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 && nk_idx != 0 && 202 regnd[nk_reg] != NULL && regnd[nk_reg] != x && 203 n2lidx(x) == n2lidx(regnd[nk_reg])) { 204 // When rematerialzing nodes and stretching lifetimes, the 205 // allocator will reuse the original def for multidef LRG instead 206 // of the current reaching def because it can't know it's safe to 207 // do so. After allocation completes if they are in the same LRG 208 // then it should use the current reaching def instead. 209 n->set_req(k, regnd[nk_reg]); 210 blk_adjust += yank_if_dead(val, current_block, &value, ®nd); 211 val = skip_copies(n->in(k)); 212 } 213 214 if( val == x ) return blk_adjust; // No progress? 215 216 bool single = is_single_register(val->ideal_reg()); 217 uint val_idx = n2lidx(val); 218 OptoReg::Name val_reg = lrgs(val_idx).reg(); 219 220 // See if it happens to already be in the correct register! 221 // (either Phi's direct register, or the common case of the name 222 // never-clobbered original-def register) 223 if( value[val_reg] == val && 224 // Doubles check both halves 225 ( single || value[val_reg-1] == val ) ) { 226 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); 227 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying 228 return blk_adjust; 229 } 230 231 // See if we can skip the copy by changing registers. Don't change from 232 // using a register to using the stack unless we know we can remove a 233 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill 234 // ops reading from memory instead of just loading once and using the 235 // register. 236 237 // Also handle duplicate copies here. 238 const Type *t = val->is_Con() ? val->bottom_type() : NULL; 239 240 // Scan all registers to see if this value is around already 241 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { 242 if (reg == (uint)nk_reg) { 243 // Found ourselves so check if there is only one user of this 244 // copy and keep on searching for a better copy if so. 245 bool ignore_self = true; 246 x = n->in(k); 247 DUIterator_Fast imax, i = x->fast_outs(imax); 248 Node* first = x->fast_out(i); i++; 249 while (i < imax && ignore_self) { 250 Node* use = x->fast_out(i); i++; 251 if (use != first) ignore_self = false; 252 } 253 if (ignore_self) continue; 254 } 255 256 Node *vv = value[reg]; 257 if( !single ) { // Doubles check for aligned-adjacent pair 258 if( (reg&1)==0 ) continue; // Wrong half of a pair 259 if( vv != value[reg-1] ) continue; // Not a complete pair 260 } 261 if( vv == val || // Got a direct hit? 262 (t && vv && vv->bottom_type() == t && vv->is_Mach() && 263 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? 264 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); 265 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR 266 OptoReg::is_reg(reg) || // turning into a register use OR 267 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use 268 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); 269 if( n->in(k) == regnd[reg] ) // Success! Quit trying 270 return blk_adjust; 271 } // End of if not degrading to a stack 272 } // End of if found value in another register 273 } // End of scan all machine registers 274 return blk_adjust; 275 } 276 277 278 // 279 // Check if nreg already contains the constant value val. Normal copy 280 // elimination doesn't doesn't work on constants because multiple 281 // nodes can represent the same constant so the type and rule of the 282 // MachNode must be checked to ensure equivalence. 283 // 284 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, 285 Block *current_block, 286 Node_List& value, Node_List& regnd, 287 OptoReg::Name nreg, OptoReg::Name nreg2) { 288 if (value[nreg] != val && val->is_Con() && 289 value[nreg] != NULL && value[nreg]->is_Con() && 290 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && 291 value[nreg]->bottom_type() == val->bottom_type() && 292 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { 293 // This code assumes that two MachNodes representing constants 294 // which have the same rule and the same bottom type will produce 295 // identical effects into a register. This seems like it must be 296 // objectively true unless there are hidden inputs to the nodes 297 // but if that were to change this code would need to updated. 