/*

 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.

 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

 *

 * This code is free software; you can redistribute it and/or modify it

 * under the terms of the GNU General Public License version 2 only, as

 * published by the Free Software Foundation.

 *

 * This code is distributed in the hope that it will be useful, but WITHOUT

 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

 * version 2 for more details (a copy is included in the LICENSE file that

 * accompanied this code).

 *

 * You should have received a copy of the GNU General Public License version

 * 2 along with this work; if not, write to the Free Software Foundation,

 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

 *

 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA

 * or visit www.oracle.com if you need additional information or have any

 * questions.

 *

 */

#include "precompiled.hpp"

#include "memory/allocation.inline.hpp"

#include "opto/block.hpp"

#include "opto/cfgnode.hpp"

#include "opto/chaitin.hpp"

#include "opto/coalesce.hpp"

#include "opto/connode.hpp"

#include "opto/indexSet.hpp"

#include "opto/machnode.hpp"

#include "opto/matcher.hpp"

#include "opto/regmask.hpp"

//=============================================================================

//------------------------------Dump-------------------------------------------

#ifndef PRODUCT

void PhaseCoalesce::dump(Node *n) const {

  // Being a const function means I cannot use 'Find'

  uint r = _phc._lrg_map.find(n);

  tty->print("L%d/N%d ",r,n->_idx);

}

//------------------------------dump-------------------------------------------

void PhaseCoalesce::dump() const {

  // I know I have a block layout now, so I can print blocks in a loop

  for( uint i=0; i<_phc._cfg._num_blocks; i++ ) {

    uint j;

    Block *b = _phc._cfg._blocks[i];

    // Print a nice block header

    tty->print("B%d: ",b->_pre_order);

    for( j=1; j<b->num_preds(); j++ )

      tty->print("B%d ", _phc._cfg._bbs[b->pred(j)->_idx]->_pre_order);

    tty->print("-> ");

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

      tty->print("B%d ",b->_succs[j]->_pre_order);

    tty->print(" IDom: B%d/#%d\n", b->_idom ? b->_idom->_pre_order : 0, b->_dom_depth);

    uint cnt = b->_nodes.size();

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

      Node *n = b->_nodes[j];

      dump( n );

      tty->print("\t%s\t",n->Name());

      // Dump the inputs

      uint k;                   // Exit value of loop

      for( k=0; k<n->req(); k++ ) // For all required inputs

        if( n->in(k) ) dump( n->in(k) );

        else tty->print("_ ");

      int any_prec = 0;

      for( ; k<n->len(); k++ )          // For all precedence inputs

        if( n->in(k) ) {

          if( !any_prec++ ) tty->print(" |");

          dump( n->in(k) );

        }

      // Dump node-specific info

      n->dump_spec(tty);

      tty->print("\n");

    }

    tty->print("\n");

  }

}

#endif

//------------------------------combine_these_two------------------------------

// Combine the live ranges def'd by these 2 Nodes.  N2 is an input to N1.

void PhaseCoalesce::combine_these_two(Node *n1, Node *n2) {

  uint lr1 = _phc._lrg_map.find(n1);

  uint lr2 = _phc._lrg_map.find(n2);

  if( lr1 != lr2 &&             // Different live ranges already AND

      !_phc._ifg->test_edge_sq( lr1, lr2 ) ) {  // Do not interfere

    LRG *lrg1 = &_phc.lrgs(lr1);

    LRG *lrg2 = &_phc.lrgs(lr2);

    // Not an oop->int cast; oop->oop, int->int, AND int->oop are OK.

    // Now, why is int->oop OK?  We end up declaring a raw-pointer as an oop

    // and in general that's a bad thing.  However, int->oop conversions only

    // happen at GC points, so the lifetime of the misclassified raw-pointer

    // is from the CheckCastPP (that converts it to an oop) backwards up

    // through a merge point and into the slow-path call, and around the

    // diamond up to the heap-top check and back down into the slow-path call.

