |
duke@435 | 1642 | |
duke@435 | 1643 | return must_recompute_live != 0; |
duke@435 | 1644 | } |
duke@435 | 1645 | |
duke@435 | 1646 | |
duke@435 | 1647 | //------------------------------add_reference---------------------------------- |
duke@435 | 1648 | // Extend the node to LRG mapping |
duke@435 | 1649 | void PhaseChaitin::add_reference( const Node *node, const Node *old_node ) { |
duke@435 | 1650 | _names.extend( node->_idx, n2lidx(old_node) ); |
duke@435 | 1651 | } |
duke@435 | 1652 | |
duke@435 | 1653 | //------------------------------dump------------------------------------------- |
duke@435 | 1654 | #ifndef PRODUCT |
duke@435 | 1655 | void PhaseChaitin::dump( const Node *n ) const { |
duke@435 | 1656 | uint r = (n->_idx < _names.Size() ) ? Find_const(n) : 0; |
duke@435 | 1657 | tty->print("L%d",r); |
duke@435 | 1658 | if( r && n->Opcode() != Op_Phi ) { |
duke@435 | 1659 | if( _node_regs ) { // Got a post-allocation copy of allocation? |
duke@435 | 1660 | tty->print("["); |
duke@435 | 1661 | OptoReg::Name second = get_reg_second(n); |
duke@435 | 1662 | if( OptoReg::is_valid(second) ) { |
duke@435 | 1663 | if( OptoReg::is_reg(second) ) |
duke@435 | 1664 | tty->print("%s:",Matcher::regName[second]); |
duke@435 | 1665 | else |
duke@435 | 1666 | tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(second)); |
duke@435 | 1667 | } |
duke@435 | 1668 | OptoReg::Name first = get_reg_first(n); |
duke@435 | 1669 | if( OptoReg::is_reg(first) ) |
duke@435 | 1670 | tty->print("%s]",Matcher::regName[first]); |
duke@435 | 1671 | else |
duke@435 | 1672 | tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), reg2offset_unchecked(first)); |
duke@435 | 1673 | } else |
duke@435 | 1674 | n->out_RegMask().dump(); |
duke@435 | 1675 | } |
duke@435 | 1676 | tty->print("/N%d\t",n->_idx); |
duke@435 | 1677 | tty->print("%s === ", n->Name()); |
duke@435 | 1678 | uint k; |
duke@435 | 1679 | for( k = 0; k < n->req(); k++) { |
duke@435 | 1680 | Node *m = n->in(k); |
duke@435 | 1681 | if( !m ) tty->print("_ "); |
duke@435 | 1682 | else { |
duke@435 | 1683 | uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0; |
duke@435 | 1684 | tty->print("L%d",r); |
duke@435 | 1685 | // Data MultiNode's can have projections with no real registers. |
duke@435 | 1686 | // Don't die while dumping them. |
duke@435 | 1687 | int op = n->Opcode(); |
duke@435 | 1688 | if( r && op != Op_Phi && op != Op_Proj && op != Op_SCMemProj) { |
duke@435 | 1689 | if( _node_regs ) { |
duke@435 | 1690 | tty->print("["); |
duke@435 | 1691 | OptoReg::Name second = get_reg_second(n->in(k)); |
duke@435 | 1692 | if( OptoReg::is_valid(second) ) { |
duke@435 | 1693 | if( OptoReg::is_reg(second) ) |
duke@435 | 1694 | tty->print("%s:",Matcher::regName[second]); |
duke@435 | 1695 | else |
duke@435 | 1696 | tty->print("%s+%d:",OptoReg::regname(OptoReg::c_frame_pointer), |
duke@435 | 1697 | reg2offset_unchecked(second)); |
duke@435 | 1698 | } |
duke@435 | 1699 | OptoReg::Name first = get_reg_first(n->in(k)); |
duke@435 | 1700 | if( OptoReg::is_reg(first) ) |
duke@435 | 1701 | tty->print("%s]",Matcher::regName[first]); |
duke@435 | 1702 | else |
