def parse_insts(f):
"""
从文件中解析出所有的指令，生成 Inst 对象并加入到 insts 列表中
假设 ins 文件中指令系统部分不会出现 \xff，故以此为循环结束条件
"""
tmp_data = handle_00_byte(f)
log(f"开始解析指令部分，当前文件指针的位置为 {f.tell()}")
while tmp_data != b'\xff':
try:
mnemonic_len = read(f)[0]
tmp_mnemonic = ''.join([
read(f).decode('ascii') for _ in range(mnemonic_len)
])
tmp_addr = read(f)[0]
op_num1 = op_num_map[read(f)]
op_num2 = op_num_map[read(f)]
tmp_inst = Inst(tmp_mnemonic, tmp_addr, op_num1, op_num2)
log(f"解析出指令 {tmp_inst}")
insts.append(tmp_inst)
handle_int_inst(f, tmp_mnemonic)
tmp_data = handle_00_byte(f)
except Exception as err:
error(f"解析失败，原因 {err}")
log("指令部分解析结束")
def decode(self):
"""Instruction decoder. Returns the number of bytes decoded."""
inst = Inst(myInput = self.input, add = self.pc, mode = self.dis_mode, syntax = self.syntax)
self.error = 0
self.input.start ()
if get_prefixes(self, inst) != 0:
pass # ;print('prefixes error') # error
elif search_itab(self, inst) != 0:
pass #; print('itab error') # error
elif do_mode(self, inst) != 0:
pass #; print('mode error') # error
elif disasm_operands(self, inst) != 0:
pass #; print('operand error') # error
elif resolve_operator(self, inst) != 0:
pass #; print('operator error') # error
# Handle decode error.
if self.error:
inst.clear()
inst.size = self.input.ctr + 1
inst.raw = self.input.buffer[0:inst.size]
inst.set_pc(inst.add + inst.size)
return inst
def load_fd(fd):
offsets = {
'const': len(const_table),
'label': len(label_table),
'inst': len(inst_table),
}
# constants block
num_constants = read_int(fd)
for _ in range(0, num_constants):
const = read_constant(fd)
const_table.register(const)
# symbols block
num_symbols = read_int(fd)
symbol_translation = [0] * num_symbols
for i in range(0, num_symbols):
val = read_str(fd)
symbol_translation[i] = sym(val)
num_labels = read_int(fd)
for _ in range(0, num_labels):
name = read_str(fd)
addr = read_int(fd) + offsets['inst']
trace = None
has_trace = read_int(fd)
trace = None
if has_trace == 1: trace = read_str(fd)
label = Label(name, addr, trace)
register_label(label)
num_insts = read_int(fd)
for i in range(0, num_insts):
command = read_str(fd)
num_args = read_int(fd)
raw_args = [999] * num_args
for i in range(0, num_args):
raw_args[i] = read_int(fd)
inst_type = inst_type_table.get(command)
args = inst_type.reindex(raw_args, offsets, symbol_translation)
inst_table.register(Inst(inst_type.id, args))
def filter_change(origin_inst, dump_file_content, base, newcode, line_index,
reversal_bit_id):
"""Compare the origin inst and redisassember inst, and record the changed part
@:return: It currently presents the judgment of operands' different
"""
tmp_result2 = code_line_reg.findall(dump_file_content)
if not tmp_result2:
logging.warning("generated file %s has no legal content" % newcode)
tmp_line = tmp_result2[line_index]
tmp_inst = Inst(tmp_line)
# confirm opcode part
if tmp_inst.op != origin_inst.op:
logging.info(
"Opcode changes: %s => %s when bit [%d] is flipped from [%d]\t\tChange to:%s",
origin_inst.op, tmp_inst.op, reversal_bit_id,
(base >> reversal_bit_id) & 0x1, tmp_line[0])
# bits = bits | (((base >> i) & 0x1) << i)
origin_inst.opcode_positions.append(reversal_bit_id)
# confirm modifier part
elif tmp_inst.modifier != origin_inst.modifier:
logging.info(
"Modifier changes: %s => %s when bit [%d] is flipped from [%d]\t\tChange to:%s",
origin_inst.modifier, tmp_inst.modifier, reversal_bit_id,
(base >> reversal_bit_id) & 0x1, tmp_line[0])
origin_inst.modifier_positions.append(reversal_bit_id)
else:
len_origin = len(origin_inst.operands)
len_tmp = len(tmp_inst.operands)
for i in range(min(len_origin, len_tmp)):
if origin_inst.operands[i] != tmp_inst.operands[i]:
return i, tmp_inst
# if len_tmp > len_origin:
# return len_origin
return None, None
def work(input_file_name, output_file, section_start):
global sass_content, ops_bits
arch = "sm_75"
# for debug
# input_file_name = 'gaussian.cubin'
init_sass_cubin_files(input_file_name, arch)
logging.basicConfig(format="%(message)s",
filename="./log/%s" % output_file,
filemode="a",
level=logging.INFO)
logging.info("Time:\t%s" %
time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
com_lib.kernel_section_start_offset = int(section_start, 16)
with open("%s.sass" % os.path.splitext(input_file_name)[0], 'r') as fin:
sass_content = fin.read()
tmp_result = code_line_reg.findall(sass_content)
if not tmp_result:
logging.error("Not found pairs")
