/*
 * Adapted from the block TEA reference code in the "Tea Extensions" paper by
 * Roger M. Needham and David J. Wheeler, which is available at:
 * http://www.cl.cam.ac.uk/ftp/users/djw3/xtea.ps
 */

#include "arch.h"
#include "util.h"
#include "btea.h"

void btea(u32 *v, u32 *k, u32 len, int decode, int swap)
{
	u32 e, p, sum = 0, z, diff;
	s32 q;

	q = 6 + 52 / len;

	if(decode) {
		sum = q * DELTA;
		if(swap) {
			while(sum != 0) {
				e = sum >> 2 & 3;
				for(p = len - 1; p > 0; p--) {
					z = swapu32(v[p - 1]);
					diff = ((z << 4 ^ z >> 5) + z) ^
						(k[(p & 3) ^ e] + sum);
					v[p] = swapu32(swapu32(v[p]) - diff);
				}
				z = swapu32(v[len - 1]);
				diff = ((z << 4 ^ z >> 5) + z) ^
					(k[(p & 3) ^ e] + sum);
				v[0] = swapu32(swapu32(v[0]) - diff);
				sum -= DELTA;
			}
		} else {
			while(sum != 0) {
				e = sum >> 2 & 3;
				for(p = len - 1; p > 0; p--) {
					z = v[p - 1];
					v[p] -= ((z << 4 ^ z >> 5) + z) ^
						(k[(p & 3) ^ e] + sum);
				}
				z = v[len - 1];
				v[0] -= ((z << 4 ^ z >> 5) + z) ^
						(k[(p & 3) ^ e] + sum);
				sum -= DELTA;
			}
		}
	} else {
		if(swap) {
			z = swapu32(v[len - 1]);
			while(q-- > 0) {
				sum += DELTA;
				e = sum >> 2 & 3;
				for(p = 0; p < len; p++) {
					diff = ((z << 4 ^ z >> 5) + z) ^
							(k[(p & 3) ^ e] + sum);
					z = swapu32(v[p]) + diff;
					v[p] = swapu32(z);
				}
			}
		} else {
			z = v[len - 1];
			while(q-- > 0) {
				sum += DELTA;
				e = sum >> 2 & 3;
				for(p = 0; p < len; p++) {
					z = v[p] += ((z << 4 ^ z >> 5) + z) ^
							(k[(p & 3) ^ e] + sum);
				}
			}
		}
	}
}