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 alignment_code(shift):
p = list(range(12))
r = muladd_code(0, P.cutoff_comments(g_mul_01), p, shift)
r += muladd_code(2, P.cutoff_comments(g_mul_2), p, shift)
r += muladd_code(3, P.cutoff_comments(g_mul_3), p, shift)
r += muladd_code(4, P.cutoff_comments(g_mul_4), p, shift)
return r
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_mul0)
muladd = P.cutoff_comments(g_muladd)
tail = P.cutoff_comments(g_tail)
for i in range(1, 6):
meat += mul1_code(i, muladd)
meat += tail
P.cook_asm(o, meat, g_var_map, True)
def do_it(o):
code = P.cutoff_comments(g_preamble)
mul_01 = P.cutoff_comments(g_mul01)
code += alignment_code(8)
code += P.g_std_end + ['retq', 'align0:']
code += P.cutoff_comments(g_load0)
code += alignment_code(0)
P.cook_asm(o, code, E.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 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 form_m9(src, amd):
if amd:
rez = P.cutoff_comments(g_mul_9_bwl_piece0) + src
p = 'adc3(rr[i+4], sC, s8)'
i = [i for i in range(len(rez)) if rez[i].find(p) != -1][0] + 1
rez = rez[:i] + P.cutoff_comments(g_mul_9_zen_piece2)
else:
rez = P.cutoff_comments(g_mul_9_bwl_piece0) + src
i = [
i for i in range(len(rez)) if rez[i].find('movq s3, rr[i+4]') != -1
]
i = i[0] + 1
rez = rez[:i] + P.cutoff_comments(g_mul_9_bwl_piece2)
rez = inject(rez, 'mulx sp[3]', g_mul_9_bwl_piece1)
return rez
def mul_code(i, jj_arg, p, align):
if i:
rr = ['# mul_add %s' % i]
else:
rr = []
if i == 5:
jj = jj_arg[:-1] + P.cutoff_comments(g_tail)
else:
jj = jj_arg
for j in jj:
m = g_patt.match(j)
if m:
u, v = m.group(1), m.group(2)
if v == 'i+1':
k = extract_v(i + 1, u, align)
else:
k = extract_v(int(v), u, align)
if k:
rr.append(k)
continue
rr.append(j)
for y in range(len(rr)):
src = E.apply_s_permutation(rr[y], p)
for x in range(1, 9):
' replace i+x with 8*(i+x) '
src = src.replace('i+%s(' % x, '%s(' % (8 * (i + x)))
' replace i with 8*i '
src = src.replace('i(', '%s(' % (8 * i))
rr[y] = src.rstrip()
return rr
def chew_code(src, amd, i, aligned, p):
if not isinstance(src, list):
src = P.cutoff_comments(src)
if i:
rr = ['# mul_add %s' % i]
if p:
e = '# '
for x in range(len(p)):
e += 's%X->W%X ' % (x, p[x])
rr.append(e)
else:
rr = []
for j in src:
k = evaluate_row(j, i, amd, aligned)
if k and (k != [None]) and (k != ['']):
rr += k
if p:
re = []
for x in rr:
if x[0] == '#':
re.append(x)
else:
re.append(E.apply_s_permutation(x, p))
return re
return rr
def cook_tail(cc):
rr = []
for c in cc:
if c.find(', t1') != -1:
break
rr.append(c)
return rr + P.cutoff_comments(g_tail)
def chew_code(src, i, aligned, p):
if not isinstance(src, list):
src = P.cutoff_comments(src)
if i:
rr = ['# mul_add %s' % i]
else:
rr = []
for j in src:
for k in evaluate_row(j, i, aligned):
if k:
rr.append(k)
