Subversion Repositories openrisc_2011-10-31

/* dcache-model.c -- data cache simulation

   Copyright (C) 1999 Damjan Lampret, lampret@opencores.org

   Copyright (C) 2008 Embecosm Limited

   Contributor Jeremy Bennett <jeremy.bennett@embecosm.com>

   This file is part of OpenRISC 1000 Architectural Simulator.

   This program 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 of the License, or (at your option)

   any later version.

   This program 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 this program.  If not, see <http://www.gnu.org/licenses/>. */

/* This program is commented throughout in a fashion suitable for processing

   with Doxygen. */

/* Cache functions.  At the moment these functions only simulate functionality

   of data caches and do not influence on fetche/decode/execute stages and

   timings.  They are here only to verify performance of various cache

   configurations. */

/* Autoconf and/or portability configuration */

#include "config.h"

/* Package includes */

#include "dcache-model.h"

#include "execute.h"

#include "spr-defs.h"

#include "abstract.h"

#include "stats.h"

#include "misc.h"

/* Data cache */

struct dc_set

{

  struct

  {

    uint32_t line[MAX_DC_BLOCK_SIZE/4];

    oraddr_t tagaddr;           /* tag address */

    int lru;                    /* least recently used */

  } way[MAX_DC_WAYS];

} dc[MAX_DC_SETS];

void

dc_info (void)

{

  if (!(cpu_state.sprs[SPR_UPR] & SPR_UPR_DCP))

    {

      PRINTF ("DCache not implemented. Set UPR[DCP].\n");

      return;

    }

  PRINTF ("Data cache %dKB: ",

          config.dc.nsets * config.dc.blocksize * config.dc.nways / 1024);

  PRINTF ("%d ways, %d sets, block size %d bytes\n", config.dc.nways,

          config.dc.nsets, config.dc.blocksize);

}

/* First check if data is already in the cache and if it is:

    - increment DC read hit stats,

    - set 'lru' at this way to config.dc.ustates - 1 and

      decrement 'lru' of other ways unless they have reached 0,

   and if not:

    - increment DC read miss stats

    - find lru way and entry and replace old tag with tag of the 'dataaddr'

    - set 'lru' with config.dc.ustates - 1 and decrement 'lru' of other

      ways unless they have reached 0

    - refill cache line

*/

uint32_t

dc_simulate_read (oraddr_t dataaddr, oraddr_t virt_addr, int width)

{

  int set, way = -1;

  int i;

  oraddr_t tagaddr;

  uint32_t tmp = 0;

  if (!(cpu_state.sprs[SPR_UPR] & SPR_UPR_DCP) ||

      !(cpu_state.sprs[SPR_SR] & SPR_SR_DCE) || data_ci)

    {

      if (width == 4)

        tmp = evalsim_mem32 (dataaddr, virt_addr);

      else if (width == 2)

        tmp = evalsim_mem16 (dataaddr, virt_addr);

      else if (width == 1)

        tmp = evalsim_mem8 (dataaddr, virt_addr);

      if (cur_area && cur_area->log)

        fprintf (cur_area->log, "[%" PRIxADDR "] -> read %08" PRIx32 "\n",

                 dataaddr, tmp);

      return tmp;

    }

  /* Which set to check out? */

  set = (dataaddr / config.dc.blocksize) % config.dc.nsets;

  tagaddr = (dataaddr / config.dc.blocksize) / config.dc.nsets;

  /* Scan all ways and try to find a matching way. */

  for (i = 0; i < config.dc.nways; i++)

    if (dc[set].way[i].tagaddr == tagaddr)

      way = i;

  /* Did we find our cached data? */

  if (way >= 0)

    {                           /* Yes, we did. */

      dc_stats.readhit++;

      for (i = 0; i < config.dc.nways; i++)

        if (dc[set].way[i].lru > dc[set].way[way].lru)

          dc[set].way[i].lru--;

      dc[set].way[way].lru = config.dc.ustates - 1;

      runtime.sim.mem_cycles += config.dc.load_hitdelay;

      tmp =

        dc[set].way[way].line[(dataaddr & (config.dc.blocksize - 1)) >> 2];

      if (width == 4)

        return tmp;

      else if (width == 2)

        {

          tmp = ((tmp >> ((dataaddr & 2) ? 0 : 16)) & 0xffff);

          return tmp;

