PR libstdc++/56278

[official-gcc.git] / gcc / tree-ssa-phiopt.c

/* Optimization of PHI nodes by converting them into straightline code.
   Copyright (C) 2004-2013 Free Software Foundation, Inc.

This file is part of GCC.

GCC is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by the
Free Software Foundation; either version 3, or (at your option) any
later version.

GCC 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
for more details.

You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3.  If not see
<http://www.gnu.org/licenses/>.  */

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "ggc.h"
#include "tree.h"
#include "flags.h"
#include "tm_p.h"
#include "basic-block.h"
#include "tree-flow.h"
#include "tree-pass.h"
#include "langhooks.h"
#include "pointer-set.h"
#include "domwalk.h"
#include "cfgloop.h"
#include "tree-data-ref.h"
#include "gimple-pretty-print.h"
#include "insn-config.h"
#include "expr.h"
#include "optabs.h"

#ifndef HAVE_conditional_move
#define HAVE_conditional_move (0)
#endif

static unsigned int tree_ssa_phiopt (void);
static unsigned int tree_ssa_phiopt_worker (bool, bool);
static bool conditional_replacement (basic_block, basic_block,
                                     edge, edge, gimple, tree, tree);
static int value_replacement (basic_block, basic_block,
                              edge, edge, gimple, tree, tree);
static bool minmax_replacement (basic_block, basic_block,
                                edge, edge, gimple, tree, tree);
static bool abs_replacement (basic_block, basic_block,
                             edge, edge, gimple, tree, tree);
static bool cond_store_replacement (basic_block, basic_block, edge, edge,
                                    struct pointer_set_t *);
static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
static struct pointer_set_t * get_non_trapping (void);
static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
static void hoist_adjacent_loads (basic_block, basic_block,
                                  basic_block, basic_block);
static bool gate_hoist_loads (void);

/* This pass tries to replaces an if-then-else block with an
   assignment.  We have four kinds of transformations.  Some of these
   transformations are also performed by the ifcvt RTL optimizer.

   Conditional Replacement
   -----------------------

   This transformation, implemented in conditional_replacement,
   replaces

     bb0:
      if (cond) goto bb2; else goto bb1;
     bb1:
     bb2:
      x = PHI <0 (bb1), 1 (bb0), ...>;

   with

     bb0:
      x' = cond;
      goto bb2;
     bb2:
      x = PHI <x' (bb0), ...>;

   We remove bb1 as it becomes unreachable.  This occurs often due to
   gimplification of conditionals.

   Value Replacement
   -----------------

   This transformation, implemented in value_replacement, replaces

     bb0:
       if (a != b) goto bb2; else goto bb1;
     bb1:
     bb2:
       x = PHI <a (bb1), b (bb0), ...>;

   with

     bb0:
     bb2:
       x = PHI <b (bb0), ...>;

   This opportunity can sometimes occur as a result of other
   optimizations.

   ABS Replacement
   ---------------

   This transformation, implemented in abs_replacement, replaces

     bb0:
       if (a >= 0) goto bb2; else goto bb1;
     bb1:
       x = -a;
     bb2:
       x = PHI <x (bb1), a (bb0), ...>;

   with

     bb0:
       x' = ABS_EXPR< a >;
     bb2:
       x = PHI <x' (bb0), ...>;

   MIN/MAX Replacement
   -------------------

   This transformation, minmax_replacement replaces

     bb0:
       if (a <= b) goto bb2; else goto bb1;
     bb1:
     bb2:
       x = PHI <b (bb1), a (bb0), ...>;

   with

     bb0:
       x' = MIN_EXPR (a, b)
     bb2:
       x = PHI <x' (bb0), ...>;

   A similar transformation is done for MAX_EXPR.


   This pass also performs a fifth transformation of a slightly different
   flavor.

   Adjacent Load Hoisting
   ----------------------
   
   This transformation replaces

     bb0:
       if (...) goto bb2; else goto bb1;
     bb1:
       x1 = (<expr>).field1;
       goto bb3;
     bb2:
       x2 = (<expr>).field2;
     bb3:
       # x = PHI <x1, x2>;

   with

     bb0:
       x1 = (<expr>).field1;
       x2 = (<expr>).field2;
       if (...) goto bb2; else goto bb1;
     bb1:
       goto bb3;
     bb2:
     bb3:
       # x = PHI <x1, x2>;

   The purpose of this transformation is to enable generation of conditional
   move instructions such as Intel CMOVE or PowerPC ISEL.  Because one of
   the loads is speculative, the transformation is restricted to very
   specific cases to avoid introducing a page fault.  We are looking for
   the common idiom:

     if (...)
       x = y->left;
     else
       x = y->right;

   where left and right are typically adjacent pointers in a tree structure.  */

static unsigned int
tree_ssa_phiopt (void)
{
  return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
}

/* This pass tries to transform conditional stores into unconditional
   ones, enabling further simplifications with the simpler then and else
   blocks.  In particular it replaces this:

     bb0:
       if (cond) goto bb2; else goto bb1;
     bb1:
       *p = RHS;
     bb2:

   with

     bb0:
       if (cond) goto bb1; else goto bb2;
     bb1:
       condtmp' = *p;
     bb2:
       condtmp = PHI <RHS, condtmp'>
       *p = condtmp;

   This transformation can only be done under several constraints,
   documented below.  It also replaces:

     bb0:
       if (cond) goto bb2; else goto bb1;
     bb1:
       *p = RHS1;
       goto bb3;
     bb2:
       *p = RHS2;
     bb3:

   with

     bb0:
       if (cond) goto bb3; else goto bb1;
     bb1:
     bb3:
       condtmp = PHI <RHS1, RHS2>
       *p = condtmp;  */

static unsigned int
tree_ssa_cs_elim (void)
{
  return tree_ssa_phiopt_worker (true, false);
}

/* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */

static gimple
single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
{
  gimple_stmt_iterator i;
  gimple phi = NULL;
  if (gimple_seq_singleton_p (seq))
    return gsi_stmt (gsi_start (seq));
  for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
    {
      gimple p = gsi_stmt (i);
      /* If the PHI arguments are equal then we can skip this PHI. */
      if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
                                       gimple_phi_arg_def (p, e1->dest_idx)))
        continue;

      /* If we already have a PHI that has the two edge arguments are
         different, then return it is not a singleton for these PHIs. */
      if (phi)
        return NULL;

      phi = p;
    }
  return phi;
}

/* The core routine of conditional store replacement and normal
   phi optimizations.  Both share much of the infrastructure in how
   to match applicable basic block patterns.  DO_STORE_ELIM is true
   when we want to do conditional store replacement, false otherwise.
   DO_HOIST_LOADS is true when we want to hoist adjacent loads out 
   of diamond control flow patterns, false otherwise.  */
static unsigned int
tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
{
  basic_block bb;
  basic_block *bb_order;
  unsigned n, i;
  bool cfgchanged = false;
  struct pointer_set_t *nontrap = 0;

  if (do_store_elim)
    /* Calculate the set of non-trapping memory accesses.  */
    nontrap = get_non_trapping ();

  /* Search every basic block for COND_EXPR we may be able to optimize.