if k == 'jmp tail':
break
if not p:
return rr
re = []
for x in rr:
if x[0] == '#':
re.append(x)
else:
re.append(E.apply_s_permutation(x, p))
return re
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, code, var_map):
code = '\n'.join(P.cutoff_comments(code))
code = P.replace_symbolic_names_wr(code, var_map)
comment = P.g_autogenerated_patt % os.path.basename(sys.argv[0])
o.write(comment.replace('//', '#'))
P.write_asm_procedure_header(o, P.guess_subroutine_name(sys.argv[1]))
P.write_asm_inside(o, code + '\nretq')
def mul1_code(i, jj, p):
rr = ['# mul_add %s' % i]
for j in jj:
if j == 'extract v[i+1]':
rr.append(extract_code(i + 1))
continue
if j == 'shift v47':
# only needed once for i = 4
if i == 4:
rr.append('vperm2i128 $0x81,v47,v47,v47')
continue
rr.append(j)
# for i=7, append tail code
if i == 7:
rr += P.cutoff_comments(g_tail)
# apply permutation p, replace i(rp)
for y in range(len(rr)):
src = rr[y]
for x in range(12):
a = '%X' % x
b = '%X' % p[x]
src = re.sub(r'\bs%s\b' % a, 'w' + b, src)
src += ' '
for x in range(1, 9):
' replace i+x with 8*(i+x) '
src = src.replace('i+%s(' % x, '%s(' % (8 * (i + x)))
' replace i with 8*i '
src = src.replace('i(', '%s(' % (8 * i)) + ' '
rr[y] = src.rstrip()
return rr
def mul1_code(i, jj, p):
rr = ['# mul_add %s' % i]
for j in jj:
if j == 'extract v[i+1]':
rr.append(extract_code(i + 1))
continue
j = replace_rz(j)
if (i == 7) and (j.find('dd') != -1):
# no need to update dd
continue
rr.append(j)
# for i=7, append tail code
if i == 7:
rr += P.cutoff_comments(g_tail)
# apply permutation p, replace i(rp)
for y in range(len(rr)):
src = rr[y]
for x in range(12):
a = '%X' % x
b = '%X' % p[x]
src = re.sub(r'\bs%s\b' % a, 'w' + b, src)
src += ' '
for x in range(1, 9):
' replace i+x with 8*(i+x) '
src = src.replace('i+%s(' % x, '%s(' % (8 * (i + x)))
' replace i with 8*i '
src = src.replace('i(', '%s(' % (8 * i)) + ' '
rr[y] = src.rstrip()
return rr
def mul1_code(i, jj, p):
rr = ['# mul_add %s' % i]
for j in jj:
if j.find(':=v[i+1]') != -1:
j = extract_v(i + 1, j[:2])
elif j == 'shift v47':
j = shift_v(i + 1)
if not j:
continue
rr.append(j)
if i == 7:
rr += P.cutoff_comments(g_tail)
# apply permutation p, replace i(rp)
for y in range(len(rr)):
src = rr[y]
for x in range(12):
a = '%X' % x
b = '%X' % p[x]
src = re.sub(r'\bs%s\b' % a, 'w' + b, src)
src += ' '
for x in range(1, 9):
' replace i+x with 8*(i+x) '
src = src.replace('i+%s(' % x, '%s(' % (8 * (i + x)))
' replace i with 8*i '
src = src.replace('i(', '%s(' % (8 * i)) + ' '
rr[y] = src.rstrip()
return rr
def do_it(o):
mul_01 = P.cutoff_comments(E.g_mul_01)
mul_01 = P.starting_from(mul_01, 'mulx')
mul_01 = P.replace_in_string_array(mul_01, 'pextrq $0x1, t0, w8',
'w8:=v[2]')
code = P.cutoff_comments(g_preamble)
code += replace_extract_v(mul_01, 8)
code += alignment_code(8)
