duke@435: /*
mikael@6198:  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
duke@435:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435:  *
duke@435:  * This code is free software; you can redistribute it and/or modify it
duke@435:  * under the terms of the GNU General Public License version 2 only, as
duke@435:  * published by the Free Software Foundation.
duke@435:  *
duke@435:  * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
duke@435:  * version 2 for more details (a copy is included in the LICENSE file that
duke@435:  * accompanied this code).
duke@435:  *
duke@435:  * You should have received a copy of the GNU General Public License version
duke@435:  * 2 along with this work; if not, write to the Free Software Foundation,
duke@435:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435:  *
trims@1907:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907:  * or visit www.oracle.com if you need additional information or have any
trims@1907:  * questions.
duke@435:  *
duke@435:  */
duke@435: 
stefank@2314: #include "precompiled.hpp"
stefank@2314: #include "libadt/vectset.hpp"
stefank@2314: #include "memory/allocation.inline.hpp"
stefank@2314: #include "opto/block.hpp"
stefank@2314: #include "opto/cfgnode.hpp"
stefank@2314: #include "opto/chaitin.hpp"
stefank@2314: #include "opto/loopnode.hpp"
stefank@2314: #include "opto/machnode.hpp"
stefank@2314: #include "opto/matcher.hpp"
stefank@2314: #include "opto/opcodes.hpp"
stefank@2314: #include "opto/rootnode.hpp"
stefank@2314: #include "utilities/copy.hpp"
stefank@2314: 
duke@435: void Block_Array::grow( uint i ) {
duke@435:   assert(i >= Max(), "must be an overflow");
duke@435:   debug_only(_limit = i+1);
duke@435:   if( i < _size )  return;
duke@435:   if( !_size ) {
duke@435:     _size = 1;
duke@435:     _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) );
duke@435:     _blocks[0] = NULL;
duke@435:   }
duke@435:   uint old = _size;
duke@435:   while( i >= _size ) _size <<= 1;      // Double to fit
duke@435:   _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*));
duke@435:   Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) );
duke@435: }
duke@435: 
duke@435: void Block_List::remove(uint i) {
duke@435:   assert(i < _cnt, "index out of bounds");
duke@435:   Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*)));
duke@435:   pop(); // shrink list by one block
duke@435: }
duke@435: 
duke@435: void Block_List::insert(uint i, Block *b) {
duke@435:   push(b); // grow list by one block
duke@435:   Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*)));
duke@435:   _blocks[i] = b;
duke@435: }
duke@435: 
rasbold@853: #ifndef PRODUCT
rasbold@853: void Block_List::print() {
rasbold@853:   for (uint i=0; i < size(); i++) {
rasbold@853:     tty->print("B%d ", _blocks[i]->_pre_order);
rasbold@853:   }
rasbold@853:   tty->print("size = %d\n", size());
rasbold@853: }
rasbold@853: #endif
duke@435: 
duke@435: uint Block::code_alignment() {
duke@435:   // Check for Root block
kvn@3049:   if (_pre_order == 0) return CodeEntryAlignment;
duke@435:   // Check for Start block
kvn@3049:   if (_pre_order == 1) return InteriorEntryAlignment;
duke@435:   // Check for loop alignment
kvn@3049:   if (has_loop_alignment()) return loop_alignment();
rasbold@853: 
kvn@3049:   return relocInfo::addr_unit(); // no particular alignment
rasbold@853: }
rasbold@853: 
rasbold@853: uint Block::compute_loop_alignment() {
duke@435:   Node *h = head();
kvn@3049:   int unit_sz = relocInfo::addr_unit();
kvn@3049:   if (h->is_Loop() && h->as_Loop()->is_inner_loop())  {
duke@435:     // Pre- and post-loops have low trip count so do not bother with
duke@435:     // NOPs for align loop head.  The constants are hidden from tuning
duke@435:     // but only because my "divide by 4" heuristic surely gets nearly
duke@435:     // all possible gain (a "do not align at all" heuristic has a
duke@435:     // chance of getting a really tiny gain).
kvn@3049:     if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() ||
kvn@3049:                                 h->as_CountedLoop()->is_post_loop())) {
kvn@3049:       return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz;
kvn@3049:     }
duke@435:     // Loops with low backedge frequency should not be aligned.
duke@435:     Node *n = h->in(LoopNode::LoopBackControl)->in(0);
kvn@3049:     if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) {
kvn@3049:       return unit_sz; // Loop does not loop, more often than not!
duke@435:     }
duke@435:     return OptoLoopAlignment; // Otherwise align loop head
duke@435:   }
rasbold@853: 
kvn@3049:   return unit_sz; // no particular alignment
duke@435: }
duke@435: 
duke@435: // Compute the size of first 'inst_cnt' instructions in this block.
duke@435: // Return the number of instructions left to compute if the block has
rasbold@853: // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size
rasbold@853: // exceeds OptoLoopAlignment.
duke@435: uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt,
duke@435:                                     PhaseRegAlloc* ra) {
adlertz@5635:   uint last_inst = number_of_nodes();
duke@435:   for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) {
adlertz@5635:     uint inst_size = get_node(j)->size(ra);
duke@435:     if( inst_size > 0 ) {
duke@435:       inst_cnt--;
duke@435:       uint sz = sum_size + inst_size;
duke@435:       if( sz <= (uint)OptoLoopAlignment ) {
duke@435:         // Compute size of instructions which fit into fetch buffer only
duke@435:         // since all inst_cnt instructions will not fit even if we align them.
duke@435:         sum_size = sz;
duke@435:       } else {
duke@435:         return 0;
duke@435:       }
duke@435:     }
duke@435:   }
duke@435:   return inst_cnt;
duke@435: }
duke@435: 
duke@435: uint Block::find_node( const Node *n ) const {
adlertz@5635:   for( uint i = 0; i < number_of_nodes(); i++ ) {
adlertz@5635:     if( get_node(i) == n )
duke@435:       return i;
duke@435:   }
duke@435:   ShouldNotReachHere();
duke@435:   return 0;
duke@435: }
duke@435: 
duke@435: // Find and remove n from block list
duke@435: void Block::find_remove( const Node *n ) {
adlertz@5635:   remove_node(find_node(n));
duke@435: }
duke@435: 
duke@435: // Return empty status of a block.  Empty blocks contain only the head, other
duke@435: // ideal nodes, and an optional trailing goto.
duke@435: int Block::is_Empty() const {
duke@435: 
duke@435:   // Root or start block is not considered empty
duke@435:   if (head()->is_Root() || head()->is_Start()) {
duke@435:     return not_empty;
duke@435:   }
duke@435: 
duke@435:   int success_result = completely_empty;
adlertz@5635:   int end_idx = number_of_nodes() - 1;
duke@435: 
duke@435:   // Check for ending goto
adlertz@5635:   if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) {
duke@435:     success_result = empty_with_goto;
duke@435:     end_idx--;
duke@435:   }
duke@435: 
duke@435:   // Unreachable blocks are considered empty
duke@435:   if (num_preds() <= 1) {
duke@435:     return success_result;
duke@435:   }
duke@435: 
duke@435:   // Ideal nodes are allowable in empty blocks: skip them  Only MachNodes
duke@435:   // turn directly into code, because only MachNodes have non-trivial
duke@435:   // emit() functions.
adlertz@5635:   while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) {
duke@435:     end_idx--;
duke@435:   }
duke@435: 
duke@435:   // No room for any interesting instructions?
duke@435:   if (end_idx == 0) {
duke@435:     return success_result;
duke@435:   }
duke@435: 
duke@435:   return not_empty;
duke@435: }
duke@435: 
twisti@1040: // Return true if the block's code implies that it is likely to be
duke@435: // executed infrequently.  Check to see if the block ends in a Halt or
duke@435: // a low probability call.
duke@435: bool Block::has_uncommon_code() const {
duke@435:   Node* en = end();
duke@435: 
kvn@3040:   if (en->is_MachGoto())
duke@435:     en = en->in(0);
duke@435:   if (en->is_Catch())
duke@435:     en = en->in(0);
kvn@3040:   if (en->is_MachProj() && en->in(0)->is_MachCall()) {
duke@435:     MachCallNode* call = en->in(0)->as_MachCall();
duke@435:     if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) {
duke@435:       // This is true for slow-path stubs like new_{instance,array},
duke@435:       // slow_arraycopy, complete_monitor_locking, uncommon_trap.
duke@435:       // The magic number corresponds to the probability of an uncommon_trap,
duke@435:       // even though it is a count not a probability.
