1 /*

     2  * Copyright 1998-2007 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 // Optimization - Graph Style

    26 

    27 #include "incls/_precompiled.incl"

    28 #include "incls/_lcm.cpp.incl"

    29 

    30 //------------------------------implicit_null_check----------------------------

    31 // Detect implicit-null-check opportunities.  Basically, find NULL checks

    32 // with suitable memory ops nearby.  Use the memory op to do the NULL check.

    33 // I can generate a memory op if there is not one nearby.

    34 // The proj is the control projection for the not-null case.

    35 // The val is the pointer being checked for nullness.

    36 void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) {

    37   // Assume if null check need for 0 offset then always needed

    38   // Intel solaris doesn't support any null checks yet and no

    39   // mechanism exists (yet) to set the switches at an os_cpu level

    40   if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return;

    41 

    42   // Make sure the ptr-is-null path appears to be uncommon!

    43   float f = end()->as_MachIf()->_prob;

    44   if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f;

    45   if( f > PROB_UNLIKELY_MAG(4) ) return;

    46 

    47   uint bidx = 0;                // Capture index of value into memop

    48   bool was_store;               // Memory op is a store op

    49 

    50   // Get the successor block for if the test ptr is non-null

    51   Block* not_null_block;  // this one goes with the proj

    52   Block* null_block;

    53   if (_nodes[_nodes.size()-1] == proj) {

    54     null_block     = _succs[0];

    55     not_null_block = _succs[1];

    56   } else {

    57     assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other");

    58     not_null_block = _succs[0];

    59     null_block     = _succs[1];

    60   }

    61 

    62   // Search the exception block for an uncommon trap.

    63   // (See Parse::do_if and Parse::do_ifnull for the reason

    64   // we need an uncommon trap.  Briefly, we need a way to

    65   // detect failure of this optimization, as in 6366351.)

    66   {

    67     bool found_trap = false;

    68     for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {

    69       Node* nn = null_block->_nodes[i1];

    70       if (nn->is_MachCall() &&

    71           nn->as_MachCall()->entry_point() ==

    72           SharedRuntime::uncommon_trap_blob()->instructions_begin()) {

    73         const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();

    74         if (trtype->isa_int() && trtype->is_int()->is_con()) {

    75           jint tr_con = trtype->is_int()->get_con();

    76           Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);

    77           Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);

    78           assert((int)reason < (int)BitsPerInt, "recode bit map");

    79           if (is_set_nth_bit(allowed_reasons, (int) reason)

    80               && action != Deoptimization::Action_none) {

    81             // This uncommon trap is sure to recompile, eventually.

    82             // When that happens, C->too_many_traps will prevent

    83             // this transformation from happening again.

    84             found_trap = true;

    85           }

    86         }

    87         break;

    88       }

    89     }

    90     if (!found_trap) {

    91       // We did not find an uncommon trap.

    92       return;

    93     }

    94   }

    95 

    96   // Search the successor block for a load or store who's base value is also

    97   // the tested value.  There may be several.

    98   Node_List *out = new Node_List(Thread::current()->resource_area());

    99   MachNode *best = NULL;        // Best found so far

   100   for (DUIterator i = val->outs(); val->has_out(i); i++) {

   101     Node *m = val->out(i);

   102     if( !m->is_Mach() ) continue;

   103     MachNode *mach = m->as_Mach();

   104     was_store = false;

   105     switch( mach->ideal_Opcode() ) {

   106     case Op_LoadB:

   107     case Op_LoadC:

   108     case Op_LoadD:

   109     case Op_LoadF:

   110     case Op_LoadI:

   111     case Op_LoadL:

   112     case Op_LoadP:

   113     case Op_LoadS:

   114     case Op_LoadKlass:

   115     case Op_LoadRange:

   116     case Op_LoadD_unaligned:

   117     case Op_LoadL_unaligned:

   118       break;

   119     case Op_StoreB:

   120     case Op_StoreC:

   121     case Op_StoreCM:

   122     case Op_StoreD:

   123     case Op_StoreF:

   124     case Op_StoreI:

   125     case Op_StoreL:

   126     case Op_StoreP:

   127       was_store = true;         // Memory op is a store op

   128       // Stores will have their address in slot 2 (memory in slot 1).

   129       // If the value being nul-checked is in another slot, it means we

   130       // are storing the checked value, which does NOT check the value!

   131       if( mach->in(2) != val ) continue;

   132       break;                    // Found a memory op?

