1 /* 2 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/chaitin.hpp" 28 #include "opto/machnode.hpp" 29 30 // see if this register kind does not requires two registers 31 static bool is_single_register(uint x) { 32 #ifdef _LP64 33 return (x != Op_RegD && x != Op_RegL && x != Op_RegP); 34 #else 35 return (x != Op_RegD && x != Op_RegL); 36 #endif 37 } 38 39 //---------------------------may_be_copy_of_callee----------------------------- 40 // Check to see if we can possibly be a copy of a callee-save value. 41 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const { 42 // Short circuit if there are no callee save registers 43 if (_matcher.number_of_saved_registers() == 0) return false; 44 45 // Expect only a spill-down and reload on exit for callee-save spills. 46 // Chains of copies cannot be deep. 47 // 5008997 - This is wishful thinking. Register allocator seems to 48 // be splitting live ranges for callee save registers to such 49 // an extent that in large methods the chains can be very long 50 // (50+). The conservative answer is to return true if we don't 51 // know as this prevents optimizations from occurring. 52 53 const int limit = 60; 54 int i; 55 for( i=0; i < limit; i++ ) { 56 if( def->is_Proj() && def->in(0)->is_Start() && 57 _matcher.is_save_on_entry(lrgs(n2lidx(def)).reg()) ) 58 return true; // Direct use of callee-save proj 59 if( def->is_Copy() ) // Copies carry value through 60 def = def->in(def->is_Copy()); 61 else if( def->is_Phi() ) // Phis can merge it from any direction 62 def = def->in(1); 63 else 64 break; 65 guarantee(def != NULL, "must not resurrect dead copy"); 66 } 67 // If we reached the end and didn't find a callee save proj 68 // then this may be a callee save proj so we return true 69 // as the conservative answer. If we didn't reach then end 70 // we must have discovered that it was not a callee save 71 // else we would have returned. 72 return i == limit; 73 } 74 75 //------------------------------yank----------------------------------- 76 // Helper function for yank_if_dead 77 int PhaseChaitin::yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { 78 int blk_adjust=0; 79 Block *oldb = _cfg._bbs[old->_idx]; 80 oldb->find_remove(old); 81 // Count 1 if deleting an instruction from the current block 82 if( oldb == current_block ) blk_adjust++; 83 _cfg._bbs.map(old->_idx,NULL); 84 OptoReg::Name old_reg = lrgs(n2lidx(old)).reg(); 85 if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available? 86 value->map(old_reg,NULL); // Yank from value/regnd maps 87 regnd->map(old_reg,NULL); // This register's value is now unknown 88 } 89 return blk_adjust; 90 } 91 92 //------------------------------yank_if_dead----------------------------------- 93 // Removed an edge from 'old'. Yank if dead. Return adjustment counts to 94 // iterators in the current block. 95 int PhaseChaitin::yank_if_dead( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) { 96 int blk_adjust=0; 97 while (old->outcnt() == 0 && old != C->top()) { 98 blk_adjust += yank(old, current_block, value, regnd); 99 100 Node *tmp = NULL; 101 for (uint i = 1; i < old->req(); i++) { 102 if (old->in(i)->is_MachTemp()) { 103 Node* machtmp = old->in(i); 104 assert(machtmp->outcnt() == 1, "expected for a MachTemp"); 105 blk_adjust += yank(machtmp, current_block, value, regnd); 106 machtmp->disconnect_inputs(NULL); 107 } else { 108 assert(tmp == NULL, "can't handle more non MachTemp inputs"); 109 tmp = old->in(i); 110 } 111 } 112 old->disconnect_inputs(NULL); 113 if( !tmp ) break; 114 old = tmp; 115 } 116 return blk_adjust; 117 } 118 119 //------------------------------use_prior_register----------------------------- 120 // Use the prior value instead of the current value, in an effort to make 121 // the current value go dead. Return block iterator adjustment, in case 122 // we yank some instructions from this block. 123 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { 124 // No effect? 125 if( def == n->in(idx) ) return 0; 126 // Def is currently dead and can be removed? Do not resurrect 127 if( def->outcnt() == 0 ) return 0; 128 129 // Not every pair of physical registers are assignment compatible, 130 // e.g. on sparc floating point registers are not assignable to integer 131 // registers. 