1 /* 2 * Copyright (c) 1998, 2009, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_postaloc.cpp.incl" 27 28 // see if this register kind does not requires two registers 29 static bool is_single_register(uint x) { 30 #ifdef _LP64 31 return (x != Op_RegD && x != Op_RegL && x != Op_RegP); 32 #else 33 return (x != Op_RegD && x != Op_RegL); 34 #endif 35 } 36 37 //---------------------------may_be_copy_of_callee----------------------------- 38 // Check to see if we can possibly be a copy of a callee-save value. 39 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { 40 // Short circuit if there are no callee save registers 41 if (_matcher.number_of_saved_registers() == 0) return false; 42 43 // Expect only a spill-down and reload on exit for callee-save spills. 44 // Chains of copies cannot be deep. 45 // 5008997 - This is wishful thinking. Register allocator seems to 46 // be splitting live ranges for callee save registers to such 47 // an extent that in large methods the chains can be very long 48 // (50+). The conservative answer is to return true if we don't 49 // know as this prevents optimizations from occurring. 50 51 const int limit = 60; 52 int i; 53 for( i=0; i < limit; i++ ) { 54 if( def->is_Proj() && def->in(0)->is_Start() && 55 _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) 56 return true; // Direct use of callee-save proj 57 if( def->is_Copy() ) // Copies carry value through 58 def = def->in(def->is_Copy()); 59 else if( def->is_Phi() ) // Phis can merge it from any direction 60 def = def->in(1); 61 else 62 break; 63 guarantee(def != NULL, "must not resurrect dead copy"); 64 } 65 // If we reached the end and didn't find a callee save proj 66 // then this may be a callee save proj so we return true 67 // as the conservative answer. If we didn't reach then end 68 // we must have discovered that it was not a callee save 69 // else we would have returned. 70 return i == limit; 71 } 72 73 74 75 //------------------------------yank_if_dead----------------------------------- 76 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to 77 // iterators in the current block. 78 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { 79 int blk_adjust=0; 80 while (old->outcnt() == 0 && old != C->top()) { 81 Block *oldb = _cfg._bbs[old->_idx]; 82 oldb->find_remove(old); 83 // Count 1 if deleting an instruction from the current block 84 if( oldb == current_block ) blk_adjust++; 85 _cfg._bbs.map(old->_idx,NULL); 86 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); 87 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? 88 value->map(old_reg,NULL); // Yank from value/regnd maps 89 regnd->map(old_reg,NULL); // This register's value is now unknown 90 } 91 assert(old->req() <= 2, "can't handle more inputs"); 92 Node *tmp = old->req() > 1 ? old->in(1) : NULL; 93 old->disconnect_inputs(NULL); 94 if( !tmp ) break; 95 old = tmp; 96 } 97 return blk_adjust; 98 } 99 100 //------------------------------use_prior_register----------------------------- 101 // Use the prior value instead of the current value, in an effort to make 102 // the current value go dead. Return block iterator adjustment, in case 103 // we yank some instructions from this block. 104 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { 105 // No effect? 106 if( def == n->in(idx) ) return 0; 107 // Def is currently dead and can be removed? Do not resurrect 108 if( def->outcnt() == 0 ) return 0; 109 110 // Not every pair of physical registers are assignment compatible, 111 // e.g. on sparc floating point registers are not assignable to integer 112 // registers. 113 const LRG &def_lrg = lrgs(n2lidx(def)); 114 OptoReg::Name def_reg = def_lrg.reg(); 115 const RegMask &use_mask = n->in_RegMask(idx); 116 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) 117 : (use_mask.is_AllStack() != 0)); 118 // Check for a copy to or from a misaligned pair. 119 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); 120 121 if (!can_use) 122 return 0; 123 124 // Capture the old def in case it goes dead... 125 Node *old = n->in(idx); 126 127 // Save-on-call copies can only be elided if the entire copy chain can go 128 // away, lest we get the same callee-save value alive in 2 locations at 129 // once. We check for the obvious trivial case here. Although it can 130 // sometimes be elided with cooperation outside our scope, here we will just 131 // miss the opportunity. :-( 132 if( may_be_copy_of_callee(def) ) { 133 if( old->outcnt() > 1 ) return 0; // We're the not last user 134 int idx = old->is_Copy(); 135 assert( idx, "chain of copies being removed" ); 136 