        }

      else if (width == 1)

        {

          tmp = ((tmp >> (8 * (3 - (dataaddr & 3)))) & 0xff);

          return tmp;

        }

    }

  else

    {                           /* No, we didn't. */

      int minlru = config.dc.ustates - 1;

      int minway = 0;

      dc_stats.readmiss++;

      for (i = 0; i < config.dc.nways; i++)

        {

          if (dc[set].way[i].lru < minlru)

            {

              minway = i;

              minlru = dc[set].way[i].lru;

            }

        }

      for (i = 0; i < (config.dc.blocksize); i += 4)

        {

          /* FIXME: What is the virtual address meant to be? (ie. What happens if

           * we read out of memory while refilling a cache line?) */

          tmp =

            evalsim_mem32 ((dataaddr & ~(config.dc.blocksize - 1)) +

                           (((dataaddr & ~ADDR_C (3)) +

                             i) & (config.dc.blocksize - 1)), 0);

          dc[set].way[minway].

            line[((dataaddr + i) & (config.dc.blocksize - 1)) >> 2] = tmp;

          if (!cur_area)

            {

              dc[set].way[minway].tagaddr = -1;

              dc[set].way[minway].lru = 0;

              return 0;

            }

          else if (cur_area->log)

            fprintf (cur_area->log, "[%" PRIxADDR "] -> read %08" PRIx32 "\n",

                     dataaddr, tmp);

        }

      dc[set].way[minway].tagaddr = tagaddr;

      for (i = 0; i < config.dc.nways; i++)

        if (dc[set].way[i].lru)

          dc[set].way[i].lru--;

      dc[set].way[minway].lru = config.dc.ustates - 1;

      runtime.sim.mem_cycles += config.dc.load_missdelay;

      tmp =

        dc[set].way[minway].line[(dataaddr & (config.dc.blocksize - 1)) >> 2];

      if (width == 4)

        return tmp;

      else if (width == 2)

        {

          tmp = (tmp >> ((dataaddr & 2) ? 0 : 16)) & 0xffff;

          return tmp;

        }

      else if (width == 1)

        {

          tmp = (tmp >> (8 * (3 - (dataaddr & 3)))) & 0xff;

          return tmp;

        }

    }

  return 0;

}

/* First check if data is already in the cache and if it is:

    - increment DC write hit stats,

    - set 'lru' at this way to config.dc.ustates - 1 and

      decrement 'lru' of other ways unless they have reached 0,

   and if not:

    - increment DC write miss stats

    - find lru way and entry and replace old tag with tag of the 'dataaddr'

    - set 'lru' with config.dc.ustates - 1 and decrement 'lru' of other

      ways unless they have reached 0

*/

void

dc_simulate_write (oraddr_t dataaddr, oraddr_t virt_addr, uint32_t data,

                   int width)

{

  int set, way = -1;

  int i;

  oraddr_t tagaddr;

  uint32_t tmp;

  if (width == 4)

    setsim_mem32 (dataaddr, virt_addr, data);

  else if (width == 2)

    setsim_mem16 (dataaddr, virt_addr, data);

  else if (width == 1)

    setsim_mem8 (dataaddr, virt_addr, data);

  if (!(cpu_state.sprs[SPR_UPR] & SPR_UPR_DCP) ||

      !(cpu_state.sprs[SPR_SR] & SPR_SR_DCE) || data_ci || !cur_area)

    return;

  /* Which set to check out? */

  set = (dataaddr / config.dc.blocksize) % config.dc.nsets;

  tagaddr = (dataaddr / config.dc.blocksize) / config.dc.nsets;

  /* Scan all ways and try to find a matching way. */

  for (i = 0; i < config.dc.nways; i++)

    if (dc[set].way[i].tagaddr == tagaddr)

      way = i;

  /* Did we find our cached data? */

  if (way >= 0)

    {                           /* Yes, we did. */

      dc_stats.writehit++;

      for (i = 0; i < config.dc.nways; i++)

        if (dc[set].way[i].lru > dc[set].way[way].lru)

          dc[set].way[i].lru--;

      dc[set].way[way].lru = config.dc.ustates - 1;

      runtime.sim.mem_cycles += config.dc.store_hitdelay;

      tmp =

        dc[set].way[way].line[(dataaddr & (config.dc.blocksize - 1)) >> 2];

      if (width == 4)

        tmp = data;

      else if (width == 2)

