1 /*
   2  * Copyright (c) 1998, 2012, 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(n2lidx(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._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   return blk_adjust;
  92 }
  93 
  94 #ifdef ASSERT
  95 static bool expected_yanked_node(Node *old, Node *orig_old) {
  96   // This code is expected only next original nodes:
  97   // - load from constant table node which may have next data input nodes:
  98   //     MachConstantBase, Phi, MachTemp, MachSpillCopy
  99   // - load constant node which may have next data input nodes:
 100   //     MachTemp, MachSpillCopy
 101   // - MachSpillCopy
 102   // - MachProj and Copy dead nodes
 103   if (old->is_MachSpillCopy()) {
 104     return true;
 105   } else if (old->is_Con()) {
 106     return true;
 107   } else if (old->is_MachProj()) { // Dead kills projection of Con node
 108     return (old == orig_old);
 109   } else if (old->is_Copy()) {     // Dead copy of a callee-save value
 110     return (old == orig_old);
 111   } else if (old->is_MachTemp()) {
 112     return orig_old->is_Con();
 113   } else if (old->is_Phi() || old->is_MachConstantBase()) {
 114     return (orig_old->is_Con() && orig_old->is_MachConstant());
 115   }
 116   return false;
 117 }
 118 #endif
 119 
 120 //------------------------------yank_if_dead-----------------------------------
 121 // Removed edges from 'old'.  Yank if dead.  Return adjustment counts to
 122 // iterators in the current block.
 123 int PhaseChaitin::yank_if_dead_recurse(Node *old, Node *orig_old, Block *current_block,
 124                                        Node_List *value, Node_List *regnd) {
 125   int blk_adjust=0;
 126   if (old->outcnt() == 0 && old != C->top()) {
 127 #ifdef ASSERT
 128     if (!expected_yanked_node(old, orig_old)) {
 129       tty->print_cr("==============================================");
 130       tty->print_cr("orig_old:");
 131       orig_old->dump();
 132       tty->print_cr("old:");
 133       old->dump();
 134       assert(false, "unexpected yanked node");
 135     }
 136     if (old->is_Con())
 137       orig_old = old; // Reset to satisfy expected nodes checks.
 138 #endif
 139     blk_adjust += yank(old, current_block, value, regnd);
 140 
 141     for (uint i = 1; i < old->req(); i++) {
 142       Node* n = old->in(i);
 143       if (n != NULL) {
 144         old->set_req(i, NULL);
 145         blk_adjust += yank_if_dead_recurse(n, orig_old, current_block, value, regnd);
 146       }
 147     }
 148     // Disconnect control and remove precedence edges if any exist
 149     old->disconnect_inputs(NULL);
 150   }
 151   return blk_adjust;
 152 }
 153 
 154 //------------------------------use_prior_register-----------------------------
 155 // Use the prior value instead of the current value, in an effort to make
 156 // the current value go dead.  Return block iterator adjustment, in case
 157 // we yank some instructions from this block.
 158 int PhaseChaitin::use_prior_register( Node *n, uint idx, Node *def, Block *current_block, Node_List &value, Node_List &regnd ) {
 159   // No effect?
 160   if( def == n->in(idx) ) return 0;
 161   // Def is currently dead and can be removed?  Do not resurrect
 162   if( def->outcnt() == 0 ) return 0;
 163 
 164   // Not every pair of physical registers are assignment compatible,
 165   // e.g. on sparc floating point registers are not assignable to integer
 166   // registers.
 167   const LRG &def_lrg = lrgs(n2lidx(def));
 168   OptoReg::Name def_reg = def_lrg.reg();
 169   const RegMask &use_mask = n->in_RegMask(idx);
 170   bool can_use = ( RegMask::can_represent(def_reg) ? (use_mask.Member(def_reg) != 0)
 171                                                    : (use_mask.is_AllStack() != 0));
 172   if (!RegMask::is_vector(def->ideal_reg())) {
 173     // Check for a copy to or from a misaligned pair.
 174     // It is workaround for a sparc with misaligned pairs.
 175     can_use = can_use && !use_mask.is_misaligned_pair() && !def_lrg.mask().is_misaligned_pair();
 176   }
 177   if (!can_use)
 178     return 0;
 179 
 180   // Capture the old def in case it goes dead...