duke@435:       return true;
duke@435:     }
duke@435:   }
duke@435: 
duke@435:   int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode();
duke@435:   return op == Op_Halt;
duke@435: }
duke@435: 
duke@435: // True if block is low enough frequency or guarded by a test which
duke@435: // mostly does not go here.
adlertz@5639: bool PhaseCFG::is_uncommon(const Block* block) {
duke@435:   // Initial blocks must never be moved, so are never uncommon.
adlertz@5639:   if (block->head()->is_Root() || block->head()->is_Start())  return false;
duke@435: 
duke@435:   // Check for way-low freq
adlertz@5639:   if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true;
duke@435: 
duke@435:   // Look for code shape indicating uncommon_trap or slow path
adlertz@5639:   if (block->has_uncommon_code()) return true;
duke@435: 
duke@435:   const float epsilon = 0.05f;
duke@435:   const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon);
duke@435:   uint uncommon_preds = 0;
duke@435:   uint freq_preds = 0;
duke@435:   uint uncommon_for_freq_preds = 0;
duke@435: 
adlertz@5639:   for( uint i=1; i< block->num_preds(); i++ ) {
adlertz@5639:     Block* guard = get_block_for_node(block->pred(i));
duke@435:     // Check to see if this block follows its guard 1 time out of 10000
duke@435:     // or less.
duke@435:     //
duke@435:     // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which
duke@435:     // we intend to be "uncommon", such as slow-path TLE allocation,
duke@435:     // predicted call failure, and uncommon trap triggers.
duke@435:     //
duke@435:     // Use an epsilon value of 5% to allow for variability in frequency
duke@435:     // predictions and floating point calculations. The net effect is
duke@435:     // that guard_factor is set to 9500.
duke@435:     //
duke@435:     // Ignore low-frequency blocks.
duke@435:     // The next check is (guard->_freq < 1.e-5 * 9500.).
duke@435:     if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) {
duke@435:       uncommon_preds++;
duke@435:     } else {
duke@435:       freq_preds++;
adlertz@5639:       if(block->_freq < guard->_freq * guard_factor ) {
duke@435:         uncommon_for_freq_preds++;
duke@435:       }
duke@435:     }
duke@435:   }
adlertz@5639:   if( block->num_preds() > 1 &&
duke@435:       // The block is uncommon if all preds are uncommon or
adlertz@5639:       (uncommon_preds == (block->num_preds()-1) ||
duke@435:       // it is uncommon for all frequent preds.
duke@435:        uncommon_for_freq_preds == freq_preds) ) {
duke@435:     return true;
duke@435:   }
duke@435:   return false;
duke@435: }
duke@435: 
duke@435: #ifndef PRODUCT
kvn@3049: void Block::dump_bidx(const Block* orig, outputStream* st) const {
kvn@3049:   if (_pre_order) st->print("B%d",_pre_order);
kvn@3049:   else st->print("N%d", head()->_idx);
duke@435: 
duke@435:   if (Verbose && orig != this) {
duke@435:     // Dump the original block's idx
kvn@3049:     st->print(" (");
kvn@3049:     orig->dump_bidx(orig, st);
kvn@3049:     st->print(")");
duke@435:   }
duke@435: }
duke@435: 
adlertz@5509: void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const {
duke@435:   if (is_connector()) {
duke@435:     for (uint i=1; i<num_preds(); i++) {
adlertz@5509:       Block *p = cfg->get_block_for_node(pred(i));
adlertz@5509:       p->dump_pred(cfg, orig, st);
duke@435:     }
duke@435:   } else {
kvn@3049:     dump_bidx(orig, st);
kvn@3049:     st->print(" ");
duke@435:   }
duke@435: }
duke@435: 
adlertz@5509: void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const {
duke@435:   // Print the basic block
kvn@3049:   dump_bidx(this, st);
kvn@3049:   st->print(": #\t");
duke@435: 
duke@435:   // Print the incoming CFG edges and the outgoing CFG edges
duke@435:   for( uint i=0; i<_num_succs; i++ ) {
kvn@3049:     non_connector_successor(i)->dump_bidx(_succs[i], st);
kvn@3049:     st->print(" ");
duke@435:   }
kvn@3049:   st->print("<- ");
duke@435:   if( head()->is_block_start() ) {
duke@435:     for (uint i=1; i<num_preds(); i++) {
duke@435:       Node *s = pred(i);
adlertz@5509:       if (cfg != NULL) {
adlertz@5509:         Block *p = cfg->get_block_for_node(s);
adlertz@5509:         p->dump_pred(cfg, p, st);
duke@435:       } else {
duke@435:         while (!s->is_block_start())
duke@435:           s = s->in(0);
kvn@3049:         st->print("N%d ", s->_idx );
duke@435:       }
duke@435:     }
adlertz@5509:   } else {
kvn@3049:     st->print("BLOCK HEAD IS JUNK  ");
adlertz@5509:   }
duke@435: 
duke@435:   // Print loop, if any
duke@435:   const Block *bhead = this;    // Head of self-loop
duke@435:   Node *bh = bhead->head();
adlertz@5509: 
adlertz@5509:   if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) {
duke@435:     LoopNode *loop = bh->as_Loop();
adlertz@5509:     const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl));
duke@435:     while (bx->is_connector()) {
adlertz@5509:       bx = cfg->get_block_for_node(bx->pred(1));
duke@435:     }
kvn@3049:     st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order);
duke@435:     // Dump any loop-specific bits, especially for CountedLoops.
kvn@3049:     loop->dump_spec(st);
rasbold@853:   } else if (has_loop_alignment()) {
kvn@3049:     st->print(" top-of-loop");
duke@435:   }
kvn@3049:   st->print(" Freq: %g",_freq);
duke@435:   if( Verbose || WizardMode ) {
kvn@3049:     st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth);
kvn@3049:     st->print(" RegPressure: %d",_reg_pressure);
kvn@3049:     st->print(" IHRP Index: %d",_ihrp_index);
kvn@3049:     st->print(" FRegPressure: %d",_freg_pressure);
kvn@3049:     st->print(" FHRP Index: %d",_fhrp_index);
duke@435:   }
kvn@3049:   st->print_cr("");
duke@435: }
duke@435: 
adlertz@5509: void Block::dump() const {
adlertz@5509:   dump(NULL);
adlertz@5509: }
duke@435: 
adlertz@5509: void Block::dump(const PhaseCFG* cfg) const {
adlertz@5509:   dump_head(cfg);
adlertz@5635:   for (uint i=0; i< number_of_nodes(); i++) {
adlertz@5635:     get_node(i)->dump();
adlertz@5509:   }
duke@435:   tty->print("\n");
duke@435: }
duke@435: #endif
duke@435: 
adlertz@5509: PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher)
adlertz@5509: : Phase(CFG)
adlertz@5509: , _block_arena(arena)
adlertz@5539: , _root(root)
adlertz@5539: , _matcher(matcher)
adlertz@5509: , _node_to_block_mapping(arena)
adlertz@5509: , _node_latency(NULL)
duke@435: #ifndef PRODUCT
adlertz@5509: , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining"))
duke@435: #endif
kvn@1268: #ifdef ASSERT
adlertz@5509: , _raw_oops(arena)
kvn@1268: #endif
duke@435: {
duke@435:   ResourceMark rm;
duke@435:   // I'll need a few machine-specific GotoNodes.  Make an Ideal GotoNode,
duke@435:   // then Match it into a machine-specific Node.  Then clone the machine
duke@435:   // Node on demand.
kvn@4115:   Node *x = new (C) GotoNode(NULL);
duke@435:   x->init_req(0, x);
adlertz@5509:   _goto = matcher.match_tree(x);
duke@435:   assert(_goto != NULL, "");
duke@435:   _goto->set_req(0,_goto);
duke@435: 
duke@435:   // Build the CFG in Reverse Post Order
adlertz@5539:   _number_of_blocks = build_cfg();
adlertz@5539:   _root_block = get_block_for_node(_root);
duke@435: }
duke@435: 
duke@435: // Build a proper looking CFG.  Make every block begin with either a StartNode
duke@435: // or a RegionNode.  Make every block end with either a Goto, If or Return.
duke@435: // The RootNode both starts and ends it's own block.  Do this with a recursive
duke@435: // backwards walk over the control edges.
