Mercurial > hg > graal-compiler
annotate src/share/vm/opto/coalesce.cpp @ 10111:8373c19be854
8011621: live_ranges_in_separate_class.patch
Reviewed-by: kvn, roland
Contributed-by: niclas.adlertz@oracle.com
| author | neliasso |
|---|---|
| date | Tue, 16 Apr 2013 10:08:41 +0200 |
| parents | c7b60b601eb4 |
| children | 693e4d04fd09 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 1972 | 2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
844
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #include "precompiled.hpp" |
| 26 #include "memory/allocation.inline.hpp" | |
| 27 #include "opto/block.hpp" | |
| 28 #include "opto/cfgnode.hpp" | |
| 29 #include "opto/chaitin.hpp" | |
| 30 #include "opto/coalesce.hpp" | |
| 31 #include "opto/connode.hpp" | |
| 32 #include "opto/indexSet.hpp" | |
| 33 #include "opto/machnode.hpp" | |
| 34 #include "opto/matcher.hpp" | |
| 35 #include "opto/regmask.hpp" | |
| 0 | 36 |
| 37 //============================================================================= | |
| 38 //------------------------------Dump------------------------------------------- | |
| 39 #ifndef PRODUCT | |
| 10111 | 40 void PhaseCoalesce::dump(Node *n) const { |
| 0 | 41 // Being a const function means I cannot use 'Find' |
| 10111 | 42 uint r = _phc._lrg_map.find(n); |
| 0 | 43 tty->print("L%d/N%d ",r,n->_idx); |
| 44 } | |
| 45 | |
| 46 //------------------------------dump------------------------------------------- | |
| 47 void PhaseCoalesce::dump() const { | |
| 48 // I know I have a block layout now, so I can print blocks in a loop | |
| 49 for( uint i=0; i<_phc._cfg._num_blocks; i++ ) { | |
| 50 uint j; | |
| 51 Block *b = _phc._cfg._blocks[i]; | |
| 52 // Print a nice block header | |
| 53 tty->print("B%d: ",b->_pre_order); | |
| 54 for( j=1; j<b->num_preds(); j++ ) | |
| 55 tty->print("B%d ", _phc._cfg._bbs[b->pred(j)->_idx]->_pre_order); | |
| 56 tty->print("-> "); | |
| 57 for( j=0; j<b->_num_succs; j++ ) | |
| 58 tty->print("B%d ",b->_succs[j]->_pre_order); | |
| 59 tty->print(" IDom: B%d/#%d\n", b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth); | |
| 60 uint cnt = b->_nodes.size(); | |
| 61 for( j=0; j<cnt; j++ ) { | |
| 62 Node *n = b->_nodes[j]; | |
| 63 dump( n ); | |
| 64 tty->print("\t%s\t",n->Name()); | |
| 65 | |
| 66 // Dump the inputs | |
| 67 uint k; // Exit value of loop | |
| 68 for( k=0; k<n->req(); k++ ) // For all required inputs | |
| 69 if( n->in(k) ) dump( n->in(k) ); | |
| 70 else tty->print("_ "); | |
| 71 int any_prec = 0; | |
| 72 for( ; k<n->len(); k++ ) // For all precedence inputs | |
| 73 if( n->in(k) ) { | |
| 74 if( !any_prec++ ) tty->print(" |"); | |
| 75 dump( n->in(k) ); | |
| 76 } | |
| 77 | |
| 78 // Dump node-specific info | |
| 79 n->dump_spec(tty); | |
| 80 tty->print("\n"); | |
| 81 | |
| 82 } | |
| 83 tty->print("\n"); | |
| 84 } | |
| 85 } | |
| 86 #endif | |
| 87 | |
| 88 //------------------------------combine_these_two------------------------------ | |
| 89 // Combine the live ranges def'd by these 2 Nodes. N2 is an input to N1. | |
| 10111 | 90 void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) { |
| 91 uint lr1 = _phc._lrg_map.find(n1); | |
| 92 uint lr2 = _phc._lrg_map.find(n2); | |
| 0 | 93 if( lr1 != lr2 && // Different live ranges already AND |
| 94 !_phc._ifg->test_edge_sq( lr1, lr2 ) ) { // Do not interfere | |
| 95 LRG *lrg1 = &_phc.lrgs(lr1); | |
| 96 LRG *lrg2 = &_phc.lrgs(lr2); | |
| 97 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. | |
| 98 | |
| 99 // Now, why is int->oop OK? We end up declaring a raw-pointer as an oop | |
| 100 // and in general that's a bad thing. However, int->oop conversions only | |
| 101 // happen at GC points, so the lifetime of the misclassified raw-pointer | |
| 102 // is from the CheckCastPP (that converts it to an oop) backwards up | |
| 103 // through a merge point and into the slow-path call, and around the | |
| 104 // diamond up to the heap-top check and back down into the slow-path call. | |
| 105 // The misclassified raw pointer is NOT live across the slow-path call, | |
| 106 // and so does not appear in any GC info, so the fact that it is | |
| 107 // misclassified is OK. | |
| 108 | |
| 109 if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND | |
| 110 // Compatible final mask | |
| 111 lrg1->mask().overlap( lrg2->mask() ) ) { | |
| 112 // Merge larger into smaller. | |
| 113 if( lr1 > lr2 ) { | |
| 114 uint tmp = lr1; lr1 = lr2; lr2 = tmp; | |
| 115 Node *n = n1; n1 = n2; n2 = n; | |
| 116 LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp; | |
