Subversion Repositories scarts

/* Nested function decomposition for trees.

   Copyright (C) 2004, 2005 Free Software Foundation, Inc.

   This file is part of GCC.

   GCC is free software; you can redistribute it and/or modify

   it under the terms of the GNU General Public License as published by

   the Free Software Foundation; either version 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.  */

#include "config.h"

#include "system.h"

#include "coretypes.h"

#include "tm.h"

#include "tree.h"

#include "rtl.h"

#include "tm_p.h"

#include "function.h"

#include "tree-dump.h"

#include "tree-inline.h"

#include "tree-gimple.h"

#include "tree-iterator.h"

#include "tree-flow.h"

#include "cgraph.h"

#include "expr.h"

#include "langhooks.h"

#include "ggc.h"

/* The object of this pass is to lower the representation of a set of nested

   functions in order to expose all of the gory details of the various

   nonlocal references.  We want to do this sooner rather than later, in

   order to give us more freedom in emitting all of the functions in question.

   Back in olden times, when gcc was young, we developed an insanely

   complicated scheme whereby variables which were referenced nonlocally

   were forced to live in the stack of the declaring function, and then

   the nested functions magically discovered where these variables were

   placed.  In order for this scheme to function properly, it required

   that the outer function be partially expanded, then we switch to

   compiling the inner function, and once done with those we switch back

   to compiling the outer function.  Such delicate ordering requirements

   makes it difficult to do whole translation unit optimizations

   involving such functions.

   The implementation here is much more direct.  Everything that can be

   referenced by an inner function is a member of an explicitly created

   structure herein called the "nonlocal frame struct".  The incoming

   static chain for a nested function is a pointer to this struct in

   the parent.  In this way, we settle on known offsets from a known

   base, and so are decoupled from the logic that places objects in the

   function's stack frame.  More importantly, we don't have to wait for

   that to happen -- since the compilation of the inner function is no

   longer tied to a real stack frame, the nonlocal frame struct can be

   allocated anywhere.  Which means that the outer function is now

   inlinable.

   Theory of operation here is very simple.  Iterate over all the

   statements in all the functions (depth first) several times,

   allocating structures and fields on demand.  In general we want to

   examine inner functions first, so that we can avoid making changes

   to outer functions which are unnecessary.

   The order of the passes matters a bit, in that later passes will be

   skipped if it is discovered that the functions don't actually interact

   at all.  That is, they're nested in the lexical sense but could have

   been written as independent functions without change.  */

struct var_map_elt GTY(())

{

  tree old;

  tree new;

};

struct nesting_info GTY ((chain_next ("%h.next")))

{

  struct nesting_info *outer;

  struct nesting_info *inner;

  struct nesting_info *next;

  htab_t GTY ((param_is (struct var_map_elt))) var_map;

  tree context;

  tree new_local_var_chain;

  tree frame_type;

  tree frame_decl;

  tree chain_field;

  tree chain_decl;

  tree nl_goto_field;

  bool any_parm_remapped;

  bool any_tramp_created;

};

/* Hashing and equality functions for nesting_info->var_map.  */

static hashval_t

var_map_hash (const void *x)

{

  const struct var_map_elt *a = x;

  return htab_hash_pointer (a->old);

}

static int

var_map_eq (const void *x, const void *y)

{

  const struct var_map_elt *a = x;

  const struct var_map_elt *b = y;

  return a->old == b->old;

}

/* We're working in so many different function contexts simultaneously,

   that create_tmp_var is dangerous.  Prevent mishap.  */

#define create_tmp_var cant_use_create_tmp_var_here_dummy

/* Like create_tmp_var, except record the variable for registration at

   the given nesting level.  */

static tree

create_tmp_var_for (struct nesting_info *info, tree type, const char *prefix)

{

  tree tmp_var;

  /* If the type is of variable size or a type which must be created by the

     frontend, something is wrong.  Note that we explicitly allow

     incomplete types here, since we create them ourselves here.  */

  gcc_assert (!TREE_ADDRESSABLE (type));

  gcc_assert (!TYPE_SIZE_UNIT (type)

              || TREE_CODE (TYPE_SIZE_UNIT (type)) == INTEGER_CST);

  tmp_var = create_tmp_var_raw (type, prefix);

  DECL_CONTEXT (tmp_var) = info->context;

  TREE_CHAIN (tmp_var) = info->new_local_var_chain;

  DECL_SEEN_IN_BIND_EXPR_P (tmp_var) = 1;

  info->new_local_var_chain = tmp_var;

  return tmp_var;

}

/* Take the address of EXP to be used within function CONTEXT.