     We walk the blocks in order that guarantees that a block with
     a single predecessor is processed before the predecessor.
     This ensures that we collapse inner ifs before visiting the
     outer ones, and also that we do not try to visit a removed
     block.  */
  bb_order = blocks_in_phiopt_order ();
  n = n_basic_blocks - NUM_FIXED_BLOCKS;

  for (i = 0; i < n; i++)
    {
      gimple cond_stmt, phi;
      basic_block bb1, bb2;
      edge e1, e2;
      tree arg0, arg1;

      bb = bb_order[i];

      cond_stmt = last_stmt (bb);
      /* Check to see if the last statement is a GIMPLE_COND.  */
      if (!cond_stmt
          || gimple_code (cond_stmt) != GIMPLE_COND)
        continue;

      e1 = EDGE_SUCC (bb, 0);
      bb1 = e1->dest;
      e2 = EDGE_SUCC (bb, 1);
      bb2 = e2->dest;

      /* We cannot do the optimization on abnormal edges.  */
      if ((e1->flags & EDGE_ABNORMAL) != 0
          || (e2->flags & EDGE_ABNORMAL) != 0)
       continue;

      /* If either bb1's succ or bb2 or bb2's succ is non NULL.  */
      if (EDGE_COUNT (bb1->succs) == 0
          || bb2 == NULL
          || EDGE_COUNT (bb2->succs) == 0)
        continue;

      /* Find the bb which is the fall through to the other.  */
      if (EDGE_SUCC (bb1, 0)->dest == bb2)
        ;
      else if (EDGE_SUCC (bb2, 0)->dest == bb1)
        {
          basic_block bb_tmp = bb1;
          edge e_tmp = e1;
          bb1 = bb2;
          bb2 = bb_tmp;
          e1 = e2;
          e2 = e_tmp;
        }
      else if (do_store_elim
               && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
        {
          basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;

          if (!single_succ_p (bb1)
              || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
              || !single_succ_p (bb2)
              || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
              || EDGE_COUNT (bb3->preds) != 2)
            continue;
          if (cond_if_else_store_replacement (bb1, bb2, bb3))
            cfgchanged = true;
          continue;
        }
      else if (do_hoist_loads
                 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
        {
          basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;

          if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
              && single_succ_p (bb1)
              && single_succ_p (bb2)
              && single_pred_p (bb1)
              && single_pred_p (bb2)
              && EDGE_COUNT (bb->succs) == 2
              && EDGE_COUNT (bb3->preds) == 2
              /* If one edge or the other is dominant, a conditional move
                 is likely to perform worse than the well-predicted branch.  */
              && !predictable_edge_p (EDGE_SUCC (bb, 0))
              && !predictable_edge_p (EDGE_SUCC (bb, 1)))
            hoist_adjacent_loads (bb, bb1, bb2, bb3);
          continue;
        }
      else
        continue;      

      e1 = EDGE_SUCC (bb1, 0);

      /* Make sure that bb1 is just a fall through.  */
      if (!single_succ_p (bb1)
          || (e1->flags & EDGE_FALLTHRU) == 0)
        continue;

      /* Also make sure that bb1 only have one predecessor and that it
         is bb.  */
      if (!single_pred_p (bb1)
          || single_pred (bb1) != bb)
        continue;

      if (do_store_elim)
        {
          /* bb1 is the middle block, bb2 the join block, bb the split block,
             e1 the fallthrough edge from bb1 to bb2.  We can't do the
             optimization if the join block has more than two predecessors.  */
          if (EDGE_COUNT (bb2->preds) > 2)
            continue;
          if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
            cfgchanged = true;
        }
      else
        {
          gimple_seq phis = phi_nodes (bb2);
          gimple_stmt_iterator gsi;
          bool candorest = true;

          /* Value replacement can work with more than one PHI
             so try that first. */
          for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
            {
              phi = gsi_stmt (gsi);
              arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
              arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
              if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
                {
                  candorest = false;
                  cfgchanged = true;
                  break;
                }
            }

          if (!candorest)
            continue;
          
          phi = single_non_singleton_phi_for_edges (phis, e1, e2);
          if (!phi)
            continue;

          arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
          arg1 = gimple_phi_arg_def (phi, e2->dest_idx);

          /* Something is wrong if we cannot find the arguments in the PHI
             node.  */
          gcc_assert (arg0 != NULL && arg1 != NULL);

          /* Do the replacement of conditional if it can be done.  */
          if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
            cfgchanged = true;
          else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
            cfgchanged = true;
          else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
            cfgchanged = true;
        }
    }

  free (bb_order);

  if (do_store_elim)
    pointer_set_destroy (nontrap);
  /* If the CFG has changed, we should cleanup the CFG.  */
  if (cfgchanged && do_store_elim)
    {
      /* In cond-store replacement we have added some loads on edges
         and new VOPS (as we moved the store, and created a load).  */
      gsi_commit_edge_inserts ();
      return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
    }
  else if (cfgchanged)
    return TODO_cleanup_cfg;
  return 0;
}

/* Returns the list of basic blocks in the function in an order that guarantees
   that if a block X has just a single predecessor Y, then Y is after X in the
   ordering.  */

basic_block *
blocks_in_phiopt_order (void)
{
  basic_block x, y;
  basic_block *order = XNEWVEC (basic_block, n_basic_blocks);
  unsigned n = n_basic_blocks - NUM_FIXED_BLOCKS;
  unsigned np, i;
  sbitmap visited = sbitmap_alloc (last_basic_block);

#define MARK_VISITED(BB) (bitmap_set_bit (visited, (BB)->index))
#define VISITED_P(BB) (bitmap_bit_p (visited, (BB)->index))

  bitmap_clear (visited);

  MARK_VISITED (ENTRY_BLOCK_PTR);
  FOR_EACH_BB (x)
    {
      if (VISITED_P (x))
        continue;

      /* Walk the predecessors of x as long as they have precisely one
         predecessor and add them to the list, so that they get stored
         after x.  */
      for (y = x, np = 1;
           single_pred_p (y) && !VISITED_P (single_pred (y));
           y = single_pred (y))
        np++;
      for (y = x, i = n - np;
           single_pred_p (y) && !VISITED_P (single_pred (y));
           y = single_pred (y), i++)
        {
          order[i] = y;
          MARK_VISITED (y);
        }
      order[i] = y;
      MARK_VISITED (y);

      gcc_assert (i == n - 1);
      n -= np;
    }

  sbitmap_free (visited);
  gcc_assert (n == 0);
  return order;

#undef MARK_VISITED
#undef VISITED_P
}

/* Replace PHI node element whose edge is E in block BB with variable NEW.
   Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
   is known to have two edges, one of which must reach BB).  */

static void
replace_phi_edge_with_variable (basic_block cond_block,
                                edge e, gimple phi, tree new_tree)
{
  basic_block bb = gimple_bb (phi);
  basic_block block_to_remove;
  gimple_stmt_iterator gsi;

  /* Change the PHI argument to new.  */
  SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);

  /* Remove the empty basic block.  */
  if (EDGE_SUCC (cond_block, 0)->dest == bb)
    {
      EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
      EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
      EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
      EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;

      block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
    }
  else
    {
      EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
      EDGE_SUCC (cond_block, 1)->flags
        &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
      EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
      EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;

      block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
    }
  delete_basic_block (block_to_remove);

  /* Eliminate the COND_EXPR at the end of COND_BLOCK.  */
  gsi = gsi_last_bb (cond_block);
  gsi_remove (&gsi, true);

  if (dump_file && (dump_flags & TDF_DETAILS))
    fprintf (dump_file,
              "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
              cond_block->index,
              bb->index);
}

