int ecc_sw_hamming_calculate()

in nand/ecc-sw-hamming.c [115:353]


int ecc_sw_hamming_calculate(const unsigned char *buf, unsigned int step_size,
			     unsigned char *code, bool sm_order)
{
	const u32 *bp = (uint32_t *)buf;
	const u32 eccsize_mult = (step_size == 256) ? 1 : 2;
	/* current value in buffer */
	u32 cur;
	/* rp0..rp17 are the various accumulated parities (per byte) */
	u32 rp0, rp1, rp2, rp3, rp4, rp5, rp6, rp7, rp8, rp9, rp10, rp11, rp12,
		rp13, rp14, rp15, rp16, rp17;
	/* Cumulative parity for all data */
	u32 par;
	/* Cumulative parity at the end of the loop (rp12, rp14, rp16) */
	u32 tmppar;
	int i;

	par = 0;
	rp4 = 0;
	rp6 = 0;
	rp8 = 0;
	rp10 = 0;
	rp12 = 0;
	rp14 = 0;
	rp16 = 0;
	rp17 = 0;

	/*
	 * The loop is unrolled a number of times;
	 * This avoids if statements to decide on which rp value to update
	 * Also we process the data by longwords.
	 * Note: passing unaligned data might give a performance penalty.
	 * It is assumed that the buffers are aligned.
	 * tmppar is the cumulative sum of this iteration.
	 * needed for calculating rp12, rp14, rp16 and par
	 * also used as a performance improvement for rp6, rp8 and rp10
	 */
	for (i = 0; i < eccsize_mult << 2; i++) {
		cur = *bp++;
		tmppar = cur;
		rp4 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp6 ^= tmppar;
		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp8 ^= tmppar;

		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		rp6 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp6 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp10 ^= tmppar;

		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		rp6 ^= cur;
		rp8 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp6 ^= cur;
		rp8 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		rp8 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp8 ^= cur;

		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		rp6 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp6 ^= cur;
		cur = *bp++;
		tmppar ^= cur;
		rp4 ^= cur;
		cur = *bp++;
		tmppar ^= cur;

		par ^= tmppar;
		if ((i & 0x1) == 0)
			rp12 ^= tmppar;
		if ((i & 0x2) == 0)
			rp14 ^= tmppar;
		if (eccsize_mult == 2 && (i & 0x4) == 0)
			rp16 ^= tmppar;
	}

	/*
	 * handle the fact that we use longword operations
	 * we'll bring rp4..rp14..rp16 back to single byte entities by
	 * shifting and xoring first fold the upper and lower 16 bits,
	 * then the upper and lower 8 bits.
	 */
	rp4 ^= (rp4 >> 16);
	rp4 ^= (rp4 >> 8);
	rp4 &= 0xff;
	rp6 ^= (rp6 >> 16);
	rp6 ^= (rp6 >> 8);
	rp6 &= 0xff;
	rp8 ^= (rp8 >> 16);
	rp8 ^= (rp8 >> 8);
	rp8 &= 0xff;
	rp10 ^= (rp10 >> 16);
	rp10 ^= (rp10 >> 8);
	rp10 &= 0xff;
	rp12 ^= (rp12 >> 16);
	rp12 ^= (rp12 >> 8);
	rp12 &= 0xff;
	rp14 ^= (rp14 >> 16);
	rp14 ^= (rp14 >> 8);
	rp14 &= 0xff;
	if (eccsize_mult == 2) {
		rp16 ^= (rp16 >> 16);
		rp16 ^= (rp16 >> 8);
		rp16 &= 0xff;
	}

	/*
	 * we also need to calculate the row parity for rp0..rp3
	 * This is present in par, because par is now
	 * rp3 rp3 rp2 rp2 in little endian and
	 * rp2 rp2 rp3 rp3 in big endian
	 * as well as
	 * rp1 rp0 rp1 rp0 in little endian and
	 * rp0 rp1 rp0 rp1 in big endian
	 * First calculate rp2 and rp3
	 */
#ifdef __BIG_ENDIAN
	rp2 = (par >> 16);
	rp2 ^= (rp2 >> 8);
	rp2 &= 0xff;
	rp3 = par & 0xffff;
	rp3 ^= (rp3 >> 8);
	rp3 &= 0xff;
#else
	rp3 = (par >> 16);
	rp3 ^= (rp3 >> 8);
	rp3 &= 0xff;
	rp2 = par & 0xffff;
	rp2 ^= (rp2 >> 8);
	rp2 &= 0xff;
#endif