return
# @todo: check is there any irregular code
# how many lines will be checked.
for j, line in enumerate(tmp_result[:1]):
logging.info("================================")
logging.info("raw line:\t\t%s" % line[0])
a_origin_inst = Inst(line)
instructions.append(a_origin_inst)
base = a_origin_inst.enc
bits = 0x0
origin_operand_types = check_operand_types(a_origin_inst)
if FLAG_CHECK_OPERAND:
if len(a_origin_inst.operands) and origin_operand_types:
logging.info(
"Original op and modifier:%s:\t%s" %
(a_origin_inst.op, "".join(a_origin_inst.modifier)))
logging.info("0b{:0128b}".format(base) + ": " +
"".join(a_origin_inst.operands))
logging.info("newcode operand:")
# if you just want to check operand, uncomment the following else branch
else:
continue
# In volta and turing, len of the instruction code is 128bitk
for i in range(0, 128):
# i from right to left
mask = 2**i
newcode = base ^ mask
dump_file_content = dump("0x{:032x}".format(newcode), arch, j)
# print(dump_file_content)
# Compare the disassemble to check which field changes: opcode, operand or modifer
if dump_file_content and dump_file_content.find(
"?") == -1 and dump_file_content.find("error") == -1:
tmp_pp, tmp_inst = filter_change(a_origin_inst,
dump_file_content, base,
newcode, j, i)
if tmp_pp is not None:
# the ith bit affects tmp_ppth operand
a_origin_inst.operand_positions[tmp_pp].append(i)
# @todo print the reverse bit
logging.info("%s: %d\t%s" %
("0b{:0128b}".format(newcode), i, " ".join(
tmp_inst.operands)))
# if len(positions) > 0:
# logging.info("0b{:0128b}".format(bits) + ": %s opcode bits %s: ", origin_inst.op, positions)
if len(a_origin_inst.opcode_positions) > 0:
ops_bits[a_origin_inst.op] = list(
set(
ops_bits.get(a_origin_inst.op, []) +
a_origin_inst.opcode_positions))
logging.info("Operand combination types: %s", origin_operand_types)
for i in range(0, len(a_origin_inst.operand_positions)):
if len(origin_operand_types) > i:
tmp_type = origin_operand_types[i]
else:
tmp_type = 'None'
logging.info("Operand type: %s", tmp_type)
logging.info("Encoding: %s", a_origin_inst.operand_positions[i])
for node in ops_bits:
logging.info("%s:\t[%s]", node,
",".join(str(x) for x in ops_bits[node]))
from sets import Set
from inst import Inst
import subprocess
import sys
if __name__ == "__main__":
opset = Set([])
with open(sys.argv[1]) as f:
for line in f:
field = line.split()
inst = Inst(field, False)
if not inst.op() in opset:
opset.add(inst.op())
sys.stdout.write(line)
if len(sys.argv) >= 5:
limit = sys.argv[4]
else:
limit = 100
count = 0
version = int(arch.split("_")[1])
with open(sass) as f:
for line in f:
pos = []
bits = 0x0
count += 1
if count == limit:
break
line_split = line.split()
# Construct instruction structure
origin = Inst(line_split)
# Find the 64-bit encodings
base = int(origin.enc(), 16)
# Bit by bit xor, observe whether opcode changes and guess what this bit represent
for i in range(0, 64):
mask = 2**i
newcode = base ^ mask
