1 /*

     2  * Copyright 1999-2006 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,

    20  * CA 95054 USA or visit www.sun.com if you need additional information or

    21  * have any questions.

    22  *

    23  */

    24 

    25 #include "incls/_precompiled.incl"

    26 #include "incls/_split_if.cpp.incl"

    27 

    28 

    29 //------------------------------split_thru_region------------------------------

    30 // Split Node 'n' through merge point.

    31 Node *PhaseIdealLoop::split_thru_region( Node *n, Node *region ) {

    32   uint wins = 0;

    33   assert( n->is_CFG(), "" );

    34   assert( region->is_Region(), "" );

    35   Node *r = new (C, region->req()) RegionNode( region->req() );

    36   IdealLoopTree *loop = get_loop( n );

    37   for( uint i = 1; i < region->req(); i++ ) {

    38     Node *x = n->clone();

    39     Node *in0 = n->in(0);

    40     if( in0->in(0) == region ) x->set_req( 0, in0->in(i) );

    41     for( uint j = 1; j < n->req(); j++ ) {

    42       Node *in = n->in(j);

    43       if( get_ctrl(in) == region )

    44         x->set_req( j, in->in(i) );

    45     }

    46     _igvn.register_new_node_with_optimizer(x);

    47     set_loop(x, loop);

    48     set_idom(x, x->in(0), dom_depth(x->in(0))+1);

    49     r->init_req(i, x);

    50   }

    51 

    52   // Record region

    53   r->set_req(0,region);         // Not a TRUE RegionNode

    54   _igvn.register_new_node_with_optimizer(r);

    55   set_loop(r, loop);

    56   if( !loop->_child )

    57     loop->_body.push(r);

    58   return r;

    59 }

    60 

    61 //------------------------------split_up---------------------------------------

    62 // Split block-local op up through the phis to empty the current block

    63 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) {

    64   if( n->is_CFG() ) {

    65     assert( n->in(0) != blk1, "Lousy candidate for split-if" );

    66     return false;

    67   }

    68   if( get_ctrl(n) != blk1 && get_ctrl(n) != blk2 )

    69     return false;               // Not block local

    70   if( n->is_Phi() ) return false; // Local PHIs are expected

    71 

    72   // Recursively split-up inputs

    73   for (uint i = 1; i < n->req(); i++) {

    74     if( split_up( n->in(i), blk1, blk2 ) ) {

    75       // Got split recursively and self went dead?

    76       if (n->outcnt() == 0)

    77         _igvn.remove_dead_node(n);

    78       return true;

    79     }

    80   }

    81 

    82   // Check for needing to clone-up a compare.  Can't do that, it forces

    83   // another (nested) split-if transform.  Instead, clone it "down".

    84   if( n->is_Cmp() ) {

    85     assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF");

    86     // Check for simple Cmp/Bool/CMove which we can clone-up.  Cmp/Bool/CMove

    87     // sequence can have no other users and it must all reside in the split-if

    88     // block.  Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below -

    89     // private, per-use versions of the Cmp and Bool are made.  These sink to

    90     // the CMove block.  If the CMove is in the split-if block, then in the

    91     // next iteration this will become a simple Cmp/Bool/CMove set to clone-up.

    92     Node *bol, *cmov;

    93     if( !(n->outcnt() == 1 && n->unique_out()->is_Bool() &&

    94           (bol = n->unique_out()->as_Bool()) &&

    95           (get_ctrl(bol) == blk1 ||

    96            get_ctrl(bol) == blk2) &&

    97           bol->outcnt() == 1 &&

    98           bol->unique_out()->is_CMove() &&

    99           (cmov = bol->unique_out()->as_CMove()) &&

   100           (get_ctrl(cmov) == blk1 ||

   101            get_ctrl(cmov) == blk2) ) ) {

   102 

   103       // Must clone down

   104 #ifndef PRODUCT

   105       if( PrintOpto && VerifyLoopOptimizations ) {

   106         tty->print("Cloning down: ");

   107         n->dump();

   108       }

   109 #endif

   110       // Clone down any block-local BoolNode uses of this CmpNode

   111       for (DUIterator i = n->outs(); n->has_out(i); i++) {

   112         Node* bol = n->out(i);

   113         assert( bol->is_Bool(), "" );

