1 /*
   2  * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/block.hpp"
  29 #include "opto/cfgnode.hpp"
  30 #include "opto/chaitin.hpp"
  31 #include "opto/loopnode.hpp"
  32 #include "opto/machnode.hpp"
  33 #include "opto/matcher.hpp"
  34 #include "opto/opcodes.hpp"
  35 #include "opto/rootnode.hpp"
  36 #include "utilities/copy.hpp"
  37 
  38 // Optimization - Graph Style
  39 
  40 
  41 //-----------------------------------------------------------------------------
  42 void Block_Array::grow( uint i ) {
  43   assert(i >= Max(), "must be an overflow");
  44   debug_only(_limit = i+1);
  45   if( i < _size )  return;
  46   if( !_size ) {
  47     _size = 1;
  48     _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
  49     _blocks[0] = NULL;
  50   }
  51   uint old = _size;
  52   while( i >= _size ) _size <<= 1;      // Double to fit
  53   _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
  54   Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
  55 }
  56 
  57 //=============================================================================
  58 void Block_List::remove(uint i) {
  59   assert(i < _cnt, "index out of bounds");
  60   Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
  61   pop(); // shrink list by one block
  62 }
  63 
  64 void Block_List::insert(uint i, Block *b) {
  65   push(b); // grow list by one block
  66   Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
  67   _blocks[i] = b;
  68 }
  69 
  70 #ifndef PRODUCT
  71 void Block_List::print() {
  72   for (uint i=0; i < size(); i++) {
  73     tty->print("B%d ", _blocks[i]->_pre_order);
  74   }
  75   tty->print("size = %d\n", size());
  76 }
  77 #endif
  78 
  79 //=============================================================================
  80 
  81 uint Block::code_alignment() {
  82   // Check for Root block
  83   if (_pre_order == 0) return CodeEntryAlignment;
  84   // Check for Start block
  85   if (_pre_order == 1) return InteriorEntryAlignment;
  86   // Check for loop alignment
  87   if (has_loop_alignment()) return loop_alignment();
  88 
  89   return relocInfo::addr_unit(); // no particular alignment
  90 }
  91 
  92 uint Block::compute_loop_alignment() {
  93   Node *h = head();
  94   int unit_sz = relocInfo::addr_unit();
  95   if (h->is_Loop() && h->as_Loop()->is_inner_loop())  {
  96     // Pre- and post-loops have low trip count so do not bother with
  97     // NOPs for align loop head.  The constants are hidden from tuning
  98     // but only because my "divide by 4" heuristic surely gets nearly
  99     // all possible gain (a "do not align at all" heuristic has a
 100     // chance of getting a really tiny gain).
 101     if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
 102                                 h->as_CountedLoop()->is_post_loop())) {
 103       return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
 104     }
 105     // Loops with low backedge frequency should not be aligned.
 106     Node *n = h->in(LoopNode::LoopBackControl)->in(0);
 107     if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
 108       return unit_sz; // Loop does not loop, more often than not!
 109     }
 110     return OptoLoopAlignment; // Otherwise align loop head
 111   }
 112 
 113   return unit_sz; // no particular alignment
 114 }
 115 
 116 //-----------------------------------------------------------------------------
 117 // Compute the size of first 'inst_cnt' instructions in this block.
 118 // Return the number of instructions left to compute if the block has
 119 // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
 120 // exceeds OptoLoopAlignment.
 121 uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
 122                                     PhaseRegAlloc* ra) {
 123   uint last_inst = _nodes.size();
 124   for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
 125     uint inst_size = _nodes[j]->size(ra);
 126     if( inst_size > 0 ) {
 127       inst_cnt--;
 128       uint sz = sum_size + inst_size;
 129       if( sz <= (uint)OptoLoopAlignment ) {
 130         // Compute size of instructions which fit into fetch buffer only
 131         // since all inst_cnt instructions will not fit even if we align them.
 132         sum_size = sz;
 133       } else {
 134         return 0;
 135       }
 136     }
 137   }
 138   return inst_cnt;
 139 }
 140 
 141 //-----------------------------------------------------------------------------
 142 uint Block::find_node( const Node *n ) const {
 143   for( uint i = 0; i < _nodes.size(); i++ ) {
 144     if( _nodes[i] == n )
 145       return i;
 146   }
 147   ShouldNotReachHere();
 148   return 0;
 149 }
 150 
 151 // Find and remove n from block list
 152 void Block::find_remove( const Node *n ) {
 153   _nodes.remove(find_node(n));
 154 }
 155 
 156 //------------------------------is_Empty---------------------------------------
 157 // Return empty status of a block.  Empty blocks contain only the head, other
 158 // ideal nodes, and an optional trailing goto.
 159 int Block::is_Empty() const {
 160 
 161   // Root or start block is not considered empty
 162   if (head()->is_Root() || head()->is_Start()) {
 163     return not_empty;
 164   }
 165 
 166   int success_result = completely_empty;
 167   int end_idx = _nodes.size()-1;
 168 
 169   // Check for ending goto
 170   if ((end_idx > 0) && (_nodes[end_idx]->is_MachGoto())) {
 171     success_result = empty_with_goto;
 172     end_idx--;
 173   }
 174 
 175   // Unreachable blocks are considered empty
 176   if (num_preds() <= 1) {
 177     return success_result;
 178   }
 179 
 180   // Ideal nodes are allowable in empty blocks: skip them  Only MachNodes
 181   // turn directly into code, because only MachNodes have non-trivial
 182   // emit() functions.
 183   while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) {
 184     end_idx--;
 185   }
 186 
 187   // No room for any interesting instructions?
 188   if (end_idx == 0) {
 189     return success_result;
 190   }
 191 
 192   return not_empty;
 193 }
 194 
 195 //------------------------------has_uncommon_code------------------------------
 196 // Return true if the block's code implies that it is likely to be
 197 // executed infrequently.  Check to see if the block ends in a Halt or
 198 // a low probability call.
 199 bool Block::has_uncommon_code() const {
 200   Node* en = end();
 201 
 202   if (en->is_MachGoto())
 203     en = en->in(0);
 204   if (en->is_Catch())
 205     en = en->in(0);
 206   if (en->is_MachProj() && en->in(0)->is_MachCall()) {
 207     MachCallNode* call = en->in(0)->as_MachCall();
 208     if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
 209       // This is true for slow-path stubs like new_{instance,array},
 210       // slow_arraycopy, complete_monitor_locking, uncommon_trap.
 211       // The magic number corresponds to the probability of an uncommon_trap,
 212       // even though it is a count not a probability.
 213       return true;
 214     }
 215   }
 216 
 217   int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
 218   return op == Op_Halt;
 219 }
 220 
 221 //------------------------------is_uncommon------------------------------------
 222 // True if block is low enough frequency or guarded by a test which
 223 // mostly does not go here.
 224 bool Block::is_uncommon(PhaseCFG* cfg) const {
 225   // Initial blocks must never be moved, so are never uncommon.