/*  The function conditional_replacement does the main work of doing the
    conditional replacement.  Return true if the replacement is done.
    Otherwise return false.
    BB is the basic block where the replacement is going to be done on.  ARG0
    is argument 0 from PHI.  Likewise for ARG1.  */

static bool
conditional_replacement (basic_block cond_bb, basic_block middle_bb,
                         edge e0, edge e1, gimple phi,
                         tree arg0, tree arg1)
{
  tree result;
  gimple stmt, new_stmt;
  tree cond;
  gimple_stmt_iterator gsi;
  edge true_edge, false_edge;
  tree new_var, new_var2;
  bool neg;

  /* FIXME: Gimplification of complex type is too hard for now.  */
  /* We aren't prepared to handle vectors either (and it is a question
     if it would be worthwhile anyway).  */
  if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
        || POINTER_TYPE_P (TREE_TYPE (arg0)))
      || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
           || POINTER_TYPE_P (TREE_TYPE (arg1))))
    return false;

  /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
     convert it to the conditional.  */
  if ((integer_zerop (arg0) && integer_onep (arg1))
      || (integer_zerop (arg1) && integer_onep (arg0)))
    neg = false;
  else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
           || (integer_zerop (arg1) && integer_all_onesp (arg0)))
    neg = true;
  else
    return false;

  if (!empty_block_p (middle_bb))
    return false;

  /* At this point we know we have a GIMPLE_COND with two successors.
     One successor is BB, the other successor is an empty block which
     falls through into BB.

     There is a single PHI node at the join point (BB) and its arguments
     are constants (0, 1) or (0, -1).

     So, given the condition COND, and the two PHI arguments, we can
     rewrite this PHI into non-branching code:

       dest = (COND) or dest = COND'

     We use the condition as-is if the argument associated with the
     true edge has the value one or the argument associated with the
     false edge as the value zero.  Note that those conditions are not
     the same since only one of the outgoing edges from the GIMPLE_COND
     will directly reach BB and thus be associated with an argument.  */

  stmt = last_stmt (cond_bb);
  result = PHI_RESULT (phi);

  /* To handle special cases like floating point comparison, it is easier and
     less error-prone to build a tree and gimplify it on the fly though it is
     less efficient.  */
  cond = fold_build2_loc (gimple_location (stmt),
                          gimple_cond_code (stmt), boolean_type_node,
                          gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));

  /* We need to know which is the true edge and which is the false
     edge so that we know when to invert the condition below.  */
  extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
  if ((e0 == true_edge && integer_zerop (arg0))
      || (e0 == false_edge && !integer_zerop (arg0))
      || (e1 == true_edge && integer_zerop (arg1))
      || (e1 == false_edge && !integer_zerop (arg1)))
    cond = fold_build1_loc (gimple_location (stmt),
                            TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);

  if (neg)
    {
      cond = fold_convert_loc (gimple_location (stmt),
                               TREE_TYPE (result), cond);
      cond = fold_build1_loc (gimple_location (stmt),
                              NEGATE_EXPR, TREE_TYPE (cond), cond);
    }

  /* Insert our new statements at the end of conditional block before the
     COND_STMT.  */
  gsi = gsi_for_stmt (stmt);
  new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
                                      GSI_SAME_STMT);

  if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
    {
      source_location locus_0, locus_1;

      new_var2 = make_ssa_name (TREE_TYPE (result), NULL);
      new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2,
                                               new_var, NULL);
      gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
      new_var = new_var2;

      /* Set the locus to the first argument, unless is doesn't have one.  */
      locus_0 = gimple_phi_arg_location (phi, 0);
      locus_1 = gimple_phi_arg_location (phi, 1);
      if (locus_0 == UNKNOWN_LOCATION)
        locus_0 = locus_1;
      gimple_set_location (new_stmt, locus_0);
    }

  replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);

  /* Note that we optimized this PHI.  */
  return true;
}

/* Update *ARG which is defined in STMT so that it contains the
   computed value if that seems profitable.  Return true if the
   statement is made dead by that rewriting.  */

static bool
jump_function_from_stmt (tree *arg, gimple stmt)
{
  enum tree_code code = gimple_assign_rhs_code (stmt);
  if (code == ADDR_EXPR)
    {
      /* For arg = &p->i transform it to p, if possible.  */
      tree rhs1 = gimple_assign_rhs1 (stmt);
      HOST_WIDE_INT offset;
      tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
                                                &offset);
      if (tem
          && TREE_CODE (tem) == MEM_REF
          && (mem_ref_offset (tem) + double_int::from_shwi (offset)).is_zero ())
        {
          *arg = TREE_OPERAND (tem, 0);
          return true;
        }
    }
  /* TODO: Much like IPA-CP jump-functions we want to handle constant
     additions symbolically here, and we'd need to update the comparison
     code that compares the arg + cst tuples in our caller.  For now the
     code above exactly handles the VEC_BASE pattern from vec.h.  */
  return false;
}

/*  The function value_replacement does the main work of doing the value
    replacement.  Return non-zero if the replacement is done.  Otherwise return
    0.  If we remove the middle basic block, return 2.
    BB is the basic block where the replacement is going to be done on.  ARG0
    is argument 0 from the PHI.  Likewise for ARG1.  */

static int
value_replacement (basic_block cond_bb, basic_block middle_bb,
                   edge e0, edge e1, gimple phi,
                   tree arg0, tree arg1)
{
  gimple_stmt_iterator gsi;
  gimple cond;
  edge true_edge, false_edge;
  enum tree_code code;
  bool emtpy_or_with_defined_p = true;

  /* If the type says honor signed zeros we cannot do this
     optimization.  */
  if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
    return 0;

  /* If there is a statement in MIDDLE_BB that defines one of the PHI
     arguments, then adjust arg0 or arg1.  */
  gsi = gsi_after_labels (middle_bb);
  if (!gsi_end_p (gsi) && is_gimple_debug (gsi_stmt (gsi)))
    gsi_next_nondebug (&gsi);
  while (!gsi_end_p (gsi))
    {
      gimple stmt = gsi_stmt (gsi);
      tree lhs;
      gsi_next_nondebug (&gsi);
      if (!is_gimple_assign (stmt))
        {
          emtpy_or_with_defined_p = false;
          continue;
        }
      /* Now try to adjust arg0 or arg1 according to the computation
         in the statement.  */
      lhs = gimple_assign_lhs (stmt);
      if (!(lhs == arg0
             && jump_function_from_stmt (&arg0, stmt))
            || (lhs == arg1
                && jump_function_from_stmt (&arg1, stmt)))
        emtpy_or_with_defined_p = false;
    }

  cond = last_stmt (cond_bb);
  code = gimple_cond_code (cond);

  /* This transformation is only valid for equality comparisons.  */
  if (code != NE_EXPR && code != EQ_EXPR)
    return 0;

  /* We need to know which is the true edge and which is the false
      edge so that we know if have abs or negative abs.  */
  extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);

  /* At this point we know we have a COND_EXPR with two successors.
     One successor is BB, the other successor is an empty block which
     falls through into BB.

     The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.

     There is a single PHI node at the join point (BB) with two arguments.

