#include"auxFormat.h"

#define aead_RATE (128 / 8)
#define PR0_ROUNDS 100
#define PR_ROUNDS 52
#define PRF_ROUNDS 56
unsigned char  constant7Format_aead[127] = { 
	/*constant7_aead_256*/
0x1,
0x4,
0x10,
0x40,
0x2,
0x8,
0x21,
0x5,
0x14,
0x50,
0x42,
0xa,
0x29,
0x24,
0x11,
0x44,
0x12,
0x48,
0x23,
0xd,
0x35,
0x55,
0x56,
0x5a,
0x6b,
0x2e,
0x38,
0x60,
0x3,
0xc,
0x31,
0x45,
0x16,
0x58,
0x63,
0xf,
0x3d,
0x74,
0x53,
0x4e,
0x3b,
0x6c,
0x32,
0x49,
0x27,
0x1d,
0x75,
0x57,
0x5e,
0x7b,
0x6e,
0x3a,
0x68,
0x22,
0x9,
0x25,
0x15,
0x54,
0x52,
0x4a,
0x2b,
0x2c,
0x30,
0x41,
0x6,
0x18,
0x61,
0x7,
0x1c,
0x71,
0x47,
0x1e,
0x79,
0x66,
0x1b,
0x6d,
0x36,
0x59,
0x67,
0x1f,
0x7d,
0x76,
0x5b,
0x6f,
0x3e,
0x78,
0x62,
0xb,
0x2d,
0x34,
0x51,
0x46,
0x1a,
0x69,
0x26,
0x19,
0x65,
0x17,
0x5c,
0x73,
0x4f,
0x3f,
0x7c,
0x72,
0x4b,
0x2f,
0x3c,
0x70,
0x43,
0xe,
0x39,
0x64,
0x13,
0x4c,
0x33,
0x4d,
0x37,
0x5d,
0x77,
0x5f,
0x7f,
0x7e,
0x7a,
0x6a,
0x2a,
0x28,
0x20, 
};




/* State
 * w12 w8  w4  w0
 * w13 w9  w5  w1
 * w14 w10 w6  w2
 * w15 w11 w7  w3
 *
 */
 static void permutation512(unsigned int *in, int rounds, unsigned char *rc) {
	uint32_t w0, w1, w2, w3, w4, w5, w6, w7;
	uint32_t w8, w9, w10, w11, w12, w13, w14, w15;
	uint32_t s0, s1, s2,s3;
	uint32_t i=0;
	__asm volatile(
		"ldr     w0,     [in]          \n\t"
		"ldr     w4,     [in, #4]      \n\t"
		"ldr     w8,     [in, #8]      \n\t"
		"ldr     w12,    [in, #12]     \n\t"
		"ldr     w1,     [in, #16]     \n\t"
		"ldr     w5,     [in, #20]     \n\t"
		"ldr     w9,     [in, #24]     \n\t"
		"ldr     w13,    [in, #28]     \n\t"
		"ldr     w2,     [in, #32]     \n\t"
		"ldr     w6,     [in, #36]     \n\t"
		"ldr     w10,    [in, #40]     \n\t"
		"ldr     w14,    [in, #44]     \n\t"
		"ldr     w3,     [in, #48]     \n\t"
		"ldr     w7,     [in, #52]     \n\t"
		"ldr     w11,    [in, #56]     \n\t"
		"ldr     w15,    [in, #60]     \n\t"
	"enc_loop:                       \n\t"
    "/*add round const   s0 s1 s2 s3*/           \n\t"
		"ldrb    s3,     [rc]          \n\t"	

		"LSR     s0,     s3, #6       \n\t"	
		"and    s0,     s0, 0x3        \n\t"

		"LSR     s1,     s3, #4       \n\t"	
		"and    s1,     s1, 0x3        \n\t"

		"LSR     s2,     s3, #2       \n\t"	
		"and    s2,     s2, 0x3        \n\t"

		"and    s3,     s3, 0x3        \n\t"