duke@435 | 1703 | tty->print("%s+%d]",OptoReg::regname(OptoReg::c_frame_pointer), |
duke@435 | 1704 | reg2offset_unchecked(first)); |
duke@435 | 1705 | } else |
duke@435 | 1706 | n->in_RegMask(k).dump(); |
duke@435 | 1707 | } |
duke@435 | 1708 | tty->print("/N%d ",m->_idx); |
duke@435 | 1709 | } |
duke@435 | 1710 | } |
duke@435 | 1711 | if( k < n->len() && n->in(k) ) tty->print("| "); |
duke@435 | 1712 | for( ; k < n->len(); k++ ) { |
duke@435 | 1713 | Node *m = n->in(k); |
duke@435 | 1714 | if( !m ) break; |
duke@435 | 1715 | uint r = (m->_idx < _names.Size() ) ? Find_const(m) : 0; |
duke@435 | 1716 | tty->print("L%d",r); |
duke@435 | 1717 | tty->print("/N%d ",m->_idx); |
duke@435 | 1718 | } |
duke@435 | 1719 | if( n->is_Mach() ) n->as_Mach()->dump_spec(tty); |
duke@435 | 1720 | else n->dump_spec(tty); |
duke@435 | 1721 | if( _spilled_once.test(n->_idx ) ) { |
duke@435 | 1722 | tty->print(" Spill_1"); |
duke@435 | 1723 | if( _spilled_twice.test(n->_idx ) ) |
duke@435 | 1724 | tty->print(" Spill_2"); |
duke@435 | 1725 | } |
duke@435 | 1726 | tty->print("\n"); |
duke@435 | 1727 | } |
duke@435 | 1728 | |
duke@435 | 1729 | void PhaseChaitin::dump( const Block * b ) const { |
duke@435 | 1730 | b->dump_head( &_cfg._bbs ); |
duke@435 | 1731 | |
duke@435 | 1732 | // For all instructions |
duke@435 | 1733 | for( uint j = 0; j < b->_nodes.size(); j++ ) |
duke@435 | 1734 | dump(b->_nodes[j]); |
duke@435 | 1735 | // Print live-out info at end of block |
duke@435 | 1736 | if( _live ) { |
duke@435 | 1737 | tty->print("Liveout: "); |
duke@435 | 1738 | IndexSet *live = _live->live(b); |
duke@435 | 1739 | IndexSetIterator elements(live); |
duke@435 | 1740 | tty->print("{"); |
duke@435 | 1741 | uint i; |
duke@435 | 1742 | while ((i = elements.next()) != 0) { |
duke@435 | 1743 | tty->print("L%d ", Find_const(i)); |
duke@435 | 1744 | } |
duke@435 | 1745 | tty->print_cr("}"); |
duke@435 | 1746 | } |
duke@435 | 1747 | tty->print("\n"); |
duke@435 | 1748 | } |
duke@435 | 1749 | |
duke@435 | 1750 | void PhaseChaitin::dump() const { |
duke@435 | 1751 | tty->print( "--- Chaitin -- argsize: %d framesize: %d ---\n", |
duke@435 | 1752 | _matcher._new_SP, _framesize ); |
duke@435 | 1753 | |
duke@435 | 1754 | // For all blocks |
duke@435 | 1755 | for( uint i = 0; i < _cfg._num_blocks; i++ ) |
duke@435 | 1756 | dump(_cfg._blocks[i]); |
duke@435 | 1757 | // End of per-block dump |
duke@435 | 1758 | tty->print("\n"); |
duke@435 | 1759 | |
duke@435 | 1760 | if (!_ifg) { |
duke@435 | 1761 | tty->print("(No IFG.)\n"); |
duke@435 | 1762 | return; |
duke@435 | 1763 | } |
duke@435 | 1764 | |
duke@435 | 1765 | // Dump LRG array |
duke@435 | 1766 | tty->print("--- Live RanGe Array ---\n"); |
duke@435 | 1767 | for(uint i2 = 1; i2 < _maxlrg; i2++ ) { |
duke@435 | 1768 | tty->print("L%d: ",i2); |
duke@435 | 1769 | if( i2 < _ifg->_maxlrg ) lrgs(i2).dump( ); |
duke@435 | 1770 | else tty->print("new LRG"); |
duke@435 | 1771 | } |
duke@435 | 1772 | tty->print_cr(""); |
duke@435 | 1773 | |
duke@435 | 1774 | // Dump lo-degree list |
duke@435 | 1775 | tty->print("Lo degree: "); |
duke@435 | 1776 | for(uint i3 = _lo_degree; i3; i3 = lrgs(i3)._next ) |
