def v_alignment_code(amd, v_alignment):
aligned = not v_alignment
if aligned:
code = chew_code(g_v_load_0, amd, 0, aligned, None)
else:
code = []
code += chew_code(g_mul_0, amd, 0, aligned, None)
if amd:
m_1 = P.cutoff_comments(g_mul_1_zen)
m_2 = P.cutoff_comments(g_mul_2_zen)
m_A = P.cutoff_comments(g_mul_10_zen)
q = [int(i, 16) for i in g_perm_zen.split(' ')]
else:
m_1 = P.cutoff_comments(g_mul_1_bwl)
m_2 = P.cutoff_comments(g_mul_2_bwl)
m_A = P.cutoff_comments(g_mul_10_bwl)
q = [int(i, 16) for i in g_perm_bwl.split(' ')]
m_9 = form_m9(m_2, amd)
p = list(range(0xD + 1))
code += chew_code(m_1, amd, 1, aligned, p)
code += chew_code(m_2, amd, 2, aligned, p)
for i in range(3, 9):
p = P.composition(p, q)
code += chew_code(m_2, amd, i, aligned, p)
p = P.composition(p, q)
code += chew_code(m_9, amd, 9, aligned, p)
if aligned:
code += chew_code(m_A, amd, 10, False, None)
code += chew_code(m_A, amd, 10, True, None)
return code
def show_postcondition():
p = list(range(0xD + 1))
q = [int(x, 16) for x in g_perm_bwl.split(' ')]
b = '''s7 sC+s9+s5 s8+sB+s1" s3+sA s2+s0' s4 == s6 sD'''
l = '''sD s7+sA+s4 s8+s5+s1" s6+s2 s9+s3' s0 == sC sB'''
l9 = '''sD sA+s4+s2 s8+s5+sB" s6+s1 s9+s7' sC'''
print 'Broadwell'
for i in range(2, 10):
print 'i=%X pre' % i, E.apply_s_permutation(b, p)
if i == 9:
print E.apply_s_permutation(l9, p)
print 'i=%X pst' % i, E.apply_s_permutation(l, p)
p = P.composition(p, q)
p = list(range(0xD + 1))
b = '''sD sC+s9+sB sA+s4+s0 s7+s5+s2" s1| s3|' == s6 s8'''
l = '''s3 sD+s8+s7 s9+s5+s6 s0+sA+sC" s2| sB|' == s1 s4'''
l9 = '''sD s9 s0+sA s2+s7+s1 sB+s3|" s6|' == s4=v[10]'''
q = [int(x, 16) for x in g_perm_zen.split(' ')]
print '\nZen'
for i in range(2, 10):
print 'i=%X pre' % i, E.apply_s_permutation(b, p)
if i == 9:
print E.apply_s_permutation(l9, p)
print 'i=%X pst' % i, E.apply_s_permutation(l, p)
p = P.composition(p, q)
def do_it(o):
#meat = P.cutoff_comments(g_mul0)
#meat = P.cutoff_comments(g_mul0_only)
#meat = P.cutoff_comments(g_mul0_only_immediate_writes)
#meat = P.cutoff_comments(g_mul0_only_adcx)
#meat = P.cutoff_comments(g_mul0_only_adox_imm_writes)
if 0:
p = list(range(12))
meat += mul1_code(3, P.cutoff_comments(g_muladd_3), p)
m4 = P.cutoff_comments(g_muladd_4)
meat += mul1_code(4, m4, p)
m5 = P.swap_adox_adcx(m4)
q = [int(x, 16) for x in g_perm.split(' ')]
p = P.composition(p, q)
meat += mul1_code(5, m5, p)
p = P.composition(p, q)
meat += mul1_code(6, m4, p)
p = P.composition(p, q)
meat += mul1_code(7, m5, p)
else:
# benchmark only mul0 and tail
#meat += P.cutoff_comments(g_tail_after_mul0_delayed_writes)
#meat += P.cutoff_comments(g_tail_after_mul0_immediate_writes)
pass
code = P.cutoff_comments(g_mul01) + P.cutoff_comments(g_mul01_tail)
code = '\n'.join(code)
code = P.replace_symbolic_names_wr(code, g_var_map)
code = code.split('\n')