   Mark it for addressability as necessary.  */

tree

build_addr (tree exp, tree context)

{

  tree base = exp;

  tree save_context;

  tree retval;

  while (handled_component_p (base))

    base = TREE_OPERAND (base, 0);

  if (DECL_P (base))

    TREE_ADDRESSABLE (base) = 1;

  /* Building the ADDR_EXPR will compute a set of properties for

     that ADDR_EXPR.  Those properties are unfortunately context

     specific.  ie, they are dependent on CURRENT_FUNCTION_DECL.

     Temporarily set CURRENT_FUNCTION_DECL to the desired context,

     build the ADDR_EXPR, then restore CURRENT_FUNCTION_DECL.  That

     way the properties are for the ADDR_EXPR are computed properly.  */

  save_context = current_function_decl;

  current_function_decl = context;

  retval = build1 (ADDR_EXPR, build_pointer_type (TREE_TYPE (exp)), exp);

  current_function_decl = save_context;;

  return retval;

}

/* Insert FIELD into TYPE, sorted by alignment requirements.  */

static void

insert_field_into_struct (tree type, tree field)

{

  tree *p;

  DECL_CONTEXT (field) = type;

  for (p = &TYPE_FIELDS (type); *p ; p = &TREE_CHAIN (*p))

    if (DECL_ALIGN (field) >= DECL_ALIGN (*p))

      break;

  TREE_CHAIN (field) = *p;

  *p = field;

}

/* Build or return the RECORD_TYPE that describes the frame state that is

   shared between INFO->CONTEXT and its nested functions.  This record will

   not be complete until finalize_nesting_tree; up until that point we'll

   be adding fields as necessary.

   We also build the DECL that represents this frame in the function.  */

static tree

get_frame_type (struct nesting_info *info)

{

  tree type = info->frame_type;

  if (!type)

    {

      char *name;

      type = make_node (RECORD_TYPE);

      name = concat ("FRAME.",

                     IDENTIFIER_POINTER (DECL_NAME (info->context)),

                     NULL);

      TYPE_NAME (type) = get_identifier (name);

      free (name);

      info->frame_type = type;

      info->frame_decl = create_tmp_var_for (info, type, "FRAME");

      /* ??? Always make it addressable for now, since it is meant to

         be pointed to by the static chain pointer.  This pessimizes

         when it turns out that no static chains are needed because

         the nested functions referencing non-local variables are not

         reachable, but the true pessimization is to create the non-

         local frame structure in the first place.  */

      TREE_ADDRESSABLE (info->frame_decl) = 1;

    }

  return type;

}

/* Return true if DECL should be referenced by pointer in the non-local

   frame structure.  */

static bool

use_pointer_in_frame (tree decl)

{

  if (TREE_CODE (decl) == PARM_DECL)

    {

      /* It's illegal to copy TREE_ADDRESSABLE, impossible to copy variable

         sized decls, and inefficient to copy large aggregates.  Don't bother

         moving anything but scalar variables.  */

      return AGGREGATE_TYPE_P (TREE_TYPE (decl));

    }

  else

    {

      /* Variable sized types make things "interesting" in the frame.  */

      return DECL_SIZE (decl) == NULL || !TREE_CONSTANT (DECL_SIZE (decl));

    }

}

/* Given DECL, a non-locally accessed variable, find or create a field

   in the non-local frame structure for the given nesting context.  */

static tree

lookup_field_for_decl (struct nesting_info *info, tree decl,

                       enum insert_option insert)

{

  struct var_map_elt *elt, dummy;

  void **slot;

  tree field;

  dummy.old = decl;

  slot = htab_find_slot (info->var_map, &dummy, insert);

  if (!slot)

    {

      gcc_assert (insert != INSERT);

      return NULL;

    }

  elt = *slot;

  if (!elt && insert == INSERT)

    {

      field = make_node (FIELD_DECL);

      DECL_NAME (field) = DECL_NAME (decl);

      if (use_pointer_in_frame (decl))

        {

          TREE_TYPE (field) = build_pointer_type (TREE_TYPE (decl));

          DECL_ALIGN (field) = TYPE_ALIGN (TREE_TYPE (field));

          DECL_NONADDRESSABLE_P (field) = 1;

        }

      else

        {

          TREE_TYPE (field) = TREE_TYPE (decl);

          DECL_SOURCE_LOCATION (field) = DECL_SOURCE_LOCATION (decl);

          DECL_ALIGN (field) = DECL_ALIGN (decl);

          DECL_USER_ALIGN (field) = DECL_USER_ALIGN (decl);

          TREE_ADDRESSABLE (field) = TREE_ADDRESSABLE (decl);