duke@435 | 1777 | tty->print("L%d ",i3); |
duke@435 | 1778 | tty->print_cr(""); |
duke@435 | 1779 | |
duke@435 | 1780 | // Dump lo-stk-degree list |
duke@435 | 1781 | tty->print("Lo stk degree: "); |
duke@435 | 1782 | for(uint i4 = _lo_stk_degree; i4; i4 = lrgs(i4)._next ) |
duke@435 | 1783 | tty->print("L%d ",i4); |
duke@435 | 1784 | tty->print_cr(""); |
duke@435 | 1785 | |
duke@435 | 1786 | // Dump lo-degree list |
duke@435 | 1787 | tty->print("Hi degree: "); |
duke@435 | 1788 | for(uint i5 = _hi_degree; i5; i5 = lrgs(i5)._next ) |
duke@435 | 1789 | tty->print("L%d ",i5); |
duke@435 | 1790 | tty->print_cr(""); |
duke@435 | 1791 | } |
duke@435 | 1792 | |
duke@435 | 1793 | //------------------------------dump_degree_lists------------------------------ |
duke@435 | 1794 | void PhaseChaitin::dump_degree_lists() const { |
duke@435 | 1795 | // Dump lo-degree list |
duke@435 | 1796 | tty->print("Lo degree: "); |
duke@435 | 1797 | for( uint i = _lo_degree; i; i = lrgs(i)._next ) |
duke@435 | 1798 | tty->print("L%d ",i); |
duke@435 | 1799 | tty->print_cr(""); |
duke@435 | 1800 | |
duke@435 | 1801 | // Dump lo-stk-degree list |
duke@435 | 1802 | tty->print("Lo stk degree: "); |
duke@435 | 1803 | for(uint i2 = _lo_stk_degree; i2; i2 = lrgs(i2)._next ) |
duke@435 | 1804 | tty->print("L%d ",i2); |
duke@435 | 1805 | tty->print_cr(""); |
duke@435 | 1806 | |
duke@435 | 1807 | // Dump lo-degree list |
duke@435 | 1808 | tty->print("Hi degree: "); |
duke@435 | 1809 | for(uint i3 = _hi_degree; i3; i3 = lrgs(i3)._next ) |
duke@435 | 1810 | tty->print("L%d ",i3); |
duke@435 | 1811 | tty->print_cr(""); |
duke@435 | 1812 | } |
duke@435 | 1813 | |
duke@435 | 1814 | //------------------------------dump_simplified-------------------------------- |
duke@435 | 1815 | void PhaseChaitin::dump_simplified() const { |
duke@435 | 1816 | tty->print("Simplified: "); |
duke@435 | 1817 | for( uint i = _simplified; i; i = lrgs(i)._next ) |
duke@435 | 1818 | tty->print("L%d ",i); |
duke@435 | 1819 | tty->print_cr(""); |
duke@435 | 1820 | } |
duke@435 | 1821 | |
duke@435 | 1822 | static char *print_reg( OptoReg::Name reg, const PhaseChaitin *pc, char *buf ) { |
duke@435 | 1823 | if ((int)reg < 0) |
duke@435 | 1824 | sprintf(buf, "<OptoReg::%d>", (int)reg); |
duke@435 | 1825 | else if (OptoReg::is_reg(reg)) |
duke@435 | 1826 | strcpy(buf, Matcher::regName[reg]); |
duke@435 | 1827 | else |
duke@435 | 1828 | sprintf(buf,"%s + #%d",OptoReg::regname(OptoReg::c_frame_pointer), |
duke@435 | 1829 | pc->reg2offset(reg)); |
duke@435 | 1830 | return buf+strlen(buf); |
duke@435 | 1831 | } |
duke@435 | 1832 | |
duke@435 | 1833 | //------------------------------dump_register---------------------------------- |
duke@435 | 1834 | // Dump a register name into a buffer. Be intelligent if we get called |
duke@435 | 1835 | // before allocation is complete. |
duke@435 | 1836 | char *PhaseChaitin::dump_register( const Node *n, char *buf ) const { |
duke@435 | 1837 | if( !this ) { // Not got anything? |
duke@435 | 1838 | sprintf(buf,"N%d",n->_idx); // Then use Node index |
duke@435 | 1839 | } else if( _node_regs ) { |