    // The misclassified raw pointer is NOT live across the slow-path call,

    // and so does not appear in any GC info, so the fact that it is

    // misclassified is OK.

    if( (lrg1->_is_oop || !lrg2->_is_oop) && // not an oop->int cast AND

        // Compatible final mask

        lrg1->mask().overlap( lrg2->mask() ) ) {

      // Merge larger into smaller.

      if( lr1 > lr2 ) {

        uint  tmp =  lr1;  lr1 =  lr2;  lr2 =  tmp;

        Node   *n =   n1;   n1 =   n2;   n2 =    n;

        LRG *ltmp = lrg1; lrg1 = lrg2; lrg2 = ltmp;

      }

      // Union lr2 into lr1

      _phc.Union( n1, n2 );

      if (lrg1->_maxfreq < lrg2->_maxfreq)

        lrg1->_maxfreq = lrg2->_maxfreq;

      // Merge in the IFG

      _phc._ifg->Union( lr1, lr2 );

      // Combine register restrictions

      lrg1->AND(lrg2->mask());

    }

  }

}

//------------------------------coalesce_driver--------------------------------

// Copy coalescing

void PhaseCoalesce::coalesce_driver( ) {

  verify();

  // Coalesce from high frequency to low

  for( uint i=0; i<_phc._cfg._num_blocks; i++ )

    coalesce( _phc._blks[i] );

}

//------------------------------insert_copy_with_overlap-----------------------

// I am inserting copies to come out of SSA form.  In the general case, I am

// doing a parallel renaming.  I'm in the Named world now, so I can't do a

// general parallel renaming.  All the copies now use  "names" (live-ranges)

// to carry values instead of the explicit use-def chains.  Suppose I need to

// insert 2 copies into the same block.  They copy L161->L128 and L128->L132.

// If I insert them in the wrong order then L128 will get clobbered before it

// can get used by the second copy.  This cannot happen in the SSA model;

// direct use-def chains get me the right value.  It DOES happen in the named

// model so I have to handle the reordering of copies.

//

// In general, I need to topo-sort the placed copies to avoid conflicts.

// Its possible to have a closed cycle of copies (e.g., recirculating the same

// values around a loop).  In this case I need a temp to break the cycle.

void PhaseAggressiveCoalesce::insert_copy_with_overlap( Block *b, Node *copy, uint dst_name, uint src_name ) {

  // Scan backwards for the locations of the last use of the dst_name.

  // I am about to clobber the dst_name, so the copy must be inserted

  // after the last use.  Last use is really first-use on a backwards scan.

  uint i = b->end_idx()-1;

  while(1) {

    Node *n = b->_nodes[i];

    // Check for end of virtual copies; this is also the end of the

    // parallel renaming effort.

    if (n->_idx < _unique) {

      break;

    }

    uint idx = n->is_Copy();

    assert( idx || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );

    if (idx && _phc._lrg_map.find(n->in(idx)) == dst_name) {

      break;

    }

    i--;

  }

  uint last_use_idx = i;

  // Also search for any kill of src_name that exits the block.

  // Since the copy uses src_name, I have to come before any kill.

  uint kill_src_idx = b->end_idx();

  // There can be only 1 kill that exits any block and that is

  // the last kill.  Thus it is the first kill on a backwards scan.

  i = b->end_idx()-1;

  while (1) {

    Node *n = b->_nodes[i];

    // Check for end of virtual copies; this is also the end of the

    // parallel renaming effort.

    if (n->_idx < _unique) {

      break;

    }

    assert( n->is_Copy() || n->is_Con() || n->is_MachProj(), "Only copies during parallel renaming" );

    if (_phc._lrg_map.find(n) == src_name) {

      kill_src_idx = i;

      break;

    }

    i--;

  }

  // Need a temp?  Last use of dst comes after the kill of src?

  if (last_use_idx >= kill_src_idx) {

    // Need to break a cycle with a temp

    uint idx = copy->is_Copy();

    Node *tmp = copy->clone();

    uint max_lrg_id = _phc._lrg_map.max_lrg_id();

    _phc.new_lrg(tmp, max_lrg_id);

    _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);

    // Insert new temp between copy and source

    tmp ->set_req(idx,copy->in(idx));

    copy->set_req(idx,tmp);

    // Save source in temp early, before source is killed

    b->_nodes.insert(kill_src_idx,tmp);

    _phc._cfg._bbs.map( tmp->_idx, b );

    last_use_idx++;

  }

  // Insert just after last use

  b->_nodes.insert(last_use_idx+1,copy);

}

//------------------------------insert_copies----------------------------------

void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) {

  // We do LRGs compressing and fix a liveout data only here since the other

  // place in Split() is guarded by the assert which we never hit.