 226   if (head()->is_Root() || head()->is_Start())  return false;
 227 
 228   // Check for way-low freq
 229   if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true;
 230 
 231   // Look for code shape indicating uncommon_trap or slow path
 232   if (has_uncommon_code()) return true;
 233 
 234   const float epsilon = 0.05f;
 235   const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
 236   uint uncommon_preds = 0;
 237   uint freq_preds = 0;
 238   uint uncommon_for_freq_preds = 0;
 239 
 240   for( uint i=1; i<num_preds(); i++ ) {
 241     Block* guard = cfg->get_block_for_node(pred(i));
 242     // Check to see if this block follows its guard 1 time out of 10000
 243     // or less.
 244     //
 245     // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
 246     // we intend to be "uncommon", such as slow-path TLE allocation,
 247     // predicted call failure, and uncommon trap triggers.
 248     //
 249     // Use an epsilon value of 5% to allow for variability in frequency
 250     // predictions and floating point calculations. The net effect is
 251     // that guard_factor is set to 9500.
 252     //
 253     // Ignore low-frequency blocks.
 254     // The next check is (guard->_freq < 1.e-5 * 9500.).
 255     if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
 256       uncommon_preds++;
 257     } else {
 258       freq_preds++;
 259       if( _freq < guard->_freq * guard_factor ) {
 260         uncommon_for_freq_preds++;
 261       }
 262     }
 263   }
 264   if( num_preds() > 1 &&
 265       // The block is uncommon if all preds are uncommon or
 266       (uncommon_preds == (num_preds()-1) ||
 267       // it is uncommon for all frequent preds.
 268        uncommon_for_freq_preds == freq_preds) ) {
 269     return true;
 270   }
 271   return false;
 272 }
 273 
 274 //------------------------------dump-------------------------------------------
 275 #ifndef PRODUCT
 276 void Block::dump_bidx(const Block* orig, outputStream* st) const {
 277   if (_pre_order) st->print("B%d",_pre_order);
 278   else st->print("N%d", head()->_idx);
 279 
 280   if (Verbose && orig != this) {
 281     // Dump the original block's idx
 282     st->print(" (");
 283     orig->dump_bidx(orig, st);
 284     st->print(")");
 285   }
 286 }
 287 
 288 void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
 289   if (is_connector()) {
 290     for (uint i=1; i<num_preds(); i++) {
 291       Block *p = cfg->get_block_for_node(pred(i));
 292       p->dump_pred(cfg, orig, st);
 293     }
 294   } else {
 295     dump_bidx(orig, st);
 296     st->print(" ");
 297   }
 298 }
 299 
 300 void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
 301   // Print the basic block
 302   dump_bidx(this, st);
 303   st->print(": #\t");
 304 
 305   // Print the incoming CFG edges and the outgoing CFG edges
 306   for( uint i=0; i<_num_succs; i++ ) {
 307     non_connector_successor(i)->dump_bidx(_succs[i], st);
 308     st->print(" ");
 309   }
 310   st->print("<- ");
 311   if( head()->is_block_start() ) {
 312     for (uint i=1; i<num_preds(); i++) {
 313       Node *s = pred(i);
 314       if (cfg != NULL) {
 315         Block *p = cfg->get_block_for_node(s);
 316         p->dump_pred(cfg, p, st);
 317       } else {
 318         while (!s->is_block_start())
 319           s = s->in(0);
 320         st->print("N%d ", s->_idx );
 321       }
 322     }
 323   } else {
 324     st->print("BLOCK HEAD IS JUNK  ");
 325   }
 326 
 327   // Print loop, if any
 328   const Block *bhead = this;    // Head of self-loop
 329   Node *bh = bhead->head();
 330 
 331   if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
 332     LoopNode *loop = bh->as_Loop();
 333     const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
 334     while (bx->is_connector()) {
 335       bx = cfg->get_block_for_node(bx->pred(1));
 336     }
 337     st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
 338     // Dump any loop-specific bits, especially for CountedLoops.
 339     loop->dump_spec(st);
 340   } else if (has_loop_alignment()) {
 341     st->print(" top-of-loop");
 342   }
 343   st->print(" Freq: %g",_freq);
 344   if( Verbose || WizardMode ) {
 345     st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
 346     st->print(" RegPressure: %d",_reg_pressure);
 347     st->print(" IHRP Index: %d",_ihrp_index);
 348     st->print(" FRegPressure: %d",_freg_pressure);
 349     st->print(" FHRP Index: %d",_fhrp_index);
 350   }
 351   st->print_cr("");
 352 }
 353 
 354 void Block::dump() const {
 355   dump(NULL);
 356 }
 357 
 358 void Block::dump(const PhaseCFG* cfg) const {
 359   dump_head(cfg);
 360   for (uint i=0; i< _nodes.size(); i++) {
 361     _nodes[i]->dump();
 362   }
 363   tty->print("\n");
 364 }
 365 #endif
 366 
 367 //=============================================================================
 368 //------------------------------PhaseCFG---------------------------------------
 369 PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
 370 : Phase(CFG)
 371 , _block_arena(arena)
 372 , _node_to_block_mapping(arena)
 373 , _root(root)
 374 , _node_latency(NULL)
 375 #ifndef PRODUCT
 376 , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
 377 #endif
 378 #ifdef ASSERT
 379 , _raw_oops(arena)
 380 #endif
 381 {
 382   ResourceMark rm;
 383   // I'll need a few machine-specific GotoNodes.  Make an Ideal GotoNode,
 384   // then Match it into a machine-specific Node.  Then clone the machine
 385   // Node on demand.
 386   Node *x = new (C) GotoNode(NULL);
 387   x->init_req(0, x);
 388   _goto = matcher.match_tree(x);
 389   assert(_goto != NULL, "");
 390   _goto->set_req(0,_goto);
 391 
 392   // Build the CFG in Reverse Post Order
 393   _num_blocks = build_cfg();
 394   _broot = get_block_for_node(_root);
 395 }
 396 
 397 //------------------------------build_cfg--------------------------------------
 398 // Build a proper looking CFG.  Make every block begin with either a StartNode
 399 // or a RegionNode.  Make every block end with either a Goto, If or Return.
 400 // The RootNode both starts and ends it's own block.  Do this with a recursive
 401 // backwards walk over the control edges.
 402 uint PhaseCFG::build_cfg() {
 403   Arena *a = Thread::current()->resource_area();
 404   VectorSet visited(a);
 405 
 406   // Allocate stack with enough space to avoid frequent realloc
 407   Node_Stack nstack(a, C->unique() >> 1);
 408   nstack.push(_root, 0);
 409   uint sum = 0;                 // Counter for blocks
 410 
 411   while (nstack.is_nonempty()) {
 412     // node and in's index from stack's top
 413     // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
 414     // only nodes which point to the start of basic block (see below).
 415     Node *np = nstack.node();
 416     // idx > 0, except for the first node (_root) pushed on stack
 417     // at the beginning when idx == 0.
 418     // We will use the condition (idx == 0) later to end the build.
