1 /*
   2  * Copyright 1998-2009 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   while (null_block->is_Empty() == Block::empty_with_goto) {
  62     null_block     = null_block->_succs[0];
  63   }
  64 
  65   // Search the exception block for an uncommon trap.
  66   // (See Parse::do_if and Parse::do_ifnull for the reason
  67   // we need an uncommon trap.  Briefly, we need a way to
  68   // detect failure of this optimization, as in 6366351.)
  69   {
  70     bool found_trap = false;
  71     for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) {
  72       Node* nn = null_block->_nodes[i1];
  73       if (nn->is_MachCall() &&
  74           nn->as_MachCall()->entry_point() ==
  75           SharedRuntime::uncommon_trap_blob()->instructions_begin()) {
  76         const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type();
  77         if (trtype->isa_int() && trtype->is_int()->is_con()) {
  78           jint tr_con = trtype->is_int()->get_con();
  79           Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con);
  80           Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con);
  81           assert((int)reason < (int)BitsPerInt, "recode bit map");
  82           if (is_set_nth_bit(allowed_reasons, (int) reason)
  83               && action != Deoptimization::Action_none) {
  84             // This uncommon trap is sure to recompile, eventually.
  85             // When that happens, C->too_many_traps will prevent
  86             // this transformation from happening again.
  87             found_trap = true;
  88           }
  89         }
  90         break;
  91       }
  92     }
  93     if (!found_trap) {
  94       // We did not find an uncommon trap.
  95       return;
  96     }
  97   }
  98 
  99   // Search the successor block for a load or store who's base value is also
 100   // the tested value.  There may be several.
 101   Node_List *out = new Node_List(Thread::current()->resource_area());
 102   MachNode *best = NULL;        // Best found so far
 103   for (DUIterator i = val->outs(); val->has_out(i); i++) {
 104     Node *m = val->out(i);
 105     if( !m->is_Mach() ) continue;
 106     MachNode *mach = m->as_Mach();
 107     was_store = false;
 108     switch( mach->ideal_Opcode() ) {
 109     case Op_LoadB:
 110     case Op_LoadUS:
 111     case Op_LoadD:
 112     case Op_LoadF:
 113     case Op_LoadI:
 114     case Op_LoadL:
 115     case Op_LoadP:
 116     case Op_LoadN:
 117     case Op_LoadS:
 118     case Op_LoadKlass:
 119     case Op_LoadNKlass:
 120     case Op_LoadRange:
 121     case Op_LoadD_unaligned:
 122     case Op_LoadL_unaligned:
 123       break;
 124     case Op_StoreB:
 125     case Op_StoreC:
 126     case Op_StoreCM:
 127     case Op_StoreD:
 128     case Op_StoreF:
 129     case Op_StoreI:
 130     case Op_StoreL:
 131     case Op_StoreP:
 132     case Op_StoreN:
 133       was_store = true;         // Memory op is a store op
 134       // Stores will have their address in slot 2 (memory in slot 1).
 135       // If the value being nul-checked is in another slot, it means we
 136       // are storing the checked value, which does NOT check the value!
 137       if( mach->in(2) != val ) continue;
 138       break;                    // Found a memory op?
 139     case Op_StrComp:
 140     case Op_StrEquals:
 141     case Op_StrIndexOf:
 142     case Op_AryEq:
 143       // Not a legit memory op for implicit null check regardless of
 144       // embedded loads
 145       continue;
 146     default:                    // Also check for embedded loads
 147       if( !mach->needs_anti_dependence_check() )
 148         continue;               // Not an memory op; skip it
 149       break;
 150     }
 151     // check if the offset is not too high for implicit exception
 152     {
 153       intptr_t offset = 0;
 154       const TypePtr *adr_type = NULL;  // Do not need this return value here
 155       const Node* base = mach->get_base_and_disp(offset, adr_type);
 156       if (base == NULL || base == NodeSentinel) {
 157         // Narrow oop address doesn't have base, only index
 158         if( val->bottom_type()->isa_narrowoop() &&
 159             MacroAssembler::needs_explicit_null_check(offset) )
 160           continue;             // Give up if offset is beyond page size
 161         // cannot reason about it; is probably not implicit null exception
 162       } else {
 163         const TypePtr* tptr;
 164         if (UseCompressedOops && Universe::narrow_oop_shift() == 0) {
 165           // 32-bits narrow oop can be the base of address expressions
 166           tptr = base->bottom_type()->make_ptr();
 167         } else {
 168           // only regular oops are expected here
 169           tptr = base->bottom_type()->is_ptr();
 170         }
 171         // Give up if offset is not a compile-time constant
 172         if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot )
 173           continue;
 174         offset += tptr->_offset; // correct if base is offseted
 175         if( MacroAssembler::needs_explicit_null_check(offset) )
 176           continue;             // Give up is reference is beyond 4K page size
 177       }
 178     }
 179 
 180     // Check ctrl input to see if the null-check dominates the memory op
 181     Block *cb = cfg->_bbs[mach->_idx];
 182     cb = cb->_idom;             // Always hoist at least 1 block
 183     if( !was_store ) {          // Stores can be hoisted only one block
 184       while( cb->_dom_depth > (_dom_depth + 1))
 185         cb = cb->_idom;         // Hoist loads as far as we want
 186       // The non-null-block should dominate the memory op, too. Live
 187       // range spilling will insert a spill in the non-null-block if it is
 188       // needs to spill the memory op for an implicit null check.
