A real implementation would probably be switch-driven, but I whipped up a terse implementation for a big-endian bijective encoder to go with my other comment (tested, but test code omitted):

    int varint_u64_encode(uint8_t** start, const uint8_t* limit, uint64_t value) {
      uint64_t offset;
      uint8_t* position;
      int bytes;

      for (bytes = 1, offset = 0x80;  value >= offset && bytes < 9; ++bytes) {
        offset = (offset << 7 ) | 0x80;
      }
      position = *start;
      value -= ( offset >> 7 ) ^ 1;
      if (position + bytes > limit) {return 0; /* not enough space */}
      *position = (((uint8_t)0xFF) << (9-bytes)) | (uint8_t)(value >> ((bytes-1) * 8 ));
      for (++position; --bytes > 0; ++position) {
        *position = (uint8_t) (value >> ((bytes-1) * 8));
      }
      *start = position;
      return 1;
    }
  
    int varint_u64_decode(const uint8_t** start, const uint8_t* limit, uint64_t* result) {
      uint64_t value;
      uint64_t offset;
      const uint8_t* position;
      int bytes;
      uint8_t mask;
  
      position = *start;  offset = 0;
      for(bytes=1, mask = 0x80; (mask & *position) == mask && bytes < 9; ++bytes) {
        mask = 0x80 | (mask >> 1);
        offset = (offset << 7) | 0x80;
      }
      if (position + bytes > limit) { return 0; /* not enough space */}
      value = (~mask) & *position;
      ++ position;
      for(; --bytes > 0; ++position) {
        value = (value << 8) + *position;
      }
      value += offset;
      if (bytes == 9 && value < (1uLL << 56)) {return 0; /* overflow, non-canonical encoding */ }
      *result = value;
      *start = position;
      return 1;
    }