        {

          tmp &= 0xffff << ((dataaddr & 2) ? 16 : 0);

          tmp |= (data & 0xffff) << ((dataaddr & 2) ? 0 : 16);

        }

      else if (width == 1)

        {

          tmp &= ~(0xff << (8 * (3 - (dataaddr & 3))));

          tmp |= (data & 0xff) << (8 * (3 - (dataaddr & 3)));

        }

      dc[set].way[way].line[(dataaddr & (config.dc.blocksize - 1)) >> 2] =

        tmp;

    }

  else

    {                           /* No, we didn't. */

      int minlru = config.dc.ustates - 1;

      int minway = 0;

      dc_stats.writemiss++;

      for (i = 0; i < config.dc.nways; i++)

        if (dc[set].way[i].lru < minlru)

          minway = i;

      for (i = 0; i < (config.dc.blocksize); i += 4)

        {

          dc[set].way[minway].

            line[((dataaddr + i) & (config.dc.blocksize - 1)) >> 2] =

            /* FIXME: Same comment as in dc_simulate_read */

            evalsim_mem32 ((dataaddr & ~(config.dc.blocksize - 1)) +

                           (((dataaddr & ~3ul) + i) & (config.dc.blocksize -

                                                       1)), 0);

          if (!cur_area)

            {

              dc[set].way[minway].tagaddr = -1;

              dc[set].way[minway].lru = 0;

              return;

            }

        }

      dc[set].way[minway].tagaddr = tagaddr;

      for (i = 0; i < config.dc.nways; i++)

        if (dc[set].way[i].lru)

          dc[set].way[i].lru--;

      dc[set].way[minway].lru = config.dc.ustates - 1;

      runtime.sim.mem_cycles += config.dc.store_missdelay;

    }

}

/* First check if data is already in the cache and if it is:

    - invalidate block if way isn't locked

   otherwise don't do anything.

*/

void

dc_inv (oraddr_t dataaddr)

{

  int set, way = -1;

  int i;

  oraddr_t tagaddr;

  if (!(cpu_state.sprs[SPR_UPR] & SPR_UPR_DCP))

    return;

  /* Which set to check out? */

  set = (dataaddr / config.dc.blocksize) % config.dc.nsets;

  tagaddr = (dataaddr / config.dc.blocksize) / config.dc.nsets;

  if (!(cpu_state.sprs[SPR_SR] & SPR_SR_DCE))

    {

      for (i = 0; i < config.dc.nways; i++)

        {

          dc[set].way[i].tagaddr = -1;

          dc[set].way[i].lru = 0;

        }

      return;

    }

  /* Scan all ways and try to find a matching way. */

  for (i = 0; i < config.dc.nways; i++)

    if (dc[set].way[i].tagaddr == tagaddr)

      way = i;

  /* Did we find our cached data? */

  if (way >= 0)

    {                           /* Yes, we did. */

      dc[set].way[way].tagaddr = -1;

      dc[set].way[way].lru = 0;

    }

}

/*-----------------------------------------------------[ DC configuration ]---*/

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

/*!Enable or disable the data cache

   Set the corresponding field in the UPR

   @param[in] val  The value to use

   @param[in] dat  The config data structure (not used here)                 */

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

static void

dc_enabled (union param_val  val,

            void            *dat)

{

  if (val.int_val)

    {

      cpu_state.sprs[SPR_UPR] |= SPR_UPR_DCP;

    }

  else

    {

      cpu_state.sprs[SPR_UPR] &= ~SPR_UPR_DCP;

    }

  config.dc.enabled = val.int_val;

    }   /* dc_enabled() */

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

/*!Set the number of data cache sets

   Value must be a power of 2 <= MAX_DC_SETS. If not issue a warning and

   ignore. Set the relevant field in the data cache config register

   @param[in] val  The value to use

   @param[in] dat  The config data structure (not used here)                 */

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

static void

dc_nsets (union param_val  val,

          void            *dat)

{

  if (is_power2 (val.int_val) && (val.int_val <= MAX_DC_SETS))

    {

      int  set_bits = log2_int (val.int_val);

      config.dc.nsets = val.int_val;

      cpu_state.sprs[SPR_DCCFGR] &= ~SPR_DCCFGR_NCS;

      cpu_state.sprs[SPR_DCCFGR] |= set_bits << SPR_DCCFGR_NCS_OFF;

