1 /*
   2  * Copyright (c) 1998, 2013, 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 (or pairs, or vector) already contains the value.
  31 static bool register_contains_value(Node* val, OptoReg::Name reg, int n_regs,
  32                                     Node_List& value) {
  33   for (int i = 0; i < n_regs; i++) {
  34     OptoReg::Name nreg = OptoReg::add(reg,-i);
  35     if (value[nreg] != val)
  36       return false;
  37   }
  38   return true;
  39 }
  40 
  41 //---------------------------may_be_copy_of_callee-----------------------------
  42 // Check to see if we can possibly be a copy of a callee-save value.
  43 bool PhaseChaitin::may_be_copy_of_callee( Node *def ) const {
  44   // Short circuit if there are no callee save registers
  45   if (_matcher.number_of_saved_registers() == 0) return false;
  46 
  47   // Expect only a spill-down and reload on exit for callee-save spills.
  48   // Chains of copies cannot be deep.
  49   // 5008997 - This is wishful thinking. Register allocator seems to
  50   // be splitting live ranges for callee save registers to such
  51   // an extent that in large methods the chains can be very long
  52   // (50+). The conservative answer is to return true if we don't
  53   // know as this prevents optimizations from occurring.
  54 
  55   const int limit = 60;
  56   int i;
  57   for( i=0; i < limit; i++ ) {
  58     if( def->is_Proj() && def->in(0)->is_Start() &&
  59         _matcher.is_save_on_entry(lrgs(_lrg_map.live_range_id(def)).reg()))
  60       return true;              // Direct use of callee-save proj
  61     if( def->is_Copy() )        // Copies carry value through
  62       def = def->in(def->is_Copy());
  63     else if( def->is_Phi() )    // Phis can merge it from any direction
  64       def = def->in(1);
  65     else
  66       break;
  67     guarantee(def != NULL, "must not resurrect dead copy");
  68   }
  69   // If we reached the end and didn't find a callee save proj
  70   // then this may be a callee save proj so we return true
  71   // as the conservative answer. If we didn't reach then end
  72   // we must have discovered that it was not a callee save
  73   // else we would have returned.
  74   return i == limit;
  75 }
  76 
  77 //------------------------------yank-----------------------------------
  78 // Helper function for yank_if_dead
  79 int PhaseChaitin::yank( Node *old, Block *current_block, Node_List *value, Node_List *regnd ) {
  80   int blk_adjust=0;
  81   Block *oldb = _cfg.get_block_for_node(old);
  82   oldb->find_remove(old);
  83   // Count 1 if deleting an instruction from the current block
  84   if (oldb == current_block) {
  85     blk_adjust++;
  86   }
  87   _cfg.unmap_node_from_block(old);
  88   OptoReg::Name old_reg = lrgs(_lrg_map.live_range_id(old)).reg();
  89   if( regnd && (*regnd)[old_reg]==old ) { // Instruction is currently available?
  90     value->map(old_reg,NULL);  // Yank from value/regnd maps
  91     regnd->map(old_reg,NULL);  // This register's value is now unknown
  92   }
  93   return blk_adjust;
  94 }
  95 
  96 #ifdef ASSERT
  97 static bool expected_yanked_node(Node *old, Node *orig_old) {
  98   // This code is expected only next original nodes:
  99   // - load from constant table node which may have next data input nodes:
 100   //     MachConstantBase, MachTemp, MachSpillCopy
 101   // - Phi nodes that are considered Junk
 102   // - load constant node which may have next data input nodes:
 103   //     MachTemp, MachSpillCopy
 104   // - MachSpillCopy
 105   // - MachProj and Copy dead nodes
 106   if (old->is_MachSpillCopy()) {
 107     return true;
 108   } else if (old->is_Con()) {
 109     return true;
 110   } else if (old->is_MachProj()) { // Dead kills projection of Con node
 111     return (old == orig_old);
 112   } else if (old->is_Copy()) {     // Dead copy of a callee-save value
 113     return (old == orig_old);
 114   } else if (old->is_MachTemp()) {
 115     return orig_old->is_Con();
 116   } else if (old->is_Phi()) { // Junk phi's
 117     return true;
 118   } else if (old->is_MachConstantBase()) {
 119     return (orig_old->is_Con() && orig_old->is_MachConstant());
 120   }
 121   return false;
 122 }
 123 #endif
 124 
 125 //------------------------------yank_if_dead-----------------------------------
 126 // Removed edges from 'old'.  Yank if dead.  Return adjustment counts to
 127 // iterators in the current block.