   133     case Op_StrComp:

   134       // Not a legit memory op for implicit null check regardless of

   135       // embedded loads

   136       continue;

   137     default:                    // Also check for embedded loads

   138       if( !mach->needs_anti_dependence_check() )

   139         continue;               // Not an memory op; skip it

   140       break;

   141     }

   142     // check if the offset is not too high for implicit exception

   143     {

   144       intptr_t offset = 0;

   145       const TypePtr *adr_type = NULL;  // Do not need this return value here

   146       const Node* base = mach->get_base_and_disp(offset, adr_type);

   147       if (base == NULL || base == NodeSentinel) {

   148         // cannot reason about it; is probably not implicit null exception

   149       } else {

   150         const TypePtr* tptr = base->bottom_type()->is_ptr();

   151         // Give up if offset is not a compile-time constant

   152         if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )

   153           continue;

   154         offset += tptr->_offset; // correct if base is offseted

   155         if( MacroAssembler::needs_explicit_null_check(offset) )

   156           continue;             // Give up is reference is beyond 4K page size

   157       }

   158     }

   159 

   160     // Check ctrl input to see if the null-check dominates the memory op

   161     Block *cb = cfg->_bbs[mach->_idx];

   162     cb = cb->_idom;             // Always hoist at least 1 block

   163     if( !was_store ) {          // Stores can be hoisted only one block

   164       while( cb->_dom_depth > (_dom_depth + 1))

   165         cb = cb->_idom;         // Hoist loads as far as we want

   166       // The non-null-block should dominate the memory op, too. Live

   167       // range spilling will insert a spill in the non-null-block if it is

   168       // needs to spill the memory op for an implicit null check.

   169       if (cb->_dom_depth == (_dom_depth + 1)) {

   170         if (cb != not_null_block) continue;

   171         cb = cb->_idom;

   172       }

   173     }

   174     if( cb != this ) continue;

   175 

   176     // Found a memory user; see if it can be hoisted to check-block

   177     uint vidx = 0;              // Capture index of value into memop

   178     uint j;

   179     for( j = mach->req()-1; j > 0; j-- ) {

   180       if( mach->in(j) == val ) vidx = j;

   181       // Block of memory-op input

   182       Block *inb = cfg->_bbs[mach->in(j)->_idx];

   183       Block *b = this;          // Start from nul check

   184       while( b != inb && b->_dom_depth > inb->_dom_depth )

   185         b = b->_idom;           // search upwards for input

   186       // See if input dominates null check

   187       if( b != inb )

   188         break;

   189     }

   190     if( j > 0 )

   191       continue;

   192     Block *mb = cfg->_bbs[mach->_idx];

   193     // Hoisting stores requires more checks for the anti-dependence case.

   194     // Give up hoisting if we have to move the store past any load.

   195     if( was_store ) {

   196       Block *b = mb;            // Start searching here for a local load

   197       // mach use (faulting) trying to hoist

   198       // n might be blocker to hoisting

   199       while( b != this ) {

   200         uint k;

   201         for( k = 1; k < b->_nodes.size(); k++ ) {

   202           Node *n = b->_nodes[k];

   203           if( n->needs_anti_dependence_check() &&

   204               n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )

   205             break;              // Found anti-dependent load

   206         }

   207         if( k < b->_nodes.size() )

   208           break;                // Found anti-dependent load

   209         // Make sure control does not do a merge (would have to check allpaths)

   210         if( b->num_preds() != 2 ) break;

   211         b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block

   212       }

   213       if( b != this ) continue;

   214     }

   215 

   216     // Make sure this memory op is not already being used for a NullCheck

   217     Node *e = mb->end();

   218     if( e->is_MachNullCheck() && e->in(1) == mach )

   219       continue;                 // Already being used as a NULL check

   220 

   221     // Found a candidate!  Pick one with least dom depth - the highest

   222     // in the dom tree should be closest to the null check.

   223     if( !best ||

   224         cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {

   225       best = mach;

   226       bidx = vidx;

   227 

   228     }

   229   }

   230   // No candidate!

   231   if( !best ) return;

   232 

   233   // ---- Found an implicit null check

   234   extern int implicit_null_checks;

   235   implicit_null_checks++;

   236 

   237   // Hoist the memory candidate up to the end of the test block.