duke@435 | 1840 | // Post allocation, use direct mappings, no LRG info available |
duke@435 | 1841 | print_reg( get_reg_first(n), this, buf ); |
duke@435 | 1842 | } else { |
duke@435 | 1843 | uint lidx = Find_const(n); // Grab LRG number |
duke@435 | 1844 | if( !_ifg ) { |
duke@435 | 1845 | sprintf(buf,"L%d",lidx); // No register binding yet |
duke@435 | 1846 | } else if( !lidx ) { // Special, not allocated value |
duke@435 | 1847 | strcpy(buf,"Special"); |
duke@435 | 1848 | } else if( (lrgs(lidx).num_regs() == 1) |
duke@435 | 1849 | ? !lrgs(lidx).mask().is_bound1() |
duke@435 | 1850 | : !lrgs(lidx).mask().is_bound2() ) { |
duke@435 | 1851 | sprintf(buf,"L%d",lidx); // No register binding yet |
duke@435 | 1852 | } else { // Hah! We have a bound machine register |
duke@435 | 1853 | print_reg( lrgs(lidx).reg(), this, buf ); |
duke@435 | 1854 | } |
duke@435 | 1855 | } |
duke@435 | 1856 | return buf+strlen(buf); |
duke@435 | 1857 | } |
duke@435 | 1858 | |
duke@435 | 1859 | //----------------------dump_for_spill_split_recycle-------------------------- |
duke@435 | 1860 | void PhaseChaitin::dump_for_spill_split_recycle() const { |
duke@435 | 1861 | if( WizardMode && (PrintCompilation || PrintOpto) ) { |
duke@435 | 1862 | // Display which live ranges need to be split and the allocator's state |
duke@435 | 1863 | tty->print_cr("Graph-Coloring Iteration %d will split the following live ranges", _trip_cnt); |
duke@435 | 1864 | for( uint bidx = 1; bidx < _maxlrg; bidx++ ) { |
duke@435 | 1865 | if( lrgs(bidx).alive() && lrgs(bidx).reg() >= LRG::SPILL_REG ) { |
duke@435 | 1866 | tty->print("L%d: ", bidx); |
duke@435 | 1867 | lrgs(bidx).dump(); |
duke@435 | 1868 | } |
duke@435 | 1869 | } |
duke@435 | 1870 | tty->cr(); |
duke@435 | 1871 | dump(); |
duke@435 | 1872 | } |
duke@435 | 1873 | } |
duke@435 | 1874 | |
duke@435 | 1875 | //------------------------------dump_frame------------------------------------ |
duke@435 | 1876 | void PhaseChaitin::dump_frame() const { |
duke@435 | 1877 | const char *fp = OptoReg::regname(OptoReg::c_frame_pointer); |
duke@435 | 1878 | const TypeTuple *domain = C->tf()->domain(); |
duke@435 | 1879 | const int argcnt = domain->cnt() - TypeFunc::Parms; |
duke@435 | 1880 | |
duke@435 | 1881 | // Incoming arguments in registers dump |
duke@435 | 1882 | for( int k = 0; k < argcnt; k++ ) { |
duke@435 | 1883 | OptoReg::Name parmreg = _matcher._parm_regs[k].first(); |
duke@435 | 1884 | if( OptoReg::is_reg(parmreg)) { |
duke@435 | 1885 | const char *reg_name = OptoReg::regname(parmreg); |
duke@435 | 1886 | tty->print("#r%3.3d %s", parmreg, reg_name); |
duke@435 | 1887 | parmreg = _matcher._parm_regs[k].second(); |
duke@435 | 1888 | if( OptoReg::is_reg(parmreg)) { |
duke@435 | 1889 | tty->print(":%s", OptoReg::regname(parmreg)); |
duke@435 | 1890 | } |
duke@435 | 1891 | tty->print(" : parm %d: ", k); |
duke@435 | 1892 | domain->field_at(k + TypeFunc::Parms)->dump(); |
duke@435 | 1893 | tty->print_cr(""); |
duke@435 | 1894 | } |
duke@435 | 1895 | } |
duke@435 | 1896 | |
duke@435 | 1897 | // Check for un-owned padding above incoming args |