| 117 } | |
| 118 // Union lr2 into lr1 | |
| 119 _phc.Union( n1, n2 ); | |
| 120 if (lrg1->_maxfreq < lrg2->_maxfreq) | |
| 121 lrg1->_maxfreq = lrg2->_maxfreq; | |
| 122 // Merge in the IFG | |
| 123 _phc._ifg->Union( lr1, lr2 ); | |
| 124 // Combine register restrictions | |
| 125 lrg1->AND(lrg2->mask()); | |
| 126 } | |
| 127 } | |
| 128 } | |
| 129 | |
| 130 //------------------------------coalesce_driver-------------------------------- | |
| 131 // Copy coalescing | |
| 132 void PhaseCoalesce::coalesce_driver( ) { | |
| 133 | |
| 134 verify(); | |
| 135 // Coalesce from high frequency to low | |
| 136 for( uint i=0; i<_phc._cfg._num_blocks; i++ ) | |
| 137 coalesce( _phc._blks[i] ); | |
| 138 | |
| 139 } | |
| 140 | |
| 141 //------------------------------insert_copy_with_overlap----------------------- | |
| 142 // I am inserting copies to come out of SSA form. In the general case, I am | |
| 143 // doing a parallel renaming. I'm in the Named world now, so I can't do a | |
| 144 // general parallel renaming. All the copies now use "names" (live-ranges) | |
| 145 // to carry values instead of the explicit use-def chains. Suppose I need to | |
| 146 // insert 2 copies into the same block. They copy L161->L128 and L128->L132. | |
| 147 // If I insert them in the wrong order then L128 will get clobbered before it | |
| 148 // can get used by the second copy. This cannot happen in the SSA model; | |
| 149 // direct use-def chains get me the right value. It DOES happen in the named | |
| 150 // model so I have to handle the reordering of copies. | |
| 151 // | |
| 152 // In general, I need to topo-sort the placed copies to avoid conflicts. | |
| 153 // Its possible to have a closed cycle of copies (e.g., recirculating the same | |
| 154 // values around a loop). In this case I need a temp to break the cycle. | |
| 155 void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) { | |
| 156 | |
| 157 // Scan backwards for the locations of the last use of the dst_name. | |
| 158 // I am about to clobber the dst_name, so the copy must be inserted | |
| 159 // after the last use. Last use is really first-use on a backwards scan. | |
| 160 uint i = b->end_idx()-1; | |
| 10111 | 161 while(1) { |
| 0 | 162 Node *n = b->_nodes[i]; |
| 163 // Check for end of virtual copies; this is also the end of the | |
| 164 // parallel renaming effort. | |
| 10111 | 165 if (n->_idx < _unique) { |
| 166 break; | |
| 167 } | |
| 0 | 168 uint idx = n->is_Copy(); |
| 3842 | 169 assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); |
| 10111 | 170 if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) { |
| 171 break; | |
| 172 } | |
| 0 | 173 i--; |
| 174 } | |
| 175 uint last_use_idx = i; | |
| 176 | |
| 177 // Also search for any kill of src_name that exits the block. | |
| 178 // Since the copy uses src_name, I have to come before any kill. | |
| 179 uint kill_src_idx = b->end_idx(); | |
| 180 // There can be only 1 kill that exits any block and that is | |
| 181 // the last kill. Thus it is the first kill on a backwards scan. | |
| 182 i = b->end_idx()-1; | |
| 10111 | 183 while (1) { |
| 0 | 184 Node *n = b->_nodes[i]; |
| 185 // Check for end of virtual copies; this is also the end of the | |
| 186 // parallel renaming effort. | |
| 10111 | 187 if (n->_idx < _unique) { |
| 188 break; | |
| 189 } | |
| 3842 | 190 assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" ); |
| 10111 | 191 if (_phc._lrg_map.find(n) == src_name) { |
| 0 | 192 kill_src_idx = i; |
| 193 break; | |
| 194 } | |
| 195 i--; | |
| 196 } | |
| 197 // Need a temp? Last use of dst comes after the kill of src? | |
| 10111 | 198 if (last_use_idx >= kill_src_idx) { |
| 0 | 199 // Need to break a cycle with a temp |
| 200 uint idx = copy->is_Copy(); | |
| 201 Node *tmp = copy->clone(); | |
| 10111 | 202 uint max_lrg_id = _phc._lrg_map.max_lrg_id(); |
| 203 _phc.new_lrg(tmp, max_lrg_id); | |
| 204 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); | |
| 205 | |
| 0 | 206 // Insert new temp between copy and source |
| 207 tmp ->set_req(idx,copy->in(idx)); | |
| 208 copy->set_req(idx,tmp); | |
| 209 // Save source in temp early, before source is killed | |
| 210 b->_nodes.insert(kill_src_idx,tmp); | |
| 211 _phc._cfg._bbs.map( tmp->_idx, b ); | |
| 212 last_use_idx++; | |
| 213 } | |
| 214 | |
| 215 // Insert just after last use | |
| 216 b->_nodes.insert(last_use_idx+1,copy); | |
| 217 } | |
| 218 | |
| 219 //------------------------------insert_copies---------------------------------- | |