 181   Node *old = n->in(idx);
 182 
 183   // Save-on-call copies can only be elided if the entire copy chain can go
 184   // away, lest we get the same callee-save value alive in 2 locations at
 185   // once.  We check for the obvious trivial case here.  Although it can
 186   // sometimes be elided with cooperation outside our scope, here we will just
 187   // miss the opportunity.  :-(
 188   if( may_be_copy_of_callee(def) ) {
 189     if( old->outcnt() > 1 ) return 0; // We're the not last user
 190     int idx = old->is_Copy();
 191     assert( idx, "chain of copies being removed" );
 192     Node *old2 = old->in(idx);  // Chain of copies
 193     if( old2->outcnt() > 1 ) return 0; // old is not the last user
 194     int idx2 = old2->is_Copy();
 195     if( !idx2 ) return 0;       // Not a chain of 2 copies
 196     if( def != old2->in(idx2) ) return 0; // Chain of exactly 2 copies
 197   }
 198 
 199   // Use the new def
 200   n->set_req(idx,def);
 201   _post_alloc++;
 202 
 203   // Is old def now dead?  We successfully yanked a copy?
 204   return yank_if_dead(old,current_block,&value,&regnd);
 205 }
 206 
 207 
 208 //------------------------------skip_copies------------------------------------
 209 // Skip through any number of copies (that don't mod oop-i-ness)
 210 Node *PhaseChaitin::skip_copies( Node *c ) {
 211   int idx = c->is_Copy();
 212   uint is_oop = lrgs(n2lidx(c))._is_oop;
 213   while (idx != 0) {
 214     guarantee(c->in(idx) != NULL, "must not resurrect dead copy");
 215     if (lrgs(n2lidx(c->in(idx)))._is_oop != is_oop)
 216       break;  // casting copy, not the same value
 217     c = c->in(idx);
 218     idx = c->is_Copy();
 219   }
 220   return c;
 221 }
 222 
 223 //------------------------------elide_copy-------------------------------------
 224 // Remove (bypass) copies along Node n, edge k.
 225 int PhaseChaitin::elide_copy( Node *n, int k, Block *current_block, Node_List &value, Node_List &regnd, bool can_change_regs ) {
 226   int blk_adjust = 0;
 227 
 228   uint nk_idx = n2lidx(n->in(k));
 229   OptoReg::Name nk_reg = lrgs(nk_idx ).reg();
 230 
 231   // Remove obvious same-register copies
 232   Node *x = n->in(k);
 233   int idx;
 234   while( (idx=x->is_Copy()) != 0 ) {
 235     Node *copy = x->in(idx);
 236     guarantee(copy != NULL, "must not resurrect dead copy");
 237     if( lrgs(n2lidx(copy)).reg() != nk_reg ) break;
 238     blk_adjust += use_prior_register(n,k,copy,current_block,value,regnd);
 239     if( n->in(k) != copy ) break; // Failed for some cutout?
 240     x = copy;                   // Progress, try again
 241   }
 242 
 243   // Phis and 2-address instructions cannot change registers so easily - their
 244   // outputs must match their input.
 245   if( !can_change_regs )
 246     return blk_adjust;          // Only check stupid copies!
 247 
 248   // Loop backedges won't have a value-mapping yet
 249   if( &value == NULL ) return blk_adjust;
 250 
 251   // Skip through all copies to the _value_ being used.  Do not change from
 252   // int to pointer.  This attempts to jump through a chain of copies, where
 253   // intermediate copies might be illegal, i.e., value is stored down to stack
 254   // then reloaded BUT survives in a register the whole way.
 255   Node *val = skip_copies(n->in(k));
 256 
 257   if (val == x && nk_idx != 0 &&
 258       regnd[nk_reg] != NULL && regnd[nk_reg] != x &&
 259       n2lidx(x) == n2lidx(regnd[nk_reg])) {
 260     // When rematerialzing nodes and stretching lifetimes, the
 261     // allocator will reuse the original def for multidef LRG instead
 262     // of the current reaching def because it can't know it's safe to
 263     // do so.  After allocation completes if they are in the same LRG
 264     // then it should use the current reaching def instead.
