2013-04-26 François Dumont <fdumont@gcc.gnu.org>

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

/* Inlining decision heuristics.
   Copyright (C) 2003-2013 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 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/>.  */

/*  Inlining decision heuristics

    The implementation of inliner is organized as follows:

    inlining heuristics limits

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

      Functions are inlined when it is obvious the result is profitable (such
      as functions called once or when inlining reduce code size).
      In addition to that we perform inlining of small functions and recursive
      inlining.

    inlining heuristics

       The inliner itself is split into two passes:

       pass_early_inlining

         Simple local inlining pass inlining callees into current function.
         This pass makes no use of whole unit analysis and thus it can do only
         very simple decisions based on local properties.

         The strength of the pass is that it is run in topological order
         (reverse postorder) on the callgraph. Functions are converted into SSA
         form just before this pass and optimized subsequently. As a result, the
         callees of the function seen by the early inliner was already optimized
         and results of early inlining adds a lot of optimization opportunities
         for the local optimization.

         The pass handle the obvious inlining decisions within the compilation
         unit - inlining auto inline functions, inlining for size and
         flattening.

         main strength of the pass is the ability to eliminate abstraction
         penalty in C++ code (via combination of inlining and early
         optimization) and thus improve quality of analysis done by real IPA
         optimizers.

         Because of lack of whole unit knowledge, the pass can not really make
         good code size/performance tradeoffs.  It however does very simple
         speculative inlining allowing code size to grow by
         EARLY_INLINING_INSNS when callee is leaf function.  In this case the
         optimizations performed later are very likely to eliminate the cost.

       pass_ipa_inline

         This is the real inliner able to handle inlining with whole program
         knowledge. It performs following steps:

         1) inlining of small functions.  This is implemented by greedy
         algorithm ordering all inlinable cgraph edges by their badness and
         inlining them in this order as long as inline limits allows doing so.

         This heuristics is not very good on inlining recursive calls. Recursive
         calls can be inlined with results similar to loop unrolling. To do so,
         special purpose recursive inliner is executed on function when
         recursive edge is met as viable candidate.

         2) Unreachable functions are removed from callgraph.  Inlining leads
         to devirtualization and other modification of callgraph so functions
         may become unreachable during the process. Also functions declared as
         extern inline or virtual functions are removed, since after inlining
         we no longer need the offline bodies.

         3) Functions called once and not exported from the unit are inlined.
         This should almost always lead to reduction of code size by eliminating
         the need for offline copy of the function.  */

#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 "gimple-pretty-print.h"
#include "params.h"
#include "fibheap.h"
#include "intl.h"
#include "tree-pass.h"
#include "coverage.h"
#include "ggc.h"
#include "rtl.h"
#include "tree-flow.h"
#include "ipa-prop.h"
#include "except.h"
#include "target.h"
#include "ipa-inline.h"
#include "ipa-utils.h"

/* Statistics we collect about inlining algorithm.  */
static int overall_size;
static gcov_type max_count;

/* Return false when inlining edge E would lead to violating
   limits on function unit growth or stack usage growth.  

   The relative function body growth limit is present generally
   to avoid problems with non-linear behavior of the compiler.
   To allow inlining huge functions into tiny wrapper, the limit
   is always based on the bigger of the two functions considered.

   For stack growth limits we always base the growth in stack usage
   of the callers.  We want to prevent applications from segfaulting
   on stack overflow when functions with huge stack frames gets
   inlined. */

static bool
caller_growth_limits (struct cgraph_edge *e)
{
  struct cgraph_node *to = e->caller;
  struct cgraph_node *what = cgraph_function_or_thunk_node (e->callee, NULL);
  int newsize;
  int limit = 0;
  HOST_WIDE_INT stack_size_limit = 0, inlined_stack;
  struct inline_summary *info, *what_info, *outer_info = inline_summary (to);

  /* Look for function e->caller is inlined to.  While doing
     so work out the largest function body on the way.  As
     described above, we want to base our function growth
     limits based on that.  Not on the self size of the
     outer function, not on the self size of inline code
     we immediately inline to.  This is the most relaxed
     interpretation of the rule "do not grow large functions
     too much in order to prevent compiler from exploding".  */
  while (true)
    {
      info = inline_summary (to);
      if (limit < info->self_size)
        limit = info->self_size;
      if (stack_size_limit < info->estimated_self_stack_size)
        stack_size_limit = info->estimated_self_stack_size;
      if (to->global.inlined_to)
        to = to->callers->caller;
      else
        break;
    }

  what_info = inline_summary (what);

  if (limit < what_info->self_size)
    limit = what_info->self_size;

  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 = estimate_size_after_inlining (to, e);
  if (newsize >= info->size
      && newsize > PARAM_VALUE (PARAM_LARGE_FUNCTION_INSNS)
      && newsize > limit)
    {
      e->inline_failed = CIF_LARGE_FUNCTION_GROWTH_LIMIT;
      return false;
    }

  if (!what_info->estimated_stack_size)
    return true;

  /* FIXME: Stack size limit often prevents inlining in Fortran programs
     due to large i/o datastructures used by the Fortran front-end.
     We ought to ignore this limit when we know that the edge is executed
     on every invocation of the caller (i.e. its call statement dominates
     exit block).  We do not track this information, yet.  */
  stack_size_limit += ((gcov_type)stack_size_limit
                       * PARAM_VALUE (PARAM_STACK_FRAME_GROWTH) / 100);

  inlined_stack = (outer_info->stack_frame_offset
                   + outer_info->estimated_self_stack_size
                   + what_info->estimated_stack_size);
  /* Check new stack consumption with stack consumption at the place
     stack is used.  */
  if (inlined_stack > stack_size_limit
      /* If function already has large stack usage from sibling
         inline call, we can inline, too.
         This bit overoptimistically assume that we are good at stack
         packing.  */
      && inlined_stack > info->estimated_stack_size
      && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
    {
      e->inline_failed = CIF_LARGE_STACK_FRAME_GROWTH_LIMIT;
      return false;
    }
  return true;
}

/* Dump info about why inlining has failed.  */

static void
report_inline_failed_reason (struct cgraph_edge *e)
{
  if (dump_file)
    {
      fprintf (dump_file, "  not inlinable: %s/%i -> %s/%i, %s\n",
               xstrdup (cgraph_node_name (e->caller)), e->caller->uid,
               xstrdup (cgraph_node_name (e->callee)), e->callee->uid,
               cgraph_inline_failed_string (e->inline_failed));
    }
}

/* Decide if we can inline the edge and possibly update
   inline_failed reason.  
   We check whether inlining is possible at all and whether
   caller growth limits allow doing so.  

   if REPORT is true, output reason to the dump file.  */

static bool
can_inline_edge_p (struct cgraph_edge *e, bool report)
{
  bool inlinable = true;
  enum availability avail;
  struct cgraph_node *callee
    = cgraph_function_or_thunk_node (e->callee, &avail);
  tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (e->caller->symbol.decl);
  tree callee_tree
    = callee ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee->symbol.decl) : NULL;
  struct function *caller_cfun = DECL_STRUCT_FUNCTION (e->caller->symbol.decl);
  struct function *callee_cfun
    = callee ? DECL_STRUCT_FUNCTION (callee->symbol.decl) : NULL;

  if (!caller_cfun && e->caller->clone_of)
    caller_cfun = DECL_STRUCT_FUNCTION (e->caller->clone_of->symbol.decl);

  if (!callee_cfun && callee && callee->clone_of)
    callee_cfun = DECL_STRUCT_FUNCTION (callee->clone_of->symbol.decl);

  gcc_assert (e->inline_failed);

