使用Cython和Nuitka编译Python脚本

0x00 前言

在Python中，可以使用py2exe或PyInstaller之类的工具将Python脚本编译成二进制文件，从而提升可移植性，并在一定程度上提升了性能。不过这类工具的实现只是将py文件编译成pyc或pyo，在安全性上还是弱了一些，存在被反编译的风险。

为了测试不同编译方式的性能差异，这里统一使用python2.7中提供的test/pystone.py作为执行脚本。由于这个脚本不支持python3，因此做了下python3的适配。完整的测试代码如下：

#! /usr/bin/env python
"""
"PYSTONE" Benchmark Program
Version:        Python/1.1 (corresponds to C/1.1 plus 2 Pystone fixes)
Author:         Reinhold P. Weicker,  CACM Vol 27, No 10, 10/84 pg. 1013.
                Translated from ADA to C by Rick Richardson.
                Every method to preserve ADA-likeness has been used,
                at the expense of C-ness.
                Translated from C to Python by Guido van Rossum.
Version History:
                Version 1.1 corrects two bugs in version 1.0:
                First, it leaked memory: in Proc1(), NextRecord ends
                up having a pointer to itself.  I have corrected this
                by zapping NextRecord.PtrComp at the end of Proc1().
                Second, Proc3() used the operator != to compare a
                record to None.  This is rather inefficient and not
                true to the intention of the original benchmark (where
                a pointer comparison to None is intended; the !=
                operator attempts to find a method __cmp__ to do value
                comparison of the record).  Version 1.1 runs 5-10
                percent faster than version 1.0, so benchmark figures
                of different versions can't be compared directly.
"""
LOOPS = 50000
try:
    from time import perf_counter as clock
except ImportError:
    from time import clock
__version__ = "1.1"
[Ident1, Ident2, Ident3, Ident4, Ident5] = range(1, 6)
class Record:
    def __init__(self, PtrComp = None, Discr = 0, EnumComp = 0,
                       IntComp = 0, StringComp = 0):
        self.PtrComp = PtrComp
        self.Discr = Discr
        self.EnumComp = EnumComp
        self.IntComp = IntComp
        self.StringComp = StringComp
    def copy(self):
        return Record(self.PtrComp, self.Discr, self.EnumComp,
                      self.IntComp, self.StringComp)
TRUE = 1
FALSE = 0
def main(loops=LOOPS):
    benchtime, stones = pystones(loops)
    print("Pystone(%s) time for %d passes = %g" % \
          (__version__, loops, benchtime))
    print("This machine benchmarks at %g pystones/second" % stones)
def pystones(loops=LOOPS):
    return Proc0(loops)
IntGlob = 0
BoolGlob = FALSE
Char1Glob = '\0'
Char2Glob = '\0'
Array1Glob = [0]*51
Array2Glob = list(map(lambda x: x[:], [Array1Glob]*51))
PtrGlb = None
PtrGlbNext = None
def Proc0(loops=LOOPS):
    global IntGlob
    global BoolGlob
    global Char1Glob
    global Char2Glob
    global Array1Glob
    global Array2Glob
    global PtrGlb
    global PtrGlbNext
    starttime = clock()
    for i in range(loops):
        pass
    nulltime = clock() - starttime
    PtrGlbNext = Record()
    PtrGlb = Record()
    PtrGlb.PtrComp = PtrGlbNext
    PtrGlb.Discr = Ident1
    PtrGlb.EnumComp = Ident3
    PtrGlb.IntComp = 40
    PtrGlb.StringComp = "DHRYSTONE PROGRAM, SOME STRING"
    String1Loc = "DHRYSTONE PROGRAM, 1'ST STRING"
    Array2Glob[8][7] = 10
    starttime = clock()
    for i in range(loops):
        Proc5()
        Proc4()
        IntLoc1 = 2
        IntLoc2 = 3
        String2Loc = "DHRYSTONE PROGRAM, 2'ND STRING"
        EnumLoc = Ident2
        BoolGlob = not Func2(String1Loc, String2Loc)
        while IntLoc1 < IntLoc2:
            IntLoc3 = 5 * IntLoc1 - IntLoc2
            IntLoc3 = Proc7(IntLoc1, IntLoc2)
            IntLoc1 = IntLoc1 + 1
        Proc8(Array1Glob, Array2Glob, IntLoc1, IntLoc3)
        PtrGlb = Proc1(PtrGlb)
        CharIndex = 'A'
        while CharIndex <= Char2Glob:
            if EnumLoc == Func1(CharIndex, 'C'):
                EnumLoc = Proc6(Ident1)
            CharIndex = chr(ord(CharIndex)+1)
        IntLoc3 = IntLoc2 * IntLoc1
        IntLoc2 = IntLoc3 / IntLoc1
        IntLoc2 = 7 * (IntLoc3 - IntLoc2) - IntLoc1
        IntLoc1 = Proc2(IntLoc1)
    benchtime = clock() - starttime - nulltime
    if benchtime == 0.0:
        loopsPerBenchtime = 0.0
    else:
        loopsPerBenchtime = (loops / benchtime)
    return benchtime, loopsPerBenchtime