   114         if (bol->outcnt() == 1) {

   115           Node* use = bol->unique_out();

   116           Node *use_c = use->is_If() ? use->in(0) : get_ctrl(use);

   117           if (use_c == blk1 || use_c == blk2) {

   118             continue;

   119           }

   120         }

   121         if (get_ctrl(bol) == blk1 || get_ctrl(bol) == blk2) {

   122           // Recursively sink any BoolNode

   123 #ifndef PRODUCT

   124           if( PrintOpto && VerifyLoopOptimizations ) {

   125             tty->print("Cloning down: ");

   126             bol->dump();

   127           }

   128 #endif

   129           for (DUIterator_Last jmin, j = bol->last_outs(jmin); j >= jmin; --j) {

   130             // Uses are either IfNodes or CMoves

   131             Node* iff = bol->last_out(j);

   132             assert( iff->in(1) == bol, "" );

   133             // Get control block of either the CMove or the If input

   134             Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff);

   135             Node *x = bol->clone();

   136             register_new_node(x, iff_ctrl);

   137             _igvn.hash_delete(iff);

   138             iff->set_req(1, x);

   139             _igvn._worklist.push(iff);

   140           }

   141           _igvn.remove_dead_node( bol );

   142           --i;

   143         }

   144       }

   145       // Clone down this CmpNode

   146       for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) {

   147         Node* bol = n->last_out(j);

   148         assert( bol->in(1) == n, "" );

   149         Node *x = n->clone();

   150         register_new_node(x, get_ctrl(bol));

   151         _igvn.hash_delete(bol);

   152         bol->set_req(1, x);

   153         _igvn._worklist.push(bol);

   154       }

   155       _igvn.remove_dead_node( n );

   156 

   157       return true;

   158     }

   159   }

   160 

   161   // See if splitting-up a Store.  Any anti-dep loads must go up as

   162   // well.  An anti-dep load might be in the wrong block, because in

   163   // this particular layout/schedule we ignored anti-deps and allow

   164   // memory to be alive twice.  This only works if we do the same

   165   // operations on anti-dep loads as we do their killing stores.

   166   if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) {

   167     // Get store's memory slice

   168     int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr());

   169 

   170     // Get memory-phi anti-dep loads will be using

   171     Node *memphi = n->in(MemNode::Memory);

   172     assert( memphi->is_Phi(), "" );

   173     // Hoist any anti-dep load to the splitting block;

   174     // it will then "split-up".

   175     for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) {

   176       Node *load = memphi->fast_out(i);

   177       if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) )

   178         set_ctrl(load,blk1);

   179     }

   180   }

   181 

   182   // Found some other Node; must clone it up

   183 #ifndef PRODUCT

   184   if( PrintOpto && VerifyLoopOptimizations ) {

   185     tty->print("Cloning up: ");

   186     n->dump();

   187   }

   188 #endif

   189 

   190   // Now actually split-up this guy.  One copy per control path merging.

   191   Node *phi = PhiNode::make_blank(blk1, n);

   192   for( uint j = 1; j < blk1->req(); j++ ) {

   193     Node *x = n->clone();

   194     if( n->in(0) && n->in(0) == blk1 )

   195       x->set_req( 0, blk1->in(j) );

   196     for( uint i = 1; i < n->req(); i++ ) {

   197       Node *m = n->in(i);

   198       if( get_ctrl(m) == blk1 ) {

   199         assert( m->in(0) == blk1, "" );

   200         x->set_req( i, m->in(j) );

   201       }

   202     }

   203     register_new_node( x, blk1->in(j) );

   204     phi->init_req( j, x );

   205   }

   206   // Announce phi to optimizer

   207   register_new_node(phi, blk1);

   208 

   209   // Remove cloned-up value from optimizer; use phi instead

   210   _igvn.hash_delete(n);

   211   _igvn.subsume_node( n, phi );

   212 

   213   // (There used to be a self-recursive call to split_up() here,

   214   // but it is not needed.  All necessary forward walking is done

   215   // by do_split_if() below.)

