[PR67828] don't unswitch on default defs of non-parms

[official-gcc.git] / gcc / tree-ssa-loop-im.c

/* Loop invariant motion.
   Copyright (C) 2003-2015 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 "backend.h"
#include "cfghooks.h"
#include "tree.h"
#include "gimple.h"
#include "hard-reg-set.h"
#include "ssa.h"
#include "alias.h"
#include "fold-const.h"
#include "tm_p.h"
#include "cfganal.h"
#include "gimple-pretty-print.h"
#include "internal-fn.h"
#include "tree-eh.h"
#include "gimplify.h"
#include "gimple-iterator.h"
#include "tree-cfg.h"
#include "tree-ssa-loop-manip.h"
#include "tree-ssa-loop.h"
#include "tree-into-ssa.h"
#include "cfgloop.h"
#include "domwalk.h"
#include "params.h"
#include "tree-pass.h"
#include "flags.h"
#include "tree-affine.h"
#include "tree-ssa-propagate.h"
#include "trans-mem.h"
#include "gimple-fold.h"

/* TODO:  Support for predicated code motion.  I.e.

   while (1)
     {
       if (cond)
         {
           a = inv;
           something;
         }
     }

   Where COND and INV are invariants, but evaluating INV may trap or be
   invalid from some other reason if !COND.  This may be transformed to

   if (cond)
     a = inv;
   while (1)
     {
       if (cond)
         something;
     }  */

/* The auxiliary data kept for each statement.  */

struct lim_aux_data
{
  struct loop *max_loop;        /* The outermost loop in that the statement
                                   is invariant.  */

  struct loop *tgt_loop;        /* The loop out of that we want to move the
                                   invariant.  */

  struct loop *always_executed_in;
                                /* The outermost loop for that we are sure
                                   the statement is executed if the loop
                                   is entered.  */

  unsigned cost;                /* Cost of the computation performed by the
                                   statement.  */

  vec<gimple *> depends;                /* Vector of statements that must be also
                                   hoisted out of the loop when this statement
                                   is hoisted; i.e. those that define the
                                   operands of the statement and are inside of
                                   the MAX_LOOP loop.  */
};

/* Maps statements to their lim_aux_data.  */

static hash_map<gimple *, lim_aux_data *> *lim_aux_data_map;

/* Description of a memory reference location.  */

struct mem_ref_loc
{
  tree *ref;                    /* The reference itself.  */
  gimple *stmt;                 /* The statement in that it occurs.  */
};


/* Description of a memory reference.  */

struct im_mem_ref
{
  unsigned id;                  /* ID assigned to the memory reference
                                   (its index in memory_accesses.refs_list)  */
  hashval_t hash;               /* Its hash value.  */

  /* The memory access itself and associated caching of alias-oracle
     query meta-data.  */
  ao_ref mem;

  bitmap stored;                /* The set of loops in that this memory location
                                   is stored to.  */
  vec<mem_ref_loc>              accesses_in_loop;
                                /* The locations of the accesses.  Vector
                                   indexed by the loop number.  */

  /* The following sets are computed on demand.  We keep both set and
     its complement, so that we know whether the information was
     already computed or not.  */
  bitmap_head indep_loop;       /* The set of loops in that the memory
                                   reference is independent, meaning:
                                   If it is stored in the loop, this store
                                     is independent on all other loads and
                                     stores.
                                   If it is only loaded, then it is independent
                                     on all stores in the loop.  */
  bitmap_head dep_loop;         /* The complement of INDEP_LOOP.  */
};

/* We use two bits per loop in the ref->{in,}dep_loop bitmaps, the first
   to record (in)dependence against stores in the loop and its subloops, the
   second to record (in)dependence against all references in the loop
   and its subloops.  */
#define LOOP_DEP_BIT(loopnum, storedp) (2 * (loopnum) + (storedp ? 1 : 0))

/* Mem_ref hashtable helpers.  */

struct mem_ref_hasher : nofree_ptr_hash <im_mem_ref>
{
  typedef tree_node *compare_type;
  static inline hashval_t hash (const im_mem_ref *);
  static inline bool equal (const im_mem_ref *, const tree_node *);
};

/* A hash function for struct im_mem_ref object OBJ.  */

inline hashval_t
mem_ref_hasher::hash (const im_mem_ref *mem)
{
  return mem->hash;
}

/* An equality function for struct im_mem_ref object MEM1 with
   memory reference OBJ2.  */

inline bool
mem_ref_hasher::equal (const im_mem_ref *mem1, const tree_node *obj2)
{
  return operand_equal_p (mem1->mem.ref, (const_tree) obj2, 0);
}


/* Description of memory accesses in loops.  */

static struct
{
  /* The hash table of memory references accessed in loops.  */
  hash_table<mem_ref_hasher> *refs;

  /* The list of memory references.  */
  vec<im_mem_ref *> refs_list;

  /* The set of memory references accessed in each loop.  */
  vec<bitmap_head> refs_in_loop;

  /* The set of memory references stored in each loop.  */
  vec<bitmap_head> refs_stored_in_loop;

  /* The set of memory references stored in each loop, including subloops .  */
  vec<bitmap_head> all_refs_stored_in_loop;

  /* Cache for expanding memory addresses.  */
  hash_map<tree, name_expansion *> *ttae_cache;
} memory_accesses;

/* Obstack for the bitmaps in the above data structures.  */
static bitmap_obstack lim_bitmap_obstack;
static obstack mem_ref_obstack;

static bool ref_indep_loop_p (struct loop *, im_mem_ref *);

/* Minimum cost of an expensive expression.  */
#define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE))

/* The outermost loop for which execution of the header guarantees that the
   block will be executed.  */
#define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
#define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL))

/* ID of the shared unanalyzable mem.  */
#define UNANALYZABLE_MEM_ID 0

/* Whether the reference was analyzable.  */
#define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID)

static struct lim_aux_data *
init_lim_data (gimple *stmt)
{
  lim_aux_data *p = XCNEW (struct lim_aux_data);
  lim_aux_data_map->put (stmt, p);

  return p;
}

static struct lim_aux_data *
get_lim_data (gimple *stmt)
{
  lim_aux_data **p = lim_aux_data_map->get (stmt);
  if (!p)
    return NULL;

  return *p;
}

/* Releases the memory occupied by DATA.  */

static void
free_lim_aux_data (struct lim_aux_data *data)
{
  data->depends.release ();
  free (data);
}

static void
clear_lim_data (gimple *stmt)
{
  lim_aux_data **p = lim_aux_data_map->get (stmt);
  if (!p)
    return;

  free_lim_aux_data (*p);
  *p = NULL;
}


/* The possibilities of statement movement.  */
enum move_pos
  {
    MOVE_IMPOSSIBLE,            /* No movement -- side effect expression.  */
    MOVE_PRESERVE_EXECUTION,    /* Must not cause the non-executed statement
                                   become executed -- memory accesses, ... */
    MOVE_POSSIBLE               /* Unlimited movement.  */
  };


/* If it is possible to hoist the statement STMT unconditionally,
   returns MOVE_POSSIBLE.
   If it is possible to hoist the statement STMT, but we must avoid making
   it executed if it would not be executed in the original program (e.g.
   because it may trap), return MOVE_PRESERVE_EXECUTION.
   Otherwise return MOVE_IMPOSSIBLE.  */

enum move_pos
movement_possibility (gimple *stmt)
{
  tree lhs;
  enum move_pos ret = MOVE_POSSIBLE;

  if (flag_unswitch_loops
      && gimple_code (stmt) == GIMPLE_COND)
    {
      /* If we perform unswitching, force the operands of the invariant
         condition to be moved out of the loop.  */
      return MOVE_POSSIBLE;
    }

  if (gimple_code (stmt) == GIMPLE_PHI
      && gimple_phi_num_args (stmt) <= 2
      && !virtual_operand_p (gimple_phi_result (stmt))
      && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt)))
    return MOVE_POSSIBLE;

  if (gimple_get_lhs (stmt) == NULL_TREE)
    return MOVE_IMPOSSIBLE;

  if (gimple_vdef (stmt))
    return MOVE_IMPOSSIBLE;

  if (stmt_ends_bb_p (stmt)
      || gimple_has_volatile_ops (stmt)
      || gimple_has_side_effects (stmt)
      || stmt_could_throw_p (stmt))
    return MOVE_IMPOSSIBLE;

  if (is_gimple_call (stmt))
    {
      /* While pure or const call is guaranteed to have no side effects, we
         cannot move it arbitrarily.  Consider code like

         char *s = something ();

         while (1)
           {
             if (s)
               t = strlen (s);
             else
               t = 0;
           }

         Here the strlen call cannot be moved out of the loop, even though
         s is invariant.  In addition to possibly creating a call with
         invalid arguments, moving out a function call that is not executed
         may cause performance regressions in case the call is costly and
         not executed at all.  */
      ret = MOVE_PRESERVE_EXECUTION;
      lhs = gimple_call_lhs (stmt);
    }
  else if (is_gimple_assign (stmt))
    lhs = gimple_assign_lhs (stmt);
  else
    return MOVE_IMPOSSIBLE;

  if (TREE_CODE (lhs) == SSA_NAME
      && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
    return MOVE_IMPOSSIBLE;

  if (TREE_CODE (lhs) != SSA_NAME
      || gimple_could_trap_p (stmt))
    return MOVE_PRESERVE_EXECUTION;