 419     uint idx = nstack.index();
 420     Node *proj = np->in(idx);
 421     const Node *x = proj->is_block_proj();
 422     // Does the block end with a proper block-ending Node?  One of Return,
 423     // If or Goto? (This check should be done for visited nodes also).
 424     if (x == NULL) {                    // Does not end right...
 425       Node *g = _goto->clone(); // Force it to end in a Goto
 426       g->set_req(0, proj);
 427       np->set_req(idx, g);
 428       x = proj = g;
 429     }
 430     if (!visited.test_set(x->_idx)) { // Visit this block once
 431       // Skip any control-pinned middle'in stuff
 432       Node *p = proj;
 433       do {
 434         proj = p;                   // Update pointer to last Control
 435         p = p->in(0);               // Move control forward
 436       } while( !p->is_block_proj() &&
 437                !p->is_block_start() );
 438       // Make the block begin with one of Region or StartNode.
 439       if( !p->is_block_start() ) {
 440         RegionNode *r = new (C) RegionNode( 2 );
 441         r->init_req(1, p);         // Insert RegionNode in the way
 442         proj->set_req(0, r);        // Insert RegionNode in the way
 443         p = r;
 444       }
 445       // 'p' now points to the start of this basic block
 446 
 447       // Put self in array of basic blocks
 448       Block *bb = new (_block_arena) Block(_block_arena, p);
 449       map_node_to_block(p, bb);
 450       map_node_to_block(x, bb);
 451       if( x != p ) {                // Only for root is x == p
 452         bb->_nodes.push((Node*)x);
 453       }
 454       // Now handle predecessors
 455       ++sum;                        // Count 1 for self block
 456       uint cnt = bb->num_preds();
 457       for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
 458         Node *prevproj = p->in(i);  // Get prior input
 459         assert( !prevproj->is_Con(), "dead input not removed" );
 460         // Check to see if p->in(i) is a "control-dependent" CFG edge -
 461         // i.e., it splits at the source (via an IF or SWITCH) and merges
 462         // at the destination (via a many-input Region).
 463         // This breaks critical edges.  The RegionNode to start the block
 464         // will be added when <p,i> is pulled off the node stack
 465         if ( cnt > 2 ) {             // Merging many things?
 466           assert( prevproj== bb->pred(i),"");
 467           if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
 468             // Force a block on the control-dependent edge
 469             Node *g = _goto->clone();       // Force it to end in a Goto
 470             g->set_req(0,prevproj);
 471             p->set_req(i,g);
 472           }
 473         }
 474         nstack.push(p, i);  // 'p' is RegionNode or StartNode
 475       }
 476     } else { // Post-processing visited nodes
 477       nstack.pop();                 // remove node from stack
 478       // Check if it the fist node pushed on stack at the beginning.
 479       if (idx == 0) break;          // end of the build
 480       // Find predecessor basic block
 481       Block *pb = get_block_for_node(x);
 482       // Insert into nodes array, if not already there
 483       if (!has_block(proj)) {
 484         assert( x != proj, "" );
 485         // Map basic block of projection
 486         map_node_to_block(proj, pb);
 487         pb->_nodes.push(proj);
 488       }
 489       // Insert self as a child of my predecessor block
 490       pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
 491       assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(),
 492               "too many control users, not a CFG?" );
 493     }
 494   }
 495   // Return number of basic blocks for all children and self
 496   return sum;
 497 }
 498 
 499 //------------------------------insert_goto_at---------------------------------
 500 // Inserts a goto & corresponding basic block between
 501 // block[block_no] and its succ_no'th successor block
 502 void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
 503   // get block with block_no
 504   assert(block_no < _num_blocks, "illegal block number");
 505   Block* in  = _blocks[block_no];
 506   // get successor block succ_no
 507   assert(succ_no < in->_num_succs, "illegal successor number");
 508   Block* out = in->_succs[succ_no];
 509   // Compute frequency of the new block. Do this before inserting
 510   // new block in case succ_prob() needs to infer the probability from
 511   // surrounding blocks.
 512   float freq = in->_freq * in->succ_prob(succ_no);
 513   // get ProjNode corresponding to the succ_no'th successor of the in block
 514   ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj();
 515   // create region for basic block
 516   RegionNode* region = new (C) RegionNode(2);
 517   region->init_req(1, proj);
 518   // setup corresponding basic block
 519   Block* block = new (_block_arena) Block(_block_arena, region);
 520   map_node_to_block(region, block);
 521   C->regalloc()->set_bad(region->_idx);
 522   // add a goto node
 523   Node* gto = _goto->clone(); // get a new goto node
 524   gto->set_req(0, region);
 525   // add it to the basic block
 526   block->_nodes.push(gto);
 527   map_node_to_block(gto, block);
 528   C->regalloc()->set_bad(gto->_idx);
 529   // hook up successor block
 530   block->_succs.map(block->_num_succs++, out);
 531   // remap successor's predecessors if necessary
 532   for (uint i = 1; i < out->num_preds(); i++) {
 533     if (out->pred(i) == proj) out->head()->set_req(i, gto);
 534   }
 535   // remap predecessor's successor to new block
 536   in->_succs.map(succ_no, block);
 537   // Set the frequency of the new block
 538   block->_freq = freq;
 539   // add new basic block to basic block list
 540   _blocks.insert(block_no + 1, block);
 541   _num_blocks++;
 542 }
 543 
 544 //------------------------------no_flip_branch---------------------------------
 545 // Does this block end in a multiway branch that cannot have the default case
 546 // flipped for another case?
 547 static bool no_flip_branch( Block *b ) {
 548   int branch_idx = b->_nodes.size() - b->_num_succs-1;
 549   if( branch_idx < 1 ) return false;
 550   Node *bra = b->_nodes[branch_idx];
 551   if( bra->is_Catch() )
 552     return true;
 553   if( bra->is_Mach() ) {
 554     if( bra->is_MachNullCheck() )
 555       return true;
 556     int iop = bra->as_Mach()->ideal_Opcode();
 557     if( iop == Op_FastLock || iop == Op_FastUnlock )
 558       return true;
 559   }
 560   return false;
 561 }
 562 
 563 //------------------------------convert_NeverBranch_to_Goto--------------------
 564 // Check for NeverBranch at block end.  This needs to become a GOTO to the
 565 // true target.  NeverBranch are treated as a conditional branch that always
 566 // goes the same direction for most of the optimizer and are used to give a
 567 // fake exit path to infinite loops.  At this late stage they need to turn
 568 // into Goto's so that when you enter the infinite loop you indeed hang.