 128 int PhaseChaitin::yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block,
 129                                        Node_List *value, Node_List *regnd) {
 130   int blk_adjust=0;
 131   if (old->outcnt() == 0 && old != C->top()) {
 132 #ifdef ASSERT
 133     if (!expected_yanked_node(old, orig_old)) {
 134       tty->print_cr("==============================================");
 135       tty->print_cr("orig_old:");
 136       orig_old->dump();
 137       tty->print_cr("old:");
 138       old->dump();
 139       assert(false, "unexpected yanked node");
 140     }
 141     if (old->is_Con())
 142       orig_old = old; // Reset to satisfy expected nodes checks.
 143 #endif
 144     blk_adjust += yank(old, current_block, value, regnd);
 145 
 146     for (uint i = 1; i < old->req(); i++) {
 147       Node* n = old->in(i);
 148       if (n != NULL) {
 149         old->set_req(i, NULL);
 150         blk_adjust += yank_if_dead_recurse(n, orig_old, current_block, value, regnd);
 151       }
 152     }
 153     // Disconnect control and remove precedence edges if any exist
 154     old->disconnect_inputs(NULL, C);
 155   }
 156   return blk_adjust;
 157 }
 158 
 159 //------------------------------use_prior_register-----------------------------
 160 // Use the prior value instead of the current value, in an effort to make
 161 // the current value go dead.  Return block iterator adjustment, in case
 162 // we yank some instructions from this block.
 163 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List &regnd ) {
 164   // No effect?
 165   if( def == n->in(idx) ) return 0;
 166   // Def is currently dead and can be removed?  Do not resurrect
 167   if( def->outcnt() == 0 ) return 0;
 168 
 169   // Not every pair of physical registers are assignment compatible,
 170   // e.g. on sparc floating point registers are not assignable to integer
 171   // registers.
 172   const LRG &def_lrg = lrgs(_lrg_map.live_range_id(def));
 173   OptoReg::Name def_reg = def_lrg.reg();
 174   const RegMask &use_mask = n->in_RegMask(idx);
 175   bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0)
 176                                                    : (use_mask.is_AllStack() != 0));
 177   if (!RegMask::is_vector(def->ideal_reg())) {
 178     // Check for a copy to or from a misaligned pair.
 179     // It is workaround for a sparc with misaligned pairs.
 180     can_use = can_use && !use_mask.is_misaligned_pair() && !def_lrg.mask().is_misaligned_pair();
 181   }
 182   if (!can_use)
 183     return 0;
 184 
 185   // Capture the old def in case it goes dead...
 186   Node *old = n->in(idx);
 187 
 188   // Save-on-call copies can only be elided if the entire copy chain can go
 189   // away, lest we get the same callee-save value alive in 2 locations at
 190   // once.  We check for the obvious trivial case here.  Although it can
 191   // sometimes be elided with cooperation outside our scope, here we will just
 192   // miss the opportunity.  :-(
 193   if( may_be_copy_of_callee(def) ) {
 194     if( old->outcnt() > 1 ) return 0; // We're the not last user
 195     int idx = old->is_Copy();
 196     assert( idx, "chain of copies being removed" );
 197     Node *old2 = old->in(idx);  // Chain of copies
 198     if( old2->outcnt() > 1 ) return 0; // old is not the last user
 199     int idx2 = old2->is_Copy();
 200     if( !idx2 ) return 0;       // Not a chain of 2 copies
 201     if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies
 202   }
 203 
 204   // Use the new def
 205   n->set_req(idx,def);
 206   _post_alloc++;
 207 
 208   // Is old def now dead?  We successfully yanked a copy?
 209   return yank_if_dead(old,current_block,&value,&regnd);
 210 }
 211 
 212 
 213 //------------------------------skip_copies------------------------------------
 214 // Skip through any number of copies (that don't mod oop-i-ness)
 215 Node *PhaseChaitin::skip_copies( Node *c ) {
 216   int idx = c->is_Copy();
 217   uint is_oop = lrgs(_lrg_map.live_range_id(c))._is_oop;
 218   while (idx != 0) {
 219     guarantee(c->in(idx) != NULL, "must not resurrect dead copy");
 220     if (lrgs(_lrg_map.live_range_id(c->in(idx)))._is_oop != is_oop) {
 221       break;  // casting copy, not the same value
 222     }
 223     c = c->in(idx);
 224     idx = c->is_Copy();
 225   }
 226   return c;
 227 }
 228 
 229 //------------------------------elide_copy-------------------------------------
 230 // Remove (bypass) copies along Node n, edge k.