  if (!callee || !callee->analyzed)
    {
      e->inline_failed = CIF_BODY_NOT_AVAILABLE;
      inlinable = false;
    }
  else if (!inline_summary (callee)->inlinable)
    {
      e->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
      inlinable = false;
    }
  else if (avail <= AVAIL_OVERWRITABLE)
    {
      e->inline_failed = CIF_OVERWRITABLE;
      inlinable = false;
    }
  else if (e->call_stmt_cannot_inline_p)
    {
      e->inline_failed = CIF_MISMATCHED_ARGUMENTS;
      inlinable = false;
    }
  /* Don't inline if the functions have different EH personalities.  */
  else if (DECL_FUNCTION_PERSONALITY (e->caller->symbol.decl)
           && DECL_FUNCTION_PERSONALITY (callee->symbol.decl)
           && (DECL_FUNCTION_PERSONALITY (e->caller->symbol.decl)
               != DECL_FUNCTION_PERSONALITY (callee->symbol.decl)))
    {
      e->inline_failed = CIF_EH_PERSONALITY;
      inlinable = false;
    }
  /* TM pure functions should not be inlined into non-TM_pure
     functions.  */
  else if (is_tm_pure (callee->symbol.decl)
           && !is_tm_pure (e->caller->symbol.decl))
    {
      e->inline_failed = CIF_UNSPECIFIED;
      inlinable = false;
    }
  /* Don't inline if the callee can throw non-call exceptions but the
     caller cannot.
     FIXME: this is obviously wrong for LTO where STRUCT_FUNCTION is missing.
     Move the flag into cgraph node or mirror it in the inline summary.  */
  else if (callee_cfun && callee_cfun->can_throw_non_call_exceptions
           && !(caller_cfun && caller_cfun->can_throw_non_call_exceptions))
    {
      e->inline_failed = CIF_NON_CALL_EXCEPTIONS;
      inlinable = false;
    }
  /* Check compatibility of target optimization options.  */
  else if (!targetm.target_option.can_inline_p (e->caller->symbol.decl,
                                                callee->symbol.decl))
    {
      e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
      inlinable = false;
    }
  /* Check if caller growth allows the inlining.  */
  else if (!DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl)
           && !lookup_attribute ("flatten",
                                 DECL_ATTRIBUTES
                                   (e->caller->global.inlined_to
                                    ? e->caller->global.inlined_to->symbol.decl
                                    : e->caller->symbol.decl))
           && !caller_growth_limits (e))
    inlinable = false;
  /* Don't inline a function with a higher optimization level than the
     caller.  FIXME: this is really just tip of iceberg of handling
     optimization attribute.  */
  else if (caller_tree != callee_tree)
    {
      struct cl_optimization *caller_opt
        = TREE_OPTIMIZATION ((caller_tree)
                             ? caller_tree
                             : optimization_default_node);

      struct cl_optimization *callee_opt
        = TREE_OPTIMIZATION ((callee_tree)
                             ? callee_tree
                             : optimization_default_node);

      if (((caller_opt->x_optimize > callee_opt->x_optimize)
           || (caller_opt->x_optimize_size != callee_opt->x_optimize_size))
          /* gcc.dg/pr43564.c.  Look at forced inline even in -O0.  */
          && !DECL_DISREGARD_INLINE_LIMITS (e->callee->symbol.decl))
        {
          e->inline_failed = CIF_OPTIMIZATION_MISMATCH;
          inlinable = false;
        }
    }

  if (!inlinable && report)
    report_inline_failed_reason (e);
  return inlinable;
}


/* Return true if the edge E is inlinable during early inlining.  */

static bool
can_early_inline_edge_p (struct cgraph_edge *e)
{
  struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee,
                                                              NULL);
  /* Early inliner might get called at WPA stage when IPA pass adds new
     function.  In this case we can not really do any of early inlining
     because function bodies are missing.  */
  if (!gimple_has_body_p (callee->symbol.decl))
    {
      e->inline_failed = CIF_BODY_NOT_AVAILABLE;
      return false;
    }
  /* In early inliner some of callees may not be in SSA form yet
     (i.e. the callgraph is cyclic and we did not process
     the callee by early inliner, yet).  We don't have CIF code for this
     case; later we will re-do the decision in the real inliner.  */
  if (!gimple_in_ssa_p (DECL_STRUCT_FUNCTION (e->caller->symbol.decl))
      || !gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->symbol.decl)))
    {
      if (dump_file)
        fprintf (dump_file, "  edge not inlinable: not in SSA form\n");
      return false;
    }
  if (!can_inline_edge_p (e, true))
    return false;
  return true;
}


/* Return number of calls in N.  Ignore cheap builtins.  */

static int
num_calls (struct cgraph_node *n)
{
  struct cgraph_edge *e;
  int num = 0;

  for (e = n->callees; e; e = e->next_callee)
    if (!is_inexpensive_builtin (e->callee->symbol.decl))
      num++;
  return num;
}


/* Return true if we are interested in inlining small function.  */

static bool
want_early_inline_function_p (struct cgraph_edge *e)
{
  bool want_inline = true;
  struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);

  if (DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl))
    ;
  else if (!DECL_DECLARED_INLINE_P (callee->symbol.decl)
           && !flag_inline_small_functions)
    {
      e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
      report_inline_failed_reason (e);
      want_inline = false;
    }
  else
    {
      int growth = estimate_edge_growth (e);
      int n;

      if (growth <= 0)
        ;
      else if (!cgraph_maybe_hot_edge_p (e)
               && growth > 0)
        {
          if (dump_file)
            fprintf (dump_file, "  will not early inline: %s/%i->%s/%i, "
                     "call is cold and code would grow by %i\n",
                     xstrdup (cgraph_node_name (e->caller)), e->caller->uid,
                     xstrdup (cgraph_node_name (callee)), callee->uid,
                     growth);
          want_inline = false;
        }
      else if (growth > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS))
        {
          if (dump_file)
            fprintf (dump_file, "  will not early inline: %s/%i->%s/%i, "
                     "growth %i exceeds --param early-inlining-insns\n",
                     xstrdup (cgraph_node_name (e->caller)), e->caller->uid,
                     xstrdup (cgraph_node_name (callee)), callee->uid,
                     growth);
          want_inline = false;
        }
      else if ((n = num_calls (callee)) != 0
               && growth * (n + 1) > PARAM_VALUE (PARAM_EARLY_INLINING_INSNS))
        {
          if (dump_file)
            fprintf (dump_file, "  will not early inline: %s/%i->%s/%i, "
                     "growth %i exceeds --param early-inlining-insns "
                     "divided by number of calls\n",
                     xstrdup (cgraph_node_name (e->caller)), e->caller->uid,
                     xstrdup (cgraph_node_name (callee)), callee->uid,
                     growth);
          want_inline = false;
        }
    }
  return want_inline;
}

/* Compute time of the edge->caller + edge->callee execution when inlining
   does not happen.  */

inline gcov_type
compute_uninlined_call_time (struct inline_summary *callee_info,
                             struct cgraph_edge *edge)
{
  gcov_type uninlined_call_time =
    RDIV ((gcov_type)callee_info->time * MAX (edge->frequency, 1),
          CGRAPH_FREQ_BASE);
  gcov_type caller_time = inline_summary (edge->caller->global.inlined_to
                                          ? edge->caller->global.inlined_to
                                          : edge->caller)->time;
  return uninlined_call_time + caller_time;
}

/* Same as compute_uinlined_call_time but compute time when inlining
   does happen.  */

inline gcov_type
compute_inlined_call_time (struct cgraph_edge *edge,
                           int edge_time)
{
  gcov_type caller_time = inline_summary (edge->caller->global.inlined_to
                                          ? edge->caller->global.inlined_to
                                          : edge->caller)->time;
  gcov_type time = (caller_time
                    + RDIV (((gcov_type) edge_time
                             - inline_edge_summary (edge)->call_stmt_time)
                    * MAX (edge->frequency, 1), CGRAPH_FREQ_BASE));
  /* Possible one roundoff error, but watch for overflows.  */
  gcc_checking_assert (time >= INT_MIN / 2);
  if (time < 0)
    time = 0;
  return time;
}

/* Return true if the speedup for inlining E is bigger than
   PARAM_MAX_INLINE_MIN_SPEEDUP.  */

static bool
big_speedup_p (struct cgraph_edge *e)
{
  gcov_type time = compute_uninlined_call_time (inline_summary (e->callee),
                                                e);
  gcov_type inlined_time = compute_inlined_call_time (e,
                                                      estimate_edge_time (e));
  if (time - inlined_time
      > RDIV (time * PARAM_VALUE (PARAM_INLINE_MIN_SPEEDUP), 100))
    return true;
  return false;
}