 569 void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
 570   // Find true target
 571   int end_idx = b->end_idx();
 572   int idx = b->_nodes[end_idx+1]->as_Proj()->_con;
 573   Block *succ = b->_succs[idx];
 574   Node* gto = _goto->clone(); // get a new goto node
 575   gto->set_req(0, b->head());
 576   Node *bp = b->_nodes[end_idx];
 577   b->_nodes.map(end_idx,gto); // Slam over NeverBranch
 578   map_node_to_block(gto, b);
 579   C->regalloc()->set_bad(gto->_idx);
 580   b->_nodes.pop();              // Yank projections
 581   b->_nodes.pop();              // Yank projections
 582   b->_succs.map(0,succ);        // Map only successor
 583   b->_num_succs = 1;
 584   // remap successor's predecessors if necessary
 585   uint j;
 586   for( j = 1; j < succ->num_preds(); j++)
 587     if( succ->pred(j)->in(0) == bp )
 588       succ->head()->set_req(j, gto);
 589   // Kill alternate exit path
 590   Block *dead = b->_succs[1-idx];
 591   for( j = 1; j < dead->num_preds(); j++)
 592     if( dead->pred(j)->in(0) == bp )
 593       break;
 594   // Scan through block, yanking dead path from
 595   // all regions and phis.
 596   dead->head()->del_req(j);
 597   for( int k = 1; dead->_nodes[k]->is_Phi(); k++ )
 598     dead->_nodes[k]->del_req(j);
 599 }
 600 
 601 //------------------------------move_to_next-----------------------------------
 602 // Helper function to move block bx to the slot following b_index. Return
 603 // true if the move is successful, otherwise false
 604 bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
 605   if (bx == NULL) return false;
 606 
 607   // Return false if bx is already scheduled.
 608   uint bx_index = bx->_pre_order;
 609   if ((bx_index <= b_index) && (_blocks[bx_index] == bx)) {
 610     return false;
 611   }
 612 
 613   // Find the current index of block bx on the block list
 614   bx_index = b_index + 1;
 615   while( bx_index < _num_blocks && _blocks[bx_index] != bx ) bx_index++;
 616   assert(_blocks[bx_index] == bx, "block not found");
 617 
 618   // If the previous block conditionally falls into bx, return false,
 619   // because moving bx will create an extra jump.
 620   for(uint k = 1; k < bx->num_preds(); k++ ) {
 621     Block* pred = get_block_for_node(bx->pred(k));
 622     if (pred == _blocks[bx_index-1]) {
 623       if (pred->_num_succs != 1) {
 624         return false;
 625       }
 626     }
 627   }
 628 
 629   // Reinsert bx just past block 'b'
 630   _blocks.remove(bx_index);
 631   _blocks.insert(b_index + 1, bx);
 632   return true;
 633 }
 634 
 635 //------------------------------move_to_end------------------------------------
 636 // Move empty and uncommon blocks to the end.
 637 void PhaseCFG::move_to_end(Block *b, uint i) {
 638   int e = b->is_Empty();
 639   if (e != Block::not_empty) {
 640     if (e == Block::empty_with_goto) {
 641       // Remove the goto, but leave the block.
 642       b->_nodes.pop();
 643     }
 644     // Mark this block as a connector block, which will cause it to be
 645     // ignored in certain functions such as non_connector_successor().
 646     b->set_connector();
 647   }
 648   // Move the empty block to the end, and don't recheck.
 649   _blocks.remove(i);
 650   _blocks.push(b);
 651 }
 652 
 653 //---------------------------set_loop_alignment--------------------------------
 654 // Set loop alignment for every block
 655 void PhaseCFG::set_loop_alignment() {
 656   uint last = _num_blocks;
 657   assert( _blocks[0] == _broot, "" );
 658 
 659   for (uint i = 1; i < last; i++ ) {
 660     Block *b = _blocks[i];
 661     if (b->head()->is_Loop()) {
 662       b->set_loop_alignment(b);
 663     }
 664   }
 665 }
 666 
 667 //-----------------------------remove_empty------------------------------------
 668 // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
 669 // to the end.
 670 void PhaseCFG::remove_empty() {
 671   // Move uncommon blocks to the end
 672   uint last = _num_blocks;
 673   assert( _blocks[0] == _broot, "" );
 674 
 675   for (uint i = 1; i < last; i++) {
 676     Block *b = _blocks[i];
 677     if (b->is_connector()) break;
 678 
 679     // Check for NeverBranch at block end.  This needs to become a GOTO to the
 680     // true target.  NeverBranch are treated as a conditional branch that
 681     // always goes the same direction for most of the optimizer and are used
 682     // to give a fake exit path to infinite loops.  At this late stage they
 683     // need to turn into Goto's so that when you enter the infinite loop you
 684     // indeed hang.
 685     if( b->_nodes[b->end_idx()]->Opcode() == Op_NeverBranch )
 686       convert_NeverBranch_to_Goto(b);
 687 
 688     // Look for uncommon blocks and move to end.
 689     if (!C->do_freq_based_layout()) {
 690       if (b->is_uncommon(this)) {
 691         move_to_end(b, i);
 692         last--;                   // No longer check for being uncommon!
 693         if( no_flip_branch(b) ) { // Fall-thru case must follow?
 694           b = _blocks[i];         // Find the fall-thru block
 695           move_to_end(b, i);
 696           last--;
 697         }
 698         i--;                      // backup block counter post-increment
 699       }
 700     }
 701   }
 702 
 703   // Move empty blocks to the end
 704   last = _num_blocks;
 705   for (uint i = 1; i < last; i++) {
 706     Block *b = _blocks[i];
 707     if (b->is_Empty() != Block::not_empty) {
 708       move_to_end(b, i);
 709       last--;
 710       i--;
 711     }
 712   } // End of for all blocks
 713 }
 714 
 715 //-----------------------------fixup_flow--------------------------------------
 716 // Fix up the final control flow for basic blocks.
 717 void PhaseCFG::fixup_flow() {
 718   // Fixup final control flow for the blocks.  Remove jump-to-next
 719   // block.  If neither arm of a IF follows the conditional branch, we
 720   // have to add a second jump after the conditional.  We place the
 721   // TRUE branch target in succs[0] for both GOTOs and IFs.
 722   for (uint i=0; i < _num_blocks; i++) {
 723     Block *b = _blocks[i];
 724     b->_pre_order = i;          // turn pre-order into block-index
 725 
 726     // Connector blocks need no further processing.
 727     if (b->is_connector()) {
 728       assert((i+1) == _num_blocks || _blocks[i+1]->is_connector(),
 729              "All connector blocks should sink to the end");
 730       continue;
 731     }
 732     assert(b->is_Empty() != Block::completely_empty,
 733            "Empty blocks should be connectors");
 734 
 735     Block *bnext = (i < _num_blocks-1) ? _blocks[i+1] : NULL;
 736     Block *bs0 = b->non_connector_successor(0);
 737 
 738     // Check for multi-way branches where I cannot negate the test to
 739     // exchange the true and false targets.
