def mask_to_register(mask):
mask = Mask.as_immediate(mask)
if mask in maskcache:
maskcache.move_to_end(mask)
return maskcache[mask]
try:
maskreg = MaskRegister(64, mask)
except AllocationError:
_, maskreg = maskcache.popitem(False)
x86.mov(maskreg, mask)
maskcache[mask] = maskreg
return maskreg
def square_350_701(dst, src):
""" Requires source and destination registers to be disjunct. """
r = src
r_out = dst
maskreg = MaskRegister()
lowbitmask = Mask('0' * 255 + '1')
x86.vmovdqa(maskreg, lowbitmask)
lowbitreg = Register()
x86.vpand(lowbitreg, maskreg, r[0])
x86.vpandn(r[0], maskreg, r[0])
rest = Register()
twobits = Register()
nexttwobits = Register()
mask0001 = Mask('0001')
x86.vmovdqa(maskreg, mask0001)
for i in range(2, -1, -1):
x86.vpsllq(rest, r[i], 2)
x86.vpsrlq(twobits, r[i], 62)
x86.vpermq(twobits, twobits, '10010011')
x86.vpand(nexttwobits, maskreg, twobits)
x86.vpandn(twobits, maskreg, twobits)
x86.vpxor(r[i], rest, twobits)
if i + 1 < 3:
x86.vpxor(r[i + 1], r[i + 1], nexttwobits)
mask_bit_in_byte = [
Mask(32 * ([ZERO] * i + [ONE] + [ZERO] * (7 - i))) for i in range(8)
]
bits = Register()
accum = Register()
for i in range(2, -1, -1):
for j in range(8):
x86.vpand(bits, r[i], mask_bit_in_byte[j])
if j == 0:
x86.vpshlq(accum, bits, 7 - 2 * j)
else:
x86.vpshlq(bits, bits, 7 - 2 * j)
if j == 7:
x86.vpxor(r[i], accum, bits)
else:
x86.vpxor(accum, accum, bits)
x86.vpermq(lowbitreg, lowbitreg, '11001111')
x86.vpshlq(lowbitreg, lowbitreg, 56)
x86.vpxor(r[2], lowbitreg, r[2])
indices = IndicesMask(
list(range(15, -1, -1)) + [None] * 8 + list(range(7, -1, -1)))
x86.vpshufb(r_out[2], r[0], indices)
x86.vpermq(r_out[2], r_out[2], '10010011')
t1 = Register()
for i in range(2):
indices = IndicesMask([None] * 24 + list(range(15, 7, -1)))
x86.vpshufb(r_out[1 - i], r[i], indices)
indices = IndicesMask(
list(range(15, -1, -1)) + list(range(7, -1, -1)) + [None] * 8)
x86.vpshufb(t1, r[i + 1], indices)
x86.vpxor(r_out[1 - i], t1, r_out[1 - i])
x86.vpermq(r_out[1 - i], r_out[1 - i], '11010010')