#ifndef _program_hpp_
#define _program_hpp_

#include "config.h"
#include "TestThread.hpp"
#include "EnsembleVariables.hpp"

#include <iostream>
#include <sstream>

using namespace std;

class Program {

   int nb_lines_uncached;
   int nb_threads;
   int tab_size;
   int nb_diff_ML, nb_diff_CL, line_size, cache_lines;
   EnsembleVariables * E;
   TestThread** threads;

   public:

   Program(int nb_procs, int nb_diff_ML, int nb_diff_CL, int nb_max_trans, int nb_max_insts, int line_size, int cache_lines) {
      nb_lines_uncached = 1;
      nb_threads = nb_procs;
      threads = new TestThread*[nb_threads];
      tab_size = (cache_lines * line_size) * ((int) ceil((double) (nb_diff_ML + nb_lines_uncached) / (double) nb_diff_CL) - 1) + (nb_diff_CL * line_size);
      this->nb_diff_ML = nb_diff_ML;
      this->nb_diff_CL = nb_diff_CL;
      this->line_size = line_size;
      this->cache_lines = cache_lines;

      E = new EnsembleVariables(nb_diff_ML * line_size, nb_procs, nb_diff_CL, cache_lines, line_size, nb_lines_uncached, tab_size); // tab_size passed for format printing

      E->setnPubRO(nb_diff_ML);
      E->setnPubWO(nb_diff_ML);
      E->setnPrivate(nb_diff_ML);
      E->setnPrivROWO();

      for (int i = 0; i < nb_threads; i++) {
         threads[i] = new TestThread(nb_diff_ML, nb_diff_CL, nb_max_trans, nb_max_insts, line_size, cache_lines, E, i);
      }
   }

   ~Program() {
      for (int i = 0; i < nb_threads; i++) {
         delete threads[i];
      }
      delete [] threads;
      delete E;
   }

   string writeOutput() {
      stringstream res;
      res << endl;
      res << "#include <pthread.h>" << endl;
      res << "#include <stdlib.h>" << endl;
      res << "#include <stdio.h>" << endl;
      res << endl;

      res << "#ifdef _ALMOS_" << endl;
      res << "#define rd_wr_unc(index) ({              \\" << endl;
      res << "   asm volatile(                         \\" << endl;
      res << "       \"li   $8, \" #index \"\\n\"           \\" << endl;
      res << "       \"sll  $8, $8, 2\\n\"                \\" << endl;
      res << "       \"addu $8, $8, %0\\n\"               \\" << endl;
      res << "       \"li   $10, 0xC\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       \"lw   $9, 0($8)\\n\"                \\" << endl;
      res << "       \"add  $9, $9, 1\\n\"                \\" << endl;
      res << "       \"sw   $9, 0($8)\\n\"                \\"<< endl;
      res << "       \"li   $10, 0xF\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       :                                 \\" << endl;
      res << "       : \"r\" (tab)                       \\" << endl;
      res << "       : \"$8\", \"$9\", \"$10\");             \\" << endl;
      res << "})" << endl;
      res << endl;

      res << "#define wr_unc(index) ({                 \\" << endl;
      res << "   asm volatile(                         \\" << endl;
      res << "       \"li   $8, \" #index \"\\n\"           \\" << endl;
      res << "       \"sll  $8, $8, 2\\n\"                \\" << endl;
      res << "       \"addu $8, $8, %0\\n\"               \\" << endl;
      res << "       \"li   $10, 0xC\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       \"sw   $0, 0($8)\\n\"                \\"<< endl;
      res << "       \"li   $10, 0xF\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       :                                 \\" << endl;
      res << "       : \"r\" (tab)                       \\" << endl;
      res << "       : \"$8\", \"$10\");                   \\" << endl;
      res << "})" << endl;
      res << endl;

      res << "#define rd_unc(index) ({                 \\" << endl;
      res << "   asm volatile(                         \\" << endl;
      res << "       \"li   $8, \" #index \"\\n\"           \\" << endl;
      res << "       \"sll  $8, $8, 2\\n\"                \\" << endl;
      res << "       \"addu $8, $8, %0\\n\"               \\" << endl;
      res << "       \"li   $10, 0xC\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       \"lw   $9, 0($8)\\n\"                \\" << endl;
      res << "       \"li   $10, 0xF\\n\"                 \\" << endl;
      //res << "       \"mtc2 $10, $1\\n\"                  \\" << endl;
      res << "       :                                 \\" << endl;
      res << "       : \"r\" (tab)                       \\" << endl;
      res << "       : \"$8\", \"$9\", \"$10\");             \\" << endl;
      res << "})" << endl;
      res << endl;
      res << "#else" << endl;
      res << "   #define rd_wr_unc(index) ({ tab[index]++; })" << endl;
      res << "   #define rd_unc(index) ({ local_var = tab[index]; })" << endl;
      res << "   #define wr_unc(index) ({ tab[index] = 0; })" << endl;
      res << "#endif" << endl;

      res << endl;
      res << "#define NB_THREADS " << nb_threads << endl;
      res << endl;
      res << "/* NB_MAX_TRANS : " << NB_MAX_TRANS << endl;
      res << " * NB_MAX_INSTS : " << NB_MAX_INSTS << endl;
      res << " * LINE_SIZE    : " << LINE_SIZE << endl;
      res << " * CACHE_LINES  : " << CACHE_LINES << endl;
      res << " */" << endl;
      res << endl;
      res << "volatile int tab[" << tab_size << "];" << endl;
      res << endl;
      res << "pthread_spinlock_t lock_tab[" << nb_diff_ML * line_size << "];" << endl;
      res << endl;
      res << E->writeVariablesNature();
      res << endl;
      res << endl;