  _phc._lrg_map.compress_uf_map_for_nodes();

  // Fix block's liveout data for compressed live ranges.

  for (uint lrg = 1; lrg < _phc._lrg_map.max_lrg_id(); lrg++) {

    uint compressed_lrg = _phc._lrg_map.find(lrg);

    if (lrg != compressed_lrg) {

      for (uint bidx = 0; bidx < _phc._cfg._num_blocks; bidx++) {

        IndexSet *liveout = _phc._live->live(_phc._cfg._blocks[bidx]);

        if (liveout->member(lrg)) {

          liveout->remove(lrg);

          liveout->insert(compressed_lrg);

        }

      }

    }

  }

  // All new nodes added are actual copies to replace virtual copies.

  // Nodes with index less than '_unique' are original, non-virtual Nodes.

  _unique = C->unique();

  for( uint i=0; i<_phc._cfg._num_blocks; i++ ) {

    C->check_node_count(NodeLimitFudgeFactor, "out of nodes in coalesce");

    if (C->failing()) return;

    Block *b = _phc._cfg._blocks[i];

    uint cnt = b->num_preds();  // Number of inputs to the Phi

    for( uint l = 1; l<b->_nodes.size(); l++ ) {

      Node *n = b->_nodes[l];

      // Do not use removed-copies, use copied value instead

      uint ncnt = n->req();

      for( uint k = 1; k<ncnt; k++ ) {

        Node *copy = n->in(k);

        uint cidx = copy->is_Copy();

        if( cidx ) {

          Node *def = copy->in(cidx);

          if (_phc._lrg_map.find(copy) == _phc._lrg_map.find(def)) {

            n->set_req(k, def);

          }

        }

      }

      // Remove any explicit copies that get coalesced.

      uint cidx = n->is_Copy();

      if( cidx ) {

        Node *def = n->in(cidx);

        if (_phc._lrg_map.find(n) == _phc._lrg_map.find(def)) {

          n->replace_by(def);

          n->set_req(cidx,NULL);

          b->_nodes.remove(l);

          l--;

          continue;

        }

      }

      if (n->is_Phi()) {

        // Get the chosen name for the Phi

        uint phi_name = _phc._lrg_map.find(n);

        // Ignore the pre-allocated specials

        if (!phi_name) {

          continue;

        }

        // Check for mismatch inputs to Phi

        for (uint j = 1; j < cnt; j++) {

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

          uint src_name = _phc._lrg_map.find(m);

          if (src_name != phi_name) {

            Block *pred = _phc._cfg._bbs[b->pred(j)->_idx];

            Node *copy;

            assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");

            // Rematerialize constants instead of copying them

            if( m->is_Mach() && m->as_Mach()->is_Con() &&

                m->as_Mach()->rematerialize() ) {

              copy = m->clone();

              // Insert the copy in the predecessor basic block

              pred->add_inst(copy);

              // Copy any flags as well

              _phc.clone_projs(pred, pred->end_idx(), m, copy, _phc._lrg_map);

            } else {

              const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];

              copy = new (C) MachSpillCopyNode(m, *rm, *rm);

              // Find a good place to insert.  Kinda tricky, use a subroutine

              insert_copy_with_overlap(pred,copy,phi_name,src_name);

            }

            // Insert the copy in the use-def chain

            n->set_req(j, copy);

            _phc._cfg._bbs.map( copy->_idx, pred );

            // Extend ("register allocate") the names array for the copy.

            _phc._lrg_map.extend(copy->_idx, phi_name);

          } // End of if Phi names do not match

        } // End of for all inputs to Phi

      } else { // End of if Phi

        // Now check for 2-address instructions

        uint idx;

        if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) {

          // Get the chosen name for the Node

          uint name = _phc._lrg_map.find(n);

          assert (name, "no 2-address specials");

          // Check for name mis-match on the 2-address input

          Node *m = n->in(idx);

          if (_phc._lrg_map.find(m) != name) {

            Node *copy;

            assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach");

            // At this point it is unsafe to extend live ranges (6550579).