duke@435 | 1898 | OptoReg::Name reg = _matcher._new_SP; |
duke@435 | 1899 | if( reg > _matcher._in_arg_limit ) { |
duke@435 | 1900 | reg = OptoReg::add(reg, -1); |
duke@435 | 1901 | tty->print_cr("#r%3.3d %s+%2d: pad0, owned by CALLER", reg, fp, reg2offset_unchecked(reg)); |
duke@435 | 1902 | } |
duke@435 | 1903 | |
duke@435 | 1904 | // Incoming argument area dump |
duke@435 | 1905 | OptoReg::Name begin_in_arg = OptoReg::add(_matcher._old_SP,C->out_preserve_stack_slots()); |
duke@435 | 1906 | while( reg > begin_in_arg ) { |
duke@435 | 1907 | reg = OptoReg::add(reg, -1); |
duke@435 | 1908 | tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1909 | int j; |
duke@435 | 1910 | for( j = 0; j < argcnt; j++) { |
duke@435 | 1911 | if( _matcher._parm_regs[j].first() == reg || |
duke@435 | 1912 | _matcher._parm_regs[j].second() == reg ) { |
duke@435 | 1913 | tty->print("parm %d: ",j); |
duke@435 | 1914 | domain->field_at(j + TypeFunc::Parms)->dump(); |
duke@435 | 1915 | tty->print_cr(""); |
duke@435 | 1916 | break; |
duke@435 | 1917 | } |
duke@435 | 1918 | } |
duke@435 | 1919 | if( j >= argcnt ) |
duke@435 | 1920 | tty->print_cr("HOLE, owned by SELF"); |
duke@435 | 1921 | } |
duke@435 | 1922 | |
duke@435 | 1923 | // Old outgoing preserve area |
duke@435 | 1924 | while( reg > _matcher._old_SP ) { |
duke@435 | 1925 | reg = OptoReg::add(reg, -1); |
duke@435 | 1926 | tty->print_cr("#r%3.3d %s+%2d: old out preserve",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1927 | } |
duke@435 | 1928 | |
duke@435 | 1929 | // Old SP |
duke@435 | 1930 | tty->print_cr("# -- Old %s -- Framesize: %d --",fp, |
duke@435 | 1931 | reg2offset_unchecked(OptoReg::add(_matcher._old_SP,-1)) - reg2offset_unchecked(_matcher._new_SP)+jintSize); |
duke@435 | 1932 | |
duke@435 | 1933 | // Preserve area dump |
duke@435 | 1934 | reg = OptoReg::add(reg, -1); |
duke@435 | 1935 | while( OptoReg::is_stack(reg)) { |
duke@435 | 1936 | tty->print("#r%3.3d %s+%2d: ",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1937 | if( _matcher.return_addr() == reg ) |
duke@435 | 1938 | tty->print_cr("return address"); |
duke@435 | 1939 | else if( _matcher.return_addr() == OptoReg::add(reg,1) && |
duke@435 | 1940 | VerifyStackAtCalls ) |
duke@435 | 1941 | tty->print_cr("0xBADB100D +VerifyStackAtCalls"); |
duke@435 | 1942 | else if ((int)OptoReg::reg2stack(reg) < C->fixed_slots()) |
duke@435 | 1943 | tty->print_cr("Fixed slot %d", OptoReg::reg2stack(reg)); |
duke@435 | 1944 | else |
duke@435 | 1945 | tty->print_cr("pad2, in_preserve"); |
duke@435 | 1946 | reg = OptoReg::add(reg, -1); |
duke@435 | 1947 | } |
duke@435 | 1948 | |
duke@435 | 1949 | // Spill area dump |
duke@435 | 1950 | reg = OptoReg::add(_matcher._new_SP, _framesize ); |
duke@435 | 1951 | while( reg > _matcher._out_arg_limit ) { |
duke@435 | 1952 | reg = OptoReg::add(reg, -1); |
duke@435 | 1953 | tty->print_cr("#r%3.3d %s+%2d: spill",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1954 | } |
duke@435 | 1955 | |
duke@435 | 1956 | // Outgoing argument area dump |
duke@435 | 1957 | while( reg > OptoReg::add(_matcher._new_SP, C->out_preserve_stack_slots()) ) { |
