Subversion Repositories scarts

/* Post reload partially redundant load elimination

   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 "toplev.h"

#include "rtl.h"

#include "tree.h"

#include "tm_p.h"

#include "regs.h"

#include "hard-reg-set.h"

#include "flags.h"

#include "real.h"

#include "insn-config.h"

#include "recog.h"

#include "basic-block.h"

#include "output.h"

#include "function.h"

#include "expr.h"

#include "except.h"

#include "intl.h"

#include "obstack.h"

#include "hashtab.h"

#include "params.h"

#include "target.h"

#include "timevar.h"

#include "tree-pass.h"

/* The following code implements gcse after reload, the purpose of this

   pass is to cleanup redundant loads generated by reload and other

   optimizations that come after gcse. It searches for simple inter-block

   redundancies and tries to eliminate them by adding moves and loads

   in cold places.

   Perform partially redundant load elimination, try to eliminate redundant

   loads created by the reload pass.  We try to look for full or partial

   redundant loads fed by one or more loads/stores in predecessor BBs,

   and try adding loads to make them fully redundant.  We also check if

   it's worth adding loads to be able to delete the redundant load.

   Algorithm:

   1. Build available expressions hash table:

       For each load/store instruction, if the loaded/stored memory didn't

       change until the end of the basic block add this memory expression to

       the hash table.

   2. Perform Redundancy elimination:

      For each load instruction do the following:

         perform partial redundancy elimination, check if it's worth adding

         loads to make the load fully redundant.  If so add loads and

         register copies and delete the load.

   3. Delete instructions made redundant in step 2.

   Future enhancement:

     If the loaded register is used/defined between load and some store,

     look for some other free register between load and all its stores,

     and replace the load with a copy from this register to the loaded

     register.

*/

/* Keep statistics of this pass.  */

static struct

{

  int moves_inserted;

  int copies_inserted;

  int insns_deleted;

} stats;

/* We need to keep a hash table of expressions.  The table entries are of

   type 'struct expr', and for each expression there is a single linked

   list of occurrences.  */

/* The table itself.  */

static htab_t expr_table;

/* Expression elements in the hash table.  */

struct expr

{

  /* The expression (SET_SRC for expressions, PATTERN for assignments).  */

  rtx expr;

  /* The same hash for this entry.  */

  hashval_t hash;

  /* List of available occurrence in basic blocks in the function.  */

  struct occr *avail_occr;

};

static struct obstack expr_obstack;

/* Occurrence of an expression.

   There is at most one occurrence per basic block.  If a pattern appears

   more than once, the last appearance is used.  */

struct occr

{

  /* Next occurrence of this expression.  */

  struct occr *next;

  /* The insn that computes the expression.  */

  rtx insn;

  /* Nonzero if this [anticipatable] occurrence has been deleted.  */

  char deleted_p;

};

static struct obstack occr_obstack;

/* The following structure holds the information about the occurrences of

   the redundant instructions.  */

struct unoccr

{

  struct unoccr *next;

  edge pred;

  rtx insn;

};

static struct obstack unoccr_obstack;

/* Array where each element is the CUID if the insn that last set the hard

   register with the number of the element, since the start of the current

   basic block.

   This array is used during the building of the hash table (step 1) to

   determine if a reg is killed before the end of a basic block.

   It is also used when eliminating partial redundancies (step 2) to see

   if a reg was modified since the start of a basic block.  */

static int *reg_avail_info;

/* A list of insns that may modify memory within the current basic block.  */

struct modifies_mem

{

  rtx insn;

  struct modifies_mem *next;

};

static struct modifies_mem *modifies_mem_list;

/* The modifies_mem structs also go on an obstack, only this obstack is

   freed each time after completing the analysis or transformations on

   a basic block.  So we allocate a dummy modifies_mem_obstack_bottom

   object on the obstack to keep track of the bottom of the obstack.  */

static struct obstack modifies_mem_obstack;

static struct modifies_mem  *modifies_mem_obstack_bottom;

/* Mapping of insn UIDs to CUIDs.