/* Return true if we are interested in inlining small function.
   When REPORT is true, report reason to dump file.  */

static bool
want_inline_small_function_p (struct cgraph_edge *e, bool report)
{
  bool want_inline = true;
  struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);

  if (DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl))
    ;
  else if (!DECL_DECLARED_INLINE_P (callee->symbol.decl)
           && !flag_inline_small_functions)
    {
      e->inline_failed = CIF_FUNCTION_NOT_INLINE_CANDIDATE;
      want_inline = false;
    }
  else
    {
      int growth = estimate_edge_growth (e);
      inline_hints hints = estimate_edge_hints (e);
      bool big_speedup = big_speedup_p (e);

      if (growth <= 0)
        ;
      /* Apply MAX_INLINE_INSNS_SINGLE limit.  Do not do so when
         hints suggests that inlining given function is very profitable.  */
      else if (DECL_DECLARED_INLINE_P (callee->symbol.decl)
               && growth >= MAX_INLINE_INSNS_SINGLE
               && !big_speedup
               && !(hints & (INLINE_HINT_indirect_call
                             | INLINE_HINT_loop_iterations
                             | INLINE_HINT_array_index
                             | INLINE_HINT_loop_stride)))
        {
          e->inline_failed = CIF_MAX_INLINE_INSNS_SINGLE_LIMIT;
          want_inline = false;
        }
      /* Before giving up based on fact that caller size will grow, allow
         functions that are called few times and eliminating the offline
         copy will lead to overall code size reduction.
         Not all of these will be handled by subsequent inlining of functions
         called once: in particular weak functions are not handled or funcitons
         that inline to multiple calls but a lot of bodies is optimized out.
         Finally we want to inline earlier to allow inlining of callbacks.

         This is slightly wrong on aggressive side:  it is entirely possible
         that function is called many times with a context where inlining
         reduces code size and few times with a context where inlining increase
         code size.  Resoluting growth estimate will be negative even if it
         would make more sense to keep offline copy and do not inline into the
         call sites that makes the code size grow.  

         When badness orders the calls in a way that code reducing calls come
         first, this situation is not a problem at all: after inlining all
         "good" calls, we will realize that keeping the function around is
         better.  */
      else if (growth <= MAX_INLINE_INSNS_SINGLE
               /* Unlike for functions called once, we play unsafe with
                  COMDATs.  We can allow that since we know functions
                  in consideration are small (and thus risk is small) and
                  moreover grow estimates already accounts that COMDAT
                  functions may or may not disappear when eliminated from
                  current unit. With good probability making aggressive
                  choice in all units is going to make overall program
                  smaller.

                  Consequently we ask cgraph_can_remove_if_no_direct_calls_p
                  instead of
                  cgraph_will_be_removed_from_program_if_no_direct_calls  */
                && !DECL_EXTERNAL (callee->symbol.decl)
                && cgraph_can_remove_if_no_direct_calls_p (callee)
                && estimate_growth (callee) <= 0)
        ;
      else if (!DECL_DECLARED_INLINE_P (callee->symbol.decl)
               && !flag_inline_functions)
        {
          e->inline_failed = CIF_NOT_DECLARED_INLINED;
          want_inline = false;
        }
      /* Apply MAX_INLINE_INSNS_AUTO limit for functions not declared inline
         Upgrade it to MAX_INLINE_INSNS_SINGLE when hints suggests that
         inlining given function is very profitable.  */
      else if (!DECL_DECLARED_INLINE_P (callee->symbol.decl)
               && !big_speedup
               && growth >= ((hints & (INLINE_HINT_indirect_call
                                       | INLINE_HINT_loop_iterations
                                       | INLINE_HINT_array_index
                                       | INLINE_HINT_loop_stride))
                             ? MAX (MAX_INLINE_INSNS_AUTO,
                                    MAX_INLINE_INSNS_SINGLE)
                             : MAX_INLINE_INSNS_AUTO))
        {
          e->inline_failed = CIF_MAX_INLINE_INSNS_AUTO_LIMIT;
          want_inline = false;
        }
      /* If call is cold, do not inline when function body would grow. */
      else if (!cgraph_maybe_hot_edge_p (e))
        {
          e->inline_failed = CIF_UNLIKELY_CALL;
          want_inline = false;
        }
    }
  if (!want_inline && report)
    report_inline_failed_reason (e);
  return want_inline;
}

/* EDGE is self recursive edge.
   We hand two cases - when function A is inlining into itself
   or when function A is being inlined into another inliner copy of function
   A within function B.  

   In first case OUTER_NODE points to the toplevel copy of A, while
   in the second case OUTER_NODE points to the outermost copy of A in B.

   In both cases we want to be extra selective since
   inlining the call will just introduce new recursive calls to appear.  */

static bool
want_inline_self_recursive_call_p (struct cgraph_edge *edge,
                                   struct cgraph_node *outer_node,
                                   bool peeling,
                                   int depth)
{
  char const *reason = NULL;
  bool want_inline = true;
  int caller_freq = CGRAPH_FREQ_BASE;
  int max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH_AUTO);

  if (DECL_DECLARED_INLINE_P (edge->caller->symbol.decl))
    max_depth = PARAM_VALUE (PARAM_MAX_INLINE_RECURSIVE_DEPTH);

  if (!cgraph_maybe_hot_edge_p (edge))
    {
      reason = "recursive call is cold";
      want_inline = false;
    }
  else if (max_count && !outer_node->count)
    {
      reason = "not executed in profile";
      want_inline = false;
    }
  else if (depth > max_depth)
    {
      reason = "--param max-inline-recursive-depth exceeded.";
      want_inline = false;
    }

  if (outer_node->global.inlined_to)
    caller_freq = outer_node->callers->frequency;

  if (!want_inline)
    ;
  /* Inlining of self recursive function into copy of itself within other function
     is transformation similar to loop peeling.

     Peeling is profitable if we can inline enough copies to make probability
     of actual call to the self recursive function very small.  Be sure that
     the probability of recursion is small.

     We ensure that the frequency of recursing is at most 1 - (1/max_depth).
     This way the expected number of recision is at most max_depth.  */
  else if (peeling)
    {
      int max_prob = CGRAPH_FREQ_BASE - ((CGRAPH_FREQ_BASE + max_depth - 1)
                                         / max_depth);
      int i;
      for (i = 1; i < depth; i++)
        max_prob = max_prob * max_prob / CGRAPH_FREQ_BASE;
      if (max_count
          && (edge->count * CGRAPH_FREQ_BASE / outer_node->count
              >= max_prob))
        {
          reason = "profile of recursive call is too large";
          want_inline = false;
        }
      if (!max_count
          && (edge->frequency * CGRAPH_FREQ_BASE / caller_freq
              >= max_prob))
        {
          reason = "frequency of recursive call is too large";
          want_inline = false;
        }
    }
  /* Recursive inlining, i.e. equivalent of unrolling, is profitable if recursion
     depth is large.  We reduce function call overhead and increase chances that
     things fit in hardware return predictor.

     Recursive inlining might however increase cost of stack frame setup
     actually slowing down functions whose recursion tree is wide rather than
     deep.

     Deciding reliably on when to do recursive inlining without profile feedback
     is tricky.  For now we disable recursive inlining when probability of self
     recursion is low. 

