#include"auxFormat.h"

#define aead_RATE (192 / 8)
#define PR0_ROUNDS 76
#define PR_ROUNDS 28
#define PRF_ROUNDS 32

unsigned char  constant7Format[127] = {
	/*constant7Format[127]:*/
	0x01,0x08,0x40,0x02,0x10,0x80,0x05,0x09,0x48,0x42,0x12,0x90,
	0x85,0x0c,0x41,0x0a,0x50,0x82,0x15,0x89,0x4d,0x4b,0x5a,0xd2,
	0x97,0x9c,0xc4,0x06,0x11,0x88,0x45,0x0b,0x58,0xc2,0x17,0x99,
	0xcd,0x4e,0x53,0x9a,0xd5,0x8e,0x54,0x83,0x1d,0xc9,0x4f,0x5b,
	0xda,0xd7,0x9e,0xd4,0x86,0x14,0x81,0x0d,0x49,0x4a,0x52,0x92,
	0x95,0x8c,0x44,0x03,0x18,0xc0,0x07,0x19,0xc8,0x47,0x1b,0xd8,
	0xc7,0x1e,0xd1,0x8f,0x5c,0xc3,0x1f,0xd9,0xcf,0x5e,0xd3,0x9f,
	0xdc,0xc6,0x16,0x91,0x8d,0x4c,0x43,0x1a,0xd0,0x87,0x1c,0xc1,
	0x0f,0x59,0xca,0x57,0x9b,0xdd,0xce,0x56,0x93,0x9d,0xcc,0x46,
	0x13,0x98,0xc5,0x0e,0x51,0x8a,0x55,0x8b,0x5d,0xcb,0x5f,0xdb,
	0xdf,0xde,0xd6,0x96,0x94,0x84,0x04, };
/* State
 * w8  w4 w0
 * w9  w5 w1
 * w10 w6 w2
 * w11 w7 w3
 */
 static void permutation384(unsigned int *in, int rounds, unsigned char *rc) {

	uint32_t w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11;
	uint32_t s0, s1, s2;
	uint32_t i=0;
	__asm volatile(
		"ldr     w0,     [in]          \n\t"
		"ldr     w4,     [in, #4]      \n\t"
		"ldr     w8,     [in, #8]      \n\t"
		"ldr     w1,     [in, #12]     \n\t"
		"ldr     w5,     [in, #16]     \n\t"
		"ldr     w9,     [in, #20]     \n\t"
		"ldr     w2,     [in, #24]     \n\t"
		"ldr     w6,     [in, #28]     \n\t"
		"ldr     w10,    [in, #32]     \n\t"
		"ldr     w3,     [in, #36]     \n\t"
		"ldr     w7,     [in, #40]     \n\t"
		"ldr     w11,    [in, #44]     \n\t"	
	"enc_loop:                       \n\t"
    "/*add round const   s0 s1*/           \n\t"
		"ldrb    s0,     [rc]          \n\t"	
		"LSR     s1,     s0, #6       \n\t"	
		"and    s1,     s1, 0x3        \n\t"
		"LSR     s2,     s0, #3       \n\t"	
		"and    s2,     s2, 0x7        \n\t"
		"and    s0,     s0, 0x7        \n\t"
	  "eors    w8,     w8, s0        \n\t"
	  "eors    w4,     w4, s2        \n\t"
	  "eors    w0,     w0, s1        \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"
    "/*rotate shift left 1 bit  [w9 w5 w1-> (w1,1) w9 w5] */   \n\t"
		"mov    s0,     w1       \n\t"
		"mov    w1,     w5       \n\t"
		"mov    w5,     w9       \n\t"
		"ROR    w9,     s0, #31        \n\t"
    "/*rotate shift left 8 bits [w10 w6 w2-> £¨w6,3)  (w2,3)  ( w10,2)]*/  \n\t"
		"mov    s0,     w10       \n\t"
		"ROR    w10,    w6 , #29      \n\t"
		"ROR    w6,     w2  , #29      \n\t"
		"ROR    w2,     s0, #30        \n\t"
    "/*rotate shift left 55 bit  [w11 w7 w3-> £¨w3,13)  (w11,14)  ( w7,14)] */   \n\t"
		"mov    s0,     w3       \n\t"
		"ROR    w3,     w7 , #14      \n\t"
		"ROR    w7,     w11 , #14      \n\t"
		"ROR    w11,     s0, #13        \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     w1,     [in, #12]     \n\t"
		"str     w5,     [in, #16]     \n\t"
		"str     w9,     [in, #20]     \n\t"
		"str     w2,     [in, #24]     \n\t"
		"str     w6,     [in, #28]     \n\t"
		"str     w10,    [in, #32]     \n\t"
		"str     w3,     [in, #36]     \n\t"
		"str     w7,     [in, #40]     \n\t"
		"str     w11,    [in, #44]     \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) {
	u8 i; 
	u32 s[12] = { 0 };
	u8 tempData[24] = { 0 };
	u32 dataFormat[6] = { 0 };
	u32 s_temp[12] = { 0 };
	u32 t1, t2, t3, t5, t6, t8, t9, t11;
	u32 t1_32, t2_64, t2_65;
	u32 temp0[3] = { 0 };
	u32 temp1[3] = { 0 };
	u32 temp2[3] = { 0 }; 