   238   Block *old_block = cfg->_bbs[best->_idx];

   239   old_block->find_remove(best);

   240   add_inst(best);

   241   cfg->_bbs.map(best->_idx,this);

   242 

   243   // Move the control dependence

   244   if (best->in(0) && best->in(0) == old_block->_nodes[0])

   245     best->set_req(0, _nodes[0]);

   246 

   247   // Check for flag-killing projections that also need to be hoisted

   248   // Should be DU safe because no edge updates.

   249   for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {

   250     Node* n = best->fast_out(j);

   251     if( n->Opcode() == Op_MachProj ) {

   252       cfg->_bbs[n->_idx]->find_remove(n);

   253       add_inst(n);

   254       cfg->_bbs.map(n->_idx,this);

   255     }

   256   }

   257 

   258   Compile *C = cfg->C;

   259   // proj==Op_True --> ne test; proj==Op_False --> eq test.

   260   // One of two graph shapes got matched:

   261   //   (IfTrue  (If (Bool NE (CmpP ptr NULL))))

   262   //   (IfFalse (If (Bool EQ (CmpP ptr NULL))))

   263   // NULL checks are always branch-if-eq.  If we see a IfTrue projection

   264   // then we are replacing a 'ne' test with a 'eq' NULL check test.

   265   // We need to flip the projections to keep the same semantics.

   266   if( proj->Opcode() == Op_IfTrue ) {

   267     // Swap order of projections in basic block to swap branch targets

   268     Node *tmp1 = _nodes[end_idx()+1];

   269     Node *tmp2 = _nodes[end_idx()+2];

   270     _nodes.map(end_idx()+1, tmp2);

   271     _nodes.map(end_idx()+2, tmp1);

   272     Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input

   273     tmp1->replace_by(tmp);

   274     tmp2->replace_by(tmp1);

   275     tmp->replace_by(tmp2);

   276     tmp->destruct();

   277   }

   278 

   279   // Remove the existing null check; use a new implicit null check instead.

   280   // Since schedule-local needs precise def-use info, we need to correct

   281   // it as well.

   282   Node *old_tst = proj->in(0);

   283   MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);

   284   _nodes.map(end_idx(),nul_chk);

   285   cfg->_bbs.map(nul_chk->_idx,this);

   286   // Redirect users of old_test to nul_chk

   287   for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)

   288     old_tst->last_out(i2)->set_req(0, nul_chk);

   289   // Clean-up any dead code

   290   for (uint i3 = 0; i3 < old_tst->req(); i3++)

   291     old_tst->set_req(i3, NULL);

   292 

   293   cfg->latency_from_uses(nul_chk);

   294   cfg->latency_from_uses(best);

   295 }

   296 

   297 

   298 //------------------------------select-----------------------------------------

   299 // Select a nice fellow from the worklist to schedule next. If there is only

   300 // one choice, then use it. Projections take top priority for correctness

   301 // reasons - if I see a projection, then it is next.  There are a number of

   302 // other special cases, for instructions that consume condition codes, et al.

   303 // These are chosen immediately. Some instructions are required to immediately

   304 // precede the last instruction in the block, and these are taken last. Of the

   305 // remaining cases (most), choose the instruction with the greatest latency

   306 // (that is, the most number of pseudo-cycles required to the end of the

   307 // routine). If there is a tie, choose the instruction with the most inputs.

   308 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {

   309 

   310   // If only a single entry on the stack, use it

   311   uint cnt = worklist.size();

   312   if (cnt == 1) {

   313     Node *n = worklist[0];

   314     worklist.map(0,worklist.pop());

   315     return n;

   316   }

   317 

   318   uint choice  = 0; // Bigger is most important

   319   uint latency = 0; // Bigger is scheduled first

   320   uint score   = 0; // Bigger is better

   321   uint idx;         // Index in worklist

   322 

   323   for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist

   324     // Order in worklist is used to break ties.

   325     // See caller for how this is used to delay scheduling

   326     // of induction variable increments to after the other

   327     // uses of the phi are scheduled.

