Subversion Repositories openrisc

/* Detection of Static Control Parts (SCoP) for Graphite.

   Copyright (C) 2009, 2010 Free Software Foundation, Inc.

   Contributed by Sebastian Pop <sebastian.pop@amd.com> and

   Tobias Grosser <grosser@fim.uni-passau.de>.

This file is part of GCC.

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

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

the Free Software Foundation; either version 3, or (at your option)

any later version.

GCC is distributed in the hope that it will be useful,

but WITHOUT ANY WARRANTY; without even the implied warranty of

MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

GNU General Public License for more details.

You should have received a copy of the GNU General Public License

along with GCC; see the file COPYING3.  If not see

<http://www.gnu.org/licenses/>.  */

#include "config.h"

#include "system.h"

#include "coretypes.h"

#include "tree-flow.h"

#include "cfgloop.h"

#include "tree-chrec.h"

#include "tree-data-ref.h"

#include "tree-scalar-evolution.h"

#include "tree-pass.h"

#include "sese.h"

#ifdef HAVE_cloog

#include "ppl_c.h"

#include "graphite-ppl.h"

#include "graphite-poly.h"

#include "graphite-scop-detection.h"

/* Forward declarations.  */

static void make_close_phi_nodes_unique (basic_block);

/* The type of the analyzed basic block.  */

typedef enum gbb_type {

  GBB_UNKNOWN,

  GBB_LOOP_SING_EXIT_HEADER,

  GBB_LOOP_MULT_EXIT_HEADER,

  GBB_LOOP_EXIT,

  GBB_COND_HEADER,

  GBB_SIMPLE,

  GBB_LAST

} gbb_type;

/* Detect the type of BB.  Loop headers are only marked, if they are

   new.  This means their loop_father is different to LAST_LOOP.

   Otherwise they are treated like any other bb and their type can be

   any other type.  */

static gbb_type

get_bb_type (basic_block bb, struct loop *last_loop)

{

  VEC (basic_block, heap) *dom;

  int nb_dom, nb_suc;

  struct loop *loop = bb->loop_father;

  /* Check, if we entry into a new loop. */

  if (loop != last_loop)

    {

      if (single_exit (loop) != NULL)

        return GBB_LOOP_SING_EXIT_HEADER;

      else if (loop->num != 0)

        return GBB_LOOP_MULT_EXIT_HEADER;

      else

        return GBB_COND_HEADER;

    }

  dom = get_dominated_by (CDI_DOMINATORS, bb);

  nb_dom = VEC_length (basic_block, dom);

  VEC_free (basic_block, heap, dom);

  if (nb_dom == 0)

    return GBB_LAST;

  nb_suc = VEC_length (edge, bb->succs);

  if (nb_dom == 1 && nb_suc == 1)

    return GBB_SIMPLE;

  return GBB_COND_HEADER;

}

/* A SCoP detection region, defined using bbs as borders.

   All control flow touching this region, comes in passing basic_block

   ENTRY and leaves passing basic_block EXIT.  By using bbs instead of

   edges for the borders we are able to represent also regions that do

   not have a single entry or exit edge.

   But as they have a single entry basic_block and a single exit

   basic_block, we are able to generate for every sd_region a single

   entry and exit edge.

   1   2

    \ /

     3  <- entry

     |

     4

    / \                 This region contains: {3, 4, 5, 6, 7, 8}

   5   6

   |   |

   7   8

    \ /

     9  <- exit  */

typedef struct sd_region_p

{

  /* The entry bb dominates all bbs in the sd_region.  It is part of

     the region.  */

  basic_block entry;

  /* The exit bb postdominates all bbs in the sd_region, but is not

     part of the region.  */

  basic_block exit;

} sd_region;

DEF_VEC_O(sd_region);

DEF_VEC_ALLOC_O(sd_region, heap);

/* Moves the scops from SOURCE to TARGET and clean up SOURCE.  */

static void

move_sd_regions (VEC (sd_region, heap) **source,

                 VEC (sd_region, heap) **target)

{

  sd_region *s;

  int i;

  FOR_EACH_VEC_ELT (sd_region, *source, i, s)

    VEC_safe_push (sd_region, heap, *target, s);

  VEC_free (sd_region, heap, *source);

}

/* Something like "n * m" is not allowed.  */

static bool

graphite_can_represent_init (tree e)