# Disassemble the new code
dump_file = dump("0x{:016x}".format(newcode), arch)
# Compare the disassemble to check which field changes: opcode, operand or modifer
if dump_file and dump_file.find("?") == -1 and dump_file.find(
"error") == -1:
line = dump_file.split("\n")
if version < 40:
line_inst = line[1].split()
else:
from inst import Inst, asm, print_asm, print_ihex
rd = []
for i in range(50):
rd.append(i)
rs = []
for i in range(50):
rs.append(i)
seg_pattern = [0xFC,0x60,0xDA,0xF2,0x66,0xB6,0xBE,0xE0,0xFE,0xF6]
program = [
Inst.LUI(5,0x10001000),
]
for i in range(10):
program.append(Inst.ADDI(rd[6],rs[0],seg_pattern[i]))
program.append(Inst.SB(rd[5],rs[6],i))
program_main = [
Inst.LUI(rd[6], 0x04000000),
Inst.ADD(rd[7],rs[0],rs[0]),
Inst.LW(rd[10], rs[6], 0x48),
'mainloop',
Inst.ADD(rd[8],rs[7],rs[5]),
Inst.ANDI(rd[11],rs[10],0x01),
Inst.LBEQ(rd[11],0,'dontpush'),
'push',
Inst.LBU(rd[9],rs[8],0),
from inst import Inst, asm, print_asm, print_ihex program = [ Inst.LUI(5, 0x10000000), # memory = 0x10000000 Inst.ADD(6, 0, 0), # x6 = 0 (counter) 'loop', Inst.SW(5, 6, 0x100), # memory[x5+0x100] = x6 Inst.ADDI(6, 6, 1), # x6++ Inst.LJAL(0, 'loop') # goto loop ] r = asm(program) print_asm(r) print() print_ihex(r)
from inst import Inst, asm, print_asm, print_ihex
program = [
Inst.LUI(5, 0x04000000), # r5 に7 セグのアドレスを代入
Inst.ADDI(10, 0, 0x60), # セグ「1」のパタンをr10 に代入
Inst.SB(5, 10, 0x00), # r5[0] = 0x60 (7 セグのアドレスに0x60 をストア)
Inst.JAL(0, -4 * 1) # 1 命令前に無条件分岐
]
r = asm(program)
print_asm(r)
print()
print_ihex(r)
_score2 = 22 _counterX = 23 _intervalX = 24 _counterY = 25 _intervalY = 26 _memAddr = 27 _memBase = 28 _count = 29 _tmp2 = 30 # 初期化部分 sys_init = [ Inst.LUI(_addrBase, 0x0400_0000), # addrBase = [0x0400_0000] Inst.ADDI(_wr, _num0, 1), # wr = 1 Inst.SLLI(_wr, _wr, 14), # wr <<= 14 Inst.ADDI(_barY1u, _num0, 1), # barY1u = 1 Inst.ADDI(_barY1d, _num0, 2), # barY1d = 2 Inst.ADDI(_barY2u, _num0, 1), # barY2u = 1 Inst.ADDI(_barY2d, _num0, 2), # barY2d = 2 Inst.ADDI(_velX, _num0, 1), # velX = 1 Inst.ADDI(_velY, _num0, 1), # velY = 1 Inst.ADDI(_score1, _num0, 0), # score1 = 0 Inst.ADDI(_score2, _num0, 0), # score2 = 0 Inst.ADDI(_counterX, _num0, 0), # counterX = 0 Inst.ADDI(_counterY, _num0, 0), # counterY = 0 Inst.ADDI(_intervalX, _num0, 3), # interval = 1 Inst.ADDI(_intervalY, _num0, 3), # interval = 1 Inst.ADDI(_num1, _num0, 1), # num1 = 1
def bin_decode(bin_line):
if bin_line == '0':
inst = Inst()
d = {
"op": "pass",
"dests": [],
"srcs": []
}
inst.fromDict(d)
return inst
op_bits = bin_line[:op_bit]
dests = []
srcs = []
op = code.op_inv[int(op_bits, 2)]
start_point = op_bit
for i in range(3):
ns_bits = bin_line[start_point:start_point + ns_bit]
ns = code.ns_inv[int(ns_bits, 2)]
if ns == "NN":
pe_pu_bit = bin_line[start_point + ns_bit + index_bit + nn_nb_bit - 1]
if pe_pu_bit == '0':
ns = "NN0_out"
else:
ns = "NN1_out"
elif ns == "NB":
pe_pu_bit = bin_line[start_point + ns_bit + index_bit + nn_nb_bit - 1]
if pe_pu_bit == '0':
ns = "NB0_out"
else:
ns = "NB1_out"
index_bits = bin_line[start_point + ns_bit:start_point + ns_bit + index_bit]
index = int(index_bits, 2)
dest = {
"dest_nid": ns,
"dest_index": index
}
dests.append(dest)
start_point += (ns_bit + index_bit + nn_nb_bit)
for i in range(3):
ns_bits = bin_line[start_point:start_point + ns_bit]
ns = code.ns_inv[int(ns_bits, 2)]
if ns == "NN":
pe_pu_bit = bin_line[start_point + ns_bit + index_bit + nn_nb_bit - 1]
if pe_pu_bit == '0':
ns = "NN0_in"
else:
ns = "NN1_in"
elif ns == "NB":
pe_pu_bit = bin_line[start_point + ns_bit + index_bit + nn_nb_bit - 1]
if pe_pu_bit == '0':
ns = "NB0_in"
else:
ns = "NB1_in"
index_bits = bin_line[start_point + ns_bit:start_point + ns_bit + index_bit]
index = int(index_bits, 2)
src = {
"src_nid": ns,
"src_index": index
}
srcs.append(src)
start_point += (ns_bit + index_bit + nn_nb_bit)
inst = Inst()
d = {
"op": op,
"dests": dests,
"srcs": srcs
}
inst.fromDict(d)
return inst
# 0~9をループするカウンタ
## わかっているクソ仕様
## SBを一回使用した後にADDIを実行すると機能しない。
## これを回避するためにはダミーのSBが必要。
from inst import Inst, asm, print_asm, print_ihex
program = [
# Store 7seg led patterns
Inst.LUI(5, 0x10001000), # memory = 0x10001000
Inst.ADDI(6, 0, 0b11111100), # 7seg: 0
Inst.SB(5, 6, 0x0),
Inst.SB(5, 6, 0x1), #dummySB
Inst.ADDI(6, 0, 0b01100000), # 1
Inst.SB(5, 6, 0x1),
Inst.SB(5, 6, 0x2), #dummySB
Inst.ADDI(6, 0, 0b11011010), # 2
Inst.SB(5, 6, 0x2),
Inst.SB(5, 6, 0x3), #dummySB
Inst.ADDI(6, 0, 0b11110010), # 3
Inst.SB(5, 6, 0x3),
Inst.SB(5, 6, 0x4), #dummySB
Inst.ADDI(6, 0, 0b01100110), # 4
Inst.SB(5, 6, 0x4),
Inst.SB(5, 6, 0x5), #dummySB
Inst.ADDI(6, 0, 0b10110110), # 5
Inst.SB(5, 6, 0x5),
Inst.SB(5, 6, 0x6), #dummySB
Inst.ADDI(6, 0, 0b10111110), # 6
Inst.SB(5, 6, 0x6),
Inst.SB(5, 6, 0x7), #dummySB
def disassemble(self, add):
try:
self.seek(add)
return self.decode()
except DecodeException:
return Inst(self.input)
from inst import Inst, asm, print_asm, print_ihex
seg_pattern = [0xFC, 0x60, 0xDA, 0xF2, 0x66, 0xB6, 0xBE, 0xE0, 0xFE, 0xF6]
x_0 = 0
pattern_mem = 5
pattern_tmp = 6
program = [Inst.LUI(pattern_mem, 0x10001000)]
for i in range(10):
print(seg_pattern[i])
program.append(Inst.ADDI(pattern_tmp, 0, seg_pattern[i]))
program.append(Inst.SB(pattern_mem, pattern_tmp, i))
seg = 6
counter = 7
pattern_mem_addr = 8
pattern = 9
max = 15
program_main = [
'main',
Inst.LUI(seg, 0x04000000),
Inst.ADDI(counter, x_0, 0),
Inst.ADDI(max, x_0, 9), 'loop',
Inst.ADD(pattern_mem_addr, counter, pattern_mem),
Inst.LBU(pattern, pattern_mem_addr, 0),
Inst.SB(seg, pattern, 0),
Inst.ADDI(counter, counter, 1),
Inst.LBEQ(counter, max, 'reset'),