Node *old2 = old->in(idx); // Chain of copies 137 if( old2->outcnt() > 1 ) return 0; // old is not the last user 138 int idx2 = old2->is_Copy(); 139 if( !idx2 ) return 0; // Not a chain of 2 copies 140 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies 141 } 142 143 // Use the new def 144 n->set_req(idx,def); 145 _post_alloc++; 146 147 // Is old def now dead? We successfully yanked a copy? 148 return yank_if_dead(old,current_block,&value,®nd); 149 } 150 151 152 //------------------------------skip_copies------------------------------------ 153 // Skip through any number of copies (that don't mod oop-i-ness) 154 Node *PhaseChaitin::skip_copies( Node *c ) { 155 int idx = c->is_Copy(); 156 uint is_oop = lrgs(n2lidx(c))._is_oop; 157 while (idx != 0) { 158 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); 159 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) 160 break; // casting copy, not the same value 161 c = c->in(idx); 162 idx = c->is_Copy(); 163 } 164 return c; 165 } 166 167 //------------------------------elide_copy------------------------------------- 168 // Remove (bypass) copies along Node n, edge k. 169 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { 170 int blk_adjust = 0; 171 172 uint nk_idx = n2lidx(n->in(k)); 173 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); 174 175 // Remove obvious same-register copies 176 Node *x = n->in(k); 177 int idx; 178 while( (idx=x->is_Copy()) != 0 ) { 179 Node *copy = x->in(idx); 180 guarantee(copy != NULL, "must not resurrect dead copy"); 181 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; 182 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); 183 if( n->in(k) != copy ) break; // Failed for some cutout? 184 x = copy; // Progress, try again 185 } 186 187 // Phis and 2-address instructions cannot change registers so easily - their 188 // outputs must match their input. 189 if( !can_change_regs ) 190 return blk_adjust; // Only check stupid copies! 191 192 // Loop backedges won't have a value-mapping yet 193 if( &value == NULL ) return blk_adjust; 194 195 // Skip through all copies to the _value_ being used. Do not change from 196 // int to pointer. This attempts to jump through a chain of copies, where 197 // intermediate copies might be illegal, i.e., value is stored down to stack 198 // then reloaded BUT survives in a register the whole way. 199 Node *val = skip_copies(n->in(k)); 200 201 if( val == x ) return blk_adjust; // No progress? 202 203 bool single = is_single_register(val->ideal_reg()); 204 uint val_idx = n2lidx(val); 205 OptoReg::Name val_reg = lrgs(val_idx).reg(); 206 207 // See if it happens to already be in the correct register! 208 // (either Phi's direct register, or the common case of the name 209 // never-clobbered original-def register) 210 if( value[val_reg] == val && 211 // Doubles check both halves 212 ( single || value[val_reg-1] == val ) ) { 213 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); 214 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying 215 return blk_adjust; 216 } 217 218 // See if we can skip the copy by changing registers. Don't change from 219 // using a register to using the stack unless we know we can remove a 220 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill 221 // ops reading from memory instead of just loading once and using the 222 // register. 223 224 // Also handle duplicate copies here. 225 const Type *t = val->is_Con() ? val->bottom_type() : NULL; 226 227 // Scan all registers to see if this value is around already 228 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { 229 if (reg == (uint)nk_reg) { 230 // Found ourselves so check if there is only one user of this 231 // copy and keep on searching for a better copy if so. 232 bool ignore_self = true; 233 x = n->in(k); 234 DUIterator_Fast imax, i = x->fast_outs(imax); 235 Node* first = x->fast_out(i); i++; 236 while (i < imax && ignore_self) { 237 Node* use = x->fast_out(i); i++; 238 if (use != first) ignore_self = false; 239 } 240 if (ignore_self) continue; 241 } 242 243 Node *vv = value[reg]; 244 if( !single ) { // Doubles check for aligned-adjacent pair 245 if( (reg&1)==0 ) continue; // Wrong half of a pair 246 if( vv != value[reg-1] ) continue; // Not a complete pair 247 } 248 if( vv == val || // Got a direct hit? 249 (t && vv && vv->bottom_type() == t && vv->is_Mach() && 250 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? 