   216 

   217   return true;

   218 }

   219 

   220 //------------------------------register_new_node------------------------------

   221 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {

   222   _igvn.register_new_node_with_optimizer(n);

   223   set_ctrl(n, blk);

   224   IdealLoopTree *loop = get_loop(blk);

   225   if( !loop->_child )

   226     loop->_body.push(n);

   227 }

   228 

   229 //------------------------------small_cache------------------------------------

   230 struct small_cache : public Dict {

   231 

   232   small_cache() : Dict( cmpkey, hashptr ) {}

   233   Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); }

   234   void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); }

   235 };

   236 

   237 //------------------------------spinup-----------------------------------------

   238 // "Spin up" the dominator tree, starting at the use site and stopping when we

   239 // find the post-dominating point.

   240 

   241 // We must be at the merge point which post-dominates 'new_false' and

   242 // 'new_true'.  Figure out which edges into the RegionNode eventually lead up

   243 // to false and which to true.  Put in a PhiNode to merge values; plug in

   244 // the appropriate false-arm or true-arm values.  If some path leads to the

   245 // original IF, then insert a Phi recursively.

   246 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) {

   247   if (use_blk->is_top())        // Handle dead uses

   248     return use_blk;

   249   Node *prior_n = (Node*)0xdeadbeef;

   250   Node *n = use_blk;            // Get path input

   251   assert( use_blk != iff_dom, "" );

   252   // Here's the "spinup" the dominator tree loop.  Do a cache-check

   253   // along the way, in case we've come this way before.

   254   while( n != iff_dom ) {       // Found post-dominating point?

   255     prior_n = n;

   256     n = idom(n);                // Search higher

   257     Node *s = cache->probe( prior_n ); // Check cache

   258     if( s ) return s;           // Cache hit!

   259   }

   260 

   261   Node *phi_post;

   262   if( prior_n == new_false || prior_n == new_true ) {

   263     phi_post = def->clone();

   264     phi_post->set_req(0, prior_n );

   265     register_new_node(phi_post, prior_n);

   266   } else {

   267     // This method handles both control uses (looking for Regions) or data

   268     // uses (looking for Phis).  If looking for a control use, then we need

   269     // to insert a Region instead of a Phi; however Regions always exist

   270     // previously (the hash_find_insert below would always hit) so we can

   271     // return the existing Region.

   272     if( def->is_CFG() ) {

   273       phi_post = prior_n;       // If looking for CFG, return prior

   274     } else {

   275       assert( def->is_Phi(), "" );

   276       assert( prior_n->is_Region(), "must be a post-dominating merge point" );

   277 

   278       // Need a Phi here

   279       phi_post = PhiNode::make_blank(prior_n, def);

   280       // Search for both true and false on all paths till find one.

   281       for( uint i = 1; i < phi_post->req(); i++ ) // For all paths

   282         phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) );

   283       Node *t = _igvn.hash_find_insert(phi_post);

   284       if( t ) {                 // See if we already have this one

   285         // phi_post will not be used, so kill it

   286         _igvn.remove_dead_node(phi_post);

   287         phi_post->destruct();

   288         phi_post = t;

   289       } else {

   290         register_new_node( phi_post, prior_n );

   291       }

   292     }

   293   }

   294 

   295   // Update cache everywhere

   296   prior_n = (Node*)0xdeadbeef;  // Reset IDOM walk

   297   n = use_blk;                  // Get path input

   298   // Spin-up the idom tree again, basically doing path-compression.

   299   // Insert cache entries along the way, so that if we ever hit this

   300   // point in the IDOM tree again we'll stop immediately on a cache hit.

   301   while( n != iff_dom ) {       // Found post-dominating point?

   302     prior_n = n;

   303     n = idom(n);                // Search higher

   304     cache->lru_insert( prior_n, phi_post ); // Fill cache

   305   } // End of while not gone high enough

   306 

   307   return phi_post;

   308 }

   309 

   310 //------------------------------find_use_block---------------------------------

   311 // Find the block a USE is in.  Normally USE's are in the same block as the

   312 // using instruction.  For Phi-USE's, the USE is in the predecessor block

   313 // along the corresponding path.

   314 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) {

   315   // CFG uses are their own block

   316   if( use->is_CFG() )

   317     return use;

   318 

   319   if( use->is_Phi() ) {         // Phi uses in prior block

   320     // Grab the first Phi use; there may be many.

   321     // Each will be handled as a seperate iteration of

   322     // the "while( phi->outcnt() )" loop.