     Recursive inlining of self recursive call within loop also results in large loop
     depths that generally optimize badly.  We may want to throttle down inlining
     in those cases.  In particular this seems to happen in one of libstdc++ rb tree
     methods.  */
  else
    {
      if (max_count
          && (edge->count * 100 / outer_node->count
              <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY)))
        {
          reason = "profile of recursive call is too small";
          want_inline = false;
        }
      else if (!max_count
               && (edge->frequency * 100 / caller_freq
                   <= PARAM_VALUE (PARAM_MIN_INLINE_RECURSIVE_PROBABILITY)))
        {
          reason = "frequency of recursive call is too small";
          want_inline = false;
        }
    }
  if (!want_inline && dump_file)
    fprintf (dump_file, "   not inlining recursively: %s\n", reason);
  return want_inline;
}

/* Return true when NODE has caller other than EDGE. 
   Worker for cgraph_for_node_and_aliases.  */

static bool
check_caller_edge (struct cgraph_node *node, void *edge)
{
  return (node->callers
          && node->callers != edge);
}


/* Decide if inlining NODE would reduce unit size by eliminating
   the offline copy of function.  
   When COLD is true the cold calls are considered, too.  */

static bool
want_inline_function_to_all_callers_p (struct cgraph_node *node, bool cold)
{
   struct cgraph_node *function = cgraph_function_or_thunk_node (node, NULL);
   struct cgraph_edge *e;
   bool has_hot_call = false;

   /* Does it have callers?  */
   if (!node->callers)
     return false;
   /* Already inlined?  */
   if (function->global.inlined_to)
     return false;
   if (cgraph_function_or_thunk_node (node, NULL) != node)
     return false;
   /* Inlining into all callers would increase size?  */
   if (estimate_growth (node) > 0)
     return false;
   /* Maybe other aliases has more direct calls.  */
   if (cgraph_for_node_and_aliases (node, check_caller_edge, node->callers, true))
     return false;
   /* All inlines must be possible.  */
   for (e = node->callers; e; e = e->next_caller)
     {
       if (!can_inline_edge_p (e, true))
         return false;
       if (!has_hot_call && cgraph_maybe_hot_edge_p (e))
         has_hot_call = 1;
     }

   if (!cold && !has_hot_call)
     return false;
   return true;
}

#define RELATIVE_TIME_BENEFIT_RANGE (INT_MAX / 64)

/* Return relative time improvement for inlining EDGE in range
   1...RELATIVE_TIME_BENEFIT_RANGE  */

static inline int
relative_time_benefit (struct inline_summary *callee_info,
                       struct cgraph_edge *edge,
                       int edge_time)
{
  gcov_type relbenefit;
  gcov_type uninlined_call_time = compute_uninlined_call_time (callee_info, edge);
  gcov_type inlined_call_time = compute_inlined_call_time (edge, edge_time);

  /* Inlining into extern inline function is not a win.  */
  if (DECL_EXTERNAL (edge->caller->global.inlined_to
                     ? edge->caller->global.inlined_to->symbol.decl
                     : edge->caller->symbol.decl))
    return 1;

  /* Watch overflows.  */
  gcc_checking_assert (uninlined_call_time >= 0);
  gcc_checking_assert (inlined_call_time >= 0);
  gcc_checking_assert (uninlined_call_time >= inlined_call_time);

  /* Compute relative time benefit, i.e. how much the call becomes faster.
     ??? perhaps computing how much the caller+calle together become faster
     would lead to more realistic results.  */
  if (!uninlined_call_time)
    uninlined_call_time = 1;
  relbenefit =
    RDIV (((gcov_type)uninlined_call_time - inlined_call_time) * RELATIVE_TIME_BENEFIT_RANGE,
          uninlined_call_time);
  relbenefit = MIN (relbenefit, RELATIVE_TIME_BENEFIT_RANGE);
  gcc_checking_assert (relbenefit >= 0);
  relbenefit = MAX (relbenefit, 1);
  return relbenefit;
}


/* 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.  */

static int
edge_badness (struct cgraph_edge *edge, bool dump)
{
  gcov_type badness;
  int growth, edge_time;
  struct cgraph_node *callee = cgraph_function_or_thunk_node (edge->callee,
                                                              NULL);
  struct inline_summary *callee_info = inline_summary (callee);
  inline_hints hints;

  if (DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl))
    return INT_MIN;

  growth = estimate_edge_growth (edge);
  edge_time = estimate_edge_time (edge);
  hints = estimate_edge_hints (edge);
  gcc_checking_assert (edge_time >= 0);
  gcc_checking_assert (edge_time <= callee_info->time);
  gcc_checking_assert (growth <= callee_info->size);

  if (dump)
    {
      fprintf (dump_file, "    Badness calculation for %s/%i -> %s/%i\n",
               xstrdup (cgraph_node_name (edge->caller)),
               edge->caller->uid,
               xstrdup (cgraph_node_name (callee)),
               edge->callee->uid);
      fprintf (dump_file, "      size growth %i, time %i ",
               growth,
               edge_time);
      dump_inline_hints (dump_file, hints);
      if (big_speedup_p (edge))
        fprintf (dump_file, " big_speedup");
      fprintf (dump_file, "\n");
    }

  /* Always prefer inlining saving code size.  */
  if (growth <= 0)
    {
      badness = INT_MIN / 2 + growth;
      if (dump)
        fprintf (dump_file, "      %i: Growth %i <= 0\n", (int) badness,
                 growth);
    }

  /* When profiling is available, compute badness as:

                relative_edge_count * relative_time_benefit
     goodness = -------------------------------------------
                growth_f_caller
     badness = -goodness  

    The fraction is upside down, because on edge counts and time beneits
    the bounds are known. Edge growth is essentially unlimited.  */

  else if (max_count)
    {
      int relbenefit = relative_time_benefit (callee_info, edge, edge_time);
      badness =
        ((int)
         ((double) edge->count * INT_MIN / 2 / max_count / RELATIVE_TIME_BENEFIT_RANGE) *
         relbenefit) / growth;
      
      /* Be sure that insanity of the profile won't lead to increasing counts
         in the scalling and thus to overflow in the computation above.  */
      gcc_assert (max_count >= edge->count);
      if (dump)
        {
          fprintf (dump_file,
                   "      %i (relative %f): profile info. Relative count %f"
                   " * Relative benefit %f\n",
                   (int) badness, (double) badness / INT_MIN,
                   (double) edge->count / max_count,
                   relbenefit * 100.0 / RELATIVE_TIME_BENEFIT_RANGE);
        }
    }

  /* When function local profile is available. Compute badness as:
     
                 relative_time_benefit
     goodness =  ---------------------------------
                 growth_of_caller * overall_growth

     badness = - goodness

     compensated by the inline hints.
  */
  else if (flag_guess_branch_prob)
    {
      badness = (relative_time_benefit (callee_info, edge, edge_time)
                 * (INT_MIN / 16 / RELATIVE_TIME_BENEFIT_RANGE));
      badness /= (MIN (65536/2, growth) * MIN (65536/2, MAX (1, callee_info->growth)));
      gcc_checking_assert (badness <=0 && badness >= INT_MIN / 16);
      if ((hints & (INLINE_HINT_indirect_call
                    | INLINE_HINT_loop_iterations
                    | INLINE_HINT_array_index
                    | INLINE_HINT_loop_stride))
          || callee_info->growth <= 0)
        badness *= 8;
      if (hints & (INLINE_HINT_same_scc))
        badness /= 16;
      else if (hints & (INLINE_HINT_in_scc))
        badness /= 8;
      else if (hints & (INLINE_HINT_cross_module))
        badness /= 2;
      gcc_checking_assert (badness <= 0 && badness >= INT_MIN / 2);
      if ((hints & INLINE_HINT_declared_inline) && badness >= INT_MIN / 32)
        badness *= 16;
      if (dump)
        {
          fprintf (dump_file,
                   "      %i: guessed profile. frequency %f,"
                   " benefit %f%%, time w/o inlining %i, time w inlining %i"
                   " overall growth %i (current) %i (original)\n",
                   (int) badness, (double)edge->frequency / CGRAPH_FREQ_BASE,
                   relative_time_benefit (callee_info, edge, edge_time) * 100.0
                   / RELATIVE_TIME_BENEFIT_RANGE, 
                   (int)compute_uninlined_call_time (callee_info, edge),
                   (int)compute_inlined_call_time (edge, edge_time),
                   estimate_growth (callee),
                   callee_info->growth);
        }
    }
  /* When function local profile is not available or it does not give
     useful information (ie frequency is zero), base the cost on
     loop nest and overall size growth, so we optimize for overall number
     of functions fully inlined in program.  */
  else
    {
      int nest = MIN (inline_edge_summary (edge)->loop_depth, 8);
      badness = growth * 256;

      /* Decrease badness if call is nested.  */
      if (badness > 0)
        badness >>= nest;
      else
        {
          badness <<= nest;
        }
      if (dump)
        fprintf (dump_file, "      %i: no profile. nest %i\n", (int) badness,
                 nest);
    }

  /* Ensure that we did not overflow in all the fixed point math above.  */
  gcc_assert (badness >= INT_MIN);
  gcc_assert (badness <= INT_MAX - 1);
  /* Make recursive inlining happen always after other inlining is done.  */
  if (cgraph_edge_recursive_p (edge))
    return badness + 1;
  else
    return badness;
}

