#ifndef LOADSTORE_H_
#define LOADSTORE_H_

#include <stdint.h>

#include "config.h"
#include "endian.h"
#include "word.h"

/* 64-bit LSB mask (undefined for n == 0) */
#define MASK(n) (~0ull >> (64 - (n)))

/* get byte from Ascon 64-bit word */
#define GETBYTE(x, i) ((uint8_t)((uint64_t)(x) >> (56 - 8 * (i))))

/* set byte in Ascon 64-bit word */
#define SETBYTE(b, i) ((uint64_t)(b) << (56 - 8 * (i)))

#if ASCON_DATA_ACCESS == 'W'

#ifndef NDEBUG
#pragma message("Using wordwise data access")
#endif

__forceinline word_t LOAD64(const uint8_t* bytes) {
  uint64_t x = U64BIG(*(uint64_t*)bytes);
  return U64TOWORD(x);
}

__forceinline void STORE64(uint8_t* bytes, word_t w) {
  uint64_t x = WORDTOU64(w);
  *(uint64_t*)bytes = U64BIG(x);
}

__forceinline word_t LOAD(const uint8_t* bytes, int n) {
  uint64_t x = U64BIG(*(uint64_t*)bytes & MASK(8 * n));
  return U64TOWORD(x);
}

__forceinline void STORE(uint8_t* bytes, word_t w, int n) {
  uint64_t x = WORDTOU64(w);
  *(uint64_t*)bytes &= ~MASK(8 * n);
  *(uint64_t*)bytes |= U64BIG(x);
}

#elif ASCON_DATA_ACCESS == 'M'

#ifndef NDEBUG
#pragma message("Using memcpy to access data")
#endif

#include <string.h>

#define LOAD64(bytes) LOAD(bytes, 8)
#define STORE64(bytes, w) STORE(bytes, w, 8)

__forceinline word_t LOAD(const uint8_t* bytes, int n) {
  uint64_t x = 0;
  memcpy((uint8_t*)&x, bytes, n);
  return U64TOWORD(U64BIG(x));
}

__forceinline void STORE(uint8_t* bytes, word_t w, int n) {
  uint64_t x = U64BIG(WORDTOU64(w));
  memcpy(bytes, (uint8_t*)&x, n);
}

#elif ASCON_DATA_ACCESS == 'B'

#ifndef NDEBUG
#pragma message("Using bytewise data access")
#endif

#define LOAD64(bytes) LOAD(bytes, 8)
#define STORE64(bytes, w) STORE(bytes, w, 8)

__forceinline word_t LOAD(const uint8_t* bytes, int n) {
  uint64_t x = 0;
  for (int i = 56; i >= 64 - n * 8; i -= 8) x |= (uint64_t)*bytes++ << i;
  return U64TOWORD(x);
}

__forceinline void STORE(uint8_t* bytes, word_t w, int n) {
  uint64_t x = WORDTOU64(w);
  for (int i = 56; i >= 64 - n * 8; i -= 8) *bytes++ = x >> i;
}

#elif ASCON_DATA_ACCESS == 'H'

#ifndef NDEBUG
#pragma message("Using hybrid data access")
#endif

#define LOAD64(bytes) LOAD(bytes, 8)
#define STORE64(bytes, w) STORE(bytes, w, 8)

__forceinline word_t LOAD(const uint8_t* bytes, int n) {
  uint64_t x = 0;
  if (n == 8)
    x = U64BIG(*(uint64_t*)bytes);
  else
    for (int i = 56; i >= 64 - n * 8; i -= 8) x |= (uint64_t)*bytes++ << i;
  return U64TOWORD(x);
}

__forceinline void STORE(uint8_t* bytes, word_t w, int n) {
  uint64_t x = WORDTOU64(w);
  if (n == 8)
    *(uint64_t*)bytes = U64BIG(x);
  else
    for (int i = 56; i >= 64 - n * 8; i -= 8) *bytes++ = x >> i;
}

#else
#error "Ascon data access macro not defined correctly"
#endif

#endif /* LOADSTORE_H_ */