#ifdef SYSTEMC
/*
 * $Id: Load_store_unit_function_speculative_load_access_transition.cpp 138 2010-05-12 17:34:01Z rosiere $
 *
 * [ Description ]
 * 
 */

#include "Behavioural/Core/Multi_Execute_loop/Execute_loop/Multi_Execute_unit/Execute_unit/Load_store_unit/include/Load_store_unit.h"
#include "Behavioural/include/Identification.h"

namespace morpheo                    {
namespace behavioural {
namespace core {
namespace multi_execute_loop {
namespace execute_loop {
namespace multi_execute_unit {
namespace execute_unit {
namespace load_store_unit {

template <typename T>
T swapBytes (T data, uint32_t size_data, uint32_t size_access)
{
  uint64_t x = static_cast<uint64_t>(data);

//   switch (size_data)
//     {
//     case 2 : // 16 bits
//       {
//         switch (size_access)
//           {
//           case 2 : 
//             {
//               x = ((((x>> 8)&0xff) <<  0) |
//                    (((x>> 0)&0xff) <<  8) );
//               break;
//             }
//           default :
//             {
//               break;
//             }
//           }
//         break;
//       }
//     case 4 : // 32 bits
//       {
//         switch (size_access)
//           {
//           case 2 : 
//             {
//               x = ((((x>> 8)&0xff) <<  0) |
//                    (((x>> 0)&0xff) <<  8) |
//                    (((x>>24)&0xff) << 16) |
//                    (((x>>16)&0xff) << 24) );
//               break;
//             }
//           case 4 : 
//             {
//               x = ((((x>>24)&0xff) <<  0) |
//                    (((x>>16)&0xff) <<  8) |
//                    (((x>> 8)&0xff) << 16) |
//                    (((x>> 0)&0xff) << 24) );
//               break;
//             }
//           default :
//             {
//               break;
//             }
//           }
//         break;
//       }
//     case 8 : // 64 bits
//       {
//         switch (size_access)
//           {
//           case 2 : 
//             {
//               x = ((((x>> 8)&0xff) <<  0) |
//                    (((x>> 0)&0xff) <<  8) |
//                    (((x>>24)&0xff) << 16) |
//                    (((x>>16)&0xff) << 24) |
//                    (((x>>40)&0xff) << 32) |
//                    (((x>>32)&0xff) << 40) |
//                    (((x>>56)&0xff) << 48) |
//                    (((x>>48)&0xff) << 56) );
//               break;
//             }
//           case 4 : 
//             {
//               x = ((((x>>24)&0xff) <<  0) |
//                    (((x>>16)&0xff) <<  8) |
//                    (((x>> 8)&0xff) << 16) |
//                    (((x>> 0)&0xff) << 24) |
//                    (((x>>56)&0xff) << 32) |
//                    (((x>>48)&0xff) << 40) |
//                    (((x>>40)&0xff) << 48) |
//                    (((x>>32)&0xff) << 56) );
//               break;
//             }
//           case 8 : 
//             {
//               x = ((((x>>56)&0xff) <<  0) |
//                    (((x>>48)&0xff) <<  8) |
//                    (((x>>40)&0xff) << 16) |
//                    (((x>>32)&0xff) << 24) |
//                    (((x>>24)&0xff) << 32) |
//                    (((x>>16)&0xff) << 40) |
//                    (((x>> 8)&0xff) << 48) |
//                    (((x>> 0)&0xff) << 56) );
//               break;
//             }
//           default :
//             {
//               break;
//             }
//           }
//         break;
//       }
//     default :
//       {
//         break;
//       }
//     }


  uint64_t y=0;

  for (uint32_t i=0; i<size_data; i+=size_access)
    {
      uint32_t offset = i<<3;

      switch (size_access)
        {
        case 1 :
          {
            y = x;
            break;
          }
        case 2 : 
          {
            y |= ((((x>>( 8+offset))&0xff) << ( 0+offset)) |
                  (((x>>( 0+offset))&0xff) << ( 8+offset)) );
            break;
          }
        case 4 : 
          {
            y |= ((((x>>(24+offset))&0xff) << ( 0+offset)) |
                  (((x>>(16+offset))&0xff) << ( 8+offset)) |
                  (((x>>( 8+offset))&0xff) << (16+offset)) |
                  (((x>>( 0+offset))&0xff) << (24+offset)) );
            break;
          }
        case 8 : 
          {
            y |= ((((x>>(56+offset))&0xff) << ( 0+offset)) |
                  (((x>>(48+offset))&0xff) << ( 8+offset)) |
                  (((x>>(40+offset))&0xff) << (16+offset)) |
                  (((x>>(32+offset))&0xff) << (24+offset)) |
                  (((x>>(24+offset))&0xff) << (32+offset)) |
                  (((x>>(16+offset))&0xff) << (40+offset)) |
                  (((x>>( 8+offset))&0xff) << (48+offset)) |
                  (((x>>( 0+offset))&0xff) << (56+offset)) );
            break;
          }
        default :
            {
              break;
            }
        }
    }

  return static_cast<T>(y);
}

template <typename T>
T swapBits (T data, uint32_t size_data, uint32_t size_access)
{
  uint8_t x = static_cast<uint8_t>(data);

  uint8_t y=0;

  for (uint32_t i=0; i<size_data; i+=size_access)
    {
      uint32_t offset = i;

      switch (size_access)
        {
        case 1 :
          {
            y = x;
            break;
          }
        case 2 : 
          {
            y |= ((((x>>( 1+offset))&0x1) << ( 0+offset)) |
                  (((x>>( 0+offset))&0x1) << ( 1+offset)) );
            break;
          }
        case 4 : 
          {
            y |= ((((x>>( 3+offset))&0x1) << ( 0+offset)) |
                  (((x>>( 2+offset))&0x1) << ( 1+offset)) |
                  (((x>>( 1+offset))&0x1) << ( 2+offset)) |
                  (((x>>( 0+offset))&0x1) << ( 3+offset)) );
            break;
          }
        case 8 : 
          {
            y |= ((((x>>( 7+offset))&0x1) << ( 0+offset)) |
                  (((x>>( 6+offset))&0x1) << ( 1+offset)) |
                  (((x>>( 5+offset))&0x1) << ( 2+offset)) |
                  (((x>>( 4+offset))&0x1) << ( 3+offset)) |
                  (((x>>( 3+offset))&0x1) << ( 4+offset)) |
                  (((x>>( 2+offset))&0x1) << ( 5+offset)) |
                  (((x>>( 1+offset))&0x1) << ( 6+offset)) |
                  (((x>>( 0+offset))&0x1) << ( 7+offset)) );
            break;
          }
        default :
            {
              break;
            }
        }
    }

  return static_cast<T>(y);
}

#undef  FUNCTION
#define FUNCTION "Load_store_unit::function_speculative_load_access_transition"
  void Load_store_unit::function_speculative_load_access_transition (void)
  {
    log_begin(Load_store_unit,FUNCTION);
    log_function(Load_store_unit,FUNCTION,_name.c_str());

    if (PORT_READ(in_NRESET) == 0)
      {
	// Reset : clear all queue
	_speculative_access_queue_control->clear();

	reg_STORE_QUEUE_PTR_READ       = 0;
	reg_LOAD_QUEUE_CHECK_PRIORITY  = 0;

	for (uint32_t i=0; i< _param->_size_store_queue             ; i++)
          {
            reg_STORE_QUEUE_NB_CHECK  [i] = 0;
//          reg_STORE_QUEUE_INVALID   [i] = 0;
  	    _store_queue              [i]._state                 = STORE_QUEUE_EMPTY;
  	    _store_queue              [i]._context_id            = 0; // not necessary
  	    _store_queue              [i]._front_end_id          = 0; // not necessary
  	    _store_queue              [i]._ooo_engine_id         = 0; // not necessary
  	    _store_queue              [i]._packet_id             = 0; // not necessary
  	    _store_queue              [i]._operation             = 0; // not necessary
  	    _store_queue              [i]._cancel                = 0; // not necessary
  	    _store_queue              [i]._load_queue_ptr_write  = 0; // not necessary
  	    _store_queue              [i]._address               = 0; // not necessary
  	    _store_queue              [i]._wdata                 = 0; // not necessary
//	    _store_queue              [i]._write_rd              = 0; // not necessary
//	    _store_queue              [i]._num_reg_rd            = 0; // not necessary
  	    _store_queue              [i]._exception             = 0; // not necessary
  	    _store_queue              [i]._send_commit           = 0; // not necessary
          }