   CUIDs are like UIDs except they increase monotonically in each basic

   block, have no gaps, and only apply to real insns.  */

static int *uid_cuid;

#define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])

/* Helpers for memory allocation/freeing.  */

static void alloc_mem (void);

static void free_mem (void);

/* Support for hash table construction and transformations.  */

static bool oprs_unchanged_p (rtx, rtx, bool);

static void record_last_reg_set_info (rtx, int);

static void record_last_mem_set_info (rtx);

static void record_last_set_info (rtx, rtx, void *);

static void record_opr_changes (rtx);

static void find_mem_conflicts (rtx, rtx, void *);

static int load_killed_in_block_p (int, rtx, bool);

static void reset_opr_set_tables (void);

/* Hash table support.  */

static hashval_t hash_expr (rtx, int *);

static hashval_t hash_expr_for_htab (const void *);

static int expr_equiv_p (const void *, const void *);

static void insert_expr_in_table (rtx, rtx);

static struct expr *lookup_expr_in_table (rtx);

static int dump_hash_table_entry (void **, void *);

static void dump_hash_table (FILE *);

/* Helpers for eliminate_partially_redundant_load.  */

static bool reg_killed_on_edge (rtx, edge);

static bool reg_used_on_edge (rtx, edge);

static rtx reg_set_between_after_reload_p (rtx, rtx, rtx);

static rtx reg_used_between_after_reload_p (rtx, rtx, rtx);

static rtx get_avail_load_store_reg (rtx);

static bool bb_has_well_behaved_predecessors (basic_block);

static struct occr* get_bb_avail_insn (basic_block, struct occr *);

static void hash_scan_set (rtx);

static void compute_hash_table (void);

/* The work horses of this pass.  */

static void eliminate_partially_redundant_load (basic_block,

                                                rtx,

                                                struct expr *);

static void eliminate_partially_redundant_loads (void);

/* Allocate memory for the CUID mapping array and register/memory

   tracking tables.  */

static void

alloc_mem (void)

{

  int i;

  basic_block bb;

  rtx insn;

  /* Find the largest UID and create a mapping from UIDs to CUIDs.  */

  uid_cuid = xcalloc (get_max_uid () + 1, sizeof (int));

  i = 1;

  FOR_EACH_BB (bb)

    FOR_BB_INSNS (bb, insn)

      {

        if (INSN_P (insn))

          uid_cuid[INSN_UID (insn)] = i++;

        else

          uid_cuid[INSN_UID (insn)] = i;

      }

  /* Allocate the available expressions hash table.  We don't want to

     make the hash table too small, but unnecessarily making it too large

     also doesn't help.  The i/4 is a gcse.c relic, and seems like a

     reasonable choice.  */

  expr_table = htab_create (MAX (i / 4, 13),

                            hash_expr_for_htab, expr_equiv_p, NULL);

  /* We allocate everything on obstacks because we often can roll back

     the whole obstack to some point.  Freeing obstacks is very fast.  */

  gcc_obstack_init (&expr_obstack);

  gcc_obstack_init (&occr_obstack);

  gcc_obstack_init (&unoccr_obstack);

  gcc_obstack_init (&modifies_mem_obstack);

  /* Working array used to track the last set for each register

     in the current block.  */

  reg_avail_info = (int *) xmalloc (FIRST_PSEUDO_REGISTER * sizeof (int));

  /* Put a dummy modifies_mem object on the modifies_mem_obstack, so we

     can roll it back in reset_opr_set_tables.  */

  modifies_mem_obstack_bottom =

    (struct modifies_mem *) obstack_alloc (&modifies_mem_obstack,

                                           sizeof (struct modifies_mem));

}

/* Free memory allocated by alloc_mem.  */

static void

free_mem (void)

{

  free (uid_cuid);

  htab_delete (expr_table);

  obstack_free (&expr_obstack, NULL);

  obstack_free (&occr_obstack, NULL);

  obstack_free (&unoccr_obstack, NULL);

  obstack_free (&modifies_mem_obstack, NULL);

  free (reg_avail_info);

}

/* Hash expression X.

   DO_NOT_RECORD_P is a boolean indicating if a volatile operand is found

   or if the expression contains something we don't want to insert in the

   table.  */

static hashval_t

hash_expr (rtx x, int *do_not_record_p)

{

  *do_not_record_p = 0;

  return hash_rtx (x, GET_MODE (x), do_not_record_p,

                   NULL,  /*have_reg_qty=*/false);

}

/* Callback for hashtab.