 231 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List &regnd, bool can_change_regs ) {
 232   int blk_adjust = 0;
 233 
 234   uint nk_idx = _lrg_map.live_range_id(n->in(k));
 235   OptoReg::Name nk_reg = lrgs(nk_idx).reg();
 236 
 237   // Remove obvious same-register copies
 238   Node *x = n->in(k);
 239   int idx;
 240   while( (idx=x->is_Copy()) != 0 ) {
 241     Node *copy = x->in(idx);
 242     guarantee(copy != NULL, "must not resurrect dead copy");
 243     if(lrgs(_lrg_map.live_range_id(copy)).reg() != nk_reg) {
 244       break;
 245     }
 246     blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd);
 247     if (n->in(k) != copy) {
 248       break; // Failed for some cutout?
 249     }
 250     x = copy;                   // Progress, try again
 251   }
 252 
 253   // Phis and 2-address instructions cannot change registers so easily - their
 254   // outputs must match their input.
 255   if( !can_change_regs )
 256     return blk_adjust;          // Only check stupid copies!
 257 
 258   // Loop backedges won't have a value-mapping yet
 259   if( &value == NULL ) return blk_adjust;
 260 
 261   // Skip through all copies to the _value_ being used.  Do not change from
 262   // int to pointer.  This attempts to jump through a chain of copies, where
 263   // intermediate copies might be illegal, i.e., value is stored down to stack
 264   // then reloaded BUT survives in a register the whole way.
 265   Node *val = skip_copies(n->in(k));
 266 
 267   if (val == x && nk_idx != 0 &&
 268       regnd[nk_reg] != NULL && regnd[nk_reg] != x &&
 269       _lrg_map.live_range_id(x) == _lrg_map.live_range_id(regnd[nk_reg])) {
 270     // When rematerialzing nodes and stretching lifetimes, the
 271     // allocator will reuse the original def for multidef LRG instead
 272     // of the current reaching def because it can't know it's safe to
 273     // do so.  After allocation completes if they are in the same LRG
 274     // then it should use the current reaching def instead.
 275     n->set_req(k, regnd[nk_reg]);
 276     blk_adjust += yank_if_dead(val, current_block, &value, &regnd);
 277     val = skip_copies(n->in(k));
 278   }
 279 
 280   if (val == x) return blk_adjust; // No progress?
 281 
 282   int n_regs = RegMask::num_registers(val->ideal_reg());
 283   uint val_idx = _lrg_map.live_range_id(val);
 284   OptoReg::Name val_reg = lrgs(val_idx).reg();
 285 
 286   // See if it happens to already be in the correct register!
 287   // (either Phi's direct register, or the common case of the name
 288   // never-clobbered original-def register)
 289   if (register_contains_value(val, val_reg, n_regs, value)) {
 290     blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd);
 291     if( n->in(k) == regnd[val_reg] ) // Success!  Quit trying
 292       return blk_adjust;
 293   }
 294 
 295   // See if we can skip the copy by changing registers.  Don't change from
 296   // using a register to using the stack unless we know we can remove a
 297   // copy-load.  Otherwise we might end up making a pile of Intel cisc-spill
 298   // ops reading from memory instead of just loading once and using the
 299   // register.
 300 
 301   // Also handle duplicate copies here.
 302   const Type *t = val->is_Con() ? val->bottom_type() : NULL;
 303 
 304   // Scan all registers to see if this value is around already
 305   for( uint reg = 0; reg < (uint)_max_reg; reg++ ) {
 306     if (reg == (uint)nk_reg) {
 307       // Found ourselves so check if there is only one user of this
 308       // copy and keep on searching for a better copy if so.