          DECL_NONADDRESSABLE_P (field) = !TREE_ADDRESSABLE (decl);

          TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (decl);

        }

      insert_field_into_struct (get_frame_type (info), field);

      elt = ggc_alloc (sizeof (*elt));

      elt->old = decl;

      elt->new = field;

      *slot = elt;

      if (TREE_CODE (decl) == PARM_DECL)

        info->any_parm_remapped = true;

    }

  else

    field = elt ? elt->new : NULL;

  return field;

}

/* Build or return the variable that holds the static chain within

   INFO->CONTEXT.  This variable may only be used within INFO->CONTEXT.  */

static tree

get_chain_decl (struct nesting_info *info)

{

  tree decl = info->chain_decl;

  if (!decl)

    {

      tree type;

      type = get_frame_type (info->outer);

      type = build_pointer_type (type);

      /* Note that this variable is *not* entered into any BIND_EXPR;

         the construction of this variable is handled specially in

         expand_function_start and initialize_inlined_parameters.

         Note also that it's represented as a parameter.  This is more

         close to the truth, since the initial value does come from

         the caller.  */

      decl = build_decl (PARM_DECL, create_tmp_var_name ("CHAIN"), type);

      DECL_ARTIFICIAL (decl) = 1;

      DECL_IGNORED_P (decl) = 1;

      TREE_USED (decl) = 1;

      DECL_CONTEXT (decl) = info->context;

      DECL_ARG_TYPE (decl) = type;

      /* Tell tree-inline.c that we never write to this variable, so

         it can copy-prop the replacement value immediately.  */

      TREE_READONLY (decl) = 1;

      info->chain_decl = decl;

    }

  return decl;

}

/* Build or return the field within the non-local frame state that holds

   the static chain for INFO->CONTEXT.  This is the way to walk back up

   multiple nesting levels.  */

static tree

get_chain_field (struct nesting_info *info)

{

  tree field = info->chain_field;

  if (!field)

    {

      tree type = build_pointer_type (get_frame_type (info->outer));

      field = make_node (FIELD_DECL);

      DECL_NAME (field) = get_identifier ("__chain");

      TREE_TYPE (field) = type;

      DECL_ALIGN (field) = TYPE_ALIGN (type);

      DECL_NONADDRESSABLE_P (field) = 1;

      insert_field_into_struct (get_frame_type (info), field);

      info->chain_field = field;

    }

  return field;

}

/* Copy EXP into a temporary.  Allocate the temporary in the context of

   INFO and insert the initialization statement before TSI.  */

static tree

init_tmp_var (struct nesting_info *info, tree exp, tree_stmt_iterator *tsi)

{

  tree t, stmt;

  t = create_tmp_var_for (info, TREE_TYPE (exp), NULL);

  stmt = build (MODIFY_EXPR, TREE_TYPE (t), t, exp);

  SET_EXPR_LOCUS (stmt, EXPR_LOCUS (tsi_stmt (*tsi)));

  tsi_link_before (tsi, stmt, TSI_SAME_STMT);

  return t;

}

/* Similarly, but only do so to force EXP to satisfy is_gimple_val.  */

static tree

tsi_gimplify_val (struct nesting_info *info, tree exp, tree_stmt_iterator *tsi)

{

  if (is_gimple_val (exp))

    return exp;

  else

    return init_tmp_var (info, exp, tsi);

}

/* Similarly, but copy from the temporary and insert the statement

   after the iterator.  */

static tree

save_tmp_var (struct nesting_info *info, tree exp,

              tree_stmt_iterator *tsi)

{

  tree t, stmt;

  t = create_tmp_var_for (info, TREE_TYPE (exp), NULL);

  stmt = build (MODIFY_EXPR, TREE_TYPE (t), exp, t);

  SET_EXPR_LOCUS (stmt, EXPR_LOCUS (tsi_stmt (*tsi)));

  tsi_link_after (tsi, stmt, TSI_SAME_STMT);

  return t;

}

/* Build or return the type used to represent a nested function trampoline.  */

static GTY(()) tree trampoline_type;

static tree

get_trampoline_type (void)

{

  tree record, t;

  unsigned align, size;

  if (trampoline_type)

    return trampoline_type;

  align = TRAMPOLINE_ALIGNMENT;

  size = TRAMPOLINE_SIZE;

  /* If we won't be able to guarantee alignment simply via TYPE_ALIGN,

     then allocate extra space so that we can do dynamic alignment.  */

  if (align > STACK_BOUNDARY)

    {

      size += ((align/BITS_PER_UNIT) - 1) & -(STACK_BOUNDARY/BITS_PER_UNIT);

      align = STACK_BOUNDARY;

