Subversion Repositories openrisc_me

/* Allocation for dataflow support routines.

   Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007

   Free Software Foundation, Inc.

   Originally contributed by Michael P. Hayes

             (m.hayes@elec.canterbury.ac.nz, mhayes@redhat.com)

   Major rewrite contributed by Danny Berlin (dberlin@dberlin.org)

             and Kenneth Zadeck (zadeck@naturalbridge.com).

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/>.  */

/*

OVERVIEW:

The files in this collection (df*.c,df.h) provide a general framework

for solving dataflow problems.  The global dataflow is performed using

a good implementation of iterative dataflow analysis.

The file df-problems.c provides problem instance for the most common

dataflow problems: reaching defs, upward exposed uses, live variables,

uninitialized variables, def-use chains, and use-def chains.  However,

the interface allows other dataflow problems to be defined as well.

USAGE:

Here is an example of using the dataflow routines.

      struct df *df;

      df = df_init (init_flags);

      df_add_problem (df, problem, flags);

      df_set_blocks (df, blocks);

      df_rescan_blocks (df, blocks);

      df_analyze (df);

      df_dump (df, stderr);

      df_finish (df);

DF_INIT simply creates a poor man's object (df) that needs to be

passed to all the dataflow routines.  df_finish destroys this object

and frees up any allocated memory.

There are three flags that can be passed to df_init, each of these

flags controls the scanning of the rtl:

DF_HARD_REGS means that the scanning is to build information about

both pseudo registers and hardware registers.  Without this

information, the problems will be solved only on pseudo registers.

DF_EQUIV_NOTES marks the uses present in EQUIV/EQUAL notes.

DF_SUBREGS return subregs rather than the inner reg.

DF_ADD_PROBLEM adds a problem, defined by an instance to struct

df_problem, to the set of problems solved in this instance of df.  All

calls to add a problem for a given instance of df must occur before

the first call to DF_RESCAN_BLOCKS, DF_SET_BLOCKS or DF_ANALYZE.

For all of the problems defined in df-problems.c, there are

convenience functions named DF_*_ADD_PROBLEM.

Problems can be dependent on other problems.  For instance, solving

def-use or use-def chains is dependent on solving reaching

definitions. As long as these dependencies are listed in the problem

definition, the order of adding the problems is not material.

Otherwise, the problems will be solved in the order of calls to

df_add_problem.  Note that it is not necessary to have a problem.  In

that case, df will just be used to do the scanning.

DF_SET_BLOCKS is an optional call used to define a region of the

function on which the analysis will be performed.  The normal case is

to analyze the entire function and no call to df_set_blocks is made.

When a subset is given, the analysis behaves as if the function only

contains those blocks and any edges that occur directly between the

blocks in the set.  Care should be taken to call df_set_blocks right

before the call to analyze in order to eliminate the possibility that

optimizations that reorder blocks invalidate the bitvector.

DF_RESCAN_BLOCKS is an optional call that causes the scanner to be

 (re)run over the set of blocks passed in.  If blocks is NULL, the entire

function (or all of the blocks defined in df_set_blocks) is rescanned.

If blocks contains blocks that were not defined in the call to

df_set_blocks, these blocks are added to the set of blocks.

DF_ANALYZE causes all of the defined problems to be (re)solved.  It

does not cause blocks to be (re)scanned at the rtl level unless no

prior call is made to df_rescan_blocks.  When DF_ANALYZE is completes,

the IN and OUT sets for each basic block contain the computer

information.  The DF_*_BB_INFO macros can be used to access these

bitvectors.

DF_DUMP can then be called to dump the information produce to some

file.

DF_FINISH causes all of the datastructures to be cleaned up and freed.

The df_instance is also freed and its pointer should be NULLed.

Scanning produces a `struct df_ref' data structure (ref) is allocated

for every register reference (def or use) and this records the insn

and bb the ref is found within.  The refs are linked together in

chains of uses and defs for each insn and for each register.  Each ref

also has a chain field that links all the use refs for a def or all

the def refs for a use.  This is used to create use-def or def-use

chains.