duke@435: uint PhaseCFG::build_cfg() {
duke@435:   Arena *a = Thread::current()->resource_area();
duke@435:   VectorSet visited(a);
duke@435: 
duke@435:   // Allocate stack with enough space to avoid frequent realloc
duke@435:   Node_Stack nstack(a, C->unique() >> 1);
duke@435:   nstack.push(_root, 0);
duke@435:   uint sum = 0;                 // Counter for blocks
duke@435: 
duke@435:   while (nstack.is_nonempty()) {
duke@435:     // node and in's index from stack's top
duke@435:     // 'np' is _root (see above) or RegionNode, StartNode: we push on stack
duke@435:     // only nodes which point to the start of basic block (see below).
duke@435:     Node *np = nstack.node();
duke@435:     // idx > 0, except for the first node (_root) pushed on stack
duke@435:     // at the beginning when idx == 0.
duke@435:     // We will use the condition (idx == 0) later to end the build.
duke@435:     uint idx = nstack.index();
duke@435:     Node *proj = np->in(idx);
duke@435:     const Node *x = proj->is_block_proj();
duke@435:     // Does the block end with a proper block-ending Node?  One of Return,
duke@435:     // If or Goto? (This check should be done for visited nodes also).
duke@435:     if (x == NULL) {                    // Does not end right...
duke@435:       Node *g = _goto->clone(); // Force it to end in a Goto
duke@435:       g->set_req(0, proj);
duke@435:       np->set_req(idx, g);
duke@435:       x = proj = g;
duke@435:     }
duke@435:     if (!visited.test_set(x->_idx)) { // Visit this block once
duke@435:       // Skip any control-pinned middle'in stuff
duke@435:       Node *p = proj;
duke@435:       do {
duke@435:         proj = p;                   // Update pointer to last Control
duke@435:         p = p->in(0);               // Move control forward
duke@435:       } while( !p->is_block_proj() &&
duke@435:                !p->is_block_start() );
duke@435:       // Make the block begin with one of Region or StartNode.
duke@435:       if( !p->is_block_start() ) {
kvn@4115:         RegionNode *r = new (C) RegionNode( 2 );
duke@435:         r->init_req(1, p);         // Insert RegionNode in the way
duke@435:         proj->set_req(0, r);        // Insert RegionNode in the way
duke@435:         p = r;
duke@435:       }
duke@435:       // 'p' now points to the start of this basic block
duke@435: 
duke@435:       // Put self in array of basic blocks
adlertz@5509:       Block *bb = new (_block_arena) Block(_block_arena, p);
adlertz@5509:       map_node_to_block(p, bb);
adlertz@5509:       map_node_to_block(x, bb);
kvn@3049:       if( x != p ) {                // Only for root is x == p
adlertz@5635:         bb->push_node((Node*)x);
kvn@3049:       }
duke@435:       // Now handle predecessors
duke@435:       ++sum;                        // Count 1 for self block
duke@435:       uint cnt = bb->num_preds();
duke@435:       for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors
duke@435:         Node *prevproj = p->in(i);  // Get prior input
duke@435:         assert( !prevproj->is_Con(), "dead input not removed" );
duke@435:         // Check to see if p->in(i) is a "control-dependent" CFG edge -
duke@435:         // i.e., it splits at the source (via an IF or SWITCH) and merges
duke@435:         // at the destination (via a many-input Region).
duke@435:         // This breaks critical edges.  The RegionNode to start the block
duke@435:         // will be added when <p,i> is pulled off the node stack
duke@435:         if ( cnt > 2 ) {             // Merging many things?
duke@435:           assert( prevproj== bb->pred(i),"");
duke@435:           if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge?
duke@435:             // Force a block on the control-dependent edge
duke@435:             Node *g = _goto->clone();       // Force it to end in a Goto
duke@435:             g->set_req(0,prevproj);
duke@435:             p->set_req(i,g);
duke@435:           }
duke@435:         }
duke@435:         nstack.push(p, i);  // 'p' is RegionNode or StartNode
duke@435:       }
duke@435:     } else { // Post-processing visited nodes
duke@435:       nstack.pop();                 // remove node from stack
duke@435:       // Check if it the fist node pushed on stack at the beginning.
duke@435:       if (idx == 0) break;          // end of the build
duke@435:       // Find predecessor basic block
adlertz@5509:       Block *pb = get_block_for_node(x);
duke@435:       // Insert into nodes array, if not already there
adlertz@5509:       if (!has_block(proj)) {
duke@435:         assert( x != proj, "" );
duke@435:         // Map basic block of projection
adlertz@5509:         map_node_to_block(proj, pb);
adlertz@5635:         pb->push_node(proj);
duke@435:       }
duke@435:       // Insert self as a child of my predecessor block
adlertz@5509:       pb->_succs.map(pb->_num_succs++, get_block_for_node(np));
adlertz@5635:       assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(),
duke@435:               "too many control users, not a CFG?" );
duke@435:     }
duke@435:   }
duke@435:   // Return number of basic blocks for all children and self
duke@435:   return sum;
duke@435: }
duke@435: 
duke@435: // Inserts a goto & corresponding basic block between
duke@435: // block[block_no] and its succ_no'th successor block
duke@435: void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) {
duke@435:   // get block with block_no
adlertz@5539:   assert(block_no < number_of_blocks(), "illegal block number");
adlertz@5539:   Block* in  = get_block(block_no);
duke@435:   // get successor block succ_no
duke@435:   assert(succ_no < in->_num_succs, "illegal successor number");
duke@435:   Block* out = in->_succs[succ_no];
rasbold@743:   // Compute frequency of the new block. Do this before inserting
rasbold@743:   // new block in case succ_prob() needs to infer the probability from
rasbold@743:   // surrounding blocks.
rasbold@743:   float freq = in->_freq * in->succ_prob(succ_no);
duke@435:   // get ProjNode corresponding to the succ_no'th successor of the in block
adlertz@5635:   ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj();
duke@435:   // create region for basic block
kvn@4115:   RegionNode* region = new (C) RegionNode(2);
duke@435:   region->init_req(1, proj);
duke@435:   // setup corresponding basic block
adlertz@5509:   Block* block = new (_block_arena) Block(_block_arena, region);
adlertz@5509:   map_node_to_block(region, block);
duke@435:   C->regalloc()->set_bad(region->_idx);
duke@435:   // add a goto node
duke@435:   Node* gto = _goto->clone(); // get a new goto node
duke@435:   gto->set_req(0, region);
duke@435:   // add it to the basic block
adlertz@5635:   block->push_node(gto);
adlertz@5509:   map_node_to_block(gto, block);
duke@435:   C->regalloc()->set_bad(gto->_idx);
duke@435:   // hook up successor block
duke@435:   block->_succs.map(block->_num_succs++, out);
duke@435:   // remap successor's predecessors if necessary
duke@435:   for (uint i = 1; i < out->num_preds(); i++) {
duke@435:     if (out->pred(i) == proj) out->head()->set_req(i, gto);
duke@435:   }
duke@435:   // remap predecessor's successor to new block
duke@435:   in->_succs.map(succ_no, block);
rasbold@743:   // Set the frequency of the new block
rasbold@743:   block->_freq = freq;
duke@435:   // add new basic block to basic block list
adlertz@5539:   add_block_at(block_no + 1, block);
duke@435: }
duke@435: 
duke@435: // Does this block end in a multiway branch that cannot have the default case
duke@435: // flipped for another case?
duke@435: static bool no_flip_branch( Block *b ) {
adlertz@5635:   int branch_idx = b->number_of_nodes() - b->_num_succs-1;
duke@435:   if( branch_idx < 1 ) return false;
adlertz@5635:   Node *bra = b->get_node(branch_idx);
rasbold@853:   if( bra->is_Catch() )
rasbold@853:     return true;
duke@435:   if( bra->is_Mach() ) {
rasbold@853:     if( bra->is_MachNullCheck() )
rasbold@853:       return true;
duke@435:     int iop = bra->as_Mach()->ideal_Opcode();
duke@435:     if( iop == Op_FastLock || iop == Op_FastUnlock )
duke@435:       return true;
duke@435:   }
duke@435:   return false;
duke@435: }
duke@435: 
duke@435: // Check for NeverBranch at block end.  This needs to become a GOTO to the
duke@435: // true target.  NeverBranch are treated as a conditional branch that always
duke@435: // goes the same direction for most of the optimizer and are used to give a
duke@435: // fake exit path to infinite loops.  At this late stage they need to turn
duke@435: // into Goto's so that when you enter the infinite loop you indeed hang.