132 const LRG &def_lrg = lrgs(n2lidx(def)); 133 OptoReg::Name def_reg = def_lrg.reg(); 134 const RegMask &use_mask = n->in_RegMask(idx); 135 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) 136 : (use_mask.is_AllStack() != 0)); 137 // Check for a copy to or from a misaligned pair. 138 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); 139 140 if (!can_use) 141 return 0; 142 143 // Capture the old def in case it goes dead... 144 Node *old = n->in(idx); 145 146 // Save-on-call copies can only be elided if the entire copy chain can go 147 // away, lest we get the same callee-save value alive in 2 locations at 148 // once. We check for the obvious trivial case here. Although it can 149 // sometimes be elided with cooperation outside our scope, here we will just 150 // miss the opportunity. :-( 151 if( may_be_copy_of_callee(def) ) { 152 if( old->outcnt() > 1 ) return 0; // We're the not last user 153 int idx = old->is_Copy(); 154 assert( idx, "chain of copies being removed" ); 155 Node *old2 = old->in(idx); // Chain of copies 156 if( old2->outcnt() > 1 ) return 0; // old is not the last user 157 int idx2 = old2->is_Copy(); 158 if( !idx2 ) return 0; // Not a chain of 2 copies 159 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies 160 } 161 162 // Use the new def 163 n->set_req(idx,def); 164 _post_alloc++; 165 166 // Is old def now dead? We successfully yanked a copy? 167 return yank_if_dead(old,current_block,&value,®nd); 168 } 169 170 171 //------------------------------skip_copies------------------------------------ 172 // Skip through any number of copies (that don't mod oop-i-ness) 173 Node *PhaseChaitin::skip_copies( Node *c ) { 174 int idx = c->is_Copy(); 175 uint is_oop = lrgs(n2lidx(c))._is_oop; 176 while (idx != 0) { 177 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); 178 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) 179 break; // casting copy, not the same value 180 c = c->in(idx); 181 idx = c->is_Copy(); 182 } 183 return c; 184 } 185 186 //------------------------------elide_copy------------------------------------- 187 // Remove (bypass) copies along Node n, edge k. 188 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { 189 int blk_adjust = 0; 190 191 uint nk_idx = n2lidx(n->in(k)); 192 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); 193 194 // Remove obvious same-register copies 195 Node *x = n->in(k); 196 int idx; 197 while( (idx=x->is_Copy()) != 0 ) { 198 Node *copy = x->in(idx); 199 guarantee(copy != NULL, "must not resurrect dead copy"); 200 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; 201 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); 202 if( n->in(k) != copy ) break; // Failed for some cutout? 203 x = copy; // Progress, try again 204 } 205 206 // Phis and 2-address instructions cannot change registers so easily - their 207 // outputs must match their input. 208 if( !can_change_regs ) 209 return blk_adjust; // Only check stupid copies! 210 211 // Loop backedges won't have a value-mapping yet 212 if( &value == NULL ) return blk_adjust; 213 214 // Skip through all copies to the _value_ being used. Do not change from 215 // int to pointer. This attempts to jump through a chain of copies, where 216 // intermediate copies might be illegal, i.e., value is stored down to stack 217 // then reloaded BUT survives in a register the whole way. 218 Node *val = skip_copies(n->in(k)); 219 220 if (val == x && nk_idx != 0 && 221 regnd[nk_reg] != NULL && regnd[nk_reg] != x && 222 n2lidx(x) == n2lidx(regnd[nk_reg])) { 223 // When rematerialzing nodes and stretching lifetimes, the 224 // allocator will reuse the original def for multidef LRG instead 225 // of the current reaching def because it can't know it's safe to 226 // do so. After allocation completes if they are in the same LRG 227 // then it should use the current reaching def instead. 228 n->set_req(k, regnd[nk_reg]); 229 blk_adjust += yank_if_dead(val, current_block, &value, ®nd); 230 val = skip_copies(n->in(k)); 231 } 232 233 if( val == x ) return blk_adjust; // No progress? 234 235 bool single = is_single_register(val->ideal_reg()); 236 uint val_idx = n2lidx(val); 237 OptoReg::Name val_reg = lrgs(val_idx).reg(); 238 239 // See if it happens to already be in the correct register! 240 // (either Phi's direct register, or the common case of the name 241 // never-clobbered original-def register) 242 if( value[val_reg] == val && 243 // Doubles check both halves 244 ( single || value[val_reg-1] == val ) ) { 245 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); 246 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying 247 return blk_adjust; 248 } 249 250 // See if we can skip the copy by changing registers. Don't change from 251 // using a register to using the stack unless we know we can remove a 252 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill 253 // ops reading from memory instead of just loading once and using the 254 // register. 