  /* Non local loads in a transaction cannot be hoisted out.  Well,
     unless the load happens on every path out of the loop, but we
     don't take this into account yet.  */
  if (flag_tm
      && gimple_in_transaction (stmt)
      && gimple_assign_single_p (stmt))
    {
      tree rhs = gimple_assign_rhs1 (stmt);
      if (DECL_P (rhs) && is_global_var (rhs))
        {
          if (dump_file)
            {
              fprintf (dump_file, "Cannot hoist conditional load of ");
              print_generic_expr (dump_file, rhs, TDF_SLIM);
              fprintf (dump_file, " because it is in a transaction.\n");
            }
          return MOVE_IMPOSSIBLE;
        }
    }

  return ret;
}

/* Suppose that operand DEF is used inside the LOOP.  Returns the outermost
   loop to that we could move the expression using DEF if it did not have
   other operands, i.e. the outermost loop enclosing LOOP in that the value
   of DEF is invariant.  */

static struct loop *
outermost_invariant_loop (tree def, struct loop *loop)
{
  gimple *def_stmt;
  basic_block def_bb;
  struct loop *max_loop;
  struct lim_aux_data *lim_data;

  if (!def)
    return superloop_at_depth (loop, 1);

  if (TREE_CODE (def) != SSA_NAME)
    {
      gcc_assert (is_gimple_min_invariant (def));
      return superloop_at_depth (loop, 1);
    }

  def_stmt = SSA_NAME_DEF_STMT (def);
  def_bb = gimple_bb (def_stmt);
  if (!def_bb)
    return superloop_at_depth (loop, 1);

  max_loop = find_common_loop (loop, def_bb->loop_father);

  lim_data = get_lim_data (def_stmt);
  if (lim_data != NULL && lim_data->max_loop != NULL)
    max_loop = find_common_loop (max_loop,
                                 loop_outer (lim_data->max_loop));
  if (max_loop == loop)
    return NULL;
  max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1);

  return max_loop;
}

/* DATA is a structure containing information associated with a statement
   inside LOOP.  DEF is one of the operands of this statement.

   Find the outermost loop enclosing LOOP in that value of DEF is invariant
   and record this in DATA->max_loop field.  If DEF itself is defined inside
   this loop as well (i.e. we need to hoist it out of the loop if we want
   to hoist the statement represented by DATA), record the statement in that
   DEF is defined to the DATA->depends list.  Additionally if ADD_COST is true,
   add the cost of the computation of DEF to the DATA->cost.

   If DEF is not invariant in LOOP, return false.  Otherwise return TRUE.  */

static bool
add_dependency (tree def, struct lim_aux_data *data, struct loop *loop,
                bool add_cost)
{
  gimple *def_stmt = SSA_NAME_DEF_STMT (def);
  basic_block def_bb = gimple_bb (def_stmt);
  struct loop *max_loop;
  struct lim_aux_data *def_data;

  if (!def_bb)
    return true;

  max_loop = outermost_invariant_loop (def, loop);
  if (!max_loop)
    return false;

  if (flow_loop_nested_p (data->max_loop, max_loop))
    data->max_loop = max_loop;

  def_data = get_lim_data (def_stmt);
  if (!def_data)
    return true;

  if (add_cost
      /* Only add the cost if the statement defining DEF is inside LOOP,
         i.e. if it is likely that by moving the invariants dependent
         on it, we will be able to avoid creating a new register for
         it (since it will be only used in these dependent invariants).  */
      && def_bb->loop_father == loop)
    data->cost += def_data->cost;

  data->depends.safe_push (def_stmt);

  return true;
}

/* Returns an estimate for a cost of statement STMT.  The values here
   are just ad-hoc constants, similar to costs for inlining.  */

static unsigned
stmt_cost (gimple *stmt)
{
  /* Always try to create possibilities for unswitching.  */
  if (gimple_code (stmt) == GIMPLE_COND
      || gimple_code (stmt) == GIMPLE_PHI)
    return LIM_EXPENSIVE;

  /* We should be hoisting calls if possible.  */
  if (is_gimple_call (stmt))
    {
      tree fndecl;

      /* Unless the call is a builtin_constant_p; this always folds to a
         constant, so moving it is useless.  */
      fndecl = gimple_call_fndecl (stmt);
      if (fndecl
          && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
          && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P)
        return 0;

      return LIM_EXPENSIVE;
    }

  /* Hoisting memory references out should almost surely be a win.  */
  if (gimple_references_memory_p (stmt))
    return LIM_EXPENSIVE;

  if (gimple_code (stmt) != GIMPLE_ASSIGN)
    return 1;

  switch (gimple_assign_rhs_code (stmt))
    {
    case MULT_EXPR:
    case WIDEN_MULT_EXPR:
    case WIDEN_MULT_PLUS_EXPR:
    case WIDEN_MULT_MINUS_EXPR:
    case DOT_PROD_EXPR:
    case FMA_EXPR:
    case TRUNC_DIV_EXPR:
    case CEIL_DIV_EXPR:
    case FLOOR_DIV_EXPR:
    case ROUND_DIV_EXPR:
    case EXACT_DIV_EXPR:
    case CEIL_MOD_EXPR:
    case FLOOR_MOD_EXPR:
    case ROUND_MOD_EXPR:
    case TRUNC_MOD_EXPR:
    case RDIV_EXPR:
      /* Division and multiplication are usually expensive.  */
      return LIM_EXPENSIVE;

    case LSHIFT_EXPR:
    case RSHIFT_EXPR:
    case WIDEN_LSHIFT_EXPR:
    case LROTATE_EXPR:
    case RROTATE_EXPR:
      /* Shifts and rotates are usually expensive.  */
      return LIM_EXPENSIVE;

    case CONSTRUCTOR:
      /* Make vector construction cost proportional to the number
         of elements.  */
      return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));

    case SSA_NAME:
    case PAREN_EXPR:
      /* Whether or not something is wrapped inside a PAREN_EXPR
         should not change move cost.  Nor should an intermediate
         unpropagated SSA name copy.  */
      return 0;

    default:
      return 1;
    }
}

/* Finds the outermost loop between OUTER and LOOP in that the memory reference
   REF is independent.  If REF is not independent in LOOP, NULL is returned
   instead.  */

static struct loop *
outermost_indep_loop (struct loop *outer, struct loop *loop, im_mem_ref *ref)
{
  struct loop *aloop;

  if (ref->stored && bitmap_bit_p (ref->stored, loop->num))
    return NULL;

  for (aloop = outer;
       aloop != loop;
       aloop = superloop_at_depth (loop, loop_depth (aloop) + 1))
    if ((!ref->stored || !bitmap_bit_p (ref->stored, aloop->num))
        && ref_indep_loop_p (aloop, ref))
      return aloop;

  if (ref_indep_loop_p (loop, ref))
    return loop;
  else
    return NULL;
}

/* If there is a simple load or store to a memory reference in STMT, returns
   the location of the memory reference, and sets IS_STORE according to whether
   it is a store or load.  Otherwise, returns NULL.  */

static tree *
simple_mem_ref_in_stmt (gimple *stmt, bool *is_store)
{
  tree *lhs, *rhs;

  /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns.  */
  if (!gimple_assign_single_p (stmt))
    return NULL;

  lhs = gimple_assign_lhs_ptr (stmt);
  rhs = gimple_assign_rhs1_ptr (stmt);

  if (TREE_CODE (*lhs) == SSA_NAME && gimple_vuse (stmt))
    {
      *is_store = false;
      return rhs;
    }
  else if (gimple_vdef (stmt)
           && (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs)))
    {
      *is_store = true;
      return lhs;
    }
  else
    return NULL;
}

/* Returns the memory reference contained in STMT.  */

static im_mem_ref *
mem_ref_in_stmt (gimple *stmt)
{
  bool store;
  tree *mem = simple_mem_ref_in_stmt (stmt, &store);
  hashval_t hash;
  im_mem_ref *ref;

  if (!mem)
    return NULL;
  gcc_assert (!store);

  hash = iterative_hash_expr (*mem, 0);
  ref = memory_accesses.refs->find_with_hash (*mem, hash);

  gcc_assert (ref != NULL);
  return ref;
}

/* From a controlling predicate in DOM determine the arguments from
   the PHI node PHI that are chosen if the predicate evaluates to
   true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if
   they are non-NULL.  Returns true if the arguments can be determined,
   else return false.  */

static bool
extract_true_false_args_from_phi (basic_block dom, gphi *phi,
                                  tree *true_arg_p, tree *false_arg_p)
{
  basic_block bb = gimple_bb (phi);
  edge true_edge, false_edge, tem;
  tree arg0 = NULL_TREE, arg1 = NULL_TREE;

  /* We have to verify that one edge into the PHI node is dominated
     by the true edge of the predicate block and the other edge
     dominated by the false edge.  This ensures that the PHI argument
     we are going to take is completely determined by the path we
     take from the predicate block.
     We can only use BB dominance checks below if the destination of
     the true/false edges are dominated by their edge, thus only
     have a single predecessor.  */
  extract_true_false_edges_from_block (dom, &true_edge, &false_edge);
  tem = EDGE_PRED (bb, 0);
  if (tem == true_edge
      || (single_pred_p (true_edge->dest)
          && (tem->src == true_edge->dest
              || dominated_by_p (CDI_DOMINATORS,
                                 tem->src, true_edge->dest))))
    arg0 = PHI_ARG_DEF (phi, tem->dest_idx);
  else if (tem == false_edge
           || (single_pred_p (false_edge->dest)
               && (tem->src == false_edge->dest
                   || dominated_by_p (CDI_DOMINATORS,
                                      tem->src, false_edge->dest))))
    arg1 = PHI_ARG_DEF (phi, tem->dest_idx);
  else
    return false;
  tem = EDGE_PRED (bb, 1);
  if (tem == true_edge
      || (single_pred_p (true_edge->dest)
          && (tem->src == true_edge->dest
              || dominated_by_p (CDI_DOMINATORS,
                                 tem->src, true_edge->dest))))
    arg0 = PHI_ARG_DEF (phi, tem->dest_idx);
  else if (tem == false_edge
           || (single_pred_p (false_edge->dest)
               && (tem->src == false_edge->dest
                   || dominated_by_p (CDI_DOMINATORS,
                                      tem->src, false_edge->dest))))
    arg1 = PHI_ARG_DEF (phi, tem->dest_idx);
  else
    return false;
  if (!arg0 || !arg1)
    return false;

  if (true_arg_p)
    *true_arg_p = arg0;
  if (false_arg_p)
    *false_arg_p = arg1;

  return true;
}

/* Determine the outermost loop to that it is possible to hoist a statement
   STMT and store it to LIM_DATA (STMT)->max_loop.  To do this we determine
   the outermost loop in that the value computed by STMT is invariant.
   If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
   we preserve the fact whether STMT is executed.  It also fills other related
   information to LIM_DATA (STMT).