            // Rematerialize only constants as we do for Phi above.

            if(m->is_Mach() && m->as_Mach()->is_Con() &&

               m->as_Mach()->rematerialize()) {

              copy = m->clone();

              // Insert the copy in the basic block, just before us

              b->_nodes.insert(l++, copy);

              if(_phc.clone_projs(b, l, m, copy, _phc._lrg_map)) {

                l++;

              }

            } else {

              const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()];

              copy = new (C) MachSpillCopyNode(m, *rm, *rm);

              // Insert the copy in the basic block, just before us

              b->_nodes.insert(l++, copy);

            }

            // Insert the copy in the use-def chain

            n->set_req(idx, copy);

            // Extend ("register allocate") the names array for the copy.

            _phc._lrg_map.extend(copy->_idx, name);

            _phc._cfg._bbs.map( copy->_idx, b );

          }

        } // End of is two-adr

        // Insert a copy at a debug use for a lrg which has high frequency

        if (b->_freq < OPTO_DEBUG_SPLIT_FREQ || b->is_uncommon(_phc._cfg._bbs)) {

          // Walk the debug inputs to the node and check for lrg freq

          JVMState* jvms = n->jvms();

          uint debug_start = jvms ? jvms->debug_start() : 999999;

          uint debug_end   = jvms ? jvms->debug_end()   : 999999;

          for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) {

            // Do not split monitors; they are only needed for debug table

            // entries and need no code.

            if (jvms->is_monitor_use(inpidx)) {

              continue;

            }

            Node *inp = n->in(inpidx);

            uint nidx = _phc._lrg_map.live_range_id(inp);

            LRG &lrg = lrgs(nidx);

            // If this lrg has a high frequency use/def

            if( lrg._maxfreq >= _phc.high_frequency_lrg() ) {

              // If the live range is also live out of this block (like it

              // would be for a fast/slow idiom), the normal spill mechanism

              // does an excellent job.  If it is not live out of this block

              // (like it would be for debug info to uncommon trap) splitting

              // the live range now allows a better allocation in the high

              // frequency blocks.

              //   Build_IFG_virtual has converted the live sets to

              // live-IN info, not live-OUT info.

              uint k;

              for( k=0; k < b->_num_succs; k++ )

                if( _phc._live->live(b->_succs[k])->member( nidx ) )

                  break;      // Live in to some successor block?

              if( k < b->_num_succs )

                continue;     // Live out; do not pre-split

              // Split the lrg at this use

              const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()];

              Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm );

              // Insert the copy in the use-def chain

              n->set_req(inpidx, copy );

              // Insert the copy in the basic block, just before us

              b->_nodes.insert( l++, copy );

              // Extend ("register allocate") the names array for the copy.

              uint max_lrg_id = _phc._lrg_map.max_lrg_id();

              _phc.new_lrg(copy, max_lrg_id);

              _phc._lrg_map.set_max_lrg_id(max_lrg_id + 1);

              _phc._cfg._bbs.map(copy->_idx, b);

              //tty->print_cr("Split a debug use in Aggressive Coalesce");

            }  // End of if high frequency use/def

          }  // End of for all debug inputs

        }  // End of if low frequency safepoint

      } // End of if Phi

    } // End of for all instructions

  } // End of for all blocks

}

//=============================================================================

//------------------------------coalesce---------------------------------------

// Aggressive (but pessimistic) copy coalescing of a single block

// The following coalesce pass represents a single round of aggressive

// pessimistic coalesce.  "Aggressive" means no attempt to preserve

// colorability when coalescing.  This occasionally means more spills, but

// it also means fewer rounds of coalescing for better code - and that means

// faster compiles.

// "Pessimistic" means we do not hit the fixed point in one pass (and we are

// reaching for the least fixed point to boot).  This is typically solved

// with a few more rounds of coalescing, but the compiler must run fast.  We

// could optimistically coalescing everything touching PhiNodes together

// into one big live range, then check for self-interference.  Everywhere

// the live range interferes with self it would have to be split.  Finding

// the right split points can be done with some heuristics (based on

// expected frequency of edges in the live range).  In short, it's a real

// research problem and the timeline is too short to allow such research.