| 220 void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) { | |
| 221 // We do LRGs compressing and fix a liveout data only here since the other | |
| 222 // place in Split() is guarded by the assert which we never hit. | |
| 10111 | 223 _phc._lrg_map.compress_uf_map_for_nodes(); |
| 0 | 224 // Fix block's liveout data for compressed live ranges. |
| 10111 | 225 for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) { |
| 226 uint compressed_lrg = _phc._lrg_map.find(lrg); | |
| 227 if (lrg != compressed_lrg) { | |
| 228 for (uint bidx = 0; bidx < _phc._cfg._num_blocks; bidx++) { | |
| 0 | 229 IndexSet *liveout = _phc._live->live(_phc._cfg._blocks[bidx]); |
| 10111 | 230 if (liveout->member(lrg)) { |
| 0 | 231 liveout->remove(lrg); |
| 232 liveout->insert(compressed_lrg); | |
| 233 } | |
| 234 } | |
| 235 } | |
| 236 } | |
| 237 | |
| 238 // All new nodes added are actual copies to replace virtual copies. | |
| 239 // Nodes with index less than '_unique' are original, non-virtual Nodes. | |
| 240 _unique = C->unique(); | |
| 241 | |
| 242 for( uint i=0; i<_phc._cfg._num_blocks; i++ ) { | |
| 243 Block *b = _phc._cfg._blocks[i]; | |
| 244 uint cnt = b->num_preds(); // Number of inputs to the Phi | |
| 245 | |
| 246 for( uint l = 1; l<b->_nodes.size(); l++ ) { | |
| 247 Node *n = b->_nodes[l]; | |
| 248 | |
| 249 // Do not use removed-copies, use copied value instead | |
| 250 uint ncnt = n->req(); | |
| 251 for( uint k = 1; k<ncnt; k++ ) { | |
| 252 Node *copy = n->in(k); | |
| 253 uint cidx = copy->is_Copy(); | |
| 254 if( cidx ) { | |
| 255 Node *def = copy->in(cidx); | |
| 10111 | 256 if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) { |
| 257 n->set_req(k, def); | |
| 258 } | |
| 0 | 259 } |
| 260 } | |
| 261 | |
| 262 // Remove any explicit copies that get coalesced. | |
| 263 uint cidx = n->is_Copy(); | |
| 264 if( cidx ) { | |
| 265 Node *def = n->in(cidx); | |
| 10111 | 266 if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) { |
| 0 | 267 n->replace_by(def); |
| 268 n->set_req(cidx,NULL); | |
| 269 b->_nodes.remove(l); | |
| 270 l--; | |
| 271 continue; | |
| 272 } | |
| 273 } | |
| 274 | |
| 10111 | 275 if (n->is_Phi()) { |
| 0 | 276 // Get the chosen name for the Phi |
| 10111 | 277 uint phi_name = _phc._lrg_map.find(n); |
| 0 | 278 // Ignore the pre-allocated specials |
| 10111 | 279 if (!phi_name) { |
| 280 continue; | |
| 281 } | |
| 0 | 282 // Check for mismatch inputs to Phi |
| 10111 | 283 for (uint j = 1; j < cnt; j++) { |
| 0 | 284 Node *m = n->in(j); |
| 10111 | 285 uint src_name = _phc._lrg_map.find(m); |
| 286 if (src_name != phi_name) { | |
| 0 | 287 Block *pred = _phc._cfg._bbs[b->pred(j)->_idx]; |
| 288 Node *copy; | |
| 289 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); | |
| 290 // Rematerialize constants instead of copying them | |
| 291 if( m->is_Mach() && m->as_Mach()->is_Con() && | |
| 292 m->as_Mach()->rematerialize() ) { | |
| 293 copy = m->clone(); | |
| 294 // Insert the copy in the predecessor basic block | |
| 295 pred->add_inst(copy); | |
| 296 // Copy any flags as well | |
| 10111 | 297 _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map); |
| 0 | 298 } else { |
| 299 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; | |
| 10111 | 300 copy = new (C) MachSpillCopyNode(m, *rm, *rm); |
| 0 | 301 // Find a good place to insert. Kinda tricky, use a subroutine |
| 302 insert_copy_with_overlap(pred,copy,phi_name,src_name); | |
| 303 } | |
| 304 // Insert the copy in the use-def chain | |
| 10111 | 305 n->set_req(j, copy); |
| 0 | 306 _phc._cfg._bbs.map( copy->_idx, pred ); |
| 307 // Extend ("register allocate") the names array for the copy. | |
| 10111 | 308 _phc._lrg_map.extend(copy->_idx, phi_name); |
| 0 | 309 } // End of if Phi names do not match |
| 310 } // End of for all inputs to Phi | |
| 311 } else { // End of if Phi | |
| 312 | |
| 313 // Now check for 2-address instructions | |
| 314 uint idx; | |
| 315 if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) { | |
| 316 // Get the chosen name for the Node | |
| 10111 | 317 uint name = _phc._lrg_map.find(n); |
| 318 assert (name, "no 2-address specials"); | |
| 0 | 319 // Check for name mis-match on the 2-address input |
| 320 Node *m = n->in(idx); | |
| 10111 | 321 if (_phc._lrg_map.find(m) != name) { |
| 0 | 322 Node *copy; |
| 323 assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); | |
| 324 // At this point it is unsafe to extend live ranges (6550579). | |
| 325 // Rematerialize only constants as we do for Phi above. | |