   328     Node *n = worklist[i];      // Get Node on worklist

   329 

   330     int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;

   331     if( n->is_Proj() ||         // Projections always win

   332         n->Opcode()== Op_Con || // So does constant 'Top'

   333         iop == Op_CreateEx ||   // Create-exception must start block

   334         iop == Op_CheckCastPP

   335         ) {

   336       worklist.map(i,worklist.pop());

   337       return n;

   338     }

   339 

   340     // Final call in a block must be adjacent to 'catch'

   341     Node *e = end();

   342     if( e->is_Catch() && e->in(0)->in(0) == n )

   343       continue;

   344 

   345     // Memory op for an implicit null check has to be at the end of the block

   346     if( e->is_MachNullCheck() && e->in(1) == n )

   347       continue;

   348 

   349     uint n_choice  = 2;

   350 

   351     // See if this instruction is consumed by a branch. If so, then (as the

   352     // branch is the last instruction in the basic block) force it to the

   353     // end of the basic block

   354     if ( must_clone[iop] ) {

   355       // See if any use is a branch

   356       bool found_machif = false;

   357 

   358       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {

   359         Node* use = n->fast_out(j);

   360 

   361         // The use is a conditional branch, make them adjacent

   362         if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {

   363           found_machif = true;

   364           break;

   365         }

   366 

   367         // More than this instruction pending for successor to be ready,

   368         // don't choose this if other opportunities are ready

   369         if (ready_cnt[use->_idx] > 1)

   370           n_choice = 1;

   371       }

   372 

   373       // loop terminated, prefer not to use this instruction

   374       if (found_machif)

   375         continue;

   376     }

   377 

   378     // See if this has a predecessor that is "must_clone", i.e. sets the

   379     // condition code. If so, choose this first

   380     for (uint j = 0; j < n->req() ; j++) {

   381       Node *inn = n->in(j);

   382       if (inn) {

   383         if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {

   384           n_choice = 3;

   385           break;

   386         }

   387       }

   388     }

   389 

   390     // MachTemps should be scheduled last so they are near their uses

   391     if (n->is_MachTemp()) {

   392       n_choice = 1;

   393     }

   394 

   395     uint n_latency = cfg->_node_latency.at_grow(n->_idx);

   396     uint n_score   = n->req();   // Many inputs get high score to break ties

   397 

   398     // Keep best latency found

   399     if( choice < n_choice ||

   400         ( choice == n_choice &&

   401           ( latency < n_latency ||

   402             ( latency == n_latency &&

   403               ( score < n_score ))))) {

   404       choice  = n_choice;

   405       latency = n_latency;

   406       score   = n_score;

   407       idx     = i;               // Also keep index in worklist

   408     }

   409   } // End of for all ready nodes in worklist

   410 

   411   Node *n = worklist[idx];      // Get the winner

   412 

   413   worklist.map(idx,worklist.pop());     // Compress worklist

   414   return n;

   415 }

   416 

   417 

   418 //------------------------------set_next_call----------------------------------

   419 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {

   420   if( next_call.test_set(n->_idx) ) return;

   421   for( uint i=0; i<n->len(); i++ ) {

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

   423     if( !m ) continue;  // must see all nodes in block that precede call

   424     if( bbs[m->_idx] == this )

   425       set_next_call( m, next_call, bbs );

   426   }

   427 }

   428 

   429 //------------------------------needed_for_next_call---------------------------

   430 // Set the flag 'next_call' for each Node that is needed for the next call to

   431 // be scheduled.  This flag lets me bias scheduling so Nodes needed for the

   432 // next subroutine call get priority - basically it moves things NOT needed

   433 // for the next call till after the call.  This prevents me from trying to

   434 // carry lots of stuff live across a call.

   435 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {

   436   // Find the next control-defining Node in this block

   437   Node* call = NULL;

   438   for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {

   439     Node* m = this_call->fast_out(i);

   440     if( bbs[m->_idx] == this && // Local-block user

   441         m != this_call &&       // Not self-start node

   442         m->is_Call() )

   443       call = m;

   444       break;

   445   }

   446   if (call == NULL)  return;    // No next call (e.g., block end is near)

   447   // Set next-call for all inputs to this call

   448   set_next_call(call, next_call, bbs);

   449 }

   450 

   451 //------------------------------sched_call-------------------------------------

   452 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {

   453   RegMask regs;

   454 

   455   // Schedule all the users of the call right now.  All the users are

   456   // projection Nodes, so they must be scheduled next to the call.

   457   // Collect all the defined registers.