 189       if (cb->_dom_depth == (_dom_depth + 1)) {
 190         if (cb != not_null_block) continue;
 191         cb = cb->_idom;
 192       }
 193     }
 194     if( cb != this ) continue;
 195 
 196     // Found a memory user; see if it can be hoisted to check-block
 197     uint vidx = 0;              // Capture index of value into memop
 198     uint j;
 199     for( j = mach->req()-1; j > 0; j-- ) {
 200       if( mach->in(j) == val ) vidx = j;
 201       // Block of memory-op input
 202       Block *inb = cfg->_bbs[mach->in(j)->_idx];
 203       Block *b = this;          // Start from nul check
 204       while( b != inb && b->_dom_depth > inb->_dom_depth )
 205         b = b->_idom;           // search upwards for input
 206       // See if input dominates null check
 207       if( b != inb )
 208         break;
 209     }
 210     if( j > 0 )
 211       continue;
 212     Block *mb = cfg->_bbs[mach->_idx];
 213     // Hoisting stores requires more checks for the anti-dependence case.
 214     // Give up hoisting if we have to move the store past any load.
 215     if( was_store ) {
 216       Block *b = mb;            // Start searching here for a local load
 217       // mach use (faulting) trying to hoist
 218       // n might be blocker to hoisting
 219       while( b != this ) {
 220         uint k;
 221         for( k = 1; k < b->_nodes.size(); k++ ) {
 222           Node *n = b->_nodes[k];
 223           if( n->needs_anti_dependence_check() &&
 224               n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) )
 225             break;              // Found anti-dependent load
 226         }
 227         if( k < b->_nodes.size() )
 228           break;                // Found anti-dependent load
 229         // Make sure control does not do a merge (would have to check allpaths)
 230         if( b->num_preds() != 2 ) break;
 231         b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block
 232       }
 233       if( b != this ) continue;
 234     }
 235 
 236     // Make sure this memory op is not already being used for a NullCheck
 237     Node *e = mb->end();
 238     if( e->is_MachNullCheck() && e->in(1) == mach )
 239       continue;                 // Already being used as a NULL check
 240 
 241     // Found a candidate!  Pick one with least dom depth - the highest
 242     // in the dom tree should be closest to the null check.
 243     if( !best ||
 244         cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) {
 245       best = mach;
 246       bidx = vidx;
 247 
 248     }
 249   }
 250   // No candidate!
 251   if( !best ) return;
 252 
 253   // ---- Found an implicit null check
 254   extern int implicit_null_checks;
 255   implicit_null_checks++;
 256 
 257   // Hoist the memory candidate up to the end of the test block.
 258   Block *old_block = cfg->_bbs[best->_idx];
 259   old_block->find_remove(best);
 260   add_inst(best);
 261   cfg->_bbs.map(best->_idx,this);
 262 
 263   // Move the control dependence
 264   if (best->in(0) && best->in(0) == old_block->_nodes[0])
 265     best->set_req(0, _nodes[0]);
 266 
 267   // Check for flag-killing projections that also need to be hoisted
 268   // Should be DU safe because no edge updates.
 269   for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) {
 270     Node* n = best->fast_out(j);
 271     if( n->Opcode() == Op_MachProj ) {
 272       cfg->_bbs[n->_idx]->find_remove(n);
 273       add_inst(n);
 274       cfg->_bbs.map(n->_idx,this);
 275     }
 276   }
 277 
 278   Compile *C = cfg->C;
 279   // proj==Op_True --> ne test; proj==Op_False --> eq test.