{

  switch (TREE_CODE (e))

    {

    case POLYNOMIAL_CHREC:

      return graphite_can_represent_init (CHREC_LEFT (e))

        && graphite_can_represent_init (CHREC_RIGHT (e));

    case MULT_EXPR:

      if (chrec_contains_symbols (TREE_OPERAND (e, 0)))

        return graphite_can_represent_init (TREE_OPERAND (e, 0))

          && host_integerp (TREE_OPERAND (e, 1), 0);

      else

        return graphite_can_represent_init (TREE_OPERAND (e, 1))

          && host_integerp (TREE_OPERAND (e, 0), 0);

    case PLUS_EXPR:

    case POINTER_PLUS_EXPR:

    case MINUS_EXPR:

      return graphite_can_represent_init (TREE_OPERAND (e, 0))

        && graphite_can_represent_init (TREE_OPERAND (e, 1));

    case NEGATE_EXPR:

    case BIT_NOT_EXPR:

    CASE_CONVERT:

    case NON_LVALUE_EXPR:

      return graphite_can_represent_init (TREE_OPERAND (e, 0));

   default:

     break;

    }

  return true;

}

/* Return true when SCEV can be represented in the polyhedral model.

   An expression can be represented, if it can be expressed as an

   affine expression.  For loops (i, j) and parameters (m, n) all

   affine expressions are of the form:

   x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z

   1 i + 20 j + (-2) m + 25

   Something like "i * n" or "n * m" is not allowed.  */

static bool

graphite_can_represent_scev (tree scev)

{

  if (chrec_contains_undetermined (scev))

    return false;

  switch (TREE_CODE (scev))

    {

    case PLUS_EXPR:

    case MINUS_EXPR:

      return graphite_can_represent_scev (TREE_OPERAND (scev, 0))

        && graphite_can_represent_scev (TREE_OPERAND (scev, 1));

    case MULT_EXPR:

      return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 0)))

        && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 1)))

        && !(chrec_contains_symbols (TREE_OPERAND (scev, 0))

             && chrec_contains_symbols (TREE_OPERAND (scev, 1)))

        && graphite_can_represent_init (scev)

        && graphite_can_represent_scev (TREE_OPERAND (scev, 0))

        && graphite_can_represent_scev (TREE_OPERAND (scev, 1));

    case POLYNOMIAL_CHREC:

      /* Check for constant strides.  With a non constant stride of

         'n' we would have a value of 'iv * n'.  Also check that the

         initial value can represented: for example 'n * m' cannot be

         represented.  */

      if (!evolution_function_right_is_integer_cst (scev)

          || !graphite_can_represent_init (scev))

        return false;

    default:

      break;

    }

  /* Only affine functions can be represented.  */

  if (!scev_is_linear_expression (scev))

    return false;

  return true;

}

/* Return true when EXPR can be represented in the polyhedral model.

   This means an expression can be represented, if it is linear with

   respect to the loops and the strides are non parametric.

   LOOP is the place where the expr will be evaluated.  SCOP_ENTRY defines the

   entry of the region we analyse.  */

static bool

graphite_can_represent_expr (basic_block scop_entry, loop_p loop,

                             tree expr)

{

  tree scev = analyze_scalar_evolution (loop, expr);

  scev = instantiate_scev (scop_entry, loop, scev);

  return graphite_can_represent_scev (scev);

}

/* Return true if the data references of STMT can be represented by

   Graphite.  */

static bool

stmt_has_simple_data_refs_p (loop_p outermost_loop ATTRIBUTE_UNUSED,

                             gimple stmt)

{

  data_reference_p dr;

  unsigned i;

  int j;

  bool res = true;

  VEC (data_reference_p, heap) *drs = NULL;

  loop_p outer;

  for (outer = loop_containing_stmt (stmt); outer; outer = loop_outer (outer))

    {

      graphite_find_data_references_in_stmt (outer,

                                             loop_containing_stmt (stmt),

                                             stmt, &drs);

      FOR_EACH_VEC_ELT (data_reference_p, drs, j, dr)

        for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)

          if (!graphite_can_represent_scev (DR_ACCESS_FN (dr, i)))

            {

              res = false;

              goto done;

            }

      free_data_refs (drs);

      drs = NULL;

    }

 done:

  free_data_refs (drs);

  return res;

}

/* Return true only when STMT is simple enough for being handled by

   Graphite.  This depends on SCOP_ENTRY, as the parameters are

   initialized relatively to this basic block, the linear functions

   are initialized to OUTERMOST_LOOP and BB is the place where we try

   to evaluate the STMT.  */

static bool

stmt_simple_for_scop_p (basic_block scop_entry, loop_p outermost_loop,

                        gimple stmt, basic_block bb)

{

  loop_p loop = bb->loop_father;

  gcc_assert (scop_entry);

  /* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects.