/* Recompute badness of EDGE and update its key in HEAP if needed.  */
static inline void
update_edge_key (fibheap_t heap, struct cgraph_edge *edge)
{
  int badness = edge_badness (edge, false);
  if (edge->aux)
    {
      fibnode_t n = (fibnode_t) edge->aux;
      gcc_checking_assert (n->data == edge);

      /* fibheap_replace_key only decrease the keys.
         When we increase the key we do not update heap
         and instead re-insert the element once it becomes
         a minimum of heap.  */
      if (badness < n->key)
        {
          if (dump_file && (dump_flags & TDF_DETAILS))
            {
              fprintf (dump_file,
                       "  decreasing badness %s/%i -> %s/%i, %i to %i\n",
                       xstrdup (cgraph_node_name (edge->caller)),
                       edge->caller->uid,
                       xstrdup (cgraph_node_name (edge->callee)),
                       edge->callee->uid,
                       (int)n->key,
                       badness);
            }
          fibheap_replace_key (heap, n, badness);
          gcc_checking_assert (n->key == badness);
        }
    }
  else
    {
       if (dump_file && (dump_flags & TDF_DETAILS))
         {
           fprintf (dump_file,
                    "  enqueuing call %s/%i -> %s/%i, badness %i\n",
                    xstrdup (cgraph_node_name (edge->caller)),
                    edge->caller->uid,
                    xstrdup (cgraph_node_name (edge->callee)),
                    edge->callee->uid,
                    badness);
         }
      edge->aux = fibheap_insert (heap, badness, edge);
    }
}


/* NODE was inlined.
   All caller edges needs to be resetted because
   size estimates change. Similarly callees needs reset
   because better context may be known.  */

static void
reset_edge_caches (struct cgraph_node *node)
{
  struct cgraph_edge *edge;
  struct cgraph_edge *e = node->callees;
  struct cgraph_node *where = node;
  int i;
  struct ipa_ref *ref;

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

  /* WHERE body size has changed, the cached growth is invalid.  */
  reset_node_growth_cache (where);

  for (edge = where->callers; edge; edge = edge->next_caller)
    if (edge->inline_failed)
      reset_edge_growth_cache (edge);
  for (i = 0; ipa_ref_list_referring_iterate (&where->symbol.ref_list,
                                              i, ref); i++)
    if (ref->use == IPA_REF_ALIAS)
      reset_edge_caches (ipa_ref_referring_node (ref));

  if (!e)
    return;

  while (true)
    if (!e->inline_failed && e->callee->callees)
      e = e->callee->callees;
    else
      {
        if (e->inline_failed)
          reset_edge_growth_cache (e);
        if (e->next_callee)
          e = e->next_callee;
        else
          {
            do
              {
                if (e->caller == node)
                  return;
                e = e->caller->callers;
              }
            while (!e->next_callee);
            e = e->next_callee;
          }
      }
}

/* Recompute HEAP nodes for each of caller of NODE.
   UPDATED_NODES track nodes we already visited, to avoid redundant work.
   When CHECK_INLINABLITY_FOR is set, re-check for specified edge that
   it is inlinable. Otherwise check all edges.  */

static void
update_caller_keys (fibheap_t heap, struct cgraph_node *node,
                    bitmap updated_nodes,
                    struct cgraph_edge *check_inlinablity_for)
{
  struct cgraph_edge *edge;
  int i;
  struct ipa_ref *ref;

  if ((!node->alias && !inline_summary (node)->inlinable)
      || cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE
      || node->global.inlined_to)
    return;
  if (!bitmap_set_bit (updated_nodes, node->uid))
    return;

  for (i = 0; ipa_ref_list_referring_iterate (&node->symbol.ref_list,
                                              i, ref); i++)
    if (ref->use == IPA_REF_ALIAS)
      {
        struct cgraph_node *alias = ipa_ref_referring_node (ref);
        update_caller_keys (heap, alias, updated_nodes, check_inlinablity_for);
      }

  for (edge = node->callers; edge; edge = edge->next_caller)
    if (edge->inline_failed)
      {
        if (!check_inlinablity_for
            || check_inlinablity_for == edge)
          {
            if (can_inline_edge_p (edge, false)
                && want_inline_small_function_p (edge, false))
              update_edge_key (heap, edge);
            else if (edge->aux)
              {
                report_inline_failed_reason (edge);
                fibheap_delete_node (heap, (fibnode_t) edge->aux);
                edge->aux = NULL;
              }
          }
        else if (edge->aux)
          update_edge_key (heap, edge);
      }
}

/* Recompute HEAP nodes for each uninlined call in NODE.
   This is used when we know that edge badnesses are going only to increase
   (we introduced new call site) and thus all we need is to insert newly
   created edges into heap.  */

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

  if (!e)
    return;
  while (true)
    if (!e->inline_failed && e->callee->callees)
      e = e->callee->callees;
    else
      {
        enum availability avail;
        struct cgraph_node *callee;
        /* We do not reset callee growth cache here.  Since we added a new call,
           growth chould have just increased and consequentely badness metric
           don't need updating.  */
        if (e->inline_failed
            && (callee = cgraph_function_or_thunk_node (e->callee, &avail))
            && inline_summary (callee)->inlinable
            && cgraph_function_body_availability (callee) >= AVAIL_AVAILABLE
            && !bitmap_bit_p (updated_nodes, callee->uid))
          {
            if (can_inline_edge_p (e, false)
                && want_inline_small_function_p (e, false))
              update_edge_key (heap, e);
            else if (e->aux)
              {
                report_inline_failed_reason (e);
                fibheap_delete_node (heap, (fibnode_t) e->aux);
                e->aux = NULL;
              }
          }
        if (e->next_callee)
          e = e->next_callee;
        else
          {
            do
              {
                if (e->caller == node)
                  return;
                e = e->caller->callers;
              }
            while (!e->next_callee);
            e = e->next_callee;
          }
      }
}

/* 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)
{
  struct cgraph_edge *e;
  enum availability avail;

  for (e = where->callees; e; e = e->next_callee)
    if (e->callee == node
        || (cgraph_function_or_thunk_node (e->callee, &avail) == node
            && avail > AVAIL_OVERWRITABLE))
      {
        /* When profile feedback is available, prioritize by expected number
           of calls.  */
        fibheap_insert (heap,
                        !max_count ? -e->frequency
                        : -(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.  If any new indirect edges
   are discovered in the process, add them to *NEW_EDGES, unless NEW_EDGES
   is NULL.  */

static bool
recursive_inlining (struct cgraph_edge *edge,
                    vec<cgraph_edge_p> *new_edges)
{
  int limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE_AUTO);
  fibheap_t heap;
  struct cgraph_node *node;
  struct cgraph_edge *e;
  struct cgraph_node *master_clone = NULL, *next;
  int depth = 0;
  int n = 0;

  node = edge->caller;
  if (node->global.inlined_to)
    node = node->global.inlined_to;

  if (DECL_DECLARED_INLINE_P (node->symbol.decl))
    limit = PARAM_VALUE (PARAM_MAX_INLINE_INSNS_RECURSIVE);

  /* Make sure that function is small enough to be considered for inlining.  */
  if (estimate_size_after_inlining (node, edge)  >= 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));

  /* Do the inlining and update list of recursive call during process.  */
  while (!fibheap_empty (heap))
    {
      struct cgraph_edge *curr
        = (struct cgraph_edge *) fibheap_extract_min (heap);
      struct cgraph_node *cnode, *dest = curr->callee;

      if (!can_inline_edge_p (curr, true))
        continue;

