src/share/vm/opto/postaloc.cpp
Index Unified diffs Context diffs Sdiffs Wdiffs Patch New Old Previous File Next File 7119644 Sdiff src/share/vm/opto

src/share/vm/opto/postaloc.cpp

Print this page


   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() &&


 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 &regnd ) {
 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


 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, &regnd);
 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 


 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,&regnd);
 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);


   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() &&


 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


 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 


 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);


src/share/vm/opto/postaloc.cpp
Index Unified diffs Context diffs Sdiffs Wdiffs Patch New Old Previous File Next File