duke@435 | 1958 | reg = OptoReg::add(reg, -1); |
duke@435 | 1959 | tty->print_cr("#r%3.3d %s+%2d: outgoing argument",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1960 | } |
duke@435 | 1961 | |
duke@435 | 1962 | // Outgoing new preserve area |
duke@435 | 1963 | while( reg > _matcher._new_SP ) { |
duke@435 | 1964 | reg = OptoReg::add(reg, -1); |
duke@435 | 1965 | tty->print_cr("#r%3.3d %s+%2d: new out preserve",reg,fp,reg2offset_unchecked(reg)); |
duke@435 | 1966 | } |
duke@435 | 1967 | tty->print_cr("#"); |
duke@435 | 1968 | } |
duke@435 | 1969 | |
duke@435 | 1970 | //------------------------------dump_bb---------------------------------------- |
duke@435 | 1971 | void PhaseChaitin::dump_bb( uint pre_order ) const { |
duke@435 | 1972 | tty->print_cr("---dump of B%d---",pre_order); |
duke@435 | 1973 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 1974 | Block *b = _cfg._blocks[i]; |
duke@435 | 1975 | if( b->_pre_order == pre_order ) |
duke@435 | 1976 | dump(b); |
duke@435 | 1977 | } |
duke@435 | 1978 | } |
duke@435 | 1979 | |
duke@435 | 1980 | //------------------------------dump_lrg--------------------------------------- |
duke@435 | 1981 | void PhaseChaitin::dump_lrg( uint lidx ) const { |
duke@435 | 1982 | tty->print_cr("---dump of L%d---",lidx); |
duke@435 | 1983 | |
duke@435 | 1984 | if( _ifg ) { |
duke@435 | 1985 | if( lidx >= _maxlrg ) { |
duke@435 | 1986 | tty->print("Attempt to print live range index beyond max live range.\n"); |
duke@435 | 1987 | return; |
duke@435 | 1988 | } |
duke@435 | 1989 | tty->print("L%d: ",lidx); |
duke@435 | 1990 | lrgs(lidx).dump( ); |
duke@435 | 1991 | } |
duke@435 | 1992 | if( _ifg ) { tty->print("Neighbors: %d - ", _ifg->neighbor_cnt(lidx)); |
duke@435 | 1993 | _ifg->neighbors(lidx)->dump(); |
duke@435 | 1994 | tty->cr(); |
duke@435 | 1995 | } |
duke@435 | 1996 | // For all blocks |
duke@435 | 1997 | for( uint i = 0; i < _cfg._num_blocks; i++ ) { |
duke@435 | 1998 | Block *b = _cfg._blocks[i]; |
duke@435 | 1999 | int dump_once = 0; |
duke@435 | 2000 | |
duke@435 | 2001 | // For all instructions |
duke@435 | 2002 | for( uint j = 0; j < b->_nodes.size(); j++ ) { |
duke@435 | 2003 | Node *n = b->_nodes[j]; |
duke@435 | 2004 | if( Find_const(n) == lidx ) { |
duke@435 | 2005 | if( !dump_once++ ) { |
duke@435 | 2006 | tty->cr(); |
duke@435 | 2007 | b->dump_head( &_cfg._bbs ); |
duke@435 | 2008 | } |
duke@435 | 2009 | dump(n); |
duke@435 | 2010 | continue; |
duke@435 | 2011 | } |
duke@435 | 2012 | uint cnt = n->req(); |
duke@435 | 2013 | for( uint k = 1; k < cnt; k++ ) { |
duke@435 | 2014 | Node *m = n->in(k); |
duke@435 | 2015 | if (!m) continue; // be robust in the dumper |
duke@435 | 2016 | if( Find_const(m) == lidx ) { |
duke@435 | 2017 | if( !dump_once++ ) { |
duke@435 | 2018 | tty->cr(); |
duke@435 | 2019 | b->dump_head( &_cfg._bbs ); |
duke@435 | 2020 | } |
duke@435 | 2021 | dump(n); |
duke@435 | 2022 | } |
duke@435 | 2023 | } |
duke@435 | 2024 | } |
duke@435 | 2025 | } // End of per-block dump |
duke@435 | 2026 | tty->cr(); |
duke@435 | 2027 | } |
duke@435 | 2028 | #endif // not PRODUCT |
duke@435 | 2029 | |