	for (uint32_t i=0; i< _param->_size_load_queue              ; i++)
          {
            _load_queue               [i]._state                 = LOAD_QUEUE_EMPTY;
            _load_queue               [i]._context_id            = 0; // not necessary
            _load_queue               [i]._front_end_id          = 0; // not necessary
            _load_queue               [i]._ooo_engine_id         = 0; // not necessary
            _load_queue               [i]._packet_id             = 0; // not necessary
            _load_queue               [i]._operation             = 0; // not necessary
            _load_queue               [i]._cancel                = 0; // not necessary
            _load_queue               [i]._store_queue_ptr_write = 0; // not necessary
            _load_queue               [i]._store_queue_ptr_read  = 0; // not necessary
            _load_queue               [i]._store_queue_empty     = 0; // not necessary
            _load_queue               [i]._address               = 0; // not necessary
            _load_queue               [i]._check_hit_byte        = 0; // not necessary 
            _load_queue               [i]._check_hit             = 0; // not necessary
            _load_queue               [i]._shift                 = 0; // not necessary
            _load_queue               [i]._is_load_signed        = 0; // not necessary
            _load_queue               [i]._access_size           = 0; // not necessary
            _load_queue               [i]._rdata                 = 0; // not necessary
            _load_queue               [i]._write_rd              = 0; // not necessary
            _load_queue               [i]._num_reg_rd            = 0; // not necessary
            _load_queue               [i]._exception             = 0; // not necessary
            _load_queue               [i]._can_speculative_commit= 0; // not necessary
          }

	for (uint32_t i=0; i< _param->_size_speculative_access_queue; i++)
          {
            _speculative_access_queue [i]._state                 = SPECULATIVE_ACCESS_QUEUE_EMPTY;
            _speculative_access_queue [i]._context_id            = 0; // not necessary
            _speculative_access_queue [i]._front_end_id          = 0; // not necessary
            _speculative_access_queue [i]._ooo_engine_id         = 0; // not necessary
            _speculative_access_queue [i]._packet_id             = 0; // not necessary
            _speculative_access_queue [i]._operation             = 0; // not necessary
            _speculative_access_queue [i]._cancel                = 0; // not necessary
            _speculative_access_queue [i]._load_queue_ptr_write  = 0; // not necessary
            _speculative_access_queue [i]._store_queue_ptr_write = 0; // not necessary
            _speculative_access_queue [i]._store_queue_ptr_read  = 0; // not necessary
            _speculative_access_queue [i]._store_queue_empty     = 0; // not necessary
            _speculative_access_queue [i]._address               = 0; // not necessary
            _speculative_access_queue [i]._write_rd              = 0; // not necessary
            _speculative_access_queue [i]._num_reg_rd            = 0; // not necessary
            _speculative_access_queue [i]._exception             = 0; // not necessary
          }
      }
    else
      {
        // load_queue_push if speculative_access_queue have access at the dcache, or they have an event
	bool load_queue_push = (_speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._state == SPECULATIVE_ACCESS_QUEUE_WAIT_LOAD_QUEUE);

	//================================================================
	// Interface "MEMORY_OUT"
	//================================================================

        if ((    internal_MEMORY_OUT_VAL  == 1) and
            (PORT_READ(in_MEMORY_OUT_ACK[0]) == 1))
          {
	    log_printf(TRACE,Load_store_unit,FUNCTION,"  * MEMORY_OUT[0] transaction");

	    switch (internal_MEMORY_OUT_SELECT_QUEUE)
	      {
	      case SELECT_STORE_QUEUE :
		{
		  // =======================
		  // ===== STORE_QUEUE =====
		  // =======================
		  
		  log_printf(TRACE,Load_store_unit,FUNCTION,"    * store_queue [%d]",internal_MEMORY_OUT_PTR);
	    
		  // Entry flush and increase the read pointer
		  if (_store_queue [internal_MEMORY_OUT_PTR]._send_commit)
		    {
		      _store_queue [internal_MEMORY_OUT_PTR]._state = STORE_QUEUE_EMPTY;
		      reg_STORE_QUEUE_PTR_READ = (reg_STORE_QUEUE_PTR_READ+1)%_param->_size_store_queue;
		    }
		  else
		    _store_queue [internal_MEMORY_OUT_PTR]._send_commit = true;

		  break;
		}
	      case SELECT_LOAD_QUEUE :
		{
		  // ======================
		  // ===== LOAD_QUEUE =====
		  // ======================
		  
		  log_printf(TRACE,Load_store_unit,FUNCTION,"    * load_queue  [%d]",internal_MEMORY_OUT_PTR);
		  
		  // Entry flush and increase the read pointer
		  
		  _load_queue [internal_MEMORY_OUT_PTR]._state = LOAD_QUEUE_EMPTY;
		  
		  // reg_LOAD_QUEUE_PTR_READ = (reg_LOAD_QUEUE_PTR_READ+1)%_param->_size_load_queue;

		  break;
		}
	      case SELECT_LOAD_QUEUE_SPECULATIVE :
		{
		  log_printf(TRACE,Load_store_unit,FUNCTION,"    * load_queue  [%d] (speculative)",internal_MEMORY_OUT_PTR);
		  
                  // !!! WARNING !!!
                  // !!! Test special case :
                  // !!! in a cycle an instruction can check the last store AND commit instruction
                  // !!! also the memory_out is before the port_check

		  _load_queue [internal_MEMORY_OUT_PTR]._state    = LOAD_QUEUE_CHECK;
		  // NOTE : a speculative load write in the register file.
		  // if the speculation is a miss, write_rd is re set at 1.
		  _load_queue [internal_MEMORY_OUT_PTR]._write_rd = 0;

#ifdef STATISTICS
                  if (usage_is_set(_usage,USE_STATISTICS))
                    (*_stat_nb_inst_load_commit_speculative) ++;
#endif

		  break;
		}

		break;
	      }
	  }

	//================================================================
	// Interface "PORT_CHECK"
	//================================================================
	
	// Plusieurs moyens de faire la verification de dépendance entre les loads et les stores.
	//  1) un load ne peut vérifier qu'un store par cycle. Dans ce cas port_check <= size_load_queue
	//  2) un load tente de vérifier le maximum de store par cycle. Dans ce cas ce n'est pas du pointeur d'écriture qu'il lui faut mais un vecteur de bit indiquant quel store à déjà été testé. De plus il faut un bit indiquant qu'il y a un match mais que ce n'est pas forcément le premier.

	// solution 1)
 	log_printf(TRACE,Load_store_unit,FUNCTION,"  * CHECK");
	for (uint32_t i=0, nb_check=0; (nb_check<_param->_nb_port_check) and (i<_param->_size_load_queue); i++)
	  {
            // Get an index from load queue
	    uint32_t index_load_queue = (i + reg_LOAD_QUEUE_CHECK_PRIORITY)%_param->_size_load_queue;

            // Test if this load must ckecked store queue
	    if (((_load_queue[index_load_queue]._state == LOAD_QUEUE_WAIT_CHECK) or
		 (_load_queue[index_load_queue]._state == LOAD_QUEUE_COMMIT_CHECK) or
		 (_load_queue[index_load_queue]._state == LOAD_QUEUE_CHECK)) and
		is_operation_memory_load(_load_queue[index_load_queue]._operation))
	      {
		log_printf(TRACE,Load_store_unit,FUNCTION,"    * Find a load : %d",index_load_queue);

		nb_check++; // use one port

		// find a entry that it need a check
		Tlsq_ptr_t store_queue_ptr_write     = _load_queue[index_load_queue]._store_queue_ptr_write;
		Tlsq_ptr_t store_queue_ptr_read      = _load_queue[index_load_queue]._store_queue_ptr_read ;
 		Tlsq_ptr_t store_queue_empty         = _load_queue[index_load_queue]._store_queue_empty    ;
// 		Tlsq_ptr_t store_queue_ptr_write_old = store_queue_ptr_write;

                // Init variable
		bool       end_check    = false;
		bool       change_state = false;
		bool       next         = false;

		// At the first store queue empty, stop check.
		// Explication :
		//  * rename logic keep a empty case in the store queue (also size_store_queue > 1)
		//  * when a store is out of store queue, also it was in head of re order buffer. Also, they are none previous load.