   458   for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {

   459     Node* n = mcall->fast_out(i);

   460     assert( n->Opcode()==Op_MachProj, "" );

   461     --ready_cnt[n->_idx];

   462     assert( !ready_cnt[n->_idx], "" );

   463     // Schedule next to call

   464     _nodes.map(node_cnt++, n);

   465     // Collect defined registers

   466     regs.OR(n->out_RegMask());

   467     // Check for scheduling the next control-definer

   468     if( n->bottom_type() == Type::CONTROL )

   469       // Warm up next pile of heuristic bits

   470       needed_for_next_call(n, next_call, bbs);

   471 

   472     // Children of projections are now all ready

   473     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {

   474       Node* m = n->fast_out(j); // Get user

   475       if( bbs[m->_idx] != this ) continue;

   476       if( m->is_Phi() ) continue;

   477       if( !--ready_cnt[m->_idx] )

   478         worklist.push(m);

   479     }

   480 

   481   }

   482 

   483   // Act as if the call defines the Frame Pointer.

   484   // Certainly the FP is alive and well after the call.

   485   regs.Insert(matcher.c_frame_pointer());

   486 

   487   // Set all registers killed and not already defined by the call.

   488   uint r_cnt = mcall->tf()->range()->cnt();

   489   int op = mcall->ideal_Opcode();

   490   MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );

   491   bbs.map(proj->_idx,this);

   492   _nodes.insert(node_cnt++, proj);

   493 

   494   // Select the right register save policy.

   495   const char * save_policy;

   496   switch (op) {

   497     case Op_CallRuntime:

   498     case Op_CallLeaf:

   499     case Op_CallLeafNoFP:

   500       // Calling C code so use C calling convention

   501       save_policy = matcher._c_reg_save_policy;

   502       break;

   503 

   504     case Op_CallStaticJava:

   505     case Op_CallDynamicJava:

   506       // Calling Java code so use Java calling convention

   507       save_policy = matcher._register_save_policy;

   508       break;

   509 

   510     default:

   511       ShouldNotReachHere();

   512   }

   513 

   514   // When using CallRuntime mark SOE registers as killed by the call

   515   // so values that could show up in the RegisterMap aren't live in a

   516   // callee saved register since the register wouldn't know where to

   517   // find them.  CallLeaf and CallLeafNoFP are ok because they can't

   518   // have debug info on them.  Strictly speaking this only needs to be

   519   // done for oops since idealreg2debugmask takes care of debug info

   520   // references but there no way to handle oops differently than other

   521   // pointers as far as the kill mask goes.

   522   bool exclude_soe = op == Op_CallRuntime;

   523 

   524   // Fill in the kill mask for the call

   525   for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {

   526     if( !regs.Member(r) ) {     // Not already defined by the call

   527       // Save-on-call register?

   528       if ((save_policy[r] == 'C') ||

   529           (save_policy[r] == 'A') ||

   530           ((save_policy[r] == 'E') && exclude_soe)) {

   531         proj->_rout.Insert(r);

   532       }

   533     }

   534   }

   535 

   536   return node_cnt;

   537 }

   538 

   539 

   540 //------------------------------schedule_local---------------------------------

   541 // Topological sort within a block.  Someday become a real scheduler.

   542 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {

   543   // Already "sorted" are the block start Node (as the first entry), and

   544   // the block-ending Node and any trailing control projections.  We leave

   545   // these alone.  PhiNodes and ParmNodes are made to follow the block start

   546   // Node.  Everything else gets topo-sorted.

   547 

   548 #ifndef PRODUCT

   549     if (cfg->trace_opto_pipelining()) {

   550       tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);

   551       for (uint i = 0;i < _nodes.size();i++) {

   552         tty->print("# ");

   553         _nodes[i]->fast_dump();

   554       }

   555       tty->print_cr("#");

   556     }

   557 #endif

   558 

   559   // RootNode is already sorted

   560   if( _nodes.size() == 1 ) return true;

   561 

   562   // Move PhiNodes and ParmNodes from 1 to cnt up to the start

   563   uint node_cnt = end_idx();

   564   uint phi_cnt = 1;

   565   uint i;

   566   for( i = 1; i<node_cnt; i++ ) { // Scan for Phi

   567     Node *n = _nodes[i];

   568     if( n->is_Phi() ||          // Found a PhiNode or ParmNode

   569         (n->is_Proj()  && n->in(0) == head()) ) {

   570       // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt

   571       _nodes.map(i,_nodes[phi_cnt]);

   572       _nodes.map(phi_cnt++,n);  // swap Phi/Parm up front

   573     } else {                    // All others

   574       // Count block-local inputs to 'n'

   575       uint cnt = n->len();      // Input count

   576       uint local = 0;

   577       for( uint j=0; j<cnt; j++ ) {

   578         Node *m = n->in(j);

   579         if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )

   580           local++;              // One more block-local input

   581       }

   582       ready_cnt[n->_idx] = local; // Count em up

   583 

   584       // A few node types require changing a required edge to a precedence edge

   585       // before allocation.