	  "eors    w12,     w12, s0        \n\t"
	  "eors    w8,     w8, s1        \n\t"
	  "eors    w4,     w4, s2        \n\t"
	  "eors    w0,     w0, s3        \n\t"
  "/*sbox first column*/         \n\t"
		"mvns    w0,     w0            \n\t"
		"ands    s0,     w1, w0        \n\t"
		"eors    s0,     w2, s0        \n\t"
		"orrs    w2,     w1, w2        \n\t"
		"eors    w0,     w3, w0        \n\t"
		"eors    w2,     w2, w0        \n\t"
		"eors    s1,     w1, w3        \n\t"
		"eors    w3,     w3, s0        \n\t"
		"ands    w0,     s0, w0        \n\t"
		"eors    w0,     s1, w0        \n\t"
		"ands    w1,     w2, s1        \n\t"
		"eors    w1,     s0, w1        \n\t"
		"/*sbox second column*/        \n\t"
		"mvns    w4,     w4            \n\t"
		"ands    s0,     w5, w4        \n\t"
		"eors    s0,     w6, s0        \n\t"
		"orrs    w6,     w5, w6        \n\t"
		"eors    w4,     w7, w4        \n\t"
		"eors    w6,     w6, w4        \n\t"
		"eors    s1,     w5, w7        \n\t"
		"eors    w7,     w7, s0        \n\t"
		"ands    w4,     s0, w4        \n\t"
		"eors    w4,     s1, w4        \n\t"
		"ands    w5,     w6, s1        \n\t"
		"eors    w5,     s0, w5        \n\t"
		"/*sbox third column*/         \n\t"
		"mvns    w8,     w8            \n\t"
		"ands    s0,     w9,  w8       \n\t"
		"eors    s0,     w10, s0       \n\t"
		"orrs    w10,    w9,  w10      \n\t"
		"eors    w8,     w11, w8       \n\t"
		"eors    w10,    w10, w8       \n\t"
		"eors    s1,     w9,  w11      \n\t"
		"eors    w11,    w11, s0       \n\t"
		"ands    w8,     s0,  w8       \n\t"
		"eors    w8,     s1,  w8       \n\t"
		"ands    w9,     w10, s1       \n\t"
		"eors    w9,     s0,  w9       \n\t"
		"/*sbox forth column*/         \n\t"
		"mvns    w12,    w12           \n\t"
		"ands    s0,     w13, w12      \n\t"
		"eors    s0,     w14, s0       \n\t"
		"orrs    w14,    w13, w14      \n\t"
		"eors    w12,    w15, w12      \n\t"
		"eors    w14,    w14, w12      \n\t"
		"eors    s1,     w13, w15      \n\t"
		"eors    w15,    w15, s0       \n\t"
		"ands    w12,    s0,  w12      \n\t"
		"eors    w12,    s1,  w12      \n\t"
		"ands    w13,    w14, s1       \n\t"
		"eors    w13,    s0,  w13      \n\t"
    "/*rotate shift left 1 bit  [ w13 w9 w5 w1->  w9 w5 w1 (w13,1)] */   \n\t"
		"mov    s0,     w13       \n\t"
		"mov    w13,    w9        \n\t"
		"mov    w9,     w5       \n\t"
		"mov    w5,     w1       \n\t"
		"ROR    w1,     s0 , #31      \n\t"
    "/*rotate shift left 8 bits [w14 w10 w6 w2->£¨w14,4)   £¨w10,4)  (w6,4)  ( w2,4)]*/  \n\t"
		"ROR    w14,    w14 , #28      \n\t"
		"ROR    w10,    w10 , #28      \n\t"
		"ROR    w6,     w6  , #28      \n\t"
		"ROR    w2,     w2  , #28        \n\t"
    "/*rotate shift left 25 bit  [w15 w11 w7 w3-> £¨w11,13)  (w7,14)  ( w3,14)  ( w15,14)] */   \n\t"
		"mov    s0,     w15       \n\t"
		"ROR    w15,    w11  , #26      \n\t"
		"ROR    w11,    w7   , #26      \n\t"
		"ROR    w7 ,    w3   , #26      \n\t"
		"ROR    w3 ,    s0   , #25      \n\t"
		"/*loop control*/              \n\t"
 		"adds    rc,     rc,  #1       \n\t"
		"subs    rounds, rounds, #1    \n\t"
		"bne     enc_loop              \n\t"
		"str     w0,     [in]          \n\t"
		"str     w4,     [in, #4]      \n\t"
		"str     w8,     [in, #8]      \n\t"
		"str     w12,    [in, #12]     \n\t"
		"str     w1,     [in, #16]     \n\t"
		"str     w5,     [in, #20]     \n\t"
		"str     w9,     [in, #24]     \n\t"
		"str     w13,    [in, #28]     \n\t"
		"str     w2,     [in, #32]     \n\t"
		"str     w6,     [in, #36]     \n\t"
		"str     w10,    [in, #40]     \n\t"
		"str     w14,    [in, #44]     \n\t"
		"str     w3,     [in, #48]     \n\t"
		"str     w7,     [in, #52]     \n\t"
		"str     w11,    [in, #56]     \n\t"
		"str     w15,    [in, #60]     \n\t"
	);
}