		log_printf(TRACE,Load_store_unit,FUNCTION,"    * store_queue_ptr_write    : %d",store_queue_ptr_write);
		log_printf(TRACE,Load_store_unit,FUNCTION,"    * store_queue_ptr_read     : %d",store_queue_ptr_read );
// 		log_printf(TRACE,Load_store_unit,FUNCTION,"    * store_queue_empty        : %d",store_queue_empty );
		log_printf(TRACE,Load_store_unit,FUNCTION,"    * reg_STORE_QUEUE_PTR_READ : %d",reg_STORE_QUEUE_PTR_READ);

 		if ((store_queue_ptr_write ==     store_queue_ptr_read) or
		    (store_queue_ptr_write == reg_STORE_QUEUE_PTR_READ))
// 		if (store_queue_empty)
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_ptr_write == store_queue_ptr_read");
		    end_check    = true;
		    change_state = true;
		  }
		else
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_ptr_write != store_queue_ptr_read");

		    store_queue_ptr_write = (store_queue_ptr_write-1)%(_param->_size_store_queue); // store_queue_ptr_write target the next slot to write, also the slot is not significatif when the load is renaming

		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_ptr_write : %d",store_queue_ptr_write);

                    bool test_thread_id = true;
                    
                    // Test thread id
                    if (_param->_have_port_context_id)
                      test_thread_id &= (_load_queue[index_load_queue]._context_id    == _store_queue[store_queue_ptr_write]._context_id);
                    if (_param->_have_port_front_end_id)
                      test_thread_id &= (_load_queue[index_load_queue]._front_end_id  == _store_queue[store_queue_ptr_write]._front_end_id);
                    if (_param->_have_port_ooo_engine_id)
                      test_thread_id &= (_load_queue[index_load_queue]._ooo_engine_id == _store_queue[store_queue_ptr_write]._ooo_engine_id);
		    
                    // switch on store_queue state
		    switch (_store_queue[store_queue_ptr_write]._state)
		      {
		      case STORE_QUEUE_VALID_NO_SPECULATIVE : 
		      case STORE_QUEUE_COMMIT :
		      case STORE_QUEUE_VALID_SPECULATIVE :
			{
                          log_printf(TRACE,Load_store_unit,FUNCTION,"      * store have a valid entry");
			  
			  // TODO : MMU - nous considérons que les adresses sont physique
			  if (test_thread_id)
			    {
			      // the load and store are in the same thread. Now, we must test address.

			      log_printf(TRACE,Load_store_unit,FUNCTION,"        * load and store is the same thread.");
			      Tdcache_address_t load_addr  = _load_queue [index_load_queue ]._address;
			      Tdcache_address_t store_addr = _store_queue[store_queue_ptr_write]._address;
			      
			      log_printf(TRACE,Load_store_unit,FUNCTION,"          * load_addr                     : %.8x.",load_addr );
			      log_printf(TRACE,Load_store_unit,FUNCTION,"          * store_addr                    : %.8x.",store_addr);
			      log_printf(TRACE,Load_store_unit,FUNCTION,"          * load_addr  & mask_address_msb : %.8x.",load_addr  & _param->_mask_address_msb);
			      log_printf(TRACE,Load_store_unit,FUNCTION,"          * store_addr & mask_address_msb : %.8x.",store_addr & _param->_mask_address_msb);
			      // Test if the both address target the same "word"
			      if ((load_addr  & _param->_mask_address_msb) == 
				  (store_addr & _param->_mask_address_msb))
				{
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * address_msb is the same.");

				  // all case - [] : store, () : load
				  // (1) store_max >= load_max and store_min <= load_min  ...[...(...)...]... Ok - inclusion in store
				  // (2) store_min >  load_max                            ...[...]...(...)... Ok - no conflit
				  // (3) store_max <  load_min                            ...(...)...[...]... Ok - no conflit
				  // (4) store_max <  load_max and store_min >  load_min  ...(...[...]...)... Ko - inclusion in load
				  // (5) store_max >= load_max and store_min >  load_min  ...[...(...]...)... Ko - conflit
				  // (6) store_max <  load_max and store_min <= load_min  ...(...[...)...]... Ko - conflit
				  // but :
				  // load in the cache is a word !
				  // the mask can be make when the load is commited. Also, the rdata content a full word.
				  // the only case is (4)
				  
                                  // Read data
                                  bool is_big_endian = true;

				  Tgeneral_data_t   load_data      = _load_queue [index_load_queue ]._rdata  ;
				  Tgeneral_data_t   store_data     = _store_queue[store_queue_ptr_write]._wdata  ;
                                  Tdcache_address_t check_hit_byte = _load_queue [index_load_queue ]._check_hit_byte;
                                  Tcontrol_t        check_hit      = _load_queue [index_load_queue ]._check_hit;
                                  uint32_t          load_size_access  = memory_size(_load_queue [index_load_queue ]._operation)>>3;
                                  uint32_t          store_size_access = memory_size(_store_queue[store_queue_ptr_write]._operation)>>3;

                                  log_printf(TRACE,Load_store_unit,FUNCTION,"            * is_big_endian           : %d",is_big_endian);
                                  log_printf(TRACE,Load_store_unit,FUNCTION,"            * load_data               : 0x%.8x",load_data);
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * store_data              : 0x%.8x",store_data);
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit_byte          : %x",check_hit_byte);
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit               : %d",check_hit);

				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * load_size_access        : %d",load_size_access );
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * store_size_access       : %d",store_size_access);

                                  if (is_big_endian)
                                    {
                                      // swap in little endian
                                      load_data      = swapBytes<Tgeneral_data_t  >(load_data     , _param->_size_general_data>>3,load_size_access);
                                      store_data     = swapBytes<Tgeneral_data_t  >(store_data    , _param->_size_general_data>>3,store_size_access);
                                      check_hit_byte = swapBits <Tdcache_address_t>(check_hit_byte, _param->_size_general_data>>3,load_size_access);
                                      
                                      
                                      log_printf(TRACE,Load_store_unit,FUNCTION,"            * load_data      (swap 1) : 0x%.8x",load_data);
                                      log_printf(TRACE,Load_store_unit,FUNCTION,"            * store_data     (swap 1) : 0x%.8x",store_data);
                                      log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit_byte (swap 1) : %x",check_hit_byte);
                                    }

                                  uint32_t store_nb_byte     = (1<<memory_access(_store_queue[store_queue_ptr_write]._operation));

                                  // Take interval to the store
				  uint32_t store_num_byte_min = (store_addr & _param->_mask_address_lsb);
				  uint32_t store_num_byte_max = store_num_byte_min+store_nb_byte;

				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * store_num_byte_min      : %d",store_num_byte_min);
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * store_num_byte_max      : %d",store_num_byte_max);


				  // The bypass is checked byte per byte
                                  // Is same endianness : because to change endianness, we must write in special register. Also the pipeline is flushed.
				  for (uint32_t num_store_byte=store_num_byte_min; num_store_byte<store_num_byte_max; num_store_byte ++)
				    {
                                      // Make a mask
                                      uint32_t num_load_byte = num_store_byte;

//                                       if (is_big_endian)
//                                         {
//                                           // sd 0 : 0 1 2 3 4 5 6 7
//                                           // ld 0 : 0 1 2 3 4 5 6 7 >>  0 
//                                           // lw 0 :         0 1 2 3 >>  0 -4
//                                           // lw 4 : 4 5 6 7         >> 32 +4
//                                           // lh 0 :             0 1 >>  0 -6
//                                           // lh 2 :         2 3     >> 16 -2
//                                           // lh 4 :     4 5         >> 32 +2
//                                           // lh 6 : 6 7             >> 48 +6
//                                           // lb 0 :               0 >>  0 -7
//                                           // lb 1 :             1   >>  8 -5
//                                           // lb 2 :           2     >> 16 -3
//                                           // lb 3 :         3       >> 24 -1
//                                           // lb 4 :       4         >> 32 +1
//                                           // lb 5 :     5           >> 40 +3
//                                           // lb 6 :   6             >> 48 +5
//                                           // lb 7 : 7               >> 56 +7

//                                           // diff : (store_nb_byte + load_nb_byte) - 2*nb_load_byte*((num_store_byte+1)

//                                           // store duplicate = all store access can be see as full size_data store
// //                                           uint32_t load_nb_byte = (1<<memory_access(_load_queue [index_load_queue ]._operation));

// //                                           int32_t diff = ((_param->_size_general_data>>3)+load_nb_byte-2*load_nb_byte*((num_store_byte/load_nb_byte)+1));

// //                                           num_load_byte =num_store_byte+diff;

// //                                           log_printf(TRACE,Load_store_unit,FUNCTION,"              * load_nb_byte   : %d",load_nb_byte);
// //                                           log_printf(TRACE,Load_store_unit,FUNCTION,"              * diff           : %d",diff);


