Multigetbits, the first

simple_multigetbits.cpp

// Returns 8 bit fields at the given positions (in bits) and of the
// given widths as 16-bit integers, with the values aligned with the
// MSB at the top and garbage in the lower-order bits.
//
// The individual lens must be <=8, the positions are bit offsets
// into the 128-bit "bytes".
template<
	int pos0, int len0,
	int pos1, int len1,
	int pos2, int len2,
	int pos3, int len3,
	int pos4, int len4,
	int pos5, int len5,
	int pos6, int len6,
	int pos7, int len7
>
static inline Vec128_U16 simd_multigetbits(const Vec128_U8& bytes)
{
	// Grab the two bytes straddling each field
	const Vec128_U8 shuf {
		(pos0 >> 3), (pos0 >> 3) + 1,
		(pos1 >> 3), (pos1 >> 3) + 1,
		(pos2 >> 3), (pos2 >> 3) + 1,
		(pos3 >> 3), (pos3 >> 3) + 1,
		(pos4 >> 3), (pos4 >> 3) + 1,
		(pos5 >> 3), (pos5 >> 3) + 1,
		(pos6 >> 3), (pos6 >> 3) + 1,
		(pos7 >> 3), (pos7 >> 3) + 1
	};

	// Use a multiply to do a per-lane variable shift to align the
	// desired bits at the top
	const Vec128_U16 mult {
		1 << (16 - len0 - (pos0 & 7)),
		1 << (16 - len1 - (pos1 & 7)),
		1 << (16 - len2 - (pos2 & 7)),
		1 << (16 - len3 - (pos3 & 7)),
		1 << (16 - len4 - (pos4 & 7)),
		1 << (16 - len5 - (pos5 & 7)),
		1 << (16 - len6 - (pos6 & 7)),
		1 << (16 - len7 - (pos7 & 7))
	};

	// So far, all we've done is set up literal constants!
	// Actual code is extremely short:
	Vec128_U8 shuffled = bytes.shuf(shuf); // PSHUFB
	Vec128_U16 result = shuffled.u16() * mult; // PMULLW

	return result;
}