 280   // One of two graph shapes got matched:
 281   //   (IfTrue  (If (Bool NE (CmpP ptr NULL))))
 282   //   (IfFalse (If (Bool EQ (CmpP ptr NULL))))
 283   // NULL checks are always branch-if-eq.  If we see a IfTrue projection
 284   // then we are replacing a 'ne' test with a 'eq' NULL check test.
 285   // We need to flip the projections to keep the same semantics.
 286   if( proj->Opcode() == Op_IfTrue ) {
 287     // Swap order of projections in basic block to swap branch targets
 288     Node *tmp1 = _nodes[end_idx()+1];
 289     Node *tmp2 = _nodes[end_idx()+2];
 290     _nodes.map(end_idx()+1, tmp2);
 291     _nodes.map(end_idx()+2, tmp1);
 292     Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input
 293     tmp1->replace_by(tmp);
 294     tmp2->replace_by(tmp1);
 295     tmp->replace_by(tmp2);
 296     tmp->destruct();
 297   }
 298 
 299   // Remove the existing null check; use a new implicit null check instead.
 300   // Since schedule-local needs precise def-use info, we need to correct
 301   // it as well.
 302   Node *old_tst = proj->in(0);
 303   MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx);
 304   _nodes.map(end_idx(),nul_chk);
 305   cfg->_bbs.map(nul_chk->_idx,this);
 306   // Redirect users of old_test to nul_chk
 307   for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2)
 308     old_tst->last_out(i2)->set_req(0, nul_chk);
 309   // Clean-up any dead code
 310   for (uint i3 = 0; i3 < old_tst->req(); i3++)
 311     old_tst->set_req(i3, NULL);
 312 
 313   cfg->latency_from_uses(nul_chk);
 314   cfg->latency_from_uses(best);
 315 }
 316 
 317 
 318 //------------------------------select-----------------------------------------
 319 // Select a nice fellow from the worklist to schedule next. If there is only
 320 // one choice, then use it. Projections take top priority for correctness
 321 // reasons - if I see a projection, then it is next.  There are a number of
 322 // other special cases, for instructions that consume condition codes, et al.
 323 // These are chosen immediately. Some instructions are required to immediately
 324 // precede the last instruction in the block, and these are taken last. Of the
 325 // remaining cases (most), choose the instruction with the greatest latency
 326 // (that is, the most number of pseudo-cycles required to the end of the
 327 // routine). If there is a tie, choose the instruction with the most inputs.
 328 Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) {
 329 
 330   // If only a single entry on the stack, use it
 331   uint cnt = worklist.size();
 332   if (cnt == 1) {
 333     Node *n = worklist[0];
 334     worklist.map(0,worklist.pop());
 335     return n;
 336   }
 337 
 338   uint choice  = 0; // Bigger is most important
 339   uint latency = 0; // Bigger is scheduled first
 340   uint score   = 0; // Bigger is better
 341   int idx = -1;     // Index in worklist
 342 
 343   for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist
 344     // Order in worklist is used to break ties.
 345     // See caller for how this is used to delay scheduling
 346     // of induction variable increments to after the other
 347     // uses of the phi are scheduled.