Different optimizations have different needs.  Ultimately, only

register allocation and schedulers should be using the bitmaps

produced for the live register and uninitialized register problems.

The rest of the backend should be upgraded to using and maintaining

the linked information such as def use or use def chains.

PHILOSOPHY:

While incremental bitmaps are not worthwhile to maintain, incremental

chains may be perfectly reasonable.  The fastest way to build chains

from scratch or after significant modifications is to build reaching

definitions (RD) and build the chains from this.

However, general algorithms for maintaining use-def or def-use chains

are not practical.  The amount of work to recompute the chain any

chain after an arbitrary change is large.  However, with a modest

amount of work it is generally possible to have the application that

uses the chains keep them up to date.  The high level knowledge of

what is really happening is essential to crafting efficient

incremental algorithms.

As for the bit vector problems, there is no interface to give a set of

blocks over with to resolve the iteration.  In general, restarting a

dataflow iteration is difficult and expensive.  Again, the best way to

keep the dataflow information up to data (if this is really what is

needed) it to formulate a problem specific solution.

There are fine grained calls for creating and deleting references from

instructions in df-scan.c.  However, these are not currently connected

to the engine that resolves the dataflow equations.

DATA STRUCTURES:

The basic object is a DF_REF (reference) and this may either be a

DEF (definition) or a USE of a register.

These are linked into a variety of lists; namely reg-def, reg-use,

insn-def, insn-use, def-use, and use-def lists.  For example, the

reg-def lists contain all the locations that define a given register

while the insn-use lists contain all the locations that use a

register.

Note that the reg-def and reg-use chains are generally short for

pseudos and long for the hard registers.

ACCESSING REFS:

There are 4 ways to obtain access to refs:

1) References are divided into two categories, REAL and ARTIFICIAL.

   REAL refs are associated with instructions.  They are linked into

   either in the insn's defs list (accessed by the DF_INSN_DEFS or

   DF_INSN_UID_DEFS macros) or the insn's uses list (accessed by the

   DF_INSN_USES or DF_INSN_UID_USES macros).  These macros produce a

   ref (or NULL), the rest of the list can be obtained by traversal of

   the NEXT_REF field (accessed by the DF_REF_NEXT_REF macro.)  There

   is no significance to the ordering of the uses or refs in an

   instruction.

   ARTIFICIAL refs are associated with basic blocks.  The heads of

   these lists can be accessed by calling get_artificial_defs or

   get_artificial_uses for the particular basic block.  Artificial

   defs and uses are only there if DF_HARD_REGS was specified when the

   df instance was created.

   Artificial defs and uses occur both at the beginning and ends of blocks.

     For blocks that area at the destination of eh edges, the

     artificial uses and defs occur at the beginning.  The defs relate

     to the registers specified in EH_RETURN_DATA_REGNO and the uses

     relate to the registers specified in ED_USES.  Logically these

     defs and uses should really occur along the eh edge, but there is

     no convenient way to do this.  Artificial edges that occur at the

     beginning of the block have the DF_REF_AT_TOP flag set.

     Artificial uses occur at the end of all blocks.  These arise from

     the hard registers that are always live, such as the stack

     register and are put there to keep the code from forgetting about

     them.

     Artificial defs occur at the end of the entry block.  These arise

     from registers that are live at entry to the function.

2) All of the uses and defs associated with each pseudo or hard

   register are linked in a bidirectional chain.  These are called

   reg-use or reg_def chains.

   The first use (or def) for a register can be obtained using the

   DF_REG_USE_GET macro (or DF_REG_DEF_GET macro).  Subsequent uses

   for the same regno can be obtained by following the next_reg field

   of the ref.

   In previous versions of this code, these chains were ordered.  It

   has not been practical to continue this practice.

3) If def-use or use-def chains are built, these can be traversed to

   get to other refs.

4) An array of all of the uses (and an array of all of the defs) can

   be built.  These arrays are indexed by the value in the id

   structure.  These arrays are only lazily kept up to date, and that

   process can be expensive.  To have these arrays built, call

   df_reorganize_refs.   Note that the values in the id field of a ref

   may change across calls to df_analyze or df_reorganize refs.