//                                           num_load_byte = num_store_byte;
//                                         }
//                                       else
//                                         {
//                                           // sd 0 : 0 1 2 3 4 5 6 7
//                                           // ld 0 : 0 1 2 3 4 5 6 7 >>  0
//                                           // lw 0 :         4 5 6 7 >>  0
//                                           // lw 4 : 0 1 2 3         >> 32
//                                           // lh 0 :             6 7 >>  0
//                                           // lh 2 :         4 5     >> 16
//                                           // lh 4 :     2 3         >> 32
//                                           // lh 6 : 0 1             >> 48
//                                           // lb 0 :               7 >>  0
//                                           // lb 1 :             6   >>  8
//                                           // lb 2 :           5     >> 16
//                                           // lb 3 :         4       >> 24
//                                           // lb 4 :       3         >> 32
//                                           // lb 5 :     2           >> 40
//                                           // lb 6 :   1             >> 48
//                                           // lb 7 : 0               >> 56
                                          
//                                           num_load_byte = num_store_byte;
//                                         }

				      uint32_t mask  = 1<<num_load_byte;

				      log_printf(TRACE,Load_store_unit,FUNCTION,"              * num_store_byte : %d",num_store_byte);
				      log_printf(TRACE,Load_store_unit,FUNCTION,"              * num_load_byte  : %d",num_load_byte);
				      log_printf(TRACE,Load_store_unit,FUNCTION,"              * mask           : %d",mask);

				      // Accept the bypass if :
                                      //   * they have not a previous bypass with an another store
                                      //   * it's a valid request of load
				      if ((check_hit_byte&mask)==0)
					{
                                          // Note : Store is duplicate = all store access can be see as full size_data store

                                          uint32_t num_store_bit_min = num_store_byte<<3; //*8
//                                        uint32_t num_store_bit_max = num_store_bit_min+8-1;
                                          uint32_t num_load_bit_min  = num_load_byte <<3; //*8
                                          uint32_t num_load_bit_max  = num_load_bit_min+8-1;

					  log_printf(TRACE,Load_store_unit,FUNCTION,"              * bypass !!!");
//                                        log_printf(TRACE,Load_store_unit,FUNCTION,"              * interval store : [%d:%d]",num_store_bit_max,num_store_bit_min);
                                          log_printf(TRACE,Load_store_unit,FUNCTION,"              * interval store : [..:%d]",num_store_bit_min);
                                          log_printf(TRACE,Load_store_unit,FUNCTION,"              * interval load  : [%d:%d]",num_load_bit_max,num_load_bit_min);
					  log_printf(TRACE,Load_store_unit,FUNCTION,"                * rdata_old : 0x%.8x", load_data);

					  load_data = ((((store_data>>num_store_bit_min) & 0xff) << num_load_bit_min) |
                                                       mask_not<Tdcache_data_t>(load_data,num_load_bit_max,num_load_bit_min));

					  check_hit_byte |= mask;
					  check_hit       = 1;
					  change_state = true;

					  log_printf(TRACE,Load_store_unit,FUNCTION,"                * rdata_new : 0x%.8x", load_data);
					}
				    }

                                  if (is_big_endian)
                                    {
                                      // swap in little endian
                                      load_data      = swapBytes<Tgeneral_data_t  >(load_data     , _param->_size_general_data>>3,load_size_access);
                                      check_hit_byte = swapBits <Tdcache_address_t>(check_hit_byte, _param->_size_general_data>>3,load_size_access);
                                      
                                      
                                      log_printf(TRACE,Load_store_unit,FUNCTION,"            * load_data      (swap 2) : 0x%.8x",load_data);
                                      log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit_byte (swap 2) : %x",check_hit_byte);
                                    }

				  _load_queue[index_load_queue]._rdata          = load_data;
                                  _load_queue[index_load_queue]._check_hit_byte = check_hit_byte;
                                  _load_queue[index_load_queue]._check_hit      = check_hit;

				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * load_data  (after) : 0x%.8x",load_data);

				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit          : %x",check_hit);
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * check_hit_byte     : %x",check_hit_byte);

				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * mask_end_check     : %x",(-1& _param->_mask_address_lsb));
				  log_printf(TRACE,Load_store_unit,FUNCTION,"            * mask_check_hit_byte: %x",_param->_mask_check_hit_byte);
				  // The check is finish if all bit is set
				  end_check = (_load_queue[index_load_queue]._check_hit_byte == _param->_mask_check_hit_byte);
				}
			    }
			  
			  next = true;
			  break;
			}
		      case STORE_QUEUE_EMPTY :
			{
			  log_printf(TRACE,Load_store_unit,FUNCTION,"      * store is empty ... wait.");
			  break;
			}

		      case STORE_QUEUE_NO_VALID_NO_SPECULATIVE :
			{
                          log_printf(TRACE,Load_store_unit,FUNCTION,"      * store have an invalid entry");

			  if (test_thread_id)
                            {
                              end_check    = true; // previous store is invalid -> exception
                              change_state = true;
                            }
                          else
                            next = true;
                            
			  break;
			}
		      }
		  }

		if (next)
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"              * next");
		    log_printf(TRACE,Load_store_unit,FUNCTION,"                * new store_queue_ptr_write : %d",store_queue_ptr_write);

		    log_printf(TRACE,Load_store_unit,FUNCTION,"              * update reg_STORE_QUEUE_NB_CHECK");
#ifdef DEBUG
                    if (reg_STORE_QUEUE_NB_CHECK [store_queue_ptr_write] == 0)
                      throw ERRORMORPHEO(FUNCTION,_("reg_STORE_QUEUE_NB_CHECK must be > 0\n"));
#endif

                    reg_STORE_QUEUE_NB_CHECK [store_queue_ptr_write] --;

		    log_printf(TRACE,Load_store_unit,FUNCTION,"              * reg_STORE_QUEUE_NB_CHECK 1 [%d] <- %d", store_queue_ptr_write,reg_STORE_QUEUE_NB_CHECK [store_queue_ptr_write]);

		    if ((store_queue_ptr_write ==     store_queue_ptr_read) or
			(store_queue_ptr_write == reg_STORE_QUEUE_PTR_READ))
		      {
// 			store_queue_empty = true;
			end_check         = true;
			change_state      = true;
		      }

// 		    if (_load_queue[index_load_queue]._store_queue_ptr_write == 0)
// 		      _load_queue[index_load_queue]._store_queue_ptr_write = _param->_size_store_queue-1;
// 		    else
// 		      _load_queue[index_load_queue]._store_queue_ptr_write --;
		    _load_queue[index_load_queue]._store_queue_ptr_write = store_queue_ptr_write; // because the index store have be decrease
 		    _load_queue[index_load_queue]._store_queue_empty     = store_queue_empty; // because the index store have be decrease

		    // FIXME : peut n'est pas obliger de faire cette comparaison. Au prochain cycle on le détectera que les pointeur sont égaux. Ceci évitera d'avoir deux comparateurs avec le registre "store_queue_ptr_read"
		  }

		if (change_state)
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"              * change_state");
                    log_printf(TRACE,Load_store_unit,FUNCTION,"              * end_check : %d",end_check);

		    log_printf(TRACE,Load_store_unit,FUNCTION,"                * state old : %s",toString(_load_queue[index_load_queue]._state).c_str());

		    switch (_load_queue[index_load_queue]._state)
		      {
		      case LOAD_QUEUE_WAIT_CHECK   : 
                        {
			  if (end_check)
                            _load_queue[index_load_queue]._state = LOAD_QUEUE_WAIT  ; 
                          break;
                        }
		      case LOAD_QUEUE_COMMIT_CHECK : 
			{
			  if (end_check)
			    _load_queue[index_load_queue]._state = LOAD_QUEUE_COMMIT; 
			  else
			    _load_queue[index_load_queue]._state = LOAD_QUEUE_CHECK; // No commit : check hit and no end
			  break;
			}
		      case LOAD_QUEUE_CHECK        : 
			{
			  if (end_check)
			    _load_queue[index_load_queue]._state     = LOAD_QUEUE_COMMIT;

			  // check find a bypass. A speculative load have been committed : report a speculation miss.
			  if ((_load_queue[index_load_queue]._check_hit != 0) and 
                              (_load_queue[index_load_queue]._write_rd  == 0) // is commit
                              )
			    {
			      _load_queue[index_load_queue]._exception = EXCEPTION_MEMORY_MISS_SPECULATION;
			      _load_queue[index_load_queue]._write_rd  = 1; // write the good result