    }

  else

    {

      fprintf (stderr, "Warning: data cache nsets not a power of 2 <= %d: "

               "ignored\n", MAX_DC_SETS);

    }

}       /* dc_nsets() */

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

/*!Set the number of data cache ways

   Value must be a power of 2 <= MAX_DC_WAYS. If not issue a warning and

   ignore. Set the relevant field in the data cache config register

   @param[in] val  The value to use

   @param[in] dat  The config data structure (not used here)                 */

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

static void

dc_nways (union param_val  val,

          void            *dat)

{

  if (is_power2 (val.int_val) && (val.int_val <= MAX_DC_WAYS))

    {

      int  way_bits = log2_int (val.int_val);

      config.dc.nways = val.int_val;

      cpu_state.sprs[SPR_DCCFGR] &= ~SPR_DCCFGR_NCW;

      cpu_state.sprs[SPR_DCCFGR] |= way_bits << SPR_DCCFGR_NCW_OFF;

    }

  else

    {

      fprintf (stderr, "Warning: data cache nways not a power of 2 <= %d: "

               "ignored\n", MAX_DC_WAYS);

    }

}        /* dc_nways() */

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

/*!Set the data cache block size

   Value must be either MIN_DC_BLOCK_SIZE or MAX_DC_BLOCK_SIZE. If not issue a

   warning and ignore. Set the relevant field in the data cache config register

   @param[in] val  The value to use

   @param[in] dat  The config data structure (not used here)                 */

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

static void

dc_blocksize (union param_val  val,

              void            *dat)

{

  switch (val.int_val)

    {

    case MIN_DC_BLOCK_SIZE:

      config.dc.blocksize         = val.int_val;

      cpu_state.sprs[SPR_DCCFGR] &= ~SPR_DCCFGR_CBS;

      break;

    case MAX_DC_BLOCK_SIZE:

      config.dc.blocksize         = val.int_val;

      cpu_state.sprs[SPR_DCCFGR] |= SPR_DCCFGR_CBS;

      break;

    default:

      fprintf (stderr, "Warning: data cache block size not %d or %d: "

               "ignored\n", MIN_DC_BLOCK_SIZE, MAX_DC_BLOCK_SIZE);

      break;

    }

}       /* dc_blocksize() */

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

/*!Set the number of data cache usage states

   Value must be 2, 3 or 4. If not issue a warning and ignore.

   @param[in] val  The value to use

   @param[in] dat  The config data structure (not used here)                 */

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

static void

dc_ustates (union param_val  val,

            void            *dat)

{

  if ((val.int_val >= 2) && (val.int_val <= 4))

    {

      config.dc.ustates = val.int_val;

    }

  else

    {

      fprintf (stderr, "Warning number of data cache usage states must be "

               "2, 3 or 4: ignored\n");

    }

}       /* dc_ustates() */

static void

dc_load_hitdelay (union param_val val, void *dat)

{

  config.dc.load_hitdelay = val.int_val;

}

static void

dc_load_missdelay (union param_val val, void *dat)

{

  config.dc.load_missdelay = val.int_val;

}

static void

dc_store_hitdelay (union param_val val, void *dat)

{

  config.dc.store_hitdelay = val.int_val;

}

static void

dc_store_missdelay (union param_val val, void *dat)

{

  config.dc.store_missdelay = val.int_val;

}

void

reg_dc_sec (void)

{

  struct config_section *sec = reg_config_sec ("dc", NULL, NULL);

  reg_config_param (sec, "enabled", paramt_int, dc_enabled);

  reg_config_param (sec, "nsets", paramt_int, dc_nsets);

  reg_config_param (sec, "nways", paramt_int, dc_nways);

  reg_config_param (sec, "blocksize", paramt_int, dc_blocksize);

  reg_config_param (sec, "ustates", paramt_int, dc_ustates);

  reg_config_param (sec, "load_hitdelay", paramt_int, dc_load_hitdelay);

  reg_config_param (sec, "load_missdelay", paramt_int, dc_load_missdelay);

  reg_config_param (sec, "store_hitdelay", paramt_int, dc_store_hitdelay);

  reg_config_param (sec, "store_missdelay", paramt_int, dc_store_missdelay);

}