 309       bool ignore_self = true;
 310       x = n->in(k);
 311       DUIterator_Fast imax, i = x->fast_outs(imax);
 312       Node* first = x->fast_out(i); i++;
 313       while (i < imax && ignore_self) {
 314         Node* use = x->fast_out(i); i++;
 315         if (use != first) ignore_self = false;
 316       }
 317       if (ignore_self) continue;
 318     }
 319 
 320     Node *vv = value[reg];
 321     if (n_regs > 1) { // Doubles and vectors check for aligned-adjacent set
 322       uint last = (n_regs-1); // Looking for the last part of a set
 323       if ((reg&last) != last) continue; // Wrong part of a set
 324       if (!register_contains_value(vv, reg, n_regs, value)) continue; // Different value
 325     }
 326     if( vv == val ||            // Got a direct hit?
 327         (t && vv && vv->bottom_type() == t && vv->is_Mach() &&
 328          vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant?
 329       assert( !n->is_Phi(), "cannot change registers at a Phi so easily" );
 330       if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR
 331           OptoReg::is_reg(reg) || // turning into a register use OR
 332           regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use
 333         blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd);
 334         if( n->in(k) == regnd[reg] ) // Success!  Quit trying
 335           return blk_adjust;
 336       } // End of if not degrading to a stack
 337     } // End of if found value in another register
 338   } // End of scan all machine registers
 339   return blk_adjust;
 340 }
 341 
 342 
 343 //
 344 // Check if nreg already contains the constant value val.  Normal copy
 345 // elimination doesn't doesn't work on constants because multiple
 346 // nodes can represent the same constant so the type and rule of the
 347 // MachNode must be checked to ensure equivalence.
 348 //
 349 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n,
 350                                               Block *current_block,
 351                                               Node_List& value, Node_List& regnd,
 352                                               OptoReg::Name nreg, OptoReg::Name nreg2) {
 353   if (value[nreg] != val && val->is_Con() &&
 354       value[nreg] != NULL && value[nreg]->is_Con() &&
 355       (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) &&
 356       value[nreg]->bottom_type() == val->bottom_type() &&
 357       value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) {
 358     // This code assumes that two MachNodes representing constants
 359     // which have the same rule and the same bottom type will produce
 360     // identical effects into a register.  This seems like it must be
 361     // objectively true unless there are hidden inputs to the nodes
 362     // but if that were to change this code would need to updated.
 363     // Since they are equivalent the second one if redundant and can
 364     // be removed.
 365     //
 366     // n will be replaced with the old value but n might have
 367     // kills projections associated with it so remove them now so that
 368     // yank_if_dead will be able to eliminate the copy once the uses
 369     // have been transferred to the old[value].
 370     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 371       Node* use = n->fast_out(i);
 372       if (use->is_Proj() && use->outcnt() == 0) {
 373         // Kill projections have no users and one input
 374         use->set_req(0, C->top());
 375         yank_if_dead(use, current_block, &value, &regnd);
 376         --i; --imax;
 377       }
 378     }
 379     _post_alloc++;
 380     return true;
 381   }
 382   return false;
 383 }
 384 
 385 
 386 //------------------------------post_allocate_copy_removal---------------------
 387 // Post-Allocation peephole copy removal.  We do this in 1 pass over the
 388 // basic blocks.  We maintain a mapping of registers to Nodes (an  array of
 389 // Nodes indexed by machine register or stack slot number).  NULL means that a
 390 // register is not mapped to any Node.  We can (want to have!) have several
 391 // registers map to the same Node.  We walk forward over the instructions
 392 // updating the mapping as we go.  At merge points we force a NULL if we have
 393 // to merge 2 different Nodes into the same register.  Phi functions will give
 394 // us a new Node if there is a proper value merging.  Since the blocks are
 395 // arranged in some RPO, we will visit all parent blocks before visiting any
 396 // successor blocks (except at loops).
 397 //
 398 // If we find a Copy we look to see if the Copy's source register is a stack
 399 // slot and that value has already been loaded into some machine register; if
 400 // so we use machine register directly.  This turns a Load into a reg-reg
 401 // Move.  We also look for reloads of identical constants.
 402 //
 403 // When we see a use from a reg-reg Copy, we will attempt to use the copy's
 404 // source directly and make the copy go dead.
 405 void PhaseChaitin::post_allocate_copy_removal() {
 406   Compile::TracePhase t3("postAllocCopyRemoval", &timers[_t_postAllocCopyRemoval]);
 407   ResourceMark rm;
 408 
 409   // Need a mapping from basic block Node_Lists.  We need a Node_List to
 410   // map from register number to value-producing Node.
 411   Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1);
 412   memset(blk2value, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1));
 413   // Need a mapping from basic block Node_Lists.  We need a Node_List to
 414   // map from register number to register-defining Node.