| 10111 | 326 if(m->is_Mach() && m->as_Mach()->is_Con() && |
| 327 m->as_Mach()->rematerialize()) { | |
| 0 | 328 copy = m->clone(); |
| 329 // Insert the copy in the basic block, just before us | |
| 10111 | 330 b->_nodes.insert(l++, copy); |
| 331 if(_phc.clone_projs(b, l, m, copy, _phc._lrg_map)) { | |
| 0 | 332 l++; |
| 10111 | 333 } |
| 0 | 334 } else { |
| 335 const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; | |
| 10111 | 336 copy = new (C) MachSpillCopyNode(m, *rm, *rm); |
| 0 | 337 // Insert the copy in the basic block, just before us |
| 10111 | 338 b->_nodes.insert(l++, copy); |
| 0 | 339 } |
| 340 // Insert the copy in the use-def chain | |
| 10111 | 341 n->set_req(idx, copy); |
| 0 | 342 // Extend ("register allocate") the names array for the copy. |
| 10111 | 343 _phc._lrg_map.extend(copy->_idx, name); |
| 0 | 344 _phc._cfg._bbs.map( copy->_idx, b ); |
| 345 } | |
| 346 | |
| 347 } // End of is two-adr | |
| 348 | |
| 349 // Insert a copy at a debug use for a lrg which has high frequency | |
| 10111 | 350 if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || b->is_uncommon(_phc._cfg._bbs)) { |
| 0 | 351 // Walk the debug inputs to the node and check for lrg freq |
| 352 JVMState* jvms = n->jvms(); | |
| 353 uint debug_start = jvms ? jvms->debug_start() : 999999; | |
| 354 uint debug_end = jvms ? jvms->debug_end() : 999999; | |
| 355 for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) { | |
| 356 // Do not split monitors; they are only needed for debug table | |
| 357 // entries and need no code. | |
| 10111 | 358 if (jvms->is_monitor_use(inpidx)) { |
| 359 continue; | |
| 360 } | |
| 0 | 361 Node *inp = n->in(inpidx); |
| 10111 | 362 uint nidx = _phc._lrg_map.live_range_id(inp); |
| 0 | 363 LRG &lrg = lrgs(nidx); |
| 364 | |
| 365 // If this lrg has a high frequency use/def | |
| 673 | 366 if( lrg._maxfreq >= _phc.high_frequency_lrg() ) { |
| 0 | 367 // If the live range is also live out of this block (like it |
| 368 // would be for a fast/slow idiom), the normal spill mechanism | |
| 369 // does an excellent job. If it is not live out of this block | |
| 370 // (like it would be for debug info to uncommon trap) splitting | |
| 371 // the live range now allows a better allocation in the high | |
| 372 // frequency blocks. | |
| 373 // Build_IFG_virtual has converted the live sets to | |
| 374 // live-IN info, not live-OUT info. | |
| 375 uint k; | |
| 376 for( k=0; k < b->_num_succs; k++ ) | |
| 377 if( _phc._live->live(b->_succs[k])->member( nidx ) ) | |
| 378 break; // Live in to some successor block? | |
| 379 if( k < b->_num_succs ) | |
| 380 continue; // Live out; do not pre-split | |
| 381 // Split the lrg at this use | |
| 382 const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()]; | |
| 383 Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm ); | |
| 384 // Insert the copy in the use-def chain | |
| 385 n->set_req(inpidx, copy ); | |
| 386 // Insert the copy in the basic block, just before us | |
| 387 b->_nodes.insert( l++, copy ); | |
| 388 // Extend ("register allocate") the names array for the copy. | |
| 10111 | 389 uint max_lrg_id = _phc._lrg_map.max_lrg_id(); |
| 390 _phc.new_lrg(copy, max_lrg_id); | |
| 391 _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1); | |
| 392 _phc._cfg._bbs.map(copy->_idx, b); | |
| 0 | 393 //tty->print_cr("Split a debug use in Aggressive Coalesce"); |
| 394 } // End of if high frequency use/def | |
| 395 } // End of for all debug inputs | |
| 396 } // End of if low frequency safepoint | |
| 397 | |
| 398 } // End of if Phi | |
| 399 | |
| 400 } // End of for all instructions | |
| 401 } // End of for all blocks | |
| 402 } | |
| 403 | |
| 404 //============================================================================= | |
| 405 //------------------------------coalesce--------------------------------------- | |
| 406 // Aggressive (but pessimistic) copy coalescing of a single block | |
| 407 | |
| 408 // The following coalesce pass represents a single round of aggressive | |
| 409 // pessimistic coalesce. "Aggressive" means no attempt to preserve | |
| 410 // colorability when coalescing. This occasionally means more spills, but | |
| 411 // it also means fewer rounds of coalescing for better code - and that means | |
| 412 // faster compiles. | |
| 413 | |
| 414 // "Pessimistic" means we do not hit the fixed point in one pass (and we are | |
| 415 // reaching for the least fixed point to boot). This is typically solved | |
| 416 // with a few more rounds of coalescing, but the compiler must run fast. We | |
