* cgraphunit.c, config/arm/arm.c, config/m68k/m68k.c,

[official-gcc.git] / gcc / ipa-inline.c

/* Inlining decision heuristics.
   Copyright (C) 2003, 2004 Free Software Foundation, Inc.
   Contributed by Jan Hubicka

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 2, 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 COPYING.  If not, write to the Free
Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA.  */

/*  Inlining decision heuristics

    We separate inlining decisions from the inliner itself and store it
    inside callgraph as so called inline plan.  Refer to cgraph.c
    documentation about particular representation of inline plans in the
    callgraph.

    There are three major parts of this file:

    cgraph_mark_inline implementation

      This function allows to mark given call inline and performs necessary
      modifications of cgraph (production of the clones and updating overall
      statistics)

    inlining heuristics limits

      These functions allow to check that particular inlining is allowed
      by the limits specified by user (allowed function growth, overall unit
      growth and so on).

    inlining heuristics

      This is implementation of IPA pass aiming to get as much of benefit
      from inlining obeying the limits checked above.

      The implementation of particular heuristics is separated from
      the rest of code to make it easier to replace it with more complicated
      implementation in the future.  The rest of inlining code acts as a
      library aimed to modify the callgraph and verify that the parameters
      on code size growth fits.

      To mark given call inline, use cgraph_mark_inline function, the
      verification is performed by cgraph_default_inline_p and
      cgraph_check_inline_limits.

      The heuristics implements simple knapsack style algorithm ordering
      all functions by their "profitability" (estimated by code size growth)
      and inlining them in priority order.

      cgraph_decide_inlining implements heuristics taking whole callgraph
      into account, while cgraph_decide_inlining_incrementally considers
      only one function at a time and is used in non-unit-at-a-time mode. 

   The inliner itself is split into several passes:

   pass_inline_parameters

     This pass computes local properties of functions that are used by inliner:
     estimated function body size, whether function is inlinable at all and
     stack frame consumption.

     Before executing any of inliner passes, this local pass has to be applied
     to each function in the callgraph (ie run as subpass of some earlier
     IPA pass).  The results are made out of date by any optimization applied
     on the function body.

   pass_early_inlining

     Simple local inlining pass inlining callees into current function.  This
     pass makes no global whole compilation unit analysis and this when allowed
     to do inlining expanding code size it might result in unbounded growth of
     whole unit.

     This is the main inlining pass in non-unit-at-a-time.

     With unit-at-a-time the pass is run during conversion into SSA form.
     Only functions already converted into SSA form are inlined, so the
     conversion must happen in topological order on the callgraph (that is
     maintained by pass manager).  The functions after inlining are early
     optimized so the early inliner sees unoptimized function itself, but
     all considered callees are already optimized allowing it to unfold
     abstraction penalty on C++ effectively and cheaply.

   pass_ipa_early_inlining

     With profiling, the early inlining is also necessary to reduce
     instrumentation costs on program with high abstraction penalty (doing
     many redundant calls).  This can't happen in parallel with early
     optimization and profile instrumentation, because we would end up
     re-instrumenting already instrumented function bodies we brought in via
     inlining.

     To avoid this, this pass is executed as IPA pass before profiling.  It is
     simple wrapper to pass_early_inlining and ensures first inlining.

   pass_ipa_inline

     This is the main pass implementing simple greedy algorithm to do inlining
     of small functions that results in overall growth of compilation unit and
     inlining of functions called once.  The pass compute just so called inline
     plan (representation of inlining to be done in callgraph) and unlike early
     inlining it is not performing the inlining itself.

   pass_apply_inline

     This pass performs actual inlining according to pass_ipa_inline on given
     function.  Possible the function body before inlining is saved when it is
     needed for further inlining later.
 */

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "tree.h"
#include "tree-inline.h"
#include "langhooks.h"
#include "flags.h"
#include "cgraph.h"
#include "diagnostic.h"
#include "timevar.h"
#include "params.h"
#include "fibheap.h"
#include "intl.h"
#include "tree-pass.h"
#include "hashtab.h"
#include "coverage.h"
#include "ggc.h"
#include "tree-flow.h"

/* Mode incremental inliner operate on:

   In ALWAYS_INLINE only functions marked
   always_inline are inlined.  This mode is used after detecting cycle during
   flattening.

   In SIZE mode, only functions that reduce function body size after inlining
   are inlined, this is used during early inlining.

   In SPEED mode, all small functions are inlined.  This might result in
   unbounded growth of compilation unit and is used only in non-unit-at-a-time
   mode.

   in ALL mode, everything is inlined.  This is used during flattening.  */
enum inlining_mode {
  INLINE_NONE = 0,
  INLINE_ALWAYS_INLINE,
  INLINE_SIZE,
  INLINE_SPEED,
  INLINE_ALL
};
static bool
cgraph_decide_inlining_incrementally (struct cgraph_node *, enum inlining_mode,
                                      int);


/* Statistics we collect about inlining algorithm.  */
static int ncalls_inlined;
static int nfunctions_inlined;
static int overall_insns;
static gcov_type max_count;

/* Estimate size of the function after inlining WHAT into TO.  */

static int
cgraph_estimate_size_after_inlining (int times, struct cgraph_node *to,
                                     struct cgraph_node *what)
{
  int size;
  tree fndecl = what->decl, arg;
  int call_insns = PARAM_VALUE (PARAM_INLINE_CALL_COST);

  for (arg = DECL_ARGUMENTS (fndecl); arg; arg = TREE_CHAIN (arg))
    call_insns += estimate_move_cost (TREE_TYPE (arg));
  size = (what->global.insns - call_insns) * times + to->global.insns;
  gcc_assert (size >= 0);
  return size;
}