code = [replace_ymm_by_xmm(x) for x in code]
cook_asm(o, code, 'mul8x2_zen')
def do_it(o, mul_01, muladd, tail, perm, var_map):
p = list(range(12))
code = mul_add_code(0, P.cutoff_comments(mul_01), p)
mm = P.cutoff_comments(muladd)
code += mul_add_code(2, mm, p)
q = [int(x, 16) for x in perm.split(' ')]
for i in range(3, 6):
p = P.composition(p, q)
code += mul_add_code(i, mm, p)
j = [j for j in range(len(mm)) if mm[j].find('mulx 48(') != -1][0]
mm = mm[:j] + P.cutoff_comments(tail)
p = P.composition(p, q)
code += mul_add_code(6, mm, p)
P.cook_asm(o, code, var_map, True)
def alignment_code(alignment):
if alignment:
code = []
else:
code = chew_code(g_load_0, None, 0, None)
code += chew_code(g_mul_01, 0, alignment, None)
code += chew_code(g_mul_2, 2, alignment, None)
p = list(range(0xC + 1))
q = [int(x, 16) for x in g_perm.split(' ')]
m3 = P.cutoff_comments(g_mul_3)
for i in range(3, 11):
if alignment and (i == 10):
break
fresh = chew_code(m3, i, alignment, p)
if i == 9:
fresh = F.remove_after_jmp(fresh)
code += fresh
if i == 10:
break
p = P.composition(p, q)
if not alignment:
code.append('# tail')
code += chew_code(g_tail, 10, alignment, p)[1:]
return code
def do_it(o):
meat = mul0_code(P.cutoff_comments(g_mul_012))
p = list(range(12))
meat += mul1_code(3, P.cutoff_comments(g_muladd_3), p)
m4 = P.cutoff_comments(g_muladd_4)
meat += mul1_code(4, m4, p)
m5 = P.swap_adox_adcx(m4)
q = [int(x, 16) for x in g_perm.split(' ')]
p = P.composition(p, q)
meat += mul1_code(5, m5, p)
p = P.composition(p, q)
meat += mul1_code(6, m4, p)
p = P.composition(p, q)
meat += mul1_code(7, m5, p)
cook_asm(o, meat)
def alignment_code(shift):
p = list(range(12))
m2 = P.cutoff_comments(E.g_muladd_2)
m3 = P.swap_adox_adcx(m2)
code = mul1_code(2, m2, p, shift)
q = [int(x, 16) for x in E.g_perm.split(' ')]
for i in range(3, 6):
p = P.composition(p, q)
if i & 1:
code += mul1_code(i, m3, p, shift)
else:
code += mul1_code(i, m2, p, shift)
tail = E.cook_tail(m2)
p = P.composition(p, q)
code += mul1_code(6, tail, p, shift)
return code
def alignment_code(shift):
p = list(range(12))
r = mul_code(0, P.cutoff_comments(g_mul01), p, shift)
q = [int(x, 16) for x in g_perm.split(' ')]
m = P.cutoff_comments(g_mul2)
for i in range(2, 6):
r += mul_code(i, m, p, shift)
p = P.composition(p, q)
return r
def do_it(o):
meat = P.cutoff_comments(g_mul_01)
p = list(range(12))
m2 = P.cutoff_comments(g_muladd_2)
m3 = P.swap_adox_adcx(m2)
meat += mul1_code(2, m2, p)
q = [int(x, 16) for x in g_perm.split(' ')]
for i in range(3, 6):
p = P.composition(p, q)
if i & 1:
meat += mul1_code(i, m3, p)
else:
meat += mul1_code(i, m2, p)
tail = cook_tail(m2)
p = P.composition(p, q)
meat += mul1_code(6, tail, p)
P.cook_asm(o, meat, g_var_map, True)
def do_it(name, o):
meat = P.cutoff_comments(g_mul_01)
p = list(range(12))