| 417 // could optimistically coalescing everything touching PhiNodes together | |
| 418 // into one big live range, then check for self-interference. Everywhere | |
| 419 // the live range interferes with self it would have to be split. Finding | |
| 420 // the right split points can be done with some heuristics (based on | |
| 421 // expected frequency of edges in the live range). In short, it's a real | |
| 422 // research problem and the timeline is too short to allow such research. | |
| 423 // Further thoughts: (1) build the LR in a pass, (2) find self-interference | |
| 424 // in another pass, (3) per each self-conflict, split, (4) split by finding | |
| 425 // the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted | |
| 426 // according to the GCM algorithm (or just exec freq on CFG edges). | |
| 427 | |
| 428 void PhaseAggressiveCoalesce::coalesce( Block *b ) { | |
| 429 // Copies are still "virtual" - meaning we have not made them explicitly | |
| 430 // copies. Instead, Phi functions of successor blocks have mis-matched | |
| 431 // live-ranges. If I fail to coalesce, I'll have to insert a copy to line | |
| 432 // up the live-ranges. Check for Phis in successor blocks. | |
| 433 uint i; | |
| 434 for( i=0; i<b->_num_succs; i++ ) { | |
| 435 Block *bs = b->_succs[i]; | |
| 436 // Find index of 'b' in 'bs' predecessors | |
| 437 uint j=1; | |
| 438 while( _phc._cfg._bbs[bs->pred(j)->_idx] != b ) j++; | |
| 439 // Visit all the Phis in successor block | |
| 440 for( uint k = 1; k<bs->_nodes.size(); k++ ) { | |
| 441 Node *n = bs->_nodes[k]; | |
| 442 if( !n->is_Phi() ) break; | |
| 443 combine_these_two( n, n->in(j) ); | |
| 444 } | |
| 445 } // End of for all successor blocks | |
| 446 | |
| 447 | |
| 448 // Check _this_ block for 2-address instructions and copies. | |
| 449 uint cnt = b->end_idx(); | |
| 450 for( i = 1; i<cnt; i++ ) { | |
| 451 Node *n = b->_nodes[i]; | |
| 452 uint idx; | |
| 453 // 2-address instructions have a virtual Copy matching their input | |
| 454 // to their output | |
| 10111 | 455 if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) { |
| 0 | 456 MachNode *mach = n->as_Mach(); |
| 10111 | 457 combine_these_two(mach, mach->in(idx)); |
| 0 | 458 } |
| 459 } // End of for all instructions in block | |
| 460 } | |
| 461 | |
| 462 //============================================================================= | |
| 463 //------------------------------PhaseConservativeCoalesce---------------------- | |
| 10111 | 464 PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) { |
| 465 _ulr.initialize(_phc._lrg_map.max_lrg_id()); | |
| 0 | 466 } |
| 467 | |
| 468 //------------------------------verify----------------------------------------- | |
| 469 void PhaseConservativeCoalesce::verify() { | |
| 470 #ifdef ASSERT | |
| 471 _phc.set_was_low(); | |
| 472 #endif | |
| 473 } | |
| 474 | |
| 475 //------------------------------union_helper----------------------------------- | |
| 476 void PhaseConservativeCoalesce::union_helper( Node *lr1_node, Node *lr2_node, uint lr1, uint lr2, Node *src_def, Node *dst_copy, Node *src_copy, Block *b, uint bindex ) { | |
| 477 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the | |
| 478 // union-find tree | |
| 479 _phc.Union( lr1_node, lr2_node ); | |
| 480 | |
| 481 // Single-def live range ONLY if both live ranges are single-def. | |
| 482 // If both are single def, then src_def powers one live range | |
| 483 // and def_copy powers the other. After merging, src_def powers | |
| 484 // the combined live range. | |
|
295
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
485 lrgs(lr1)._def = (lrgs(lr1).is_multidef() || |
|
ea18057223c4
6732194: Data corruption dependent on -server/-client/-Xbatch
never
parents:
0
diff
changeset
|
486 lrgs(lr2).is_multidef() ) |
| 0 | 487 ? NodeSentinel : src_def; |
| 488 lrgs(lr2)._def = NULL; // No def for lrg 2 | |
| 489 lrgs(lr2).Clear(); // Force empty mask for LRG 2 | |
| 490 //lrgs(lr2)._size = 0; // Live-range 2 goes dead | |
| 491 lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop; | |
| 492 lrgs(lr2)._is_oop = 0; // In particular, not an oop for GC info | |
| 493 | |
| 494 if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq) | |
| 495 lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq; | |
| 496 | |
| 497 // Copy original value instead. Intermediate copies go dead, and | |
| 498 // the dst_copy becomes useless. | |
| 499 int didx = dst_copy->is_Copy(); | |
| 500 dst_copy->set_req( didx, src_def ); | |
| 501 // Add copy to free list | |