 740     if( no_flip_branch( b ) ) {
 741       // Find fall through case - if must fall into its target
 742       int branch_idx = b->_nodes.size() - b->_num_succs;
 743       for (uint j2 = 0; j2 < b->_num_succs; j2++) {
 744         const ProjNode* p = b->_nodes[branch_idx + j2]->as_Proj();
 745         if (p->_con == 0) {
 746           // successor j2 is fall through case
 747           if (b->non_connector_successor(j2) != bnext) {
 748             // but it is not the next block => insert a goto
 749             insert_goto_at(i, j2);
 750           }
 751           // Put taken branch in slot 0
 752           if( j2 == 0 && b->_num_succs == 2) {
 753             // Flip targets in succs map
 754             Block *tbs0 = b->_succs[0];
 755             Block *tbs1 = b->_succs[1];
 756             b->_succs.map( 0, tbs1 );
 757             b->_succs.map( 1, tbs0 );
 758           }
 759           break;
 760         }
 761       }
 762       // Remove all CatchProjs
 763       for (uint j1 = 0; j1 < b->_num_succs; j1++) b->_nodes.pop();
 764 
 765     } else if (b->_num_succs == 1) {
 766       // Block ends in a Goto?
 767       if (bnext == bs0) {
 768         // We fall into next block; remove the Goto
 769         b->_nodes.pop();
 770       }
 771 
 772     } else if( b->_num_succs == 2 ) { // Block ends in a If?
 773       // Get opcode of 1st projection (matches _succs[0])
 774       // Note: Since this basic block has 2 exits, the last 2 nodes must
 775       //       be projections (in any order), the 3rd last node must be
 776       //       the IfNode (we have excluded other 2-way exits such as
 777       //       CatchNodes already).
 778       MachNode *iff   = b->_nodes[b->_nodes.size()-3]->as_Mach();
 779       ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
 780       ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
 781 
 782       // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
 783       assert(proj0->raw_out(0) == b->_succs[0]->head(), "Mismatch successor 0");
 784       assert(proj1->raw_out(0) == b->_succs[1]->head(), "Mismatch successor 1");
 785 
 786       Block *bs1 = b->non_connector_successor(1);
 787 
 788       // Check for neither successor block following the current
 789       // block ending in a conditional. If so, move one of the
 790       // successors after the current one, provided that the
 791       // successor was previously unscheduled, but moveable
 792       // (i.e., all paths to it involve a branch).
 793       if( !C->do_freq_based_layout() && bnext != bs0 && bnext != bs1 ) {
 794         // Choose the more common successor based on the probability
 795         // of the conditional branch.
 796         Block *bx = bs0;
 797         Block *by = bs1;
 798 
 799         // _prob is the probability of taking the true path. Make
 800         // p the probability of taking successor #1.
 801         float p = iff->as_MachIf()->_prob;
 802         if( proj0->Opcode() == Op_IfTrue ) {
 803           p = 1.0 - p;
 804         }
 805 
 806         // Prefer successor #1 if p > 0.5
 807         if (p > PROB_FAIR) {
 808           bx = bs1;
 809           by = bs0;
 810         }
 811 
 812         // Attempt the more common successor first
 813         if (move_to_next(bx, i)) {
 814           bnext = bx;
 815         } else if (move_to_next(by, i)) {
 816           bnext = by;
 817         }
 818       }
 819 
 820       // Check for conditional branching the wrong way.  Negate
 821       // conditional, if needed, so it falls into the following block
 822       // and branches to the not-following block.
 823 
 824       // Check for the next block being in succs[0].  We are going to branch
 825       // to succs[0], so we want the fall-thru case as the next block in
 826       // succs[1].
 827       if (bnext == bs0) {
 828         // Fall-thru case in succs[0], so flip targets in succs map
 829         Block *tbs0 = b->_succs[0];
 830         Block *tbs1 = b->_succs[1];
 831         b->_succs.map( 0, tbs1 );
 832         b->_succs.map( 1, tbs0 );
 833         // Flip projection for each target
 834         { ProjNode *tmp = proj0; proj0 = proj1; proj1 = tmp; }
 835 
 836       } else if( bnext != bs1 ) {
 837         // Need a double-branch
 838         // The existing conditional branch need not change.
 839         // Add a unconditional branch to the false target.
 840         // Alas, it must appear in its own block and adding a
 841         // block this late in the game is complicated.  Sigh.
 842         insert_goto_at(i, 1);
 843       }
 844 
 845       // Make sure we TRUE branch to the target
 846       if( proj0->Opcode() == Op_IfFalse ) {
 847         iff->as_MachIf()->negate();
 848       }
 849 
 850       b->_nodes.pop();          // Remove IfFalse & IfTrue projections
 851       b->_nodes.pop();
 852 
 853     } else {
 854       // Multi-exit block, e.g. a switch statement
 855       // But we don't need to do anything here
 856     }
 857   } // End of for all blocks
 858 }
 859 
 860 
 861 //------------------------------dump-------------------------------------------
 862 #ifndef PRODUCT
 863 void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited  ) const {
 864   const Node *x = end->is_block_proj();
 865   assert( x, "not a CFG" );
 866 
 867   // Do not visit this block again
 868   if( visited.test_set(x->_idx) ) return;
 869 
 870   // Skip through this block
 871   const Node *p = x;
 872   do {
 873     p = p->in(0);               // Move control forward
 874     assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
 875   } while( !p->is_block_start() );
 876 
 877   // Recursively visit
 878   for (uint i = 1; i < p->req(); i++) {
 879     _dump_cfg(p->in(i), visited);
 880   }
 881 
 882   // Dump the block
 883   get_block_for_node(p)->dump(this);
 884 }
 885 
 886 void PhaseCFG::dump( ) const {
 887   tty->print("\n--- CFG --- %d BBs\n",_num_blocks);
 888   if (_blocks.size()) {        // Did we do basic-block layout?
 889     for (uint i = 0; i < _num_blocks; i++) {
 890       _blocks[i]->dump(this);
 891     }
 892   } else {                      // Else do it with a DFS
 893     VectorSet visited(_block_arena);
 894     _dump_cfg(_root,visited);
 895   }
 896 }
 897 
 898 void PhaseCFG::dump_headers() {
 899   for( uint i = 0; i < _num_blocks; i++ ) {
 900     if (_blocks[i]) {
 901       _blocks[i]->dump_head(this);
 902     }
 903   }
 904 }
 905 
 906 void PhaseCFG::verify( ) const {
 907 #ifdef ASSERT
 908   // Verify sane CFG
 909   for (uint i = 0; i < _num_blocks; i++) {
 910     Block *b = _blocks[i];
 911     uint cnt = b->_nodes.size();
 912     uint j;
 913     for (j = 0; j < cnt; j++)  {
 914       Node *n = b->_nodes[j];
 915       assert(get_block_for_node(n) == b, "");
 916       if (j >= 1 && n->is_Mach() &&
 917           n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
 918         assert(j == 1 || b->_nodes[j-1]->is_Phi(),
 919                "CreateEx must be first instruction in block");
 920       }
 921       for (uint k = 0; k < n->req(); k++) {
 922         Node *def = n->in(k);
 923         if (def && def != n) {
 924           assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
 925           // Verify that instructions in the block is in correct order.
 926           // Uses must follow their definition if they are at the same block.