   323     uint j;

   324     for( j = 1; j < use->req(); j++ )

   325       if( use->in(j) == def )

   326         break;

   327     assert( j < use->req(), "def should be among use's inputs" );

   328     return use->in(0)->in(j);

   329   }

   330   // Normal (non-phi) use

   331   Node *use_blk = get_ctrl(use);

   332   // Some uses are directly attached to the old (and going away)

   333   // false and true branches.

   334   if( use_blk == old_false ) {

   335     use_blk = new_false;

   336     set_ctrl(use, new_false);

   337   }

   338   if( use_blk == old_true ) {

   339     use_blk = new_true;

   340     set_ctrl(use, new_true);

   341   }

   342 

   343   if (use_blk == NULL) {        // He's dead, Jim

   344     _igvn.hash_delete(use);

   345     _igvn.subsume_node(use, C->top());

   346   }

   347 

   348   return use_blk;

   349 }

   350 

   351 //------------------------------handle_use-------------------------------------

   352 // Handle uses of the merge point.  Basically, split-if makes the merge point

   353 // go away so all uses of the merge point must go away as well.  Most block

   354 // local uses have already been split-up, through the merge point.  Uses from

   355 // far below the merge point can't always be split up (e.g., phi-uses are

   356 // pinned) and it makes too much stuff live.  Instead we use a path-based

   357 // solution to move uses down.

   358 //

   359 // If the use is along the pre-split-CFG true branch, then the new use will

   360 // be from the post-split-CFG true merge point.  Vice-versa for the false

   361 // path.  Some uses will be along both paths; then we sink the use to the

   362 // post-dominating location; we may need to insert a Phi there.

   363 void PhaseIdealLoop::handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ) {

   364 

   365   Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true);

   366   if( !use_blk ) return;        // He's dead, Jim

   367 

   368   // Walk up the dominator tree until I hit either the old IfFalse, the old

   369   // IfTrue or the old If.  Insert Phis where needed.

   370   Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache );

   371 

   372   // Found where this USE goes.  Re-point him.

   373   uint i;

   374   for( i = 0; i < use->req(); i++ )

   375     if( use->in(i) == def )

   376       break;

   377   assert( i < use->req(), "def should be among use's inputs" );

   378   _igvn.hash_delete(use);

   379   use->set_req(i, new_def);

   380   _igvn._worklist.push(use);

   381 }

   382 

   383 //------------------------------do_split_if------------------------------------

   384 // Found an If getting its condition-code input from a Phi in the same block.

   385 // Split thru the Region.

   386 void PhaseIdealLoop::do_split_if( Node *iff ) {

   387 #ifndef PRODUCT

   388   if( PrintOpto && VerifyLoopOptimizations )

   389     tty->print_cr("Split-if");

   390 #endif

   391   C->set_major_progress();

   392   Node *region = iff->in(0);

   393   Node *region_dom = idom(region);

   394 

   395   // We are going to clone this test (and the control flow with it) up through

   396   // the incoming merge point.  We need to empty the current basic block.

   397   // Clone any instructions which must be in this block up through the merge

   398   // point.

   399   DUIterator i, j;

   400   bool progress = true;

   401   while (progress) {

   402     progress = false;

   403     for (i = region->outs(); region->has_out(i); i++) {

   404       Node* n = region->out(i);

   405       if( n == region ) continue;

   406       // The IF to be split is OK.

   407       if( n == iff ) continue;

   408       if( !n->is_Phi() ) {      // Found pinned memory op or such

   409         if (split_up(n, region, iff)) {

   410           i = region->refresh_out_pos(i);

   411           progress = true;

   412         }

   413         continue;

   414       }

   415       assert( n->in(0) == region, "" );

   416 

   417       // Recursively split up all users of a Phi

   418       for (j = n->outs(); n->has_out(j); j++) {

   419         Node* m = n->out(j);

   420         // If m is dead, throw it away, and declare progress

   421         if (_nodes[m->_idx] == NULL) {

   422           _igvn.remove_dead_node(m);

   423           // fall through

   424         }

   425         else if (m != iff && split_up(m, region, iff)) {

   426           // fall through

   427         } else {

   428           continue;

   429         }

   430         // Something unpredictable changed.

   431         // Tell the iterators to refresh themselves, and rerun the loop.