| 502 // _phc.free_spillcopy(b->_nodes[bindex]); | |
| 503 assert( b->_nodes[bindex] == dst_copy, "" ); | |
| 504 dst_copy->replace_by( dst_copy->in(didx) ); | |
| 505 dst_copy->set_req( didx, NULL); | |
| 506 b->_nodes.remove(bindex); | |
| 507 if( bindex < b->_ihrp_index ) b->_ihrp_index--; | |
| 508 if( bindex < b->_fhrp_index ) b->_fhrp_index--; | |
| 509 | |
| 510 // Stretched lr1; add it to liveness of intermediate blocks | |
| 511 Block *b2 = _phc._cfg._bbs[src_copy->_idx]; | |
| 512 while( b != b2 ) { | |
| 513 b = _phc._cfg._bbs[b->pred(1)->_idx]; | |
| 514 _phc._live->live(b)->insert(lr1); | |
| 515 } | |
| 516 } | |
| 517 | |
| 518 //------------------------------compute_separating_interferences--------------- | |
| 519 // Factored code from copy_copy that computes extra interferences from | |
| 520 // lengthening a live range by double-coalescing. | |
| 521 uint PhaseConservativeCoalesce::compute_separating_interferences(Node *dst_copy, Node *src_copy, Block *b, uint bindex, RegMask &rm, uint reg_degree, uint rm_size, uint lr1, uint lr2 ) { | |
| 522 | |
| 523 assert(!lrgs(lr1)._fat_proj, "cannot coalesce fat_proj"); | |
| 524 assert(!lrgs(lr2)._fat_proj, "cannot coalesce fat_proj"); | |
| 525 Node *prev_copy = dst_copy->in(dst_copy->is_Copy()); | |
| 526 Block *b2 = b; | |
| 527 uint bindex2 = bindex; | |
| 528 while( 1 ) { | |
| 529 // Find previous instruction | |
| 530 bindex2--; // Chain backwards 1 instruction | |
| 531 while( bindex2 == 0 ) { // At block start, find prior block | |
| 532 assert( b2->num_preds() == 2, "cannot double coalesce across c-flow" ); | |
| 533 b2 = _phc._cfg._bbs[b2->pred(1)->_idx]; | |
| 534 bindex2 = b2->end_idx()-1; | |
| 535 } | |
| 536 // Get prior instruction | |
| 537 assert(bindex2 < b2->_nodes.size(), "index out of bounds"); | |
| 538 Node *x = b2->_nodes[bindex2]; | |
| 539 if( x == prev_copy ) { // Previous copy in copy chain? | |
| 540 if( prev_copy == src_copy)// Found end of chain and all interferences | |
| 541 break; // So break out of loop | |
| 542 // Else work back one in copy chain | |
| 543 prev_copy = prev_copy->in(prev_copy->is_Copy()); | |
| 544 } else { // Else collect interferences | |
| 10111 | 545 uint lidx = _phc._lrg_map.find(x); |
| 0 | 546 // Found another def of live-range being stretched? |
| 10111 | 547 if(lidx == lr1) { |
| 548 return max_juint; | |
| 549 } | |
| 550 if(lidx == lr2) { | |
| 551 return max_juint; | |
| 552 } | |
| 0 | 553 |
| 554 // If we attempt to coalesce across a bound def | |
| 555 if( lrgs(lidx).is_bound() ) { | |
| 556 // Do not let the coalesced LRG expect to get the bound color | |
| 557 rm.SUBTRACT( lrgs(lidx).mask() ); | |
| 558 // Recompute rm_size | |
| 559 rm_size = rm.Size(); | |
| 560 //if( rm._flags ) rm_size += 1000000; | |
| 561 if( reg_degree >= rm_size ) return max_juint; | |
| 562 } | |
| 563 if( rm.overlap(lrgs(lidx).mask()) ) { | |
| 564 // Insert lidx into union LRG; returns TRUE if actually inserted | |
| 565 if( _ulr.insert(lidx) ) { | |
| 566 // Infinite-stack neighbors do not alter colorability, as they | |
| 567 // can always color to some other color. | |
| 568 if( !lrgs(lidx).mask().is_AllStack() ) { | |
| 569 // If this coalesce will make any new neighbor uncolorable, | |
| 570 // do not coalesce. | |
| 571 if( lrgs(lidx).just_lo_degree() ) | |
| 572 return max_juint; | |
| 573 // Bump our degree | |
| 574 if( ++reg_degree >= rm_size ) | |
| 575 return max_juint; | |
| 576 } // End of if not infinite-stack neighbor | |
| 577 } // End of if actually inserted | |
| 578 } // End of if live range overlaps | |
| 605 | 579 } // End of else collect interferences for 1 node |
| 580 } // End of while forever, scan back for interferences | |
| 0 | 581 return reg_degree; |
| 582 } | |
| 583 | |
| 584 //------------------------------update_ifg------------------------------------- | |
| 585 void PhaseConservativeCoalesce::update_ifg(uint lr1, uint lr2, IndexSet *n_lr1, IndexSet *n_lr2) { | |
| 586 // Some original neighbors of lr1 might have gone away | |
| 587 // because the constrained register mask prevented them. | |
| 588 // Remove lr1 from such neighbors. | |
| 589 IndexSetIterator one(n_lr1); | |
| 590 uint neighbor; | |
| 591 LRG &lrg1 = lrgs(lr1); | |
| 592 while ((neighbor = one.next()) != 0) | |
| 593 if( !_ulr.member(neighbor) ) | |
| 594 if( _phc._ifg->neighbors(neighbor)->remove(lr1) ) | |
| 595 lrgs(neighbor).inc_degree( -lrg1.compute_degree(lrgs(neighbor)) ); | |
| 596 | |
| 597 | |
| 598 // lr2 is now called (coalesced into) lr1. | |