251 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); 252 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR 253 OptoReg::is_reg(reg) || // turning into a register use OR 254 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use 255 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); 256 if( n->in(k) == regnd[reg] ) // Success! Quit trying 257 return blk_adjust; 258 } // End of if not degrading to a stack 259 } // End of if found value in another register 260 } // End of scan all machine registers 261 return blk_adjust; 262 } 263 264 265 // 266 // Check if nreg already contains the constant value val. Normal copy 267 // elimination doesn't doesn't work on constants because multiple 268 // nodes can represent the same constant so the type and rule of the 269 // MachNode must be checked to ensure equivalence. 270 // 271 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, 272 Block *current_block, 273 Node_List& value, Node_List& regnd, 274 OptoReg::Name nreg, OptoReg::Name nreg2) { 275 if (value[nreg] != val && val->is_Con() && 276 value[nreg] != NULL && value[nreg]->is_Con() && 277 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && 278 value[nreg]->bottom_type() == val->bottom_type() && 279 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { 280 // This code assumes that two MachNodes representing constants 281 // which have the same rule and the same bottom type will produce 282 // identical effects into a register. This seems like it must be 283 // objectively true unless there are hidden inputs to the nodes 284 // but if that were to change this code would need to updated. 285 // Since they are equivalent the second one if redundant and can 286 // be removed. 287 // 288 // n will be replaced with the old value but n might have 289 // kills projections associated with it so remove them now so that 290 // yank_if_dead will be able to eliminate the copy once the uses 291 // have been transferred to the old[value]. 292 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 293 Node* use = n->fast_out(i); 294 if (use->is_Proj() && use->outcnt() == 0) { 295 // Kill projections have no users and one input 296 use->set_req(0, C->top()); 297 yank_if_dead(use, current_block, &value, ®nd); 298 --i; --imax; 299 } 300 } 301 _post_alloc++; 302 return true; 303 } 304 return false; 305 } 306 307 308 //------------------------------post_allocate_copy_removal--------------------- 309 // Post-Allocation peephole copy removal. We do this in 1 pass over the 310 // basic blocks. We maintain a mapping of registers to Nodes (an array of 311 // Nodes indexed by machine register or stack slot number). NULL means that a 312 // register is not mapped to any Node. We can (want to have!) have several 313 // registers map to the same Node. We walk forward over the instructions 314 // updating the mapping as we go. At merge points we force a NULL if we have 315 // to merge 2 different Nodes into the same register. Phi functions will give 316 // us a new Node if there is a proper value merging. Since the blocks are 317 // arranged in some RPO, we will visit all parent blocks before visiting any 318 // successor blocks (except at loops). 319 // 320 // If we find a Copy we look to see if the Copy's source register is a stack 321 // slot and that value has already been loaded into some machine register; if 322 // so we use machine register directly. This turns a Load into a reg-reg 323 // Move. We also look for reloads of identical constants. 324 // 325 // When we see a use from a reg-reg Copy, we will attempt to use the copy's 326 // source directly and make the copy go dead. 327 void PhaseChaitin::post_allocate_copy_removal() { 328 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) 329 ResourceMark rm; 330 331 // Need a mapping from basic block Node_Lists. We need a Node_List to 332 // map from register number to value-producing Node. 333 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 334 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 335 // Need a mapping from basic block Node_Lists. We need a Node_List to 336 // map from register number to register-defining Node. 337 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 338 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 339 340 // We keep unused Node_Lists on a free_list to avoid wasting 341 // memory. 