 265     n->set_req(k, regnd[nk_reg]);
 266     blk_adjust += yank_if_dead(val, current_block, &value, &regnd);
 267     val = skip_copies(n->in(k));
 268   }
 269 
 270   if (val == x) return blk_adjust; // No progress?
 271 
 272   int n_regs = RegMask::num_registers(val->ideal_reg());
 273   uint val_idx = n2lidx(val);
 274   OptoReg::Name val_reg = lrgs(val_idx).reg();
 275 
 276   // See if it happens to already be in the correct register!
 277   // (either Phi's direct register, or the common case of the name
 278   // never-clobbered original-def register)
 279   if (register_contains_value(val, val_reg, n_regs, value)) {
 280     blk_adjust += use_prior_register(n,k,regnd[val_reg],current_block,value,regnd);
 281     if( n->in(k) == regnd[val_reg] ) // Success!  Quit trying
 282       return blk_adjust;
 283   }
 284 
 285   // See if we can skip the copy by changing registers.  Don't change from
 286   // using a register to using the stack unless we know we can remove a
 287   // copy-load.  Otherwise we might end up making a pile of Intel cisc-spill
 288   // ops reading from memory instead of just loading once and using the
 289   // register.
 290 
 291   // Also handle duplicate copies here.
 292   const Type *t = val->is_Con() ? val->bottom_type() : NULL;
 293 
 294   // Scan all registers to see if this value is around already
 295   for( uint reg = 0; reg < (uint)_max_reg; reg++ ) {
 296     if (reg == (uint)nk_reg) {
 297       // Found ourselves so check if there is only one user of this
 298       // copy and keep on searching for a better copy if so.
 299       bool ignore_self = true;
 300       x = n->in(k);
 301       DUIterator_Fast imax, i = x->fast_outs(imax);
 302       Node* first = x->fast_out(i); i++;
 303       while (i < imax && ignore_self) {
 304         Node* use = x->fast_out(i); i++;
 305         if (use != first) ignore_self = false;
 306       }
 307       if (ignore_self) continue;
 308     }
 309 
 310     Node *vv = value[reg];
 311     if (n_regs > 1) {             // Doubles check for aligned-adjacent pair
 312       if( (reg&1)==0 ) continue;  // Wrong half of a pair
 313       if( vv != value[reg-1] ) continue; // Not a complete pair
 314     }
 315     if( vv == val ||            // Got a direct hit?
 316         (t && vv && vv->bottom_type() == t && vv->is_Mach() &&
 317          vv->as_Mach()->rule() == val->as_Mach()->rule()) ) { // Or same constant?
 318       assert( !n->is_Phi(), "cannot change registers at a Phi so easily" );
 319       if( OptoReg::is_stack(nk_reg) || // CISC-loading from stack OR
 320           OptoReg::is_reg(reg) || // turning into a register use OR
 321           regnd[reg]->outcnt()==1 ) { // last use of a spill-load turns into a CISC use
 322         blk_adjust += use_prior_register(n,k,regnd[reg],current_block,value,regnd);
 323         if( n->in(k) == regnd[reg] ) // Success!  Quit trying
 324           return blk_adjust;
 325       } // End of if not degrading to a stack
 326     } // End of if found value in another register
 327   } // End of scan all machine registers
 328   return blk_adjust;
 329 }
 330 
 331 
 332 //
 333 // Check if nreg already contains the constant value val.  Normal copy
 334 // elimination doesn't doesn't work on constants because multiple
 335 // nodes can represent the same constant so the type and rule of the
 336 // MachNode must be checked to ensure equivalence.
 337 //
 338 bool PhaseChaitin::eliminate_copy_of_constant(Node* val, Node* n,
 339                                               Block *current_block,
 340                                               Node_List& value, Node_List& regnd,
 341                                               OptoReg::Name nreg, OptoReg::Name nreg2) {
 342   if (value[nreg] != val && val->is_Con() &&
 343       value[nreg] != NULL && value[nreg]->is_Con() &&
 344       (nreg2 == OptoReg::Bad || value[nreg] == value[nreg2]) &&
 345       value[nreg]->bottom_type() == val->bottom_type() &&
 346       value[nreg]->as_Mach()->rule() == val->as_Mach()->rule()) {
 347     // This code assumes that two MachNodes representing constants
 348     // which have the same rule and the same bottom type will produce
 349     // identical effects into a register.  This seems like it must be
 350     // objectively true unless there are hidden inputs to the nodes
 351     // but if that were to change this code would need to updated.