     We now need to verify that the two arguments in the PHI node match
     the two arguments to the equality comparison.  */

  if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond))
       && operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond)))
      || (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond))
          && operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond))))
    {
      edge e;
      tree arg;

      /* For NE_EXPR, we want to build an assignment result = arg where
         arg is the PHI argument associated with the true edge.  For
         EQ_EXPR we want the PHI argument associated with the false edge.  */
      e = (code == NE_EXPR ? true_edge : false_edge);

      /* Unfortunately, E may not reach BB (it may instead have gone to
         OTHER_BLOCK).  If that is the case, then we want the single outgoing
         edge from OTHER_BLOCK which reaches BB and represents the desired
         path from COND_BLOCK.  */
      if (e->dest == middle_bb)
        e = single_succ_edge (e->dest);

      /* Now we know the incoming edge to BB that has the argument for the
         RHS of our new assignment statement.  */
      if (e0 == e)
        arg = arg0;
      else
        arg = arg1;

      /* If the middle basic block was empty or is defining the
         PHI arguments and this is a single phi where the args are different
         for the edges e0 and e1 then we can remove the middle basic block. */
      if (emtpy_or_with_defined_p
          && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
                                                            e0, e1))
        {
          replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
          /* Note that we optimized this PHI.  */
          return 2;
        }
      else
        {
          /* Replace the PHI arguments with arg. */
          SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
          SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
          if (dump_file && (dump_flags & TDF_DETAILS))
            {
              fprintf (dump_file, "PHI ");
              print_generic_expr (dump_file, gimple_phi_result (phi), 0);
              fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
                       cond_bb->index);
              print_generic_expr (dump_file, arg, 0);
              fprintf (dump_file, ".\n");
            }
          return 1;
        }

    }
  return 0;
}

/*  The function minmax_replacement does the main work of doing the minmax
    replacement.  Return true if the replacement is done.  Otherwise return
    false.
    BB is the basic block where the replacement is going to be done on.  ARG0
    is argument 0 from the PHI.  Likewise for ARG1.  */

static bool
minmax_replacement (basic_block cond_bb, basic_block middle_bb,
                    edge e0, edge e1, gimple phi,
                    tree arg0, tree arg1)
{
  tree result, type;
  gimple cond, new_stmt;
  edge true_edge, false_edge;
  enum tree_code cmp, minmax, ass_code;
  tree smaller, larger, arg_true, arg_false;
  gimple_stmt_iterator gsi, gsi_from;

  type = TREE_TYPE (PHI_RESULT (phi));

  /* The optimization may be unsafe due to NaNs.  */
  if (HONOR_NANS (TYPE_MODE (type)))
    return false;

  cond = last_stmt (cond_bb);
  cmp = gimple_cond_code (cond);

  /* This transformation is only valid for order comparisons.  Record which
     operand is smaller/larger if the result of the comparison is true.  */
  if (cmp == LT_EXPR || cmp == LE_EXPR)
    {
      smaller = gimple_cond_lhs (cond);
      larger = gimple_cond_rhs (cond);
    }
  else if (cmp == GT_EXPR || cmp == GE_EXPR)
    {
      smaller = gimple_cond_rhs (cond);
      larger = gimple_cond_lhs (cond);
    }
  else
    return false;

  /* We need to know which is the true edge and which is the false
      edge so that we know if have abs or negative abs.  */
  extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);

  /* Forward the edges over the middle basic block.  */
  if (true_edge->dest == middle_bb)
    true_edge = EDGE_SUCC (true_edge->dest, 0);
  if (false_edge->dest == middle_bb)
    false_edge = EDGE_SUCC (false_edge->dest, 0);

  if (true_edge == e0)
    {
      gcc_assert (false_edge == e1);
      arg_true = arg0;
      arg_false = arg1;
    }
  else
    {
      gcc_assert (false_edge == e0);
      gcc_assert (true_edge == e1);
      arg_true = arg1;
      arg_false = arg0;
    }

  if (empty_block_p (middle_bb))
    {
      if (operand_equal_for_phi_arg_p (arg_true, smaller)
          && operand_equal_for_phi_arg_p (arg_false, larger))
        {
          /* Case

             if (smaller < larger)
             rslt = smaller;
             else
             rslt = larger;  */
          minmax = MIN_EXPR;
        }
      else if (operand_equal_for_phi_arg_p (arg_false, smaller)
               && operand_equal_for_phi_arg_p (arg_true, larger))
        minmax = MAX_EXPR;
      else
        return false;
    }
  else
    {
      /* Recognize the following case, assuming d <= u:

         if (a <= u)
           b = MAX (a, d);
         x = PHI <b, u>

         This is equivalent to

         b = MAX (a, d);
         x = MIN (b, u);  */

      gimple assign = last_and_only_stmt (middle_bb);
      tree lhs, op0, op1, bound;

      if (!assign
          || gimple_code (assign) != GIMPLE_ASSIGN)
        return false;

      lhs = gimple_assign_lhs (assign);
      ass_code = gimple_assign_rhs_code (assign);
      if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
        return false;
      op0 = gimple_assign_rhs1 (assign);
      op1 = gimple_assign_rhs2 (assign);

      if (true_edge->src == middle_bb)
        {
          /* We got here if the condition is true, i.e., SMALLER < LARGER.  */
          if (!operand_equal_for_phi_arg_p (lhs, arg_true))
            return false;

          if (operand_equal_for_phi_arg_p (arg_false, larger))
            {
              /* Case

                 if (smaller < larger)
                   {
                     r' = MAX_EXPR (smaller, bound)
                   }
                 r = PHI <r', larger>  --> to be turned to MIN_EXPR.  */
              if (ass_code != MAX_EXPR)
                return false;

              minmax = MIN_EXPR;
              if (operand_equal_for_phi_arg_p (op0, smaller))
                bound = op1;
              else if (operand_equal_for_phi_arg_p (op1, smaller))
                bound = op0;
              else
                return false;

              /* We need BOUND <= LARGER.  */
              if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
                                                  bound, larger)))
                return false;
            }
          else if (operand_equal_for_phi_arg_p (arg_false, smaller))
            {
              /* Case

                 if (smaller < larger)
                   {
                     r' = MIN_EXPR (larger, bound)
                   }
                 r = PHI <r', smaller>  --> to be turned to MAX_EXPR.  */
              if (ass_code != MIN_EXPR)
                return false;

              minmax = MAX_EXPR;
              if (operand_equal_for_phi_arg_p (op0, larger))
                bound = op1;
              else if (operand_equal_for_phi_arg_p (op1, larger))
                bound = op0;
              else
                return false;

              /* We need BOUND >= SMALLER.  */
              if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
                                                  bound, smaller)))
                return false;
            }
          else
            return false;
        }
      else
        {
          /* We got here if the condition is false, i.e., SMALLER > LARGER.  */
          if (!operand_equal_for_phi_arg_p (lhs, arg_false))
            return false;

          if (operand_equal_for_phi_arg_p (arg_true, larger))
            {
              /* Case

                 if (smaller > larger)
                   {
                     r' = MIN_EXPR (smaller, bound)
                   }
                 r = PHI <r', larger>  --> to be turned to MAX_EXPR.  */
              if (ass_code != MIN_EXPR)
                return false;

              minmax = MAX_EXPR;
              if (operand_equal_for_phi_arg_p (op0, smaller))
                bound = op1;
              else if (operand_equal_for_phi_arg_p (op1, smaller))
                bound = op0;
              else
                return false;