 415   Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg.number_of_blocks() + 1);
 416   memset(blk2regnd, 0, sizeof(Node_List*) * (_cfg.number_of_blocks() + 1));
 417 
 418   // We keep unused Node_Lists on a free_list to avoid wasting
 419   // memory.
 420   GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16);
 421 
 422   // For all blocks
 423   for (uint i = 0; i < _cfg.number_of_blocks(); i++) {
 424     uint j;
 425     Block* block = _cfg.get_block(i);
 426 
 427     // Count of Phis in block
 428     uint phi_dex;
 429     for (phi_dex = 1; phi_dex < block->number_of_nodes(); phi_dex++) {
 430       Node* phi = block->get_node(phi_dex);
 431       if (!phi->is_Phi()) {
 432         break;
 433       }
 434     }
 435 
 436     // If any predecessor has not been visited, we do not know the state
 437     // of registers at the start.  Check for this, while updating copies
 438     // along Phi input edges
 439     bool missing_some_inputs = false;
 440     Block *freed = NULL;
 441     for (j = 1; j < block->num_preds(); j++) {
 442       Block* pb = _cfg.get_block_for_node(block->pred(j));
 443       // Remove copies along phi edges
 444       for (uint k = 1; k < phi_dex; k++) {
 445         elide_copy(block->get_node(k), j, block, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false);
 446       }
 447       if (blk2value[pb->_pre_order]) { // Have a mapping on this edge?
 448         // See if this predecessor's mappings have been used by everybody
 449         // who wants them.  If so, free 'em.
 450         uint k;
 451         for (k = 0; k < pb->_num_succs; k++) {
 452           Block* pbsucc = pb->_succs[k];
 453           if (!blk2value[pbsucc->_pre_order] && pbsucc != block) {
 454             break;              // Found a future user
 455           }
 456         }
 457         if (k >= pb->_num_succs) { // No more uses, free!
 458           freed = pb;           // Record last block freed
 459           free_list.push(blk2value[pb->_pre_order]);
 460           free_list.push(blk2regnd[pb->_pre_order]);
 461         }
 462       } else {                  // This block has unvisited (loopback) inputs
 463         missing_some_inputs = true;
 464       }
 465     }
 466 
 467 
 468     // Extract Node_List mappings.  If 'freed' is non-zero, we just popped
 469     // 'freed's blocks off the list
 470     Node_List &regnd = *(free_list.is_empty() ? new Node_List() : free_list.pop());
 471     Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop());
 472     assert( !freed || blk2value[freed->_pre_order] == &value, "" );
 473     value.map(_max_reg,NULL);
 474     regnd.map(_max_reg,NULL);
 475     // Set mappings as OUR mappings
 476     blk2value[block->_pre_order] = &value;
 477     blk2regnd[block->_pre_order] = &regnd;
 478 
 479     // Initialize value & regnd for this block
 480     if (missing_some_inputs) {
 481       // Some predecessor has not yet been visited; zap map to empty
 482       for (uint k = 0; k < (uint)_max_reg; k++) {
 483         value.map(k,NULL);
 484         regnd.map(k,NULL);
 485       }
 486     } else {
 487       if( !freed ) {            // Didn't get a freebie prior block
 488         // Must clone some data
 489         freed = _cfg.get_block_for_node(block->pred(1));
 490         Node_List &f_value = *blk2value[freed->_pre_order];
 491         Node_List &f_regnd = *blk2regnd[freed->_pre_order];
 492         for( uint k = 0; k < (uint)_max_reg; k++ ) {
 493           value.map(k,f_value[k]);
 494           regnd.map(k,f_regnd[k]);
 495         }
 496       }
 497       // Merge all inputs together, setting to NULL any conflicts.
 498       for (j = 1; j < block->num_preds(); j++) {
 499         Block* pb = _cfg.get_block_for_node(block->pred(j));
 500         if (pb == freed) {
 501           continue; // Did self already via freelist
 502         }
 503         Node_List &p_regnd = *blk2regnd[pb->_pre_order];
 504         for( uint k = 0; k < (uint)_max_reg; k++ ) {
 505           if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs?