	/* reduce par to 16 bits then calculate rp1 and rp0 */
	par ^= (par >> 16);
#ifdef __BIG_ENDIAN
	rp0 = (par >> 8) & 0xff;
	rp1 = (par & 0xff);
#else
	rp1 = (par >> 8) & 0xff;
	rp0 = (par & 0xff);
#endif

	/* finally reduce par to 8 bits */
	par ^= (par >> 8);
	par &= 0xff;

	/*
	 * and calculate rp5..rp15..rp17
	 * note that par = rp4 ^ rp5 and due to the commutative property
	 * of the ^ operator we can say:
	 * rp5 = (par ^ rp4);
	 * The & 0xff seems superfluous, but benchmarking learned that
	 * leaving it out gives slightly worse results. No idea why, probably
	 * it has to do with the way the pipeline in pentium is organized.
	 */
	rp5 = (par ^ rp4) & 0xff;
	rp7 = (par ^ rp6) & 0xff;
	rp9 = (par ^ rp8) & 0xff;
	rp11 = (par ^ rp10) & 0xff;
	rp13 = (par ^ rp12) & 0xff;
	rp15 = (par ^ rp14) & 0xff;
	if (eccsize_mult == 2)
		rp17 = (par ^ rp16) & 0xff;

	/*
	 * Finally calculate the ECC bits.
	 * Again here it might seem that there are performance optimisations
	 * possible, but benchmarks showed that on the system this is developed
	 * the code below is the fastest
	 */
	if (sm_order) {
		code[0] = (invparity[rp7] << 7) | (invparity[rp6] << 6) |
			  (invparity[rp5] << 5) | (invparity[rp4] << 4) |
			  (invparity[rp3] << 3) | (invparity[rp2] << 2) |
			  (invparity[rp1] << 1) | (invparity[rp0]);
		code[1] = (invparity[rp15] << 7) | (invparity[rp14] << 6) |
			  (invparity[rp13] << 5) | (invparity[rp12] << 4) |
			  (invparity[rp11] << 3) | (invparity[rp10] << 2) |
			  (invparity[rp9] << 1) | (invparity[rp8]);
	} else {
		code[1] = (invparity[rp7] << 7) | (invparity[rp6] << 6) |
			  (invparity[rp5] << 5) | (invparity[rp4] << 4) |
			  (invparity[rp3] << 3) | (invparity[rp2] << 2) |
			  (invparity[rp1] << 1) | (invparity[rp0]);
		code[0] = (invparity[rp15] << 7) | (invparity[rp14] << 6) |
			  (invparity[rp13] << 5) | (invparity[rp12] << 4) |
			  (invparity[rp11] << 3) | (invparity[rp10] << 2) |
			  (invparity[rp9] << 1) | (invparity[rp8]);
	}

	if (eccsize_mult == 1)
		code[2] =
		    (invparity[par & 0xf0] << 7) |
		    (invparity[par & 0x0f] << 6) |
		    (invparity[par & 0xcc] << 5) |
		    (invparity[par & 0x33] << 4) |
		    (invparity[par & 0xaa] << 3) |
		    (invparity[par & 0x55] << 2) |
		    3;
	else
		code[2] =
		    (invparity[par & 0xf0] << 7) |
		    (invparity[par & 0x0f] << 6) |
		    (invparity[par & 0xcc] << 5) |
		    (invparity[par & 0x33] << 4) |
		    (invparity[par & 0xaa] << 3) |
		    (invparity[par & 0x55] << 2) |
		    (invparity[rp17] << 1) |
		    (invparity[rp16] << 0);

	return 0;
}