	*clen = mlen + CRYPTO_ABYTES;
	// initialization
	packU96FormatToThreePacket(s, npub);
	memcpy(tempData, npub+12, sizeof(unsigned char)*4);
	memcpy(tempData+4, k, sizeof(unsigned char) * 16);
	packU96FormatToThreePacket((s + 3), tempData);
	packU96FormatToThreePacket((s + 6), (tempData+12));

	s[9] = 0x80000000;
	permutation384(s,PR0_ROUNDS,constant7Format);
	// process associated data
	if (adlen) {
		//	rlen = adlen;
		while (adlen >= aead_RATE) {
			packU96FormatToThreePacket(dataFormat, ad);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			packU96FormatToThreePacket((dataFormat+3), (ad+12));
			s[3] ^= dataFormat[3];
			s[4] ^= dataFormat[4];
			s[5] ^= dataFormat[5];
	permutation384(s,PR_ROUNDS,constant7Format);
			adlen -= aead_RATE;
			ad += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));
		memcpy(tempData, ad, adlen * sizeof(unsigned char));
		tempData[adlen] = 0x01;
		packU96FormatToThreePacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		packU96FormatToThreePacket((dataFormat + 3), (tempData + 12));
		s[3] ^= dataFormat[3];
		s[4] ^= dataFormat[4];
		s[5] ^= dataFormat[5];
	permutation384(s,PR_ROUNDS,constant7Format);
	}
	s[9] ^= 0x80000000;
	if (mlen) {
		while (mlen >= aead_RATE) {
			packU96FormatToThreePacket(dataFormat, m);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			packU96FormatToThreePacket((dataFormat + 3), (m + 12));
			s[3] ^= dataFormat[3];
			s[4] ^= dataFormat[4];
			s[5] ^= dataFormat[5];
			unpackU96FormatToThreePacket(c, s);
			unpackU96FormatToThreePacket((c+12), (s+3));
	permutation384(s,PR_ROUNDS,constant7Format);
			mlen -= aead_RATE;
			m += aead_RATE;
			c += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));
		memcpy(tempData, m, mlen * sizeof(unsigned char));
    tempData[mlen]= 0x01;
		packU96FormatToThreePacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		packU96FormatToThreePacket((dataFormat + 3), (tempData + 12));
		s[3] ^= dataFormat[3];
		s[4] ^= dataFormat[4];
		s[5] ^= dataFormat[5];
		unpackU96FormatToThreePacket(tempData, s);
		unpackU96FormatToThreePacket((tempData+12), (s+3));
    memcpy(c, tempData, mlen * sizeof(unsigned char));
		c += mlen;
	}
	// finalization
	permutation384(s,PRF_ROUNDS,constant7Format);
	// return tag
	unpackU96FormatToThreePacket(c, s);
	unpackU96FormatToThreePacket(tempData, (s + 3));
	memcpy(c+12, tempData, sizeof(unsigned char) * 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) {