   432         i = region->refresh_out_pos(i);

   433         j = region->refresh_out_pos(j);

   434         progress = true;

   435       }

   436     }

   437   }

   438 

   439   // Now we have no instructions in the block containing the IF.

   440   // Split the IF.

   441   Node *new_iff = split_thru_region( iff, region );

   442 

   443   // Replace both uses of 'new_iff' with Regions merging True/False

   444   // paths.  This makes 'new_iff' go dead.

   445   Node *old_false, *old_true;

   446   Node *new_false, *new_true;

   447   for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) {

   448     Node *ifp = iff->last_out(j2);

   449     assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" );

   450     ifp->set_req(0, new_iff);

   451     Node *ifpx = split_thru_region( ifp, region );

   452 

   453     // Replace 'If' projection of a Region with a Region of

   454     // 'If' projections.

   455     ifpx->set_req(0, ifpx);       // A TRUE RegionNode

   456 

   457     // Setup dominator info

   458     set_idom(ifpx, region_dom, dom_depth(region_dom) + 1);

   459 

   460     // Check for splitting loop tails

   461     if( get_loop(iff)->tail() == ifp )

   462       get_loop(iff)->_tail = ifpx;

   463 

   464     // Replace in the graph with lazy-update mechanism

   465     new_iff->set_req(0, new_iff); // hook self so it does not go dead

   466     lazy_replace_proj( ifp, ifpx );

   467     new_iff->set_req(0, region);

   468 

   469     // Record bits for later xforms

   470     if( ifp->Opcode() == Op_IfFalse ) {

   471       old_false = ifp;

   472       new_false = ifpx;

   473     } else {

   474       old_true = ifp;

   475       new_true = ifpx;

   476     }

   477   }

   478   _igvn.remove_dead_node(new_iff);

   479   // Lazy replace IDOM info with the region's dominator

   480   lazy_replace( iff, region_dom );

   481 

   482   // Now make the original merge point go dead, by handling all its uses.

   483   small_cache region_cache;

   484   // Preload some control flow in region-cache

   485   region_cache.lru_insert( new_false, new_false );

   486   region_cache.lru_insert( new_true , new_true  );

   487   // Now handle all uses of the splitting block

   488   for (DUIterator_Last kmin, k = region->last_outs(kmin); k >= kmin; --k) {

   489     Node* phi = region->last_out(k);

   490     if( !phi->in(0) ) {         // Dead phi?  Remove it

   491       _igvn.remove_dead_node(phi);

   492       continue;

   493     }

   494     assert( phi->in(0) == region, "" );

   495     if( phi == region ) {       // Found the self-reference

   496       phi->set_req(0, NULL);

   497       continue;                 // Break the self-cycle

   498     }

   499     // Expected common case: Phi hanging off of Region

   500     if( phi->is_Phi() ) {

   501       // Need a per-def cache.  Phi represents a def, so make a cache

   502       small_cache phi_cache;

   503 

   504       // Inspect all Phi uses to make the Phi go dead

   505       for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) {

   506         Node* use = phi->last_out(l);

   507         // Compute the new DEF for this USE.  New DEF depends on the path

   508         // taken from the original DEF to the USE.  The new DEF may be some

   509         // collection of PHI's merging values from different paths.  The Phis

   510         // inserted depend only on the location of the USE.  We use a

   511         // 2-element cache to handle multiple uses from the same block.

   512         handle_use( use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true );

   513       } // End of while phi has uses

   514 

   515       // Because handle_use might relocate region->_out,

   516       // we must refresh the iterator.

   517       k = region->last_outs(kmin);

   518 

   519       // Remove the dead Phi

   520       _igvn.remove_dead_node( phi );

   521 

   522     } else {

   523       // Random memory op guarded by Region.  Compute new DEF for USE.

   524       handle_use( phi, region, &region_cache, region_dom, new_false, new_true, old_false, old_true );

   525     }

   526 

   527   } // End of while merge point has phis

   528 

   529   // Any leftover bits in the splitting block must not have depended on local

   530   // Phi inputs (these have already been split-up).  Hence it's safe to hoist

   531   // these guys to the dominating point.

   532   lazy_replace( region, region_dom );

   533 #ifndef PRODUCT

   534   if( VerifyLoopOptimizations ) verify();

   535 #endif

   536 }