| 599 // Remove lr2 from the IFG. | |
| 600 IndexSetIterator two(n_lr2); | |
| 601 LRG &lrg2 = lrgs(lr2); | |
| 602 while ((neighbor = two.next()) != 0) | |
| 603 if( _phc._ifg->neighbors(neighbor)->remove(lr2) ) | |
| 604 lrgs(neighbor).inc_degree( -lrg2.compute_degree(lrgs(neighbor)) ); | |
| 605 | |
| 606 // Some neighbors of intermediate copies now interfere with the | |
| 607 // combined live range. | |
| 608 IndexSetIterator three(&_ulr); | |
| 609 while ((neighbor = three.next()) != 0) | |
| 610 if( _phc._ifg->neighbors(neighbor)->insert(lr1) ) | |
| 611 lrgs(neighbor).inc_degree( lrg1.compute_degree(lrgs(neighbor)) ); | |
| 612 } | |
| 613 | |
| 614 //------------------------------record_bias------------------------------------ | |
| 615 static void record_bias( const PhaseIFG *ifg, int lr1, int lr2 ) { | |
| 616 // Tag copy bias here | |
| 617 if( !ifg->lrgs(lr1)._copy_bias ) | |
| 618 ifg->lrgs(lr1)._copy_bias = lr2; | |
| 619 if( !ifg->lrgs(lr2)._copy_bias ) | |
| 620 ifg->lrgs(lr2)._copy_bias = lr1; | |
| 621 } | |
| 622 | |
| 623 //------------------------------copy_copy-------------------------------------- | |
| 624 // See if I can coalesce a series of multiple copies together. I need the | |
| 625 // final dest copy and the original src copy. They can be the same Node. | |
| 626 // Compute the compatible register masks. | |
| 10111 | 627 bool PhaseConservativeCoalesce::copy_copy(Node *dst_copy, Node *src_copy, Block *b, uint bindex) { |
| 0 | 628 |
| 10111 | 629 if (!dst_copy->is_SpillCopy()) { |
| 630 return false; | |
| 631 } | |
| 632 if (!src_copy->is_SpillCopy()) { | |
| 633 return false; | |
| 634 } | |
| 0 | 635 Node *src_def = src_copy->in(src_copy->is_Copy()); |
| 10111 | 636 uint lr1 = _phc._lrg_map.find(dst_copy); |
| 637 uint lr2 = _phc._lrg_map.find(src_def); | |
| 0 | 638 |
| 639 // Same live ranges already? | |
| 10111 | 640 if (lr1 == lr2) { |
| 641 return false; | |
| 642 } | |
| 0 | 643 |
| 644 // Interfere? | |
| 10111 | 645 if (_phc._ifg->test_edge_sq(lr1, lr2)) { |
| 646 return false; | |
| 647 } | |
| 0 | 648 |
| 649 // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK. | |
| 10111 | 650 if (!lrgs(lr1)._is_oop && lrgs(lr2)._is_oop) { // not an oop->int cast |
| 0 | 651 return false; |
| 10111 | 652 } |
| 0 | 653 |
| 654 // Coalescing between an aligned live range and a mis-aligned live range? | |
| 655 // No, no! Alignment changes how we count degree. | |
| 10111 | 656 if (lrgs(lr1)._fat_proj != lrgs(lr2)._fat_proj) { |
| 0 | 657 return false; |
| 10111 | 658 } |
| 0 | 659 |
| 660 // Sort; use smaller live-range number | |
| 661 Node *lr1_node = dst_copy; | |
| 662 Node *lr2_node = src_def; | |
| 10111 | 663 if (lr1 > lr2) { |
| 0 | 664 uint tmp = lr1; lr1 = lr2; lr2 = tmp; |
| 665 lr1_node = src_def; lr2_node = dst_copy; | |
| 666 } | |
| 667 | |
| 668 // Check for compatibility of the 2 live ranges by | |
| 669 // intersecting their allowed register sets. | |
| 670 RegMask rm = lrgs(lr1).mask(); | |
| 671 rm.AND(lrgs(lr2).mask()); | |
| 672 // Number of bits free | |
| 673 uint rm_size = rm.Size(); | |
| 674 | |
|
1730
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
675 if (UseFPUForSpilling && rm.is_AllStack() ) { |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
676 // Don't coalesce when frequency difference is large |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
677 Block *dst_b = _phc._cfg._bbs[dst_copy->_idx]; |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
678 Block *src_def_b = _phc._cfg._bbs[src_def->_idx]; |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
679 if (src_def_b->_freq > 10*dst_b->_freq ) |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
680 return false; |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
681 } |
|
f55c4f82ab9d
6978249: spill between cpu and fpu registers when those moves are fast
never
parents:
1552
diff
changeset
|
682 |
| 0 | 683 // If we can use any stack slot, then effective size is infinite |
| 684 if( rm.is_AllStack() ) rm_size += 1000000; | |
| 685 // Incompatible masks, no way to coalesce | |
| 686 if( rm_size == 0 ) return false; | |
| 687 | |
| 688 // Another early bail-out test is when we are double-coalescing and the | |
| 605 | 689 // 2 copies are separated by some control flow. |
| 0 | 690 if( dst_copy != src_copy ) { |
| 691 Block *src_b = _phc._cfg._bbs[src_copy->_idx]; | |
| 692 Block *b2 = b; | |
| 693 while( b2 != src_b ) { | |
| 694 if( b2->num_preds() > 2 ){// Found merge-point | |
| 695 _phc._lost_opp_cflow_coalesce++; | |