298 // Since they are equivalent the second one if redundant and can 299 // be removed. 300 // 301 // n will be replaced with the old value but n might have 302 // kills projections associated with it so remove them now so that 303 // yank_if_dead will be able to eliminate the copy once the uses 304 // have been transferred to the old[value]. 305 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 306 Node* use = n->fast_out(i); 307 if (use->is_Proj() && use->outcnt() == 0) { 308 // Kill projections have no users and one input 309 use->set_req(0, C->top()); 310 yank_if_dead(use, current_block, &value, ®nd); 311 --i; --imax; 312 } 313 } 314 _post_alloc++; 315 return true; 316 } 317 return false; 318 } 319 320 321 //------------------------------post_allocate_copy_removal--------------------- 322 // Post-Allocation peephole copy removal. We do this in 1 pass over the 323 // basic blocks. We maintain a mapping of registers to Nodes (an array of 324 // Nodes indexed by machine register or stack slot number). NULL means that a 325 // register is not mapped to any Node. We can (want to have!) have several 326 // registers map to the same Node. We walk forward over the instructions 327 // updating the mapping as we go. At merge points we force a NULL if we have 328 // to merge 2 different Nodes into the same register. Phi functions will give 329 // us a new Node if there is a proper value merging. Since the blocks are 330 // arranged in some RPO, we will visit all parent blocks before visiting any 331 // successor blocks (except at loops). 332 // 333 // If we find a Copy we look to see if the Copy's source register is a stack 334 // slot and that value has already been loaded into some machine register; if 335 // so we use machine register directly. This turns a Load into a reg-reg 336 // Move. We also look for reloads of identical constants. 337 // 338 // When we see a use from a reg-reg Copy, we will attempt to use the copy's 339 // source directly and make the copy go dead. 340 void PhaseChaitin::post_allocate_copy_removal() { 341 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) 342 ResourceMark rm; 343 344 // Need a mapping from basic block Node_Lists. We need a Node_List to 345 // map from register number to value-producing Node. 346 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 347 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 348 // Need a mapping from basic block Node_Lists. We need a Node_List to 349 // map from register number to register-defining Node. 350 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 351 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 352 353 // We keep unused Node_Lists on a free_list to avoid wasting 354 // memory. 355 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); 356 357 // For all blocks 358 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 359 uint j; 360 Block *b = _cfg._blocks[i]; 361 362 // Count of Phis in block 363 uint phi_dex; 364 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { 365 Node *phi = b->_nodes[phi_dex]; 366 if( !phi->is_Phi() ) 367 break; 368 } 369 370 // If any predecessor has not been visited, we do not know the state 371 // of registers at the start. Check for this, while updating copies 372 // along Phi input edges 373 bool missing_some_inputs = false; 374 Block *freed = NULL; 375 for( j = 1; j < b->num_preds(); j++ ) { 376 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 377 // Remove copies along phi edges 378 for( uint k=1; k<phi_dex; k++ ) 379 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); 380 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? 