from inst import Inst, asm, print_asm, print_ihex program = [ # Store 7seg led patterns Inst.LUI(5, 0x10001000), # memory = 0x10001000 Inst.ADDI(6, 0, 0b11111100), # 7seg: 0 Inst.SB(5, 6, 0x0), Inst.ADDI(6, 0, 0b01100000), # 1 Inst.SB(5, 6, 0x1), Inst.ADDI(6, 0, 0b11011010), # 2 Inst.SB(5, 6, 0x2), Inst.ADDI(6, 0, 0b11110010), # 3 Inst.SB(5, 6, 0x3), Inst.ADDI(6, 0, 0b01100110), # 4 Inst.SB(5, 6, 0x4), Inst.ADDI(6, 0, 0b10110110), # 5 Inst.SB(5, 6, 0x5), Inst.ADDI(6, 0, 0b10111110), # 6 Inst.SB(5, 6, 0x6), Inst.ADDI(6, 0, 0b11100000), # 7 Inst.SB(5, 6, 0x7), Inst.ADDI(6, 0, 0b11111110), # 8 Inst.SB(5, 6, 0x8), Inst.ADDI(6, 0, 0b11110110), # 9 Inst.SB(5, 6, 0x9), # Main Inst.LUI(6, 0x04000000), Inst.ADD(7, 0, 0), # x7 = 0 (counter) ’loop’, Inst.ADD(8, 7, 5), # address of memory[x7] Inst.LBU(9, 8, 0), # x9 = memory[x7]
#! /usr/bin/env python3
from inst import Inst
filename = 'sample1_7seg.hex'
program = []
program.append(Inst.LUI(5, 0x04000000))
program.append(Inst.ADDI(10, 0, 0xff))
program.append(Inst.SB(5, 10, 0x00))
program.append(Inst.JAL(0, -4 * 2))
# generate assembly
#print('')
print('sample program 1')
for i in program:
print('{:08x} | {}'.format(i.gen_code(), i.gen_mnemonic()))
# generate intel hex format
with open(filename, 'w', encoding='utf-8') as file:
for offset, inst in enumerate(program):
file.write(inst.gen_HEX(offset))
file.write("\n")
file.write(':00000001FF\n')
arch = sys.argv[2]
if len(sys.argv) >= 5:
limit = sys.argv[4]
else:
limit = 100
count = 0
version = int(arch.split("_")[1])
with open(sass) as f:
for line in f:
pos = []
count += 1
if count == limit:
break
line_split = line.split()
# Construct instruction structure
origin = Inst(line_split)
# Find the 64-bit encodings
base = int(origin.enc(), 16)
# Bit by bit xor, observe whether opcode changes and guess what this bit represent
for i in range(0, 64):
mask = 2**i
newcode = base ^ mask
# Disassemble the new code
dump_file = dump("0x{:016x}".format(newcode), arch)
# Compare the disassemble to check which field changes: opcode, operand or modifer
if dump_file and dump_file.find("?") == -1 and dump_file.find(
"error") == -1:
line = dump_file.split("\n")
if version < 40:
line_inst = line[1].split()
else:
arch = sys.argv[2]
if len(sys.argv) >= 5:
limit = sys.argv[4]
else:
limit = 100
count = 0
version = int(arch.split("_")[1])
with open(sys.argv[1]) as f:
for line in f:
pos = []
count += 1
if count == limit:
break
line_split = line.split()
# Construct instruction structure
origin = Inst(line_split)
# Find the 64-bit encodings
base = int(origin.enc(), 16)
origin_operand_types = check_operand_types(origin)
if len(origin.operands()) and origin_operand_types.find('X') == -1:
pp = [[] for i in range(len(origin_operand_types))]
logging.info(origin.op() + " " + origin.modifier())
logging.info("0b{:064b}".format(base) + ": " +
origin.operands())
for i in range(0, 64):
mask = 2**i
newcode = base ^ mask
# Disassemble the new code
dump_file = dump("0x{:016x}".format(newcode), arch)
if dump_file and dump_file.find(
"?") == -1 and dump_file.find("error") == -1:
from inst import Inst, asm, print_asm, print_ihex
program = [
Inst.LUI(5, 0x04000000),
Inst.ADDI(6, 0, 0xFF),
Inst.LW(10, 5, 0x48),
Inst.ANDI(11, 10, 0x01),
Inst.BEQ(11, 0, 0x08),
Inst.SB(5, 6, 0x000),
Inst.JAL(0, 4),
Inst.SB(5, 0, 0x000),
Inst.JAL(0, -28)
]
r = asm(program)
print_asm(r)
print()
print_ihex(r)