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 #ifdef ASSERT 93 static bool expected_yanked_node(Node *old, Node *orig_old) { 94 // This code is expected only next original nodes: 95 // - load from constant table node which may have next data input nodes: 96 // MachConstantBase, Phi, MachTemp, MachSpillCopy 97 // - load constant node which may have next data input nodes: 98 // MachTemp, MachSpillCopy 99 // - MachSpillCopy 100 // - MachProj and Copy dead nodes 101 if (old->is_MachSpillCopy()) { 102 return true; 103 } else if (old->is_Con()) { 104 return true; 105 } else if (old->is_MachProj()) { // Dead kills projection of Con node 106 return (old == orig_old); 107 } else if (old->is_Copy()) { // Dead copy of a callee-save value 108 return (old == orig_old); 109 } else if (old->is_MachTemp()) { 110 return orig_old->is_Con(); 111 } else if (old->is_Phi() || old->is_MachConstantBase()) { 112 return (orig_old->is_Con() && orig_old->is_MachConstant()); 113 } 114 return false; 115 } 116 #endif 117 118 //------------------------------yank_if_dead----------------------------------- 119 // Removed edges from 'old'. Yank if dead. Return adjustment counts to 120 // iterators in the current block. 121 int PhaseChaitin::yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block, 122 Node_List *value, Node_List *regnd) { 123 int blk_adjust=0; 124 if (old->outcnt() == 0 && old != C->top()) { 125 #ifdef ASSERT 126 if (!expected_yanked_node(old, orig_old)) { 127 tty->print_cr("=============================================="); 128 tty->print_cr("orig_old:"); 129 orig_old->dump(); 130 tty->print_cr("old:"); 131 old->dump(); 132 assert(false, "unexpected yanked node"); 133 } 134 if (old->is_Con()) 135 orig_old = old; // Reset to satisfy expected nodes checks. 136 #endif 137 blk_adjust += yank(old, current_block, value, regnd); 138 139 for (uint i = 1; i < old->req(); i++) { 140 Node* n = old->in(i); 141 if (n != NULL) { 142 old->set_req(i, NULL); 143 blk_adjust += yank_if_dead_recurse(n, orig_old, current_block, value, regnd); 144 } 145 } 146 // Disconnect control and remove precedence edges if any exist 147 old->disconnect_inputs(NULL); 148 } 149 return blk_adjust; 150 } 151 152 //------------------------------use_prior_register----------------------------- 153 // Use the prior value instead of the current value, in an effort to make 154 // the current value go dead. Return block iterator adjustment, in case 155 // we yank some instructions from this block. 156 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List ®nd ) { 157 // No effect? 158 if( def == n->in(idx) ) return 0; 159 // Def is currently dead and can be removed? Do not resurrect 160 if( def->outcnt() == 0 ) return 0; 161 162 // Not every pair of physical registers are assignment compatible, 163 // e.g. on sparc floating point registers are not assignable to integer 164 // registers. 165 const LRG &def_lrg = lrgs(n2lidx(def)); 166 OptoReg::Name def_reg = def_lrg.reg(); 167 const RegMask &use_mask = n->in_RegMask(idx); 168 bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0) 169 : (use_mask.is_AllStack() != 0)); 170 // Check for a copy to or from a misaligned pair. 171 can_use = can_use && !use_mask.is_misaligned_Pair() && !def_lrg.mask().is_misaligned_Pair(); 172 173 if (!can_use) 174 return 0; 175 176 // Capture the old def in case it goes dead... 177 Node *old = n->in(idx); 178 179 // Save-on-call copies can only be elided if the entire copy chain can go 180 // away, lest we get the same callee-save value alive in 2 locations at 181 // once. We check for the obvious trivial case here. Although it can 182 // sometimes be elided with cooperation outside our scope, here we will just 183 // miss the opportunity. :-( 184 if( may_be_copy_of_callee(def) ) { 185 if( old->outcnt() > 1 ) return 0; // We're the not last user 