   Return the hash value for expression EXP.  We don't actually hash

   here, we just return the cached hash value.  */

static hashval_t

hash_expr_for_htab (const void *expp)

{

  struct expr *exp = (struct expr *) expp;

  return exp->hash;

}

/* Callback for hashtab.

   Return nonzero if exp1 is equivalent to exp2.  */

static int

expr_equiv_p (const void *exp1p, const void *exp2p)

{

  struct expr *exp1 = (struct expr *) exp1p;

  struct expr *exp2 = (struct expr *) exp2p;

  int equiv_p = exp_equiv_p (exp1->expr, exp2->expr, 0, true);

  gcc_assert (!equiv_p || exp1->hash == exp2->hash);

  return equiv_p;

}

/* Insert expression X in INSN in the hash TABLE.

   If it is already present, record it as the last occurrence in INSN's

   basic block.  */

static void

insert_expr_in_table (rtx x, rtx insn)

{

  int do_not_record_p;

  hashval_t hash;

  struct expr *cur_expr, **slot;

  struct occr *avail_occr, *last_occr = NULL;

  hash = hash_expr (x, &do_not_record_p);

  /* Do not insert expression in the table if it contains volatile operands,

     or if hash_expr determines the expression is something we don't want

     to or can't handle.  */

  if (do_not_record_p)

    return;

  /* We anticipate that redundant expressions are rare, so for convenience

     allocate a new hash table element here already and set its fields.

     If we don't do this, we need a hack with a static struct expr.  Anyway,

     obstack_free is really fast and one more obstack_alloc doesn't hurt if

     we're going to see more expressions later on.  */

  cur_expr = (struct expr *) obstack_alloc (&expr_obstack,

                                            sizeof (struct expr));

  cur_expr->expr = x;

  cur_expr->hash = hash;

  cur_expr->avail_occr = NULL;

  slot = (struct expr **) htab_find_slot_with_hash (expr_table, cur_expr,

                                                    hash, INSERT);

  if (! (*slot))

    /* The expression isn't found, so insert it.  */

    *slot = cur_expr;

  else

    {

      /* The expression is already in the table, so roll back the

         obstack and use the existing table entry.  */

      obstack_free (&expr_obstack, cur_expr);

      cur_expr = *slot;

    }

  /* Search for another occurrence in the same basic block.  */

  avail_occr = cur_expr->avail_occr;

  while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))

    {

      /* If an occurrence isn't found, save a pointer to the end of

         the list.  */

      last_occr = avail_occr;

      avail_occr = avail_occr->next;

    }

  if (avail_occr)

    /* Found another instance of the expression in the same basic block.

       Prefer this occurrence to the currently recorded one.  We want

       the last one in the block and the block is scanned from start

       to end.  */

    avail_occr->insn = insn;

  else

    {

      /* First occurrence of this expression in this basic block.  */

      avail_occr = (struct occr *) obstack_alloc (&occr_obstack,

                                                  sizeof (struct occr));

      /* First occurrence of this expression in any block?  */

      if (cur_expr->avail_occr == NULL)

        cur_expr->avail_occr = avail_occr;

      else

        last_occr->next = avail_occr;

      avail_occr->insn = insn;

      avail_occr->next = NULL;

      avail_occr->deleted_p = 0;

    }

}

/* Lookup pattern PAT in the expression hash table.

   The result is a pointer to the table entry, or NULL if not found.  */

static struct expr *

lookup_expr_in_table (rtx pat)

{

  int do_not_record_p;

  struct expr **slot, *tmp_expr;

  hashval_t hash = hash_expr (pat, &do_not_record_p);

  if (do_not_record_p)

    return NULL;

  tmp_expr = (struct expr *) obstack_alloc (&expr_obstack,

                                            sizeof (struct expr));

  tmp_expr->expr = pat;

  tmp_expr->hash = hash;

  tmp_expr->avail_occr = NULL;

  slot = (struct expr **) htab_find_slot_with_hash (expr_table, tmp_expr,

                                                    hash, INSERT);

  obstack_free (&expr_obstack, tmp_expr);

  if (!slot)

    return NULL;

  else

    return (*slot);

}

/* Dump all expressions and occurrences that are currently in the

   expression hash table to FILE.  */

/* This helper is called via htab_traverse.  */

static int

dump_hash_table_entry (void **slot, void *filep)