 927           // Mostly done to check that MachSpillCopy nodes are placed correctly
 928           // when CreateEx node is moved in build_ifg_physical().
 929           if (get_block_for_node(def) == b &&
 930               !(b->head()->is_Loop() && n->is_Phi()) &&
 931               // See (+++) comment in reg_split.cpp
 932               !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
 933             bool is_loop = false;
 934             if (n->is_Phi()) {
 935               for (uint l = 1; l < def->req(); l++) {
 936                 if (n == def->in(l)) {
 937                   is_loop = true;
 938                   break; // Some kind of loop
 939                 }
 940               }
 941             }
 942             assert(is_loop || b->find_node(def) < j, "uses must follow definitions");
 943           }
 944         }
 945       }
 946     }
 947 
 948     j = b->end_idx();
 949     Node *bp = (Node*)b->_nodes[b->_nodes.size()-1]->is_block_proj();
 950     assert( bp, "last instruction must be a block proj" );
 951     assert( bp == b->_nodes[j], "wrong number of successors for this block" );
 952     if (bp->is_Catch()) {
 953       while (b->_nodes[--j]->is_MachProj()) ;
 954       assert(b->_nodes[j]->is_MachCall(), "CatchProj must follow call");
 955     } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
 956       assert(b->_num_succs == 2, "Conditional branch must have two targets");
 957     }
 958   }
 959 #endif
 960 }
 961 #endif
 962 
 963 //=============================================================================
 964 //------------------------------UnionFind--------------------------------------
 965 UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
 966   Copy::zero_to_bytes( _indices, sizeof(uint)*max );
 967 }
 968 
 969 void UnionFind::extend( uint from_idx, uint to_idx ) {
 970   _nesting.check();
 971   if( from_idx >= _max ) {
 972     uint size = 16;
 973     while( size <= from_idx ) size <<=1;
 974     _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
 975     _max = size;
 976   }
 977   while( _cnt <= from_idx ) _indices[_cnt++] = 0;
 978   _indices[from_idx] = to_idx;
 979 }
 980 
 981 void UnionFind::reset( uint max ) {
 982   assert( max <= max_uint, "Must fit within uint" );
 983   // Force the Union-Find mapping to be at least this large
 984   extend(max,0);
 985   // Initialize to be the ID mapping.
 986   for( uint i=0; i<max; i++ ) map(i,i);
 987 }
 988 
 989 //------------------------------Find_compress----------------------------------
 990 // Straight out of Tarjan's union-find algorithm
 991 uint UnionFind::Find_compress( uint idx ) {
 992   uint cur  = idx;
 993   uint next = lookup(cur);
 994   while( next != cur ) {        // Scan chain of equivalences
 995     assert( next < cur, "always union smaller" );
 996     cur = next;                 // until find a fixed-point
 997     next = lookup(cur);
 998   }
 999   // Core of union-find algorithm: update chain of
1000   // equivalences to be equal to the root.
1001   while( idx != next ) {
1002     uint tmp = lookup(idx);
1003     map(idx, next);
1004     idx = tmp;
1005   }
1006   return idx;
1007 }
1008 
1009 //------------------------------Find_const-------------------------------------
1010 // Like Find above, but no path compress, so bad asymptotic behavior
1011 uint UnionFind::Find_const( uint idx ) const {
1012   if( idx == 0 ) return idx;    // Ignore the zero idx
1013   // Off the end?  This can happen during debugging dumps
1014   // when data structures have not finished being updated.
1015   if( idx >= _max ) return idx;
1016   uint next = lookup(idx);
1017   while( next != idx ) {        // Scan chain of equivalences
1018     idx = next;                 // until find a fixed-point
1019     next = lookup(idx);
1020   }
1021   return next;
1022 }
1023 
1024 //------------------------------Union------------------------------------------
1025 // union 2 sets together.
1026 void UnionFind::Union( uint idx1, uint idx2 ) {
1027   uint src = Find(idx1);
1028   uint dst = Find(idx2);
1029   assert( src, "" );
1030   assert( dst, "" );
1031   assert( src < _max, "oob" );
1032   assert( dst < _max, "oob" );
1033   assert( src < dst, "always union smaller" );
1034   map(dst,src);
1035 }
1036 
1037 #ifndef PRODUCT
1038 void Trace::dump( ) const {
1039   tty->print_cr("Trace (freq %f)", first_block()->_freq);
1040   for (Block *b = first_block(); b != NULL; b = next(b)) {
1041     tty->print("  B%d", b->_pre_order);
1042     if (b->head()->is_Loop()) {
1043       tty->print(" (L%d)", b->compute_loop_alignment());
1044     }
1045     if (b->has_loop_alignment()) {
1046       tty->print(" (T%d)", b->code_alignment());
1047     }
1048   }
1049   tty->cr();
1050 }
1051 
1052 void CFGEdge::dump( ) const {
1053   tty->print(" B%d  -->  B%d  Freq: %f  out:%3d%%  in:%3d%%  State: ",
1054              from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
1055   switch(state()) {
1056   case connected:
1057     tty->print("connected");
1058     break;
1059   case open:
1060     tty->print("open");
1061     break;
1062   case interior:
1063     tty->print("interior");
1064     break;
1065   }
1066   if (infrequent()) {
1067     tty->print("  infrequent");
1068   }
1069   tty->cr();
1070 }
1071 #endif
1072 
1073 //=============================================================================
1074 
1075 //------------------------------edge_order-------------------------------------
1076 // Comparison function for edges
1077 static int edge_order(CFGEdge **e0, CFGEdge **e1) {
1078   float freq0 = (*e0)->freq();
1079   float freq1 = (*e1)->freq();
1080   if (freq0 != freq1) {
1081     return freq0 > freq1 ? -1 : 1;
1082   }
1083 
1084   int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
1085   int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
1086 
1087   return dist1 - dist0;
1088 }
1089 
1090 //------------------------------trace_frequency_order--------------------------
1091 // Comparison function for edges
1092 extern "C" int trace_frequency_order(const void *p0, const void *p1) {
1093   Trace *tr0 = *(Trace **) p0;
1094   Trace *tr1 = *(Trace **) p1;
1095   Block *b0 = tr0->first_block();
1096   Block *b1 = tr1->first_block();
1097 
1098   // The trace of connector blocks goes at the end;
1099   // we only expect one such trace
1100   if (b0->is_connector() != b1->is_connector()) {
1101     return b1->is_connector() ? -1 : 1;
1102   }
1103 
1104   // Pull more frequently executed blocks to the beginning
1105   float freq0 = b0->_freq;
1106   float freq1 = b1->_freq;
1107   if (freq0 != freq1) {
1108     return freq0 > freq1 ? -1 : 1;
1109   }
1110 
1111   int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
1112 
1113   return diff;
1114 }
1115 
1116 //------------------------------find_edges-------------------------------------
1117 // Find edges of interest, i.e, those which can fall through. Presumes that
1118 // edges which don't fall through are of low frequency and can be generally
1119 // ignored.  Initialize the list of traces.