186 int idx = old->is_Copy(); 187 assert( idx, "chain of copies being removed" ); 188 Node *old2 = old->in(idx); // Chain of copies 189 if( old2->outcnt() > 1 ) return 0; // old is not the last user 190 int idx2 = old2->is_Copy(); 191 if( !idx2 ) return 0; // Not a chain of 2 copies 192 if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies 193 } 194 195 // Use the new def 196 n->set_req(idx,def); 197 _post_alloc++; 198 199 // Is old def now dead? We successfully yanked a copy? 200 return yank_if_dead(old,current_block,&value,®nd); 201 } 202 203 204 //------------------------------skip_copies------------------------------------ 205 // Skip through any number of copies (that don't mod oop-i-ness) 206 Node *PhaseChaitin::skip_copies( Node *c ) { 207 int idx = c->is_Copy(); 208 uint is_oop = lrgs(n2lidx(c))._is_oop; 209 while (idx != 0) { 210 guarantee(c->in(idx) != NULL, "must not resurrect dead copy"); 211 if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop) 212 break; // casting copy, not the same value 213 c = c->in(idx); 214 idx = c->is_Copy(); 215 } 216 return c; 217 } 218 219 //------------------------------elide_copy------------------------------------- 220 // Remove (bypass) copies along Node n, edge k. 221 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List ®nd, bool can_change_regs ) { 222 int blk_adjust = 0; 223 224 uint nk_idx = n2lidx(n->in(k)); 225 OptoReg::Name nk_reg = lrgs(nk_idx ).reg(); 226 227 // Remove obvious same-register copies 228 Node *x = n->in(k); 229 int idx; 230 while( (idx=x->is_Copy()) != 0 ) { 231 Node *copy = x->in(idx); 232 guarantee(copy != NULL, "must not resurrect dead copy"); 233 if( lrgs(n2lidx(copy)).reg() != nk_reg ) break; 234 blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd); 235 if( n->in(k) != copy ) break; // Failed for some cutout? 236 x = copy; // Progress, try again 237 } 238 239 // Phis and 2-address instructions cannot change registers so easily - their 240 // outputs must match their input. 241 if( !can_change_regs ) 242 return blk_adjust; // Only check stupid copies! 243 244 // Loop backedges won't have a value-mapping yet 245 if( &value == NULL ) return blk_adjust; 246 247 // Skip through all copies to the _value_ being used. Do not change from 248 // int to pointer. This attempts to jump through a chain of copies, where 249 // intermediate copies might be illegal, i.e., value is stored down to stack 250 // then reloaded BUT survives in a register the whole way. 251 Node *val = skip_copies(n->in(k)); 252 253 if (val == x && nk_idx != 0 && 254 regnd[nk_reg] != NULL && regnd[nk_reg] != x && 255 n2lidx(x) == n2lidx(regnd[nk_reg])) { 256 // When rematerialzing nodes and stretching lifetimes, the 257 // allocator will reuse the original def for multidef LRG instead 258 // of the current reaching def because it can't know it's safe to 259 // do so. After allocation completes if they are in the same LRG 260 // then it should use the current reaching def instead. 261 n->set_req(k, regnd[nk_reg]); 262 blk_adjust += yank_if_dead(val, current_block, &value, ®nd); 263 val = skip_copies(n->in(k)); 264 } 265 266 if( val == x ) return blk_adjust; // No progress? 267 268 bool single = is_single_register(val->ideal_reg()); 269 uint val_idx = n2lidx(val); 270 OptoReg::Name val_reg = lrgs(val_idx).reg(); 271 272 // See if it happens to already be in the correct register! 273 // (either Phi's direct register, or the common case of the name 274 // never-clobbered original-def register) 275 if( value[val_reg] == val && 276 // Doubles check both halves 277 ( single || value[val_reg-1] == val ) ) { 278 blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd); 279 if( n->in(k) == regnd[val_reg] ) // Success! Quit trying 280 return blk_adjust; 281 } 282 283 // See if we can skip the copy by changing registers. Don't change from 284 // using a register to using the stack unless we know we can remove a 285 // copy-load. Otherwise we might end up making a pile of Intel cisc-spill 286 // ops reading from memory instead of just loading once and using the 287 // register. 