381 // See if this predecessor's mappings have been used by everybody 382 // who wants them. If so, free 'em. 383 uint k; 384 for( k=0; k<pb->_num_succs; k++ ) { 385 Block *pbsucc = pb->_succs[k]; 386 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) 387 break; // Found a future user 388 } 389 if( k >= pb->_num_succs ) { // No more uses, free! 390 freed = pb; // Record last block freed 391 free_list.push(blk2value[pb->_pre_order]); 392 free_list.push(blk2regnd[pb->_pre_order]); 393 } 394 } else { // This block has unvisited (loopback) inputs 395 missing_some_inputs = true; 396 } 397 } 398 399 400 // Extract Node_List mappings. If 'freed' is non-zero, we just popped 401 // 'freed's blocks off the list 402 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 403 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 404 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); 405 value.map(_max_reg,NULL); 406 regnd.map(_max_reg,NULL); 407 // Set mappings as OUR mappings 408 blk2value[b->_pre_order] = &value; 409 blk2regnd[b->_pre_order] = ®nd; 410 411 // Initialize value & regnd for this block 412 if( missing_some_inputs ) { 413 // Some predecessor has not yet been visited; zap map to empty 414 for( uint k = 0; k < (uint)_max_reg; k++ ) { 415 value.map(k,NULL); 416 regnd.map(k,NULL); 417 } 418 } else { 419 if( !freed ) { // Didn't get a freebie prior block 420 // Must clone some data 421 freed = _cfg._bbs[b->pred(1)->_idx]; 422 Node_List &f_value = *blk2value[freed->_pre_order]; 423 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; 424 for( uint k = 0; k < (uint)_max_reg; k++ ) { 425 value.map(k,f_value[k]); 426 regnd.map(k,f_regnd[k]); 427 } 428 } 429 // Merge all inputs together, setting to NULL any conflicts. 430 for( j = 1; j < b->num_preds(); j++ ) { 431 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 432 if( pb == freed ) continue; // Did self already via freelist 433 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; 434 for( uint k = 0; k < (uint)_max_reg; k++ ) { 435 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? 436 value.map(k,NULL); // Then no value handy 437 regnd.map(k,NULL); 438 } 439 } 440 } 441 } 442 443 // For all Phi's 444 for( j = 1; j < phi_dex; j++ ) { 445 uint k; 446 Node *phi = b->_nodes[j]; 447 uint pidx = n2lidx(phi); 448 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); 449 450 // Remove copies remaining on edges. Check for junk phi. 451 Node *u = NULL; 452 for( k=1; k<phi->req(); k++ ) { 453 Node *x = phi->in(k); 454 if( phi != x && u != x ) // Found a different input 455 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input 456 } 457 if( u != NodeSentinel ) { // Junk Phi. Remove 458 b->_nodes.remove(j--); phi_dex--; 459 _cfg._bbs.map(phi->_idx,NULL); 460 phi->replace_by(u); 461 phi->disconnect_inputs(NULL); 462 continue; 463 } 464 // Note that if value[pidx] exists, then we merged no new values here 465 // and the phi is useless. This can happen even with the above phi 466 // removal for complex flows. I cannot keep the better known value here 467 // because locally the phi appears to define a new merged value. If I 468 // keep the better value then a copy of the phi, being unable to use the 469 // global flow analysis, can't "peek through" the phi to the original 470 // reaching value and so will act like it's defining a new value. This 471 // can lead to situations where some uses are from the old and some from 472 // the new values. Not illegal by itself but throws the over-strong 473 // assert in scheduling. 474 if( pidx ) { 475 value.map(preg,phi); 476 regnd.map(preg,phi); 477 OptoReg::Name preg_lo = OptoReg::add(preg,-1); 478 if( !is_single_register(phi->ideal_reg()) ) { 479 value.map(preg_lo,phi); 480 regnd.map(preg_lo,phi); 481 } 482 } 483 } 484 485 // For all remaining instructions 486 for( j = phi_dex; j < b->_nodes.size(); j++ ) { 487 Node *n = b->_nodes[j]; 488 489 if( n->outcnt() == 0 && // Dead? 490 n != C->top() && // (ignore TOP, it has no du info) 491 !n->is_Proj() ) { // fat-proj kills 492 j -= yank_if_dead(n,b,&value,®nd); 493 continue; 494 } 495 496 // Improve reaching-def info. Occasionally post-alloc's liveness gives 497 // up (at loop backedges, because we aren't doing a full flow pass). 