     Calls have side-effects, except those to const or pure

     functions.  */

  if (gimple_has_volatile_ops (stmt)

      || (gimple_code (stmt) == GIMPLE_CALL

          && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))

      || (gimple_code (stmt) == GIMPLE_ASM))

    return false;

  if (is_gimple_debug (stmt))

    return true;

  if (!stmt_has_simple_data_refs_p (outermost_loop, stmt))

    return false;

  switch (gimple_code (stmt))

    {

    case GIMPLE_RETURN:

    case GIMPLE_LABEL:

      return true;

    case GIMPLE_COND:

      {

        tree op;

        ssa_op_iter op_iter;

        enum tree_code code = gimple_cond_code (stmt);

        /* We can handle all binary comparisons.  Inequalities are

           also supported as they can be represented with union of

           polyhedra.  */

        if (!(code == LT_EXPR

              || code == GT_EXPR

              || code == LE_EXPR

              || code == GE_EXPR

              || code == EQ_EXPR

              || code == NE_EXPR))

          return false;

        FOR_EACH_SSA_TREE_OPERAND (op, stmt, op_iter, SSA_OP_ALL_USES)

          if (!graphite_can_represent_expr (scop_entry, loop, op)

              /* We can not handle REAL_TYPE. Failed for pr39260.  */

              || TREE_CODE (TREE_TYPE (op)) == REAL_TYPE)

            return false;

        return true;

      }

    case GIMPLE_ASSIGN:

    case GIMPLE_CALL:

      return true;

    default:

      /* These nodes cut a new scope.  */

      return false;

    }

  return false;

}

/* Returns the statement of BB that contains a harmful operation: that

   can be a function call with side effects, the induction variables

   are not linear with respect to SCOP_ENTRY, etc.  The current open

   scop should end before this statement.  The evaluation is limited using

   OUTERMOST_LOOP as outermost loop that may change.  */

static gimple

harmful_stmt_in_bb (basic_block scop_entry, loop_p outer_loop, basic_block bb)

{

  gimple_stmt_iterator gsi;

  for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))

    if (!stmt_simple_for_scop_p (scop_entry, outer_loop, gsi_stmt (gsi), bb))

      return gsi_stmt (gsi);

  return NULL;

}

/* Return true if LOOP can be represented in the polyhedral

   representation.  This is evaluated taking SCOP_ENTRY and

   OUTERMOST_LOOP in mind.  */

static bool

graphite_can_represent_loop (basic_block scop_entry, loop_p loop)

{

  tree niter;

  struct tree_niter_desc niter_desc;

  /* FIXME: For the moment, graphite cannot be used on loops that

     iterate using induction variables that wrap.  */

  return number_of_iterations_exit (loop, single_exit (loop), &niter_desc, false)

    && niter_desc.control.no_overflow

    && (niter = number_of_latch_executions (loop))

    && !chrec_contains_undetermined (niter)

    && graphite_can_represent_expr (scop_entry, loop, niter);

}

/* Store information needed by scopdet_* functions.  */

struct scopdet_info

{

  /* Exit of the open scop would stop if the current BB is harmful.  */

  basic_block exit;

  /* Where the next scop would start if the current BB is harmful.  */

  basic_block next;

  /* The bb or one of its children contains open loop exits.  That means

     loop exit nodes that are not surrounded by a loop dominated by bb.  */

  bool exits;

  /* The bb or one of its children contains only structures we can handle.  */

  bool difficult;

};

static struct scopdet_info build_scops_1 (basic_block, loop_p,

                                          VEC (sd_region, heap) **, loop_p);

/* Calculates BB infos. If bb is difficult we add valid SCoPs dominated by BB

   to SCOPS.  TYPE is the gbb_type of BB.  */

static struct scopdet_info

scopdet_basic_block_info (basic_block bb, loop_p outermost_loop,

                          VEC (sd_region, heap) **scops, gbb_type type)