288 289 // Also handle duplicate copies here. 290 const Type *t = val->is_Con() ? val->bottom_type() : NULL; 291 292 // Scan all registers to see if this value is around already 293 for( uint reg = 0; reg < (uint)_max_reg; reg++ ) { 294 if (reg == (uint)nk_reg) { 295 // Found ourselves so check if there is only one user of this 296 // copy and keep on searching for a better copy if so. 297 bool ignore_self = true; 298 x = n->in(k); 299 DUIterator_Fast imax, i = x->fast_outs(imax); 300 Node* first = x->fast_out(i); i++; 301 while (i < imax && ignore_self) { 302 Node* use = x->fast_out(i); i++; 303 if (use != first) ignore_self = false; 304 } 305 if (ignore_self) continue; 306 } 307 308 Node *vv = value[reg]; 309 if( !single ) { // Doubles check for aligned-adjacent pair 310 if( (reg&1)==0 ) continue; // Wrong half of a pair 311 if( vv != value[reg-1] ) continue; // Not a complete pair 312 } 313 if( vv == val || // Got a direct hit? 314 (t && vv && vv->bottom_type() == t && vv->is_Mach() && 315 vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant? 316 assert( !n->is_Phi(), "cannot change registers at a Phi so easily" ); 317 if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR 318 OptoReg::is_reg(reg) || // turning into a register use OR 319 regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use 320 blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd); 321 if( n->in(k) == regnd[reg] ) // Success! Quit trying 322 return blk_adjust; 323 } // End of if not degrading to a stack 324 } // End of if found value in another register 325 } // End of scan all machine registers 326 return blk_adjust; 327 } 328 329 330 // 331 // Check if nreg already contains the constant value val. Normal copy 332 // elimination doesn't doesn't work on constants because multiple 333 // nodes can represent the same constant so the type and rule of the 334 // MachNode must be checked to ensure equivalence. 335 // 336 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n, 337 Block *current_block, 338 Node_List& value, Node_List& regnd, 339 OptoReg::Name nreg, OptoReg::Name nreg2) { 340 if (value[nreg] != val && val->is_Con() && 341 value[nreg] != NULL && value[nreg]->is_Con() && 342 (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) && 343 value[nreg]->bottom_type() == val->bottom_type() && 344 value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) { 345 // This code assumes that two MachNodes representing constants 346 // which have the same rule and the same bottom type will produce 347 // identical effects into a register. This seems like it must be 348 // objectively true unless there are hidden inputs to the nodes 349 // but if that were to change this code would need to updated. 350 // Since they are equivalent the second one if redundant and can 351 // be removed. 352 // 353 // n will be replaced with the old value but n might have 354 // kills projections associated with it so remove them now so that 355 // yank_if_dead will be able to eliminate the copy once the uses 356 // have been transferred to the old[value]. 