int crypto_aead_encrypt(
	unsigned char *c, unsigned long long *clen,
	const unsigned char *m, unsigned long long mlen,
	const unsigned char *ad, unsigned long long adlen,
	const unsigned char *nsec,
	const unsigned char *npub,
	const unsigned char *k
) {
	u32 i, j;
	u32 s_temp[16] = { 0 };
	u32 t1, t2, t3, t5, t6, t8, t9, t11;
	// initialization
	u32 s[16] = { 0 };
	u32 dataFormat[4] = { 0 };
	u8 tempData[16] = {0};
	*clen = mlen + CRYPTO_ABYTES;
	//initialization
	packU128FormatToFourPacket(s, npub);
	packU128FormatToFourPacket((s + 4), (npub + 16));
	packU128FormatToFourPacket((s + 8), k);
	packU128FormatToFourPacket((s + 12), (k + 16));
	permutation512(s,PR0_ROUNDS,constant7Format_aead);
	// process associated data
	if (adlen) {
		while (adlen >= aead_RATE) {
			packU128FormatToFourPacket(dataFormat, ad);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			s[3] ^= dataFormat[3];
	permutation512(s,PR_ROUNDS,constant7Format_aead);
			adlen -= aead_RATE;
			ad += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));
		memcpy(tempData, ad, adlen * sizeof(unsigned char));
		tempData[adlen] = 0x01;
		packU128FormatToFourPacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		s[3] ^= dataFormat[3];
	permutation512(s,PR_ROUNDS,constant7Format_aead);
	}
	s[15] ^= 0x80000000;
	if (mlen) {
		while (mlen >= aead_RATE) {
			packU128FormatToFourPacket(dataFormat, m);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			s[3] ^= dataFormat[3];
			unpackU128FormatToFourPacket(c, s);
	permutation512(s,PR_ROUNDS,constant7Format_aead);
			mlen -= aead_RATE;
			m += aead_RATE;
			c += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));
		memcpy(tempData, m, mlen * sizeof(unsigned char));
		tempData[mlen]= 0x01;
		packU128FormatToFourPacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		s[3] ^= dataFormat[3];
		unpackU128FormatToFourPacket(tempData, s);
		memcpy(c, tempData, mlen * sizeof(unsigned char));
		c += mlen;
	}
	// finalization
	permutation512(s,PRF_ROUNDS,constant7Format_aead);
	// return tag
	unpackU128FormatToFourPacket(c, s);
	unpackU128FormatToFourPacket((c+16), (s+4));
	return 0;
}

int crypto_aead_decrypt(
	unsigned char *m, unsigned long long *mlen,
	unsigned char *nsec,
	const unsigned char *c, unsigned long long clen,
	const unsigned char *ad, unsigned long long adlen,
	const unsigned char *npub,
	const unsigned char *k
){
	u32 s_temp[16] = { 0 };
	u32 t1, t2, t3, t5, t6, t8, t9, t11;
	u8 i, j;
	// initialization
	u32 s[16] = { 0 };
	u32 dataFormat[4] = { 0 };
	u32 dataFormat_1[4] = { 0 };
	u32 dataFormat_2[4] = { 0 };
	u8 tempData[16] = { 0 };
	u8 tempU8[64] = { 0 };
	
	if (clen < CRYPTO_ABYTES)
		return -1;
	*mlen = clen - CRYPTO_ABYTES;
	//initialization
	packU128FormatToFourPacket(s, npub);
	packU128FormatToFourPacket((s + 4), (npub + 16));
	packU128FormatToFourPacket((s + 8), k);
	packU128FormatToFourPacket((s + 12), (k + 16));
	permutation512(s,PR0_ROUNDS,constant7Format_aead);
	// process associated data
	if (adlen) {
		while (adlen >= aead_RATE) {
			packU128FormatToFourPacket(dataFormat, ad);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			s[3] ^= dataFormat[3];
	permutation512(s,PR_ROUNDS,constant7Format_aead);
			adlen -= aead_RATE;
			ad += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));

		memcpy(tempData, ad, adlen * sizeof(unsigned char));
		tempData[adlen] = 0x01;
		packU128FormatToFourPacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		s[3] ^= dataFormat[3];
	permutation512(s,PR_ROUNDS,constant7Format_aead);
	}
	s[15] ^= 0x80000000;
	clen = clen - CRYPTO_KEYBYTES;

	if (clen) {
		while (clen >= aead_RATE) {
			packU128FormatToFourPacket(dataFormat_2, c);
			dataFormat_1[0] = s[0] ^ dataFormat_2[0];
			dataFormat_1[1] = s[1] ^ dataFormat_2[1];
			dataFormat_1[2] = s[2] ^ dataFormat_2[2];
			dataFormat_1[3] = s[3] ^ dataFormat_2[3];
			unpackU128FormatToFourPacket(m, dataFormat_1);
			s[0] = dataFormat_2[0];
			s[1] = dataFormat_2[1];
			s[2] = dataFormat_2[2];
			s[3] = dataFormat_2[3];
	permutation512(s,PR_ROUNDS,constant7Format_aead);
			clen -= aead_RATE;
			m += aead_RATE;
			c += aead_RATE;
		}
		unpackU128FormatToFourPacket(tempU8, s);
		for (i = 0; i < clen; ++i, ++m, ++c)
		{
			*m = tempU8[i] ^ *c;
			tempU8[i] = *c;
		}
		tempU8[i] ^= 0x01;
		packU128FormatToFourPacket(s, tempU8);
	}
	// finalization
	permutation512(s,PRF_ROUNDS,constant7Format_aead);
	// return tag	
	packU128FormatToFourPacket(dataFormat, c);
	packU128FormatToFourPacket(dataFormat_1, (c + 16));
	if (dataFormat[0] != s[0] || dataFormat[1] != s[1] || dataFormat[2] != s[2] || dataFormat[3] != s[3]
		|| dataFormat_1[0] != s[4] || dataFormat_1[1] != s[5] || dataFormat_1[2] != s[6] || dataFormat_1[3] != s[7]) {
		return -1;
	}
	return 0;

}