	u8 i, j;
	u32 s[12] = { 0 };
	u32 s_temp[12] = { 0 };
	u32 dataFormat[12] = { 0 };
	u32 dataFormat_1[12] = { 0 };
	u8 tempData[24] = { 0 };
	u8 tempU8[24] = { 0 };
	u32 t1, t2, t3, t5, t6, t8, t9, t11;
	u32 t1_32, t2_64, t2_65;
	u32 temp0[3] = { 0 };
	u32 temp1[3] = { 0 };
	u32 temp2[3] = { 0 };	*mlen = clen - CRYPTO_ABYTES;
	if (clen < CRYPTO_ABYTES)
		return -1;
	// initialization
	packU96FormatToThreePacket(s, npub);
	memcpy(tempData, npub + 12, sizeof(unsigned char) * 4);
	memcpy(tempData + 4, k, sizeof(unsigned char) * 16);
	packU96FormatToThreePacket((s + 3), tempData);
	packU96FormatToThreePacket((s + 6), (tempData + 12));

	s[9] = 0x80000000;
	permutation384(s,PR0_ROUNDS,constant7Format);
	// process associated data
	if (adlen) {
		while (adlen >= aead_RATE) {
			packU96FormatToThreePacket(dataFormat, ad);
			s[0] ^= dataFormat[0];
			s[1] ^= dataFormat[1];
			s[2] ^= dataFormat[2];
			packU96FormatToThreePacket((dataFormat + 3), (ad + 12));
			s[3] ^= dataFormat[3];
			s[4] ^= dataFormat[4];
			s[5] ^= dataFormat[5];
	permutation384(s,PR_ROUNDS,constant7Format);
			adlen -= aead_RATE;
			ad += aead_RATE;
		}
		memset(tempData, 0, sizeof(tempData));
		memcpy(tempData, ad, adlen * sizeof(unsigned char));
		tempData[adlen] = 0x01;
		packU96FormatToThreePacket(dataFormat, tempData);
		s[0] ^= dataFormat[0];
		s[1] ^= dataFormat[1];
		s[2] ^= dataFormat[2];
		packU96FormatToThreePacket((dataFormat + 3), (tempData + 12));
		s[3] ^= dataFormat[3];
		s[4] ^= dataFormat[4];
		s[5] ^= dataFormat[5];
	permutation384(s,PR_ROUNDS,constant7Format);
	}
	s[9] ^= 0x80000000;
	clen -= CRYPTO_ABYTES;
	if (clen) {
		while (clen >= aead_RATE) {
			packU96FormatToThreePacket(dataFormat, c);
			dataFormat_1[0] = s[0] ^ dataFormat[0];
			dataFormat_1[1] = s[1] ^ dataFormat[1];
			dataFormat_1[2] = s[2] ^ dataFormat[2];
			packU96FormatToThreePacket((dataFormat+3), (c+12));
			dataFormat_1[3] = s[3] ^ dataFormat[3];
			dataFormat_1[4] = s[4] ^ dataFormat[4];
			dataFormat_1[5] = s[5] ^ dataFormat[5];
			unpackU96FormatToThreePacket(m, dataFormat_1);
			unpackU96FormatToThreePacket((m + 12), (dataFormat_1 + 3));
			s[0] = dataFormat[0];
			s[1] = dataFormat[1];
			s[2] = dataFormat[2];
			s[3] = dataFormat[3];
			s[4] = dataFormat[4];
			s[5] = dataFormat[5];
	permutation384(s,PR_ROUNDS,constant7Format);
			clen -= aead_RATE;
			m += aead_RATE;
			c += aead_RATE;
		}
		unpackU96FormatToThreePacket(tempU8, s);
		unpackU96FormatToThreePacket((tempU8+12), (s+3));
		for (i = 0; i < clen; ++i, ++m, ++c)
		{
			*m = tempU8[i] ^ *c;
			tempU8[i] = *c;
		}
		tempU8[i] ^= 0x01;
		packU96FormatToThreePacket(s, tempU8);
		packU96FormatToThreePacket((s + 3), (tempU8 + 12));
	}
	// finalization		
	permutation384(s,PRF_ROUNDS,constant7Format);
	// return tag	

	unpackU96FormatToThreePacket(tempU8, s);
	unpackU96FormatToThreePacket((tempU8+12), (s+3));
	if (U32BIG(((u32*)tempU8)[0]) != U32BIG(((u32*)c)[0]) ||
		U32BIG(((u32*)tempU8)[1]) != U32BIG(((u32*)c)[1]) || 
		U32BIG(((u32*)tempU8)[2]) != U32BIG(((u32*)c)[2]) || 
		U32BIG(((u32*)tempU8)[3]) != U32BIG(((u32*)c)[3]) ){
		return -1;
	}
	return 0;
}