   If the only use of this array is to find all of the refs, it is

   better to traverse all of the registers and then traverse all of

   reg-use or reg-def chains.

NOTES:

Embedded addressing side-effects, such as POST_INC or PRE_INC, generate

both a use and a def.  These are both marked read/write to show that they

are dependent. For example, (set (reg 40) (mem (post_inc (reg 42))))

will generate a use of reg 42 followed by a def of reg 42 (both marked

read/write).  Similarly, (set (reg 40) (mem (pre_dec (reg 41))))

generates a use of reg 41 then a def of reg 41 (both marked read/write),

even though reg 41 is decremented before it is used for the memory

address in this second example.

A set to a REG inside a ZERO_EXTRACT, or a set to a non-paradoxical SUBREG

for which the number of word_mode units covered by the outer mode is

smaller than that covered by the inner mode, invokes a read-modify-write.

operation.  We generate both a use and a def and again mark them

read/write.

Paradoxical subreg writes do not leave a trace of the old content, so they

are write-only operations.

*/

#include "config.h"

#include "system.h"

#include "coretypes.h"

#include "tm.h"

#include "rtl.h"

#include "tm_p.h"

#include "insn-config.h"

#include "recog.h"

#include "function.h"

#include "regs.h"

#include "output.h"

#include "alloc-pool.h"

#include "flags.h"

#include "hard-reg-set.h"

#include "basic-block.h"

#include "sbitmap.h"

#include "bitmap.h"

#include "timevar.h"

#include "df.h"

#include "tree-pass.h"

static struct df *ddf = NULL;

struct df *shared_df = NULL;

static void *df_get_bb_info (struct dataflow *, unsigned int);

static void df_set_bb_info (struct dataflow *, unsigned int, void *);

/*----------------------------------------------------------------------------

  Functions to create, destroy and manipulate an instance of df.

----------------------------------------------------------------------------*/

/* Initialize dataflow analysis and allocate and initialize dataflow

   memory.  */

struct df *

df_init (int flags)

{

  struct df *df = XCNEW (struct df);

  /* This is executed once per compilation to initialize platform

     specific data structures. */

  df_hard_reg_init ();

  /* All df instance must define the scanning problem.  */

  df_scan_add_problem (df, flags);

  ddf = df;

  return df;

}

/* Add PROBLEM to the DF instance.  */

struct dataflow *

df_add_problem (struct df *df, struct df_problem *problem, int flags)

{

  struct dataflow *dflow;

  /* First try to add the dependent problem. */

  if (problem->dependent_problem_fun)

    (problem->dependent_problem_fun) (df, 0);

  /* Check to see if this problem has already been defined.  If it

     has, just return that instance, if not, add it to the end of the

     vector.  */

  dflow = df->problems_by_index[problem->id];

  if (dflow)

    return dflow;

  /* Make a new one and add it to the end.  */

  dflow = XCNEW (struct dataflow);

  dflow->flags = flags;

  dflow->df = df;

  dflow->problem = problem;

  df->problems_in_order[df->num_problems_defined++] = dflow;

  df->problems_by_index[dflow->problem->id] = dflow;

  return dflow;

}

/* Set the MASK flags in the DFLOW problem.  The old flags are

   returned.  If a flag is not allowed to be changed this will fail if

   checking is enabled.  */

int

df_set_flags (struct dataflow *dflow, int mask)

{

  int old_flags = dflow->flags;

  gcc_assert (!(mask & (~dflow->problem->changeable_flags)));

  dflow->flags |= mask;

  return old_flags;

}

/* Clear the MASK flags in the DFLOW problem.  The old flags are

   returned.  If a flag is not allowed to be changed this will fail if

   checking is enabled.  */

int

df_clear_flags (struct dataflow *dflow, int mask)

{

  int old_flags = dflow->flags;

  gcc_assert (!(mask & (~dflow->problem->changeable_flags)));

  dflow->flags &= !mask;

  return old_flags;