   The function returns false if STMT cannot be hoisted outside of the loop it
   is defined in, and true otherwise.  */

static bool
determine_max_movement (gimple *stmt, bool must_preserve_exec)
{
  basic_block bb = gimple_bb (stmt);
  struct loop *loop = bb->loop_father;
  struct loop *level;
  struct lim_aux_data *lim_data = get_lim_data (stmt);
  tree val;
  ssa_op_iter iter;

  if (must_preserve_exec)
    level = ALWAYS_EXECUTED_IN (bb);
  else
    level = superloop_at_depth (loop, 1);
  lim_data->max_loop = level;

  if (gphi *phi = dyn_cast <gphi *> (stmt))
    {
      use_operand_p use_p;
      unsigned min_cost = UINT_MAX;
      unsigned total_cost = 0;
      struct lim_aux_data *def_data;

      /* We will end up promoting dependencies to be unconditionally
         evaluated.  For this reason the PHI cost (and thus the
         cost we remove from the loop by doing the invariant motion)
         is that of the cheapest PHI argument dependency chain.  */
      FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE)
        {
          val = USE_FROM_PTR (use_p);

          if (TREE_CODE (val) != SSA_NAME)
            {
              /* Assign const 1 to constants.  */
              min_cost = MIN (min_cost, 1);
              total_cost += 1;
              continue;
            }
          if (!add_dependency (val, lim_data, loop, false))
            return false;

          gimple *def_stmt = SSA_NAME_DEF_STMT (val);
          if (gimple_bb (def_stmt)
              && gimple_bb (def_stmt)->loop_father == loop)
            {
              def_data = get_lim_data (def_stmt);
              if (def_data)
                {
                  min_cost = MIN (min_cost, def_data->cost);
                  total_cost += def_data->cost;
                }
            }
        }

      min_cost = MIN (min_cost, total_cost);
      lim_data->cost += min_cost;

      if (gimple_phi_num_args (phi) > 1)
        {
          basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
          gimple *cond;
          if (gsi_end_p (gsi_last_bb (dom)))
            return false;
          cond = gsi_stmt (gsi_last_bb (dom));
          if (gimple_code (cond) != GIMPLE_COND)
            return false;
          /* Verify that this is an extended form of a diamond and
             the PHI arguments are completely controlled by the
             predicate in DOM.  */
          if (!extract_true_false_args_from_phi (dom, phi, NULL, NULL))
            return false;

          /* Fold in dependencies and cost of the condition.  */
          FOR_EACH_SSA_TREE_OPERAND (val, cond, iter, SSA_OP_USE)
            {
              if (!add_dependency (val, lim_data, loop, false))
                return false;
              def_data = get_lim_data (SSA_NAME_DEF_STMT (val));
              if (def_data)
                total_cost += def_data->cost;
            }

          /* We want to avoid unconditionally executing very expensive
             operations.  As costs for our dependencies cannot be
             negative just claim we are not invariand for this case.
             We also are not sure whether the control-flow inside the
             loop will vanish.  */
          if (total_cost - min_cost >= 2 * LIM_EXPENSIVE
              && !(min_cost != 0
                   && total_cost / min_cost <= 2))
            return false;

          /* Assume that the control-flow in the loop will vanish.
             ???  We should verify this and not artificially increase
             the cost if that is not the case.  */
          lim_data->cost += stmt_cost (stmt);
        }

      return true;
    }
  else
    FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE)
      if (!add_dependency (val, lim_data, loop, true))
        return false;

  if (gimple_vuse (stmt))
    {
      im_mem_ref *ref = mem_ref_in_stmt (stmt);

      if (ref)
        {
          lim_data->max_loop
                  = outermost_indep_loop (lim_data->max_loop, loop, ref);
          if (!lim_data->max_loop)
            return false;
        }
      else
        {
          if ((val = gimple_vuse (stmt)) != NULL_TREE)
            {
              if (!add_dependency (val, lim_data, loop, false))
                return false;
            }
        }
    }

  lim_data->cost += stmt_cost (stmt);

  return true;
}

/* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
   and that one of the operands of this statement is computed by STMT.
   Ensure that STMT (together with all the statements that define its
   operands) is hoisted at least out of the loop LEVEL.  */

static void
set_level (gimple *stmt, struct loop *orig_loop, struct loop *level)
{
  struct loop *stmt_loop = gimple_bb (stmt)->loop_father;
  struct lim_aux_data *lim_data;
  gimple *dep_stmt;
  unsigned i;

  stmt_loop = find_common_loop (orig_loop, stmt_loop);
  lim_data = get_lim_data (stmt);
  if (lim_data != NULL && lim_data->tgt_loop != NULL)
    stmt_loop = find_common_loop (stmt_loop,
                                  loop_outer (lim_data->tgt_loop));
  if (flow_loop_nested_p (stmt_loop, level))
    return;

  gcc_assert (level == lim_data->max_loop
              || flow_loop_nested_p (lim_data->max_loop, level));

  lim_data->tgt_loop = level;
  FOR_EACH_VEC_ELT (lim_data->depends, i, dep_stmt)
    set_level (dep_stmt, orig_loop, level);
}

/* Determines an outermost loop from that we want to hoist the statement STMT.
   For now we chose the outermost possible loop.  TODO -- use profiling
   information to set it more sanely.  */

static void
set_profitable_level (gimple *stmt)
{
  set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop);
}

/* Returns true if STMT is a call that has side effects.  */

static bool
nonpure_call_p (gimple *stmt)
{
  if (gimple_code (stmt) != GIMPLE_CALL)
    return false;

  return gimple_has_side_effects (stmt);
}

/* Rewrite a/b to a*(1/b).  Return the invariant stmt to process.  */

static gimple *
rewrite_reciprocal (gimple_stmt_iterator *bsi)
{
  gassign *stmt, *stmt1, *stmt2;
  tree name, lhs, type;
  tree real_one;
  gimple_stmt_iterator gsi;

  stmt = as_a <gassign *> (gsi_stmt (*bsi));
  lhs = gimple_assign_lhs (stmt);
  type = TREE_TYPE (lhs);

  real_one = build_one_cst (type);

  name = make_temp_ssa_name (type, NULL, "reciptmp");
  stmt1 = gimple_build_assign (name, RDIV_EXPR, real_one,
                               gimple_assign_rhs2 (stmt));
  stmt2 = gimple_build_assign (lhs, MULT_EXPR, name,
                               gimple_assign_rhs1 (stmt));

  /* Replace division stmt with reciprocal and multiply stmts.
     The multiply stmt is not invariant, so update iterator
     and avoid rescanning.  */
  gsi = *bsi;
  gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
  gsi_replace (&gsi, stmt2, true);

  /* Continue processing with invariant reciprocal statement.  */
  return stmt1;
}

/* Check if the pattern at *BSI is a bittest of the form
   (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0.  */

static gimple *
rewrite_bittest (gimple_stmt_iterator *bsi)
{
  gassign *stmt;
  gimple *stmt1;
  gassign *stmt2;
  gimple *use_stmt;
  gcond *cond_stmt;
  tree lhs, name, t, a, b;
  use_operand_p use;

  stmt = as_a <gassign *> (gsi_stmt (*bsi));
  lhs = gimple_assign_lhs (stmt);

  /* Verify that the single use of lhs is a comparison against zero.  */
  if (TREE_CODE (lhs) != SSA_NAME
      || !single_imm_use (lhs, &use, &use_stmt))
    return stmt;
  cond_stmt = dyn_cast <gcond *> (use_stmt);
  if (!cond_stmt)
    return stmt;
  if (gimple_cond_lhs (cond_stmt) != lhs
      || (gimple_cond_code (cond_stmt) != NE_EXPR
          && gimple_cond_code (cond_stmt) != EQ_EXPR)
      || !integer_zerop (gimple_cond_rhs (cond_stmt)))
    return stmt;

  /* Get at the operands of the shift.  The rhs is TMP1 & 1.  */
  stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
  if (gimple_code (stmt1) != GIMPLE_ASSIGN)
    return stmt;

  /* There is a conversion in between possibly inserted by fold.  */
  if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1)))
    {
      t = gimple_assign_rhs1 (stmt1);
      if (TREE_CODE (t) != SSA_NAME
          || !has_single_use (t))
        return stmt;
      stmt1 = SSA_NAME_DEF_STMT (t);
      if (gimple_code (stmt1) != GIMPLE_ASSIGN)
        return stmt;
    }

  /* Verify that B is loop invariant but A is not.  Verify that with
     all the stmt walking we are still in the same loop.  */
  if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR
      || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt))
    return stmt;

  a = gimple_assign_rhs1 (stmt1);
  b = gimple_assign_rhs2 (stmt1);

  if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL
      && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL)
    {
      gimple_stmt_iterator rsi;

      /* 1 << B */
      t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a),
                       build_int_cst (TREE_TYPE (a), 1), b);
      name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
      stmt1 = gimple_build_assign (name, t);

      /* A & (1 << B) */
      t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name);
      name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
      stmt2 = gimple_build_assign (name, t);

      /* Replace the SSA_NAME we compare against zero.  Adjust
         the type of zero accordingly.  */
      SET_USE (use, name);
      gimple_cond_set_rhs (cond_stmt,
                           build_int_cst_type (TREE_TYPE (name),
                                               0));