342 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); 343 344 // For all blocks 345 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 346 uint j; 347 Block *b = _cfg._blocks[i]; 348 349 // Count of Phis in block 350 uint phi_dex; 351 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { 352 Node *phi = b->_nodes[phi_dex]; 353 if( !phi->is_Phi() ) 354 break; 355 } 356 357 // If any predecessor has not been visited, we do not know the state 358 // of registers at the start. Check for this, while updating copies 359 // along Phi input edges 360 bool missing_some_inputs = false; 361 Block *freed = NULL; 362 for( j = 1; j < b->num_preds(); j++ ) { 363 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 364 // Remove copies along phi edges 365 for( uint k=1; k<phi_dex; k++ ) 366 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); 367 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? 368 // See if this predecessor's mappings have been used by everybody 369 // who wants them. If so, free 'em. 370 uint k; 371 for( k=0; k<pb->_num_succs; k++ ) { 372 Block *pbsucc = pb->_succs[k]; 373 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) 374 break; // Found a future user 375 } 376 if( k >= pb->_num_succs ) { // No more uses, free! 377 freed = pb; // Record last block freed 378 free_list.push(blk2value[pb->_pre_order]); 379 free_list.push(blk2regnd[pb->_pre_order]); 380 } 381 } else { // This block has unvisited (loopback) inputs 382 missing_some_inputs = true; 383 } 384 } 385 386 387 // Extract Node_List mappings. If 'freed' is non-zero, we just popped 388 // 'freed's blocks off the list 389 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 390 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 391 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); 392 value.map(_max_reg,NULL); 393 regnd.map(_max_reg,NULL); 394 // Set mappings as OUR mappings 395 blk2value[b->_pre_order] = &value; 396 blk2regnd[b->_pre_order] = ®nd; 397 398 // Initialize value & regnd for this block 399 if( missing_some_inputs ) { 400 // Some predecessor has not yet been visited; zap map to empty 401 for( uint k = 0; k < (uint)_max_reg; k++ ) { 402 value.map(k,NULL); 403 regnd.map(k,NULL); 404 } 405 } else { 406 if( !freed ) { // Didn't get a freebie prior block 407 // Must clone some data 408 freed = _cfg._bbs[b->pred(1)->_idx]; 409 Node_List &f_value = *blk2value[freed->_pre_order]; 410 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; 411 for( uint k = 0; k < (uint)_max_reg; k++ ) { 412 value.map(k,f_value[k]); 413 regnd.map(k,f_regnd[k]); 414 } 415 } 416 // Merge all inputs together, setting to NULL any conflicts. 417 for( j = 1; j < b->num_preds(); j++ ) { 418 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 419 if( pb == freed ) continue; // Did self already via freelist 420 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; 421 for( uint k = 0; k < (uint)_max_reg; k++ ) { 422 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? 423 value.map(k,NULL); // Then no value handy 424 regnd.map(k,NULL); 425 } 426 } 427 } 428 } 429 430 // For all Phi's 431 for( j = 1; j < phi_dex; j++ ) { 432 uint k; 433 Node *phi = b->_nodes[j]; 434 uint pidx = n2lidx(phi); 435 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); 436 437 // Remove copies remaining on edges. Check for junk phi. 438 Node *u = NULL; 439 for( k=1; k<phi->req(); k++ ) { 440 Node *x = phi->in(k); 441 if( phi != x && u != x ) // Found a different input 442 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input 443 } 444 if( u != NodeSentinel ) { // Junk Phi. Remove 445 b->_nodes.remove(j--); phi_dex--; 446 _cfg._bbs.map(phi->_idx,NULL); 447 phi->replace_by(u); 448 phi->disconnect_inputs(NULL); 449 continue; 450 } 451 // Note that if value[pidx] exists, then we merged no new values here 452 // and the phi is useless. This can happen even with the above phi 453 // removal for complex flows. I cannot keep the better known value here 454 // because locally the phi appears to define a new merged value. If I 455 // keep the better value then a copy of the phi, being unable to use the 456 // global flow analysis, can't "peek through" the phi to the original 457 // reaching value and so will act like it's defining a new value. This 458 // can lead to situations where some uses are from the old and some from 459 // the new values. Not illegal by itself but throws the over-strong 460 // assert in scheduling. 461 if( pidx ) { 462 value.map(preg,phi); 463 regnd.map(preg,phi); 464 OptoReg::Name preg_lo = OptoReg::add(preg,-1); 465 if( !is_single_register(phi->ideal_reg()) ) { 466 value.map(preg_lo,phi); 467 regnd.map(preg_lo,phi); 468 } 469 } 470 } 471 472 // For all remaining instructions 473 for( j = phi_dex; j < b->_nodes.size(); j++ ) { 474 Node *n = b->_nodes[j]; 475 476 if( n->outcnt() == 0 && // Dead? 477 n != C->top() && // (ignore TOP, it has no du info) 478 !n->is_Proj() ) { // fat-proj kills 479 j -= yank_if_dead(n,b,&value,®nd); 480 continue; 481 } 482 483 // Improve reaching-def info. Occasionally post-alloc's liveness gives 484 // up (at loop backedges, because we aren't doing a full flow pass). 