meat += mul1_code(2, P.cutoff_comments(g_muladd_2), p)
q = [int(x, 16) for x in g_perm.split(' ')]
for i in range(3, 8):
p = P.composition(p, q)
meat += mul1_code(i, P.cutoff_comments(g_muladd_2), p)
cook_asm(name, o, meat)
def show_postcondition():
p = list(range(0xB + 1))
q = [int(x, 16) for x in g_perm.split(' ')]
b = '''s7 ^4+sB+s5 ^3+s8+sA" s6+s3 s2+s1' s0'''
l = '''s6 ^5+sB+s0 ^4+s9+s2" s7+s8 sA+s4' s1'''
for i in range(2, 8):
print 'i=%X pre' % i, E.apply_s_permutation(b, p)
if i < 7:
print 'i=%X pst' % i, E.apply_s_permutation(l, p)
else:
k = ' '.join(['s%X' % j for j in range(0xB + 1)])
print k, '\n' + E.apply_s_permutation(k, p)
p = P.composition(p, q)
def show_postcondition():
p = list(range(0xC + 1))
q = [int(x, 16) for x in g_perm.split(' ')]
assert len(p) == len(q)
b = '''sA sB+s2+s8 s7+sC+s0' s4+s3 s1+s5"'''
l = '''sB s9+sC+s4 s2+s8+s5' s0+s7 s3+s6"'''
for i in range(3, 11):
print 'i=%X pre' % i, A.apply_s_permutation(b, p)
if i == 10:
break
pst = A.apply_s_permutation(l, p)
print 'i=%X pst' % i, pst
pst = P.replace_symbolic_names_wr(pst, g_var_map).replace('%', '')
print 'pst again', pst
p = P.composition(p, q)
def alignment_code(alignment, extra):
if alignment:
code = chew_code(g_preamble, 0, extra, None, None)
else:
code = chew_code(g_load_0, None, extra, True, None)
code += chew_code(g_mul_01, 0, extra, not alignment, None)
m2 = P.cutoff_comments(g_mul_2)
m7 = form_tail(m2, extra)
p = list(range(0xB + 1))
q = [int(i, 16) for i in g_perm.split(' ')]
for i in range(2, 7):
code += chew_code(m2, i, extra, not alignment, p)
p = P.composition(p, q)
if not alignment:
code += chew_code(m7, 7, extra, None, p)
return code
def alignment_code(alignment):
if alignment:
code = chew_code(g_preamble, 0, None, None)
else:
code = chew_code(g_load_0, None, True, None)
code += chew_code(g_mul_01, 0, alignment, None)
m2 = P.cutoff_comments(g_mul_2)
tt = form_tail(m2)
p = list(range(0xC + 1))
q = [int(i, 16) for i in g_perm.split(' ')]
for i in range(2, 10):
code += chew_code(m2, i, alignment, p)
p = P.composition(p, q)
if not alignment:
code += chew_code(tt, 10, None, p)
return code
def do_it(o):
preamble = P.cutoff_comments(g_preamble)
xmm_save = {}
if not g_writing_macro:
xmm_save = P.save_registers_in_xmm(preamble, 11)
p = list(range(14))
code = mul_code(0, preamble, p)
m1 = P.cutoff_comments(g_mul_1)
m2 = P.cutoff_comments(g_mul_2)
code += mul_code(1, m1, p)
q = [int(x, 16) for x in g_perm.split(' ')]
for i in range(2, 5):
code += mul_code(i, m2, p)
p = P.composition(p, q)
tail = m2[:-1]
# extract rp as soon as s6 becomes useless
tail = [tail[0], 'movq rp, s6'] + tail[1:] + P.cutoff_comments(g_tail)
code += mul_code(5, tail, p)
if not g_writing_macro:
P.save_in_xmm(code, xmm_save)
P.insert_restore(code, xmm_save)
cook_asm(o, code, xmm_save, g_var_map)