duke@435: void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) {
duke@435:   // Find true target
duke@435:   int end_idx = b->end_idx();
adlertz@5635:   int idx = b->get_node(end_idx+1)->as_Proj()->_con;
duke@435:   Block *succ = b->_succs[idx];
duke@435:   Node* gto = _goto->clone(); // get a new goto node
duke@435:   gto->set_req(0, b->head());
adlertz@5635:   Node *bp = b->get_node(end_idx);
adlertz@5635:   b->map_node(gto, end_idx); // Slam over NeverBranch
adlertz@5509:   map_node_to_block(gto, b);
duke@435:   C->regalloc()->set_bad(gto->_idx);
adlertz@5635:   b->pop_node();              // Yank projections
adlertz@5635:   b->pop_node();              // Yank projections
duke@435:   b->_succs.map(0,succ);        // Map only successor
duke@435:   b->_num_succs = 1;
duke@435:   // remap successor's predecessors if necessary
duke@435:   uint j;
duke@435:   for( j = 1; j < succ->num_preds(); j++)
duke@435:     if( succ->pred(j)->in(0) == bp )
duke@435:       succ->head()->set_req(j, gto);
duke@435:   // Kill alternate exit path
duke@435:   Block *dead = b->_succs[1-idx];
duke@435:   for( j = 1; j < dead->num_preds(); j++)
duke@435:     if( dead->pred(j)->in(0) == bp )
duke@435:       break;
duke@435:   // Scan through block, yanking dead path from
duke@435:   // all regions and phis.
duke@435:   dead->head()->del_req(j);
adlertz@5635:   for( int k = 1; dead->get_node(k)->is_Phi(); k++ )
adlertz@5635:     dead->get_node(k)->del_req(j);
duke@435: }
duke@435: 
duke@435: // Helper function to move block bx to the slot following b_index. Return
duke@435: // true if the move is successful, otherwise false
rasbold@853: bool PhaseCFG::move_to_next(Block* bx, uint b_index) {
duke@435:   if (bx == NULL) return false;
duke@435: 
duke@435:   // Return false if bx is already scheduled.
duke@435:   uint bx_index = bx->_pre_order;
adlertz@5539:   if ((bx_index <= b_index) && (get_block(bx_index) == bx)) {
duke@435:     return false;
duke@435:   }
duke@435: 
duke@435:   // Find the current index of block bx on the block list
duke@435:   bx_index = b_index + 1;
adlertz@5539:   while (bx_index < number_of_blocks() && get_block(bx_index) != bx) {
adlertz@5539:     bx_index++;
adlertz@5539:   }
adlertz@5539:   assert(get_block(bx_index) == bx, "block not found");
duke@435: 
duke@435:   // If the previous block conditionally falls into bx, return false,
duke@435:   // because moving bx will create an extra jump.
duke@435:   for(uint k = 1; k < bx->num_preds(); k++ ) {
adlertz@5509:     Block* pred = get_block_for_node(bx->pred(k));
adlertz@5539:     if (pred == get_block(bx_index - 1)) {
duke@435:       if (pred->_num_succs != 1) {
duke@435:         return false;
duke@435:       }
duke@435:     }
duke@435:   }
duke@435: 
duke@435:   // Reinsert bx just past block 'b'
duke@435:   _blocks.remove(bx_index);
duke@435:   _blocks.insert(b_index + 1, bx);
duke@435:   return true;
duke@435: }
duke@435: 
duke@435: // Move empty and uncommon blocks to the end.
rasbold@853: void PhaseCFG::move_to_end(Block *b, uint i) {
duke@435:   int e = b->is_Empty();
duke@435:   if (e != Block::not_empty) {
duke@435:     if (e == Block::empty_with_goto) {
duke@435:       // Remove the goto, but leave the block.
adlertz@5635:       b->pop_node();
duke@435:     }
duke@435:     // Mark this block as a connector block, which will cause it to be
duke@435:     // ignored in certain functions such as non_connector_successor().
duke@435:     b->set_connector();
duke@435:   }
duke@435:   // Move the empty block to the end, and don't recheck.
duke@435:   _blocks.remove(i);
duke@435:   _blocks.push(b);
duke@435: }
duke@435: 
rasbold@853: // Set loop alignment for every block
rasbold@853: void PhaseCFG::set_loop_alignment() {
adlertz@5539:   uint last = number_of_blocks();
adlertz@5539:   assert(get_block(0) == get_root_block(), "");
rasbold@853: 
adlertz@5539:   for (uint i = 1; i < last; i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5539:     if (block->head()->is_Loop()) {
adlertz@5539:       block->set_loop_alignment(block);
rasbold@853:     }
rasbold@853:   }
rasbold@853: }
rasbold@853: 
rasbold@853: // Make empty basic blocks to be "connector" blocks, Move uncommon blocks
rasbold@853: // to the end.
adlertz@5539: void PhaseCFG::remove_empty_blocks() {
duke@435:   // Move uncommon blocks to the end
adlertz@5539:   uint last = number_of_blocks();
adlertz@5539:   assert(get_block(0) == get_root_block(), "");
rasbold@853: 
rasbold@853:   for (uint i = 1; i < last; i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5539:     if (block->is_connector()) {
adlertz@5539:       break;
adlertz@5539:     }
duke@435: 
duke@435:     // Check for NeverBranch at block end.  This needs to become a GOTO to the
duke@435:     // true target.  NeverBranch are treated as a conditional branch that
duke@435:     // always goes the same direction for most of the optimizer and are used
duke@435:     // to give a fake exit path to infinite loops.  At this late stage they
duke@435:     // need to turn into Goto's so that when you enter the infinite loop you
duke@435:     // indeed hang.
adlertz@5635:     if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) {
adlertz@5539:       convert_NeverBranch_to_Goto(block);
adlertz@5539:     }
duke@435: 
duke@435:     // Look for uncommon blocks and move to end.
rasbold@853:     if (!C->do_freq_based_layout()) {
adlertz@5639:       if (is_uncommon(block)) {
adlertz@5539:         move_to_end(block, i);
rasbold@853:         last--;                   // No longer check for being uncommon!
adlertz@5539:         if (no_flip_branch(block)) { // Fall-thru case must follow?
adlertz@5539:           // Find the fall-thru block
adlertz@5539:           block = get_block(i);
adlertz@5539:           move_to_end(block, i);
rasbold@853:           last--;
rasbold@853:         }
adlertz@5539:         // backup block counter post-increment
adlertz@5539:         i--;
duke@435:       }
duke@435:     }
duke@435:   }
duke@435: 
rasbold@853:   // Move empty blocks to the end
adlertz@5539:   last = number_of_blocks();
rasbold@853:   for (uint i = 1; i < last; i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5539:     if (block->is_Empty() != Block::not_empty) {
adlertz@5539:       move_to_end(block, i);
rasbold@853:       last--;
rasbold@853:       i--;
duke@435:     }
duke@435:   } // End of for all blocks
rasbold@853: }
duke@435: 
rasbold@853: // Fix up the final control flow for basic blocks.
rasbold@853: void PhaseCFG::fixup_flow() {
duke@435:   // Fixup final control flow for the blocks.  Remove jump-to-next
duke@435:   // block.  If neither arm of a IF follows the conditional branch, we
duke@435:   // have to add a second jump after the conditional.  We place the
duke@435:   // TRUE branch target in succs[0] for both GOTOs and IFs.
adlertz@5539:   for (uint i = 0; i < number_of_blocks(); i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5539:     block->_pre_order = i;          // turn pre-order into block-index
duke@435: 
duke@435:     // Connector blocks need no further processing.
adlertz@5539:     if (block->is_connector()) {
adlertz@5539:       assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end");
duke@435:       continue;
duke@435:     }
adlertz@5539:     assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors");
duke@435: 
adlertz@5539:     Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL;
adlertz@5539:     Block* bs0 = block->non_connector_successor(0);
duke@435: 
duke@435:     // Check for multi-way branches where I cannot negate the test to
duke@435:     // exchange the true and false targets.
adlertz@5539:     if (no_flip_branch(block)) {
duke@435:       // Find fall through case - if must fall into its target
adlertz@5635:       int branch_idx = block->number_of_nodes() - block->_num_succs;
adlertz@5539:       for (uint j2 = 0; j2 < block->_num_succs; j2++) {
adlertz@5635:         const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj();
duke@435:         if (p->_con == 0) {
duke@435:           // successor j2 is fall through case
adlertz@5539:           if (block->non_connector_successor(j2) != bnext) {
duke@435:             // but it is not the next block => insert a goto
duke@435:             insert_goto_at(i, j2);
duke@435:           }
duke@435:           // Put taken branch in slot 0
adlertz@5539:           if (j2 == 0 && block->_num_succs == 2) {
duke@435:             // Flip targets in succs map
adlertz@5539:             Block *tbs0 = block->_succs[0];
adlertz@5539:             Block *tbs1 = block->_succs[1];
adlertz@5539:             block->_succs.map(0, tbs1);
adlertz@5539:             block->_succs.map(1, tbs0);
duke@435:           }
duke@435:           break;
duke@435:         }
duke@435:       }
adlertz@5539: 
duke@435:       // Remove all CatchProjs
adlertz@5539:       for (uint j = 0; j < block->_num_succs; j++) {
adlertz@5635:         block->pop_node();
adlertz@5539:       }
duke@435: 
adlertz@5539:     } else if (block->_num_succs == 1) {
duke@435:       // Block ends in a Goto?