 348     Node *n = worklist[i];      // Get Node on worklist
 349 
 350     int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0;
 351     if( n->is_Proj() ||         // Projections always win
 352         n->Opcode()== Op_Con || // So does constant 'Top'
 353         iop == Op_CreateEx ||   // Create-exception must start block
 354         iop == Op_CheckCastPP
 355         ) {
 356       worklist.map(i,worklist.pop());
 357       return n;
 358     }
 359 
 360     // Final call in a block must be adjacent to 'catch'
 361     Node *e = end();
 362     if( e->is_Catch() && e->in(0)->in(0) == n )
 363       continue;
 364 
 365     // Memory op for an implicit null check has to be at the end of the block
 366     if( e->is_MachNullCheck() && e->in(1) == n )
 367       continue;
 368 
 369     uint n_choice  = 2;
 370 
 371     // See if this instruction is consumed by a branch. If so, then (as the
 372     // branch is the last instruction in the basic block) force it to the
 373     // end of the basic block
 374     if ( must_clone[iop] ) {
 375       // See if any use is a branch
 376       bool found_machif = false;
 377 
 378       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 379         Node* use = n->fast_out(j);
 380 
 381         // The use is a conditional branch, make them adjacent
 382         if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) {
 383           found_machif = true;
 384           break;
 385         }
 386 
 387         // More than this instruction pending for successor to be ready,
 388         // don't choose this if other opportunities are ready
 389         if (ready_cnt[use->_idx] > 1)
 390           n_choice = 1;
 391       }
 392 
 393       // loop terminated, prefer not to use this instruction
 394       if (found_machif)
 395         continue;
 396     }
 397 
 398     // See if this has a predecessor that is "must_clone", i.e. sets the
 399     // condition code. If so, choose this first
 400     for (uint j = 0; j < n->req() ; j++) {
 401       Node *inn = n->in(j);
 402       if (inn) {
 403         if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) {
 404           n_choice = 3;
 405           break;
 406         }
 407       }
 408     }
 409 
 410     // MachTemps should be scheduled last so they are near their uses
 411     if (n->is_MachTemp()) {
 412       n_choice = 1;
 413     }
 414 
 415     uint n_latency = cfg->_node_latency.at_grow(n->_idx);
 416     uint n_score   = n->req();   // Many inputs get high score to break ties
 417 
 418     // Keep best latency found
 419     if( choice < n_choice ||
 420         ( choice == n_choice &&
 421           ( latency < n_latency ||
 422             ( latency == n_latency &&
 423               ( score < n_score ))))) {
 424       choice  = n_choice;
 425       latency = n_latency;
 426       score   = n_score;
 427       idx     = i;               // Also keep index in worklist
 428     }
 429   } // End of for all ready nodes in worklist
 430 
 431   assert(idx >= 0, "index should be set");
 432   Node *n = worklist[(uint)idx];      // Get the winner
 433 
 434   worklist.map((uint)idx, worklist.pop());     // Compress worklist
 435   return n;
 436 }
 437 
 438 
 439 //------------------------------set_next_call----------------------------------
 440 void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) {
 441   if( next_call.test_set(n->_idx) ) return;
 442   for( uint i=0; i<n->len(); i++ ) {
 443     Node *m = n->in(i);
 444     if( !m ) continue;  // must see all nodes in block that precede call
 445     if( bbs[m->_idx] == this )
 446       set_next_call( m, next_call, bbs );
 447   }
 448 }
 449 
 450 //------------------------------needed_for_next_call---------------------------
 451 // Set the flag 'next_call' for each Node that is needed for the next call to
 452 // be scheduled.  This flag lets me bias scheduling so Nodes needed for the
 453 // next subroutine call get priority - basically it moves things NOT needed
 454 // for the next call till after the call.  This prevents me from trying to
 455 // carry lots of stuff live across a call.
 456 void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) {
 457   // Find the next control-defining Node in this block
 458   Node* call = NULL;
 459   for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) {
 460     Node* m = this_call->fast_out(i);
 461     if( bbs[m->_idx] == this && // Local-block user
 462         m != this_call &&       // Not self-start node
 463         m->is_Call() )
 464       call = m;
 465       break;
 466   }
 467   if (call == NULL)  return;    // No next call (e.g., block end is near)
 468   // Set next-call for all inputs to this call
 469   set_next_call(call, next_call, bbs);
 470 }
 471 
 472 //------------------------------sched_call-------------------------------------
 473 uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) {
 474   RegMask regs;
 475 
 476   // Schedule all the users of the call right now.  All the users are
 477   // projection Nodes, so they must be scheduled next to the call.
 478   // Collect all the defined registers.
 479   for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) {
 480     Node* n = mcall->fast_out(i);
 481     assert( n->Opcode()==Op_MachProj, "" );
 482     --ready_cnt[n->_idx];
 483     assert( !ready_cnt[n->_idx], "" );
 484     // Schedule next to call
 485     _nodes.map(node_cnt++, n);
 486     // Collect defined registers
 487     regs.OR(n->out_RegMask());
 488     // Check for scheduling the next control-definer
 489     if( n->bottom_type() == Type::CONTROL )
 490       // Warm up next pile of heuristic bits
 491       needed_for_next_call(n, next_call, bbs);
 492 
 493     // Children of projections are now all ready
 494     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 495       Node* m = n->fast_out(j); // Get user
 496       if( bbs[m->_idx] != this ) continue;
 497       if( m->is_Phi() ) continue;
 498       if( !--ready_cnt[m->_idx] )
 499         worklist.push(m);
 500     }
 501 
 502   }
 503 
 504   // Act as if the call defines the Frame Pointer.
 505   // Certainly the FP is alive and well after the call.
 506   regs.Insert(matcher.c_frame_pointer());
 507 
 508   // Set all registers killed and not already defined by the call.