   586       if( UseConcMarkSweepGC ) {

   587         if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {

   588           // Note: Required edges with an index greater than oper_input_base

   589           // are not supported by the allocator.

   590           // Note2: Can only depend on unmatched edge being last,

   591           // can not depend on its absolute position.

   592           Node *oop_store = n->in(n->req() - 1);

   593           n->del_req(n->req() - 1);

   594           n->add_prec(oop_store);

   595           assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");

   596         }

   597       }

   598       if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire ) {

   599         Node *x = n->in(TypeFunc::Parms);

   600         n->del_req(TypeFunc::Parms);

   601         n->add_prec(x);

   602       }

   603     }

   604   }

   605   for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count

   606     ready_cnt[_nodes[i2]->_idx] = 0;

   607 

   608   // All the prescheduled guys do not hold back internal nodes

   609   uint i3;

   610   for(i3 = 0; i3<phi_cnt; i3++ ) {  // For all pre-scheduled

   611     Node *n = _nodes[i3];       // Get pre-scheduled

   612     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {

   613       Node* m = n->fast_out(j);

   614       if( cfg->_bbs[m->_idx] ==this ) // Local-block user

   615         ready_cnt[m->_idx]--;   // Fix ready count

   616     }

   617   }

   618 

   619   Node_List delay;

   620   // Make a worklist

   621   Node_List worklist;

   622   for(uint i4=i3; i4<node_cnt; i4++ ) {    // Put ready guys on worklist

   623     Node *m = _nodes[i4];

   624     if( !ready_cnt[m->_idx] ) {   // Zero ready count?

   625       if (m->is_iteratively_computed()) {

   626         // Push induction variable increments last to allow other uses

   627         // of the phi to be scheduled first. The select() method breaks

   628         // ties in scheduling by worklist order.

   629         delay.push(m);

   630       } else {

   631         worklist.push(m);         // Then on to worklist!

   632       }

   633     }

   634   }

   635   while (delay.size()) {

   636     Node* d = delay.pop();

   637     worklist.push(d);

   638   }

   639 

   640   // Warm up the 'next_call' heuristic bits

   641   needed_for_next_call(_nodes[0], next_call, cfg->_bbs);

   642 

   643 #ifndef PRODUCT

   644     if (cfg->trace_opto_pipelining()) {

   645       for (uint j=0; j<_nodes.size(); j++) {

   646         Node     *n = _nodes[j];

   647         int     idx = n->_idx;

   648         tty->print("#   ready cnt:%3d  ", ready_cnt[idx]);

   649         tty->print("latency:%3d  ", cfg->_node_latency.at_grow(idx));

   650         tty->print("%4d: %s\n", idx, n->Name());

   651       }

   652     }

   653 #endif

   654 

   655   // Pull from worklist and schedule

   656   while( worklist.size() ) {    // Worklist is not ready

   657 

   658 #ifndef PRODUCT

   659     if (cfg->trace_opto_pipelining()) {

   660       tty->print("#   ready list:");

   661       for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist

   662         Node *n = worklist[i];      // Get Node on worklist

   663         tty->print(" %d", n->_idx);

   664       }

   665       tty->cr();

   666     }

   667 #endif

   668 

   669     // Select and pop a ready guy from worklist

   670     Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);

   671     _nodes.map(phi_cnt++,n);    // Schedule him next

   672 

   673 #ifndef PRODUCT

   674     if (cfg->trace_opto_pipelining()) {

   675       tty->print("#    select %d: %s", n->_idx, n->Name());

   676       tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx));

   677       n->dump();

   678       if (Verbose) {

   679         tty->print("#   ready list:");

   680         for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist

   681           Node *n = worklist[i];      // Get Node on worklist

   682           tty->print(" %d", n->_idx);

   683         }

   684         tty->cr();

   685       }

   686     }

   687 

   688 #endif

   689     if( n->is_MachCall() ) {

   690       MachCallNode *mcall = n->as_MachCall();

   691       phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);

   692       continue;

   693     }

   694     // Children are now all ready

   695     for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {

   696       Node* m = n->fast_out(i5); // Get user

   697       if( cfg->_bbs[m->_idx] != this ) continue;

   698       if( m->is_Phi() ) continue;

   699       if( !--ready_cnt[m->_idx] )

   700         worklist.push(m);

   701     }

   702   }

   703 

   704   if( phi_cnt != end_idx() ) {

   705     // did not schedule all.  Retry, Bailout, or Die

   706     Compile* C = matcher.C;

   707     if (C->subsume_loads() == true && !C->failing()) {

   708       // Retry with subsume_loads == false

   709       // If this is the first failure, the sentinel string will "stick"

   710       // to the Compile object, and the C2Compiler will see it and retry.