      for (int i = 0; i < nb_threads; i++) {
         res << threads[i]->writeOutput(i);
      }

      res << "int main() {" << endl;
      res << endl;
      res << "   unsigned int start;" << endl;
      res << "   unsigned int end;" << endl;
      res << "   unsigned int result;" << endl;
      res << "   int i;" << endl;
      res << "   void (*main_run) (); // fonction run du main" << endl;
      res << endl;
      res << "   #ifdef _ALMOS_" << endl;
      res << "      printf(\"\\n\");" << endl;
      res << "   #endif" << endl;
      res << endl;
      res << "   printf(\"dans le main\\n\");" << endl;
      res << endl;
      res << "   for (i = 0; i < " << nb_diff_ML * line_size << "; i++) {" << endl;
      res << "      pthread_spin_init(&lock_tab[i], 0);" << endl;
      res << "   }" << endl;
      res << "   pthread_t** threads;" << endl;
      res << "   threads = malloc(sizeof(pthread_t *) * NB_THREADS);" << endl;
      res << endl;
      res << "   pthread_attr_t* attr;" << endl;
      res << "   attr = malloc(sizeof(pthread_attr_t) * NB_THREADS);" << endl;
      res << endl;
      res << "   int my_cpu;" << endl;
      res << "   #ifdef _ALMOS_" << endl;
      res << "      pthread_attr_getcpuid_np(&my_cpu);" << endl;
      res << "   #else" << endl;
      res << "      my_cpu = NB_THREADS - 1;" << endl;
      res << "   #endif" << endl;
      res << "   for (i = 0; i < NB_THREADS; i++) {" << endl;
      res << "      if (i != my_cpu) { " << endl;
      res << "         threads[i] = malloc(sizeof(pthread_t));" << endl;
      res << "         pthread_attr_init(&attr[i]);" << endl;
      res << "         #ifdef _ALMOS_" << endl;
      res << "            pthread_attr_setcpuid_np(&attr[i], i, NULL);" << endl;
      res << "         #endif" << endl;
      res << "      } " << endl;
      res << "   }" << endl;
      res << endl;

      int index;
      for (int i = 0; i < nb_diff_ML * line_size; i++) {
         index = index2realindex(i,nb_diff_CL,cache_lines,line_size);
         res << "   tab[" << index << "] = " << index << ";" << endl;
      }
      res << endl;
      for (int i = 0; i < nb_threads; i++) {
        res << "   run" << i << "();" << endl;
      }
      res << endl;
      for (int i = 0; i < nb_diff_ML * line_size; i++) {
         index = index2realindex(i,nb_diff_CL,cache_lines,line_size);
         res << "   tab[" << index << "] = " << index << ";" << endl;
      }

      //res << "   printf("%s --> create run\n",__func__);" << endl;
      //res << "   start = GetTimeUS();" << endl;
      res << endl;
      res << "   for (i = 0; i < NB_THREADS; i++) {" << endl;
      res << "      if (i == 0) {" << endl;
      res << "         if (i != my_cpu) { " << endl;
      res << "            int error = pthread_create(threads[i], &attr[i], (void *) run0, 0);" << endl;
      res << "            if (error != 0) {" << endl;
      res << "               printf(\"*** Error in pthread_create\\n\");" << endl;
      res << "            }" << endl;
      res << "         }" << endl;
      res << "         else {" << endl;
      res << "            main_run = run0;" << endl;
      res << "         }" << endl;
      res << "      }" << endl;
      for (int i = 1; i < nb_threads; i++) {
         res << "      else if (i == " << i << ") {" << endl;
         res << "         if (i != my_cpu) { " << endl;
         res << "            int error = pthread_create(threads[i], &attr[i], (void *) run" << i << ", 0);" << endl;
         res << "            if (error != 0) {" << endl;
         res << "               printf(\"*** Error in pthread_create\\n\");" << endl;
         res << "            }" << endl;
         res << "         }" << endl;
         res << "         else {" << endl;
         res << "            main_run = run" << i << ";" << endl;
         res << "         }" << endl;
         res << "      }" << endl;
      }
      res << "      else {" << endl;
      res << "         printf(\"Erreur : pas de fonction run correspondant au threads/processeur %d\\n\",i);" << endl;
      res << "      }" << endl;
      res << "   }" << endl;
      res << endl;
      res << "   main_run();" << endl;
      res << endl;
      res << "   for (i = 0; i < NB_THREADS; i++) {" << endl;
      res << "      if (i != my_cpu) {" << endl;
      res << "         pthread_join(*threads[i], NULL);" << endl;
      res << "      }" << endl;
      res << "   }" << endl;
      res << endl;
      //res << "   end = GetTimeUS();" << endl;
      //res << "   printf("%s <-- join run\n",__func__);" << endl;
      res << endl;
      res << "   result = end - start;" << endl;
      //res << "   printf(\"temps ecoule : %u\n\",result);" << endl;

      for (int i = 0; i < nb_diff_ML*line_size; i++) {
         index = index2realindex(i,nb_diff_CL,cache_lines,line_size);
         res << "   printf(\"tab[" << index << "] final : %d\\n\",tab[" << index << "]);" << endl;
      }
      res << "   #ifdef _ALMOS_" << endl;
      res << "      *(int *) 0x0 = 0xDEADDEAD;" << endl;
      res << "   #endif" << endl;
      res << endl;
      res << "   return 0;" << endl;
      res << "}   " << endl;
      return res.str();
    }

};

#endif