 509   uint r_cnt = mcall->tf()->range()->cnt();
 510   int op = mcall->ideal_Opcode();
 511   MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
 512   bbs.map(proj->_idx,this);
 513   _nodes.insert(node_cnt++, proj);
 514 
 515   // Select the right register save policy.
 516   const char * save_policy;
 517   switch (op) {
 518     case Op_CallRuntime:
 519     case Op_CallLeaf:
 520     case Op_CallLeafNoFP:
 521       // Calling C code so use C calling convention
 522       save_policy = matcher._c_reg_save_policy;
 523       break;
 524 
 525     case Op_CallStaticJava:
 526     case Op_CallDynamicJava:
 527       // Calling Java code so use Java calling convention
 528       save_policy = matcher._register_save_policy;
 529       break;
 530 
 531     default:
 532       ShouldNotReachHere();
 533   }
 534 
 535   // When using CallRuntime mark SOE registers as killed by the call
 536   // so values that could show up in the RegisterMap aren't live in a
 537   // callee saved register since the register wouldn't know where to
 538   // find them.  CallLeaf and CallLeafNoFP are ok because they can't
 539   // have debug info on them.  Strictly speaking this only needs to be
 540   // done for oops since idealreg2debugmask takes care of debug info
 541   // references but there no way to handle oops differently than other
 542   // pointers as far as the kill mask goes.
 543   bool exclude_soe = op == Op_CallRuntime;
 544 
 545   proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask());
 546 
 547   // Fill in the kill mask for the call
 548   for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) {
 549     if( !regs.Member(r) ) {     // Not already defined by the call
 550       // Save-on-call register?
 551       if ((save_policy[r] == 'C') ||
 552           (save_policy[r] == 'A') ||
 553           ((save_policy[r] == 'E') && exclude_soe)) {
 554         proj->_rout.Insert(r);
 555       }
 556     }
 557   }
 558 
 559   return node_cnt;
 560 }
 561 
 562 
 563 //------------------------------schedule_local---------------------------------
 564 // Topological sort within a block.  Someday become a real scheduler.
 565 bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) {
 566   // Already "sorted" are the block start Node (as the first entry), and
 567   // the block-ending Node and any trailing control projections.  We leave
 568   // these alone.  PhiNodes and ParmNodes are made to follow the block start
 569   // Node.  Everything else gets topo-sorted.
 570 
 571 #ifndef PRODUCT
 572     if (cfg->trace_opto_pipelining()) {
 573       tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order);
 574       for (uint i = 0;i < _nodes.size();i++) {
 575         tty->print("# ");
 576         _nodes[i]->fast_dump();
 577       }
 578       tty->print_cr("#");
 579     }
 580 #endif
 581 
 582   // RootNode is already sorted
 583   if( _nodes.size() == 1 ) return true;
 584 
 585   // Move PhiNodes and ParmNodes from 1 to cnt up to the start
 586   uint node_cnt = end_idx();
 587   uint phi_cnt = 1;
 588   uint i;
 589   for( i = 1; i<node_cnt; i++ ) { // Scan for Phi
 590     Node *n = _nodes[i];
 591     if( n->is_Phi() ||          // Found a PhiNode or ParmNode
 592         (n->is_Proj()  && n->in(0) == head()) ) {
 593       // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt
 594       _nodes.map(i,_nodes[phi_cnt]);
 595       _nodes.map(phi_cnt++,n);  // swap Phi/Parm up front
 596     } else {                    // All others
 597       // Count block-local inputs to 'n'
 598       uint cnt = n->len();      // Input count
 599       uint local = 0;
 600       for( uint j=0; j<cnt; j++ ) {
 601         Node *m = n->in(j);
 602         if( m && cfg->_bbs[m->_idx] == this && !m->is_top() )
 603           local++;              // One more block-local input
 604       }
 605       ready_cnt[n->_idx] = local; // Count em up
 606 
 607       // A few node types require changing a required edge to a precedence edge
 608       // before allocation.
 609       if( UseConcMarkSweepGC || UseG1GC ) {
 610         if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) {
 611           // Note: Required edges with an index greater than oper_input_base
 612           // are not supported by the allocator.
 613           // Note2: Can only depend on unmatched edge being last,
 614           // can not depend on its absolute position.
 615           Node *oop_store = n->in(n->req() - 1);
 616           n->del_req(n->req() - 1);
 617           n->add_prec(oop_store);
 618           assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark");
 619         }
 620       }
 621       if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire &&
 622           n->req() > TypeFunc::Parms ) {
 623         // MemBarAcquire could be created without Precedent edge.