}

/* Set the blocks that are to be considered for analysis.  If this is

   not called or is called with null, the entire function in

   analyzed.  */

void

df_set_blocks (struct df *df, bitmap blocks)

{

  if (blocks)

    {

      if (df->blocks_to_analyze)

        {

          int p;

          bitmap diff = BITMAP_ALLOC (NULL);

          bitmap_and_compl (diff, df->blocks_to_analyze, blocks);

          for (p = df->num_problems_defined - 1; p >= 0 ;p--)

            {

              struct dataflow *dflow = df->problems_in_order[p];

              if (dflow->problem->reset_fun)

                dflow->problem->reset_fun (dflow, df->blocks_to_analyze);

              else if (dflow->problem->free_bb_fun)

                {

                  bitmap_iterator bi;

                  unsigned int bb_index;

                  EXECUTE_IF_SET_IN_BITMAP (diff, 0, bb_index, bi)

                    {

                      basic_block bb = BASIC_BLOCK (bb_index);

                      if (bb)

                        {

                          dflow->problem->free_bb_fun

                            (dflow, bb, df_get_bb_info (dflow, bb_index));

                          df_set_bb_info (dflow, bb_index, NULL);

                        }

                    }

                }

            }

          BITMAP_FREE (diff);

        }

      else

        {

          /* If we have not actually run scanning before, do not try

             to clear anything.  */

          struct dataflow *scan_dflow = df->problems_by_index [DF_SCAN];

          if (scan_dflow->problem_data)

            {

              bitmap blocks_to_reset = NULL;

              int p;

              for (p = df->num_problems_defined - 1; p >= 0 ;p--)

                {

                  struct dataflow *dflow = df->problems_in_order[p];

                  if (dflow->problem->reset_fun)

                    {

                      if (!blocks_to_reset)

                        {

                          basic_block bb;

                          blocks_to_reset = BITMAP_ALLOC (NULL);

                          FOR_ALL_BB(bb)

                            {

                              bitmap_set_bit (blocks_to_reset, bb->index);

                            }

                        }

                      dflow->problem->reset_fun (dflow, blocks_to_reset);

                    }

                }

              if (blocks_to_reset)

                BITMAP_FREE (blocks_to_reset);

            }

          df->blocks_to_analyze = BITMAP_ALLOC (NULL);

        }

      bitmap_copy (df->blocks_to_analyze, blocks);

    }

  else

    {

      if (df->blocks_to_analyze)

        {

          BITMAP_FREE (df->blocks_to_analyze);

          df->blocks_to_analyze = NULL;

        }

    }

}

/* Free all of the per basic block dataflow from all of the problems.

   This is typically called before a basic block is deleted and the

   problem will be reanalyzed.  */

void

df_delete_basic_block (struct df *df, int bb_index)

{

  basic_block bb = BASIC_BLOCK (bb_index);

  int i;

  for (i = 0; i < df->num_problems_defined; i++)

    {

      struct dataflow *dflow = df->problems_in_order[i];

      if (dflow->problem->free_bb_fun)

        dflow->problem->free_bb_fun

          (dflow, bb, df_get_bb_info (dflow, bb_index));

    }

}

/* Free all the dataflow info and the DF structure.  This should be

   called from the df_finish macro which also NULLs the parm.  */

void

df_finish1 (struct df *df)

{

  int i;

  for (i = 0; i < df->num_problems_defined; i++)

    df->problems_in_order[i]->problem->free_fun (df->problems_in_order[i]);

  free (df);

}

/*----------------------------------------------------------------------------

   The general data flow analysis engine.