 352     // Since they are equivalent the second one if redundant and can
 353     // be removed.
 354     //
 355     // n will be replaced with the old value but n might have
 356     // kills projections associated with it so remove them now so that
 357     // yank_if_dead will be able to eliminate the copy once the uses
 358     // have been transferred to the old[value].
 359     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 360       Node* use = n->fast_out(i);
 361       if (use->is_Proj() && use->outcnt() == 0) {
 362         // Kill projections have no users and one input
 363         use->set_req(0, C->top());
 364         yank_if_dead(use, current_block, &value, &regnd);
 365         --i; --imax;
 366       }
 367     }
 368     _post_alloc++;
 369     return true;
 370   }
 371   return false;
 372 }
 373 
 374 
 375 //------------------------------post_allocate_copy_removal---------------------
 376 // Post-Allocation peephole copy removal.  We do this in 1 pass over the
 377 // basic blocks.  We maintain a mapping of registers to Nodes (an  array of
 378 // Nodes indexed by machine register or stack slot number).  NULL means that a
 379 // register is not mapped to any Node.  We can (want to have!) have several
 380 // registers map to the same Node.  We walk forward over the instructions
 381 // updating the mapping as we go.  At merge points we force a NULL if we have
 382 // to merge 2 different Nodes into the same register.  Phi functions will give
 383 // us a new Node if there is a proper value merging.  Since the blocks are
 384 // arranged in some RPO, we will visit all parent blocks before visiting any
 385 // successor blocks (except at loops).
 386 //
 387 // If we find a Copy we look to see if the Copy's source register is a stack
 388 // slot and that value has already been loaded into some machine register; if
 389 // so we use machine register directly.  This turns a Load into a reg-reg
 390 // Move.  We also look for reloads of identical constants.
 391 //
 392 // When we see a use from a reg-reg Copy, we will attempt to use the copy's
 393 // source directly and make the copy go dead.
 394 void PhaseChaitin::post_allocate_copy_removal() {
 395   NOT_PRODUCT( Compile::TracePhase t3("postAllocCopyRemoval", &_t_postAllocCopyRemoval, TimeCompiler); )
 396   ResourceMark rm;
 397 
 398   // Need a mapping from basic block Node_Lists.  We need a Node_List to
 399   // map from register number to value-producing Node.
 400   Node_List **blk2value = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1);
 401   memset( blk2value, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) );
 402   // Need a mapping from basic block Node_Lists.  We need a Node_List to
 403   // map from register number to register-defining Node.
 404   Node_List **blk2regnd = NEW_RESOURCE_ARRAY( Node_List *, _cfg._num_blocks+1);
 405   memset( blk2regnd, 0, sizeof(Node_List*)*(_cfg._num_blocks+1) );
 406 
 407   // We keep unused Node_Lists on a free_list to avoid wasting
 408   // memory.
 409   GrowableArray<Node_List*> free_list = GrowableArray<Node_List*>(16);
 410 
 411   // For all blocks
 412   for( uint i = 0; i < _cfg._num_blocks; i++ ) {
 413     uint j;
 414     Block *b = _cfg._blocks[i];
 415 
 416     // Count of Phis in block
 417     uint phi_dex;
 418     for( phi_dex = 1; phi_dex < b->_nodes.size(); phi_dex++ ) {
 419       Node *phi = b->_nodes[phi_dex];
 420       if( !phi->is_Phi() )
 421         break;
 422     }
 423 
 424     // If any predecessor has not been visited, we do not know the state
 425     // of registers at the start.  Check for this, while updating copies
 426     // along Phi input edges
 427     bool missing_some_inputs = false;
 428     Block *freed = NULL;
 429     for( j = 1; j < b->num_preds(); j++ ) {
 430       Block *pb = _cfg._bbs[b->pred(j)->_idx];
 431       // Remove copies along phi edges
 432       for( uint k=1; k<phi_dex; k++ )
 433         elide_copy( b->_nodes[k], j, b, *blk2value[pb->_pre_order], *blk2regnd[pb->_pre_order], false );
 434       if( blk2value[pb->_pre_order] ) { // Have a mapping on this edge?