              /* We need BOUND >= LARGER.  */
              if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
                                                  bound, larger)))
                return false;
            }
          else if (operand_equal_for_phi_arg_p (arg_true, smaller))
            {
              /* Case

                 if (smaller > larger)
                   {
                     r' = MAX_EXPR (larger, bound)
                   }
                 r = PHI <r', smaller>  --> to be turned to MIN_EXPR.  */
              if (ass_code != MAX_EXPR)
                return false;

              minmax = MIN_EXPR;
              if (operand_equal_for_phi_arg_p (op0, larger))
                bound = op1;
              else if (operand_equal_for_phi_arg_p (op1, larger))
                bound = op0;
              else
                return false;

              /* We need BOUND <= SMALLER.  */
              if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
                                                  bound, smaller)))
                return false;
            }
          else
            return false;
        }

      /* Move the statement from the middle block.  */
      gsi = gsi_last_bb (cond_bb);
      gsi_from = gsi_last_nondebug_bb (middle_bb);
      gsi_move_before (&gsi_from, &gsi);
    }

  /* Emit the statement to compute min/max.  */
  result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
  new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1);
  gsi = gsi_last_bb (cond_bb);
  gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);

  replace_phi_edge_with_variable (cond_bb, e1, phi, result);
  return true;
}

/*  The function absolute_replacement does the main work of doing the absolute
    replacement.  Return true if the replacement is done.  Otherwise return
    false.
    bb is the basic block where the replacement is going to be done on.  arg0
    is argument 0 from the phi.  Likewise for arg1.  */

static bool
abs_replacement (basic_block cond_bb, basic_block middle_bb,
                 edge e0 ATTRIBUTE_UNUSED, edge e1,
                 gimple phi, tree arg0, tree arg1)
{
  tree result;
  gimple new_stmt, cond;
  gimple_stmt_iterator gsi;
  edge true_edge, false_edge;
  gimple assign;
  edge e;
  tree rhs, lhs;
  bool negate;
  enum tree_code cond_code;

  /* If the type says honor signed zeros we cannot do this
     optimization.  */
  if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
    return false;

  /* OTHER_BLOCK must have only one executable statement which must have the
     form arg0 = -arg1 or arg1 = -arg0.  */

  assign = last_and_only_stmt (middle_bb);
  /* If we did not find the proper negation assignment, then we can not
     optimize.  */
  if (assign == NULL)
    return false;

  /* If we got here, then we have found the only executable statement
     in OTHER_BLOCK.  If it is anything other than arg = -arg1 or
     arg1 = -arg0, then we can not optimize.  */
  if (gimple_code (assign) != GIMPLE_ASSIGN)
    return false;

  lhs = gimple_assign_lhs (assign);

  if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
    return false;

  rhs = gimple_assign_rhs1 (assign);

  /* The assignment has to be arg0 = -arg1 or arg1 = -arg0.  */
  if (!(lhs == arg0 && rhs == arg1)
      && !(lhs == arg1 && rhs == arg0))
    return false;

  cond = last_stmt (cond_bb);
  result = PHI_RESULT (phi);

  /* Only relationals comparing arg[01] against zero are interesting.  */
  cond_code = gimple_cond_code (cond);
  if (cond_code != GT_EXPR && cond_code != GE_EXPR
      && cond_code != LT_EXPR && cond_code != LE_EXPR)
    return false;

  /* Make sure the conditional is arg[01] OP y.  */
  if (gimple_cond_lhs (cond) != rhs)
    return false;

  if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
               ? real_zerop (gimple_cond_rhs (cond))
               : integer_zerop (gimple_cond_rhs (cond)))
    ;
  else
    return false;

  /* We need to know which is the true edge and which is the false
     edge so that we know if have abs or negative abs.  */
  extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);

  /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
     will need to negate the result.  Similarly for LT_EXPR/LE_EXPR if
     the false edge goes to OTHER_BLOCK.  */
  if (cond_code == GT_EXPR || cond_code == GE_EXPR)
    e = true_edge;
  else
    e = false_edge;

  if (e->dest == middle_bb)
    negate = true;
  else
    negate = false;

  result = duplicate_ssa_name (result, NULL);

  if (negate)
    lhs = make_ssa_name (TREE_TYPE (result), NULL);
  else
    lhs = result;

  /* Build the modify expression with abs expression.  */
  new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL);

  gsi = gsi_last_bb (cond_bb);
  gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);

  if (negate)
    {
      /* Get the right GSI.  We want to insert after the recently
         added ABS_EXPR statement (which we know is the first statement
         in the block.  */
      new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL);

      gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
    }

  replace_phi_edge_with_variable (cond_bb, e1, phi, result);

  /* Note that we optimized this PHI.  */
  return true;
}

/* Auxiliary functions to determine the set of memory accesses which
   can't trap because they are preceded by accesses to the same memory
   portion.  We do that for MEM_REFs, so we only need to track
   the SSA_NAME of the pointer indirectly referenced.  The algorithm
   simply is a walk over all instructions in dominator order.  When
   we see an MEM_REF we determine if we've already seen a same
   ref anywhere up to the root of the dominator tree.  If we do the
   current access can't trap.  If we don't see any dominating access
   the current access might trap, but might also make later accesses
   non-trapping, so we remember it.  We need to be careful with loads
   or stores, for instance a load might not trap, while a store would,
   so if we see a dominating read access this doesn't mean that a later
   write access would not trap.  Hence we also need to differentiate the
   type of access(es) seen.

   ??? We currently are very conservative and assume that a load might
   trap even if a store doesn't (write-only memory).  This probably is
   overly conservative.  */

/* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
   through it was seen, which would constitute a no-trap region for
   same accesses.  */
struct name_to_bb
{
  unsigned int ssa_name_ver;
  unsigned int phase;
  bool store;
  HOST_WIDE_INT offset, size;
  basic_block bb;
};

/* The hash table for remembering what we've seen.  */
static htab_t seen_ssa_names;

/* Used for quick clearing of the hash-table when we see calls.
   Hash entries with phase < nt_call_phase are invalid.  */
static unsigned int nt_call_phase;

/* The set of MEM_REFs which can't trap.  */
static struct pointer_set_t *nontrap_set;

/* The hash function.  */
static hashval_t
name_to_bb_hash (const void *p)
{
  const struct name_to_bb *n = (const struct name_to_bb *) p;
  return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
         ^ (n->offset << 6) ^ (n->size << 3);
}

/* The equality function of *P1 and *P2.  */
static int
name_to_bb_eq (const void *p1, const void *p2)
{
  const struct name_to_bb *n1 = (const struct name_to_bb *)p1;
  const struct name_to_bb *n2 = (const struct name_to_bb *)p2;

  return n1->ssa_name_ver == n2->ssa_name_ver
         && n1->store == n2->store
         && n1->offset == n2->offset
         && n1->size == n2->size;
}

/* We see the expression EXP in basic block BB.  If it's an interesting
   expression (an MEM_REF through an SSA_NAME) possibly insert the
   expression into the set NONTRAP or the hash table of seen expressions.
   STORE is true if this expression is on the LHS, otherwise it's on
   the RHS.  */
static void
add_or_mark_expr (basic_block bb, tree exp,
                  struct pointer_set_t *nontrap, bool store)
{
  HOST_WIDE_INT size;

  if (TREE_CODE (exp) == MEM_REF
      && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
      && host_integerp (TREE_OPERAND (exp, 1), 0)
      && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
    {
      tree name = TREE_OPERAND (exp, 0);
      struct name_to_bb map;
      void **slot;
      struct name_to_bb *n2bb;
      basic_block found_bb = 0;