      /* Don't use gsi_replace here, none of the new assignments sets
         the variable originally set in stmt.  Move bsi to stmt1, and
         then remove the original stmt, so that we get a chance to
         retain debug info for it.  */
      rsi = *bsi;
      gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
      gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT);
      gimple *to_release = gsi_stmt (rsi);
      gsi_remove (&rsi, true);
      release_defs (to_release);

      return stmt1;
    }

  return stmt;
}

/* For each statement determines the outermost loop in that it is invariant,
   -   statements on whose motion it depends and the cost of the computation.
   -   This information is stored to the LIM_DATA structure associated with
   -   each statement.  */
class invariantness_dom_walker : public dom_walker
{
public:
  invariantness_dom_walker (cdi_direction direction)
    : dom_walker (direction) {}

  virtual void before_dom_children (basic_block);
};

/* Determine the outermost loops in that statements in basic block BB are
   invariant, and record them to the LIM_DATA associated with the statements.
   Callback for dom_walker.  */

void
invariantness_dom_walker::before_dom_children (basic_block bb)
{
  enum move_pos pos;
  gimple_stmt_iterator bsi;
  gimple *stmt;
  bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL;
  struct loop *outermost = ALWAYS_EXECUTED_IN (bb);
  struct lim_aux_data *lim_data;

  if (!loop_outer (bb->loop_father))
    return;

  if (dump_file && (dump_flags & TDF_DETAILS))
    fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n",
             bb->index, bb->loop_father->num, loop_depth (bb->loop_father));

  /* Look at PHI nodes, but only if there is at most two.
     ???  We could relax this further by post-processing the inserted
     code and transforming adjacent cond-exprs with the same predicate
     to control flow again.  */
  bsi = gsi_start_phis (bb);
  if (!gsi_end_p (bsi)
      && ((gsi_next (&bsi), gsi_end_p (bsi))
          || (gsi_next (&bsi), gsi_end_p (bsi))))
    for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
      {
        stmt = gsi_stmt (bsi);

        pos = movement_possibility (stmt);
        if (pos == MOVE_IMPOSSIBLE)
          continue;

        lim_data = init_lim_data (stmt);
        lim_data->always_executed_in = outermost;

        if (!determine_max_movement (stmt, false))
          {
            lim_data->max_loop = NULL;
            continue;
          }

        if (dump_file && (dump_flags & TDF_DETAILS))
          {
            print_gimple_stmt (dump_file, stmt, 2, 0);
            fprintf (dump_file, "  invariant up to level %d, cost %d.\n\n",
                     loop_depth (lim_data->max_loop),
                     lim_data->cost);
          }

        if (lim_data->cost >= LIM_EXPENSIVE)
          set_profitable_level (stmt);
      }

  for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
    {
      stmt = gsi_stmt (bsi);

      pos = movement_possibility (stmt);
      if (pos == MOVE_IMPOSSIBLE)
        {
          if (nonpure_call_p (stmt))
            {
              maybe_never = true;
              outermost = NULL;
            }
          /* Make sure to note always_executed_in for stores to make
             store-motion work.  */
          else if (stmt_makes_single_store (stmt))
            {
              struct lim_aux_data *lim_data = init_lim_data (stmt);
              lim_data->always_executed_in = outermost;
            }
          continue;
        }

      if (is_gimple_assign (stmt)
          && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
              == GIMPLE_BINARY_RHS))
        {
          tree op0 = gimple_assign_rhs1 (stmt);
          tree op1 = gimple_assign_rhs2 (stmt);
          struct loop *ol1 = outermost_invariant_loop (op1,
                                        loop_containing_stmt (stmt));

          /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal
             to be hoisted out of loop, saving expensive divide.  */
          if (pos == MOVE_POSSIBLE
              && gimple_assign_rhs_code (stmt) == RDIV_EXPR
              && flag_unsafe_math_optimizations
              && !flag_trapping_math
              && ol1 != NULL
              && outermost_invariant_loop (op0, ol1) == NULL)
            stmt = rewrite_reciprocal (&bsi);

          /* If the shift count is invariant, convert (A >> B) & 1 to
             A & (1 << B) allowing the bit mask to be hoisted out of the loop
             saving an expensive shift.  */
          if (pos == MOVE_POSSIBLE
              && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR
              && integer_onep (op1)
              && TREE_CODE (op0) == SSA_NAME
              && has_single_use (op0))
            stmt = rewrite_bittest (&bsi);
        }

      lim_data = init_lim_data (stmt);
      lim_data->always_executed_in = outermost;

      if (maybe_never && pos == MOVE_PRESERVE_EXECUTION)
        continue;

      if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION))
        {
          lim_data->max_loop = NULL;
          continue;
        }

      if (dump_file && (dump_flags & TDF_DETAILS))
        {
          print_gimple_stmt (dump_file, stmt, 2, 0);
          fprintf (dump_file, "  invariant up to level %d, cost %d.\n\n",
                   loop_depth (lim_data->max_loop),
                   lim_data->cost);
        }

      if (lim_data->cost >= LIM_EXPENSIVE)
        set_profitable_level (stmt);
    }
}

class move_computations_dom_walker : public dom_walker
{
public:
  move_computations_dom_walker (cdi_direction direction)
    : dom_walker (direction), todo_ (0) {}

  virtual void before_dom_children (basic_block);

  unsigned int todo_;
};

/* Hoist the statements in basic block BB out of the loops prescribed by
   data stored in LIM_DATA structures associated with each statement.  Callback
   for walk_dominator_tree.  */

void
move_computations_dom_walker::before_dom_children (basic_block bb)
{
  struct loop *level;
  unsigned cost = 0;
  struct lim_aux_data *lim_data;

  if (!loop_outer (bb->loop_father))
    return;

  for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); )
    {
      gassign *new_stmt;
      gphi *stmt = bsi.phi ();

      lim_data = get_lim_data (stmt);
      if (lim_data == NULL)
        {
          gsi_next (&bsi);
          continue;
        }

      cost = lim_data->cost;
      level = lim_data->tgt_loop;
      clear_lim_data (stmt);

      if (!level)
        {
          gsi_next (&bsi);
          continue;
        }

      if (dump_file && (dump_flags & TDF_DETAILS))
        {
          fprintf (dump_file, "Moving PHI node\n");
          print_gimple_stmt (dump_file, stmt, 0, 0);
          fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
                   cost, level->num);
        }

      if (gimple_phi_num_args (stmt) == 1)
        {
          tree arg = PHI_ARG_DEF (stmt, 0);
          new_stmt = gimple_build_assign (gimple_phi_result (stmt),
                                          TREE_CODE (arg), arg);
        }
      else
        {
          basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
          gimple *cond = gsi_stmt (gsi_last_bb (dom));
          tree arg0 = NULL_TREE, arg1 = NULL_TREE, t;
          /* Get the PHI arguments corresponding to the true and false
             edges of COND.  */
          extract_true_false_args_from_phi (dom, stmt, &arg0, &arg1);
          gcc_assert (arg0 && arg1);
          t = build2 (gimple_cond_code (cond), boolean_type_node,
                      gimple_cond_lhs (cond), gimple_cond_rhs (cond));
          new_stmt = gimple_build_assign (gimple_phi_result (stmt),
                                          COND_EXPR, t, arg0, arg1);
          todo_ |= TODO_cleanup_cfg;
        }
      if (INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (new_stmt)))
          && (!ALWAYS_EXECUTED_IN (bb)
              || (ALWAYS_EXECUTED_IN (bb) != level
                  && !flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
        {
          tree lhs = gimple_assign_lhs (new_stmt);
          SSA_NAME_RANGE_INFO (lhs) = NULL;
        }
      gsi_insert_on_edge (loop_preheader_edge (level), new_stmt);
      remove_phi_node (&bsi, false);
    }

  for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); )
    {
      edge e;

      gimple *stmt = gsi_stmt (bsi);

      lim_data = get_lim_data (stmt);
      if (lim_data == NULL)
        {
          gsi_next (&bsi);
          continue;
        }

      cost = lim_data->cost;
      level = lim_data->tgt_loop;
      clear_lim_data (stmt);

      if (!level)
        {
          gsi_next (&bsi);
          continue;
        }

      /* We do not really want to move conditionals out of the loop; we just
         placed it here to force its operands to be moved if necessary.  */
      if (gimple_code (stmt) == GIMPLE_COND)
        continue;

      if (dump_file && (dump_flags & TDF_DETAILS))
        {
          fprintf (dump_file, "Moving statement\n");
          print_gimple_stmt (dump_file, stmt, 0, 0);
          fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
                   cost, level->num);
        }

      e = loop_preheader_edge (level);
      gcc_assert (!gimple_vdef (stmt));
      if (gimple_vuse (stmt))
        {
          /* The new VUSE is the one from the virtual PHI in the loop
             header or the one already present.  */
          gphi_iterator gsi2;
          for (gsi2 = gsi_start_phis (e->dest);
               !gsi_end_p (gsi2); gsi_next (&gsi2))
            {
              gphi *phi = gsi2.phi ();
              if (virtual_operand_p (gimple_phi_result (phi)))
                {
                  gimple_set_vuse (stmt, PHI_ARG_DEF_FROM_EDGE (phi, e));
                  break;
                }
            }
        }
      gsi_remove (&bsi, false);
      if (gimple_has_lhs (stmt)
          && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
          && INTEGRAL_TYPE_P (TREE_TYPE (gimple_get_lhs (stmt)))
          && (!ALWAYS_EXECUTED_IN (bb)
              || !(ALWAYS_EXECUTED_IN (bb) == level
                   || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
        {
          tree lhs = gimple_get_lhs (stmt);
          SSA_NAME_RANGE_INFO (lhs) = NULL;
        }
      /* In case this is a stmt that is not unconditionally executed
         when the target loop header is executed and the stmt may
         invoke undefined integer or pointer overflow rewrite it to
         unsigned arithmetic.  */
      if (is_gimple_assign (stmt)
          && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt)))
          && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (gimple_assign_lhs (stmt)))
          && arith_code_with_undefined_signed_overflow
               (gimple_assign_rhs_code (stmt))
          && (!ALWAYS_EXECUTED_IN (bb)
              || !(ALWAYS_EXECUTED_IN (bb) == level
                   || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
        gsi_insert_seq_on_edge (e, rewrite_to_defined_overflow (stmt));
      else
        gsi_insert_on_edge (e, stmt);
    }
}