duke@435:       if (bnext == bs0) {
duke@435:         // We fall into next block; remove the Goto
adlertz@5635:         block->pop_node();
duke@435:       }
duke@435: 
adlertz@5539:     } else if(block->_num_succs == 2) { // Block ends in a If?
duke@435:       // Get opcode of 1st projection (matches _succs[0])
duke@435:       // Note: Since this basic block has 2 exits, the last 2 nodes must
duke@435:       //       be projections (in any order), the 3rd last node must be
duke@435:       //       the IfNode (we have excluded other 2-way exits such as
duke@435:       //       CatchNodes already).
adlertz@5635:       MachNode* iff   = block->get_node(block->number_of_nodes() - 3)->as_Mach();
adlertz@5635:       ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj();
adlertz@5635:       ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj();
duke@435: 
duke@435:       // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1].
adlertz@5539:       assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0");
adlertz@5539:       assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1");
duke@435: 
adlertz@5539:       Block* bs1 = block->non_connector_successor(1);
duke@435: 
duke@435:       // Check for neither successor block following the current
duke@435:       // block ending in a conditional. If so, move one of the
duke@435:       // successors after the current one, provided that the
duke@435:       // successor was previously unscheduled, but moveable
duke@435:       // (i.e., all paths to it involve a branch).
adlertz@5539:       if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) {
duke@435:         // Choose the more common successor based on the probability
duke@435:         // of the conditional branch.
adlertz@5539:         Block* bx = bs0;
adlertz@5539:         Block* by = bs1;
duke@435: 
duke@435:         // _prob is the probability of taking the true path. Make
duke@435:         // p the probability of taking successor #1.
duke@435:         float p = iff->as_MachIf()->_prob;
adlertz@5539:         if (proj0->Opcode() == Op_IfTrue) {
duke@435:           p = 1.0 - p;
duke@435:         }
duke@435: 
duke@435:         // Prefer successor #1 if p > 0.5
duke@435:         if (p > PROB_FAIR) {
duke@435:           bx = bs1;
duke@435:           by = bs0;
duke@435:         }
duke@435: 
duke@435:         // Attempt the more common successor first
rasbold@853:         if (move_to_next(bx, i)) {
duke@435:           bnext = bx;
rasbold@853:         } else if (move_to_next(by, i)) {
duke@435:           bnext = by;
duke@435:         }
duke@435:       }
duke@435: 
duke@435:       // Check for conditional branching the wrong way.  Negate
duke@435:       // conditional, if needed, so it falls into the following block
duke@435:       // and branches to the not-following block.
duke@435: 
duke@435:       // Check for the next block being in succs[0].  We are going to branch
duke@435:       // to succs[0], so we want the fall-thru case as the next block in
duke@435:       // succs[1].
duke@435:       if (bnext == bs0) {
duke@435:         // Fall-thru case in succs[0], so flip targets in succs map
adlertz@5539:         Block* tbs0 = block->_succs[0];
adlertz@5539:         Block* tbs1 = block->_succs[1];
adlertz@5539:         block->_succs.map(0, tbs1);
adlertz@5539:         block->_succs.map(1, tbs0);
duke@435:         // Flip projection for each target
adlertz@5539:         ProjNode* tmp = proj0;
adlertz@5539:         proj0 = proj1;
adlertz@5539:         proj1 = tmp;
duke@435: 
adlertz@5539:       } else if(bnext != bs1) {
rasbold@853:         // Need a double-branch
duke@435:         // The existing conditional branch need not change.
duke@435:         // Add a unconditional branch to the false target.
duke@435:         // Alas, it must appear in its own block and adding a
duke@435:         // block this late in the game is complicated.  Sigh.
duke@435:         insert_goto_at(i, 1);
duke@435:       }
duke@435: 
duke@435:       // Make sure we TRUE branch to the target
adlertz@5539:       if (proj0->Opcode() == Op_IfFalse) {
kvn@3051:         iff->as_MachIf()->negate();
rasbold@853:       }
duke@435: 
adlertz@5635:       block->pop_node();          // Remove IfFalse & IfTrue projections
adlertz@5635:       block->pop_node();
duke@435: 
duke@435:     } else {
duke@435:       // Multi-exit block, e.g. a switch statement
duke@435:       // But we don't need to do anything here
duke@435:     }
duke@435:   } // End of for all blocks
duke@435: }
duke@435: 
duke@435: 
duke@435: #ifndef PRODUCT
duke@435: void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited  ) const {
duke@435:   const Node *x = end->is_block_proj();
duke@435:   assert( x, "not a CFG" );
duke@435: 
duke@435:   // Do not visit this block again
duke@435:   if( visited.test_set(x->_idx) ) return;
duke@435: 
duke@435:   // Skip through this block
duke@435:   const Node *p = x;
duke@435:   do {
duke@435:     p = p->in(0);               // Move control forward
duke@435:     assert( !p->is_block_proj() || p->is_Root(), "not a CFG" );
duke@435:   } while( !p->is_block_start() );
duke@435: 
duke@435:   // Recursively visit
adlertz@5509:   for (uint i = 1; i < p->req(); i++) {
adlertz@5509:     _dump_cfg(p->in(i), visited);
adlertz@5509:   }
duke@435: 
duke@435:   // Dump the block
adlertz@5509:   get_block_for_node(p)->dump(this);
duke@435: }
duke@435: 
duke@435: void PhaseCFG::dump( ) const {
adlertz@5539:   tty->print("\n--- CFG --- %d BBs\n", number_of_blocks());
adlertz@5509:   if (_blocks.size()) {        // Did we do basic-block layout?
adlertz@5539:     for (uint i = 0; i < number_of_blocks(); i++) {
adlertz@5539:       const Block* block = get_block(i);
adlertz@5539:       block->dump(this);
adlertz@5509:     }
duke@435:   } else {                      // Else do it with a DFS
adlertz@5509:     VectorSet visited(_block_arena);
duke@435:     _dump_cfg(_root,visited);
duke@435:   }
duke@435: }
duke@435: 
duke@435: void PhaseCFG::dump_headers() {
adlertz@5539:   for (uint i = 0; i < number_of_blocks(); i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5539:     if (block != NULL) {
adlertz@5539:       block->dump_head(this);
adlertz@5509:     }
duke@435:   }
duke@435: }
duke@435: 
adlertz@5539: void PhaseCFG::verify() const {
kvn@1001: #ifdef ASSERT
duke@435:   // Verify sane CFG
adlertz@5539:   for (uint i = 0; i < number_of_blocks(); i++) {
adlertz@5539:     Block* block = get_block(i);
adlertz@5635:     uint cnt = block->number_of_nodes();
duke@435:     uint j;
kvn@3311:     for (j = 0; j < cnt; j++)  {
adlertz@5635:       Node *n = block->get_node(j);
adlertz@5539:       assert(get_block_for_node(n) == block, "");
adlertz@5539:       if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) {
adlertz@5635:         assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block");
duke@435:       }
kvn@3311:       for (uint k = 0; k < n->req(); k++) {
kvn@1001:         Node *def = n->in(k);
kvn@3311:         if (def && def != n) {
adlertz@5509:           assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok");
kvn@1001:           // Verify that instructions in the block is in correct order.
kvn@1001:           // Uses must follow their definition if they are at the same block.
kvn@1001:           // Mostly done to check that MachSpillCopy nodes are placed correctly
kvn@1001:           // when CreateEx node is moved in build_ifg_physical().