#ifdef STATISTICS
                              if (usage_is_set(_usage,USE_STATISTICS))
                                (*_stat_nb_inst_load_commit_miss) ++;
#endif
			    }
			  
			  break;
			}
		      default : break;
		      }
		    log_printf(TRACE,Load_store_unit,FUNCTION,"                * state new : %s",toString(_load_queue[index_load_queue]._state).c_str());
		    log_printf(TRACE,Load_store_unit,FUNCTION,"                * exception : %d",_load_queue[index_load_queue]._exception);

                    if (end_check//  and not store_queue_empty
			)
                      {
                        log_printf(TRACE,Load_store_unit,FUNCTION,"                * end check, decrease all nb_check");
                        
                        uint32_t i=store_queue_ptr_write;
                        while (i!=store_queue_ptr_read)
                          {
                            i=((i==0)?_param->_size_store_queue:i)-1;
                            
#ifdef DEBUG
                            if (reg_STORE_QUEUE_NB_CHECK [i] == 0)
                              throw ERRORMORPHEO(FUNCTION,_("reg_STORE_QUEUE_NB_CHECK must be > 0\n"));
#endif
			    
                            reg_STORE_QUEUE_NB_CHECK [i] --;

			    log_printf(TRACE,Load_store_unit,FUNCTION,"              * reg_STORE_QUEUE_NB_CHECK 2 [%d] <- %d", i,reg_STORE_QUEUE_NB_CHECK [i]);

                            //i=(i+1)%_param->_size_store_queue;
                          }
			
			_load_queue[index_load_queue]._store_queue_empty = true; // end of check
                      }
		  }
	      }
	    // else : don't use a port
	  }
	
	//================================================================
	// Interface "MEMORY_IN"
	//================================================================
	
        if ((PORT_READ(in_MEMORY_IN_VAL [internal_MEMORY_IN_PORT]) == 1) and
            (    internal_MEMORY_IN_ACK  == 1))
          {
            log_printf(TRACE,Load_store_unit,FUNCTION,"  * MEMORY_IN [%d]",internal_MEMORY_IN_PORT);

	    // Test operation :
	    //~~~~~~~~~~~~~~~~~
	    //  store  in store_queue
	    //  load   in speculation_access_queue
	    //  others in speculation_access_queue

#ifdef DEBUG_TEST
	    if (PORT_READ(in_MEMORY_IN_TYPE [internal_MEMORY_IN_PORT]) != TYPE_MEMORY)
	      throw ERRORMORPHEO(FUNCTION,"The type is different at 'TYPE_MEMORY'");
#endif
	    Toperation_t    operation            = PORT_READ(in_MEMORY_IN_OPERATION[internal_MEMORY_IN_PORT]);
	    Tcontrol_t      cancel               = PORT_READ(in_MEMORY_IN_CANCEL   [internal_MEMORY_IN_PORT]);
	    Tgeneral_data_t address              = (PORT_READ(in_MEMORY_IN_IMMEDIAT[internal_MEMORY_IN_PORT]) +
						    PORT_READ(in_MEMORY_IN_DATA_RA [internal_MEMORY_IN_PORT]));
	    bool            exception_alignement = (mask_memory_access(operation) & address) != 0;
						    
	    if (is_operation_memory_store(operation) == true)
	      {
		// =======================
		// ===== STORE_QUEUE =====
		// =======================
		// There a two store request type :
		//   - first is operation with address and data
		//   - second is the information of re order buffer : the store become not speculative and can access at the data cache

		log_printf(TRACE,Load_store_unit,FUNCTION,"    * store_queue");
		log_printf(TRACE,Load_store_unit,FUNCTION,"      * PUSH");
		
		// Write pointer is define in rename stage :
		Tlsq_ptr_t           index         = PORT_READ(in_MEMORY_IN_STORE_QUEUE_PTR_WRITE[internal_MEMORY_IN_PORT]);
		log_printf(TRACE,Load_store_unit,FUNCTION,"      * index         : %d",index);
		
		// Need read : state and exception.
		Tstore_queue_state_t old_state     = _store_queue [index]._state;
		Tstore_queue_state_t new_state     = old_state;
		bool                 update_info   = false;

		Texception_t         old_exception = _store_queue [index]._exception;
		Texception_t         new_exception = old_exception;

		// Compute next state
		switch (old_state)
		  {
		  case STORE_QUEUE_EMPTY                   :
		    {
//                    reg_STORE_QUEUE_INVALID [index] = false; // new first store instruction

		      if (is_operation_memory_store_head(operation) == true)
			{
			  new_state = STORE_QUEUE_NO_VALID_NO_SPECULATIVE;

			  // test if is a speculation
			  if (operation == OPERATION_MEMORY_STORE_HEAD_KO)
                            {
                              new_exception = EXCEPTION_MEMORY_MISS_SPECULATION;
//                            reg_STORE_QUEUE_INVALID [index] = true;
                            }
			  else
			    new_exception = EXCEPTION_MEMORY_NONE;
			}
		      else
			{
			  new_state = STORE_QUEUE_VALID_SPECULATIVE;

			  // Test if have an exception
			  if (exception_alignement == true)
			    new_exception = EXCEPTION_MEMORY_ALIGNMENT;
			  else
			    new_exception = EXCEPTION_MEMORY_NONE;

			  update_info = true;
			}
		      break;
		    }
		  case STORE_QUEUE_NO_VALID_NO_SPECULATIVE :
		    {
#ifdef DEBUG_TEST
		      if (is_operation_memory_store_head(operation) == true)
			throw ERRORMORPHEO(FUNCTION,_("Transaction in memory_in's interface, actual state of store_queue is \"STORE_QUEUE_NO_VALID_NO_SPECULATIVE\", also a previous store_head have been receiveid. But this operation is a store_head."));
#endif
		      // Test if have a new exception (priority : miss_speculation)
		      if ((exception_alignement == true) and (old_exception == EXCEPTION_MEMORY_NONE))
			new_exception = EXCEPTION_MEMORY_ALIGNMENT;
		      
		      if (new_exception != EXCEPTION_MEMORY_NONE)
			new_state = STORE_QUEUE_COMMIT;
		      else
			new_state = STORE_QUEUE_VALID_NO_SPECULATIVE;
		      
		      update_info = true;
		      break;
		    }
		  case STORE_QUEUE_VALID_SPECULATIVE       :
		    {
#ifdef DEBUG_TEST
		      if (is_operation_memory_store_head(operation) == false)
			throw ERRORMORPHEO(FUNCTION,_("Transaction in memory_in's interface, actual state of store_queue is \"STORE_QUEUE_VALID_SPECULATIVE\", also a previous access with register and address have been receiveid. But this operation is a not store_head."));
#endif
		      if (operation == OPERATION_MEMORY_STORE_HEAD_KO)
                        {
                          new_exception = EXCEPTION_MEMORY_MISS_SPECULATION; // great prioritary
//                        reg_STORE_QUEUE_INVALID [index] = true;
                        }

		      if (new_exception != EXCEPTION_MEMORY_NONE)
			new_state = STORE_QUEUE_COMMIT;
		      else
			new_state = STORE_QUEUE_VALID_NO_SPECULATIVE;
		      
		      break;
		    }
		  case STORE_QUEUE_VALID_NO_SPECULATIVE    :
		  case STORE_QUEUE_COMMIT                  :
		    {
		      throw ERRORMORPHEO(FUNCTION,"<Load_store_unit::function_speculative_load_access_transition> Invalid state and operation");
		    }
		  }

		_store_queue [index]._state     = new_state;
		_store_queue [index]._exception = new_exception;
		
		if (update_info == true)
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * Update information");

		    _store_queue [index]._context_id           = (not _param->_have_port_context_id   )?0:PORT_READ(in_MEMORY_IN_CONTEXT_ID   [internal_MEMORY_IN_PORT]);
		    _store_queue [index]._front_end_id         = (not _param->_have_port_front_end_id )?0:PORT_READ(in_MEMORY_IN_FRONT_END_ID [internal_MEMORY_IN_PORT]);
		    _store_queue [index]._ooo_engine_id        = (not _param->_have_port_ooo_engine_id)?0:PORT_READ(in_MEMORY_IN_OOO_ENGINE_ID[internal_MEMORY_IN_PORT]);
		    _store_queue [index]._packet_id            = (not _param->_have_port_rob_ptr      )?0:PORT_READ(in_MEMORY_IN_PACKET_ID    [internal_MEMORY_IN_PORT]);
		    _store_queue [index]._operation            = operation;
		    _store_queue [index]._cancel               = cancel   ;
		    _store_queue [index]._load_queue_ptr_write = (not _param->_have_port_load_queue_ptr)?0:PORT_READ(in_MEMORY_IN_LOAD_QUEUE_PTR_WRITE[internal_MEMORY_IN_PORT]);
		    _store_queue [index]._address              = address;
		    _store_queue [index]._send_commit          = false;