      /* MASTER_CLONE is produced in the case we already started modified
         the function. Be sure to redirect edge to the original body before
         estimating growths otherwise we will be seeing growths after inlining
         the already modified body.  */
      if (master_clone)
        {
          cgraph_redirect_edge_callee (curr, master_clone);
          reset_edge_growth_cache (curr);
        }

      if (estimate_size_after_inlining (node, curr) > limit)
        {
          cgraph_redirect_edge_callee (curr, dest);
          reset_edge_growth_cache (curr);
          break;
        }

      depth = 1;
      for (cnode = curr->caller;
           cnode->global.inlined_to; cnode = cnode->callers->caller)
        if (node->symbol.decl
            == cgraph_function_or_thunk_node (curr->callee, NULL)->symbol.decl)
          depth++;

      if (!want_inline_self_recursive_call_p (curr, node, false, depth))
        {
          cgraph_redirect_edge_callee (curr, dest);
          reset_edge_growth_cache (curr);
          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");
        }
      if (!master_clone)
        {
          /* We need original clone to copy around.  */
          master_clone = cgraph_clone_node (node, node->symbol.decl,
                                            node->count, CGRAPH_FREQ_BASE,
                                            false, vNULL, true);
          for (e = master_clone->callees; e; e = e->next_callee)
            if (!e->inline_failed)
              clone_inlined_nodes (e, true, false, NULL);
          cgraph_redirect_edge_callee (curr, master_clone);
          reset_edge_growth_cache (curr);
        }

      inline_call (curr, false, new_edges, &overall_size, true);
      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 (!master_clone)
    return false;

  if (dump_file)
    fprintf (dump_file,
             "\n   Inlined %i times, "
             "body grown from size %i to %i, time %i to %i\n", n,
             inline_summary (master_clone)->size, inline_summary (node)->size,
             inline_summary (master_clone)->time, inline_summary (node)->time);

  /* 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_first_function (); node != master_clone;
       node = next)
    {
      next = cgraph_next_function (node);
      if (node->global.inlined_to == master_clone)
        cgraph_remove_node (node);
    }
  cgraph_remove_node (master_clone);
  return true;
}


/* 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);
}


/* Compute badness of all edges in NEW_EDGES and add them to the HEAP.  */

static void
add_new_edges_to_heap (fibheap_t heap, vec<cgraph_edge_p> new_edges)
{
  while (new_edges.length () > 0)
    {
      struct cgraph_edge *edge = new_edges.pop ();

      gcc_assert (!edge->aux);
      if (edge->inline_failed
          && can_inline_edge_p (edge, true)
          && want_inline_small_function_p (edge, true))
        edge->aux = fibheap_insert (heap, edge_badness (edge, false), edge);
    }
}


/* We use greedy algorithm for inlining of small functions:
   All inline candidates are put into prioritized heap ordered in
   increasing badness.

   The inlining of small functions is bounded by unit growth parameters.  */

static void
inline_small_functions (void)
{
  struct cgraph_node *node;
  struct cgraph_edge *edge;
  fibheap_t edge_heap = fibheap_new ();
  bitmap updated_nodes = BITMAP_ALLOC (NULL);
  int min_size, max_size;
  vec<cgraph_edge_p> new_indirect_edges = vNULL;
  int initial_size = 0;
  struct cgraph_node **order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);

  if (flag_indirect_inlining)
    new_indirect_edges.create (8);

  /* Compute overall unit size and other global parameters used by badness
     metrics.  */

  max_count = 0;
  ipa_reduced_postorder (order, true, true, NULL);
  free (order);

  FOR_EACH_DEFINED_FUNCTION (node)
    if (!node->global.inlined_to)
      {
        if (cgraph_function_with_gimple_body_p (node)
            || node->thunk.thunk_p)
          {
            struct inline_summary *info = inline_summary (node);
            struct ipa_dfs_info *dfs = (struct ipa_dfs_info *) node->symbol.aux;

            if (!DECL_EXTERNAL (node->symbol.decl))
              initial_size += info->size;
            info->growth = estimate_growth (node);
            if (dfs && dfs->next_cycle)
              {
                struct cgraph_node *n2;
                int id = dfs->scc_no + 1;
                for (n2 = node; n2;
                     n2 = ((struct ipa_dfs_info *) node->symbol.aux)->next_cycle)
                  {
                    struct inline_summary *info2 = inline_summary (n2);
                    if (info2->scc_no)
                      break;
                    info2->scc_no = id;
                  }
              }
          }

        for (edge = node->callers; edge; edge = edge->next_caller)
          if (max_count < edge->count)
            max_count = edge->count;
      }
  ipa_free_postorder_info ();
  initialize_growth_caches ();

  if (dump_file)
    fprintf (dump_file,
             "\nDeciding on inlining of small functions.  Starting with size %i.\n",
             initial_size);

  overall_size = initial_size;
  max_size = compute_max_insns (overall_size);
  min_size = overall_size;

  /* Populate the heeap with all edges we might inline.  */

  FOR_EACH_DEFINED_FUNCTION (node)
    if (!node->global.inlined_to)
      {
        if (dump_file)
          fprintf (dump_file, "Enqueueing calls of %s/%i.\n",
                   cgraph_node_name (node), node->uid);

        for (edge = node->callers; edge; edge = edge->next_caller)
          if (edge->inline_failed
              && can_inline_edge_p (edge, true)
              && want_inline_small_function_p (edge, true)
              && edge->inline_failed)
            {
              gcc_assert (!edge->aux);
              update_edge_key (edge_heap, edge);
            }
      }

  gcc_assert (in_lto_p
              || !max_count
              || (profile_info && flag_branch_probabilities));

  while (!fibheap_empty (edge_heap))
    {
      int old_size = overall_size;
      struct cgraph_node *where, *callee;
      int badness = fibheap_min_key (edge_heap);
      int current_badness;
      int cached_badness;
      int growth;

      edge = (struct cgraph_edge *) fibheap_extract_min (edge_heap);
      gcc_assert (edge->aux);
      edge->aux = NULL;
      if (!edge->inline_failed)
        continue;

      /* Be sure that caches are maintained consistent.  
         We can not make this ENABLE_CHECKING only because it cause different
         updates of the fibheap queue.  */
      cached_badness = edge_badness (edge, false);
      reset_edge_growth_cache (edge);
      reset_node_growth_cache (edge->callee);

      /* When updating the edge costs, we only decrease badness in the keys.
         Increases of badness are handled lazilly; when we see key with out
         of date value on it, we re-insert it now.  */
      current_badness = edge_badness (edge, false);
      gcc_assert (cached_badness == current_badness);
      gcc_assert (current_badness >= badness);
      if (current_badness != badness)
        {
          edge->aux = fibheap_insert (edge_heap, current_badness, edge);
          continue;
        }

      if (!can_inline_edge_p (edge, true))
        continue;
      
      callee = cgraph_function_or_thunk_node (edge->callee, NULL);
      growth = estimate_edge_growth (edge);
      if (dump_file)
        {
          fprintf (dump_file,
                   "\nConsidering %s with %i size\n",
                   cgraph_node_name (callee),
                   inline_summary (callee)->size);
          fprintf (dump_file,
                   " to be inlined into %s in %s:%i\n"
                   " Estimated growth after inlined into all is %+i insns.\n"
                   " Estimated badness is %i, frequency %.2f.\n",
                   cgraph_node_name (edge->caller),
                   flag_wpa ? "unknown"
                   : gimple_filename ((const_gimple) edge->call_stmt),
                   flag_wpa ? -1
                   : gimple_lineno ((const_gimple) edge->call_stmt),
                   estimate_growth (callee),
                   badness,
                   edge->frequency / (double)CGRAPH_FREQ_BASE);
          if (edge->count)
            fprintf (dump_file," Called "HOST_WIDEST_INT_PRINT_DEC"x\n",
                     edge->count);
          if (dump_flags & TDF_DETAILS)
            edge_badness (edge, true);
        }

      if (overall_size + growth > max_size
          && !DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl))
        {
          edge->inline_failed = CIF_INLINE_UNIT_GROWTH_LIMIT;
          report_inline_failed_reason (edge);
          continue;
        }

      if (!want_inline_small_function_p (edge, true))
        continue;

      /* Heuristics for inlining small functions works poorly for
         recursive calls where we do efect similar to loop unrolling.
         When inliing such edge seems profitable, leave decision on
         specific inliner.  */
      if (cgraph_edge_recursive_p (edge))
        {
          where = edge->caller;
          if (where->global.inlined_to)
            where = where->global.inlined_to;
          if (!recursive_inlining (edge,
                                   flag_indirect_inlining
                                   ? &new_indirect_edges : NULL))
            {
              edge->inline_failed = CIF_RECURSIVE_INLINING;
              continue;
            }
          reset_edge_caches (where);
          /* Recursive inliner inlines all recursive calls of the function
             at once. Consequently we need to update all callee keys.  */
          if (flag_indirect_inlining)
            add_new_edges_to_heap (edge_heap, new_indirect_edges);
          update_callee_keys (edge_heap, where, updated_nodes);
        }
      else
        {
          struct cgraph_node *outer_node = NULL;
          int depth = 0;