255 256 // Also handle duplicate copies here. 257 const Type *t = val->is_Con() ? val->bottom_type() : NULL; 258 259 // Scan all registers to see if this value is around already 260 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { 261 if (reg == (uint)nk_reg) { 262 // Found ourselves so check if there is only one user of this 263 // copy and keep on searching for a better copy if so. 264 bool ignore_self = true; 265 x = n->in(k); 266 DUIterator_Fast imax, i = x->fast_outs(imax); 267 Node* first = x->fast_out(i); i++; 268 while (i < imax && ignore_self) { 269 Node* use = x->fast_out(i); i++; 270 if (use != first) ignore_self = false; 271 } 272 if (ignore_self) continue; 273 } 274 275 Node *vv = value[reg]; 276 if( !single ) { // Doubles check for aligned-adjacent pair 277 if( (reg&1)==0 ) continue; // Wrong half of a pair 278 if( vv != value[reg-1] ) continue; // Not a complete pair 279 } 280 if( vv == val || // Got a direct hit? 281 (t && vv && vv->bottom_type() == t && vv->is_Mach() && 282 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? 283 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); 284 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR 285 OptoReg::is_reg(reg) || // turning into a register use OR 286 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use 287 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); 288 if( n->in(k) == regnd[reg] ) // Success! Quit trying 289 return blk_adjust; 290 } // End of if not degrading to a stack 291 } // End of if found value in another register 292 } // End of scan all machine registers 293 return blk_adjust; 294 } 295 296 297 // 298 // Check if nreg already contains the constant value val. Normal copy 299 // elimination doesn't doesn't work on constants because multiple 300 // nodes can represent the same constant so the type and rule of the 301 // MachNode must be checked to ensure equivalence. 302 // 303 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, 304 Block *current_block, 305 Node_List& value, Node_List& regnd, 306 OptoReg::Name nreg, OptoReg::Name nreg2) { 307 if (value[nreg] != val && val->is_Con() && 308 value[nreg] != NULL && value[nreg]->is_Con() && 309 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && 310 value[nreg]->bottom_type() == val->bottom_type() && 311 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { 312 // This code assumes that two MachNodes representing constants 313 // which have the same rule and the same bottom type will produce 314 // identical effects into a register. This seems like it must be 315 // objectively true unless there are hidden inputs to the nodes 316 // but if that were to change this code would need to updated. 317 // Since they are equivalent the second one if redundant and can 318 // be removed. 319 // 320 // n will be replaced with the old value but n might have 321 // kills projections associated with it so remove them now so that 322 // yank_if_dead will be able to eliminate the copy once the uses 323 // have been transferred to the old[value]. 324 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 325 Node* use = n->fast_out(i); 326 if (use->is_Proj() && use->outcnt() == 0) { 327 // Kill projections have no users and one input 328 use->set_req(0, C->top()); 329 yank_if_dead(use, current_block, &value, ®nd); 330 --i; --imax; 331 } 332 } 333 _post_alloc++; 334 return true; 335 } 336 return false; 337 } 338 339 340 //------------------------------post_allocate_copy_removal--------------------- 341 // Post-Allocation peephole copy removal. We do this in 1 pass over the 342 // basic blocks. We maintain a mapping of registers to Nodes (an array of 343 // Nodes indexed by machine register or stack slot number). NULL means that a 344 // register is not mapped to any Node. We can (want to have!) have several 345 // registers map to the same Node. We walk forward over the instructions 346 // updating the mapping as we go. At merge points we force a NULL if we have 347 // to merge 2 different Nodes into the same register. Phi functions will give 348 // us a new Node if there is a proper value merging. Since the blocks are 349 // arranged in some RPO, we will visit all parent blocks before visiting any 350 // successor blocks (except at loops). 351 // 352 // If we find a Copy we look to see if the Copy's source register is a stack 353 // slot and that value has already been loaded into some machine register; if 354 // so we use machine register directly. This turns a Load into a reg-reg 355 // Move. We also look for reloads of identical constants. 356 // 357 // When we see a use from a reg-reg Copy, we will attempt to use the copy's 358 // source directly and make the copy go dead. 359 void PhaseChaitin::post_allocate_copy_removal() { 360 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) 361 ResourceMark rm; 362 363 // Need a mapping from basic block Node_Lists. We need a Node_List to 364 // map from register number to value-producing Node. 365 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 366 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 367 // Need a mapping from basic block Node_Lists. We need a Node_List to 368 // map from register number to register-defining Node. 369 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 370 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 371 372 // We keep unused Node_Lists on a free_list to avoid wasting 373 // memory. 