		    // reordering data
		    _store_queue [index]._wdata                = duplicate<Tgeneral_data_t>(_param->_size_general_data,PORT_READ(in_MEMORY_IN_DATA_RB[internal_MEMORY_IN_PORT]), memory_size(operation), 0); 
//                  _store_queue [index]._num_reg_rd           = PORT_READ(in_MEMORY_IN_NUM_REG_RD  [internal_MEMORY_IN_PORT]);
		  }
	      }
	    else
	      {
		// ====================================
		// ===== SPECULATIVE_ACCESS_QUEUE =====
		// ====================================

		// In speculative access queue, they are many type's request
		log_printf(TRACE,Load_store_unit,FUNCTION,"    * speculative_access_queue");
		log_printf(TRACE,Load_store_unit,FUNCTION,"      * PUSH");
		
		// Write in reservation station
		uint32_t     index = _speculative_access_queue_control->push();

		log_printf(TRACE,Load_store_unit,FUNCTION,"      * index : %d", index);

		Texception_t exception;

		if (exception_alignement == true)
		  exception = EXCEPTION_MEMORY_ALIGNMENT;
		else
		  exception = EXCEPTION_MEMORY_NONE;
				
		// if exception, don't access at the cache
		// NOTE : type "other" (lock, invalidate, flush and sync) can't make an alignement exception (access is equivalent at a 8 bits)
		_speculative_access_queue [index]._state                = (exception == EXCEPTION_MEMORY_NONE)?SPECULATIVE_ACCESS_QUEUE_WAIT_CACHE:SPECULATIVE_ACCESS_QUEUE_WAIT_LOAD_QUEUE;
		_speculative_access_queue [index]._context_id           = (not _param->_have_port_context_id   )?0:PORT_READ(in_MEMORY_IN_CONTEXT_ID   [internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._front_end_id         = (not _param->_have_port_front_end_id )?0:PORT_READ(in_MEMORY_IN_FRONT_END_ID [internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._ooo_engine_id        = (not _param->_have_port_ooo_engine_id)?0:PORT_READ(in_MEMORY_IN_OOO_ENGINE_ID[internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._packet_id            = (not _param->_have_port_rob_ptr      )?0:PORT_READ(in_MEMORY_IN_PACKET_ID    [internal_MEMORY_IN_PORT]);

		_speculative_access_queue [index]._operation            = operation;
		_speculative_access_queue [index]._cancel               = cancel   ;
		_speculative_access_queue [index]._load_queue_ptr_write = (not _param->_have_port_load_queue_ptr)?0:PORT_READ(in_MEMORY_IN_LOAD_QUEUE_PTR_WRITE[internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._store_queue_ptr_write= PORT_READ(in_MEMORY_IN_STORE_QUEUE_PTR_WRITE[internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._store_queue_ptr_read = PORT_READ(in_MEMORY_IN_STORE_QUEUE_PTR_READ [internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._store_queue_empty    = PORT_READ(in_MEMORY_IN_STORE_QUEUE_EMPTY    [internal_MEMORY_IN_PORT]);
		_speculative_access_queue [index]._address              = address;
		// NOTE : is operation is a load, then they are a result and must write in the register file
		_speculative_access_queue [index]._write_rd             = is_operation_memory_load(operation);
		_speculative_access_queue [index]._num_reg_rd           = PORT_READ(in_MEMORY_IN_NUM_REG_RD  [internal_MEMORY_IN_PORT]);

		_speculative_access_queue [index]._exception            = exception;
	      }
	  }

	//================================================================
	// Interface "DCACHE_REQ"
	//================================================================
        if ((    internal_DCACHE_REQ_VAL  == 1) and
            (PORT_READ(in_DCACHE_REQ_ACK[0]) == 1))
          {
	    log_printf(TRACE,Load_store_unit,FUNCTION,"  * DCACHE_REQ [0]");

	    switch (internal_DCACHE_REQ_SELECT_QUEUE)
	      {
	      case SELECT_STORE_QUEUE :
		{
		  // =======================
		  // ===== STORE_QUEUE =====
		  // =======================
		  
		  // Entry flush and increase the read pointer
		  
		  _store_queue [reg_STORE_QUEUE_PTR_READ]._state = STORE_QUEUE_COMMIT;

#if defined(DEBUG) and defined(DEBUG_Load_store_unit) and (DEBUG_Load_store_unit == true)
		  if (log_file_generate)
	            {
	              // log file
                      Tcontext_t num_thread    = get_num_thread(_store_queue [reg_STORE_QUEUE_PTR_READ]._context_id   , _param->_size_context_id   ,
                                                                _store_queue [reg_STORE_QUEUE_PTR_READ]._front_end_id , _param->_size_front_end_id ,
                                                                _store_queue [reg_STORE_QUEUE_PTR_READ]._ooo_engine_id, _param->_size_ooo_engine_id);

	              memory_log_file [num_thread] 
	            	<< "[" << simulation_cycle() << "] "
	            	<< std::hex
                        << "@ 0x" << std::setfill('0') << std::setw(_param->_size_general_data/4) << _store_queue [reg_STORE_QUEUE_PTR_READ]._address << " -[ Write ]-> 0x" << std::setfill('0') << std::setw(_param->_size_general_data/4) << _store_queue [reg_STORE_QUEUE_PTR_READ]._wdata << " "
	            	<< std::dec
	            	<< "{" << toString(_store_queue [reg_STORE_QUEUE_PTR_READ]._operation) << "}";
	              
	              memory_log_file [num_thread] << std::endl;
	            }
#endif	    


		  break;
		}
	      case SELECT_LOAD_QUEUE_SPECULATIVE :
		{
		  // =========================================
		  // ===== SELECT_LOAD_QUEUE_SPECULATIVE =====
		  // =========================================

		  load_queue_push = true;
		  break;
		}
	      case SELECT_LOAD_QUEUE :
		{
		  throw ERRORMORPHEO(FUNCTION,_("Invalid selection"));
		  break;
		}

		break;
	      }
	  }

	if (load_queue_push)
	  {
            log_printf(TRACE,Load_store_unit,FUNCTION,"  * load_queue_push");

	    Tlsq_ptr_t   ptr_write = _speculative_access_queue[internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._load_queue_ptr_write;
	    Toperation_t operation = _speculative_access_queue[internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._operation;
	    Texception_t exception = _speculative_access_queue[internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._exception;
	    bool         have_exception = (exception != EXCEPTION_MEMORY_NONE);
            bool         need_check= false;
            Tlsq_ptr_t   store_queue_ptr_write = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._store_queue_ptr_write;
            Tlsq_ptr_t   store_queue_ptr_read  = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._store_queue_ptr_read ;
            Tcontrol_t   store_queue_empty     = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._store_queue_empty    ;
	    
	    if (have_exception)
	      _load_queue [ptr_write]._state = LOAD_QUEUE_COMMIT;
	    else
	      {
		if (have_dcache_rsp(operation))
		  {
		    // load and synchronisation
		    if (must_check(operation) and not store_queue_empty)
		      {
			// load
                        need_check = true;
			
			_load_queue [ptr_write]._state = LOAD_QUEUE_WAIT_CHECK;
		      }
		    else
		      {
			// synchronisation
			_load_queue [ptr_write]._state = LOAD_QUEUE_WAIT;
		      }
		  }
		else
		  {
		    // lock, prefecth, flush and invalidate
		    _load_queue [ptr_write]._state = LOAD_QUEUE_COMMIT;
		  }
	      }

	    Tdcache_address_t address        = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._address;
	    Tdcache_address_t address_lsb    = (address & _param->_mask_address_lsb);
	    Tdcache_address_t check_hit_byte = gen_mask_not<Tdcache_address_t>(address_lsb+(memory_size(operation)>>3)-1,address_lsb) & _param->_mask_check_hit_byte;

	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * address                 : 0x%.8x", address);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * address_lsb             : 0x%.8x", address_lsb);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * operation               : %d", operation);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * memory_size             : %d", memory_size(operation));
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * check_hit_byte          : 0x%x", check_hit_byte);