/* E is expected to be an edge being inlined.  Clone destination node of
   the edge and redirect it to the new clone.
   DUPLICATE is used for bookkeeping on whether we are actually creating new
   clones or re-using node originally representing out-of-line function call.
   */
void
cgraph_clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, bool update_original)
{
  HOST_WIDE_INT peak;
  if (duplicate)
    {
      /* We may eliminate the need for out-of-line copy to be output.
         In that case just go ahead and re-use it.  */
      if (!e->callee->callers->next_caller
          && !e->callee->needed
          && flag_unit_at_a_time)
        {
          gcc_assert (!e->callee->global.inlined_to);
          if (DECL_SAVED_TREE (e->callee->decl))
            overall_insns -= e->callee->global.insns, nfunctions_inlined++;
          duplicate = false;
        }
      else
        {
          struct cgraph_node *n;
          n = cgraph_clone_node (e->callee, e->count, e->loop_nest, 
                                 update_original);
          cgraph_redirect_edge_callee (e, n);
        }
    }

  if (e->caller->global.inlined_to)
    e->callee->global.inlined_to = e->caller->global.inlined_to;
  else
    e->callee->global.inlined_to = e->caller;
  e->callee->global.stack_frame_offset
    = e->caller->global.stack_frame_offset + e->caller->local.estimated_self_stack_size;
  peak = e->callee->global.stack_frame_offset + e->callee->local.estimated_self_stack_size;
  if (e->callee->global.inlined_to->global.estimated_stack_size < peak)
    e->callee->global.inlined_to->global.estimated_stack_size = peak;

  /* Recursively clone all bodies.  */
  for (e = e->callee->callees; e; e = e->next_callee)
    if (!e->inline_failed)
      cgraph_clone_inlined_nodes (e, duplicate, update_original);
}

/* Mark edge E as inlined and update callgraph accordingly. 
   UPDATE_ORIGINAL specify whether profile of original function should be
   updated. */

void
cgraph_mark_inline_edge (struct cgraph_edge *e, bool update_original)
{
  int old_insns = 0, new_insns = 0;
  struct cgraph_node *to = NULL, *what;

  if (e->callee->inline_decl)
    cgraph_redirect_edge_callee (e, cgraph_node (e->callee->inline_decl));

  gcc_assert (e->inline_failed);
  e->inline_failed = NULL;

  if (!e->callee->global.inlined && flag_unit_at_a_time)
    DECL_POSSIBLY_INLINED (e->callee->decl) = true;
  e->callee->global.inlined = true;

  cgraph_clone_inlined_nodes (e, true, update_original);

  what = e->callee;

  /* Now update size of caller and all functions caller is inlined into.  */
  for (;e && !e->inline_failed; e = e->caller->callers)
    {
      old_insns = e->caller->global.insns;
      new_insns = cgraph_estimate_size_after_inlining (1, e->caller,
                                                       what);
      gcc_assert (new_insns >= 0);
      to = e->caller;
      to->global.insns = new_insns;
    }
  gcc_assert (what->global.inlined_to == to);
  if (new_insns > old_insns)
    overall_insns += new_insns - old_insns;
  ncalls_inlined++;
}

/* Mark all calls of EDGE->CALLEE inlined into EDGE->CALLER.
   Return following unredirected edge in the list of callers
   of EDGE->CALLEE  */

static struct cgraph_edge *
cgraph_mark_inline (struct cgraph_edge *edge)
{
  struct cgraph_node *to = edge->caller;
  struct cgraph_node *what = edge->callee;
  struct cgraph_edge *e, *next;
  int times = 0;

  /* Look for all calls, mark them inline and clone recursively
     all inlined functions.  */
  for (e = what->callers; e; e = next)
    {
      next = e->next_caller;
      if (e->caller == to && e->inline_failed)
        {
          cgraph_mark_inline_edge (e, true);
          if (e == edge)
            edge = next;
          times++;
        }
    }
  gcc_assert (times);
  return edge;
}

/* Estimate the growth caused by inlining NODE into all callees.  */

static int
cgraph_estimate_growth (struct cgraph_node *node)
{
  int growth = 0;
  struct cgraph_edge *e;
  if (node->global.estimated_growth != INT_MIN)
    return node->global.estimated_growth;

  for (e = node->callers; e; e = e->next_caller)
    if (e->inline_failed)
      growth += (cgraph_estimate_size_after_inlining (1, e->caller, node)
                 - e->caller->global.insns);

  /* ??? Wrong for self recursive functions or cases where we decide to not
     inline for different reasons, but it is not big deal as in that case
     we will keep the body around, but we will also avoid some inlining.  */
  if (!node->needed && !DECL_EXTERNAL (node->decl))
    growth -= node->global.insns;

  node->global.estimated_growth = growth;
  return growth;
}

/* Return false when inlining WHAT into TO is not good idea
   as it would cause too large growth of function bodies.  
   When ONE_ONLY is true, assume that only one call site is going
   to be inlined, otherwise figure out how many call sites in
   TO calls WHAT and verify that all can be inlined.
   */

static bool
cgraph_check_inline_limits (struct cgraph_node *to, struct cgraph_node *what,
                            const char **reason, bool one_only)
{
  int times = 0;
  struct cgraph_edge *e;
  int newsize;
  int limit;
  HOST_WIDE_INT stack_size_limit, inlined_stack;

  if (one_only)
    times = 1;
  else
    for (e = to->callees; e; e = e->next_callee)
      if (e->callee == what)
        times++;

  if (to->global.inlined_to)
    to = to->global.inlined_to;

  /* When inlining large function body called once into small function,
     take the inlined function as base for limiting the growth.  */
  if (to->local.self_insns > what->local.self_insns)
    limit = to->local.self_insns;
  else
    limit = what->local.self_insns;

  limit += limit * PARAM_VALUE (PARAM_LARGE_FUNCTION_GROWTH) / 100;