374 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); 375 376 // For all blocks 377 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 378 uint j; 379 Block *b = _cfg._blocks[i]; 380 381 // Count of Phis in block 382 uint phi_dex; 383 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { 384 Node *phi = b->_nodes[phi_dex]; 385 if( !phi->is_Phi() ) 386 break; 387 } 388 389 // If any predecessor has not been visited, we do not know the state 390 // of registers at the start. Check for this, while updating copies 391 // along Phi input edges 392 bool missing_some_inputs = false; 393 Block *freed = NULL; 394 for( j = 1; j < b->num_preds(); j++ ) { 395 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 396 // Remove copies along phi edges 397 for( uint k=1; k<phi_dex; k++ ) 398 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); 399 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? 400 // See if this predecessor's mappings have been used by everybody 401 // who wants them. If so, free 'em. 402 uint k; 403 for( k=0; k<pb->_num_succs; k++ ) { 404 Block *pbsucc = pb->_succs[k]; 405 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) 406 break; // Found a future user 407 } 408 if( k >= pb->_num_succs ) { // No more uses, free! 409 freed = pb; // Record last block freed 410 free_list.push(blk2value[pb->_pre_order]); 411 free_list.push(blk2regnd[pb->_pre_order]); 412 } 413 } else { // This block has unvisited (loopback) inputs 414 missing_some_inputs = true; 415 } 416 } 417 418 419 // Extract Node_List mappings. If 'freed' is non-zero, we just popped 420 // 'freed's blocks off the list 421 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 422 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 423 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); 424 value.map(_max_reg,NULL); 425 regnd.map(_max_reg,NULL); 426 // Set mappings as OUR mappings 427 blk2value[b->_pre_order] = &value; 428 blk2regnd[b->_pre_order] = ®nd; 429 430 // Initialize value & regnd for this block 431 if( missing_some_inputs ) { 432 // Some predecessor has not yet been visited; zap map to empty 433 for( uint k = 0; k < (uint)_max_reg; k++ ) { 434 value.map(k,NULL); 435 regnd.map(k,NULL); 436 } 437 } else { 438 if( !freed ) { // Didn't get a freebie prior block 439 // Must clone some data 440 freed = _cfg._bbs[b->pred(1)->_idx]; 441 Node_List &f_value = *blk2value[freed->_pre_order]; 442 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; 443 for( uint k = 0; k < (uint)_max_reg; k++ ) { 444 value.map(k,f_value[k]); 445 regnd.map(k,f_regnd[k]); 446 } 447 } 448 // Merge all inputs together, setting to NULL any conflicts. 449 for( j = 1; j < b->num_preds(); j++ ) { 450 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 451 if( pb == freed ) continue; // Did self already via freelist 452 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; 453 for( uint k = 0; k < (uint)_max_reg; k++ ) { 454 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? 455 value.map(k,NULL); // Then no value handy 456 regnd.map(k,NULL); 457 } 458 } 459 } 460 } 461 462 // For all Phi's 463 for( j = 1; j < phi_dex; j++ ) { 464 uint k; 465 Node *phi = b->_nodes[j]; 466 uint pidx = n2lidx(phi); 467 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); 468 469 // Remove copies remaining on edges. Check for junk phi. 470 Node *u = NULL; 471 for( k=1; k<phi->req(); k++ ) { 472 Node *x = phi->in(k); 473 if( phi != x && u != x ) // Found a different input 474 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input 475 } 476 if( u != NodeSentinel ) { // Junk Phi. Remove 477 b->_nodes.remove(j--); phi_dex--; 478 _cfg._bbs.map(phi->_idx,NULL); 479 phi->replace_by(u); 480 phi->disconnect_inputs(NULL); 481 continue; 482 } 483 // Note that if value[pidx] exists, then we merged no new values here 484 // and the phi is useless. This can happen even with the above phi 485 // removal for complex flows. I cannot keep the better known value here 486 // because locally the phi appears to define a new merged value. If I 487 // keep the better value then a copy of the phi, being unable to use the 488 // global flow analysis, can't "peek through" the phi to the original 489 // reaching value and so will act like it's defining a new value. This 490 // can lead to situations where some uses are from the old and some from 491 // the new values. Not illegal by itself but throws the over-strong 492 // assert in scheduling. 