      /* Try to find the last seen MEM_REF through the same
         SSA_NAME, which can trap.  */
      map.ssa_name_ver = SSA_NAME_VERSION (name);
      map.phase = 0;
      map.bb = 0;
      map.store = store;
      map.offset = tree_low_cst (TREE_OPERAND (exp, 1), 0);
      map.size = size;

      slot = htab_find_slot (seen_ssa_names, &map, INSERT);
      n2bb = (struct name_to_bb *) *slot;
      if (n2bb && n2bb->phase >= nt_call_phase)
        found_bb = n2bb->bb;

      /* If we've found a trapping MEM_REF, _and_ it dominates EXP
         (it's in a basic block on the path from us to the dominator root)
         then we can't trap.  */
      if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
        {
          pointer_set_insert (nontrap, exp);
        }
      else
        {
          /* EXP might trap, so insert it into the hash table.  */
          if (n2bb)
            {
              n2bb->phase = nt_call_phase;
              n2bb->bb = bb;
            }
          else
            {
              n2bb = XNEW (struct name_to_bb);
              n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
              n2bb->phase = nt_call_phase;
              n2bb->bb = bb;
              n2bb->store = store;
              n2bb->offset = map.offset;
              n2bb->size = size;
              *slot = n2bb;
            }
        }
    }
}

/* Return true when CALL is a call stmt that definitely doesn't
   free any memory or makes it unavailable otherwise.  */
static bool
nonfreeing_call_p (gimple call)
{
  if (gimple_call_builtin_p (call, BUILT_IN_NORMAL)
      && gimple_call_flags (call) & ECF_LEAF)
    switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call)))
      {
        /* Just in case these become ECF_LEAF in the future.  */
        case BUILT_IN_FREE:
        case BUILT_IN_TM_FREE:
        case BUILT_IN_REALLOC:
        case BUILT_IN_STACK_RESTORE:
          return false;
        default:
          return true;
      }

  return false;
}

/* Called by walk_dominator_tree, when entering the block BB.  */
static void
nt_init_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
{
  edge e;
  edge_iterator ei;
  gimple_stmt_iterator gsi;

  /* If we haven't seen all our predecessors, clear the hash-table.  */
  FOR_EACH_EDGE (e, ei, bb->preds)
    if ((((size_t)e->src->aux) & 2) == 0)
      {
        nt_call_phase++;
        break;
      }

  /* Mark this BB as being on the path to dominator root and as visited.  */
  bb->aux = (void*)(1 | 2);

  /* And walk the statements in order.  */
  for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
    {
      gimple stmt = gsi_stmt (gsi);

      if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
        nt_call_phase++;
      else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
        {
          add_or_mark_expr (bb, gimple_assign_lhs (stmt), nontrap_set, true);
          add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), nontrap_set, false);
        }
    }
}

/* Called by walk_dominator_tree, when basic block BB is exited.  */
static void
nt_fini_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
{
  /* This BB isn't on the path to dominator root anymore.  */
  bb->aux = (void*)2;
}

/* This is the entry point of gathering non trapping memory accesses.
   It will do a dominator walk over the whole function, and it will
   make use of the bb->aux pointers.  It returns a set of trees
   (the MEM_REFs itself) which can't trap.  */
static struct pointer_set_t *
get_non_trapping (void)
{
  struct pointer_set_t *nontrap;
  struct dom_walk_data walk_data;

  nt_call_phase = 0;
  nontrap = pointer_set_create ();
  seen_ssa_names = htab_create (128, name_to_bb_hash, name_to_bb_eq,
                                free);
  /* We're going to do a dominator walk, so ensure that we have
     dominance information.  */
  calculate_dominance_info (CDI_DOMINATORS);

  /* Setup callbacks for the generic dominator tree walker.  */
  nontrap_set = nontrap;
  walk_data.dom_direction = CDI_DOMINATORS;
  walk_data.initialize_block_local_data = NULL;
  walk_data.before_dom_children = nt_init_block;
  walk_data.after_dom_children = nt_fini_block;
  walk_data.global_data = NULL;
  walk_data.block_local_data_size = 0;

  init_walk_dominator_tree (&walk_data);
  walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
  fini_walk_dominator_tree (&walk_data);
  htab_delete (seen_ssa_names);

  clear_aux_for_blocks ();
  return nontrap;
}

/* Do the main work of conditional store replacement.  We already know
   that the recognized pattern looks like so:

   split:
     if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
   MIDDLE_BB:
     something
     fallthrough (edge E0)
   JOIN_BB:
     some more

   We check that MIDDLE_BB contains only one store, that that store
   doesn't trap (not via NOTRAP, but via checking if an access to the same
   memory location dominates us) and that the store has a "simple" RHS.  */

static bool
cond_store_replacement (basic_block middle_bb, basic_block join_bb,
                        edge e0, edge e1, struct pointer_set_t *nontrap)
{
  gimple assign = last_and_only_stmt (middle_bb);
  tree lhs, rhs, name, name2;
  gimple newphi, new_stmt;
  gimple_stmt_iterator gsi;
  source_location locus;

  /* Check if middle_bb contains of only one store.  */
  if (!assign
      || !gimple_assign_single_p (assign)
      || gimple_has_volatile_ops (assign))
    return false;

  locus = gimple_location (assign);
  lhs = gimple_assign_lhs (assign);
  rhs = gimple_assign_rhs1 (assign);
  if (TREE_CODE (lhs) != MEM_REF
      || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
      || !is_gimple_reg_type (TREE_TYPE (lhs)))
    return false;

  /* Prove that we can move the store down.  We could also check
     TREE_THIS_NOTRAP here, but in that case we also could move stores,
     whose value is not available readily, which we want to avoid.  */
  if (!pointer_set_contains (nontrap, lhs))
    return false;

  /* Now we've checked the constraints, so do the transformation:
     1) Remove the single store.  */
  gsi = gsi_for_stmt (assign);
  unlink_stmt_vdef (assign);
  gsi_remove (&gsi, true);
  release_defs (assign);

  /* 2) Insert a load from the memory of the store to the temporary
        on the edge which did not contain the store.  */
  lhs = unshare_expr (lhs);
  name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
  new_stmt = gimple_build_assign (name, lhs);
  gimple_set_location (new_stmt, locus);
  gsi_insert_on_edge (e1, new_stmt);

  /* 3) Create a PHI node at the join block, with one argument
        holding the old RHS, and the other holding the temporary
        where we stored the old memory contents.  */
  name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
  newphi = create_phi_node (name2, join_bb);
  add_phi_arg (newphi, rhs, e0, locus);
  add_phi_arg (newphi, name, e1, locus);

  lhs = unshare_expr (lhs);
  new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));