  /* Check the size after inlining against the function limits.  But allow
     the function to shrink if it went over the limits by forced inlining.  */
  newsize = cgraph_estimate_size_after_inlining (times, to, what);
  if (newsize >= to->global.insns
      && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
      && newsize > limit)
    {
      if (reason)
        *reason = N_("--param large-function-growth limit reached");
      return false;
    }

  stack_size_limit = to->local.estimated_self_stack_size;

  stack_size_limit += stack_size_limit * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100;

  inlined_stack = (to->global.stack_frame_offset
                   + to->local.estimated_self_stack_size
                   + what->global.estimated_stack_size);
  if (inlined_stack  > stack_size_limit
      && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
    {
      if (reason)
        *reason = N_("--param large-stack-frame-growth limit reached");
      return false;
    }
  return true;
}

/* Return true when function N is small enough to be inlined.  */

bool
cgraph_default_inline_p (struct cgraph_node *n, const char **reason)
{
  tree decl = n->decl;

  if (n->inline_decl)
    decl = n->inline_decl;
  if (!DECL_INLINE (decl))
    {
      if (reason)
        *reason = N_("function not inlinable");
      return false;
    }

  if (!DECL_STRUCT_FUNCTION (decl)->cfg)
    {
      if (reason)
        *reason = N_("function body not available");
      return false;
    }

  if (DECL_DECLARED_INLINE_P (decl))
    {
      if (n->global.insns >= MAX_INLINE_INSNS_SINGLE)
        {
          if (reason)
            *reason = N_("--param max-inline-insns-single limit reached");
          return false;
        }
    }
  else
    {
      if (n->global.insns >= MAX_INLINE_INSNS_AUTO)
        {
          if (reason)
            *reason = N_("--param max-inline-insns-auto limit reached");
          return false;
        }
    }

  return true;
}

/* Return true when inlining WHAT would create recursive inlining.
   We call recursive inlining all cases where same function appears more than
   once in the single recursion nest path in the inline graph.  */

static bool
cgraph_recursive_inlining_p (struct cgraph_node *to,
                             struct cgraph_node *what,
                             const char **reason)
{
  bool recursive;
  if (to->global.inlined_to)
    recursive = what->decl == to->global.inlined_to->decl;
  else
    recursive = what->decl == to->decl;
  /* Marking recursive function inline has sane semantic and thus we should
     not warn on it.  */
  if (recursive && reason)
    *reason = (what->local.disregard_inline_limits
               ? N_("recursive inlining") : "");
  return recursive;
}

/* Return true if the call can be hot.  */
static bool
cgraph_maybe_hot_edge_p (struct cgraph_edge *edge)
{
  if (profile_info && flag_branch_probabilities
      && (edge->count
          <= profile_info->sum_max / PARAM_VALUE (HOT_BB_COUNT_FRACTION)))
    return false;
  return true;
}

/* A cost model driving the inlining heuristics in a way so the edges with
   smallest badness are inlined first.  After each inlining is performed
   the costs of all caller edges of nodes affected are recomputed so the
   metrics may accurately depend on values such as number of inlinable callers
   of the function or function body size.

   With profiling we use number of executions of each edge to drive the cost.
   We also should distinguish hot and cold calls where the cold calls are
   inlined into only when code size is overall improved.  
   */

static int
cgraph_edge_badness (struct cgraph_edge *edge)
{
  if (max_count)
    {
      int growth =
        cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee);
      growth -= edge->caller->global.insns;

      /* Always prefer inlining saving code size.  */
      if (growth <= 0)
        return INT_MIN - growth;
      return ((int)((double)edge->count * INT_MIN / max_count)) / growth;
    }
  else
  {
    int nest = MIN (edge->loop_nest, 8);
    int badness = cgraph_estimate_growth (edge->callee) * 256;

    /* Decrease badness if call is nested.  */
    if (badness > 0)    
      badness >>= nest;
    else
      badness <<= nest;

    /* Make recursive inlining happen always after other inlining is done.  */
    if (cgraph_recursive_inlining_p (edge->caller, edge->callee, NULL))
      return badness + 1;
    else
      return badness;
  }
}

/* Recompute heap nodes for each of caller edge.  */

static void
update_caller_keys (fibheap_t heap, struct cgraph_node *node,
                    bitmap updated_nodes)
{
  struct cgraph_edge *edge;
  const char *failed_reason;

  if (!node->local.inlinable || node->local.disregard_inline_limits
      || node->global.inlined_to)
    return;
  if (bitmap_bit_p (updated_nodes, node->uid))
    return;
  bitmap_set_bit (updated_nodes, node->uid);
  node->global.estimated_growth = INT_MIN;

  if (!node->local.inlinable)
    return;
  /* Prune out edges we won't inline into anymore.  */
  if (!cgraph_default_inline_p (node, &failed_reason))
    {
      for (edge = node->callers; edge; edge = edge->next_caller)
        if (edge->aux)
          {
            fibheap_delete_node (heap, edge->aux);
            edge->aux = NULL;
            if (edge->inline_failed)
              edge->inline_failed = failed_reason;
          }
      return;
    }

  for (edge = node->callers; edge; edge = edge->next_caller)
    if (edge->inline_failed)
      {
        int badness = cgraph_edge_badness (edge);
        if (edge->aux)
          {
            fibnode_t n = edge->aux;
            gcc_assert (n->data == edge);
            if (n->key == badness)
              continue;

            /* fibheap_replace_key only increase the keys.  */
            if (fibheap_replace_key (heap, n, badness))
              continue;
            fibheap_delete_node (heap, edge->aux);
          }
        edge->aux = fibheap_insert (heap, badness, edge);
      }
}