{

  struct expr *expr = (struct expr *) *slot;

  FILE *file = (FILE *) filep;

  struct occr *occr;

  fprintf (file, "expr: ");

  print_rtl (file, expr->expr);

  fprintf (file,"\nhashcode: %u\n", expr->hash);

  fprintf (file,"list of occurrences:\n");

  occr = expr->avail_occr;

  while (occr)

    {

      rtx insn = occr->insn;

      print_rtl_single (file, insn);

      fprintf (file, "\n");

      occr = occr->next;

    }

  fprintf (file, "\n");

  return 1;

}

static void

dump_hash_table (FILE *file)

{

  fprintf (file, "\n\nexpression hash table\n");

  fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",

           (long) htab_size (expr_table),

           (long) htab_elements (expr_table),

           htab_collisions (expr_table));

  if (htab_elements (expr_table) > 0)

    {

      fprintf (file, "\n\ntable entries:\n");

      htab_traverse (expr_table, dump_hash_table_entry, file);

    }

  fprintf (file, "\n");

}

/* Return nonzero if the operands of expression X are unchanged

   1) from the start of INSN's basic block up to but not including INSN

      if AFTER_INSN is false, or

   2) from INSN to the end of INSN's basic block if AFTER_INSN is true.  */

static bool

oprs_unchanged_p (rtx x, rtx insn, bool after_insn)

{

  int i, j;

  enum rtx_code code;

  const char *fmt;

  if (x == 0)

    return 1;

  code = GET_CODE (x);

  switch (code)

    {

    case REG:

      /* We are called after register allocation.  */

      gcc_assert (REGNO (x) < FIRST_PSEUDO_REGISTER);

      if (after_insn)

        /* If the last CUID setting the insn is less than the CUID of

           INSN, then reg X is not changed in or after INSN.  */

        return reg_avail_info[REGNO (x)] < INSN_CUID (insn);

      else

        /* Reg X is not set before INSN in the current basic block if

           we have not yet recorded the CUID of an insn that touches

           the reg.  */

        return reg_avail_info[REGNO (x)] == 0;

    case MEM:

      if (load_killed_in_block_p (INSN_CUID (insn), x, after_insn))

        return 0;

      else

        return oprs_unchanged_p (XEXP (x, 0), insn, after_insn);

    case PC:

    case CC0: /*FIXME*/

    case CONST:

    case CONST_INT:

    case CONST_DOUBLE:

    case CONST_VECTOR:

    case SYMBOL_REF:

    case LABEL_REF:

    case ADDR_VEC:

    case ADDR_DIFF_VEC:

      return 1;

    case PRE_DEC:

    case PRE_INC:

    case POST_DEC:

    case POST_INC:

    case PRE_MODIFY:

    case POST_MODIFY:

      if (after_insn)

        return 0;

      break;

    default:

      break;

    }

  for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)

    {

      if (fmt[i] == 'e')

        {

          if (! oprs_unchanged_p (XEXP (x, i), insn, after_insn))

            return 0;

        }

      else if (fmt[i] == 'E')

        for (j = 0; j < XVECLEN (x, i); j++)

          if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, after_insn))

            return 0;

    }

  return 1;

}

/* Used for communication between find_mem_conflicts and

   load_killed_in_block_p.  Nonzero if find_mem_conflicts finds a

   conflict between two memory references.

   This is a bit of a hack to work around the limitations of note_stores.  */

static int mems_conflict_p;

/* DEST is the output of an instruction.  If it is a memory reference, and

   possibly conflicts with the load found in DATA, then set mems_conflict_p

   to a nonzero value.  */

static void

find_mem_conflicts (rtx dest, rtx setter ATTRIBUTE_UNUSED,

                    void *data)

{

  rtx mem_op = (rtx) data;

  while (GET_CODE (dest) == SUBREG

         || GET_CODE (dest) == ZERO_EXTRACT

         || GET_CODE (dest) == STRICT_LOW_PART)

    dest = XEXP (dest, 0);

  /* If DEST is not a MEM, then it will not conflict with the load.  Note

     that function calls are assumed to clobber memory, but are handled

     elsewhere.  */

  if (! MEM_P (dest))

    return;

  if (true_dependence (dest, GET_MODE (dest), mem_op,

                       rtx_addr_varies_p))

    mems_conflict_p = 1;

}

/* Return nonzero if the expression in X (a memory reference) is killed

   in the current basic block before (if AFTER_INSN is false) or after

   (if AFTER_INSN is true) the insn with the CUID in UID_LIMIT.