 624         // del_req() replaces the specified edge with the last input edge
 625         // and then removes the last edge. If the specified edge > number of
 626         // edges the last edge will be moved outside of the input edges array
 627         // and the edge will be lost. This is why this code should be
 628         // executed only when Precedent (== TypeFunc::Parms) edge is present.
 629         Node *x = n->in(TypeFunc::Parms);
 630         n->del_req(TypeFunc::Parms);
 631         n->add_prec(x);
 632       }
 633     }
 634   }
 635   for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count
 636     ready_cnt[_nodes[i2]->_idx] = 0;
 637 
 638   // All the prescheduled guys do not hold back internal nodes
 639   uint i3;
 640   for(i3 = 0; i3<phi_cnt; i3++ ) {  // For all pre-scheduled
 641     Node *n = _nodes[i3];       // Get pre-scheduled
 642     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
 643       Node* m = n->fast_out(j);
 644       if( cfg->_bbs[m->_idx] ==this ) // Local-block user
 645         ready_cnt[m->_idx]--;   // Fix ready count
 646     }
 647   }
 648 
 649   Node_List delay;
 650   // Make a worklist
 651   Node_List worklist;
 652   for(uint i4=i3; i4<node_cnt; i4++ ) {    // Put ready guys on worklist
 653     Node *m = _nodes[i4];
 654     if( !ready_cnt[m->_idx] ) {   // Zero ready count?
 655       if (m->is_iteratively_computed()) {
 656         // Push induction variable increments last to allow other uses
 657         // of the phi to be scheduled first. The select() method breaks
 658         // ties in scheduling by worklist order.
 659         delay.push(m);
 660       } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) {
 661         // Force the CreateEx to the top of the list so it's processed
 662         // first and ends up at the start of the block.
 663         worklist.insert(0, m);
 664       } else {
 665         worklist.push(m);         // Then on to worklist!
 666       }
 667     }
 668   }
 669   while (delay.size()) {
 670     Node* d = delay.pop();
 671     worklist.push(d);
 672   }
 673 
 674   // Warm up the 'next_call' heuristic bits
 675   needed_for_next_call(_nodes[0], next_call, cfg->_bbs);
 676 
 677 #ifndef PRODUCT
 678     if (cfg->trace_opto_pipelining()) {
 679       for (uint j=0; j<_nodes.size(); j++) {
 680         Node     *n = _nodes[j];
 681         int     idx = n->_idx;
 682         tty->print("#   ready cnt:%3d  ", ready_cnt[idx]);
 683         tty->print("latency:%3d  ", cfg->_node_latency.at_grow(idx));
 684         tty->print("%4d: %s\n", idx, n->Name());
 685       }
 686     }
 687 #endif
 688 
 689   // Pull from worklist and schedule
 690   while( worklist.size() ) {    // Worklist is not ready
 691 
 692 #ifndef PRODUCT
 693     if (cfg->trace_opto_pipelining()) {
 694       tty->print("#   ready list:");
 695       for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
 696         Node *n = worklist[i];      // Get Node on worklist
 697         tty->print(" %d", n->_idx);
 698       }
 699       tty->cr();
 700     }
 701 #endif
 702 
 703     // Select and pop a ready guy from worklist
 704     Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt);
 705     _nodes.map(phi_cnt++,n);    // Schedule him next
 706 
 707 #ifndef PRODUCT
 708     if (cfg->trace_opto_pipelining()) {
 709       tty->print("#    select %d: %s", n->_idx, n->Name());
 710       tty->print(", latency:%d", cfg->_node_latency.at_grow(n->_idx));
 711       n->dump();
 712       if (Verbose) {
 713         tty->print("#   ready list:");
 714         for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist
 715           Node *n = worklist[i];      // Get Node on worklist
 716           tty->print(" %d", n->_idx);
 717         }
 718         tty->cr();
 719       }
 720     }
 721 
 722 #endif
 723     if( n->is_MachCall() ) {
 724       MachCallNode *mcall = n->as_MachCall();
 725       phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call);
 726       continue;
 727     }
 728     // Children are now all ready
 729     for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) {
 730       Node* m = n->fast_out(i5); // Get user
 731       if( cfg->_bbs[m->_idx] != this ) continue;
 732       if( m->is_Phi() ) continue;
 733       if( !--ready_cnt[m->_idx] )
 734         worklist.push(m);
 735     }
 736   }
 737 
 738   if( phi_cnt != end_idx() ) {
 739     // did not schedule all.  Retry, Bailout, or Die
 740     Compile* C = matcher.C;
 741     if (C->subsume_loads() == true && !C->failing()) {
 742       // Retry with subsume_loads == false
 743       // If this is the first failure, the sentinel string will "stick"
 744       // to the Compile object, and the C2Compiler will see it and retry.