357 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 358 Node* use = n->fast_out(i); 359 if (use->is_Proj() && use->outcnt() == 0) { 360 // Kill projections have no users and one input 361 use->set_req(0, C->top()); 362 yank_if_dead(use, current_block, &value, ®nd); 363 --i; --imax; 364 } 365 } 366 _post_alloc++; 367 return true; 368 } 369 return false; 370 } 371 372 373 //------------------------------post_allocate_copy_removal--------------------- 374 // Post-Allocation peephole copy removal. We do this in 1 pass over the 375 // basic blocks. We maintain a mapping of registers to Nodes (an array of 376 // Nodes indexed by machine register or stack slot number). NULL means that a 377 // register is not mapped to any Node. We can (want to have!) have several 378 // registers map to the same Node. We walk forward over the instructions 379 // updating the mapping as we go. At merge points we force a NULL if we have 380 // to merge 2 different Nodes into the same register. Phi functions will give 381 // us a new Node if there is a proper value merging. Since the blocks are 382 // arranged in some RPO, we will visit all parent blocks before visiting any 383 // successor blocks (except at loops). 384 // 385 // If we find a Copy we look to see if the Copy's source register is a stack 386 // slot and that value has already been loaded into some machine register; if 387 // so we use machine register directly. This turns a Load into a reg-reg 388 // Move. We also look for reloads of identical constants. 389 // 390 // When we see a use from a reg-reg Copy, we will attempt to use the copy's 391 // source directly and make the copy go dead. 392 void PhaseChaitin::post_allocate_copy_removal() { 393 NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); ) 394 ResourceMark rm; 395 396 // Need a mapping from basic block Node_Lists. We need a Node_List to 397 // map from register number to value-producing Node. 398 Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 399 memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 400 // Need a mapping from basic block Node_Lists. We need a Node_List to 401 // map from register number to register-defining Node. 402 Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1); 403 memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) ); 404 405 // We keep unused Node_Lists on a free_list to avoid wasting 406 // memory. 407 GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16); 408 409 // For all blocks 410 for( uint i = 0; i < _cfg._num_blocks; i++ ) { 411 uint j; 412 Block *b = _cfg._blocks[i]; 413 414 // Count of Phis in block 415 uint phi_dex; 416 for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) { 417 Node *phi = b->_nodes[phi_dex]; 418 if( !phi->is_Phi() ) 419 break; 420 } 421 422 // If any predecessor has not been visited, we do not know the state 423 // of registers at the start. Check for this, while updating copies 424 // along Phi input edges 425 bool missing_some_inputs = false; 426 Block *freed = NULL; 427 for( j = 1; j < b->num_preds(); j++ ) { 428 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 429 // Remove copies along phi edges 430 for( uint k=1; k<phi_dex; k++ ) 431 elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false ); 432 if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge? 433 // See if this predecessor's mappings have been used by everybody 434 // who wants them. If so, free 'em. 435 uint k; 436 for( k=0; k<pb->_num_succs; k++ ) { 437 Block *pbsucc = pb->_succs[k]; 438 if( !blk2value[pbsucc->_pre_order] && pbsucc != b ) 439 break; // Found a future user 440 } 441 if( k >= pb->_num_succs ) { // No more uses, free! 442 freed = pb; // Record last block freed 443 free_list.push(blk2value[pb->_pre_order]); 444 free_list.push(blk2regnd[pb->_pre_order]); 445 } 446 } else { // This block has unvisited (loopback) inputs 447 missing_some_inputs = true; 448 } 449 } 450 451 452 // Extract Node_List mappings. If 'freed' is non-zero, we just popped 453 // 'freed's blocks off the list 454 Node_List ®nd = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 455 Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop()); 456 assert( !freed || blk2value[freed->_pre_order] == &value, "" ); 457 value.map(_max_reg,NULL); 458 regnd.map(_max_reg,NULL); 459 // Set mappings as OUR mappings 460 blk2value[b->_pre_order] = &value; 