{

  loop_p loop = bb->loop_father;

  struct scopdet_info result;

  gimple stmt;

  /* XXX: ENTRY_BLOCK_PTR could be optimized in later steps.  */

  basic_block entry_block = ENTRY_BLOCK_PTR;

  stmt = harmful_stmt_in_bb (entry_block, outermost_loop, bb);

  result.difficult = (stmt != NULL);

  result.exit = NULL;

  switch (type)

    {

    case GBB_LAST:

      result.next = NULL;

      result.exits = false;

      /* Mark bbs terminating a SESE region difficult, if they start

         a condition.  */

      if (!single_succ_p (bb))

        result.difficult = true;

      else

        result.exit = single_succ (bb);

      break;

    case GBB_SIMPLE:

      result.next = single_succ (bb);

      result.exits = false;

      result.exit = single_succ (bb);

      break;

    case GBB_LOOP_SING_EXIT_HEADER:

      {

        VEC (sd_region, heap) *regions = VEC_alloc (sd_region, heap, 3);

        struct scopdet_info sinfo;

        edge exit_e = single_exit (loop);

        sinfo = build_scops_1 (bb, outermost_loop, &regions, loop);

        if (!graphite_can_represent_loop (entry_block, loop))

          result.difficult = true;

        result.difficult |= sinfo.difficult;

        /* Try again with another loop level.  */

        if (result.difficult

            && loop_depth (outermost_loop) + 1 == loop_depth (loop))

          {

            outermost_loop = loop;

            VEC_free (sd_region, heap, regions);

            regions = VEC_alloc (sd_region, heap, 3);

            sinfo = scopdet_basic_block_info (bb, outermost_loop, scops, type);

            result = sinfo;

            result.difficult = true;

            if (sinfo.difficult)

              move_sd_regions (&regions, scops);

            else

              {

                sd_region open_scop;

                open_scop.entry = bb;

                open_scop.exit = exit_e->dest;

                VEC_safe_push (sd_region, heap, *scops, &open_scop);

                VEC_free (sd_region, heap, regions);

              }

          }

        else

          {

            result.exit = exit_e->dest;

            result.next = exit_e->dest;

            /* If we do not dominate result.next, remove it.  It's either

               the EXIT_BLOCK_PTR, or another bb dominates it and will

               call the scop detection for this bb.  */

            if (!dominated_by_p (CDI_DOMINATORS, result.next, bb))

              result.next = NULL;

            if (exit_e->src->loop_father != loop)

              result.next = NULL;

            result.exits = false;

            if (result.difficult)

              move_sd_regions (&regions, scops);

            else

              VEC_free (sd_region, heap, regions);

          }

        break;

      }

    case GBB_LOOP_MULT_EXIT_HEADER:

      {

        /* XXX: For now we just do not join loops with multiple exits.  If the

           exits lead to the same bb it may be possible to join the loop.  */

        VEC (sd_region, heap) *regions = VEC_alloc (sd_region, heap, 3);

        VEC (edge, heap) *exits = get_loop_exit_edges (loop);

        edge e;

        int i;

        build_scops_1 (bb, loop, &regions, loop);

        /* Scan the code dominated by this loop.  This means all bbs, that are

           are dominated by a bb in this loop, but are not part of this loop.

           The easiest case:

             - The loop exit destination is dominated by the exit sources.

           TODO: We miss here the more complex cases:

                  - The exit destinations are dominated by another bb inside

                    the loop.

                  - The loop dominates bbs, that are not exit destinations.  */

        FOR_EACH_VEC_ELT (edge, exits, i, e)

          if (e->src->loop_father == loop

              && dominated_by_p (CDI_DOMINATORS, e->dest, e->src))

            {

              if (loop_outer (outermost_loop))

                outermost_loop = loop_outer (outermost_loop);

              /* Pass loop_outer to recognize e->dest as loop header in

                 build_scops_1.  */

              if (e->dest->loop_father->header == e->dest)

                build_scops_1 (e->dest, outermost_loop, &regions,

                               loop_outer (e->dest->loop_father));

              else

                build_scops_1 (e->dest, outermost_loop, &regions,

                               e->dest->loop_father);