   711       C->record_failure(C2Compiler::retry_no_subsuming_loads());

   712     }

   713     // assert( phi_cnt == end_idx(), "did not schedule all" );

   714     return false;

   715   }

   716 

   717 #ifndef PRODUCT

   718   if (cfg->trace_opto_pipelining()) {

   719     tty->print_cr("#");

   720     tty->print_cr("# after schedule_local");

   721     for (uint i = 0;i < _nodes.size();i++) {

   722       tty->print("# ");

   723       _nodes[i]->fast_dump();

   724     }

   725     tty->cr();

   726   }

   727 #endif

   728 

   729 

   730   return true;

   731 }

   732 

   733 //--------------------------catch_cleanup_fix_all_inputs-----------------------

   734 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {

   735   for (uint l = 0; l < use->len(); l++) {

   736     if (use->in(l) == old_def) {

   737       if (l < use->req()) {

   738         use->set_req(l, new_def);

   739       } else {

   740         use->rm_prec(l);

   741         use->add_prec(new_def);

   742         l--;

   743       }

   744     }

   745   }

   746 }

   747 

   748 //------------------------------catch_cleanup_find_cloned_def------------------

   749 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {

   750   assert( use_blk != def_blk, "Inter-block cleanup only");

   751 

   752   // The use is some block below the Catch.  Find and return the clone of the def

   753   // that dominates the use. If there is no clone in a dominating block, then

   754   // create a phi for the def in a dominating block.

   755 

   756   // Find which successor block dominates this use.  The successor

   757   // blocks must all be single-entry (from the Catch only; I will have

   758   // split blocks to make this so), hence they all dominate.

   759   while( use_blk->_dom_depth > def_blk->_dom_depth+1 )

   760     use_blk = use_blk->_idom;

   761 

   762   // Find the successor

   763   Node *fixup = NULL;

   764 

   765   uint j;

   766   for( j = 0; j < def_blk->_num_succs; j++ )

   767     if( use_blk == def_blk->_succs[j] )

   768       break;

   769 

   770   if( j == def_blk->_num_succs ) {

   771     // Block at same level in dom-tree is not a successor.  It needs a

   772     // PhiNode, the PhiNode uses from the def and IT's uses need fixup.

   773     Node_Array inputs = new Node_List(Thread::current()->resource_area());

   774     for(uint k = 1; k < use_blk->num_preds(); k++) {

   775       inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));

   776     }

   777 

   778     // Check to see if the use_blk already has an identical phi inserted.

   779     // If it exists, it will be at the first position since all uses of a

   780     // def are processed together.

   781     Node *phi = use_blk->_nodes[1];

   782     if( phi->is_Phi() ) {

   783       fixup = phi;

   784       for (uint k = 1; k < use_blk->num_preds(); k++) {

   785         if (phi->in(k) != inputs[k]) {

   786           // Not a match

   787           fixup = NULL;

   788           break;

   789         }

   790       }

   791     }

   792 

   793     // If an existing PhiNode was not found, make a new one.

   794     if (fixup == NULL) {

   795       Node *new_phi = PhiNode::make(use_blk->head(), def);

   796       use_blk->_nodes.insert(1, new_phi);

   797       bbs.map(new_phi->_idx, use_blk);

   798       for (uint k = 1; k < use_blk->num_preds(); k++) {

   799         new_phi->set_req(k, inputs[k]);

   800       }

   801       fixup = new_phi;

   802     }

   803 

   804   } else {

   805     // Found the use just below the Catch.  Make it use the clone.

   806     fixup = use_blk->_nodes[n_clone_idx];

   807   }

   808 

   809   return fixup;

   810 }

   811 

   812 //--------------------------catch_cleanup_intra_block--------------------------

   813 // Fix all input edges in use that reference "def".  The use is in the same

   814 // block as the def and both have been cloned in each successor block.

   815 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {

   816 

   817   // Both the use and def have been cloned. For each successor block,

   818   // get the clone of the use, and make its input the clone of the def

   819   // found in that block.