485 // The presence of a live use essentially asserts that the use's def is 486 // alive and well at the use (or else the allocator fubar'd). Take 487 // advantage of this info to set a reaching def for the use-reg. 488 uint k; 489 for( k = 1; k < n->req(); k++ ) { 490 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE 491 guarantee(def != NULL, "no disconnected nodes at this point"); 492 uint useidx = n2lidx(def); // useidx is the live range index for this USE 493 494 if( useidx ) { 495 OptoReg::Name ureg = lrgs(useidx).reg(); 496 if( !value[ureg] ) { 497 int idx; // Skip occasional useless copy 498 while( (idx=def->is_Copy()) != 0 && 499 def->in(idx) != NULL && // NULL should not happen 500 ureg == lrgs(n2lidx(def->in(idx))).reg() ) 501 def = def->in(idx); 502 Node *valdef = skip_copies(def); // tighten up val through non-useless copies 503 value.map(ureg,valdef); // record improved reaching-def info 504 regnd.map(ureg, def); 505 // Record other half of doubles 506 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); 507 if( !is_single_register(def->ideal_reg()) && 508 ( !RegMask::can_represent(ureg_lo) || 509 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent 510 !value[ureg_lo] ) { 511 value.map(ureg_lo,valdef); // record improved reaching-def info 512 regnd.map(ureg_lo, def); 513 } 514 } 515 } 516 } 517 518 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; 519 520 // Remove copies along input edges 521 for( k = 1; k < n->req(); k++ ) 522 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); 523 524 // Unallocated Nodes define no registers 525 uint lidx = n2lidx(n); 526 if( !lidx ) continue; 527 528 // Update the register defined by this instruction 529 OptoReg::Name nreg = lrgs(lidx).reg(); 530 // Skip through all copies to the _value_ being defined. 531 // Do not change from int to pointer 532 Node *val = skip_copies(n); 533 534 // Clear out a dead definition before starting so that the 535 // elimination code doesn't have to guard against it. The 536 // definition could in fact be a kill projection with a count of 537 // 0 which is safe but since those are uninteresting for copy 538 // elimination just delete them as well. 539 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { 540 regnd.map(nreg, NULL); 541 value.map(nreg, NULL); 542 } 543 544 uint n_ideal_reg = n->ideal_reg(); 545 if( is_single_register(n_ideal_reg) ) { 546 // If Node 'n' does not change the value mapped by the register, 547 // then 'n' is a useless copy. Do not update the register->node 548 // mapping so 'n' will go dead. 549 if( value[nreg] != val ) { 550 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { 551 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 552 } else { 553 // Update the mapping: record new Node defined by the register 554 regnd.map(nreg,n); 555 // Update mapping for defined *value*, which is the defined 556 // Node after skipping all copies. 557 value.map(nreg,val); 558 } 559 } else if( !may_be_copy_of_callee(n) ) { 560 assert( n->is_Copy(), "" ); 561 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 562 } 563 } else { 564 // If the value occupies a register pair, record same info 565 // in both registers. 566 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); 567 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or 568 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent 569 // Sparc occasionally has non-adjacent pairs. 570 // Find the actual other value 571 RegMask tmp = lrgs(lidx).mask(); 572 tmp.Remove(nreg); 573 nreg_lo = tmp.find_first_elem(); 574 } 575 if( value[nreg] != val || value[nreg_lo] != val ) { 576 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { 577 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 578 } else { 579 regnd.map(nreg , n ); 580 regnd.map(nreg_lo, n ); 581 value.map(nreg ,val); 582 value.map(nreg_lo,val); 583 } 584 } else if( !may_be_copy_of_callee(n) ) { 585 assert( n->is_Copy(), "" ); 586 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 587 } 588 } 589 590 // Fat projections kill many registers 591 if( n_ideal_reg == MachProjNode::fat_proj ) { 592 RegMask rm = n->out_RegMask(); 593 // wow, what an expensive iterator... 594 nreg = rm.find_first_elem(); 595 while( OptoReg::is_valid(nreg)) { 596 rm.Remove(nreg); 597 value.map(nreg,n); 598 regnd.map(nreg,n); 599 nreg = rm.find_first_elem(); 600 } 601 } 602 603 } // End of for all instructions in the block 604 605 } // End for all blocks 606 }