            }

        result.next = NULL;

        result.exit = NULL;

        result.difficult = true;

        result.exits = false;

        move_sd_regions (&regions, scops);

        VEC_free (edge, heap, exits);

        break;

      }

    case GBB_COND_HEADER:

      {

        VEC (sd_region, heap) *regions = VEC_alloc (sd_region, heap, 3);

        struct scopdet_info sinfo;

        VEC (basic_block, heap) *dominated;

        int i;

        basic_block dom_bb;

        basic_block last_exit = NULL;

        edge e;

        result.exits = false;

        /* First check the successors of BB, and check if it is

           possible to join the different branches.  */

        FOR_EACH_VEC_ELT (edge, bb->succs, i, e)

          {

            /* Ignore loop exits.  They will be handled after the loop

               body.  */

            if (loop_exits_to_bb_p (loop, e->dest))

              {

                result.exits = true;

                continue;

              }

            /* Do not follow edges that lead to the end of the

               conditions block.  For example, in

               |   0

               |  /|\

               | 1 2 |

               | | | |

               | 3 4 |

               |  \|/

               |   6

               the edge from 0 => 6.  Only check if all paths lead to

               the same node 6.  */

            if (!single_pred_p (e->dest))

              {

                /* Check, if edge leads directly to the end of this

                   condition.  */

                if (!last_exit)

                  last_exit = e->dest;

                if (e->dest != last_exit)

                  result.difficult = true;

                continue;

              }

            if (!dominated_by_p (CDI_DOMINATORS, e->dest, bb))

              {

                result.difficult = true;

                continue;

              }

            sinfo = build_scops_1 (e->dest, outermost_loop, &regions, loop);

            result.exits |= sinfo.exits;

            result.difficult |= sinfo.difficult;

            /* Checks, if all branches end at the same point.

               If that is true, the condition stays joinable.

               Have a look at the example above.  */

            if (sinfo.exit)

              {

                if (!last_exit)

                  last_exit = sinfo.exit;

                if (sinfo.exit != last_exit)

                  result.difficult = true;

              }

            else

              result.difficult = true;

          }

        if (!last_exit)

          result.difficult = true;

        /* Join the branches of the condition if possible.  */

        if (!result.exits && !result.difficult)

          {

            /* Only return a next pointer if we dominate this pointer.

               Otherwise it will be handled by the bb dominating it.  */

            if (dominated_by_p (CDI_DOMINATORS, last_exit, bb)

                && last_exit != bb)

              result.next = last_exit;

            else

              result.next = NULL;

            result.exit = last_exit;

            VEC_free (sd_region, heap, regions);

            break;

          }

        /* Scan remaining bbs dominated by BB.  */

        dominated = get_dominated_by (CDI_DOMINATORS, bb);

        FOR_EACH_VEC_ELT (basic_block, dominated, i, dom_bb)

          {

            /* Ignore loop exits: they will be handled after the loop body.  */

            if (loop_depth (find_common_loop (loop, dom_bb->loop_father))

                < loop_depth (loop))

              {

                result.exits = true;

                continue;

              }

            /* Ignore the bbs processed above.  */

            if (single_pred_p (dom_bb) && single_pred (dom_bb) == bb)

              continue;

            if (loop_depth (loop) > loop_depth (dom_bb->loop_father))

              sinfo = build_scops_1 (dom_bb, outermost_loop, &regions,

                                     loop_outer (loop));

            else

              sinfo = build_scops_1 (dom_bb, outermost_loop, &regions, loop);

            result.exits |= sinfo.exits;

            result.difficult = true;

            result.exit = NULL;

          }

        VEC_free (basic_block, heap, dominated);

        result.next = NULL;

        move_sd_regions (&regions, scops);

        break;

      }

    default:

      gcc_unreachable ();

    }

  return result;

}

/* Starting from CURRENT we walk the dominance tree and add new sd_regions to

   SCOPS. The analyse if a sd_region can be handled is based on the value

   of OUTERMOST_LOOP. Only loops inside OUTERMOST loops may change.  LOOP

   is the loop in which CURRENT is handled.

   TODO: These functions got a little bit big. They definitely should be cleaned

         up.  */

static struct scopdet_info

build_scops_1 (basic_block current, loop_p outermost_loop,

               VEC (sd_region, heap) **scops, loop_p loop)

{

  bool in_scop = false;

  sd_region open_scop;