/* Recompute heap nodes for each of caller edges of each of callees.  */

static void
update_callee_keys (fibheap_t heap, struct cgraph_node *node,
                    bitmap updated_nodes)
{
  struct cgraph_edge *e;
  node->global.estimated_growth = INT_MIN;

  for (e = node->callees; e; e = e->next_callee)
    if (e->inline_failed)
      update_caller_keys (heap, e->callee, updated_nodes);
    else if (!e->inline_failed)
      update_callee_keys (heap, e->callee, updated_nodes);
}

/* Enqueue all recursive calls from NODE into priority queue depending on
   how likely we want to recursively inline the call.  */

static void
lookup_recursive_calls (struct cgraph_node *node, struct cgraph_node *where,
                        fibheap_t heap)
{
  static int priority;
  struct cgraph_edge *e;
  for (e = where->callees; e; e = e->next_callee)
    if (e->callee == node)
      {
        /* When profile feedback is available, prioritize by expected number
           of calls.  Without profile feedback we maintain simple queue
           to order candidates via recursive depths.  */
        fibheap_insert (heap,
                        !max_count ? priority++
                        : -(e->count / ((max_count + (1<<24) - 1) / (1<<24))),
                        e);
      }
  for (e = where->callees; e; e = e->next_callee)
    if (!e->inline_failed)
      lookup_recursive_calls (node, e->callee, heap);
}

/* Decide on recursive inlining: in the case function has recursive calls,
   inline until body size reaches given argument.  */

static bool
cgraph_decide_recursive_inlining (struct cgraph_node *node)
{
  int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO);
  int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO);
  int probability = PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY);
  fibheap_t heap;
  struct cgraph_edge *e;
  struct cgraph_node *master_clone, *next;
  int depth = 0;
  int n = 0;

  if (DECL_DECLARED_INLINE_P (node->decl))
    {
      limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE);
      max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH);
    }

  /* Make sure that function is small enough to be considered for inlining.  */
  if (!max_depth
      || cgraph_estimate_size_after_inlining (1, node, node)  >= limit)
    return false;
  heap = fibheap_new ();
  lookup_recursive_calls (node, node, heap);
  if (fibheap_empty (heap))
    {
      fibheap_delete (heap);
      return false;
    }

  if (dump_file)
    fprintf (dump_file, 
             "  Performing recursive inlining on %s\n",
             cgraph_node_name (node));

  /* We need original clone to copy around.  */
  master_clone = cgraph_clone_node (node, node->count, 1, false);
  master_clone->needed = true;
  for (e = master_clone->callees; e; e = e->next_callee)
    if (!e->inline_failed)
      cgraph_clone_inlined_nodes (e, true, false);

  /* Do the inlining and update list of recursive call during process.  */
  while (!fibheap_empty (heap)
         && (cgraph_estimate_size_after_inlining (1, node, master_clone)
             <= limit))
    {
      struct cgraph_edge *curr = fibheap_extract_min (heap);
      struct cgraph_node *cnode;

      depth = 1;
      for (cnode = curr->caller;
           cnode->global.inlined_to; cnode = cnode->callers->caller)
        if (node->decl == curr->callee->decl)
          depth++;
      if (depth > max_depth)
        {
          if (dump_file)
            fprintf (dump_file, 
                     "   maxmal depth reached\n");
          continue;
        }

      if (max_count)
        {
          if (!cgraph_maybe_hot_edge_p (curr))
            {
              if (dump_file)
                fprintf (dump_file, "   Not inlining cold call\n");
              continue;
            }
          if (curr->count * 100 / node->count < probability)
            {
              if (dump_file)
                fprintf (dump_file, 
                         "   Probability of edge is too small\n");
              continue;
            }
        }

      if (dump_file)
        {
          fprintf (dump_file, 
                   "   Inlining call of depth %i", depth);
          if (node->count)
            {
              fprintf (dump_file, " called approx. %.2f times per call",
                       (double)curr->count / node->count);
            }
          fprintf (dump_file, "\n");
        }
      cgraph_redirect_edge_callee (curr, master_clone);
      cgraph_mark_inline_edge (curr, false);
      lookup_recursive_calls (node, curr->callee, heap);
      n++;
    }
  if (!fibheap_empty (heap) && dump_file)
    fprintf (dump_file, "    Recursive inlining growth limit met.\n");

  fibheap_delete (heap);
  if (dump_file)
    fprintf (dump_file, 
             "\n   Inlined %i times, body grown from %i to %i insns\n", n,
             master_clone->global.insns, node->global.insns);

  /* Remove master clone we used for inlining.  We rely that clones inlined
     into master clone gets queued just before master clone so we don't
     need recursion.  */
  for (node = cgraph_nodes; node != master_clone;
       node = next)
    {
      next = node->next;
      if (node->global.inlined_to == master_clone)
        cgraph_remove_node (node);
    }
  cgraph_remove_node (master_clone);
  /* FIXME: Recursive inlining actually reduces number of calls of the
     function.  At this place we should probably walk the function and
     inline clones and compensate the counts accordingly.  This probably
     doesn't matter much in practice.  */
  return n > 0;
}

/* Set inline_failed for all callers of given function to REASON.  */

static void
cgraph_set_inline_failed (struct cgraph_node *node, const char *reason)
{
  struct cgraph_edge *e;

  if (dump_file)
    fprintf (dump_file, "Inlining failed: %s\n", reason);
  for (e = node->callers; e; e = e->next_caller)
    if (e->inline_failed)
      e->inline_failed = reason;
}

/* Given whole compilation unit estimate of INSNS, compute how large we can
   allow the unit to grow.  */
static int
compute_max_insns (int insns)
{
  int max_insns = insns;
  if (max_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
    max_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);

  return ((HOST_WIDEST_INT) max_insns
          * (100 + PARAM_VALUE (PARAM_INLINE_UNIT_GROWTH)) / 100);
}

/* We use greedy algorithm for inlining of small functions:
   All inline candidates are put into prioritized heap based on estimated
   growth of the overall number of instructions and then update the estimates.