1120 void PhaseBlockLayout::find_edges()
1121 {
1122   // Walk the blocks, creating edges and Traces
1123   uint i;
1124   Trace *tr = NULL;
1125   for (i = 0; i < _cfg._num_blocks; i++) {
1126     Block *b = _cfg._blocks[i];
1127     tr = new Trace(b, next, prev);
1128     traces[tr->id()] = tr;
1129 
1130     // All connector blocks should be at the end of the list
1131     if (b->is_connector()) break;
1132 
1133     // If this block and the next one have a one-to-one successor
1134     // predecessor relationship, simply append the next block
1135     int nfallthru = b->num_fall_throughs();
1136     while (nfallthru == 1 &&
1137            b->succ_fall_through(0)) {
1138       Block *n = b->_succs[0];
1139 
1140       // Skip over single-entry connector blocks, we don't want to
1141       // add them to the trace.
1142       while (n->is_connector() && n->num_preds() == 1) {
1143         n = n->_succs[0];
1144       }
1145 
1146       // We see a merge point, so stop search for the next block
1147       if (n->num_preds() != 1) break;
1148 
1149       i++;
1150       assert(n = _cfg._blocks[i], "expecting next block");
1151       tr->append(n);
1152       uf->map(n->_pre_order, tr->id());
1153       traces[n->_pre_order] = NULL;
1154       nfallthru = b->num_fall_throughs();
1155       b = n;
1156     }
1157 
1158     if (nfallthru > 0) {
1159       // Create a CFGEdge for each outgoing
1160       // edge that could be a fall-through.
1161       for (uint j = 0; j < b->_num_succs; j++ ) {
1162         if (b->succ_fall_through(j)) {
1163           Block *target = b->non_connector_successor(j);
1164           float freq = b->_freq * b->succ_prob(j);
1165           int from_pct = (int) ((100 * freq) / b->_freq);
1166           int to_pct = (int) ((100 * freq) / target->_freq);
1167           edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
1168         }
1169       }
1170     }
1171   }
1172 
1173   // Group connector blocks into one trace
1174   for (i++; i < _cfg._num_blocks; i++) {
1175     Block *b = _cfg._blocks[i];
1176     assert(b->is_connector(), "connector blocks at the end");
1177     tr->append(b);
1178     uf->map(b->_pre_order, tr->id());
1179     traces[b->_pre_order] = NULL;
1180   }
1181 }
1182 
1183 //------------------------------union_traces----------------------------------
1184 // Union two traces together in uf, and null out the trace in the list
1185 void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace)
1186 {
1187   uint old_id = old_trace->id();
1188   uint updated_id = updated_trace->id();
1189 
1190   uint lo_id = updated_id;
1191   uint hi_id = old_id;
1192 
1193   // If from is greater than to, swap values to meet
1194   // UnionFind guarantee.
1195   if (updated_id > old_id) {
1196     lo_id = old_id;
1197     hi_id = updated_id;
1198 
1199     // Fix up the trace ids
1200     traces[lo_id] = traces[updated_id];
1201     updated_trace->set_id(lo_id);
1202   }
1203 
1204   // Union the lower with the higher and remove the pointer
1205   // to the higher.
1206   uf->Union(lo_id, hi_id);
1207   traces[hi_id] = NULL;
1208 }
1209 
1210 //------------------------------grow_traces-------------------------------------
1211 // Append traces together via the most frequently executed edges
1212 void PhaseBlockLayout::grow_traces()
1213 {
1214   // Order the edges, and drive the growth of Traces via the most
1215   // frequently executed edges.
1216   edges->sort(edge_order);
1217   for (int i = 0; i < edges->length(); i++) {
1218     CFGEdge *e = edges->at(i);
1219 
1220     if (e->state() != CFGEdge::open) continue;
1221 
1222     Block *src_block = e->from();
1223     Block *targ_block = e->to();
1224 
1225     // Don't grow traces along backedges?
1226     if (!BlockLayoutRotateLoops) {
1227       if (targ_block->_rpo <= src_block->_rpo) {
1228         targ_block->set_loop_alignment(targ_block);
1229         continue;
1230       }
1231     }
1232 
1233     Trace *src_trace = trace(src_block);
1234     Trace *targ_trace = trace(targ_block);
1235 
1236     // If the edge in question can join two traces at their ends,
1237     // append one trace to the other.
1238    if (src_trace->last_block() == src_block) {
1239       if (src_trace == targ_trace) {
1240         e->set_state(CFGEdge::interior);
1241         if (targ_trace->backedge(e)) {
1242           // Reset i to catch any newly eligible edge
1243           // (Or we could remember the first "open" edge, and reset there)
1244           i = 0;
1245         }
1246       } else if (targ_trace->first_block() == targ_block) {
1247         e->set_state(CFGEdge::connected);
1248         src_trace->append(targ_trace);
1249         union_traces(src_trace, targ_trace);
1250       }
1251     }
1252   }
1253 }
1254 
1255 //------------------------------merge_traces-----------------------------------
1256 // Embed one trace into another, if the fork or join points are sufficiently
1257 // balanced.
1258 void PhaseBlockLayout::merge_traces(bool fall_thru_only)
1259 {
1260   // Walk the edge list a another time, looking at unprocessed edges.
1261   // Fold in diamonds
1262   for (int i = 0; i < edges->length(); i++) {
1263     CFGEdge *e = edges->at(i);
1264 
1265     if (e->state() != CFGEdge::open) continue;
1266     if (fall_thru_only) {
1267       if (e->infrequent()) continue;
1268     }
1269 
1270     Block *src_block = e->from();
1271     Trace *src_trace = trace(src_block);
1272     bool src_at_tail = src_trace->last_block() == src_block;
1273 
1274     Block *targ_block  = e->to();
1275     Trace *targ_trace  = trace(targ_block);
1276     bool targ_at_start = targ_trace->first_block() == targ_block;
1277 
1278     if (src_trace == targ_trace) {
1279       // This may be a loop, but we can't do much about it.
1280       e->set_state(CFGEdge::interior);
1281       continue;
1282     }
1283 
1284     if (fall_thru_only) {
1285       // If the edge links the middle of two traces, we can't do anything.
1286       // Mark the edge and continue.
1287       if (!src_at_tail & !targ_at_start) {
1288         continue;
1289       }
1290 
1291       // Don't grow traces along backedges?
1292       if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
1293           continue;
1294       }
1295 
1296       // If both ends of the edge are available, why didn't we handle it earlier?
1297       assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
1298 
1299       if (targ_at_start) {
1300         // Insert the "targ" trace in the "src" trace if the insertion point
1301         // is a two way branch.
1302         // Better profitability check possible, but may not be worth it.
1303         // Someday, see if the this "fork" has an associated "join";
1304         // then make a policy on merging this trace at the fork or join.
1305         // For example, other things being equal, it may be better to place this
1306         // trace at the join point if the "src" trace ends in a two-way, but
1307         // the insertion point is one-way.