          /* Consider the case where self recursive function A is inlined into B.
             This is desired optimization in some cases, since it leads to effect
             similar of loop peeling and we might completely optimize out the
             recursive call.  However we must be extra selective.  */

          where = edge->caller;
          while (where->global.inlined_to)
            {
              if (where->symbol.decl == callee->symbol.decl)
                outer_node = where, depth++;
              where = where->callers->caller;
            }
          if (outer_node
              && !want_inline_self_recursive_call_p (edge, outer_node,
                                                     true, depth))
            {
              edge->inline_failed
                = (DECL_DISREGARD_INLINE_LIMITS (edge->callee->symbol.decl)
                   ? CIF_RECURSIVE_INLINING : CIF_UNSPECIFIED);
              continue;
            }
          else if (depth && dump_file)
            fprintf (dump_file, " Peeling recursion with depth %i\n", depth);

          gcc_checking_assert (!callee->global.inlined_to);
          inline_call (edge, true, &new_indirect_edges, &overall_size, true);
          if (flag_indirect_inlining)
            add_new_edges_to_heap (edge_heap, new_indirect_edges);

          reset_edge_caches (edge->callee);
          reset_node_growth_cache (callee);

          update_callee_keys (edge_heap, where, 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 (edge_heap, where, updated_nodes, NULL);
      bitmap_clear (updated_nodes);

      if (dump_file)
        {
          fprintf (dump_file,
                   " Inlined into %s which now has time %i and size %i,"
                   "net change of %+i.\n",
                   cgraph_node_name (edge->caller),
                   inline_summary (edge->caller)->time,
                   inline_summary (edge->caller)->size,
                   overall_size - old_size);
        }
      if (min_size > overall_size)
        {
          min_size = overall_size;
          max_size = compute_max_insns (min_size);

          if (dump_file)
            fprintf (dump_file, "New minimal size reached: %i\n", min_size);
        }
    }

  free_growth_caches ();
  new_indirect_edges.release ();
  fibheap_delete (edge_heap);
  if (dump_file)
    fprintf (dump_file,
             "Unit growth for small function inlining: %i->%i (%i%%)\n",
             initial_size, overall_size,
             initial_size ? overall_size * 100 / (initial_size) - 100: 0);
  BITMAP_FREE (updated_nodes);
}

/* Flatten NODE.  Performed both during early inlining and
   at IPA inlining time.  */

static void
flatten_function (struct cgraph_node *node, bool early)
{
  struct cgraph_edge *e;

  /* We shouldn't be called recursively when we are being processed.  */
  gcc_assert (node->symbol.aux == NULL);

  node->symbol.aux = (void *) node;

  for (e = node->callees; e; e = e->next_callee)
    {
      struct cgraph_node *orig_callee;
      struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);

      /* We've hit cycle?  It is time to give up.  */
      if (callee->symbol.aux)
        {
          if (dump_file)
            fprintf (dump_file,
                     "Not inlining %s into %s to avoid cycle.\n",
                     xstrdup (cgraph_node_name (callee)),
                     xstrdup (cgraph_node_name (e->caller)));
          e->inline_failed = CIF_RECURSIVE_INLINING;
          continue;
        }

      /* When the edge is already inlined, we just need to recurse into
         it in order to fully flatten the leaves.  */
      if (!e->inline_failed)
        {
          flatten_function (callee, early);
          continue;
        }

      /* Flatten attribute needs to be processed during late inlining. For
         extra code quality we however do flattening during early optimization,
         too.  */
      if (!early
          ? !can_inline_edge_p (e, true)
          : !can_early_inline_edge_p (e))
        continue;

      if (cgraph_edge_recursive_p (e))
        {
          if (dump_file)
            fprintf (dump_file, "Not inlining: recursive call.\n");
          continue;
        }

      if (gimple_in_ssa_p (DECL_STRUCT_FUNCTION (node->symbol.decl))
          != gimple_in_ssa_p (DECL_STRUCT_FUNCTION (callee->symbol.decl)))
        {
          if (dump_file)
            fprintf (dump_file, "Not inlining: SSA form does not match.\n");
          continue;
        }

      /* Inline the edge and flatten the inline clone.  Avoid
         recursing through the original node if the node was cloned.  */
      if (dump_file)
        fprintf (dump_file, " Inlining %s into %s.\n",
                 xstrdup (cgraph_node_name (callee)),
                 xstrdup (cgraph_node_name (e->caller)));
      orig_callee = callee;
      inline_call (e, true, NULL, NULL, false);
      if (e->callee != orig_callee)
        orig_callee->symbol.aux = (void *) node;
      flatten_function (e->callee, early);
      if (e->callee != orig_callee)
        orig_callee->symbol.aux = NULL;
    }

  node->symbol.aux = NULL;
  if (!node->global.inlined_to)
    inline_update_overall_summary (node);
}

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

static unsigned int
ipa_inline (void)
{
  struct cgraph_node *node;
  int nnodes;
  struct cgraph_node **order =
    XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
  int i;

  if (in_lto_p && optimize)
    ipa_update_after_lto_read ();

  if (dump_file)
    dump_inline_summaries (dump_file);

  nnodes = ipa_reverse_postorder (order);

  FOR_EACH_FUNCTION (node)
    node->symbol.aux = 0;

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

  /* In the first pass handle functions to be flattened.  Do this with
     a priority so none of our later choices will make this impossible.  */
  for (i = nnodes - 1; i >= 0; i--)
    {
      node = order[i];

      /* Handle nodes to be flattened.
         Ideally when processing callees we stop inlining at the
         entry of cycles, possibly cloning that entry point and
         try to flatten itself turning it into a self-recursive
         function.  */
      if (lookup_attribute ("flatten",
                            DECL_ATTRIBUTES (node->symbol.decl)) != NULL)
        {
          if (dump_file)
            fprintf (dump_file,
                     "Flattening %s\n", cgraph_node_name (node));
          flatten_function (node, false);
        }
    }

  inline_small_functions ();
  symtab_remove_unreachable_nodes (false, dump_file);
  free (order);

  /* Inline functions with a property that after inlining into all callers the
     code size will shrink because the out-of-line copy is eliminated. 
     We do this regardless on the callee size as long as function growth limits
     are met.  */
  if (flag_inline_functions_called_once)
    {
      int cold;
      if (dump_file)
        fprintf (dump_file,
                 "\nDeciding on functions to be inlined into all callers:\n");

      /* Inlining one function called once has good chance of preventing
         inlining other function into the same callee.  Ideally we should
         work in priority order, but probably inlining hot functions first
         is good cut without the extra pain of maintaining the queue.

         ??? this is not really fitting the bill perfectly: inlining function
         into callee often leads to better optimization of callee due to
         increased context for optimization.
         For example if main() function calls a function that outputs help
         and then function that does the main optmization, we should inline
         the second with priority even if both calls are cold by themselves.