	    _load_queue [ptr_write]._context_id            = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._context_id;
	    _load_queue [ptr_write]._front_end_id          = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._front_end_id;
	    _load_queue [ptr_write]._ooo_engine_id         = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._ooo_engine_id;
	    _load_queue [ptr_write]._packet_id             = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._packet_id;
	    _load_queue [ptr_write]._operation             = operation;
	    _load_queue [ptr_write]._cancel                = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._cancel;
	    _load_queue [ptr_write]._store_queue_ptr_write = store_queue_ptr_write;
	    _load_queue [ptr_write]._store_queue_ptr_read  = store_queue_ptr_read ;
 	    _load_queue [ptr_write]._store_queue_empty     = store_queue_empty    ;
	    _load_queue [ptr_write]._address               = address;
	    _load_queue [ptr_write]._check_hit_byte        = check_hit_byte;
	    _load_queue [ptr_write]._check_hit             = 0;
	    _load_queue [ptr_write]._shift                 = address_lsb<<3;// *8
	    _load_queue [ptr_write]._is_load_signed        = is_operation_memory_load_signed(operation);
	    _load_queue [ptr_write]._access_size           = memory_size(operation);
	    // NOTE : if have an exception, must write in register, because a depend instruction wait the load data.
	    _load_queue [ptr_write]._write_rd              = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._write_rd             ;
	    _load_queue [ptr_write]._num_reg_rd            = _speculative_access_queue [internal_SPECULATIVE_ACCESS_QUEUE_PTR_READ]._num_reg_rd           ;
	    _load_queue [ptr_write]._exception             = exception;
	    _load_queue [ptr_write]._rdata                 = address; // to the exception

            switch (_param->_speculative_commit_predictor_scheme)
              {

              case PREDICTOR_NEVER_TAKE  : {_load_queue [ptr_write]._can_speculative_commit = false; break;}
              case PREDICTOR_ALWAYS_TAKE : {_load_queue [ptr_write]._can_speculative_commit = true ; break;}
              case PREDICTOR_COUNTER     : {_load_queue [ptr_write]._can_speculative_commit = false; break;} // TODO
              default :
                {
                  // throw ERRORMORPHEO(FUNCTION,_("Invalid predictor scheme.\n"));
                  _load_queue [ptr_write]._can_speculative_commit = false;
                  break;
                }
              }
	    
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * speculative_access_queue");
	    log_printf(TRACE,Load_store_unit,FUNCTION,"      * POP[%d]",(*_speculative_access_queue_control)[0]);
	    
	    _speculative_access_queue [(*_speculative_access_queue_control)[0]]._state = SPECULATIVE_ACCESS_QUEUE_EMPTY;
	    
	    _speculative_access_queue_control->pop();

#ifdef STATISTICS
            if (usage_is_set(_usage,USE_STATISTICS))
              (*_stat_nb_inst_load) ++;
#endif

            // Only load need check
            if (need_check//  and not store_queue_empty
		)
//             if (is_operation_memory_load(_load_queue [ptr_write]._operation))
              {
                log_printf(TRACE,Load_store_unit,FUNCTION,"    * update nb_check");
                log_printf(TRACE,Load_store_unit,FUNCTION,"      *     store_queue_ptr_write : %d",store_queue_ptr_write);
                log_printf(TRACE,Load_store_unit,FUNCTION,"      *     store_queue_ptr_read  : %d",store_queue_ptr_read );
		log_printf(TRACE,Load_store_unit,FUNCTION,"      *     store_queue_empty     : %d",store_queue_empty    );
                log_printf(TRACE,Load_store_unit,FUNCTION,"      * reg_STORE_QUEUE_PTR_READ  : %d",reg_STORE_QUEUE_PTR_READ);
                
                uint32_t i=store_queue_ptr_write;
                while (i!=store_queue_ptr_read)
                  {
                    i=((i==0)?_param->_size_store_queue:i)-1;
                    
                    log_printf(TRACE,Load_store_unit,FUNCTION,"      * i                         : %d",i);
                    
                    reg_STORE_QUEUE_NB_CHECK [i] ++;

		    log_printf(TRACE,Load_store_unit,FUNCTION,"              * reg_STORE_QUEUE_NB_CHECK 3 [%d] <- %d", i,reg_STORE_QUEUE_NB_CHECK [i]);
                  }
              }
          }

	//================================================================
	// Interface "DCACHE_RSP"
	//================================================================
        if ((PORT_READ(in_DCACHE_RSP_VAL[0])== 1) and
            (    internal_DCACHE_RSP_ACK == 1))
          {
	    log_printf(TRACE,Load_store_unit,FUNCTION,"  * DCACHE_RSP [0]");

	    // don't use context_id : because there are one queue for all thread
	    //Tcontext_t      context_id = PORT_READ(in_DCACHE_RSP_CONTEXT_ID[0]); 
	    Tpacket_t       packet_id  = PORT_READ(in_DCACHE_RSP_PACKET_ID [0]);
	    Tdcache_data_t  rdata      = PORT_READ(in_DCACHE_RSP_RDATA     [0]);
	    Tdcache_error_t error      = PORT_READ(in_DCACHE_RSP_ERROR     [0]);

	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * original packet_id : %d"  , packet_id);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * packet_id          : %d"  , packet_id>>1);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * rdata              : %.8x", rdata);
	    log_printf(TRACE,Load_store_unit,FUNCTION,"    * error              : %d"  , error);
	    
	    if (DCACHE_RSP_IS_LOAD(packet_id) == 1)
	      {
		packet_id >>= 1;

		log_printf(TRACE,Load_store_unit,FUNCTION,"    * packet is a LOAD");
 
#ifdef DEBUG_TEST
		if (not have_dcache_rsp(_load_queue [packet_id]._operation))
		  throw ERRORMORPHEO(FUNCTION,_("Receive of respons, but the corresponding operation don't wait a respons."));
#endif

                Tdcache_data_t data = _load_queue [packet_id]._rdata;

                log_printf(TRACE,Load_store_unit,FUNCTION,"    * data construction");
                log_printf(TRACE,Load_store_unit,FUNCTION,"      * data from cache     : 0x%.8x",rdata);
                log_printf(TRACE,Load_store_unit,FUNCTION,"      * data (before)       : 0x%.8x", data);
                log_printf(TRACE,Load_store_unit,FUNCTION,"      * check_hit_byte      : 0x%x"  ,_load_queue [packet_id]._check_hit_byte);
                for (uint32_t i=0;i<(_param->_size_general_data>>3)/*8*/; ++i)
                  // Test if this byte has been checked
                  if ((_load_queue [packet_id]._check_hit_byte & (1<<i)) == 0)
                    {
                      log_printf(TRACE,Load_store_unit,FUNCTION,"      * no previous check ]%d:%d]",(i+1)<<3,i<<3);
                      data = insert<Tdcache_data_t>(data,rdata,((i+1)<<3)-1,i<<3);
                    }
                log_printf(TRACE,Load_store_unit,FUNCTION,"      * data (after)        : 0x%.8x", data);
                
                _load_queue [packet_id]._rdata = data;
		
#if defined(DEBUG) and defined(DEBUG_Load_store_unit) and (DEBUG_Load_store_unit == true)
		if (log_file_generate)
		  {
		    // log file
		    Tcontext_t num_thread    = get_num_thread(_load_queue [packet_id]._context_id   , _param->_size_context_id   ,
							      _load_queue [packet_id]._front_end_id , _param->_size_front_end_id ,
							      _load_queue [packet_id]._ooo_engine_id, _param->_size_ooo_engine_id);
		    
		    memory_log_file [num_thread] 
		      << "[" << simulation_cycle() << "] "
		      << std::hex
		      << "@ 0x" << std::setfill('0') << std::setw(_param->_size_general_data/4) << _load_queue [packet_id]._address << " -[ Read  ]-> 0x" << std::setfill('0') << std::setw(_param->_size_general_data/4) << _load_queue [packet_id]._rdata << " "
		      << std::dec
		      << "{" << toString(_load_queue [packet_id]._operation) << "}";
		    
		    memory_log_file [num_thread] << std::endl;
		  }
#endif	    

		if (error != DCACHE_ERROR_NONE)
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"    * have a bus error !!!");

		    _load_queue [packet_id]._exception = EXCEPTION_MEMORY_BUS_ERROR;
		    _load_queue [packet_id]._state     = LOAD_QUEUE_COMMIT;

