MONTMUL

via = cpuid_leaf(0xC0000001, 0)[3]
if via >> 12 & 3 != 3:
    fault(UD)                    # CPUID: no PMM, or it is off
if not cr4 >> 9 & 1:
    fault(UD)                    # no SSE
x = aSI                          # the context, any alignment:
z = load(x, 32)                  # m' = -1 / M modulo 2^32,
pa = load(x + 4, 32)             # then where A, B, T and M are,
pb = load(x + 8, 32)             # 16-byte aligned (and a scratch
pt = load(x + 12, 32)            # buffer at x + 20)
pm = load(x + 16, 32)
n = ecx                          # the numbers' length in bits
if 256 <= n <= 32768 and n % 128 == 0:
    a = load(pa, n)
    b = load(pb, n)
    m = load(pm, n)
    s = 0                        # A's words in turn, the lowest
    for i in range(n // 32):     # first: s stays below 2M
        s += (a >> 32 * i & 0xFFFFFFFF) * b
        q = s * z & 0xFFFFFFFF   # s + q * M: a multiple of 2^32
        s = s + q * m >> 32
    store(pt, n + 128, s)        # A * B / 2^n modulo M, or that
    ecx = 0                      # plus M: software subtracts M;
    edx = 0                      # T has 128 bits more for it
else:                            # a length the guide does not
    ecx = edx = UNDEFINED        # allow
eax = UNDEFINED                  # the multiplier's progress
# EAX must be 0 at the start, and T all zeros: the multiplier keeps
# its partial product there, so that an interrupt resumes it.
# Pointers and lengths are 32 bits, whatever the prefixes; in
# 64-bit mode XED decodes the form only with a 67 prefix