// Further thoughts: (1) build the LR in a pass, (2) find self-interference

// in another pass, (3) per each self-conflict, split, (4) split by finding

// the low-cost cut (min-cut) of the LR, (5) edges in the LR are weighted

// according to the GCM algorithm (or just exec freq on CFG edges).

void PhaseAggressiveCoalesce::coalesce( Block *b ) {

  // Copies are still "virtual" - meaning we have not made them explicitly

  // copies.  Instead, Phi functions of successor blocks have mis-matched

  // live-ranges.  If I fail to coalesce, I'll have to insert a copy to line

  // up the live-ranges.  Check for Phis in successor blocks.

  uint i;

  for( i=0; i<b->_num_succs; i++ ) {

    Block *bs = b->_succs[i];

    // Find index of 'b' in 'bs' predecessors

    uint j=1;

    while( _phc._cfg._bbs[bs->pred(j)->_idx] != b ) j++;

    // Visit all the Phis in successor block

    for( uint k = 1; k<bs->_nodes.size(); k++ ) {

      Node *n = bs->_nodes[k];

      if( !n->is_Phi() ) break;

      combine_these_two( n, n->in(j) );

    }

  } // End of for all successor blocks

  // Check _this_ block for 2-address instructions and copies.

  uint cnt = b->end_idx();

  for( i = 1; i<cnt; i++ ) {

    Node *n = b->_nodes[i];

    uint idx;

    // 2-address instructions have a virtual Copy matching their input

    // to their output

    if (n->is_Mach() && (idx = n->as_Mach()->two_adr())) {

      MachNode *mach = n->as_Mach();

      combine_these_two(mach, mach->in(idx));

    }

  } // End of for all instructions in block

}

//=============================================================================

//------------------------------PhaseConservativeCoalesce----------------------

PhaseConservativeCoalesce::PhaseConservativeCoalesce(PhaseChaitin &chaitin) : PhaseCoalesce(chaitin) {

  _ulr.initialize(_phc._lrg_map.max_lrg_id());

}

//------------------------------verify-----------------------------------------

void PhaseConservativeCoalesce::verify() {

#ifdef ASSERT

  _phc.set_was_low();

#endif

}

//------------------------------union_helper-----------------------------------

void PhaseConservativeCoalesce::union_helper( Node *lr1_node, Node *lr2_node, uint lr1, uint lr2, Node *src_def, Node *dst_copy, Node *src_copy, Block *b, uint bindex ) {

  // Join live ranges.  Merge larger into smaller.  Union lr2 into lr1 in the

  // union-find tree

  _phc.Union( lr1_node, lr2_node );

  // Single-def live range ONLY if both live ranges are single-def.

  // If both are single def, then src_def powers one live range

  // and def_copy powers the other.  After merging, src_def powers

  // the combined live range.

  lrgs(lr1)._def = (lrgs(lr1).is_multidef() ||

                        lrgs(lr2).is_multidef() )

    ? NodeSentinel : src_def;

  lrgs(lr2)._def = NULL;    // No def for lrg 2

  lrgs(lr2).Clear();        // Force empty mask for LRG 2

  //lrgs(lr2)._size = 0;      // Live-range 2 goes dead

  lrgs(lr1)._is_oop |= lrgs(lr2)._is_oop;

  lrgs(lr2)._is_oop = 0;    // In particular, not an oop for GC info

  if (lrgs(lr1)._maxfreq < lrgs(lr2)._maxfreq)

    lrgs(lr1)._maxfreq = lrgs(lr2)._maxfreq;

  // Copy original value instead.  Intermediate copies go dead, and

  // the dst_copy becomes useless.

  int didx = dst_copy->is_Copy();

  dst_copy->set_req( didx, src_def );

  // Add copy to free list

  // _phc.free_spillcopy(b->_nodes[bindex]);

  assert( b->_nodes[bindex] == dst_copy, "" );

  dst_copy->replace_by( dst_copy->in(didx) );

  dst_copy->set_req( didx, NULL);

  b->_nodes.remove(bindex);

  if( bindex < b->_ihrp_index ) b->_ihrp_index--;

  if( bindex < b->_fhrp_index ) b->_fhrp_index--;

  // Stretched lr1; add it to liveness of intermediate blocks

  Block *b2 = _phc._cfg._bbs[src_copy->_idx];

  while( b != b2 ) {

    b = _phc._cfg._bbs[b->pred(1)->_idx];

    _phc._live->live(b)->insert(lr1);