 435         // See if this predecessor's mappings have been used by everybody
 436         // who wants them.  If so, free 'em.
 437         uint k;
 438         for( k=0; k<pb->_num_succs; k++ ) {
 439           Block *pbsucc = pb->_succs[k];
 440           if( !blk2value[pbsucc->_pre_order] && pbsucc != b )
 441             break;              // Found a future user
 442         }
 443         if( k >= pb->_num_succs ) { // No more uses, free!
 444           freed = pb;           // Record last block freed
 445           free_list.push(blk2value[pb->_pre_order]);
 446           free_list.push(blk2regnd[pb->_pre_order]);
 447         }
 448       } else {                  // This block has unvisited (loopback) inputs
 449         missing_some_inputs = true;
 450       }
 451     }
 452 
 453 
 454     // Extract Node_List mappings.  If 'freed' is non-zero, we just popped
 455     // 'freed's blocks off the list
 456     Node_List &regnd = *(free_list.is_empty() ? new Node_List() : free_list.pop());
 457     Node_List &value = *(free_list.is_empty() ? new Node_List() : free_list.pop());
 458     assert( !freed || blk2value[freed->_pre_order] == &value, "" );
 459     value.map(_max_reg,NULL);
 460     regnd.map(_max_reg,NULL);
 461     // Set mappings as OUR mappings
 462     blk2value[b->_pre_order] = &value;
 463     blk2regnd[b->_pre_order] = &regnd;
 464 
 465     // Initialize value & regnd for this block
 466     if( missing_some_inputs ) {
 467       // Some predecessor has not yet been visited; zap map to empty
 468       for( uint k = 0; k < (uint)_max_reg; k++ ) {
 469         value.map(k,NULL);
 470         regnd.map(k,NULL);
 471       }
 472     } else {
 473       if( !freed ) {            // Didn't get a freebie prior block
 474         // Must clone some data
 475         freed = _cfg._bbs[b->pred(1)->_idx];
 476         Node_List &f_value = *blk2value[freed->_pre_order];
 477         Node_List &f_regnd = *blk2regnd[freed->_pre_order];
 478         for( uint k = 0; k < (uint)_max_reg; k++ ) {
 479           value.map(k,f_value[k]);
 480           regnd.map(k,f_regnd[k]);
 481         }
 482       }
 483       // Merge all inputs together, setting to NULL any conflicts.
 484       for( j = 1; j < b->num_preds(); j++ ) {
 485         Block *pb = _cfg._bbs[b->pred(j)->_idx];
 486         if( pb == freed ) continue; // Did self already via freelist
 487         Node_List &p_regnd = *blk2regnd[pb->_pre_order];
 488         for( uint k = 0; k < (uint)_max_reg; k++ ) {
 489           if( regnd[k] != p_regnd[k] ) { // Conflict on reaching defs?
 490             value.map(k,NULL); // Then no value handy
 491             regnd.map(k,NULL);
 492           }
 493         }
 494       }
 495     }
 496 
 497     // For all Phi's
 498     for( j = 1; j < phi_dex; j++ ) {
 499       uint k;
 500       Node *phi = b->_nodes[j];
 501       uint pidx = n2lidx(phi);
 502       OptoReg::Name preg = lrgs(n2lidx(phi)).reg();
 503 
 504       // Remove copies remaining on edges.  Check for junk phi.
 505       Node *u = NULL;
 506       for( k=1; k<phi->req(); k++ ) {
 507         Node *x = phi->in(k);
 508         if( phi != x && u != x ) // Found a different input
 509           u = u ? NodeSentinel : x; // Capture unique input, or NodeSentinel for 2nd input
 510       }
 511       if( u != NodeSentinel ) {    // Junk Phi.  Remove
 512         b->_nodes.remove(j--); phi_dex--;
 513         _cfg._bbs.map(phi->_idx,NULL);
 514         phi->replace_by(u);
 515         phi->disconnect_inputs(NULL);
 516         continue;
 517       }
 518       // Note that if value[pidx] exists, then we merged no new values here
 519       // and the phi is useless.  This can happen even with the above phi
 520       // removal for complex flows.  I cannot keep the better known value here
 521       // because locally the phi appears to define a new merged value.  If I
 522       // keep the better value then a copy of the phi, being unable to use the
 523       // global flow analysis, can't "peek through" the phi to the original
 524       // reaching value and so will act like it's defining a new value.  This
 525       // can lead to situations where some uses are from the old and some from
 526       // the new values.  Not illegal by itself but throws the over-strong
 527       // assert in scheduling.