/* Hoist the statements out of the loops prescribed by data stored in
   LIM_DATA structures associated with each statement.*/

static unsigned int
move_computations (void)
{
  move_computations_dom_walker walker (CDI_DOMINATORS);
  walker.walk (cfun->cfg->x_entry_block_ptr);

  gsi_commit_edge_inserts ();
  if (need_ssa_update_p (cfun))
    rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);

  return walker.todo_;
}

/* Checks whether the statement defining variable *INDEX can be hoisted
   out of the loop passed in DATA.  Callback for for_each_index.  */

static bool
may_move_till (tree ref, tree *index, void *data)
{
  struct loop *loop = (struct loop *) data, *max_loop;

  /* If REF is an array reference, check also that the step and the lower
     bound is invariant in LOOP.  */
  if (TREE_CODE (ref) == ARRAY_REF)
    {
      tree step = TREE_OPERAND (ref, 3);
      tree lbound = TREE_OPERAND (ref, 2);

      max_loop = outermost_invariant_loop (step, loop);
      if (!max_loop)
        return false;

      max_loop = outermost_invariant_loop (lbound, loop);
      if (!max_loop)
        return false;
    }

  max_loop = outermost_invariant_loop (*index, loop);
  if (!max_loop)
    return false;

  return true;
}

/* If OP is SSA NAME, force the statement that defines it to be
   moved out of the LOOP.  ORIG_LOOP is the loop in that EXPR is used.  */

static void
force_move_till_op (tree op, struct loop *orig_loop, struct loop *loop)
{
  gimple *stmt;

  if (!op
      || is_gimple_min_invariant (op))
    return;

  gcc_assert (TREE_CODE (op) == SSA_NAME);

  stmt = SSA_NAME_DEF_STMT (op);
  if (gimple_nop_p (stmt))
    return;

  set_level (stmt, orig_loop, loop);
}

/* Forces statement defining invariants in REF (and *INDEX) to be moved out of
   the LOOP.  The reference REF is used in the loop ORIG_LOOP.  Callback for
   for_each_index.  */

struct fmt_data
{
  struct loop *loop;
  struct loop *orig_loop;
};

static bool
force_move_till (tree ref, tree *index, void *data)
{
  struct fmt_data *fmt_data = (struct fmt_data *) data;

  if (TREE_CODE (ref) == ARRAY_REF)
    {
      tree step = TREE_OPERAND (ref, 3);
      tree lbound = TREE_OPERAND (ref, 2);

      force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop);
      force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop);
    }

  force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop);

  return true;
}

/* A function to free the mem_ref object OBJ.  */

static void
memref_free (struct im_mem_ref *mem)
{
  mem->accesses_in_loop.release ();
}

/* Allocates and returns a memory reference description for MEM whose hash
   value is HASH and id is ID.  */

static im_mem_ref *
mem_ref_alloc (tree mem, unsigned hash, unsigned id)
{
  im_mem_ref *ref = XOBNEW (&mem_ref_obstack, struct im_mem_ref);
  ao_ref_init (&ref->mem, mem);
  ref->id = id;
  ref->hash = hash;
  ref->stored = NULL;
  bitmap_initialize (&ref->indep_loop, &lim_bitmap_obstack);
  bitmap_initialize (&ref->dep_loop, &lim_bitmap_obstack);
  ref->accesses_in_loop.create (1);

  return ref;
}

/* Records memory reference location *LOC in LOOP to the memory reference
   description REF.  The reference occurs in statement STMT.  */

static void
record_mem_ref_loc (im_mem_ref *ref, gimple *stmt, tree *loc)
{
  mem_ref_loc aref;
  aref.stmt = stmt;
  aref.ref = loc;
  ref->accesses_in_loop.safe_push (aref);
}

/* Set the LOOP bit in REF stored bitmap and allocate that if
   necessary.  Return whether a bit was changed.  */

static bool
set_ref_stored_in_loop (im_mem_ref *ref, struct loop *loop)
{
  if (!ref->stored)
    ref->stored = BITMAP_ALLOC (&lim_bitmap_obstack);
  return bitmap_set_bit (ref->stored, loop->num);
}

/* Marks reference REF as stored in LOOP.  */

static void
mark_ref_stored (im_mem_ref *ref, struct loop *loop)
{
  while (loop != current_loops->tree_root
         && set_ref_stored_in_loop (ref, loop))
    loop = loop_outer (loop);
}

/* Gathers memory references in statement STMT in LOOP, storing the
   information about them in the memory_accesses structure.  Marks
   the vops accessed through unrecognized statements there as
   well.  */

static void
gather_mem_refs_stmt (struct loop *loop, gimple *stmt)
{
  tree *mem = NULL;
  hashval_t hash;
  im_mem_ref **slot;
  im_mem_ref *ref;
  bool is_stored;
  unsigned id;

  if (!gimple_vuse (stmt))
    return;

  mem = simple_mem_ref_in_stmt (stmt, &is_stored);
  if (!mem)
    {
      /* We use the shared mem_ref for all unanalyzable refs.  */
      id = UNANALYZABLE_MEM_ID;
      ref = memory_accesses.refs_list[id];
      if (dump_file && (dump_flags & TDF_DETAILS))
        {
          fprintf (dump_file, "Unanalyzed memory reference %u: ", id);
          print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
        }
      is_stored = gimple_vdef (stmt);
    }
  else
    {
      hash = iterative_hash_expr (*mem, 0);
      slot = memory_accesses.refs->find_slot_with_hash (*mem, hash, INSERT);
      if (*slot)
        {
          ref = *slot;
          id = ref->id;
        }
      else
        {
          id = memory_accesses.refs_list.length ();
          ref = mem_ref_alloc (*mem, hash, id);
          memory_accesses.refs_list.safe_push (ref);
          *slot = ref;

          if (dump_file && (dump_flags & TDF_DETAILS))
            {
              fprintf (dump_file, "Memory reference %u: ", id);
              print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM);
              fprintf (dump_file, "\n");
            }
        }

      record_mem_ref_loc (ref, stmt, mem);
    }
  bitmap_set_bit (&memory_accesses.refs_in_loop[loop->num], ref->id);
  if (is_stored)
    {
      bitmap_set_bit (&memory_accesses.refs_stored_in_loop[loop->num], ref->id);
      mark_ref_stored (ref, loop);
    }
  return;
}

static unsigned *bb_loop_postorder;

/* qsort sort function to sort blocks after their loop fathers postorder.  */

static int
sort_bbs_in_loop_postorder_cmp (const void *bb1_, const void *bb2_)
{
  basic_block bb1 = *(basic_block *)const_cast<void *>(bb1_);
  basic_block bb2 = *(basic_block *)const_cast<void *>(bb2_);
  struct loop *loop1 = bb1->loop_father;
  struct loop *loop2 = bb2->loop_father;
  if (loop1->num == loop2->num)
    return 0;
  return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1;
}

/* qsort sort function to sort ref locs after their loop fathers postorder.  */

static int
sort_locs_in_loop_postorder_cmp (const void *loc1_, const void *loc2_)
{
  mem_ref_loc *loc1 = (mem_ref_loc *)const_cast<void *>(loc1_);
  mem_ref_loc *loc2 = (mem_ref_loc *)const_cast<void *>(loc2_);
  struct loop *loop1 = gimple_bb (loc1->stmt)->loop_father;
  struct loop *loop2 = gimple_bb (loc2->stmt)->loop_father;
  if (loop1->num == loop2->num)
    return 0;
  return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1;
}

/* Gathers memory references in loops.  */

static void
analyze_memory_references (void)
{
  gimple_stmt_iterator bsi;
  basic_block bb, *bbs;
  struct loop *loop, *outer;
  unsigned i, n;

  /* Collect all basic-blocks in loops and sort them after their
     loops postorder.  */
  i = 0;
  bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS);
  FOR_EACH_BB_FN (bb, cfun)
    if (bb->loop_father != current_loops->tree_root)
      bbs[i++] = bb;
  n = i;
  qsort (bbs, n, sizeof (basic_block), sort_bbs_in_loop_postorder_cmp);

  /* Visit blocks in loop postorder and assign mem-ref IDs in that order.
     That results in better locality for all the bitmaps.  */
  for (i = 0; i < n; ++i)
    {
      basic_block bb = bbs[i];
      for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
        gather_mem_refs_stmt (bb->loop_father, gsi_stmt (bsi));
    }

  /* Sort the location list of gathered memory references after their
     loop postorder number.  */
  im_mem_ref *ref;
  FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref)
    ref->accesses_in_loop.qsort (sort_locs_in_loop_postorder_cmp);

  free (bbs);
//  free (bb_loop_postorder);

  /* Propagate the information about accessed memory references up
     the loop hierarchy.  */
  FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
    {
      /* Finalize the overall touched references (including subloops).  */
      bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[loop->num],
                       &memory_accesses.refs_stored_in_loop[loop->num]);

      /* Propagate the information about accessed memory references up
         the loop hierarchy.  */
      outer = loop_outer (loop);
      if (outer == current_loops->tree_root)
        continue;

      bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[outer->num],
                       &memory_accesses.all_refs_stored_in_loop[loop->num]);
    }
}

/* Returns true if MEM1 and MEM2 may alias.  TTAE_CACHE is used as a cache in
   tree_to_aff_combination_expand.  */

static bool
mem_refs_may_alias_p (im_mem_ref *mem1, im_mem_ref *mem2,
                      hash_map<tree, name_expansion *> **ttae_cache)
{
  /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same
     object and their offset differ in such a way that the locations cannot
     overlap, then they cannot alias.  */
  widest_int size1, size2;
  aff_tree off1, off2;