                    
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_ptr_write : %d", _load_queue[packet_id]._store_queue_ptr_write);
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_ptr_read  : %d", _load_queue[packet_id]._store_queue_ptr_read );
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * store_queue_empty     : %d", _load_queue[packet_id]._store_queue_empty    );

 		    if (not _load_queue[packet_id]._store_queue_empty)
		      {
			uint32_t i=_load_queue[packet_id]._store_queue_ptr_write;
			
			while (i!=_load_queue[packet_id]._store_queue_ptr_read)
			  {
			    i=((i==0)?_param->_size_store_queue:i)-1;
			    
#ifdef DEBUG
			    if (reg_STORE_QUEUE_NB_CHECK [i] == 0)
			      throw ERRORMORPHEO(FUNCTION,_("reg_STORE_QUEUE_NB_CHECK must be > 0\n"));
#endif
			    
			    reg_STORE_QUEUE_NB_CHECK [i] --;

			    log_printf(TRACE,Load_store_unit,FUNCTION,"              * reg_STORE_QUEUE_NB_CHECK 4 [%d] <- %d", i,reg_STORE_QUEUE_NB_CHECK [i]);

			    //i=(i+1)%_param->_size_store_queue;
			  }
			_load_queue[packet_id]._store_queue_empty = true; // end of check

		      }
		  }
		else
		  {
		    log_printf(TRACE,Load_store_unit,FUNCTION,"    * have no bus error.");
		    log_printf(TRACE,Load_store_unit,FUNCTION,"      * previous state : %s",toString(_load_queue [packet_id]._state).c_str());

		    // FIXME : convention : if bus error, the cache return the fautive address !
		    // But, the load's address is aligned !

		    switch (_load_queue [packet_id]._state)
		      {
		      case LOAD_QUEUE_WAIT_CHECK : _load_queue [packet_id]._state = LOAD_QUEUE_COMMIT_CHECK; break;
		      case LOAD_QUEUE_WAIT       : _load_queue [packet_id]._state = LOAD_QUEUE_COMMIT      ; break;
		      default : throw ERRORMORPHEO(FUNCTION,_("Illegal state (dcache_rsp).")); break;
		      }
		  }
	      }
	    else
	      {
		log_printf(TRACE,Load_store_unit,FUNCTION,"    * packet is a STORE");
		
		// TODO : les stores ne génére pas de réponse sauf quand c'est un bus error !!!
		throw ERRORMORPHEO(FUNCTION,_("dcache_rsp : no respons to a write. (TODO : manage bus error to the store operation.)"));
	      }
	    
	  }
	
	// this register is to manage the priority of check -> Round robin
	reg_LOAD_QUEUE_CHECK_PRIORITY = (reg_LOAD_QUEUE_CHECK_PRIORITY+1)%_param->_size_load_queue;
	
	
#if defined(DEBUG) and (DEBUG>=DEBUG_TRACE)
	// ***** dump store queue
        log_printf(TRACE,Load_store_unit,FUNCTION,"  * Dump STORE_QUEUE");
        log_printf(TRACE,Load_store_unit,FUNCTION,"    * ptr_read : %d",reg_STORE_QUEUE_PTR_READ);
	
	for (uint32_t i=0; i<_param->_size_store_queue; i++)
	  {
	    uint32_t j = (reg_STORE_QUEUE_PTR_READ+i)%_param->_size_store_queue;

            log_printf(TRACE,Load_store_unit,FUNCTION,"    [%.4d] %.4d %.4d %.4d, %.4d, %.4d %.1d, %.4d, %.8x %.8x, %.2d %.1d, %.2d %s",
                       j,
                       _store_queue[j]._context_id          ,
                       _store_queue[j]._front_end_id        ,
                       _store_queue[j]._ooo_engine_id       ,
                       _store_queue[j]._packet_id           ,
                       _store_queue[j]._operation           ,
                       _store_queue[j]._cancel              ,
                       _store_queue[j]._load_queue_ptr_write,
                       _store_queue[j]._address             ,
                       _store_queue[j]._wdata               ,
                     //_store_queue[j]._write_rd            ,
                     //_store_queue[j]._num_reg_rd          ,
                       _store_queue[j]._exception           ,
                       _store_queue[j]._send_commit         ,
                       reg_STORE_QUEUE_NB_CHECK  [j]        ,
                       toString(_store_queue[j]._state).c_str());
	  }

	// ***** dump speculative_access queue
	log_printf(TRACE,Load_store_unit,FUNCTION,"  * Dump SPECULATIVE_ACCESS_QUEUE");
	
	for (uint32_t i=0; i<_param->_size_speculative_access_queue; i++)
	  {
	    uint32_t j = (*_speculative_access_queue_control)[i];

            log_printf(TRACE,Load_store_unit,FUNCTION,"    [%.4d] %.4d %.4d %.4d, %.4d, %.4d %.1d, %.4d %.4d %.4d %.1d, %.8x, %.1d %.4d, %.2d, %s",
                       j,
                       _speculative_access_queue[j]._context_id          ,
                       _speculative_access_queue[j]._front_end_id        ,
                       _speculative_access_queue[j]._ooo_engine_id       ,
                       _speculative_access_queue[j]._packet_id           ,
                       _speculative_access_queue[j]._operation           ,
                       _speculative_access_queue[j]._cancel              ,
                       _speculative_access_queue[j]._load_queue_ptr_write,
                       _speculative_access_queue[j]._store_queue_ptr_write,
                       _speculative_access_queue[j]._store_queue_ptr_read ,
                       _speculative_access_queue[j]._store_queue_empty    ,
                       _speculative_access_queue[j]._address             ,
                       _speculative_access_queue[j]._write_rd            ,
                       _speculative_access_queue[j]._num_reg_rd          ,
                       _speculative_access_queue[j]._exception           ,
                       toString(_speculative_access_queue[j]._state).c_str());
	  }

	// ***** dump load queue
	log_printf(TRACE,Load_store_unit,FUNCTION,"  * Dump LOAD_QUEUE");
        log_printf(TRACE,Load_store_unit,FUNCTION,"    * ptr_read_check_priority : %d",reg_LOAD_QUEUE_CHECK_PRIORITY);
	
	for (uint32_t i=0; i<_param->_size_load_queue; i++)
	  {
	    uint32_t j = i;

            log_printf(TRACE,Load_store_unit,FUNCTION,"    [%.4d] %.4d %.4d %.4d, %.4d, %.4d %.1d, %.4d %.4d %.1d, %.8x %.1x %.1d %.2d %.1d %.2d, %.8x, %.1d %.4d, %.2d, %s",
                       j,
                       _load_queue[j]._context_id          ,
                       _load_queue[j]._front_end_id        ,
                       _load_queue[j]._ooo_engine_id       ,
                       _load_queue[j]._packet_id           ,
                       _load_queue[j]._operation           ,
                       _load_queue[j]._cancel              ,
                       _load_queue[j]._store_queue_ptr_write,
                       _load_queue[j]._store_queue_ptr_read ,
		       _load_queue[j]._store_queue_empty    ,
                       _load_queue[j]._address             ,
                       _load_queue[j]._check_hit_byte      , 
                       _load_queue[j]._check_hit           ,
                       _load_queue[j]._shift               ,
                       _load_queue[j]._is_load_signed      ,
                       _load_queue[j]._access_size         ,
                       _load_queue[j]._rdata               ,
                       _load_queue[j]._write_rd            ,
                       _load_queue[j]._num_reg_rd          ,
                       _load_queue[j]._exception           ,
                       toString(_load_queue[j]._state).c_str());
	  }
#endif
	
#ifdef STATISTICS
        if (usage_is_set(_usage,USE_STATISTICS))
          {
            for (uint32_t i=0; i<_param->_size_store_queue; i++)
              if (_store_queue[i]._state != STORE_QUEUE_EMPTY)
                (*_stat_use_store_queue) ++;
            for (uint32_t i=0; i<_param->_size_speculative_access_queue; i++)
              if (_speculative_access_queue[i]._state != SPECULATIVE_ACCESS_QUEUE_EMPTY)
                (*_stat_use_speculative_access_queue) ++;
            for (uint32_t i=0; i<_param->_size_load_queue; i++)
              if (_load_queue[i]._state != LOAD_QUEUE_EMPTY)
                (*_stat_use_load_queue) ++;
          }
#endif
      }

    log_end(Load_store_unit,FUNCTION);
  };

}; // end namespace load_store_unit
}; // end namespace execute_unit
}; // end namespace multi_execute_unit
}; // end namespace execute_loop
}; // end namespace multi_execute_loop
}; // end namespace core

}; // end namespace behavioural
}; // end namespace morpheo              
#endif