| 696 // extra record_bias commented out because Chris believes it is not | |
| 697 // productive. Since we can record only 1 bias, we want to choose one | |
| 698 // that stands a chance of working and this one probably does not. | |
| 699 //record_bias( _phc._lrgs, lr1, lr2 ); | |
| 700 return false; // To hard to find all interferences | |
| 701 } | |
| 702 b2 = _phc._cfg._bbs[b2->pred(1)->_idx]; | |
| 703 } | |
| 704 } | |
| 705 | |
| 706 // Union the two interference sets together into '_ulr' | |
| 707 uint reg_degree = _ulr.lrg_union( lr1, lr2, rm_size, _phc._ifg, rm ); | |
| 708 | |
| 709 if( reg_degree >= rm_size ) { | |
| 710 record_bias( _phc._ifg, lr1, lr2 ); | |
| 711 return false; | |
| 712 } | |
| 713 | |
| 714 // Now I need to compute all the interferences between dst_copy and | |
| 715 // src_copy. I'm not willing visit the entire interference graph, so | |
| 716 // I limit my search to things in dst_copy's block or in a straight | |
| 717 // line of previous blocks. I give up at merge points or when I get | |
| 718 // more interferences than my degree. I can stop when I find src_copy. | |
| 719 if( dst_copy != src_copy ) { | |
| 720 reg_degree = compute_separating_interferences(dst_copy, src_copy, b, bindex, rm, rm_size, reg_degree, lr1, lr2 ); | |
| 721 if( reg_degree == max_juint ) { | |
| 722 record_bias( _phc._ifg, lr1, lr2 ); | |
| 723 return false; | |
| 724 } | |
| 725 } // End of if dst_copy & src_copy are different | |
| 726 | |
| 727 | |
| 728 // ---- THE COMBINED LRG IS COLORABLE ---- | |
| 729 | |
| 730 // YEAH - Now coalesce this copy away | |
| 731 assert( lrgs(lr1).num_regs() == lrgs(lr2).num_regs(), "" ); | |
| 732 | |
| 733 IndexSet *n_lr1 = _phc._ifg->neighbors(lr1); | |
| 734 IndexSet *n_lr2 = _phc._ifg->neighbors(lr2); | |
| 735 | |
| 736 // Update the interference graph | |
| 737 update_ifg(lr1, lr2, n_lr1, n_lr2); | |
| 738 | |
| 739 _ulr.remove(lr1); | |
| 740 | |
| 741 // Uncomment the following code to trace Coalescing in great detail. | |
| 742 // | |
| 743 //if (false) { | |
| 744 // tty->cr(); | |
| 745 // tty->print_cr("#######################################"); | |
| 746 // tty->print_cr("union %d and %d", lr1, lr2); | |
| 747 // n_lr1->dump(); | |
| 748 // n_lr2->dump(); | |
| 749 // tty->print_cr("resulting set is"); | |
| 750 // _ulr.dump(); | |
| 751 //} | |
| 752 | |
| 753 // Replace n_lr1 with the new combined live range. _ulr will use | |
| 754 // n_lr1's old memory on the next iteration. n_lr2 is cleared to | |
| 755 // send its internal memory to the free list. | |
| 756 _ulr.swap(n_lr1); | |
| 757 _ulr.clear(); | |
| 758 n_lr2->clear(); | |
| 759 | |
| 760 lrgs(lr1).set_degree( _phc._ifg->effective_degree(lr1) ); | |
| 761 lrgs(lr2).set_degree( 0 ); | |
| 762 | |
| 763 // Join live ranges. Merge larger into smaller. Union lr2 into lr1 in the | |
| 764 // union-find tree | |
| 765 union_helper( lr1_node, lr2_node, lr1, lr2, src_def, dst_copy, src_copy, b, bindex ); | |
| 766 // Combine register restrictions | |
| 767 lrgs(lr1).set_mask(rm); | |
| 768 lrgs(lr1).compute_set_mask_size(); | |
| 769 lrgs(lr1)._cost += lrgs(lr2)._cost; | |
| 770 lrgs(lr1)._area += lrgs(lr2)._area; | |
| 771 | |
| 772 // While its uncommon to successfully coalesce live ranges that started out | |
| 773 // being not-lo-degree, it can happen. In any case the combined coalesced | |
| 774 // live range better Simplify nicely. | |
| 775 lrgs(lr1)._was_lo = 1; | |
| 776 | |
| 777 // kinda expensive to do all the time | |
| 778 //tty->print_cr("warning: slow verify happening"); | |
| 779 //_phc._ifg->verify( &_phc ); | |
| 780 return true; | |
| 781 } | |
| 782 | |
| 783 //------------------------------coalesce--------------------------------------- | |
| 784 // Conservative (but pessimistic) copy coalescing of a single block | |
| 785 void PhaseConservativeCoalesce::coalesce( Block *b ) { | |
| 786 // Bail out on infrequent blocks | |
| 787 if( b->is_uncommon(_phc._cfg._bbs) ) | |
| 788 return; | |
| 789 // Check this block for copies. | |
| 790 for( uint i = 1; i<b->end_idx(); i++ ) { | |
| 791 // Check for actual copies on inputs. Coalesce a copy into its | |
| 792 // input if use and copy's input are compatible. | |
| 793 Node *copy1 = b->_nodes[i]; | |
| 794 uint idx1 = copy1->is_Copy(); | |
| 795 if( !idx1 ) continue; // Not a copy | |
| 796 | |
| 797 if( copy_copy(copy1,copy1,b,i) ) { | |
| 798 i--; // Retry, same location in block | |
| 799 PhaseChaitin::_conserv_coalesce++; // Collect stats on success | |
| 800 continue; | |
| 801 } | |
| 802 } | |
| 803 } |