1308         assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
1309         e->set_state(CFGEdge::connected);
1310         src_trace->insert_after(src_block, targ_trace);
1311         union_traces(src_trace, targ_trace);
1312       } else if (src_at_tail) {
1313         if (src_trace != trace(_cfg._broot)) {
1314           e->set_state(CFGEdge::connected);
1315           targ_trace->insert_before(targ_block, src_trace);
1316           union_traces(targ_trace, src_trace);
1317         }
1318       }
1319     } else if (e->state() == CFGEdge::open) {
1320       // Append traces, even without a fall-thru connection.
1321       // But leave root entry at the beginning of the block list.
1322       if (targ_trace != trace(_cfg._broot)) {
1323         e->set_state(CFGEdge::connected);
1324         src_trace->append(targ_trace);
1325         union_traces(src_trace, targ_trace);
1326       }
1327     }
1328   }
1329 }
1330 
1331 //----------------------------reorder_traces-----------------------------------
1332 // Order the sequence of the traces in some desirable way, and fixup the
1333 // jumps at the end of each block.
1334 void PhaseBlockLayout::reorder_traces(int count)
1335 {
1336   ResourceArea *area = Thread::current()->resource_area();
1337   Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
1338   Block_List worklist;
1339   int new_count = 0;
1340 
1341   // Compact the traces.
1342   for (int i = 0; i < count; i++) {
1343     Trace *tr = traces[i];
1344     if (tr != NULL) {
1345       new_traces[new_count++] = tr;
1346     }
1347   }
1348 
1349   // The entry block should be first on the new trace list.
1350   Trace *tr = trace(_cfg._broot);
1351   assert(tr == new_traces[0], "entry trace misplaced");
1352 
1353   // Sort the new trace list by frequency
1354   qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
1355 
1356   // Patch up the successor blocks
1357   _cfg._blocks.reset();
1358   _cfg._num_blocks = 0;
1359   for (int i = 0; i < new_count; i++) {
1360     Trace *tr = new_traces[i];
1361     if (tr != NULL) {
1362       tr->fixup_blocks(_cfg);
1363     }
1364   }
1365 }
1366 
1367 //------------------------------PhaseBlockLayout-------------------------------
1368 // Order basic blocks based on frequency
1369 PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) :
1370   Phase(BlockLayout),
1371   _cfg(cfg)
1372 {
1373   ResourceMark rm;
1374   ResourceArea *area = Thread::current()->resource_area();
1375 
1376   // List of traces
1377   int size = _cfg._num_blocks + 1;
1378   traces = NEW_ARENA_ARRAY(area, Trace *, size);
1379   memset(traces, 0, size*sizeof(Trace*));
1380   next = NEW_ARENA_ARRAY(area, Block *, size);
1381   memset(next,   0, size*sizeof(Block *));
1382   prev = NEW_ARENA_ARRAY(area, Block *, size);
1383   memset(prev  , 0, size*sizeof(Block *));
1384 
1385   // List of edges
1386   edges = new GrowableArray<CFGEdge*>;
1387 
1388   // Mapping block index --> block_trace
1389   uf = new UnionFind(size);
1390   uf->reset(size);
1391 
1392   // Find edges and create traces.
1393   find_edges();
1394 
1395   // Grow traces at their ends via most frequent edges.
1396   grow_traces();
1397 
1398   // Merge one trace into another, but only at fall-through points.
1399   // This may make diamonds and other related shapes in a trace.
1400   merge_traces(true);
1401 
1402   // Run merge again, allowing two traces to be catenated, even if
1403   // one does not fall through into the other. This appends loosely
1404   // related traces to be near each other.
1405   merge_traces(false);
1406 
1407   // Re-order all the remaining traces by frequency
1408   reorder_traces(size);
1409 
1410   assert(_cfg._num_blocks >= (uint) (size - 1), "number of blocks can not shrink");
1411 }
1412 
1413 
1414 //------------------------------backedge---------------------------------------
1415 // Edge e completes a loop in a trace. If the target block is head of the
1416 // loop, rotate the loop block so that the loop ends in a conditional branch.
1417 bool Trace::backedge(CFGEdge *e) {
1418   bool loop_rotated = false;
1419   Block *src_block  = e->from();
1420   Block *targ_block    = e->to();
1421 
1422   assert(last_block() == src_block, "loop discovery at back branch");
1423   if (first_block() == targ_block) {
1424     if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
1425       // Find the last block in the trace that has a conditional
1426       // branch.
1427       Block *b;
1428       for (b = last_block(); b != NULL; b = prev(b)) {
1429         if (b->num_fall_throughs() == 2) {
1430           break;
1431         }
1432       }
1433 
1434       if (b != last_block() && b != NULL) {
1435         loop_rotated = true;
1436 
1437         // Rotate the loop by doing two-part linked-list surgery.
1438         append(first_block());
1439         break_loop_after(b);
1440       }
1441     }
1442 
1443     // Backbranch to the top of a trace
1444     // Scroll forward through the trace from the targ_block. If we find
1445     // a loop head before another loop top, use the the loop head alignment.
1446     for (Block *b = targ_block; b != NULL; b = next(b)) {
1447       if (b->has_loop_alignment()) {
1448         break;
1449       }
1450       if (b->head()->is_Loop()) {
1451         targ_block = b;
1452         break;
1453       }
1454     }
1455 
1456     first_block()->set_loop_alignment(targ_block);
1457 
1458   } else {
1459     // Backbranch into the middle of a trace
1460     targ_block->set_loop_alignment(targ_block);
1461   }
1462 
1463   return loop_rotated;
1464 }
1465 
1466 //------------------------------fixup_blocks-----------------------------------
1467 // push blocks onto the CFG list
1468 // ensure that blocks have the correct two-way branch sense
1469 void Trace::fixup_blocks(PhaseCFG &cfg) {
1470   Block *last = last_block();
1471   for (Block *b = first_block(); b != NULL; b = next(b)) {
1472     cfg._blocks.push(b);
1473     cfg._num_blocks++;
1474     if (!b->is_connector()) {
1475       int nfallthru = b->num_fall_throughs();
1476       if (b != last) {
1477         if (nfallthru == 2) {
1478           // Ensure that the sense of the branch is correct
1479           Block *bnext = next(b);
1480           Block *bs0 = b->non_connector_successor(0);
1481 
1482           MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach();
1483           ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj();
1484           ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj();
1485 
1486           if (bnext == bs0) {
1487             // Fall-thru case in succs[0], should be in succs[1]
1488 
1489             // Flip targets in _succs map
1490             Block *tbs0 = b->_succs[0];
1491             Block *tbs1 = b->_succs[1];
1492             b->_succs.map( 0, tbs1 );
1493             b->_succs.map( 1, tbs0 );
1494 
1495             // Flip projections to match targets
1496             b->_nodes.map(b->_nodes.size()-2, proj1);
1497             b->_nodes.map(b->_nodes.size()-1, proj0);
1498           }
1499         }
1500       }
1501     }
1502   }
1503 }