   INLINED and INLINED_CALEES are just pointers to arrays large enough
   to be passed to cgraph_inlined_into and cgraph_inlined_callees.  */

static void
cgraph_decide_inlining_of_small_functions (void)
{
  struct cgraph_node *node;
  struct cgraph_edge *edge;
  const char *failed_reason;
  fibheap_t heap = fibheap_new ();
  bitmap updated_nodes = BITMAP_ALLOC (NULL);
  int min_insns, max_insns;

  if (dump_file)
    fprintf (dump_file, "\nDeciding on smaller functions:\n");

  /* Put all inline candidates into the heap.  */

  for (node = cgraph_nodes; node; node = node->next)
    {
      if (!node->local.inlinable || !node->callers
          || node->local.disregard_inline_limits)
        continue;
      if (dump_file)
        fprintf (dump_file, "Considering inline candidate %s.\n", cgraph_node_name (node));

      node->global.estimated_growth = INT_MIN;
      if (!cgraph_default_inline_p (node, &failed_reason))
        {
          cgraph_set_inline_failed (node, failed_reason);
          continue;
        }

      for (edge = node->callers; edge; edge = edge->next_caller)
        if (edge->inline_failed)
          {
            gcc_assert (!edge->aux);
            edge->aux = fibheap_insert (heap, cgraph_edge_badness (edge), edge);
          }
    }

  max_insns = compute_max_insns (overall_insns);
  min_insns = overall_insns;

  while (overall_insns <= max_insns && (edge = fibheap_extract_min (heap)))
    {
      int old_insns = overall_insns;
      struct cgraph_node *where;
      int growth =
        cgraph_estimate_size_after_inlining (1, edge->caller, edge->callee);

      growth -= edge->caller->global.insns;

      if (dump_file)
        {
          fprintf (dump_file, 
                   "\nConsidering %s with %i insns\n",
                   cgraph_node_name (edge->callee),
                   edge->callee->global.insns);
          fprintf (dump_file, 
                   " to be inlined into %s\n"
                   " Estimated growth after inlined into all callees is %+i insns.\n"
                   " Estimated badness is %i.\n",
                   cgraph_node_name (edge->caller),
                   cgraph_estimate_growth (edge->callee),
                   cgraph_edge_badness (edge));
          if (edge->count)
            fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n", edge->count);
        }
      gcc_assert (edge->aux);
      edge->aux = NULL;
      if (!edge->inline_failed)
        continue;

      /* When not having profile info ready we don't weight by any way the
         position of call in procedure itself.  This means if call of
         function A from function B seems profitable to inline, the recursive
         call of function A in inline copy of A in B will look profitable too
         and we end up inlining until reaching maximal function growth.  This
         is not good idea so prohibit the recursive inlining.

         ??? When the frequencies are taken into account we might not need this
         restriction.   */
      if (!max_count)
        {
          where = edge->caller;
          while (where->global.inlined_to)
            {
              if (where->decl == edge->callee->decl)
                break;
              where = where->callers->caller;
            }
          if (where->global.inlined_to)
            {
              edge->inline_failed
                = (edge->callee->local.disregard_inline_limits ? N_("recursive inlining") : "");
              if (dump_file)
                fprintf (dump_file, " inline_failed:Recursive inlining performed only for function itself.\n");
              continue;
            }
        }

      if (!cgraph_maybe_hot_edge_p (edge) && growth > 0)
        {
          if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
                                            &edge->inline_failed))
            {
              edge->inline_failed = 
                N_("call is unlikely");
              if (dump_file)
                fprintf (dump_file, " inline_failed:%s.\n", edge->inline_failed);
            }
          continue;
        }
      if (!cgraph_default_inline_p (edge->callee, &edge->inline_failed))
        {
          if (!cgraph_recursive_inlining_p (edge->caller, edge->callee,
                                            &edge->inline_failed))
            {
              if (dump_file)
                fprintf (dump_file, " inline_failed:%s.\n", edge->inline_failed);
            }
          continue;
        }
      if (cgraph_recursive_inlining_p (edge->caller, edge->callee,
                                       &edge->inline_failed))
        {
          where = edge->caller;
          if (where->global.inlined_to)
            where = where->global.inlined_to;
          if (!cgraph_decide_recursive_inlining (where))
            continue;
          update_callee_keys (heap, where, updated_nodes);
        }
      else
        {
          struct cgraph_node *callee;
          if (!cgraph_check_inline_limits (edge->caller, edge->callee,
                                           &edge->inline_failed, true))
            {
              if (dump_file)
                fprintf (dump_file, " Not inlining into %s:%s.\n",
                         cgraph_node_name (edge->caller), edge->inline_failed);
              continue;
            }
          callee = edge->callee;
          cgraph_mark_inline_edge (edge, true);
          update_callee_keys (heap, callee, updated_nodes);
        }
      where = edge->caller;
      if (where->global.inlined_to)
        where = where->global.inlined_to;