# TODO: is it possible to shorten binary code by jmp to common part?
code += P.g_std_end + ['retq', 'align0:']
code += P.cutoff_comments(g_load0)
code += replace_extract_v(mul_01, 0)
code += alignment_code(0)
P.cook_asm(o, code, E.g_var_map, True)
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 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_4arg(o):
comment = P.g_autogenerated_patt % os.path.basename(sys.argv[0])
o.write(comment.replace('//', '#'))
P.write_asm_procedure_header(o, g_subroutine)
code = P.cutoff_comments(g_code)
code = '\n'.join(code)
symb = dict()
for v_k in g_vars_map.split(' '):
v, k = v_k.split(',')
symb[k] = '%' + v
code = P.replace_symbolic_vars_name(code, symb)
P.write_asm_inside(o, code)
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 do_it(o):
meat = mul0_code(P.cutoff_comments(g_mul_01))
p = list(range(12))
meat += mul1_code(2, P.cutoff_comments(g_muladd_2), p)
meat += mul1_code(3, P.cutoff_comments(g_muladd_3), p)
meat += mul1_code(4, P.cutoff_comments(g_muladd_4), p)
meat += mul1_code(5, P.cutoff_comments(g_muladd_5), p)
meat += mul1_code(6, P.cutoff_comments(g_muladd_6), p)
meat += mul1_code(7, P.cutoff_comments(g_muladd_7), p)
cook_asm(o, meat)
def do_it(out):
meat = P.cutoff_comments(g_mul1)
muladd = P.cutoff_comments(g_muladd)
tail = P.cutoff_comments(g_tail)
xmm_save = P.save_registers(meat)
permutation = list(range(11))
s = [int(y) for y in g_permutation.split(' ')]
for i in range(1, 7):
meat += mul1_code(i, muladd, permutation).split('\n')
# yy := composition of permutation and s: yy(i) == s(permutation(i))
yy = [s[j] for j in permutation]
permutation = yy
tail = mul1_code(7, tail, permutation)
for k,v in xmm_save.items():
if v is None:
tail = tail.replace('!restore ' + k, 'pop %s | restore' % k)
else:
tail = tail.replace('!restore ' + k, 'movq %s, %s | restore' % (v, k))
tail = tail.replace('!restore', '|restore')
meat += tail.split('\n')
cook_asm(out, '\n'.join(meat), xmm_save)
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)
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 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 do_it(tgt, code, b_ofs):
data = {
'macro_name': P.guess_subroutine_name(sys.argv[1]),
'scratch': ['w%s s%s' % (i, i) for i in range(4)],
'vars_type': dict([('s%s' % i, 0) for i in range(4)]),
'default_type': 'mp_limb_t',
'input': ['rp r r_p', 'ap r a_p'],
'clobber': 'memory cc',
'source': os.path.basename(sys.argv[0]),
'code_language': 'asm',
'macro_parameters': 'r_p a_p',
}
all_vars = P.extract_int_vars_name(data['scratch']) + ['ap', 'rp']
code = '\n'.join([chew_line(x, b_ofs) for x in P.cutoff_comments(code)])
for v in all_vars:
code = re.sub(r'\b%s\b' % v, '%%[%s]' % v, code)
P.write_cpp_code(tgt, code, data)
def do_it(o, i_name):
with open(i_name, 'rb') as i:
code = '\n'.join(chew_code(i))
comment = P.g_autogenerated_patt % os.path.basename(sys.argv[0])
o.write(comment.replace('//', '#'))
P.write_asm_procedure_header(o, g_hooligan)
P.write_asm_inside(o, code)
code = P.cutoff_comments(g_code)
xmm_save = P.save_registers_in_xmm(code, 5)
P.save_in_xmm(code, xmm_save)
code = '\n'.join(code).replace('@call', 'call ' + g_hooligan)
for k, v in xmm_save.items():
code = code.replace('!restore ' + k, 'movq %s, %s' % (v, k))
code = P.replace_symbolic_names_wr(code, g_var_map)
o.write('\n')
P.write_asm_procedure_header(o, P.guess_subroutine_name(sys.argv[2]))
P.write_asm_inside(o, code + '\nretq')