 528       if( pidx ) {
 529         value.map(preg,phi);
 530         regnd.map(preg,phi);
 531         int n_regs = RegMask::num_registers(phi->ideal_reg());
 532         for (int l = 1; l < n_regs; l++) {
 533           OptoReg::Name preg_lo = OptoReg::add(preg,-l);
 534           value.map(preg_lo,phi);
 535           regnd.map(preg_lo,phi);
 536         }
 537       }
 538     }
 539 
 540     // For all remaining instructions
 541     for( j = phi_dex; j < b->_nodes.size(); j++ ) {
 542       Node *n = b->_nodes[j];
 543 
 544       if( n->outcnt() == 0 &&   // Dead?
 545           n != C->top() &&      // (ignore TOP, it has no du info)
 546           !n->is_Proj() ) {     // fat-proj kills
 547         j -= yank_if_dead(n,b,&value,&regnd);
 548         continue;
 549       }
 550 
 551       // Improve reaching-def info.  Occasionally post-alloc's liveness gives
 552       // up (at loop backedges, because we aren't doing a full flow pass).
 553       // The presence of a live use essentially asserts that the use's def is
 554       // alive and well at the use (or else the allocator fubar'd).  Take
 555       // advantage of this info to set a reaching def for the use-reg.
 556       uint k;
 557       for( k = 1; k < n->req(); k++ ) {
 558         Node *def = n->in(k);   // n->in(k) is a USE; def is the DEF for this USE
 559         guarantee(def != NULL, "no disconnected nodes at this point");
 560         uint useidx = n2lidx(def); // useidx is the live range index for this USE
 561 
 562         if( useidx ) {
 563           OptoReg::Name ureg = lrgs(useidx).reg();
 564           if( !value[ureg] ) {
 565             int idx;            // Skip occasional useless copy
 566             while( (idx=def->is_Copy()) != 0 &&
 567                    def->in(idx) != NULL &&  // NULL should not happen
 568                    ureg == lrgs(n2lidx(def->in(idx))).reg() )
 569               def = def->in(idx);
 570             Node *valdef = skip_copies(def); // tighten up val through non-useless copies
 571             value.map(ureg,valdef); // record improved reaching-def info
 572             regnd.map(ureg,   def);
 573             // Record other half of doubles
 574             uint def_ideal_reg = def->ideal_reg();
 575             int n_regs = RegMask::num_registers(def_ideal_reg);
 576             bool is_vec = RegMask::is_vector(def_ideal_reg);
 577             for (int l = 1; l < n_regs; l++) {
 578               OptoReg::Name ureg_lo = OptoReg::add(ureg,-l);
 579               if (!value[ureg_lo] &&
 580                   (!RegMask::can_represent(ureg_lo) ||
 581                    lrgs(useidx).mask().Member(ureg_lo))) { // Nearly always adjacent
 582                 value.map(ureg_lo,valdef); // record improved reaching-def info
 583                 regnd.map(ureg_lo,   def);
 584               }
 585             }
 586           }
 587         }
 588       }
 589 
 590       const uint two_adr = n->is_Mach() ? n->as_Mach()->two_adr() : 0;
 591 
 592       // Remove copies along input edges
 593       for( k = 1; k < n->req(); k++ )
 594         j -= elide_copy( n, k, b, value, regnd, two_adr!=k );
 595 
 596       // Unallocated Nodes define no registers
 597       uint lidx = n2lidx(n);
 598       if( !lidx ) continue;
 599 
 600       // Update the register defined by this instruction
 601       OptoReg::Name nreg = lrgs(lidx).reg();
 602       // Skip through all copies to the _value_ being defined.
 603       // Do not change from int to pointer
 604       Node *val = skip_copies(n);
 605 
 606       // Clear out a dead definition before starting so that the
 607       // elimination code doesn't have to guard against it.  The
 608       // definition could in fact be a kill projection with a count of
 609       // 0 which is safe but since those are uninteresting for copy
 610       // elimination just delete them as well.