----------------------------------------------------------------------------*/

/* Hybrid search algorithm from "Implementation Techniques for

   Efficient Data-Flow Analysis of Large Programs".  */

static void

df_hybrid_search_forward (basic_block bb,

                          struct dataflow *dataflow,

                          bool single_pass)

{

  int result_changed;

  int i = bb->index;

  edge e;

  edge_iterator ei;

  SET_BIT (dataflow->visited, bb->index);

  gcc_assert (TEST_BIT (dataflow->pending, bb->index));

  RESET_BIT (dataflow->pending, i);

  /*  Calculate <conf_op> of predecessor_outs.  */

  if (EDGE_COUNT (bb->preds) > 0)

    FOR_EACH_EDGE (e, ei, bb->preds)

      {

        if (!TEST_BIT (dataflow->considered, e->src->index))

          continue;

        dataflow->problem->con_fun_n (dataflow, e);

      }

  else if (dataflow->problem->con_fun_0)

    dataflow->problem->con_fun_0 (dataflow, bb);

  result_changed = dataflow->problem->trans_fun (dataflow, i);

  if (!result_changed || single_pass)

    return;

  FOR_EACH_EDGE (e, ei, bb->succs)

    {

      if (e->dest->index == i)

        continue;

      if (!TEST_BIT (dataflow->considered, e->dest->index))

        continue;

      SET_BIT (dataflow->pending, e->dest->index);

    }

  FOR_EACH_EDGE (e, ei, bb->succs)

    {

      if (e->dest->index == i)

        continue;

      if (!TEST_BIT (dataflow->considered, e->dest->index))

        continue;

      if (!TEST_BIT (dataflow->visited, e->dest->index))

        df_hybrid_search_forward (e->dest, dataflow, single_pass);

    }

}

static void

df_hybrid_search_backward (basic_block bb,

                           struct dataflow *dataflow,

                           bool single_pass)

{

  int result_changed;

  int i = bb->index;

  edge e;

  edge_iterator ei;

  SET_BIT (dataflow->visited, bb->index);

  gcc_assert (TEST_BIT (dataflow->pending, bb->index));

  RESET_BIT (dataflow->pending, i);

  /*  Calculate <conf_op> of predecessor_outs.  */

  if (EDGE_COUNT (bb->succs) > 0)

    FOR_EACH_EDGE (e, ei, bb->succs)

      {

        if (!TEST_BIT (dataflow->considered, e->dest->index))

          continue;

        dataflow->problem->con_fun_n (dataflow, e);

      }

  else if (dataflow->problem->con_fun_0)

    dataflow->problem->con_fun_0 (dataflow, bb);

  result_changed = dataflow->problem->trans_fun (dataflow, i);

  if (!result_changed || single_pass)

    return;

  FOR_EACH_EDGE (e, ei, bb->preds)

    {

      if (e->src->index == i)

        continue;

      if (!TEST_BIT (dataflow->considered, e->src->index))

        continue;

      SET_BIT (dataflow->pending, e->src->index);

    }

  FOR_EACH_EDGE (e, ei, bb->preds)

    {

      if (e->src->index == i)

        continue;

      if (!TEST_BIT (dataflow->considered, e->src->index))

        continue;

      if (!TEST_BIT (dataflow->visited, e->src->index))

        df_hybrid_search_backward (e->src, dataflow, single_pass);

    }

}

/* This function will perform iterative bitvector dataflow described

   by DATAFLOW, producing the in and out sets.  Only the part of the

   cfg induced by blocks in DATAFLOW->order is taken into account.

   SINGLE_PASS is true if you just want to make one pass over the

   blocks.  */

void

df_iterative_dataflow (struct dataflow *dataflow,

                       bitmap blocks_to_consider, bitmap blocks_to_init,

                       int *blocks_in_postorder, int n_blocks,

                       bool single_pass)

{

  unsigned int idx;

  int i;

  sbitmap visited = sbitmap_alloc (last_basic_block);

  sbitmap pending = sbitmap_alloc (last_basic_block);

  sbitmap considered = sbitmap_alloc (last_basic_block);

  bitmap_iterator bi;

  dataflow->visited = visited;

  dataflow->pending = pending;

  dataflow->considered = considered;

  sbitmap_zero (visited);

  sbitmap_zero (pending);

  sbitmap_zero (considered);

  gcc_assert (dataflow->problem->dir);

  EXECUTE_IF_SET_IN_BITMAP (blocks_to_consider, 0, idx, bi)

    {

      SET_BIT (considered, idx);

    }

  for (i = 0; i < n_blocks; i++)

    {

      idx = blocks_in_postorder[i];

      SET_BIT (pending, idx);

    };

  dataflow->problem->init_fun (dataflow, blocks_to_init);

  while (1)

    {

      /* For forward problems, you want to pass in reverse postorder

         and for backward problems you want postorder.  This has been

         shown to be as good as you can do by several people, the

         first being Mathew Hecht in his phd dissertation.