  /* 4) Insert that PHI node.  */
  gsi = gsi_after_labels (join_bb);
  if (gsi_end_p (gsi))
    {
      gsi = gsi_last_bb (join_bb);
      gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
    }
  else
    gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);

  return true;
}

/* Do the main work of conditional store replacement.  */

static bool
cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
                                  basic_block join_bb, gimple then_assign,
                                  gimple else_assign)
{
  tree lhs_base, lhs, then_rhs, else_rhs, name;
  source_location then_locus, else_locus;
  gimple_stmt_iterator gsi;
  gimple newphi, new_stmt;

  if (then_assign == NULL
      || !gimple_assign_single_p (then_assign)
      || gimple_clobber_p (then_assign)
      || gimple_has_volatile_ops (then_assign)
      || else_assign == NULL
      || !gimple_assign_single_p (else_assign)
      || gimple_clobber_p (else_assign)
      || gimple_has_volatile_ops (else_assign))
    return false;

  lhs = gimple_assign_lhs (then_assign);
  if (!is_gimple_reg_type (TREE_TYPE (lhs))
      || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
    return false;

  lhs_base = get_base_address (lhs);
  if (lhs_base == NULL_TREE
      || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
    return false;

  then_rhs = gimple_assign_rhs1 (then_assign);
  else_rhs = gimple_assign_rhs1 (else_assign);
  then_locus = gimple_location (then_assign);
  else_locus = gimple_location (else_assign);

  /* Now we've checked the constraints, so do the transformation:
     1) Remove the stores.  */
  gsi = gsi_for_stmt (then_assign);
  unlink_stmt_vdef (then_assign);
  gsi_remove (&gsi, true);
  release_defs (then_assign);

  gsi = gsi_for_stmt (else_assign);
  unlink_stmt_vdef (else_assign);
  gsi_remove (&gsi, true);
  release_defs (else_assign);

  /* 2) Create a PHI node at the join block, with one argument
        holding the old RHS, and the other holding the temporary
        where we stored the old memory contents.  */
  name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
  newphi = create_phi_node (name, join_bb);
  add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
  add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);

  new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));

  /* 3) Insert that PHI node.  */
  gsi = gsi_after_labels (join_bb);
  if (gsi_end_p (gsi))
    {
      gsi = gsi_last_bb (join_bb);
      gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
    }
  else
    gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);

  return true;
}

/* Conditional store replacement.  We already know
   that the recognized pattern looks like so:

   split:
     if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
   THEN_BB:
     ...
     X = Y;
     ...
     goto JOIN_BB;
   ELSE_BB:
     ...
     X = Z;
     ...
     fallthrough (edge E0)
   JOIN_BB:
     some more

   We check that it is safe to sink the store to JOIN_BB by verifying that
   there are no read-after-write or write-after-write dependencies in
   THEN_BB and ELSE_BB.  */

static bool
cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
                                basic_block join_bb)
{
  gimple then_assign = last_and_only_stmt (then_bb);
  gimple else_assign = last_and_only_stmt (else_bb);
  vec<data_reference_p> then_datarefs, else_datarefs;
  vec<ddr_p> then_ddrs, else_ddrs;
  gimple then_store, else_store;
  bool found, ok = false, res;
  struct data_dependence_relation *ddr;
  data_reference_p then_dr, else_dr;
  int i, j;
  tree then_lhs, else_lhs;
  vec<gimple> then_stores, else_stores;
  basic_block blocks[3];

  if (MAX_STORES_TO_SINK == 0)
    return false;

  /* Handle the case with single statement in THEN_BB and ELSE_BB.  */
  if (then_assign && else_assign)
    return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
                                             then_assign, else_assign);

  /* Find data references.  */
  then_datarefs.create (1);
  else_datarefs.create (1);
  if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
        == chrec_dont_know)
      || !then_datarefs.length ()
      || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
        == chrec_dont_know)
      || !else_datarefs.length ())
    {
      free_data_refs (then_datarefs);
      free_data_refs (else_datarefs);
      return false;
    }

  /* Find pairs of stores with equal LHS.  */
  then_stores.create (1);
  else_stores.create (1);
  FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
    {
      if (DR_IS_READ (then_dr))
        continue;

      then_store = DR_STMT (then_dr);
      then_lhs = gimple_get_lhs (then_store);
      found = false;

      FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
        {
          if (DR_IS_READ (else_dr))
            continue;

          else_store = DR_STMT (else_dr);
          else_lhs = gimple_get_lhs (else_store);

          if (operand_equal_p (then_lhs, else_lhs, 0))
            {
              found = true;
              break;
            }
        }

      if (!found)
        continue;

      then_stores.safe_push (then_store);
      else_stores.safe_push (else_store);
    }

  /* No pairs of stores found.  */
  if (!then_stores.length ()
      || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
    {
      free_data_refs (then_datarefs);
      free_data_refs (else_datarefs);
      then_stores.release ();
      else_stores.release ();
      return false;
    }

  /* Compute and check data dependencies in both basic blocks.  */
  then_ddrs.create (1);
  else_ddrs.create (1);
  if (!compute_all_dependences (then_datarefs, &then_ddrs,
                                vNULL, false)
      || !compute_all_dependences (else_datarefs, &else_ddrs,
                                   vNULL, false))
    {
      free_dependence_relations (then_ddrs);
      free_dependence_relations (else_ddrs);
      free_data_refs (then_datarefs);
      free_data_refs (else_datarefs);
      then_stores.release ();
      else_stores.release ();
      return false;
    }
  blocks[0] = then_bb;
  blocks[1] = else_bb;
  blocks[2] = join_bb;
  renumber_gimple_stmt_uids_in_blocks (blocks, 3);

  /* Check that there are no read-after-write or write-after-write dependencies
     in THEN_BB.  */
  FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
    {
      struct data_reference *dra = DDR_A (ddr);
      struct data_reference *drb = DDR_B (ddr);

      if (DDR_ARE_DEPENDENT (ddr) != chrec_known
          && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
               && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
              || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
                  && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
              || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
        {
          free_dependence_relations (then_ddrs);
          free_dependence_relations (else_ddrs);
          free_data_refs (then_datarefs);
          free_data_refs (else_datarefs);
          then_stores.release ();
          else_stores.release ();
          return false;
        }
    }

  /* Check that there are no read-after-write or write-after-write dependencies
     in ELSE_BB.  */
  FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
    {
      struct data_reference *dra = DDR_A (ddr);
      struct data_reference *drb = DDR_B (ddr);

      if (DDR_ARE_DEPENDENT (ddr) != chrec_known
          && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
               && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
              || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
                  && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
              || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
        {
          free_dependence_relations (then_ddrs);
          free_dependence_relations (else_ddrs);
          free_data_refs (then_datarefs);
          free_data_refs (else_datarefs);
          then_stores.release ();
          else_stores.release ();
          return false;
        }
    }

  /* Sink stores with same LHS.  */
  FOR_EACH_VEC_ELT (then_stores, i, then_store)
    {
      else_store = else_stores[i];
      res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
                                              then_store, else_store);
      ok = ok || res;
    }

  free_dependence_relations (then_ddrs);
  free_dependence_relations (else_ddrs);
  free_data_refs (then_datarefs);
  free_data_refs (else_datarefs);
  then_stores.release ();
  else_stores.release ();

  return ok;
}

/* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB.  */

static bool
local_mem_dependence (gimple stmt, basic_block bb)
{
  tree vuse = gimple_vuse (stmt);
  gimple def;

  if (!vuse)
    return false;

  def = SSA_NAME_DEF_STMT (vuse);
  return (def && gimple_bb (def) == bb);
}

/* Given a "diamond" control-flow pattern where BB0 tests a condition,
   BB1 and BB2 are "then" and "else" blocks dependent on this test,
   and BB3 rejoins control flow following BB1 and BB2, look for 
   opportunities to hoist loads as follows.  If BB3 contains a PHI of
   two loads, one each occurring in BB1 and BB2, and the loads are
   provably of adjacent fields in the same structure, then move both
   loads into BB0.  Of course this can only be done if there are no
   dependencies preventing such motion.