    }

  t = build_index_type (build_int_cst (NULL_TREE, size - 1));

  t = build_array_type (char_type_node, t);

  t = build_decl (FIELD_DECL, get_identifier ("__data"), t);

  DECL_ALIGN (t) = align;

  DECL_USER_ALIGN (t) = 1;

  record = make_node (RECORD_TYPE);

  TYPE_NAME (record) = get_identifier ("__builtin_trampoline");

  TYPE_FIELDS (record) = t;

  layout_type (record);

  return record;

}

/* Given DECL, a nested function, find or create a field in the non-local

   frame structure for a trampoline for this function.  */

static tree

lookup_tramp_for_decl (struct nesting_info *info, tree decl,

                       enum insert_option insert)

{

  struct var_map_elt *elt, dummy;

  void **slot;

  tree field;

  dummy.old = decl;

  slot = htab_find_slot (info->var_map, &dummy, insert);

  if (!slot)

    {

      gcc_assert (insert != INSERT);

      return NULL;

    }

  elt = *slot;

  if (!elt && insert == INSERT)

    {

      field = make_node (FIELD_DECL);

      DECL_NAME (field) = DECL_NAME (decl);

      TREE_TYPE (field) = get_trampoline_type ();

      TREE_ADDRESSABLE (field) = 1;

      insert_field_into_struct (get_frame_type (info), field);

      elt = ggc_alloc (sizeof (*elt));

      elt->old = decl;

      elt->new = field;

      *slot = elt;

      info->any_tramp_created = true;

    }

  else

    field = elt ? elt->new : NULL;

  return field;

}

/* Build or return the field within the non-local frame state that holds

   the non-local goto "jmp_buf".  The buffer itself is maintained by the

   rtl middle-end as dynamic stack space is allocated.  */

static tree

get_nl_goto_field (struct nesting_info *info)

{

  tree field = info->nl_goto_field;

  if (!field)

    {

      unsigned size;

      tree type;

      /* For __builtin_nonlocal_goto, we need N words.  The first is the

         frame pointer, the rest is for the target's stack pointer save

         area.  The number of words is controlled by STACK_SAVEAREA_MODE;

         not the best interface, but it'll do for now.  */

      if (Pmode == ptr_mode)

        type = ptr_type_node;

      else

        type = lang_hooks.types.type_for_mode (Pmode, 1);

      size = GET_MODE_SIZE (STACK_SAVEAREA_MODE (SAVE_NONLOCAL));

      size = size / GET_MODE_SIZE (Pmode);

      size = size + 1;

      type = build_array_type

        (type, build_index_type (build_int_cst (NULL_TREE, size)));

      field = make_node (FIELD_DECL);

      DECL_NAME (field) = get_identifier ("__nl_goto_buf");

      TREE_TYPE (field) = type;

      DECL_ALIGN (field) = TYPE_ALIGN (type);

      TREE_ADDRESSABLE (field) = 1;

      insert_field_into_struct (get_frame_type (info), field);

      info->nl_goto_field = field;

    }

  return field;

}

/* Convenience routines to walk all statements of a gimple function.

   For each statement, we invoke CALLBACK via walk_tree.  The passed

   data is a walk_stmt_info structure.  Of note here is a TSI that

   points to the current statement being walked.  The VAL_ONLY flag

   that indicates whether the *TP being examined may be replaced

   with something that matches is_gimple_val (if true) or something

   slightly more complicated (if false).  "Something" technically

   means the common subset of is_gimple_lvalue and is_gimple_rhs,

   but we never try to form anything more complicated than that, so

   we don't bother checking.  */

struct walk_stmt_info

{

  walk_tree_fn callback;

  tree_stmt_iterator tsi;

  struct nesting_info *info;

  bool val_only;

  bool is_lhs;

  bool changed;

};

/* A subroutine of walk_function.  Iterate over all sub-statements of *TP.  */

static void

walk_stmts (struct walk_stmt_info *wi, tree *tp)

{

  tree t = *tp;

  if (!t)

    return;

  switch (TREE_CODE (t))

    {

    case STATEMENT_LIST:

      {

        tree_stmt_iterator i;

        for (i = tsi_start (t); !tsi_end_p (i); tsi_next (&i))

          {

            wi->tsi = i;

            walk_stmts (wi, tsi_stmt_ptr (i));