adlertz@5539:           if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) &&
kvn@1001:               // See (+++) comment in reg_split.cpp
kvn@3311:               !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) {
kvn@1328:             bool is_loop = false;
kvn@1328:             if (n->is_Phi()) {
kvn@3311:               for (uint l = 1; l < def->req(); l++) {
kvn@1328:                 if (n == def->in(l)) {
kvn@1328:                   is_loop = true;
kvn@1328:                   break; // Some kind of loop
kvn@1328:                 }
kvn@1328:               }
kvn@1328:             }
adlertz@5539:             assert(is_loop || block->find_node(def) < j, "uses must follow definitions");
kvn@1036:           }
duke@435:         }
duke@435:       }
duke@435:     }
duke@435: 
adlertz@5539:     j = block->end_idx();
adlertz@5635:     Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj();
adlertz@5539:     assert(bp, "last instruction must be a block proj");
adlertz@5635:     assert(bp == block->get_node(j), "wrong number of successors for this block");
kvn@3311:     if (bp->is_Catch()) {
adlertz@5635:       while (block->get_node(--j)->is_MachProj()) {
adlertz@5539:         ;
adlertz@5539:       }
adlertz@5635:       assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call");
kvn@3311:     } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) {
adlertz@5539:       assert(block->_num_succs == 2, "Conditional branch must have two targets");
duke@435:     }
duke@435:   }
kvn@1001: #endif
duke@435: }
duke@435: #endif
duke@435: 
duke@435: UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) {
duke@435:   Copy::zero_to_bytes( _indices, sizeof(uint)*max );
duke@435: }
duke@435: 
duke@435: void UnionFind::extend( uint from_idx, uint to_idx ) {
duke@435:   _nesting.check();
duke@435:   if( from_idx >= _max ) {
duke@435:     uint size = 16;
duke@435:     while( size <= from_idx ) size <<=1;
duke@435:     _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size );
duke@435:     _max = size;
duke@435:   }
duke@435:   while( _cnt <= from_idx ) _indices[_cnt++] = 0;
duke@435:   _indices[from_idx] = to_idx;
duke@435: }
duke@435: 
duke@435: void UnionFind::reset( uint max ) {
duke@435:   assert( max <= max_uint, "Must fit within uint" );
duke@435:   // Force the Union-Find mapping to be at least this large
duke@435:   extend(max,0);
duke@435:   // Initialize to be the ID mapping.
rasbold@853:   for( uint i=0; i<max; i++ ) map(i,i);
duke@435: }
duke@435: 
duke@435: // Straight out of Tarjan's union-find algorithm
duke@435: uint UnionFind::Find_compress( uint idx ) {
duke@435:   uint cur  = idx;
duke@435:   uint next = lookup(cur);
duke@435:   while( next != cur ) {        // Scan chain of equivalences
duke@435:     assert( next < cur, "always union smaller" );
duke@435:     cur = next;                 // until find a fixed-point
duke@435:     next = lookup(cur);
duke@435:   }
duke@435:   // Core of union-find algorithm: update chain of
duke@435:   // equivalences to be equal to the root.
duke@435:   while( idx != next ) {
duke@435:     uint tmp = lookup(idx);
duke@435:     map(idx, next);
duke@435:     idx = tmp;
duke@435:   }
duke@435:   return idx;
duke@435: }
duke@435: 
duke@435: // Like Find above, but no path compress, so bad asymptotic behavior
duke@435: uint UnionFind::Find_const( uint idx ) const {
duke@435:   if( idx == 0 ) return idx;    // Ignore the zero idx
duke@435:   // Off the end?  This can happen during debugging dumps
duke@435:   // when data structures have not finished being updated.
duke@435:   if( idx >= _max ) return idx;
duke@435:   uint next = lookup(idx);
duke@435:   while( next != idx ) {        // Scan chain of equivalences
duke@435:     idx = next;                 // until find a fixed-point
duke@435:     next = lookup(idx);
duke@435:   }
duke@435:   return next;
duke@435: }
duke@435: 
duke@435: // union 2 sets together.
duke@435: void UnionFind::Union( uint idx1, uint idx2 ) {
duke@435:   uint src = Find(idx1);
duke@435:   uint dst = Find(idx2);
duke@435:   assert( src, "" );
duke@435:   assert( dst, "" );
duke@435:   assert( src < _max, "oob" );
duke@435:   assert( dst < _max, "oob" );
duke@435:   assert( src < dst, "always union smaller" );
duke@435:   map(dst,src);
duke@435: }
rasbold@853: 
rasbold@853: #ifndef PRODUCT
rasbold@853: void Trace::dump( ) const {
rasbold@853:   tty->print_cr("Trace (freq %f)", first_block()->_freq);
rasbold@853:   for (Block *b = first_block(); b != NULL; b = next(b)) {
rasbold@853:     tty->print("  B%d", b->_pre_order);
rasbold@853:     if (b->head()->is_Loop()) {
rasbold@853:       tty->print(" (L%d)", b->compute_loop_alignment());
rasbold@853:     }
rasbold@853:     if (b->has_loop_alignment()) {
rasbold@853:       tty->print(" (T%d)", b->code_alignment());
rasbold@853:     }
rasbold@853:   }
rasbold@853:   tty->cr();
rasbold@853: }
rasbold@853: 
rasbold@853: void CFGEdge::dump( ) const {
rasbold@853:   tty->print(" B%d  -->  B%d  Freq: %f  out:%3d%%  in:%3d%%  State: ",
rasbold@853:              from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct);
rasbold@853:   switch(state()) {
rasbold@853:   case connected:
rasbold@853:     tty->print("connected");
rasbold@853:     break;
rasbold@853:   case open:
rasbold@853:     tty->print("open");
rasbold@853:     break;
rasbold@853:   case interior:
rasbold@853:     tty->print("interior");
rasbold@853:     break;
rasbold@853:   }
rasbold@853:   if (infrequent()) {
rasbold@853:     tty->print("  infrequent");
rasbold@853:   }
rasbold@853:   tty->cr();
rasbold@853: }
rasbold@853: #endif
rasbold@853: 
rasbold@853: // Comparison function for edges
rasbold@853: static int edge_order(CFGEdge **e0, CFGEdge **e1) {
rasbold@853:   float freq0 = (*e0)->freq();
rasbold@853:   float freq1 = (*e1)->freq();
rasbold@853:   if (freq0 != freq1) {
rasbold@853:     return freq0 > freq1 ? -1 : 1;
rasbold@853:   }
rasbold@853: 
rasbold@853:   int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo;
rasbold@853:   int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo;
rasbold@853: 
rasbold@853:   return dist1 - dist0;
rasbold@853: }
rasbold@853: 
rasbold@853: // Comparison function for edges
kvn@3128: extern "C" int trace_frequency_order(const void *p0, const void *p1) {
rasbold@853:   Trace *tr0 = *(Trace **) p0;
rasbold@853:   Trace *tr1 = *(Trace **) p1;
rasbold@853:   Block *b0 = tr0->first_block();
rasbold@853:   Block *b1 = tr1->first_block();
rasbold@853: 
rasbold@853:   // The trace of connector blocks goes at the end;
rasbold@853:   // we only expect one such trace
rasbold@853:   if (b0->is_connector() != b1->is_connector()) {
rasbold@853:     return b1->is_connector() ? -1 : 1;
rasbold@853:   }
rasbold@853: 
rasbold@853:   // Pull more frequently executed blocks to the beginning
rasbold@853:   float freq0 = b0->_freq;
rasbold@853:   float freq1 = b1->_freq;
rasbold@853:   if (freq0 != freq1) {
rasbold@853:     return freq0 > freq1 ? -1 : 1;
rasbold@853:   }
rasbold@853: 
rasbold@853:   int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo;
rasbold@853: 
rasbold@853:   return diff;
rasbold@853: }
rasbold@853: 
rasbold@853: // Find edges of interest, i.e, those which can fall through. Presumes that
rasbold@853: // edges which don't fall through are of low frequency and can be generally
rasbold@853: // ignored.  Initialize the list of traces.
adlertz@5539: void PhaseBlockLayout::find_edges() {
rasbold@853:   // Walk the blocks, creating edges and Traces
rasbold@853:   uint i;
rasbold@853:   Trace *tr = NULL;
adlertz@5539:   for (i = 0; i < _cfg.number_of_blocks(); i++) {
adlertz@5539:     Block* b = _cfg.get_block(i);
rasbold@853:     tr = new Trace(b, next, prev);
rasbold@853:     traces[tr->id()] = tr;
rasbold@853: 
rasbold@853:     // All connector blocks should be at the end of the list
rasbold@853:     if (b->is_connector()) break;
rasbold@853: 
rasbold@853:     // If this block and the next one have a one-to-one successor
rasbold@853:     // predecessor relationship, simply append the next block
rasbold@853:     int nfallthru = b->num_fall_throughs();
rasbold@853:     while (nfallthru == 1 &&
rasbold@853:            b->succ_fall_through(0)) {
rasbold@853:       Block *n = b->_succs[0];
rasbold@853: 
rasbold@853:       // Skip over single-entry connector blocks, we don't want to
rasbold@853:       // add them to the trace.