   One of the hoisted loads will always be speculative, so the
   transformation is currently conservative:

    - The fields must be strictly adjacent.
    - The two fields must occupy a single memory block that is
      guaranteed to not cross a page boundary.

    The last is difficult to prove, as such memory blocks should be
    aligned on the minimum of the stack alignment boundary and the
    alignment guaranteed by heap allocation interfaces.  Thus we rely
    on a parameter for the alignment value.

    Provided a good value is used for the last case, the first
    restriction could possibly be relaxed.  */

static void
hoist_adjacent_loads (basic_block bb0, basic_block bb1,
                      basic_block bb2, basic_block bb3)
{
  int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
  unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
  gimple_stmt_iterator gsi;

  /* Walk the phis in bb3 looking for an opportunity.  We are looking
     for phis of two SSA names, one each of which is defined in bb1 and
     bb2.  */
  for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
    {
      gimple phi_stmt = gsi_stmt (gsi);
      gimple def1, def2, defswap;
      tree arg1, arg2, ref1, ref2, field1, field2, fieldswap;
      tree tree_offset1, tree_offset2, tree_size2, next;
      int offset1, offset2, size2;
      unsigned align1;
      gimple_stmt_iterator gsi2;
      basic_block bb_for_def1, bb_for_def2;

      if (gimple_phi_num_args (phi_stmt) != 2
          || virtual_operand_p (gimple_phi_result (phi_stmt)))
        continue;

      arg1 = gimple_phi_arg_def (phi_stmt, 0);
      arg2 = gimple_phi_arg_def (phi_stmt, 1);
      
      if (TREE_CODE (arg1) != SSA_NAME
          || TREE_CODE (arg2) != SSA_NAME
          || SSA_NAME_IS_DEFAULT_DEF (arg1)
          || SSA_NAME_IS_DEFAULT_DEF (arg2))
        continue;

      def1 = SSA_NAME_DEF_STMT (arg1);
      def2 = SSA_NAME_DEF_STMT (arg2);

      if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
          && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
        continue;

      /* Check the mode of the arguments to be sure a conditional move
         can be generated for it.  */
      if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
          == CODE_FOR_nothing)
        continue;

      /* Both statements must be assignments whose RHS is a COMPONENT_REF.  */
      if (!gimple_assign_single_p (def1)
          || !gimple_assign_single_p (def2)
          || gimple_has_volatile_ops (def1)
          || gimple_has_volatile_ops (def2))
        continue;

      ref1 = gimple_assign_rhs1 (def1);
      ref2 = gimple_assign_rhs1 (def2);

      if (TREE_CODE (ref1) != COMPONENT_REF
          || TREE_CODE (ref2) != COMPONENT_REF)
        continue;

      /* The zeroth operand of the two component references must be
         identical.  It is not sufficient to compare get_base_address of
         the two references, because this could allow for different
         elements of the same array in the two trees.  It is not safe to
         assume that the existence of one array element implies the
         existence of a different one.  */
      if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
        continue;

      field1 = TREE_OPERAND (ref1, 1);
      field2 = TREE_OPERAND (ref2, 1);

      /* Check for field adjacency, and ensure field1 comes first.  */
      for (next = DECL_CHAIN (field1);
           next && TREE_CODE (next) != FIELD_DECL;
           next = DECL_CHAIN (next))
        ;

      if (next != field2)
        {
          for (next = DECL_CHAIN (field2);
               next && TREE_CODE (next) != FIELD_DECL;
               next = DECL_CHAIN (next))
            ;

          if (next != field1)
            continue;

          fieldswap = field1;
          field1 = field2;
          field2 = fieldswap;
          defswap = def1;
          def1 = def2;
          def2 = defswap;
        }

      bb_for_def1 = gimple_bb (def1);
      bb_for_def2 = gimple_bb (def2);

      /* Check for proper alignment of the first field.  */
      tree_offset1 = bit_position (field1);
      tree_offset2 = bit_position (field2);
      tree_size2 = DECL_SIZE (field2);

      if (!host_integerp (tree_offset1, 1)
          || !host_integerp (tree_offset2, 1)
          || !host_integerp (tree_size2, 1))
        continue;

      offset1 = TREE_INT_CST_LOW (tree_offset1);
      offset2 = TREE_INT_CST_LOW (tree_offset2);
      size2 = TREE_INT_CST_LOW (tree_size2);
      align1 = DECL_ALIGN (field1) % param_align_bits;

      if (offset1 % BITS_PER_UNIT != 0)
        continue;

      /* For profitability, the two field references should fit within
         a single cache line.  */
      if (align1 + offset2 - offset1 + size2 > param_align_bits)
        continue;

      /* The two expressions cannot be dependent upon vdefs defined
         in bb1/bb2.  */
      if (local_mem_dependence (def1, bb_for_def1)
          || local_mem_dependence (def2, bb_for_def2))
        continue;

      /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
         bb0.  We hoist the first one first so that a cache miss is handled
         efficiently regardless of hardware cache-fill policy.  */
      gsi2 = gsi_for_stmt (def1);
      gsi_move_to_bb_end (&gsi2, bb0);
      gsi2 = gsi_for_stmt (def2);
      gsi_move_to_bb_end (&gsi2, bb0);

      if (dump_file && (dump_flags & TDF_DETAILS))
        {
          fprintf (dump_file,
                   "\nHoisting adjacent loads from %d and %d into %d: \n",
                   bb_for_def1->index, bb_for_def2->index, bb0->index);
          print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
          print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
        }
    }
}

/* Determine whether we should attempt to hoist adjacent loads out of
   diamond patterns in pass_phiopt.  Always hoist loads if
   -fhoist-adjacent-loads is specified and the target machine has
   both a conditional move instruction and a defined cache line size.  */

static bool
gate_hoist_loads (void)
{
  return (flag_hoist_adjacent_loads == 1
          && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
          && HAVE_conditional_move);
}

/* Always do these optimizations if we have SSA
   trees to work on.  */
static bool
gate_phiopt (void)
{
  return 1;
}

struct gimple_opt_pass pass_phiopt =
{
 {
  GIMPLE_PASS,
  "phiopt",                             /* name */
  OPTGROUP_NONE,                        /* optinfo_flags */
  gate_phiopt,                          /* gate */
  tree_ssa_phiopt,                      /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_TREE_PHIOPT,                       /* tv_id */
  PROP_cfg | PROP_ssa,                  /* properties_required */
  0,                                    /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  TODO_ggc_collect
    | TODO_verify_ssa
    | TODO_verify_flow
    | TODO_verify_stmts                 /* todo_flags_finish */
 }
};

static bool
gate_cselim (void)
{
  return flag_tree_cselim;
}

struct gimple_opt_pass pass_cselim =
{
 {
  GIMPLE_PASS,
  "cselim",                             /* name */
  OPTGROUP_NONE,                        /* optinfo_flags */
  gate_cselim,                          /* gate */
  tree_ssa_cs_elim,                     /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_TREE_PHIOPT,                       /* tv_id */
  PROP_cfg | PROP_ssa,                  /* properties_required */
  0,                                    /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  TODO_ggc_collect
    | TODO_verify_ssa
    | TODO_verify_flow
    | TODO_verify_stmts                 /* todo_flags_finish */
 }
};