          }

      }

      break;

    case COND_EXPR:

      walk_tree (&COND_EXPR_COND (t), wi->callback, wi, NULL);

      walk_stmts (wi, &COND_EXPR_THEN (t));

      walk_stmts (wi, &COND_EXPR_ELSE (t));

      break;

    case CATCH_EXPR:

      walk_stmts (wi, &CATCH_BODY (t));

      break;

    case EH_FILTER_EXPR:

      walk_stmts (wi, &EH_FILTER_FAILURE (t));

      break;

    case TRY_CATCH_EXPR:

    case TRY_FINALLY_EXPR:

      walk_stmts (wi, &TREE_OPERAND (t, 0));

      walk_stmts (wi, &TREE_OPERAND (t, 1));

      break;

    case BIND_EXPR:

      walk_stmts (wi, &BIND_EXPR_BODY (t));

      break;

    case RETURN_EXPR:

      walk_stmts (wi, &TREE_OPERAND (t, 0));

      break;

    case MODIFY_EXPR:

      /* A formal temporary lhs may use a COMPONENT_REF rhs.  */

      wi->val_only = !is_gimple_formal_tmp_var (TREE_OPERAND (t, 0));

      walk_tree (&TREE_OPERAND (t, 1), wi->callback, wi, NULL);

      /* If the rhs is appropriate for a memory, we may use a

         COMPONENT_REF on the lhs.  */

      wi->val_only = !is_gimple_mem_rhs (TREE_OPERAND (t, 1));

      wi->is_lhs = true;

      walk_tree (&TREE_OPERAND (t, 0), wi->callback, wi, NULL);

      wi->val_only = true;

      wi->is_lhs = false;

      break;

    default:

      wi->val_only = true;

      walk_tree (tp, wi->callback, wi, NULL);

      break;

    }

}

/* Invoke CALLBACK on all statements of INFO->CONTEXT.  */

static void

walk_function (walk_tree_fn callback, struct nesting_info *info)

{

  struct walk_stmt_info wi;

  memset (&wi, 0, sizeof (wi));

  wi.callback = callback;

  wi.info = info;

  wi.val_only = true;

  walk_stmts (&wi, &DECL_SAVED_TREE (info->context));

}

/* Similarly for ROOT and all functions nested underneath, depth first.  */

static void

walk_all_functions (walk_tree_fn callback, struct nesting_info *root)

{

  do

    {

      if (root->inner)

        walk_all_functions (callback, root->inner);

      walk_function (callback, root);

      root = root->next;

    }

  while (root);

}

/* We have to check for a fairly pathological case.  The operands of function

   nested function are to be interpreted in the context of the enclosing

   function.  So if any are variably-sized, they will get remapped when the

   enclosing function is inlined.  But that remapping would also have to be

   done in the types of the PARM_DECLs of the nested function, meaning the

   argument types of that function will disagree with the arguments in the

   calls to that function.  So we'd either have to make a copy of the nested

   function corresponding to each time the enclosing function was inlined or

   add a VIEW_CONVERT_EXPR to each such operand for each call to the nested

   function.  The former is not practical.  The latter would still require

   detecting this case to know when to add the conversions.  So, for now at

   least, we don't inline such an enclosing function.

   We have to do that check recursively, so here return indicating whether

   FNDECL has such a nested function.  ORIG_FN is the function we were

   trying to inline to use for checking whether any argument is variably

   modified by anything in it.

   It would be better to do this in tree-inline.c so that we could give

   the appropriate warning for why a function can't be inlined, but that's

   too late since the nesting structure has already been flattened and

   adding a flag just to record this fact seems a waste of a flag.  */

static bool

check_for_nested_with_variably_modified (tree fndecl, tree orig_fndecl)

{

  struct cgraph_node *cgn = cgraph_node (fndecl);

  tree arg;

  for (cgn = cgn->nested; cgn ; cgn = cgn->next_nested)

    {

      for (arg = DECL_ARGUMENTS (cgn->decl); arg; arg = TREE_CHAIN (arg))

        if (variably_modified_type_p (TREE_TYPE (arg), 0), orig_fndecl)

          return true;

      if (check_for_nested_with_variably_modified (cgn->decl, orig_fndecl))

        return true;

    }

  return false;

}

/* Construct our local datastructure describing the function nesting

   tree rooted by CGN.  */

static struct nesting_info *

create_nesting_tree (struct cgraph_node *cgn)

{

  struct nesting_info *info = ggc_calloc (1, sizeof (*info));

  info->var_map = htab_create_ggc (7, var_map_hash, var_map_eq, ggc_free);

  info->context = cgn->decl;

  for (cgn = cgn->nested; cgn ; cgn = cgn->next_nested)

    {

      struct nesting_info *sub = create_nesting_tree (cgn);

      sub->outer = info;

      sub->next = info->inner;