         We probably want to implement new predicate replacing our use of
         maybe_hot_edge interpreted as maybe_hot_edge || callee is known
         to be hot.  */
      for (cold = 0; cold <= 1; cold ++)
        {
          FOR_EACH_DEFINED_FUNCTION (node)
            {
              if (want_inline_function_to_all_callers_p (node, cold))
                {
                  int num_calls = 0;
                  struct cgraph_edge *e;
                  for (e = node->callers; e; e = e->next_caller)
                    num_calls++;
                  while (node->callers && !node->global.inlined_to)
                    {
                      struct cgraph_node *caller = node->callers->caller;

                      if (dump_file)
                        {
                          fprintf (dump_file,
                                   "\nInlining %s size %i.\n",
                                   cgraph_node_name (node),
                                   inline_summary (node)->size);
                          fprintf (dump_file,
                                   " Called once from %s %i insns.\n",
                                   cgraph_node_name (node->callers->caller),
                                   inline_summary (node->callers->caller)->size);
                        }

                      inline_call (node->callers, true, NULL, NULL, true);
                      if (dump_file)
                        fprintf (dump_file,
                                 " Inlined into %s which now has %i size\n",
                                 cgraph_node_name (caller),
                                 inline_summary (caller)->size);
                      if (!num_calls--)
                        {
                          if (dump_file)
                            fprintf (dump_file, "New calls found; giving up.\n");
                          break;
                        }
                    }
                }
            }
        }
    }

  /* Free ipa-prop structures if they are no longer needed.  */
  if (optimize)
    ipa_free_all_structures_after_iinln ();

  if (dump_file)
    fprintf (dump_file,
             "\nInlined %i calls, eliminated %i functions\n\n",
             ncalls_inlined, nfunctions_inlined);

  if (dump_file)
    dump_inline_summaries (dump_file);
  /* In WPA we use inline summaries for partitioning process.  */
  if (!flag_wpa)
    inline_free_summary ();
  return 0;
}

/* Inline always-inline function calls in NODE.  */

static bool
inline_always_inline_functions (struct cgraph_node *node)
{
  struct cgraph_edge *e;
  bool inlined = false;

  for (e = node->callees; e; e = e->next_callee)
    {
      struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);
      if (!DECL_DISREGARD_INLINE_LIMITS (callee->symbol.decl))
        continue;

      if (cgraph_edge_recursive_p (e))
        {
          if (dump_file)
            fprintf (dump_file, "  Not inlining recursive call to %s.\n",
                     cgraph_node_name (e->callee));
          e->inline_failed = CIF_RECURSIVE_INLINING;
          continue;
        }

      if (!can_early_inline_edge_p (e))
        continue;

      if (dump_file)
        fprintf (dump_file, "  Inlining %s into %s (always_inline).\n",
                 xstrdup (cgraph_node_name (e->callee)),
                 xstrdup (cgraph_node_name (e->caller)));
      inline_call (e, true, NULL, NULL, false);
      inlined = true;
    }
  if (inlined)
    inline_update_overall_summary (node);

  return inlined;
}

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

static bool
early_inline_small_functions (struct cgraph_node *node)
{
  struct cgraph_edge *e;
  bool inlined = false;

  for (e = node->callees; e; e = e->next_callee)
    {
      struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);
      if (!inline_summary (callee)->inlinable
          || !e->inline_failed)
        continue;

      /* Do not consider functions not declared inline.  */
      if (!DECL_DECLARED_INLINE_P (callee->symbol.decl)
          && !flag_inline_small_functions
          && !flag_inline_functions)
        continue;

      if (dump_file)
        fprintf (dump_file, "Considering inline candidate %s.\n",
                 cgraph_node_name (callee));

      if (!can_early_inline_edge_p (e))
        continue;

      if (cgraph_edge_recursive_p (e))
        {
          if (dump_file)
            fprintf (dump_file, "  Not inlining: recursive call.\n");
          continue;
        }

      if (!want_early_inline_function_p (e))
        continue;

      if (dump_file)
        fprintf (dump_file, " Inlining %s into %s.\n",
                 xstrdup (cgraph_node_name (callee)),
                 xstrdup (cgraph_node_name (e->caller)));
      inline_call (e, true, NULL, NULL, true);
      inlined = true;
    }

  return inlined;
}

/* 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
early_inliner (void)
{
  struct cgraph_node *node = cgraph_get_node (current_function_decl);
  struct cgraph_edge *edge;
  unsigned int todo = 0;
  int iterations = 0;
  bool inlined = false;

  if (seen_error ())
    return 0;

  /* Do nothing if datastructures for ipa-inliner are already computed.  This
     happens when some pass decides to construct new function and
     cgraph_add_new_function calls lowering passes and early optimization on
     it.  This may confuse ourself when early inliner decide to inline call to
     function clone, because function clones don't have parameter list in
     ipa-prop matching their signature.  */
  if (ipa_node_params_vector.exists ())
    return 0;

#ifdef ENABLE_CHECKING
  verify_cgraph_node (node);
#endif

  /* Even when not optimizing or not inlining inline always-inline
     functions.  */
  inlined = inline_always_inline_functions (node);

  if (!optimize
      || flag_no_inline
      || !flag_early_inlining
      /* Never inline regular functions into always-inline functions
         during incremental inlining.  This sucks as functions calling
         always inline functions will get less optimized, but at the
         same time inlining of functions calling always inline
         function into an always inline function might introduce
         cycles of edges to be always inlined in the callgraph.

         We might want to be smarter and just avoid this type of inlining.  */
      || DECL_DISREGARD_INLINE_LIMITS (node->symbol.decl))
    ;
  else if (lookup_attribute ("flatten",
                             DECL_ATTRIBUTES (node->symbol.decl)) != NULL)
    {
      /* When the function is marked to be flattened, recursively inline
         all calls in it.  */
      if (dump_file)
        fprintf (dump_file,
                 "Flattening %s\n", cgraph_node_name (node));
      flatten_function (node, true);
      inlined = true;
    }
  else
    {
      /* We iterate incremental inlining to get trivial cases of indirect
         inlining.  */
      while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
             && early_inline_small_functions (node))
        {
          timevar_push (TV_INTEGRATION);
          todo |= optimize_inline_calls (current_function_decl);

          /* Technically we ought to recompute inline parameters so the new
             iteration of early inliner works as expected.  We however have
             values approximately right and thus we only need to update edge
             info that might be cleared out for newly discovered edges.  */
          for (edge = node->callees; edge; edge = edge->next_callee)
            {
              struct inline_edge_summary *es = inline_edge_summary (edge);
              es->call_stmt_size
                = estimate_num_insns (edge->call_stmt, &eni_size_weights);
              es->call_stmt_time
                = estimate_num_insns (edge->call_stmt, &eni_time_weights);
              if (edge->callee->symbol.decl
                  && !gimple_check_call_matching_types (edge->call_stmt,
                                                        edge->callee->symbol.decl))
                edge->call_stmt_cannot_inline_p = true;
            }
          timevar_pop (TV_INTEGRATION);
          iterations++;
          inlined = false;
        }
      if (dump_file)
        fprintf (dump_file, "Iterations: %i\n", iterations);
    }

  if (inlined)
    {
      timevar_push (TV_INTEGRATION);
      todo |= optimize_inline_calls (current_function_decl);
      timevar_pop (TV_INTEGRATION);
    }

  cfun->always_inline_functions_inlined = true;

  return todo;
}

struct gimple_opt_pass pass_early_inline =
{
 {
  GIMPLE_PASS,
  "einline",                            /* name */
  OPTGROUP_INLINE,                      /* optinfo_flags */
  NULL,                                 /* gate */
  early_inliner,                        /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_EARLY_INLINING,                    /* tv_id */
  PROP_ssa,                             /* properties_required */
  0,                                    /* properties_provided */
  0,                                    /* properties_destroyed */
  0,                                    /* todo_flags_start */
  0                                     /* todo_flags_finish */
 }
};


/* When to run IPA inlining.  Inlining of always-inline functions
   happens during early inlining.

   Enable inlining unconditoinally at -flto.  We need size estimates to
   drive partitioning.  */

static bool
gate_ipa_inline (void)
{
  return optimize || flag_lto || flag_wpa;
}

struct ipa_opt_pass_d pass_ipa_inline =
{
 {
  IPA_PASS,
  "inline",                             /* name */
  OPTGROUP_INLINE,                      /* optinfo_flags */
  gate_ipa_inline,                      /* gate */
  ipa_inline,                           /* execute */
  NULL,                                 /* sub */
  NULL,                                 /* next */
  0,                                    /* static_pass_number */
  TV_IPA_INLINING,                      /* tv_id */
  0,                                    /* properties_required */
  0,                                    /* properties_provided */
  0,                                    /* properties_destroyed */
  TODO_remove_functions,                /* todo_flags_finish */
  TODO_dump_symtab 
  | TODO_remove_functions               /* todo_flags_finish */
 },
 inline_generate_summary,               /* generate_summary */
 inline_write_summary,                  /* write_summary */
 inline_read_summary,                   /* read_summary */
 NULL,                                  /* write_optimization_summary */
 NULL,                                  /* read_optimization_summary */
 NULL,                                  /* stmt_fixup */
 0,                                     /* TODOs */
 inline_transform,                      /* function_transform */
 NULL,                                  /* variable_transform */
};