   820 

   821   uint use_idx = blk->find_node(use);

   822   uint offset_idx = use_idx - beg;

   823   for( uint k = 0; k < blk->_num_succs; k++ ) {

   824     // Get clone in each successor block

   825     Block *sb = blk->_succs[k];

   826     Node *clone = sb->_nodes[offset_idx+1];

   827     assert( clone->Opcode() == use->Opcode(), "" );

   828 

   829     // Make use-clone reference the def-clone

   830     catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);

   831   }

   832 }

   833 

   834 //------------------------------catch_cleanup_inter_block---------------------

   835 // Fix all input edges in use that reference "def".  The use is in a different

   836 // block than the def.

   837 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {

   838   if( !use_blk ) return;        // Can happen if the use is a precedence edge

   839 

   840   Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);

   841   catch_cleanup_fix_all_inputs(use, def, new_def);

   842 }

   843 

   844 //------------------------------call_catch_cleanup-----------------------------

   845 // If we inserted any instructions between a Call and his CatchNode,

   846 // clone the instructions on all paths below the Catch.

   847 void Block::call_catch_cleanup(Block_Array &bbs) {

   848 

   849   // End of region to clone

   850   uint end = end_idx();

   851   if( !_nodes[end]->is_Catch() ) return;

   852   // Start of region to clone

   853   uint beg = end;

   854   while( _nodes[beg-1]->Opcode() != Op_MachProj ||

   855         !_nodes[beg-1]->in(0)->is_Call() ) {

   856     beg--;

   857     assert(beg > 0,"Catch cleanup walking beyond block boundary");

   858   }

   859   // Range of inserted instructions is [beg, end)

   860   if( beg == end ) return;

   861 

   862   // Clone along all Catch output paths.  Clone area between the 'beg' and

   863   // 'end' indices.

   864   for( uint i = 0; i < _num_succs; i++ ) {

   865     Block *sb = _succs[i];

   866     // Clone the entire area; ignoring the edge fixup for now.

   867     for( uint j = end; j > beg; j-- ) {

   868       Node *clone = _nodes[j-1]->clone();

   869       sb->_nodes.insert( 1, clone );

   870       bbs.map(clone->_idx,sb);

   871     }

   872   }

   873 

   874 

   875   // Fixup edges.  Check the def-use info per cloned Node

   876   for(uint i2 = beg; i2 < end; i2++ ) {

   877     uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block

   878     Node *n = _nodes[i2];        // Node that got cloned

   879     // Need DU safe iterator because of edge manipulation in calls.

   880     Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());

   881     for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {

   882       out->push(n->fast_out(j1));

   883     }

   884     uint max = out->size();

   885     for (uint j = 0; j < max; j++) {// For all users

   886       Node *use = out->pop();

   887       Block *buse = bbs[use->_idx];

   888       if( use->is_Phi() ) {

   889         for( uint k = 1; k < use->req(); k++ )

   890           if( use->in(k) == n ) {

   891             Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);

   892             use->set_req(k, fixup);

   893           }

   894       } else {

   895         if (this == buse) {

   896           catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);

   897         } else {

   898           catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);

   899         }

   900       }

   901     } // End for all users

   902 

   903   } // End of for all Nodes in cloned area

   904 

   905   // Remove the now-dead cloned ops

   906   for(uint i3 = beg; i3 < end; i3++ ) {

   907     _nodes[beg]->disconnect_inputs(NULL);

   908     _nodes.remove(beg);

   909   }

   910 

   911   // If the successor blocks have a CreateEx node, move it back to the top

   912   for(uint i4 = 0; i4 < _num_succs; i4++ ) {

   913     Block *sb = _succs[i4];

   914     uint new_cnt = end - beg;

   915     // Remove any newly created, but dead, nodes.

   916     for( uint j = new_cnt; j > 0; j-- ) {

   917       Node *n = sb->_nodes[j];

   918       if (n->outcnt() == 0 &&

   919           (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){

   920         n->disconnect_inputs(NULL);

   921         sb->_nodes.remove(j);

   922         new_cnt--;

   923       }

   924     }

   925     // If any newly created nodes remain, move the CreateEx node to the top

   926     if (new_cnt > 0) {

   927       Node *cex = sb->_nodes[1+new_cnt];

   928       if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {

   929         sb->_nodes.remove(1+new_cnt);

   930         sb->_nodes.insert(1,cex);

   931       }

   932     }

   933   }

   934 }