493 if( pidx ) { 494 value.map(preg,phi); 495 regnd.map(preg,phi); 496 OptoReg::Name preg_lo = OptoReg::add(preg,-1); 497 if( !is_single_register(phi->ideal_reg()) ) { 498 value.map(preg_lo,phi); 499 regnd.map(preg_lo,phi); 500 } 501 } 502 } 503 504 // For all remaining instructions 505 for( j = phi_dex; j < b->_nodes.size(); j++ ) { 506 Node *n = b->_nodes[j]; 507 508 if( n->outcnt() == 0 && // Dead? 509 n != C->top() && // (ignore TOP, it has no du info) 510 !n->is_Proj() ) { // fat-proj kills 511 j -= yank_if_dead(n,b,&value,®nd); 512 continue; 513 } 514 515 // Improve reaching-def info. Occasionally post-alloc's liveness gives 516 // up (at loop backedges, because we aren't doing a full flow pass). 517 // The presence of a live use essentially asserts that the use's def is 518 // alive and well at the use (or else the allocator fubar'd). Take 519 // advantage of this info to set a reaching def for the use-reg. 520 uint k; 521 for( k = 1; k < n->req(); k++ ) { 522 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE 523 guarantee(def != NULL, "no disconnected nodes at this point"); 524 uint useidx = n2lidx(def); // useidx is the live range index for this USE 525 526 if( useidx ) { 527 OptoReg::Name ureg = lrgs(useidx).reg(); 528 if( !value[ureg] ) { 529 int idx; // Skip occasional useless copy 530 while( (idx=def->is_Copy()) != 0 && 531 def->in(idx) != NULL && // NULL should not happen 532 ureg == lrgs(n2lidx(def->in(idx))).reg() ) 533 def = def->in(idx); 534 Node *valdef = skip_copies(def); // tighten up val through non-useless copies 535 value.map(ureg,valdef); // record improved reaching-def info 536 regnd.map(ureg, def); 537 // Record other half of doubles 538 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); 539 if( !is_single_register(def->ideal_reg()) && 540 ( !RegMask::can_represent(ureg_lo) || 541 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent 542 !value[ureg_lo] ) { 543 value.map(ureg_lo,valdef); // record improved reaching-def info 544 regnd.map(ureg_lo, def); 545 } 546 } 547 } 548 } 549 550 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; 551 552 // Remove copies along input edges 553 for( k = 1; k < n->req(); k++ ) 554 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); 555 556 // Unallocated Nodes define no registers 557 uint lidx = n2lidx(n); 558 if( !lidx ) continue; 559 560 // Update the register defined by this instruction 561 OptoReg::Name nreg = lrgs(lidx).reg(); 562 // Skip through all copies to the _value_ being defined. 563 // Do not change from int to pointer 564 Node *val = skip_copies(n); 565 566 // Clear out a dead definition before starting so that the 567 // elimination code doesn't have to guard against it. The 568 // definition could in fact be a kill projection with a count of 569 // 0 which is safe but since those are uninteresting for copy 570 // elimination just delete them as well. 571 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { 572 regnd.map(nreg, NULL); 573 value.map(nreg, NULL); 574 } 575 576 uint n_ideal_reg = n->ideal_reg(); 577 if( is_single_register(n_ideal_reg) ) { 578 // If Node 'n' does not change the value mapped by the register, 579 // then 'n' is a useless copy. Do not update the register->node 580 // mapping so 'n' will go dead. 581 if( value[nreg] != val ) { 582 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { 583 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 584 } else { 585 // Update the mapping: record new Node defined by the register 586 regnd.map(nreg,n); 587 // Update mapping for defined *value*, which is the defined 588 // Node after skipping all copies. 589 value.map(nreg,val); 590 } 591 } else if( !may_be_copy_of_callee(n) ) { 592 assert( n->is_Copy(), "" ); 593 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 594 } 595 } else { 596 // If the value occupies a register pair, record same info 597 // in both registers. 598 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); 599 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or 600 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent 601 // Sparc occasionally has non-adjacent pairs. 602 // Find the actual other value 603 RegMask tmp = lrgs(lidx).mask(); 604 tmp.Remove(nreg); 605 nreg_lo = tmp.find_first_elem(); 606 } 607 if( value[nreg] != val || value[nreg_lo] != val ) { 608 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { 609 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 610 } else { 611 regnd.map(nreg , n ); 612 regnd.map(nreg_lo, n ); 613 value.map(nreg ,val); 614 value.map(nreg_lo,val); 615 } 616 } else if( !may_be_copy_of_callee(n) ) { 617 assert( n->is_Copy(), "" ); 618 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 619 } 620 } 621 622 // Fat projections kill many registers 623 if( n_ideal_reg == MachProjNode::fat_proj ) { 624 RegMask rm = n->out_RegMask(); 625 // wow, what an expensive iterator... 626 nreg = rm.find_first_elem(); 627 while( OptoReg::is_valid(nreg)) { 628 rm.Remove(nreg); 629 value.map(nreg,n); 630 regnd.map(nreg,n); 631 nreg = rm.find_first_elem(); 632 } 633 } 634 635 } // End of for all instructions in the block 636 637 } // End for all blocks 638 }