  /* Perform basic offset and type-based disambiguation.  */
  if (!refs_may_alias_p_1 (&mem1->mem, &mem2->mem, true))
    return false;

  /* The expansion of addresses may be a bit expensive, thus we only do
     the check at -O2 and higher optimization levels.  */
  if (optimize < 2)
    return true;

  get_inner_reference_aff (mem1->mem.ref, &off1, &size1);
  get_inner_reference_aff (mem2->mem.ref, &off2, &size2);
  aff_combination_expand (&off1, ttae_cache);
  aff_combination_expand (&off2, ttae_cache);
  aff_combination_scale (&off1, -1);
  aff_combination_add (&off2, &off1);

  if (aff_comb_cannot_overlap_p (&off2, size1, size2))
    return false;

  return true;
}

/* Compare function for bsearch searching for reference locations
   in a loop.  */

static int
find_ref_loc_in_loop_cmp (const void *loop_, const void *loc_)
{
  struct loop *loop = (struct loop *)const_cast<void *>(loop_);
  mem_ref_loc *loc = (mem_ref_loc *)const_cast<void *>(loc_);
  struct loop *loc_loop = gimple_bb (loc->stmt)->loop_father;
  if (loop->num  == loc_loop->num
      || flow_loop_nested_p (loop, loc_loop))
    return 0;
  return (bb_loop_postorder[loop->num] < bb_loop_postorder[loc_loop->num]
          ? -1 : 1);
}

/* Iterates over all locations of REF in LOOP and its subloops calling
   fn.operator() with the location as argument.  When that operator
   returns true the iteration is stopped and true is returned.
   Otherwise false is returned.  */

template <typename FN>
static bool
for_all_locs_in_loop (struct loop *loop, im_mem_ref *ref, FN fn)
{
  unsigned i;
  mem_ref_loc *loc;

  /* Search for the cluster of locs in the accesses_in_loop vector
     which is sorted after postorder index of the loop father.  */
  loc = ref->accesses_in_loop.bsearch (loop, find_ref_loc_in_loop_cmp);
  if (!loc)
    return false;

  /* We have found one location inside loop or its sub-loops.  Iterate
     both forward and backward to cover the whole cluster.  */
  i = loc - ref->accesses_in_loop.address ();
  while (i > 0)
    {
      --i;
      mem_ref_loc *l = &ref->accesses_in_loop[i];
      if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt)))
        break;
      if (fn (l))
        return true;
    }
  for (i = loc - ref->accesses_in_loop.address ();
       i < ref->accesses_in_loop.length (); ++i)
    {
      mem_ref_loc *l = &ref->accesses_in_loop[i];
      if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt)))
        break;
      if (fn (l))
        return true;
    }

  return false;
}

/* Rewrites location LOC by TMP_VAR.  */

struct rewrite_mem_ref_loc
{
  rewrite_mem_ref_loc (tree tmp_var_) : tmp_var (tmp_var_) {}
  bool operator () (mem_ref_loc *loc);
  tree tmp_var;
};

bool
rewrite_mem_ref_loc::operator () (mem_ref_loc *loc)
{
  *loc->ref = tmp_var;
  update_stmt (loc->stmt);
  return false;
}

/* Rewrites all references to REF in LOOP by variable TMP_VAR.  */

static void
rewrite_mem_refs (struct loop *loop, im_mem_ref *ref, tree tmp_var)
{
  for_all_locs_in_loop (loop, ref, rewrite_mem_ref_loc (tmp_var));
}

/* Stores the first reference location in LOCP.  */

struct first_mem_ref_loc_1
{
  first_mem_ref_loc_1 (mem_ref_loc **locp_) : locp (locp_) {}
  bool operator () (mem_ref_loc *loc);
  mem_ref_loc **locp;
};

bool
first_mem_ref_loc_1::operator () (mem_ref_loc *loc)
{
  *locp = loc;
  return true;
}

/* Returns the first reference location to REF in LOOP.  */

static mem_ref_loc *
first_mem_ref_loc (struct loop *loop, im_mem_ref *ref)
{
  mem_ref_loc *locp = NULL;
  for_all_locs_in_loop (loop, ref, first_mem_ref_loc_1 (&locp));
  return locp;
}

struct prev_flag_edges {
  /* Edge to insert new flag comparison code.  */
  edge append_cond_position;

  /* Edge for fall through from previous flag comparison.  */
  edge last_cond_fallthru;
};

/* Helper function for execute_sm.  Emit code to store TMP_VAR into
   MEM along edge EX.

   The store is only done if MEM has changed.  We do this so no
   changes to MEM occur on code paths that did not originally store
   into it.

   The common case for execute_sm will transform:

     for (...) {
       if (foo)
         stuff;
       else
         MEM = TMP_VAR;
     }

   into:

     lsm = MEM;
     for (...) {
       if (foo)
         stuff;
       else
         lsm = TMP_VAR;
     }
     MEM = lsm;

  This function will generate:

     lsm = MEM;

     lsm_flag = false;
     ...
     for (...) {
       if (foo)
         stuff;
       else {
         lsm = TMP_VAR;
         lsm_flag = true;
       }
     }
     if (lsm_flag)      <--
       MEM = lsm;       <--
*/

static void
execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag)
{
  basic_block new_bb, then_bb, old_dest;
  bool loop_has_only_one_exit;
  edge then_old_edge, orig_ex = ex;
  gimple_stmt_iterator gsi;
  gimple *stmt;
  struct prev_flag_edges *prev_edges = (struct prev_flag_edges *) ex->aux;
  bool irr = ex->flags & EDGE_IRREDUCIBLE_LOOP;

  /* ?? Insert store after previous store if applicable.  See note
     below.  */
  if (prev_edges)
    ex = prev_edges->append_cond_position;

  loop_has_only_one_exit = single_pred_p (ex->dest);

  if (loop_has_only_one_exit)
    ex = split_block_after_labels (ex->dest);
  else
    {
      for (gphi_iterator gpi = gsi_start_phis (ex->dest);
           !gsi_end_p (gpi); gsi_next (&gpi))
        {
          gphi *phi = gpi.phi ();
          if (virtual_operand_p (gimple_phi_result (phi)))
            continue;

          /* When the destination has a non-virtual PHI node with multiple
             predecessors make sure we preserve the PHI structure by
             forcing a forwarder block so that hoisting of that PHI will
             still work.  */
          split_edge (ex);
          break;
        }
    }

  old_dest = ex->dest;
  new_bb = split_edge (ex);
  then_bb = create_empty_bb (new_bb);
  if (irr)
    then_bb->flags = BB_IRREDUCIBLE_LOOP;
  add_bb_to_loop (then_bb, new_bb->loop_father);

  gsi = gsi_start_bb (new_bb);
  stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node,
                            NULL_TREE, NULL_TREE);
  gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);

  gsi = gsi_start_bb (then_bb);
  /* Insert actual store.  */
  stmt = gimple_build_assign (unshare_expr (mem), tmp_var);
  gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);

  make_edge (new_bb, then_bb,
             EDGE_TRUE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0));
  make_edge (new_bb, old_dest,
             EDGE_FALSE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0));
  then_old_edge = make_edge (then_bb, old_dest,
                             EDGE_FALLTHRU | (irr ? EDGE_IRREDUCIBLE_LOOP : 0));

  set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb);

  if (prev_edges)
    {
      basic_block prevbb = prev_edges->last_cond_fallthru->src;
      redirect_edge_succ (prev_edges->last_cond_fallthru, new_bb);
      set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb);
      set_immediate_dominator (CDI_DOMINATORS, old_dest,
                               recompute_dominator (CDI_DOMINATORS, old_dest));
    }

  /* ?? Because stores may alias, they must happen in the exact
     sequence they originally happened.  Save the position right after
     the (_lsm) store we just created so we can continue appending after
     it and maintain the original order.  */
  {
    struct prev_flag_edges *p;

    if (orig_ex->aux)
      orig_ex->aux = NULL;
    alloc_aux_for_edge (orig_ex, sizeof (struct prev_flag_edges));
    p = (struct prev_flag_edges *) orig_ex->aux;
    p->append_cond_position = then_old_edge;
    p->last_cond_fallthru = find_edge (new_bb, old_dest);
    orig_ex->aux = (void *) p;
  }

  if (!loop_has_only_one_exit)
    for (gphi_iterator gpi = gsi_start_phis (old_dest);
         !gsi_end_p (gpi); gsi_next (&gpi))
      {
        gphi *phi = gpi.phi ();
        unsigned i;

        for (i = 0; i < gimple_phi_num_args (phi); i++)
          if (gimple_phi_arg_edge (phi, i)->src == new_bb)
            {
              tree arg = gimple_phi_arg_def (phi, i);
              add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION);
              update_stmt (phi);
            }
      }
  /* Remove the original fall through edge.  This was the
     single_succ_edge (new_bb).  */
  EDGE_SUCC (new_bb, 0)->flags &= ~EDGE_FALLTHRU;
}

/* When REF is set on the location, set flag indicating the store.  */

struct sm_set_flag_if_changed
{
  sm_set_flag_if_changed (tree flag_) : flag (flag_) {}
  bool operator () (mem_ref_loc *loc);
  tree flag;
};

bool
sm_set_flag_if_changed::operator () (mem_ref_loc *loc)
{
  /* Only set the flag for writes.  */
  if (is_gimple_assign (loc->stmt)
      && gimple_assign_lhs_ptr (loc->stmt) == loc->ref)
    {
      gimple_stmt_iterator gsi = gsi_for_stmt (loc->stmt);
      gimple *stmt = gimple_build_assign (flag, boolean_true_node);
      gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
    }
  return false;
}