      /* Our profitability metric can depend on local properties
         such as number of inlinable calls and size of the function body.
         After inlining these properties might change for the function we
         inlined into (since it's body size changed) and for the functions
         called by function we inlined (since number of it inlinable callers
         might change).  */
      update_caller_keys (heap, where, updated_nodes);
      bitmap_clear (updated_nodes);

      if (dump_file)
        {
          fprintf (dump_file, 
                   " Inlined into %s which now has %i insns,"
                   "net change of %+i insns.\n",
                   cgraph_node_name (edge->caller),
                   edge->caller->global.insns,
                   overall_insns - old_insns);
        }
      if (min_insns > overall_insns)
        {
          min_insns = overall_insns;
          max_insns = compute_max_insns (min_insns);

          if (dump_file)
            fprintf (dump_file, "New minimal insns reached: %i\n", min_insns);
        }
    }
  while ((edge = fibheap_extract_min (heap)) != NULL)
    {
      gcc_assert (edge->aux);
      edge->aux = NULL;
      if (!edge->callee->local.disregard_inline_limits && edge->inline_failed
          && !cgraph_recursive_inlining_p (edge->caller, edge->callee,
                                           &edge->inline_failed))
        edge->inline_failed = N_("--param inline-unit-growth limit reached");
    }
  fibheap_delete (heap);
  BITMAP_FREE (updated_nodes);
}

/* Decide on the inlining.  We do so in the topological order to avoid
   expenses on updating data structures.  */

static unsigned int
cgraph_decide_inlining (void)
{
  struct cgraph_node *node;
  int nnodes;
  struct cgraph_node **order =
    XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
  int old_insns = 0;
  int i;
  int initial_insns = 0;

  max_count = 0;
  for (node = cgraph_nodes; node; node = node->next)
    if (node->analyzed && (node->needed || node->reachable))
      {
        struct cgraph_edge *e;

        initial_insns += node->local.self_insns;
        gcc_assert (node->local.self_insns == node->global.insns);
        for (e = node->callees; e; e = e->next_callee)
          if (max_count < e->count)
            max_count = e->count;
      }
  overall_insns = initial_insns;
  gcc_assert (!max_count || (profile_info && flag_branch_probabilities));

  nnodes = cgraph_postorder (order);

  if (dump_file)
    fprintf (dump_file,
             "\nDeciding on inlining.  Starting with %i insns.\n",
             initial_insns);

  for (node = cgraph_nodes; node; node = node->next)
    node->aux = 0;

  if (dump_file)
    fprintf (dump_file, "\nInlining always_inline functions:\n");

  /* In the first pass mark all always_inline edges.  Do this with a priority
     so none of our later choices will make this impossible.  */
  for (i = nnodes - 1; i >= 0; i--)
    {
      struct cgraph_edge *e, *next;

      node = order[i];

      /* Handle nodes to be flattened, but don't update overall unit size.  */
      if (lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) != NULL)
        {
          if (dump_file)
            fprintf (dump_file,
                     "Flattening %s\n", cgraph_node_name (node));
          cgraph_decide_inlining_incrementally (node, INLINE_ALL, 0);
        }

      if (!node->local.disregard_inline_limits)
        continue;
      if (dump_file)
        fprintf (dump_file,
                 "\nConsidering %s %i insns (always inline)\n",
                 cgraph_node_name (node), node->global.insns);
      old_insns = overall_insns;
      for (e = node->callers; e; e = next)
        {
          next = e->next_caller;
          if (!e->inline_failed)
            continue;
          if (cgraph_recursive_inlining_p (e->caller, e->callee,
                                           &e->inline_failed))
            continue;
          cgraph_mark_inline_edge (e, true);
          if (dump_file)
            fprintf (dump_file, 
                     " Inlined into %s which now has %i insns.\n",
                     cgraph_node_name (e->caller),
                     e->caller->global.insns);
        }
      /* Inlining self recursive function might introduce new calls to
         themselves we didn't see in the loop above.  Fill in the proper
         reason why inline failed.  */
      for (e = node->callers; e; e = e->next_caller)
        if (e->inline_failed)
          e->inline_failed = N_("recursive inlining");
      if (dump_file)
        fprintf (dump_file, 
                 " Inlined for a net change of %+i insns.\n",
                 overall_insns - old_insns);
    }

  if (!flag_really_no_inline)
    cgraph_decide_inlining_of_small_functions ();

  if (!flag_really_no_inline
      && flag_inline_functions_called_once)
    {
      if (dump_file)
        fprintf (dump_file, "\nDeciding on functions called once:\n");

      /* And finally decide what functions are called once.  */

      for (i = nnodes - 1; i >= 0; i--)
        {
          node = order[i];

          if (node->callers && !node->callers->next_caller && !node->needed
              && node->local.inlinable && node->callers->inline_failed
              && !DECL_EXTERNAL (node->decl) && !DECL_COMDAT (node->decl))
            {
              if (dump_file)
                {
                  fprintf (dump_file,
                           "\nConsidering %s %i insns.\n",
                           cgraph_node_name (node), node->global.insns);
                  fprintf (dump_file,
                           " Called once from %s %i insns.\n",
                           cgraph_node_name (node->callers->caller),
                           node->callers->caller->global.insns);
                }

              old_insns = overall_insns;

              if (cgraph_check_inline_limits (node->callers->caller, node,
                                              NULL, false))
                {
                  cgraph_mark_inline (node->callers);
                  if (dump_file)
                    fprintf (dump_file,
                             " Inlined into %s which now has %i insns"
                             " for a net change of %+i insns.\n",
                             cgraph_node_name (node->callers->caller),
                             node->callers->caller->global.insns,
                             overall_insns - old_insns);
                }
              else
                {
                  if (dump_file)
                    fprintf (dump_file,
                             " Inline limit reached, not inlined.\n");
                }
            }
        }
    }

  if (dump_file)
    fprintf (dump_file,
             "\nInlined %i calls, eliminated %i functions, "
             "%i insns turned to %i insns.\n\n",
             ncalls_inlined, nfunctions_inlined, initial_insns,
             overall_insns);
  free (order);
  return 0;
}

/* Try to inline edge E from incremental inliner.  MODE specifies mode
   of inliner.