 506             value.map(k,NULL); // Then no value handy
 507             regnd.map(k,NULL);
 508           }
 509         }
 510       }
 511     }
 512 
 513     // For all Phi's
 514     for (j = 1; j < phi_dex; j++) {
 515       uint k;
 516       Node *phi = block->get_node(j);
 517       uint pidx = _lrg_map.live_range_id(phi);
 518       OptoReg::Name preg = lrgs(_lrg_map.live_range_id(phi)).reg();
 519 
 520       // Remove copies remaining on edges.  Check for junk phi.
 521       Node *u = NULL;
 522       for (k = 1; k < phi->req(); k++) {
 523         Node *x = phi->in(k);
 524         if( phi != x && u != x ) // Found a different input
 525           u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input
 526       }
 527       if (u != NodeSentinel) {    // Junk Phi.  Remove
 528         phi->replace_by(u);
 529         j -= yank_if_dead(phi, block, &value, &regnd);
 530         phi_dex--;
 531         continue;
 532       }
 533       // Note that if value[pidx] exists, then we merged no new values here
 534       // and the phi is useless.  This can happen even with the above phi
 535       // removal for complex flows.  I cannot keep the better known value here
 536       // because locally the phi appears to define a new merged value.  If I
 537       // keep the better value then a copy of the phi, being unable to use the
 538       // global flow analysis, can't "peek through" the phi to the original
 539       // reaching value and so will act like it's defining a new value.  This
 540       // can lead to situations where some uses are from the old and some from
 541       // the new values.  Not illegal by itself but throws the over-strong
 542       // assert in scheduling.
 543       if( pidx ) {
 544         value.map(preg,phi);
 545         regnd.map(preg,phi);
 546         int n_regs = RegMask::num_registers(phi->ideal_reg());
 547         for (int l = 1; l < n_regs; l++) {
 548           OptoReg::Name preg_lo = OptoReg::add(preg,-l);
 549           value.map(preg_lo,phi);
 550           regnd.map(preg_lo,phi);
 551         }
 552       }
 553     }
 554 
 555     // For all remaining instructions
 556     for (j = phi_dex; j < block->number_of_nodes(); j++) {
 557       Node* n = block->get_node(j);
 558 
 559       if(n->outcnt() == 0 &&   // Dead?
 560          n != C->top() &&      // (ignore TOP, it has no du info)
 561          !n->is_Proj() ) {     // fat-proj kills
 562         j -= yank_if_dead(n, block, &value, &regnd);
 563         continue;
 564       }
 565 
 566       // Improve reaching-def info.  Occasionally post-alloc's liveness gives
 567       // up (at loop backedges, because we aren't doing a full flow pass).
 568       // The presence of a live use essentially asserts that the use's def is
 569       // alive and well at the use (or else the allocator fubar'd).  Take
 570       // advantage of this info to set a reaching def for the use-reg.
 571       uint k;
 572       for (k = 1; k < n->req(); k++) {
 573         Node *def = n->in(k);   // n->in(k) is a USE; def is the DEF for this USE
 574         guarantee(def != NULL, "no disconnected nodes at this point");
 575         uint useidx = _lrg_map.live_range_id(def); // useidx is the live range index for this USE
 576 
 577         if( useidx ) {
 578           OptoReg::Name ureg = lrgs(useidx).reg();
 579           if( !value[ureg] ) {
 580             int idx;            // Skip occasional useless copy
 581             while( (idx=def->is_Copy()) != 0 &&
 582                    def->in(idx) != NULL &&  // NULL should not happen
 583                    ureg == lrgs(_lrg_map.live_range_id(def->in(idx))).reg())
 584               def = def->in(idx);
 585             Node *valdef = skip_copies(def); // tighten up val through non-useless copies
 586             value.map(ureg,valdef); // record improved reaching-def info
 587             regnd.map(ureg,   def);
 588             // Record other half of doubles
 589             uint def_ideal_reg = def->ideal_reg();
 590             int n_regs = RegMask::num_registers(def_ideal_reg);
 591             for (int l = 1; l < n_regs; l++) {
 592               OptoReg::Name ureg_lo = OptoReg::add(ureg,-l);
 593               if (!value[ureg_lo] &&
 594                   (!RegMask::can_represent(ureg_lo) ||
 595                    lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent
 596                 value.map(ureg_lo,valdef); // record improved reaching-def info
 597                 regnd.map(ureg_lo,   def);
 598               }
 599             }
 600           }
 601         }
 602       }
 603 
 604       const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0;
 605 
 606       // Remove copies along input edges
 607       for (k = 1; k < n->req(); k++) {
 608         j -= elide_copy(n, k, block, value, regnd, two_adr != k);
 609       }
 610 
 611       // Unallocated Nodes define no registers
 612       uint lidx = _lrg_map.live_range_id(n);
 613       if (!lidx) {
 614         continue;
 615       }
 616 
 617       // Update the register defined by this instruction
 618       OptoReg::Name nreg = lrgs(lidx).reg();
 619       // Skip through all copies to the _value_ being defined.