/* Helper function for execute_sm.  On every location where REF is
   set, set an appropriate flag indicating the store.  */

static tree
execute_sm_if_changed_flag_set (struct loop *loop, im_mem_ref *ref)
{
  tree flag;
  char *str = get_lsm_tmp_name (ref->mem.ref, ~0, "_flag");
  flag = create_tmp_reg (boolean_type_node, str);
  for_all_locs_in_loop (loop, ref, sm_set_flag_if_changed (flag));
  return flag;
}

/* Executes store motion of memory reference REF from LOOP.
   Exits from the LOOP are stored in EXITS.  The initialization of the
   temporary variable is put to the preheader of the loop, and assignments
   to the reference from the temporary variable are emitted to exits.  */

static void
execute_sm (struct loop *loop, vec<edge> exits, im_mem_ref *ref)
{
  tree tmp_var, store_flag = NULL_TREE;
  unsigned i;
  gassign *load;
  struct fmt_data fmt_data;
  edge ex;
  struct lim_aux_data *lim_data;
  bool multi_threaded_model_p = false;
  gimple_stmt_iterator gsi;

  if (dump_file && (dump_flags & TDF_DETAILS))
    {
      fprintf (dump_file, "Executing store motion of ");
      print_generic_expr (dump_file, ref->mem.ref, 0);
      fprintf (dump_file, " from loop %d\n", loop->num);
    }

  tmp_var = create_tmp_reg (TREE_TYPE (ref->mem.ref),
                            get_lsm_tmp_name (ref->mem.ref, ~0));

  fmt_data.loop = loop;
  fmt_data.orig_loop = loop;
  for_each_index (&ref->mem.ref, force_move_till, &fmt_data);

  if (bb_in_transaction (loop_preheader_edge (loop)->src)
      || !PARAM_VALUE (PARAM_ALLOW_STORE_DATA_RACES))
    multi_threaded_model_p = true;

  if (multi_threaded_model_p)
    store_flag = execute_sm_if_changed_flag_set (loop, ref);

  rewrite_mem_refs (loop, ref, tmp_var);

  /* Emit the load code on a random exit edge or into the latch if
     the loop does not exit, so that we are sure it will be processed
     by move_computations after all dependencies.  */
  gsi = gsi_for_stmt (first_mem_ref_loc (loop, ref)->stmt);

  /* FIXME/TODO: For the multi-threaded variant, we could avoid this
     load altogether, since the store is predicated by a flag.  We
     could, do the load only if it was originally in the loop.  */
  load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref));
  lim_data = init_lim_data (load);
  lim_data->max_loop = loop;
  lim_data->tgt_loop = loop;
  gsi_insert_before (&gsi, load, GSI_SAME_STMT);

  if (multi_threaded_model_p)
    {
      load = gimple_build_assign (store_flag, boolean_false_node);
      lim_data = init_lim_data (load);
      lim_data->max_loop = loop;
      lim_data->tgt_loop = loop;
      gsi_insert_before (&gsi, load, GSI_SAME_STMT);
    }

  /* Sink the store to every exit from the loop.  */
  FOR_EACH_VEC_ELT (exits, i, ex)
    if (!multi_threaded_model_p)
      {
        gassign *store;
        store = gimple_build_assign (unshare_expr (ref->mem.ref), tmp_var);
        gsi_insert_on_edge (ex, store);
      }
    else
      execute_sm_if_changed (ex, ref->mem.ref, tmp_var, store_flag);
}

/* Hoists memory references MEM_REFS out of LOOP.  EXITS is the list of exit
   edges of the LOOP.  */

static void
hoist_memory_references (struct loop *loop, bitmap mem_refs,
                         vec<edge> exits)
{
  im_mem_ref *ref;
  unsigned  i;
  bitmap_iterator bi;

  EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
    {
      ref = memory_accesses.refs_list[i];
      execute_sm (loop, exits, ref);
    }
}

struct ref_always_accessed
{
  ref_always_accessed (struct loop *loop_, bool stored_p_)
      : loop (loop_), stored_p (stored_p_) {}
  bool operator () (mem_ref_loc *loc);
  struct loop *loop;
  bool stored_p;
};

bool
ref_always_accessed::operator () (mem_ref_loc *loc)
{
  struct loop *must_exec;

  if (!get_lim_data (loc->stmt))
    return false;

  /* If we require an always executed store make sure the statement
     stores to the reference.  */
  if (stored_p)
    {
      tree lhs = gimple_get_lhs (loc->stmt);
      if (!lhs
          || lhs != *loc->ref)
        return false;
    }

  must_exec = get_lim_data (loc->stmt)->always_executed_in;
  if (!must_exec)
    return false;

  if (must_exec == loop
      || flow_loop_nested_p (must_exec, loop))
    return true;

  return false;
}

/* Returns true if REF is always accessed in LOOP.  If STORED_P is true
   make sure REF is always stored to in LOOP.  */

static bool
ref_always_accessed_p (struct loop *loop, im_mem_ref *ref, bool stored_p)
{
  return for_all_locs_in_loop (loop, ref,
                               ref_always_accessed (loop, stored_p));
}

/* Returns true if REF1 and REF2 are independent.  */

static bool
refs_independent_p (im_mem_ref *ref1, im_mem_ref *ref2)
{
  if (ref1 == ref2)
    return true;

  if (dump_file && (dump_flags & TDF_DETAILS))
    fprintf (dump_file, "Querying dependency of refs %u and %u: ",
             ref1->id, ref2->id);

  if (mem_refs_may_alias_p (ref1, ref2, &memory_accesses.ttae_cache))
    {
      if (dump_file && (dump_flags & TDF_DETAILS))
        fprintf (dump_file, "dependent.\n");
      return false;
    }
  else
    {
      if (dump_file && (dump_flags & TDF_DETAILS))
        fprintf (dump_file, "independent.\n");
      return true;
    }
}

/* Mark REF dependent on stores or loads (according to STORED_P) in LOOP
   and its super-loops.  */

static void
record_dep_loop (struct loop *loop, im_mem_ref *ref, bool stored_p)
{
  /* We can propagate dependent-in-loop bits up the loop
     hierarchy to all outer loops.  */
  while (loop != current_loops->tree_root
         && bitmap_set_bit (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
    loop = loop_outer (loop);
}

/* Returns true if REF is independent on all other memory references in
   LOOP.  */

static bool
ref_indep_loop_p_1 (struct loop *loop, im_mem_ref *ref, bool stored_p)
{
  bitmap refs_to_check;
  unsigned i;
  bitmap_iterator bi;
  im_mem_ref *aref;

  if (stored_p)
    refs_to_check = &memory_accesses.refs_in_loop[loop->num];
  else
    refs_to_check = &memory_accesses.refs_stored_in_loop[loop->num];

  if (bitmap_bit_p (refs_to_check, UNANALYZABLE_MEM_ID))
    return false;

  EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi)
    {
      aref = memory_accesses.refs_list[i];
      if (!refs_independent_p (ref, aref))
        return false;
    }

  return true;
}

/* Returns true if REF is independent on all other memory references in
   LOOP.  Wrapper over ref_indep_loop_p_1, caching its results.  */

static bool
ref_indep_loop_p_2 (struct loop *loop, im_mem_ref *ref, bool stored_p)
{
  stored_p |= (ref->stored && bitmap_bit_p (ref->stored, loop->num));

  if (bitmap_bit_p (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
    return true;
  if (bitmap_bit_p (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
    return false;

  struct loop *inner = loop->inner;
  while (inner)
    {
      if (!ref_indep_loop_p_2 (inner, ref, stored_p))
        return false;
      inner = inner->next;
    }

  bool indep_p = ref_indep_loop_p_1 (loop, ref, stored_p);

  if (dump_file && (dump_flags & TDF_DETAILS))
    fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n",
             ref->id, loop->num, indep_p ? "independent" : "dependent");

  /* Record the computed result in the cache.  */
  if (indep_p)
    {
      if (bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p))
          && stored_p)
        {
          /* If it's independend against all refs then it's independent
             against stores, too.  */
          bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, false));
        }
    }
  else
    {
      record_dep_loop (loop, ref, stored_p);
      if (!stored_p)
        {
          /* If it's dependent against stores it's dependent against
             all refs, too.  */
          record_dep_loop (loop, ref, true);
        }
    }

  return indep_p;
}

/* Returns true if REF is independent on all other memory references in
   LOOP.  */

static bool
ref_indep_loop_p (struct loop *loop, im_mem_ref *ref)
{
  gcc_checking_assert (MEM_ANALYZABLE (ref));

  return ref_indep_loop_p_2 (loop, ref, false);
}

/* Returns true if we can perform store motion of REF from LOOP.  */

static bool
can_sm_ref_p (struct loop *loop, im_mem_ref *ref)
{
  tree base;

  /* Can't hoist unanalyzable refs.  */
  if (!MEM_ANALYZABLE (ref))
    return false;

  /* It should be movable.  */
  if (!is_gimple_reg_type (TREE_TYPE (ref->mem.ref))
      || TREE_THIS_VOLATILE (ref->mem.ref)
      || !for_each_index (&ref->mem.ref, may_move_till, loop))
    return false;

  /* If it can throw fail, we do not properly update EH info.  */
  if (tree_could_throw_p (ref->mem.ref))
    return false;

  /* If it can trap, it must be always executed in LOOP.
     Readonly memory locations may trap when storing to them, but
     tree_could_trap_p is a predicate for rvalues, so check that
     explicitly.  */
  base = get_base_address (ref->mem.ref);
  if ((tree_could_trap_p (ref->mem.ref)
       || (DECL_P (base) && TREE_READONLY (base)))
      && !ref_always_accessed_p (loop, ref, true))
    return false;