   We are detecting cycles by storing mode of inliner into cgraph_node last
   time we visited it in the recursion.  In general when mode is set, we have
   recursive inlining, but as an special case, we want to try harder inline
   ALWAYS_INLINE functions: consider callgraph a->b->c->b, with a being
   flatten, b being always inline.  Flattening 'a' will collapse
   a->b->c before hitting cycle.  To accommodate always inline, we however
   need to inline a->b->c->b.

   So after hitting cycle first time, we switch into ALWAYS_INLINE mode and
   stop inlining only after hitting ALWAYS_INLINE in ALWAY_INLINE mode.  */
static bool
try_inline (struct cgraph_edge *e, enum inlining_mode mode, int depth)
{
  struct cgraph_node *callee = e->callee;
  enum inlining_mode callee_mode = (size_t) callee->aux;
  bool always_inline = e->callee->local.disregard_inline_limits;

  /* We've hit cycle?  */
  if (callee_mode)
    {
      /* It is first time we see it and we are not in ALWAY_INLINE only
         mode yet.  and the function in question is always_inline.  */
      if (always_inline && mode != INLINE_ALWAYS_INLINE)
        mode = INLINE_ALWAYS_INLINE;
      /* Otherwise it is time to give up.  */
      else
        {
          if (dump_file)
            {
              indent_to (dump_file, depth);
              fprintf (dump_file,
                       "Not inlining %s into %s to avoid cycle.\n",
                       cgraph_node_name (callee),
                       cgraph_node_name (e->caller));
            }
          e->inline_failed = (e->callee->local.disregard_inline_limits
                              ? N_("recursive inlining") : "");
          return false;
        }
    }
      
  callee->aux = (void *)(size_t) mode;
  if (dump_file)
    {
      indent_to (dump_file, depth);
      fprintf (dump_file, " Inlining %s into %s.\n",
               cgraph_node_name (e->callee),
               cgraph_node_name (e->caller));
    }
  cgraph_mark_inline (e);

  /* In order to fully inline always_inline functions at -O0, we need to
     recurse here, since the inlined functions might not be processed by
     incremental inlining at all yet.  

     Also flattening needs to be done recursively.  */

  if (!flag_unit_at_a_time || mode == INLINE_ALL || always_inline)
    cgraph_decide_inlining_incrementally (e->callee, mode, depth + 1);
  callee->aux = (void *)(size_t) callee_mode;
  return true;
}

/* Decide on the inlining.  We do so in the topological order to avoid
   expenses on updating data structures.  
   DEPTH is depth of recursion, used only for debug output.  */

static bool
cgraph_decide_inlining_incrementally (struct cgraph_node *node, enum inlining_mode mode,
                                      int depth)
{
  struct cgraph_edge *e;
  bool inlined = false;
  const char *failed_reason;
  enum inlining_mode old_mode;

#ifdef ENABLE_CHECKING
  verify_cgraph_node (node);
#endif

  old_mode = (size_t)node->aux;

  if (mode != INLINE_ALWAYS_INLINE
      && lookup_attribute ("flatten", DECL_ATTRIBUTES (node->decl)) != NULL)
    {
      if (dump_file)
        fprintf (dump_file, " Flattening %s\n", cgraph_node_name (node));
      mode = INLINE_ALL;
    }

  node->aux = (void *)(size_t) mode;

  /* First of all look for always inline functions.  */
  for (e = node->callees; e; e = e->next_callee)
    {
      if (dump_file && e->callee->local.inlinable
          && (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
              != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl))))
        {
          fprintf (dump_file, "  Ignoring %s: SSA form not computed yet.\n",
                   cgraph_node_name (e->callee));
        }
      if ((e->callee->local.disregard_inline_limits
           || (mode == INLINE_ALL && e->callee->local.inlinable))
          && e->inline_failed
          && (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
              == gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
          && !cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed)
          /* ??? It is possible that renaming variable removed the function body
             in duplicate_decls. See gcc.c-torture/compile/20011119-2.c  */
          && (DECL_SAVED_TREE (e->callee->decl) || e->callee->inline_decl))
        {
          inlined |= try_inline (e, mode, depth);
        }
    }

  /* Now do the automatic inlining.  */
  if (!flag_really_no_inline && mode != INLINE_ALL
      && mode != INLINE_ALWAYS_INLINE)
    for (e = node->callees; e; e = e->next_callee)
      if (e->callee->local.inlinable
          && e->inline_failed
          && !e->callee->local.disregard_inline_limits
          && !cgraph_recursive_inlining_p (node, e->callee, &e->inline_failed)
          && (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->decl))
              == gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->callee->decl)))
          && (mode != INLINE_SIZE
              || (cgraph_estimate_size_after_inlining (1, e->caller, e->callee)
                  <= e->caller->global.insns))
          && cgraph_check_inline_limits (node, e->callee, &e->inline_failed,
                                         false)
          && (DECL_SAVED_TREE (e->callee->decl) || e->callee->inline_decl))
        {
          if (cgraph_default_inline_p (e->callee, &failed_reason))
            inlined |= try_inline (e, mode, depth);
          else if (!flag_unit_at_a_time)
            e->inline_failed = failed_reason;
        }
  node->aux = (void *)(size_t) old_mode;
  return inlined;
}

/* When inlining shall be performed.  */
static bool
cgraph_gate_inlining (void)
{
  return flag_inline_trees;
}

struct tree_opt_pass pass_ipa_inline = 
{
  "inline",                             /* name */
  cgraph_gate_inlining,                 /* gate */
  cgraph_decide_inlining,               /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_INLINE_HEURISTICS,                 /* tv_id */
  0,                                    /* properties_required */
  PROP_cfg,                             /* properties_provided */
  0,                                    /* properties_destroyed */
  TODO_remove_functions,                /* todo_flags_finish */
  TODO_dump_cgraph | TODO_dump_func
  | TODO_remove_functions,              /* todo_flags_finish */
  0                                     /* letter */
};