 745       C->record_failure(C2Compiler::retry_no_subsuming_loads());
 746     }
 747     // assert( phi_cnt == end_idx(), "did not schedule all" );
 748     return false;
 749   }
 750 
 751 #ifndef PRODUCT
 752   if (cfg->trace_opto_pipelining()) {
 753     tty->print_cr("#");
 754     tty->print_cr("# after schedule_local");
 755     for (uint i = 0;i < _nodes.size();i++) {
 756       tty->print("# ");
 757       _nodes[i]->fast_dump();
 758     }
 759     tty->cr();
 760   }
 761 #endif
 762 
 763 
 764   return true;
 765 }
 766 
 767 //--------------------------catch_cleanup_fix_all_inputs-----------------------
 768 static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) {
 769   for (uint l = 0; l < use->len(); l++) {
 770     if (use->in(l) == old_def) {
 771       if (l < use->req()) {
 772         use->set_req(l, new_def);
 773       } else {
 774         use->rm_prec(l);
 775         use->add_prec(new_def);
 776         l--;
 777       }
 778     }
 779   }
 780 }
 781 
 782 //------------------------------catch_cleanup_find_cloned_def------------------
 783 static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
 784   assert( use_blk != def_blk, "Inter-block cleanup only");
 785 
 786   // The use is some block below the Catch.  Find and return the clone of the def
 787   // that dominates the use. If there is no clone in a dominating block, then
 788   // create a phi for the def in a dominating block.
 789 
 790   // Find which successor block dominates this use.  The successor
 791   // blocks must all be single-entry (from the Catch only; I will have
 792   // split blocks to make this so), hence they all dominate.
 793   while( use_blk->_dom_depth > def_blk->_dom_depth+1 )
 794     use_blk = use_blk->_idom;
 795 
 796   // Find the successor
 797   Node *fixup = NULL;
 798 
 799   uint j;
 800   for( j = 0; j < def_blk->_num_succs; j++ )
 801     if( use_blk == def_blk->_succs[j] )
 802       break;
 803 
 804   if( j == def_blk->_num_succs ) {
 805     // Block at same level in dom-tree is not a successor.  It needs a
 806     // PhiNode, the PhiNode uses from the def and IT's uses need fixup.
 807     Node_Array inputs = new Node_List(Thread::current()->resource_area());
 808     for(uint k = 1; k < use_blk->num_preds(); k++) {
 809       inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx));
 810     }
 811 
 812     // Check to see if the use_blk already has an identical phi inserted.
 813     // If it exists, it will be at the first position since all uses of a
 814     // def are processed together.
 815     Node *phi = use_blk->_nodes[1];
 816     if( phi->is_Phi() ) {
 817       fixup = phi;
 818       for (uint k = 1; k < use_blk->num_preds(); k++) {
 819         if (phi->in(k) != inputs[k]) {
 820           // Not a match
 821           fixup = NULL;
 822           break;
 823         }
 824       }
 825     }
 826 
 827     // If an existing PhiNode was not found, make a new one.
 828     if (fixup == NULL) {
 829       Node *new_phi = PhiNode::make(use_blk->head(), def);
 830       use_blk->_nodes.insert(1, new_phi);
 831       bbs.map(new_phi->_idx, use_blk);
 832       for (uint k = 1; k < use_blk->num_preds(); k++) {
 833         new_phi->set_req(k, inputs[k]);
 834       }
 835       fixup = new_phi;
 836     }
 837 
 838   } else {
 839     // Found the use just below the Catch.  Make it use the clone.
 840     fixup = use_blk->_nodes[n_clone_idx];
 841   }
 842 
 843   return fixup;
 844 }
 845 
 846 //--------------------------catch_cleanup_intra_block--------------------------
 847 // Fix all input edges in use that reference "def".  The use is in the same
 848 // block as the def and both have been cloned in each successor block.
 849 static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) {
 850 
 851   // Both the use and def have been cloned. For each successor block,
 852   // get the clone of the use, and make its input the clone of the def
 853   // found in that block.