461 blk2regnd[b->_pre_order] = ®nd; 462 463 // Initialize value & regnd for this block 464 if( missing_some_inputs ) { 465 // Some predecessor has not yet been visited; zap map to empty 466 for( uint k = 0; k < (uint)_max_reg; k++ ) { 467 value.map(k,NULL); 468 regnd.map(k,NULL); 469 } 470 } else { 471 if( !freed ) { // Didn't get a freebie prior block 472 // Must clone some data 473 freed = _cfg._bbs[b->pred(1)->_idx]; 474 Node_List &f_value = *blk2value[freed->_pre_order]; 475 Node_List &f_regnd = *blk2regnd[freed->_pre_order]; 476 for( uint k = 0; k < (uint)_max_reg; k++ ) { 477 value.map(k,f_value[k]); 478 regnd.map(k,f_regnd[k]); 479 } 480 } 481 // Merge all inputs together, setting to NULL any conflicts. 482 for( j = 1; j < b->num_preds(); j++ ) { 483 Block *pb = _cfg._bbs[b->pred(j)->_idx]; 484 if( pb == freed ) continue; // Did self already via freelist 485 Node_List &p_regnd = *blk2regnd[pb->_pre_order]; 486 for( uint k = 0; k < (uint)_max_reg; k++ ) { 487 if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs? 488 value.map(k,NULL); // Then no value handy 489 regnd.map(k,NULL); 490 } 491 } 492 } 493 } 494 495 // For all Phi's 496 for( j = 1; j < phi_dex; j++ ) { 497 uint k; 498 Node *phi = b->_nodes[j]; 499 uint pidx = n2lidx(phi); 500 OptoReg::Name preg = lrgs(n2lidx(phi)).reg(); 501 502 // Remove copies remaining on edges. Check for junk phi. 503 Node *u = NULL; 504 for( k=1; k<phi->req(); k++ ) { 505 Node *x = phi->in(k); 506 if( phi != x && u != x ) // Found a different input 507 u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input 508 } 509 if( u != NodeSentinel ) { // Junk Phi. Remove 510 b->_nodes.remove(j--); phi_dex--; 511 _cfg._bbs.map(phi->_idx,NULL); 512 phi->replace_by(u); 513 phi->disconnect_inputs(NULL); 514 continue; 515 } 516 // Note that if value[pidx] exists, then we merged no new values here 517 // and the phi is useless. This can happen even with the above phi 518 // removal for complex flows. I cannot keep the better known value here 519 // because locally the phi appears to define a new merged value. If I 520 // keep the better value then a copy of the phi, being unable to use the 521 // global flow analysis, can't "peek through" the phi to the original 522 // reaching value and so will act like it's defining a new value. This 523 // can lead to situations where some uses are from the old and some from 524 // the new values. Not illegal by itself but throws the over-strong 525 // assert in scheduling. 526 if( pidx ) { 527 value.map(preg,phi); 528 regnd.map(preg,phi); 529 OptoReg::Name preg_lo = OptoReg::add(preg,-1); 530 if( !is_single_register(phi->ideal_reg()) ) { 531 value.map(preg_lo,phi); 532 regnd.map(preg_lo,phi); 533 } 534 } 535 } 536 537 // For all remaining instructions 538 for( j = phi_dex; j < b->_nodes.size(); j++ ) { 539 Node *n = b->_nodes[j]; 540 541 if( n->outcnt() == 0 && // Dead? 542 n != C->top() && // (ignore TOP, it has no du info) 543 !n->is_Proj() ) { // fat-proj kills 544 j -= yank_if_dead(n,b,&value,®nd); 545 continue; 546 } 547 548 // Improve reaching-def info. Occasionally post-alloc's liveness gives 549 // up (at loop backedges, because we aren't doing a full flow pass). 550 // The presence of a live use essentially asserts that the use's def is 551 // alive and well at the use (or else the allocator fubar'd). Take 552 // advantage of this info to set a reaching def for the use-reg. 553 uint k; 554 for( k = 1; k < n->req(); k++ ) { 555 Node *def = n->in(k); // n->in(k) is a USE; def is the DEF for this USE 556 guarantee(def != NULL, "no disconnected nodes at this point"); 557 uint useidx = n2lidx(def); // useidx is the live range index for this USE 558 559 if( useidx ) { 560 OptoReg::Name ureg = lrgs(useidx).reg(); 561 if( !value[ureg] ) { 562 int idx; // Skip occasional useless copy 563 while( (idx=def->is_Copy()) != 