rasbold@853:       while (n->is_connector() && n->num_preds() == 1) {
rasbold@853:         n = n->_succs[0];
rasbold@853:       }
rasbold@853: 
rasbold@853:       // We see a merge point, so stop search for the next block
rasbold@853:       if (n->num_preds() != 1) break;
rasbold@853: 
rasbold@853:       i++;
adlertz@5539:       assert(n = _cfg.get_block(i), "expecting next block");
rasbold@853:       tr->append(n);
rasbold@853:       uf->map(n->_pre_order, tr->id());
rasbold@853:       traces[n->_pre_order] = NULL;
rasbold@853:       nfallthru = b->num_fall_throughs();
rasbold@853:       b = n;
rasbold@853:     }
rasbold@853: 
rasbold@853:     if (nfallthru > 0) {
rasbold@853:       // Create a CFGEdge for each outgoing
rasbold@853:       // edge that could be a fall-through.
rasbold@853:       for (uint j = 0; j < b->_num_succs; j++ ) {
rasbold@853:         if (b->succ_fall_through(j)) {
rasbold@853:           Block *target = b->non_connector_successor(j);
rasbold@853:           float freq = b->_freq * b->succ_prob(j);
rasbold@853:           int from_pct = (int) ((100 * freq) / b->_freq);
rasbold@853:           int to_pct = (int) ((100 * freq) / target->_freq);
rasbold@853:           edges->append(new CFGEdge(b, target, freq, from_pct, to_pct));
rasbold@853:         }
rasbold@853:       }
rasbold@853:     }
rasbold@853:   }
rasbold@853: 
rasbold@853:   // Group connector blocks into one trace
adlertz@5539:   for (i++; i < _cfg.number_of_blocks(); i++) {
adlertz@5539:     Block *b = _cfg.get_block(i);
rasbold@853:     assert(b->is_connector(), "connector blocks at the end");
rasbold@853:     tr->append(b);
rasbold@853:     uf->map(b->_pre_order, tr->id());
rasbold@853:     traces[b->_pre_order] = NULL;
rasbold@853:   }
rasbold@853: }
rasbold@853: 
rasbold@853: // Union two traces together in uf, and null out the trace in the list
adlertz@5539: void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) {
rasbold@853:   uint old_id = old_trace->id();
rasbold@853:   uint updated_id = updated_trace->id();
rasbold@853: 
rasbold@853:   uint lo_id = updated_id;
rasbold@853:   uint hi_id = old_id;
rasbold@853: 
rasbold@853:   // If from is greater than to, swap values to meet
rasbold@853:   // UnionFind guarantee.
rasbold@853:   if (updated_id > old_id) {
rasbold@853:     lo_id = old_id;
rasbold@853:     hi_id = updated_id;
rasbold@853: 
rasbold@853:     // Fix up the trace ids
rasbold@853:     traces[lo_id] = traces[updated_id];
rasbold@853:     updated_trace->set_id(lo_id);
rasbold@853:   }
rasbold@853: 
rasbold@853:   // Union the lower with the higher and remove the pointer
rasbold@853:   // to the higher.
rasbold@853:   uf->Union(lo_id, hi_id);
rasbold@853:   traces[hi_id] = NULL;
rasbold@853: }
rasbold@853: 
rasbold@853: // Append traces together via the most frequently executed edges
adlertz@5539: void PhaseBlockLayout::grow_traces() {
rasbold@853:   // Order the edges, and drive the growth of Traces via the most
rasbold@853:   // frequently executed edges.
rasbold@853:   edges->sort(edge_order);
rasbold@853:   for (int i = 0; i < edges->length(); i++) {
rasbold@853:     CFGEdge *e = edges->at(i);
rasbold@853: 
rasbold@853:     if (e->state() != CFGEdge::open) continue;
rasbold@853: 
rasbold@853:     Block *src_block = e->from();
rasbold@853:     Block *targ_block = e->to();
rasbold@853: 
rasbold@853:     // Don't grow traces along backedges?
rasbold@853:     if (!BlockLayoutRotateLoops) {
rasbold@853:       if (targ_block->_rpo <= src_block->_rpo) {
rasbold@853:         targ_block->set_loop_alignment(targ_block);
rasbold@853:         continue;
rasbold@853:       }
rasbold@853:     }
rasbold@853: 
rasbold@853:     Trace *src_trace = trace(src_block);
rasbold@853:     Trace *targ_trace = trace(targ_block);
rasbold@853: 
rasbold@853:     // If the edge in question can join two traces at their ends,
rasbold@853:     // append one trace to the other.
rasbold@853:    if (src_trace->last_block() == src_block) {
rasbold@853:       if (src_trace == targ_trace) {
rasbold@853:         e->set_state(CFGEdge::interior);
rasbold@853:         if (targ_trace->backedge(e)) {
rasbold@853:           // Reset i to catch any newly eligible edge
rasbold@853:           // (Or we could remember the first "open" edge, and reset there)
rasbold@853:           i = 0;
rasbold@853:         }
rasbold@853:       } else if (targ_trace->first_block() == targ_block) {
rasbold@853:         e->set_state(CFGEdge::connected);
rasbold@853:         src_trace->append(targ_trace);
rasbold@853:         union_traces(src_trace, targ_trace);
rasbold@853:       }
rasbold@853:     }
rasbold@853:   }
rasbold@853: }
rasbold@853: 
rasbold@853: // Embed one trace into another, if the fork or join points are sufficiently
rasbold@853: // balanced.
adlertz@5539: void PhaseBlockLayout::merge_traces(bool fall_thru_only) {
rasbold@853:   // Walk the edge list a another time, looking at unprocessed edges.
rasbold@853:   // Fold in diamonds
rasbold@853:   for (int i = 0; i < edges->length(); i++) {
rasbold@853:     CFGEdge *e = edges->at(i);
rasbold@853: 
rasbold@853:     if (e->state() != CFGEdge::open) continue;
rasbold@853:     if (fall_thru_only) {
rasbold@853:       if (e->infrequent()) continue;
rasbold@853:     }
rasbold@853: 
rasbold@853:     Block *src_block = e->from();
rasbold@853:     Trace *src_trace = trace(src_block);
rasbold@853:     bool src_at_tail = src_trace->last_block() == src_block;
rasbold@853: 
rasbold@853:     Block *targ_block  = e->to();
rasbold@853:     Trace *targ_trace  = trace(targ_block);
rasbold@853:     bool targ_at_start = targ_trace->first_block() == targ_block;
rasbold@853: 
rasbold@853:     if (src_trace == targ_trace) {
rasbold@853:       // This may be a loop, but we can't do much about it.
rasbold@853:       e->set_state(CFGEdge::interior);
rasbold@853:       continue;
rasbold@853:     }
rasbold@853: 
rasbold@853:     if (fall_thru_only) {
rasbold@853:       // If the edge links the middle of two traces, we can't do anything.
rasbold@853:       // Mark the edge and continue.
rasbold@853:       if (!src_at_tail & !targ_at_start) {
rasbold@853:         continue;
rasbold@853:       }
rasbold@853: 
rasbold@853:       // Don't grow traces along backedges?
rasbold@853:       if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) {
rasbold@853:           continue;
rasbold@853:       }
rasbold@853: 
rasbold@853:       // If both ends of the edge are available, why didn't we handle it earlier?
rasbold@853:       assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier.");
rasbold@853: 
rasbold@853:       if (targ_at_start) {
rasbold@853:         // Insert the "targ" trace in the "src" trace if the insertion point
rasbold@853:         // is a two way branch.
rasbold@853:         // Better profitability check possible, but may not be worth it.
rasbold@853:         // Someday, see if the this "fork" has an associated "join";
rasbold@853:         // then make a policy on merging this trace at the fork or join.
rasbold@853:         // For example, other things being equal, it may be better to place this
rasbold@853:         // trace at the join point if the "src" trace ends in a two-way, but
rasbold@853:         // the insertion point is one-way.
rasbold@853:         assert(src_block->num_fall_throughs() == 2, "unexpected diamond");
rasbold@853:         e->set_state(CFGEdge::connected);
rasbold@853:         src_trace->insert_after(src_block, targ_trace);
rasbold@853:         union_traces(src_trace, targ_trace);
rasbold@853:       } else if (src_at_tail) {
adlertz@5539:         if (src_trace != trace(_cfg.get_root_block())) {
rasbold@853:           e->set_state(CFGEdge::connected);
rasbold@853:           targ_trace->insert_before(targ_block, src_trace);
rasbold@853:           union_traces(targ_trace, src_trace);
rasbold@853:         }
rasbold@853:       }
rasbold@853:     } else if (e->state() == CFGEdge::open) {
rasbold@853:       // Append traces, even without a fall-thru connection.