 611       if (regnd[nreg] != NULL && regnd[nreg]->outcnt() == 0) {
 612         regnd.map(nreg, NULL);
 613         value.map(nreg, NULL);
 614       }
 615 
 616       uint n_ideal_reg = n->ideal_reg();
 617       int n_regs = RegMask::num_registers(n_ideal_reg);
 618       if (n_regs == 1) {
 619         // If Node 'n' does not change the value mapped by the register,
 620         // then 'n' is a useless copy.  Do not update the register->node
 621         // mapping so 'n' will go dead.
 622         if( value[nreg] != val ) {
 623           if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, OptoReg::Bad)) {
 624             j -= replace_and_yank_if_dead(n, nreg, b, value, regnd);
 625           } else {
 626             // Update the mapping: record new Node defined by the register
 627             regnd.map(nreg,n);
 628             // Update mapping for defined *value*, which is the defined
 629             // Node after skipping all copies.
 630             value.map(nreg,val);
 631           }
 632         } else if( !may_be_copy_of_callee(n) ) {
 633           assert( n->is_Copy(), "" );
 634           j -= replace_and_yank_if_dead(n, nreg, b, value, regnd);
 635         }
 636       } else if (RegMask::is_vector(n_ideal_reg)) {
 637         // If Node 'n' does not change the value mapped by the register,
 638         // then 'n' is a useless copy.  Do not update the register->node
 639         // mapping so 'n' will go dead.
 640         if (!register_contains_value(val, nreg, n_regs, value)) {
 641           // Update the mapping: record new Node defined by the register
 642           regnd.map(nreg,n);
 643           // Update mapping for defined *value*, which is the defined
 644           // Node after skipping all copies.
 645           value.map(nreg,val);
 646           for (int l = 1; l < n_regs; l++) {
 647             OptoReg::Name nreg_lo = OptoReg::add(nreg,-l);
 648             regnd.map(nreg_lo, n );
 649             value.map(nreg_lo,val);
 650           }
 651         } else if (n->is_Copy()) {
 652           // Note: vector can't be constant and can't be copy of calee.
 653           j -= replace_and_yank_if_dead(n, nreg, b, value, regnd);
 654         }
 655       } else {
 656         // If the value occupies a register pair, record same info
 657         // in both registers.
 658         OptoReg::Name nreg_lo = OptoReg::add(nreg,-1);
 659         if( RegMask::can_represent(nreg_lo) &&     // Either a spill slot, or
 660             !lrgs(lidx).mask().Member(nreg_lo) ) { // Nearly always adjacent
 661           // Sparc occasionally has non-adjacent pairs.
 662           // Find the actual other value
 663           RegMask tmp = lrgs(lidx).mask();
 664           tmp.Remove(nreg);
 665           nreg_lo = tmp.find_first_elem();
 666         }
 667         if( value[nreg] != val || value[nreg_lo] != val ) {
 668           if (eliminate_copy_of_constant(val, n, b, value, regnd, nreg, nreg_lo)) {
 669             j -= replace_and_yank_if_dead(n, nreg, b, value, regnd);
 670           } else {
 671             regnd.map(nreg   , n );
 672             regnd.map(nreg_lo, n );
 673             value.map(nreg   ,val);
 674             value.map(nreg_lo,val);
 675           }
 676         } else if( !may_be_copy_of_callee(n) ) {
 677           assert( n->is_Copy(), "" );
 678           j -= replace_and_yank_if_dead(n, nreg, b, value, regnd);
 679         }
 680       }
 681 
 682       // Fat projections kill many registers
 683       if( n_ideal_reg == MachProjNode::fat_proj ) {
 684         RegMask rm = n->out_RegMask();
 685         // wow, what an expensive iterator...
 686         nreg = rm.find_first_elem();
 687         while( OptoReg::is_valid(nreg)) {
 688           rm.Remove(nreg);
 689           value.map(nreg,n);
 690           regnd.map(nreg,n);
 691           nreg = rm.find_first_elem();
 692         }
 693       }
 694 
 695     } // End of for all instructions in the block
 696 
 697   } // End for all blocks
 698 }