duke@435 | 2030 | //------------------------------print_chaitin_statistics------------------------------- |
duke@435 | 2031 | int PhaseChaitin::_final_loads = 0; |
duke@435 | 2032 | int PhaseChaitin::_final_stores = 0; |
duke@435 | 2033 | int PhaseChaitin::_final_memoves= 0; |
duke@435 | 2034 | int PhaseChaitin::_final_copies = 0; |
duke@435 | 2035 | double PhaseChaitin::_final_load_cost = 0; |
duke@435 | 2036 | double PhaseChaitin::_final_store_cost = 0; |
duke@435 | 2037 | double PhaseChaitin::_final_memove_cost= 0; |
duke@435 | 2038 | double PhaseChaitin::_final_copy_cost = 0; |
duke@435 | 2039 | int PhaseChaitin::_conserv_coalesce = 0; |
duke@435 | 2040 | int PhaseChaitin::_conserv_coalesce_pair = 0; |
duke@435 | 2041 | int PhaseChaitin::_conserv_coalesce_trie = 0; |
duke@435 | 2042 | int PhaseChaitin::_conserv_coalesce_quad = 0; |
duke@435 | 2043 | int PhaseChaitin::_post_alloc = 0; |
duke@435 | 2044 | int PhaseChaitin::_lost_opp_pp_coalesce = 0; |
duke@435 | 2045 | int PhaseChaitin::_lost_opp_cflow_coalesce = 0; |
duke@435 | 2046 | int PhaseChaitin::_used_cisc_instructions = 0; |
duke@435 | 2047 | int PhaseChaitin::_unused_cisc_instructions = 0; |
duke@435 | 2048 | int PhaseChaitin::_allocator_attempts = 0; |
duke@435 | 2049 | int PhaseChaitin::_allocator_successes = 0; |
duke@435 | 2050 | |
duke@435 | 2051 | #ifndef PRODUCT |
duke@435 | 2052 | uint PhaseChaitin::_high_pressure = 0; |
duke@435 | 2053 | uint PhaseChaitin::_low_pressure = 0; |
duke@435 | 2054 | |
duke@435 | 2055 | void PhaseChaitin::print_chaitin_statistics() { |
duke@435 | 2056 | 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); |
duke@435 | 2057 | 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); |
duke@435 | 2058 | tty->print_cr("Adjusted spill cost = %7.0f.", |
duke@435 | 2059 | _final_load_cost*4.0 + _final_store_cost * 2.0 + |
duke@435 | 2060 | _final_copy_cost*1.0 + _final_memove_cost*12.0); |
duke@435 | 2061 | tty->print("Conservatively coalesced %d copies, %d pairs", |
duke@435 | 2062 | _conserv_coalesce, _conserv_coalesce_pair); |
duke@435 | 2063 | if( _conserv_coalesce_trie || _conserv_coalesce_quad ) |
duke@435 | 2064 | tty->print(", %d tries, %d quads", _conserv_coalesce_trie, _conserv_coalesce_quad); |
duke@435 | 2065 | tty->print_cr(", %d post alloc.", _post_alloc); |
duke@435 | 2066 | if( _lost_opp_pp_coalesce || _lost_opp_cflow_coalesce ) |
duke@435 | 2067 | tty->print_cr("Lost coalesce opportunity, %d private-private, and %d cflow interfered.", |
duke@435 | 2068 | _lost_opp_pp_coalesce, _lost_opp_cflow_coalesce ); |
duke@435 | 2069 | if( _used_cisc_instructions || _unused_cisc_instructions ) |
duke@435 | 2070 | tty->print_cr("Used cisc instruction %d, remained in register %d", |
duke@435 | 2071 | _used_cisc_instructions, _unused_cisc_instructions); |
duke@435 | 2072 | if( _allocator_successes != 0 ) |
duke@435 | 2073 | tty->print_cr("Average allocation trips %f", (float)_allocator_attempts/(float)_allocator_successes); |
duke@435 | 2074 | tty->print_cr("High Pressure Blocks = %d, Low Pressure Blocks = %d", _high_pressure, _low_pressure); |
duke@435 | 2075 | } |
duke@435 | 2076 | #endif // not PRODUCT |