0 && 564 def->in(idx) != NULL && // NULL should not happen 565 ureg == lrgs(n2lidx(def->in(idx))).reg() ) 566 def = def->in(idx); 567 Node *valdef = skip_copies(def); // tighten up val through non-useless copies 568 value.map(ureg,valdef); // record improved reaching-def info 569 regnd.map(ureg, def); 570 // Record other half of doubles 571 OptoReg::Name ureg_lo = OptoReg::add(ureg,-1); 572 if( !is_single_register(def->ideal_reg()) && 573 ( !RegMask::can_represent(ureg_lo) || 574 lrgs(useidx).mask().Member(ureg_lo) ) && // Nearly always adjacent 575 !value[ureg_lo] ) { 576 value.map(ureg_lo,valdef); // record improved reaching-def info 577 regnd.map(ureg_lo, def); 578 } 579 } 580 } 581 } 582 583 const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0; 584 585 // Remove copies along input edges 586 for( k = 1; k < n->req(); k++ ) 587 j -= elide_copy( n, k, b, value, regnd, two_adr!=k ); 588 589 // Unallocated Nodes define no registers 590 uint lidx = n2lidx(n); 591 if( !lidx ) continue; 592 593 // Update the register defined by this instruction 594 OptoReg::Name nreg = lrgs(lidx).reg(); 595 // Skip through all copies to the _value_ being defined. 596 // Do not change from int to pointer 597 Node *val = skip_copies(n); 598 599 // Clear out a dead definition before starting so that the 600 // elimination code doesn't have to guard against it. The 601 // definition could in fact be a kill projection with a count of 602 // 0 which is safe but since those are uninteresting for copy 603 // elimination just delete them as well. 604 if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) { 605 regnd.map(nreg, NULL); 606 value.map(nreg, NULL); 607 } 608 609 uint n_ideal_reg = n->ideal_reg(); 610 if( is_single_register(n_ideal_reg) ) { 611 // If Node 'n' does not change the value mapped by the register, 612 // then 'n' is a useless copy. Do not update the register->node 613 // mapping so 'n' will go dead. 614 if( value[nreg] != val ) { 615 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) { 616 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 617 } else { 618 // Update the mapping: record new Node defined by the register 619 regnd.map(nreg,n); 620 // Update mapping for defined *value*, which is the defined 621 // Node after skipping all copies. 622 value.map(nreg,val); 623 } 624 } else if( !may_be_copy_of_callee(n) ) { 625 assert( n->is_Copy(), "" ); 626 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 627 } 628 } else { 629 // If the value occupies a register pair, record same info 630 // in both registers. 631 OptoReg::Name nreg_lo = OptoReg::add(nreg,-1); 632 if( RegMask::can_represent(nreg_lo) && // Either a spill slot, or 633 !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent 634 // Sparc occasionally has non-adjacent pairs. 635 // Find the actual other value 636 RegMask tmp = lrgs(lidx).mask(); 637 tmp.Remove(nreg); 638 nreg_lo = tmp.find_first_elem(); 639 } 640 if( value[nreg] != val || value[nreg_lo] != val ) { 641 if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) { 642 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 643 } else { 644 regnd.map(nreg , n ); 645 regnd.map(nreg_lo, n ); 646 value.map(nreg ,val); 647 value.map(nreg_lo,val); 648 } 649 } else if( !may_be_copy_of_callee(n) ) { 650 assert( n->is_Copy(), "" ); 651 j -= replace_and_yank_if_dead(n, nreg, b, value, regnd); 652 } 653 } 654 655 // Fat projections kill many registers 656 if( n_ideal_reg == MachProjNode::fat_proj ) { 657 RegMask rm = n->out_RegMask(); 658 // wow, what an expensive iterator... 659 nreg = rm.find_first_elem(); 660 while( OptoReg::is_valid(nreg)) { 661 rm.Remove(nreg); 662 value.map(nreg,n); 663 regnd.map(nreg,n); 664 nreg = rm.find_first_elem(); 665 } 666 } 667 668 } // End of for all instructions in the block 669 670 } // End for all blocks 671 }