498 // The presence of a live use essentially asserts that the use's def is 499 // alive and well at the use (or else the allocator fubar'd). Take 500 // advantage of this info to set a reaching def for the use-reg. 501 uint k; 502 for( k = 1; k < n->req(); k++ ) { 503 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE 504 guarantee(def != NULL, "no disconnected nodes at this point"); 505 uint useidx = n2lidx(def); // useidx is the live range index for this USE 506 507 if( useidx ) { 508 OptoReg::Name ureg = lrgs(useidx).reg(); 509 if( !value[ureg] ) { 510 int idx; // Skip occasional useless copy 511 while( (idx=def->is_Copy()) != 0 && 512 def->in(idx) != NULL && // NULL should not happen 513 ureg == lrgs(n2lidx(def->in(idx))).reg() ) 514 def = def->in(idx); 515 Node *valdef = skip_copies(def); // tighten up val through non-useless copies 516 value.map(ureg,valdef); // record improved reaching-def info 517 regnd.map(ureg, def); 518 // Record other half of doubles 519 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); 520 if( !is_single_register(def->ideal_reg()) && 521 ( !RegMask::can_represent(ureg_lo) || 522 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent 523 !value[ureg_lo] ) { 524 value.map(ureg_lo,valdef); // record improved reaching-def info 525 regnd.map(ureg_lo, def); 526 } 527 } 528 } 529 } 530 531 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; 532 533 // Remove copies along input edges 534 for( k = 1; k < n->req(); k++ ) 535 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); 536 537 // Unallocated Nodes define no registers 538 uint lidx = n2lidx(n); 539 if( !lidx ) continue; 540 541 // Update the register defined by this instruction 542 OptoReg::Name nreg = lrgs(lidx).reg(); 543 // Skip through all copies to the _value_ being defined. 544 // Do not change from int to pointer 545 Node *val = skip_copies(n); 546 547 // Clear out a dead definition before starting so that the 548 // elimination code doesn't have to guard against it. The 549 // definition could in fact be a kill projection with a count of 550 // 0 which is safe but since those are uninteresting for copy 551 // elimination just delete them as well. 552 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { 553 regnd.map(nreg, NULL); 554 value.map(nreg, NULL); 555 } 556 557 uint n_ideal_reg = n->ideal_reg(); 558 if( is_single_register(n_ideal_reg) ) { 559 // If Node 'n' does not change the value mapped by the register, 560 // then 'n' is a useless copy. Do not update the register->node 561 // mapping so 'n' will go dead. 562 if( value[nreg] != val ) { 563 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { 564 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 565 } else { 566 // Update the mapping: record new Node defined by the register 567 regnd.map(nreg,n); 568 // Update mapping for defined *value*, which is the defined 569 // Node after skipping all copies. 570 value.map(nreg,val); 571 } 572 } else if( !may_be_copy_of_callee(n) ) { 573 assert( n->is_Copy(), "" ); 574 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 575 } 576 } else { 577 // If the value occupies a register pair, record same info 578 // in both registers. 579 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); 580 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or 581 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent 582 // Sparc occasionally has non-adjacent pairs. 583 // Find the actual other value 584 RegMask tmp = lrgs(lidx).mask(); 585 tmp.Remove(nreg); 586 nreg_lo = tmp.find_first_elem(); 587 } 588 if( value[nreg] != val || value[nreg_lo] != val ) { 589 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { 590 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 591 } else { 592 regnd.map(nreg , n ); 593 regnd.map(nreg_lo, n ); 594 value.map(nreg ,val); 595 value.map(nreg_lo,val); 596 } 597 } else if( !may_be_copy_of_callee(n) ) { 598 assert( n->is_Copy(), "" ); 599 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 600 } 601 } 602 603 // Fat projections kill many registers 604 if( n_ideal_reg == MachProjNode::fat_proj ) { 605 RegMask rm = n->out_RegMask(); 606 // wow, what an expensive iterator... 607 nreg = rm.find_first_elem(); 608 while( OptoReg::is_valid(nreg)) { 609 rm.Remove(nreg); 610 value.map(nreg,n); 611 regnd.map(nreg,n); 612 nreg = rm.find_first_elem(); 613 } 614 } 615 616 } // End of for all instructions in the block 617 618 } // End for all blocks 619 }