 620       // Do not change from int to pointer
 621       Node *val = skip_copies(n);
 622 
 623       // Clear out a dead definition before starting so that the
 624       // elimination code doesn't have to guard against it.  The
 625       // definition could in fact be a kill projection with a count of
 626       // 0 which is safe but since those are uninteresting for copy
 627       // elimination just delete them as well.
 628       if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) {
 629         regnd.map(nreg, NULL);
 630         value.map(nreg, NULL);
 631       }
 632 
 633       uint n_ideal_reg = n->ideal_reg();
 634       int n_regs = RegMask::num_registers(n_ideal_reg);
 635       if (n_regs == 1) {
 636         // If Node 'n' does not change the value mapped by the register,
 637         // then 'n' is a useless copy.  Do not update the register->node
 638         // mapping so 'n' will go dead.
 639         if( value[nreg] != val ) {
 640           if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, OptoReg::Bad)) {
 641             j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
 642           } else {
 643             // Update the mapping: record new Node defined by the register
 644             regnd.map(nreg,n);
 645             // Update mapping for defined *value*, which is the defined
 646             // Node after skipping all copies.
 647             value.map(nreg,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, block, value, regnd);
 652         }
 653       } else if (RegMask::is_vector(n_ideal_reg)) {
 654         // If Node 'n' does not change the value mapped by the register,
 655         // then 'n' is a useless copy.  Do not update the register->node
 656         // mapping so 'n' will go dead.
 657         if (!register_contains_value(val, nreg, n_regs, value)) {
 658           // Update the mapping: record new Node defined by the register
 659           regnd.map(nreg,n);
 660           // Update mapping for defined *value*, which is the defined
 661           // Node after skipping all copies.
 662           value.map(nreg,val);
 663           for (int l = 1; l < n_regs; l++) {
 664             OptoReg::Name nreg_lo = OptoReg::add(nreg,-l);
 665             regnd.map(nreg_lo, n );
 666             value.map(nreg_lo,val);
 667           }
 668         } else if (n->is_Copy()) {
 669           // Note: vector can't be constant and can't be copy of calee.
 670           j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
 671         }
 672       } else {
 673         // If the value occupies a register pair, record same info
 674         // in both registers.
 675         OptoReg::Name nreg_lo = OptoReg::add(nreg,-1);
 676         if( RegMask::can_represent(nreg_lo) &&     // Either a spill slot, or
 677             !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent
 678           // Sparc occasionally has non-adjacent pairs.
 679           // Find the actual other value
 680           RegMask tmp = lrgs(lidx).mask();
 681           tmp.Remove(nreg);
 682           nreg_lo = tmp.find_first_elem();
 683         }
 684         if (value[nreg] != val || value[nreg_lo] != val) {
 685           if (eliminate_copy_of_constant(val, n, block, value, regnd, nreg, nreg_lo)) {
 686             j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
 687           } else {
 688             regnd.map(nreg   , n );
 689             regnd.map(nreg_lo, n );
 690             value.map(nreg   ,val);
 691             value.map(nreg_lo,val);
 692           }
 693         } else if (!may_be_copy_of_callee(n)) {
 694           assert(n->is_Copy(), "");
 695           j -= replace_and_yank_if_dead(n, nreg, block, value, regnd);
 696         }
 697       }
 698 
 699       // Fat projections kill many registers
 700       if( n_ideal_reg == MachProjNode::fat_proj ) {
 701         RegMask rm = n->out_RegMask();
 702         // wow, what an expensive iterator...
 703         nreg = rm.find_first_elem();
 704         while( OptoReg::is_valid(nreg)) {
 705           rm.Remove(nreg);
 706           value.map(nreg,n);
 707           regnd.map(nreg,n);
 708           nreg = rm.find_first_elem();
 709         }
 710       }
 711 
 712     } // End of for all instructions in the block
 713 
 714   } // End for all blocks
 715 }