MurmurHash3 128-bit (x86) for CC

local _and32, _xor32 = bit32.band, bit32.bxor
local _shl32, _shr32 = bit32.lshift, bit32.rshift
local _byte = string.byte

local function rol32(a, b)
    return _shl32(a, b) + _shr32(a, 32 - b)
end

local function modmul32(a, b)
    return _and32(
        _and32(a, 0xFFFF) * _and32(b, 0xFFFF) +
        _and32(_and32(a, 0xFFFF) * _and32(b, 0xFFFF0000), 0xFFFF0000) +
        _and32(_and32(a, 0xFFFF0000) * _and32(b, 0xFFFF), 0xFFFF0000),
    0xFFFFFFFF)
end

local function modadd32(a, b)
    return _and32(
        a + b,
    0xFFFFFFFF)
end

local function str2le4_x4(str, pos) -- this is where things become 1-based
    return
        _byte(str, pos +  1) + _shl32(_byte(str, pos +  2), 8) + _shl32(_byte(str, pos +  3), 16) + _shl32(_byte(str, pos +  4), 24),
        _byte(str, pos +  5) + _shl32(_byte(str, pos +  6), 8) + _shl32(_byte(str, pos +  7), 16) + _shl32(_byte(str, pos +  8), 24),
        _byte(str, pos +  9) + _shl32(_byte(str, pos + 10), 8) + _shl32(_byte(str, pos + 11), 16) + _shl32(_byte(str, pos + 12), 24),
        _byte(str, pos + 13) + _shl32(_byte(str, pos + 14), 8) + _shl32(_byte(str, pos + 15), 16) + _shl32(_byte(str, pos + 16), 24)
end

local function murmur3_mix(a)
    a = _xor32(a, _shr32(a, 16))
    a = modmul32(a, 0x85EBCA6B)
    a = _xor32(a, _shr32(a, 13))
    a = modmul32(a, 0xC2B2AE35)
    a = _xor32(a, _shr32(a, 16))
    return a
end

local c1, c2, c3, c4 = 0x239B961B, 0xAB0E9789, 0x38B34AE5, 0xA1E38B93;
function hash(data, seed)
    local datalen = #data
    local nblocks_x16 = _and32(datalen, 0xFFFFFFF0)
    data = data .. ("\0"):rep(0x10)

    local h1, h2, h3, h4 = seed, seed, seed, seed
    for ix = 0, nblocks_x16 - 16, 16 do
        local k1, k2, k3, k4 = str2le4_x4(data, ix)
        k1 = modmul32(k1, c1); k1 = rol32(k1, 15); k1 = modmul32(k1, c2); h1 = _xor32(h1, k1)
        h1 = rol32(h1, 19); h1 = modadd32(h1, h2); h1 = modadd32(modmul32(h1, 5), 0x561CCD1B)
        k2 = modmul32(k2, c2); k2 = rol32(k2, 16); k2 = modmul32(k2, c3); h2 = _xor32(h2, k2)
        h2 = rol32(h2, 17); h2 = modadd32(h2, h3); h2 = modadd32(modmul32(h2, 5), 0x0BCAA747)
        k3 = modmul32(k3, c3); k3 = rol32(k3, 17); k3 = modmul32(k3, c4); h3 = _xor32(h3, k3)
        h3 = rol32(h3, 15); h3 = modadd32(h3, h4); h3 = modadd32(modmul32(h3, 5), 0x96CD1C35)
        k4 = modmul32(k4, c4); k4 = rol32(k4, 18); k4 = modmul32(k4, c1); h4 = _xor32(h4, k4)
        h4 = rol32(h4, 13); h4 = modadd32(h4, h1); h4 = modadd32(modmul32(h4, 5), 0x32AC3B17)
    end

    local k1, k2, k3, k4 = str2le4_x4(data, nblocks_x16)
    k1 = modmul32(k1, c1); k1 = rol32(k1, 15); k1 = modmul32(k1, c2); h1 = _xor32(h1, k1)
    k2 = modmul32(k2, c2); k2 = rol32(k2, 16); k2 = modmul32(k2, c3); h2 = _xor32(h2, k2)
    k3 = modmul32(k3, c3); k3 = rol32(k3, 17); k3 = modmul32(k3, c4); h3 = _xor32(h3, k3)
    k4 = modmul32(k4, c4); k4 = rol32(k4, 18); k4 = modmul32(k4, c1); h4 = _xor32(h4, k4)

    h1, h2, h3, h4 = _xor32(h1, datalen), _xor32(h2, datalen), _xor32(h3, datalen), _xor32(h4, datalen)

    h1 = modadd32(h1, h2); h1 = modadd32(h1, h3); h1 = modadd32(h1, h4)
    h2 = modadd32(h2, h1); h3 = modadd32(h3, h1); h4 = modadd32(h4, h1)

    h1, h2, h3, h4 = murmur3_mix(h1), murmur3_mix(h2), murmur3_mix(h3), murmur3_mix(h4)

    h1 = modadd32(h1, h2); h1 = modadd32(h1, h3); h1 = modadd32(h1, h4)
    h2 = modadd32(h2, h1); h3 = modadd32(h3, h1); h4 = modadd32(h4, h1)

    return h1, h2, h3, h4
end