  /* And it must be independent on all other memory references
     in LOOP.  */
  if (!ref_indep_loop_p (loop, ref))
    return false;

  return true;
}

/* Marks the references in LOOP for that store motion should be performed
   in REFS_TO_SM.  SM_EXECUTED is the set of references for that store
   motion was performed in one of the outer loops.  */

static void
find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm)
{
  bitmap refs = &memory_accesses.all_refs_stored_in_loop[loop->num];
  unsigned i;
  bitmap_iterator bi;
  im_mem_ref *ref;

  EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi)
    {
      ref = memory_accesses.refs_list[i];
      if (can_sm_ref_p (loop, ref))
        bitmap_set_bit (refs_to_sm, i);
    }
}

/* Checks whether LOOP (with exits stored in EXITS array) is suitable
   for a store motion optimization (i.e. whether we can insert statement
   on its exits).  */

static bool
loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED,
                      vec<edge> exits)
{
  unsigned i;
  edge ex;

  FOR_EACH_VEC_ELT (exits, i, ex)
    if (ex->flags & (EDGE_ABNORMAL | EDGE_EH))
      return false;

  return true;
}

/* Try to perform store motion for all memory references modified inside
   LOOP.  SM_EXECUTED is the bitmap of the memory references for that
   store motion was executed in one of the outer loops.  */

static void
store_motion_loop (struct loop *loop, bitmap sm_executed)
{
  vec<edge> exits = get_loop_exit_edges (loop);
  struct loop *subloop;
  bitmap sm_in_loop = BITMAP_ALLOC (&lim_bitmap_obstack);

  if (loop_suitable_for_sm (loop, exits))
    {
      find_refs_for_sm (loop, sm_executed, sm_in_loop);
      hoist_memory_references (loop, sm_in_loop, exits);
    }
  exits.release ();

  bitmap_ior_into (sm_executed, sm_in_loop);
  for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
    store_motion_loop (subloop, sm_executed);
  bitmap_and_compl_into (sm_executed, sm_in_loop);
  BITMAP_FREE (sm_in_loop);
}

/* Try to perform store motion for all memory references modified inside
   loops.  */

static void
store_motion (void)
{
  struct loop *loop;
  bitmap sm_executed = BITMAP_ALLOC (&lim_bitmap_obstack);

  for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
    store_motion_loop (loop, sm_executed);

  BITMAP_FREE (sm_executed);
  gsi_commit_edge_inserts ();
}

/* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
   for each such basic block bb records the outermost loop for that execution
   of its header implies execution of bb.  CONTAINS_CALL is the bitmap of
   blocks that contain a nonpure call.  */

static void
fill_always_executed_in_1 (struct loop *loop, sbitmap contains_call)
{
  basic_block bb = NULL, *bbs, last = NULL;
  unsigned i;
  edge e;
  struct loop *inn_loop = loop;

  if (ALWAYS_EXECUTED_IN (loop->header) == NULL)
    {
      bbs = get_loop_body_in_dom_order (loop);

      for (i = 0; i < loop->num_nodes; i++)
        {
          edge_iterator ei;
          bb = bbs[i];

          if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
            last = bb;

          if (bitmap_bit_p (contains_call, bb->index))
            break;

          FOR_EACH_EDGE (e, ei, bb->succs)
            if (!flow_bb_inside_loop_p (loop, e->dest))
              break;
          if (e)
            break;

          /* A loop might be infinite (TODO use simple loop analysis
             to disprove this if possible).  */
          if (bb->flags & BB_IRREDUCIBLE_LOOP)
            break;

          if (!flow_bb_inside_loop_p (inn_loop, bb))
            break;

          if (bb->loop_father->header == bb)
            {
              if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
                break;

              /* In a loop that is always entered we may proceed anyway.
                 But record that we entered it and stop once we leave it.  */
              inn_loop = bb->loop_father;
            }
        }

      while (1)
        {
          SET_ALWAYS_EXECUTED_IN (last, loop);
          if (last == loop->header)
            break;
          last = get_immediate_dominator (CDI_DOMINATORS, last);
        }

      free (bbs);
    }

  for (loop = loop->inner; loop; loop = loop->next)
    fill_always_executed_in_1 (loop, contains_call);
}

/* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e.
   for each such basic block bb records the outermost loop for that execution
   of its header implies execution of bb.  */

static void
fill_always_executed_in (void)
{
  sbitmap contains_call = sbitmap_alloc (last_basic_block_for_fn (cfun));
  basic_block bb;
  struct loop *loop;

  bitmap_clear (contains_call);
  FOR_EACH_BB_FN (bb, cfun)
    {
      gimple_stmt_iterator gsi;
      for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
        {
          if (nonpure_call_p (gsi_stmt (gsi)))
            break;
        }

      if (!gsi_end_p (gsi))
        bitmap_set_bit (contains_call, bb->index);
    }

  for (loop = current_loops->tree_root->inner; loop; loop = loop->next)
    fill_always_executed_in_1 (loop, contains_call);

  sbitmap_free (contains_call);
}


/* Compute the global information needed by the loop invariant motion pass.  */

static void
tree_ssa_lim_initialize (void)
{
  struct loop *loop;
  unsigned i;

  bitmap_obstack_initialize (&lim_bitmap_obstack);
  gcc_obstack_init (&mem_ref_obstack);
  lim_aux_data_map = new hash_map<gimple *, lim_aux_data *>;

  if (flag_tm)
    compute_transaction_bits ();

  alloc_aux_for_edges (0);

  memory_accesses.refs = new hash_table<mem_ref_hasher> (100);
  memory_accesses.refs_list.create (100);
  /* Allocate a special, unanalyzable mem-ref with ID zero.  */
  memory_accesses.refs_list.quick_push
    (mem_ref_alloc (error_mark_node, 0, UNANALYZABLE_MEM_ID));

  memory_accesses.refs_in_loop.create (number_of_loops (cfun));
  memory_accesses.refs_in_loop.quick_grow (number_of_loops (cfun));
  memory_accesses.refs_stored_in_loop.create (number_of_loops (cfun));
  memory_accesses.refs_stored_in_loop.quick_grow (number_of_loops (cfun));
  memory_accesses.all_refs_stored_in_loop.create (number_of_loops (cfun));
  memory_accesses.all_refs_stored_in_loop.quick_grow (number_of_loops (cfun));

  for (i = 0; i < number_of_loops (cfun); i++)
    {
      bitmap_initialize (&memory_accesses.refs_in_loop[i],
                         &lim_bitmap_obstack);
      bitmap_initialize (&memory_accesses.refs_stored_in_loop[i],
                         &lim_bitmap_obstack);
      bitmap_initialize (&memory_accesses.all_refs_stored_in_loop[i],
                         &lim_bitmap_obstack);
    }

  memory_accesses.ttae_cache = NULL;

  /* Initialize bb_loop_postorder with a mapping from loop->num to
     its postorder index.  */
  i = 0;
  bb_loop_postorder = XNEWVEC (unsigned, number_of_loops (cfun));
  FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
    bb_loop_postorder[loop->num] = i++;
}

/* Cleans up after the invariant motion pass.  */

static void
tree_ssa_lim_finalize (void)
{
  basic_block bb;
  unsigned i;
  im_mem_ref *ref;

  free_aux_for_edges ();

  FOR_EACH_BB_FN (bb, cfun)
    SET_ALWAYS_EXECUTED_IN (bb, NULL);

  bitmap_obstack_release (&lim_bitmap_obstack);
  delete lim_aux_data_map;

  delete memory_accesses.refs;
  memory_accesses.refs = NULL;

  FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref)
    memref_free (ref);
  memory_accesses.refs_list.release ();
  obstack_free (&mem_ref_obstack, NULL);

  memory_accesses.refs_in_loop.release ();
  memory_accesses.refs_stored_in_loop.release ();
  memory_accesses.all_refs_stored_in_loop.release ();

  if (memory_accesses.ttae_cache)
    free_affine_expand_cache (&memory_accesses.ttae_cache);

  free (bb_loop_postorder);
}

/* Moves invariants from loops.  Only "expensive" invariants are moved out --
   i.e. those that are likely to be win regardless of the register pressure.  */

unsigned int
tree_ssa_lim (void)
{
  unsigned int todo;

  tree_ssa_lim_initialize ();

  /* Gathers information about memory accesses in the loops.  */
  analyze_memory_references ();

  /* Fills ALWAYS_EXECUTED_IN information for basic blocks.  */
  fill_always_executed_in ();

  /* For each statement determine the outermost loop in that it is
     invariant and cost for computing the invariant.  */
  invariantness_dom_walker (CDI_DOMINATORS)
    .walk (cfun->cfg->x_entry_block_ptr);

  /* Execute store motion.  Force the necessary invariants to be moved
     out of the loops as well.  */
  store_motion ();

  /* Move the expressions that are expensive enough.  */
  todo = move_computations ();

  tree_ssa_lim_finalize ();

  return todo;
}

/* Loop invariant motion pass.  */

namespace {

const pass_data pass_data_lim =
{
  GIMPLE_PASS, /* type */
  "lim", /* name */
  OPTGROUP_LOOP, /* optinfo_flags */
  TV_LIM, /* tv_id */
  PROP_cfg, /* properties_required */
  0, /* properties_provided */
  0, /* properties_destroyed */
  0, /* todo_flags_start */
  0, /* todo_flags_finish */
};

class pass_lim : public gimple_opt_pass
{
public:
  pass_lim (gcc::context *ctxt)
    : gimple_opt_pass (pass_data_lim, ctxt)
  {}

  /* opt_pass methods: */
  opt_pass * clone () { return new pass_lim (m_ctxt); }
  virtual bool gate (function *) { return flag_tree_loop_im != 0; }
  virtual unsigned int execute (function *);

}; // class pass_lim

unsigned int
pass_lim::execute (function *fun)
{
  if (number_of_loops (fun) <= 1)
    return 0;

  return tree_ssa_lim ();
}

} // anon namespace

gimple_opt_pass *
make_pass_lim (gcc::context *ctxt)
{
  return new pass_lim (ctxt);
}