 854 
 855   uint use_idx = blk->find_node(use);
 856   uint offset_idx = use_idx - beg;
 857   for( uint k = 0; k < blk->_num_succs; k++ ) {
 858     // Get clone in each successor block
 859     Block *sb = blk->_succs[k];
 860     Node *clone = sb->_nodes[offset_idx+1];
 861     assert( clone->Opcode() == use->Opcode(), "" );
 862 
 863     // Make use-clone reference the def-clone
 864     catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]);
 865   }
 866 }
 867 
 868 //------------------------------catch_cleanup_inter_block---------------------
 869 // Fix all input edges in use that reference "def".  The use is in a different
 870 // block than the def.
 871 static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) {
 872   if( !use_blk ) return;        // Can happen if the use is a precedence edge
 873 
 874   Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx);
 875   catch_cleanup_fix_all_inputs(use, def, new_def);
 876 }
 877 
 878 //------------------------------call_catch_cleanup-----------------------------
 879 // If we inserted any instructions between a Call and his CatchNode,
 880 // clone the instructions on all paths below the Catch.
 881 void Block::call_catch_cleanup(Block_Array &bbs) {
 882 
 883   // End of region to clone
 884   uint end = end_idx();
 885   if( !_nodes[end]->is_Catch() ) return;
 886   // Start of region to clone
 887   uint beg = end;
 888   while( _nodes[beg-1]->Opcode() != Op_MachProj ||
 889         !_nodes[beg-1]->in(0)->is_Call() ) {
 890     beg--;
 891     assert(beg > 0,"Catch cleanup walking beyond block boundary");
 892   }
 893   // Range of inserted instructions is [beg, end)
 894   if( beg == end ) return;
 895 
 896   // Clone along all Catch output paths.  Clone area between the 'beg' and
 897   // 'end' indices.
 898   for( uint i = 0; i < _num_succs; i++ ) {
 899     Block *sb = _succs[i];
 900     // Clone the entire area; ignoring the edge fixup for now.
 901     for( uint j = end; j > beg; j-- ) {
 902       Node *clone = _nodes[j-1]->clone();
 903       sb->_nodes.insert( 1, clone );
 904       bbs.map(clone->_idx,sb);
 905     }
 906   }
 907 
 908 
 909   // Fixup edges.  Check the def-use info per cloned Node
 910   for(uint i2 = beg; i2 < end; i2++ ) {
 911     uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block
 912     Node *n = _nodes[i2];        // Node that got cloned
 913     // Need DU safe iterator because of edge manipulation in calls.
 914     Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area());
 915     for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) {
 916       out->push(n->fast_out(j1));
 917     }
 918     uint max = out->size();
 919     for (uint j = 0; j < max; j++) {// For all users
 920       Node *use = out->pop();
 921       Block *buse = bbs[use->_idx];
 922       if( use->is_Phi() ) {
 923         for( uint k = 1; k < use->req(); k++ )
 924           if( use->in(k) == n ) {
 925             Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx);
 926             use->set_req(k, fixup);
 927           }
 928       } else {
 929         if (this == buse) {
 930           catch_cleanup_intra_block(use, n, this, beg, n_clone_idx);
 931         } else {
 932           catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx);
 933         }
 934       }
 935     } // End for all users
 936 
 937   } // End of for all Nodes in cloned area
 938 
 939   // Remove the now-dead cloned ops
 940   for(uint i3 = beg; i3 < end; i3++ ) {
 941     _nodes[beg]->disconnect_inputs(NULL);
 942     _nodes.remove(beg);
 943   }
 944 
 945   // If the successor blocks have a CreateEx node, move it back to the top
 946   for(uint i4 = 0; i4 < _num_succs; i4++ ) {
 947     Block *sb = _succs[i4];
 948     uint new_cnt = end - beg;
 949     // Remove any newly created, but dead, nodes.
 950     for( uint j = new_cnt; j > 0; j-- ) {
 951       Node *n = sb->_nodes[j];
 952       if (n->outcnt() == 0 &&
 953           (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){
 954         n->disconnect_inputs(NULL);
 955         sb->_nodes.remove(j);
 956         new_cnt--;
 957       }
 958     }
 959     // If any newly created nodes remain, move the CreateEx node to the top
 960     if (new_cnt > 0) {
 961       Node *cex = sb->_nodes[1+new_cnt];
 962       if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) {
 963         sb->_nodes.remove(1+new_cnt);
 964         sb->_nodes.insert(1,cex);
 965       }
 966     }
 967   }
 968 }