/* Because inlining might remove no-longer reachable nodes, we need to
   keep the array visible to garbage collector to avoid reading collected
   out nodes.  */
static int nnodes;
static GTY ((length ("nnodes"))) struct cgraph_node **order;

/* Do inlining of small functions.  Doing so early helps profiling and other
   passes to be somewhat more effective and avoids some code duplication in
   later real inlining pass for testcases with very many function calls.  */
static unsigned int
cgraph_early_inlining (void)
{
  struct cgraph_node *node = cgraph_node (current_function_decl);
  unsigned int todo = 0;

  if (sorrycount || errorcount)
    return 0;
  if (cgraph_decide_inlining_incrementally (node,
                                            flag_unit_at_a_time
                                            ? INLINE_SIZE : INLINE_SPEED, 0))
    {
      timevar_push (TV_INTEGRATION);
      todo = optimize_inline_calls (current_function_decl);
      timevar_pop (TV_INTEGRATION);
    }
  return todo;
}

/* When inlining shall be performed.  */
static bool
cgraph_gate_early_inlining (void)
{
  return flag_inline_trees && flag_early_inlining;
}

struct tree_opt_pass pass_early_inline = 
{
  "einline",                            /* name */
  cgraph_gate_early_inlining,           /* gate */
  cgraph_early_inlining,                /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_INLINE_HEURISTICS,                 /* tv_id */
  0,                                    /* properties_required */
  PROP_cfg,                             /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  TODO_dump_func,                       /* todo_flags_finish */
  0                                     /* letter */
};

/* When inlining shall be performed.  */
static bool
cgraph_gate_ipa_early_inlining (void)
{
  return (flag_inline_trees && flag_early_inlining
          && (flag_branch_probabilities || flag_test_coverage
              || profile_arc_flag));
}

/* IPA pass wrapper for early inlining pass.  We need to run early inlining
   before tree profiling so we have stand alone IPA pass for doing so.  */
struct tree_opt_pass pass_ipa_early_inline = 
{
  "einline_ipa",                        /* name */
  cgraph_gate_ipa_early_inlining,       /* gate */
  NULL,                                 /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_INLINE_HEURISTICS,                 /* tv_id */
  0,                                    /* properties_required */
  PROP_cfg,                             /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  TODO_dump_cgraph,                     /* todo_flags_finish */
  0                                     /* letter */
};

/* Compute parameters of functions used by inliner.  */
static unsigned int
compute_inline_parameters (void)
{
  struct cgraph_node *node = cgraph_node (current_function_decl);

  gcc_assert (!node->global.inlined_to);
  node->local.estimated_self_stack_size = estimated_stack_frame_size ();
  node->global.estimated_stack_size = node->local.estimated_self_stack_size;
  node->global.stack_frame_offset = 0;
  node->local.inlinable = tree_inlinable_function_p (current_function_decl);
  node->local.self_insns = estimate_num_insns (current_function_decl,
                                               &eni_inlining_weights);
  if (node->local.inlinable)
    node->local.disregard_inline_limits
      = lang_hooks.tree_inlining.disregard_inline_limits (current_function_decl);
  if (flag_really_no_inline && !node->local.disregard_inline_limits)
    node->local.inlinable = 0;
  /* Inlining characteristics are maintained by the cgraph_mark_inline.  */
  node->global.insns = node->local.self_insns;
  return 0;
}

/* When inlining shall be performed.  */
static bool
gate_inline_passes (void)
{
  return flag_inline_trees;
}

struct tree_opt_pass pass_inline_parameters = 
{
  NULL,                                 /* name */
  gate_inline_passes,                   /* gate */
  compute_inline_parameters,            /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_INLINE_HEURISTICS,                 /* tv_id */
  0,                                    /* properties_required */
  PROP_cfg,                             /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  0,                                    /* todo_flags_finish */
  0                                     /* letter */
};

/* Apply inline plan to the function.  */
static unsigned int
apply_inline (void)
{
  unsigned int todo = 0;
  struct cgraph_edge *e;
  struct cgraph_node *node = cgraph_node (current_function_decl);

  /* Even when not optimizing, ensure that always_inline functions get inlined.
   */
  if (!optimize)
   cgraph_decide_inlining_incrementally (node, INLINE_SPEED, 0);

  /* We might need the body of this function so that we can expand
     it inline somewhere else.  */
  if (cgraph_preserve_function_body_p (current_function_decl))
    save_inline_function_body (node);

  for (e = node->callees; e; e = e->next_callee)
    if (!e->inline_failed || warn_inline)
      break;
  if (e)
    {
      timevar_push (TV_INTEGRATION);
      todo = optimize_inline_calls (current_function_decl);
      timevar_pop (TV_INTEGRATION);
    }
  /* In non-unit-at-a-time we must mark all referenced functions as needed.  */
  if (!flag_unit_at_a_time)
    {
      struct cgraph_edge *e;
      for (e = node->callees; e; e = e->next_callee)
        if (e->callee->analyzed)
          cgraph_mark_needed_node (e->callee);
    }
  return todo | execute_fixup_cfg ();
}

struct tree_opt_pass pass_apply_inline = 
{
  "apply_inline",                       /* name */
  NULL,                                 /* gate */
  apply_inline,                         /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_INLINE_HEURISTICS,                 /* tv_id */
  0,                                    /* properties_required */
  PROP_cfg,                             /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  TODO_dump_func | TODO_verify_flow
  | TODO_verify_stmts,                  /* todo_flags_finish */
  0                                     /* letter */
};

#include "gt-ipa-inline.h"