twisti@1040:       // But leave root entry at the beginning of the block list.
adlertz@5539:       if (targ_trace != trace(_cfg.get_root_block())) {
rasbold@853:         e->set_state(CFGEdge::connected);
rasbold@853:         src_trace->append(targ_trace);
rasbold@853:         union_traces(src_trace, targ_trace);
rasbold@853:       }
rasbold@853:     }
rasbold@853:   }
rasbold@853: }
rasbold@853: 
rasbold@853: // Order the sequence of the traces in some desirable way, and fixup the
rasbold@853: // jumps at the end of each block.
adlertz@5539: void PhaseBlockLayout::reorder_traces(int count) {
rasbold@853:   ResourceArea *area = Thread::current()->resource_area();
rasbold@853:   Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count);
rasbold@853:   Block_List worklist;
rasbold@853:   int new_count = 0;
rasbold@853: 
rasbold@853:   // Compact the traces.
rasbold@853:   for (int i = 0; i < count; i++) {
rasbold@853:     Trace *tr = traces[i];
rasbold@853:     if (tr != NULL) {
rasbold@853:       new_traces[new_count++] = tr;
rasbold@853:     }
rasbold@853:   }
rasbold@853: 
rasbold@853:   // The entry block should be first on the new trace list.
adlertz@5539:   Trace *tr = trace(_cfg.get_root_block());
rasbold@853:   assert(tr == new_traces[0], "entry trace misplaced");
rasbold@853: 
rasbold@853:   // Sort the new trace list by frequency
rasbold@853:   qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order);
rasbold@853: 
rasbold@853:   // Patch up the successor blocks
adlertz@5539:   _cfg.clear_blocks();
rasbold@853:   for (int i = 0; i < new_count; i++) {
rasbold@853:     Trace *tr = new_traces[i];
rasbold@853:     if (tr != NULL) {
rasbold@853:       tr->fixup_blocks(_cfg);
rasbold@853:     }
rasbold@853:   }
rasbold@853: }
rasbold@853: 
rasbold@853: // Order basic blocks based on frequency
adlertz@5539: PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg)
adlertz@5539: : Phase(BlockLayout)
adlertz@5539: , _cfg(cfg) {
rasbold@853:   ResourceMark rm;
rasbold@853:   ResourceArea *area = Thread::current()->resource_area();
rasbold@853: 
rasbold@853:   // List of traces
adlertz@5539:   int size = _cfg.number_of_blocks() + 1;
rasbold@853:   traces = NEW_ARENA_ARRAY(area, Trace *, size);
rasbold@853:   memset(traces, 0, size*sizeof(Trace*));
rasbold@853:   next = NEW_ARENA_ARRAY(area, Block *, size);
rasbold@853:   memset(next,   0, size*sizeof(Block *));
rasbold@853:   prev = NEW_ARENA_ARRAY(area, Block *, size);
rasbold@853:   memset(prev  , 0, size*sizeof(Block *));
rasbold@853: 
rasbold@853:   // List of edges
rasbold@853:   edges = new GrowableArray<CFGEdge*>;
rasbold@853: 
rasbold@853:   // Mapping block index --> block_trace
rasbold@853:   uf = new UnionFind(size);
rasbold@853:   uf->reset(size);
rasbold@853: 
rasbold@853:   // Find edges and create traces.
rasbold@853:   find_edges();
rasbold@853: 
rasbold@853:   // Grow traces at their ends via most frequent edges.
rasbold@853:   grow_traces();
rasbold@853: 
rasbold@853:   // Merge one trace into another, but only at fall-through points.
rasbold@853:   // This may make diamonds and other related shapes in a trace.
rasbold@853:   merge_traces(true);
rasbold@853: 
rasbold@853:   // Run merge again, allowing two traces to be catenated, even if
rasbold@853:   // one does not fall through into the other. This appends loosely
rasbold@853:   // related traces to be near each other.
rasbold@853:   merge_traces(false);
rasbold@853: 
rasbold@853:   // Re-order all the remaining traces by frequency
rasbold@853:   reorder_traces(size);
rasbold@853: 
adlertz@5539:   assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink");
rasbold@853: }
rasbold@853: 
rasbold@853: 
rasbold@853: // Edge e completes a loop in a trace. If the target block is head of the
rasbold@853: // loop, rotate the loop block so that the loop ends in a conditional branch.
rasbold@853: bool Trace::backedge(CFGEdge *e) {
rasbold@853:   bool loop_rotated = false;
rasbold@853:   Block *src_block  = e->from();
rasbold@853:   Block *targ_block    = e->to();
rasbold@853: 
rasbold@853:   assert(last_block() == src_block, "loop discovery at back branch");
rasbold@853:   if (first_block() == targ_block) {
rasbold@853:     if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) {
rasbold@853:       // Find the last block in the trace that has a conditional
rasbold@853:       // branch.
rasbold@853:       Block *b;
rasbold@853:       for (b = last_block(); b != NULL; b = prev(b)) {
rasbold@853:         if (b->num_fall_throughs() == 2) {
rasbold@853:           break;
rasbold@853:         }
rasbold@853:       }
rasbold@853: 
rasbold@853:       if (b != last_block() && b != NULL) {
rasbold@853:         loop_rotated = true;
rasbold@853: 
rasbold@853:         // Rotate the loop by doing two-part linked-list surgery.
rasbold@853:         append(first_block());
rasbold@853:         break_loop_after(b);
rasbold@853:       }
rasbold@853:     }
rasbold@853: 
rasbold@853:     // Backbranch to the top of a trace
twisti@1040:     // Scroll forward through the trace from the targ_block. If we find
rasbold@853:     // a loop head before another loop top, use the the loop head alignment.
rasbold@853:     for (Block *b = targ_block; b != NULL; b = next(b)) {
rasbold@853:       if (b->has_loop_alignment()) {
rasbold@853:         break;
rasbold@853:       }
rasbold@853:       if (b->head()->is_Loop()) {
rasbold@853:         targ_block = b;
rasbold@853:         break;
rasbold@853:       }
rasbold@853:     }
rasbold@853: 
rasbold@853:     first_block()->set_loop_alignment(targ_block);
rasbold@853: 
rasbold@853:   } else {
rasbold@853:     // Backbranch into the middle of a trace
rasbold@853:     targ_block->set_loop_alignment(targ_block);
rasbold@853:   }
rasbold@853: 
rasbold@853:   return loop_rotated;
rasbold@853: }
rasbold@853: 
rasbold@853: // push blocks onto the CFG list
rasbold@853: // ensure that blocks have the correct two-way branch sense
rasbold@853: void Trace::fixup_blocks(PhaseCFG &cfg) {
rasbold@853:   Block *last = last_block();
rasbold@853:   for (Block *b = first_block(); b != NULL; b = next(b)) {
adlertz@5539:     cfg.add_block(b);
rasbold@853:     if (!b->is_connector()) {
rasbold@853:       int nfallthru = b->num_fall_throughs();
rasbold@853:       if (b != last) {
rasbold@853:         if (nfallthru == 2) {
rasbold@853:           // Ensure that the sense of the branch is correct
rasbold@853:           Block *bnext = next(b);
rasbold@853:           Block *bs0 = b->non_connector_successor(0);
rasbold@853: 
adlertz@5635:           MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach();
adlertz@5635:           ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj();
adlertz@5635:           ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj();
rasbold@853: 
rasbold@853:           if (bnext == bs0) {
rasbold@853:             // Fall-thru case in succs[0], should be in succs[1]
rasbold@853: 
rasbold@853:             // Flip targets in _succs map
rasbold@853:             Block *tbs0 = b->_succs[0];
rasbold@853:             Block *tbs1 = b->_succs[1];
rasbold@853:             b->_succs.map( 0, tbs1 );
rasbold@853:             b->_succs.map( 1, tbs0 );
rasbold@853: 
rasbold@853:             // Flip projections to match targets
adlertz@5635:             b->map_node(proj1, b->number_of_nodes() - 2);
adlertz@5635:             b->map_node(proj0, b->number_of_nodes() - 1);
rasbold@853:           }
rasbold@853:         }
rasbold@853:       }
rasbold@853:     }
rasbold@853:   }
rasbold@853: }