         The nodes are passed into this function in postorder.  */

      if (dataflow->problem->dir == DF_FORWARD)

        {

          for (i = n_blocks - 1 ; i >= 0 ; i--)

            {

              idx = blocks_in_postorder[i];

              if (TEST_BIT (pending, idx) && !TEST_BIT (visited, idx))

                df_hybrid_search_forward (BASIC_BLOCK (idx), dataflow, single_pass);

            }

        }

      else

        {

          for (i = 0; i < n_blocks; i++)

            {

              idx = blocks_in_postorder[i];

              if (TEST_BIT (pending, idx) && !TEST_BIT (visited, idx))

                df_hybrid_search_backward (BASIC_BLOCK (idx), dataflow, single_pass);

            }

        }

      if (sbitmap_first_set_bit (pending) == -1)

        break;

      sbitmap_zero (visited);

    }

  sbitmap_free (pending);

  sbitmap_free (visited);

  sbitmap_free (considered);

}

/* Remove the entries not in BLOCKS from the LIST of length LEN, preserving

   the order of the remaining entries.  Returns the length of the resulting

   list.  */

static unsigned

df_prune_to_subcfg (int list[], unsigned len, bitmap blocks)

{

  unsigned act, last;

  for (act = 0, last = 0; act < len; act++)

    if (bitmap_bit_p (blocks, list[act]))

      list[last++] = list[act];

  return last;

}

/* Execute dataflow analysis on a single dataflow problem.

   There are three sets of blocks passed in:

   BLOCKS_TO_CONSIDER are the blocks whose solution can either be

   examined or will be computed.  For calls from DF_ANALYZE, this is

   the set of blocks that has been passed to DF_SET_BLOCKS.  For calls

   from DF_ANALYZE_SIMPLE_CHANGE_SOME_BLOCKS, this is the set of

   blocks in the fringe (the set of blocks passed in plus the set of

   immed preds and succs of those blocks).

   BLOCKS_TO_INIT are the blocks whose solution will be changed by

   this iteration.  For calls from DF_ANALYZE, this is the set of

   blocks that has been passed to DF_SET_BLOCKS.  For calls from

   DF_ANALYZE_SIMPLE_CHANGE_SOME_BLOCKS, this is the set of blocks

   passed in.

   BLOCKS_TO_SCAN are the set of blocks that need to be rescanned.

   For calls from DF_ANALYZE, this is the accumulated set of blocks

   that has been passed to DF_RESCAN_BLOCKS since the last call to

   DF_ANALYZE.  For calls from DF_ANALYZE_SIMPLE_CHANGE_SOME_BLOCKS,

   this is the set of blocks passed in.

                   blocks_to_consider    blocks_to_init    blocks_to_scan

   full redo       all                   all               all

   partial redo    all                   all               sub

   small fixup     fringe                sub               sub

*/

void

df_analyze_problem (struct dataflow *dflow,

                    bitmap blocks_to_consider,

                    bitmap blocks_to_init,

                    bitmap blocks_to_scan,

                    int *postorder, int n_blocks, bool single_pass)

{

  /* (Re)Allocate the datastructures necessary to solve the problem.  */

  if (dflow->problem->alloc_fun)

    dflow->problem->alloc_fun (dflow, blocks_to_scan, blocks_to_init);

  /* Set up the problem and compute the local information.  This

     function is passed both the blocks_to_consider and the

     blocks_to_scan because the RD and RU problems require the entire

     function to be rescanned if they are going to be updated.  */

  if (dflow->problem->local_compute_fun)

    dflow->problem->local_compute_fun (dflow, blocks_to_consider, blocks_to_scan);

  /* Solve the equations.  */

  if (dflow->problem->dataflow_fun)

    dflow->problem->dataflow_fun (dflow, blocks_to_consider, blocks_to_init,