   This function assumes that the modifies_mem table is flushed when

   the hash table construction or redundancy elimination phases start

   processing a new basic block.  */

static int

load_killed_in_block_p (int uid_limit, rtx x, bool after_insn)

{

  struct modifies_mem *list_entry = modifies_mem_list;

  while (list_entry)

    {

      rtx setter = list_entry->insn;

      /* Ignore entries in the list that do not apply.  */

      if ((after_insn

           && INSN_CUID (setter) < uid_limit)

          || (! after_insn

              && INSN_CUID (setter) > uid_limit))

        {

          list_entry = list_entry->next;

          continue;

        }

      /* If SETTER is a call everything is clobbered.  Note that calls

         to pure functions are never put on the list, so we need not

         worry about them.  */

      if (CALL_P (setter))

        return 1;

      /* SETTER must be an insn of some kind that sets memory.  Call

         note_stores to examine each hunk of memory that is modified.

         It will set mems_conflict_p to nonzero if there may be a

         conflict between X and SETTER.  */

      mems_conflict_p = 0;

      note_stores (PATTERN (setter), find_mem_conflicts, x);

      if (mems_conflict_p)

        return 1;

      list_entry = list_entry->next;

    }

  return 0;

}

/* Record register first/last/block set information for REGNO in INSN.  */

static inline void

record_last_reg_set_info (rtx insn, int regno)

{

  reg_avail_info[regno] = INSN_CUID (insn);

}

/* Record memory modification information for INSN.  We do not actually care

   about the memory location(s) that are set, or even how they are set (consider

   a CALL_INSN).  We merely need to record which insns modify memory.  */

static void

record_last_mem_set_info (rtx insn)

{

  struct modifies_mem *list_entry;

  list_entry = (struct modifies_mem *) obstack_alloc (&modifies_mem_obstack,

                                                      sizeof (struct modifies_mem));

  list_entry->insn = insn;

  list_entry->next = modifies_mem_list;

  modifies_mem_list = list_entry;

}

/* Called from compute_hash_table via note_stores to handle one

   SET or CLOBBER in an insn.  DATA is really the instruction in which

   the SET is taking place.  */

static void

record_last_set_info (rtx dest, rtx setter ATTRIBUTE_UNUSED, void *data)

{

  rtx last_set_insn = (rtx) data;

  if (GET_CODE (dest) == SUBREG)

    dest = SUBREG_REG (dest);

  if (REG_P (dest))

    record_last_reg_set_info (last_set_insn, REGNO (dest));

  else if (MEM_P (dest))

    {

      /* Ignore pushes, they don't clobber memory.  They may still

         clobber the stack pointer though.  Some targets do argument

         pushes without adding REG_INC notes.  See e.g. PR25196,

         where a pushsi2 on i386 doesn't have REG_INC notes.  Note

         such changes here too.  */

      if (! push_operand (dest, GET_MODE (dest)))

        record_last_mem_set_info (last_set_insn);

      else

        record_last_reg_set_info (last_set_insn, STACK_POINTER_REGNUM);

    }

}

/* Reset tables used to keep track of what's still available since the

   start of the block.  */

static void

reset_opr_set_tables (void)

{

  memset (reg_avail_info, 0, FIRST_PSEUDO_REGISTER * sizeof (int));

  obstack_free (&modifies_mem_obstack, modifies_mem_obstack_bottom);

  modifies_mem_list = NULL;

}

/* Record things set by INSN.

   This data is used by oprs_unchanged_p.  */

static void

record_opr_changes (rtx insn)

{

  rtx note;

  /* Find all stores and record them.  */

  note_stores (PATTERN (insn), record_last_set_info, insn);

  /* Also record autoincremented REGs for this insn as changed.  */

  for (note = REG_NOTES (insn); note; note = XEXP (note, 1))

    if (REG_NOTE_KIND (note) == REG_INC)

      record_last_reg_set_info (insn, REGNO (XEXP (note, 0)));

  /* Finally, if this is a call, record all call clobbers.  */

  if (CALL_P (insn))

    {

      unsigned int regno;

      for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)