#include <stdint.h>
#include "fe.h"

int x25519(unsigned char *q, const unsigned char *n, const unsigned
		char *p) {
	unsigned char e[32];
	unsigned int i;
	int32_t x1[10];
	int32_t x2[10];
	int32_t z2[10];
	int32_t x3[10];
	int32_t z3[10];
	int32_t tmp0[10];
	int32_t tmp1[10];
	int pos;
	unsigned int swap;
	unsigned int b;

	for (i = 0;i < 32;++i) e[i] = n[i];
	e[0] &= 248;
	e[31] &= 127;
	e[31] |= 64;
	fe_frombytes(x1,p);
	fe_1(x2);
	fe_0(z2);
	fe_copy(x3,x1);
	fe_1(z3);

	swap = 0;
	for (pos = 254;pos >= 0;--pos) {
		b = e[pos / 8] >> (pos & 7);
		b &= 1;
		swap ^= b;
		fe_cswap(x2,x3,swap);
		fe_cswap(z2,z3,swap);
		swap = b;

		fe_sub(tmp0,x3,z3); /* qhasm: D = X3-Z3 */
		fe_sub(tmp1,x2,z2); /* qhasm: B = X2-Z2 */
		fe_add(x2,x2,z2); /* qhasm: A = X2+Z2 */
		fe_add(z2,x3,z3); /* qhasm: C = X3+Z3 */
		fe_mul(z3,tmp0,x2); /* qhasm: DA = D*A */
		fe_mul(z2,z2,tmp1); /* qhasm: CB = C*B */
		fe_sq(tmp0,tmp1); /* qhasm: BB = B^2 */
		fe_sq(tmp1,x2); /* qhasm: AA = A^2 */
		fe_add(x3,z3,z2); /* qhasm: t0 = DA+CB */
		/* qhasm: assign x3 to t0 */
		fe_sub(z2,z3,z2); /* qhasm: t1 = DA-CB */
		fe_mul(x2,tmp1,tmp0); /* qhasm: X4 = AA*BB */
		fe_sub(tmp1,tmp1,tmp0); /* qhasm: E = AA-BB */
		fe_sq(z2,z2); /* qhasm: t2 = t1^2 */
		fe_mul121666(z3,tmp1); /* qhasm: t3 = a24*E */
		fe_sq(x3,x3); /* qhasm: X5 = t0^2 */
		fe_add(tmp0,tmp0,z3); /* qhasm: t4 = BB+t3 */
		fe_mul(z3,x1,z2); /* qhasm: Z5 = X1*t2 */
		fe_mul(z2,tmp1,tmp0); /* qhasm: Z4 = E*t4 */
	}
	fe_cswap(x2,x3,swap);
	fe_cswap(z2,z3,swap);

	fe_invert(z2,z2);
	fe_mul(x2,x2,z2);
	fe_tobytes(q,x2);
	return 0;
}