def Proc1(PtrParIn):
    PtrParIn.PtrComp = NextRecord = PtrGlb.copy()
    PtrParIn.IntComp = 5
    NextRecord.IntComp = PtrParIn.IntComp
    NextRecord.PtrComp = PtrParIn.PtrComp
    NextRecord.PtrComp = Proc3(NextRecord.PtrComp)
    if NextRecord.Discr == Ident1:
        NextRecord.IntComp = 6
        NextRecord.EnumComp = Proc6(PtrParIn.EnumComp)
        NextRecord.PtrComp = PtrGlb.PtrComp
        NextRecord.IntComp = Proc7(NextRecord.IntComp, 10)
    else:
        PtrParIn = NextRecord.copy()
    NextRecord.PtrComp = None
    return PtrParIn
def Proc2(IntParIO):
    IntLoc = IntParIO + 10
    while 1:
        if Char1Glob == 'A':
            IntLoc = IntLoc - 1
            IntParIO = IntLoc - IntGlob
            EnumLoc = Ident1
        if EnumLoc == Ident1:
            break
    return IntParIO
def Proc3(PtrParOut):
    global IntGlob
    if PtrGlb is not None:
        PtrParOut = PtrGlb.PtrComp
    else:
        IntGlob = 100
    PtrGlb.IntComp = Proc7(10, IntGlob)
    return PtrParOut
def Proc4():
    global Char2Glob
    BoolLoc = Char1Glob == 'A'
    BoolLoc = BoolLoc or BoolGlob
    Char2Glob = 'B'
def Proc5():
    global Char1Glob
    global BoolGlob
    Char1Glob = 'A'
    BoolGlob = FALSE
def Proc6(EnumParIn):
    EnumParOut = EnumParIn
    if not Func3(EnumParIn):
        EnumParOut = Ident4
    if EnumParIn == Ident1:
        EnumParOut = Ident1
    elif EnumParIn == Ident2:
        if IntGlob > 100:
            EnumParOut = Ident1
        else:
            EnumParOut = Ident4
    elif EnumParIn == Ident3:
        EnumParOut = Ident2
    elif EnumParIn == Ident4:
        pass
    elif EnumParIn == Ident5:
        EnumParOut = Ident3
    return EnumParOut
def Proc7(IntParI1, IntParI2):
    IntLoc = IntParI1 + 2
    IntParOut = IntParI2 + IntLoc
    return IntParOut
def Proc8(Array1Par, Array2Par, IntParI1, IntParI2):
    global IntGlob
    IntLoc = IntParI1 + 5
    Array1Par[IntLoc] = IntParI2
    Array1Par[IntLoc+1] = Array1Par[IntLoc]
    Array1Par[IntLoc+30] = IntLoc
    for IntIndex in range(IntLoc, IntLoc+2):
        Array2Par[IntLoc][IntIndex] = IntLoc
    Array2Par[IntLoc][IntLoc-1] = Array2Par[IntLoc][IntLoc-1] + 1
    Array2Par[IntLoc+20][IntLoc] = Array1Par[IntLoc]
    IntGlob = 5
def Func1(CharPar1, CharPar2):
    CharLoc1 = CharPar1
    CharLoc2 = CharLoc1
    if CharLoc2 != CharPar2:
        return Ident1
    else:
        return Ident2
def Func2(StrParI1, StrParI2):
    IntLoc = 1
    while IntLoc <= 1:
        if Func1(StrParI1[IntLoc], StrParI2[IntLoc+1]) == Ident1:
            CharLoc = 'A'
            IntLoc = IntLoc + 1
    if CharLoc >= 'W' and CharLoc <= 'Z':
        IntLoc = 7
    if CharLoc == 'X':
        return TRUE
    else:
        if StrParI1 > StrParI2:
            IntLoc = IntLoc + 7
            return TRUE
        else:
            return FALSE
def Func3(EnumParIn):
    EnumLoc = EnumParIn
    if EnumLoc == Ident3: return TRUE
    return FALSE
if __name__ == '__main__':
    import sys
    def error(msg):
        print >>sys.stderr, msg,
        print >>sys.stderr, "usage: %s [number_of_loops]" % sys.argv[0]
        sys.exit(100)
    nargs = len(sys.argv) - 1
    if nargs > 1:
        error("%d arguments are too many;" % nargs)
    elif nargs == 1:
        try: loops = int(sys.argv[1])
        except ValueError:
            error("Invalid argument %r;" % sys.argv[1])
    else:
        loops = LOOPS
    main(loops)
  COPY

以下测试数据均为连续执行3次，取最大值。

0x01 不同Python版本的测试数据

• Python 2.7

  Pystone(1.1) time for 50000 passes = 0.178948
  This machine benchmarks at 279411 pystones/second
    COPY

• Python 3.7

  Pystone(1.1) time for 50000 passes = 0.201795
  This machine benchmarks at 247777 pystones/second
    COPY

• Python 3.8

  Pystone(1.1) time for 50000 passes = 0.222014
  This machine benchmarks at 225211 pystones/second
    COPY

• Python 3.9

  Pystone(1.1) time for 50000 passes = 0.223407
  This machine benchmarks at 223807 pystones/second
    COPY

• Python 3.10

  Pystone(1.1) time for 50000 passes = 0.265725
  This machine benchmarks at 188164 pystones/second
    COPY

• Python 3.11

  Pystone(1.1) time for 50000 passes = 0.104691
  This machine benchmarks at 477596 pystones/second
    COPY

可以看到，Python 3.11版本有了明显的性能提升，这个与官方的宣传也是一致的。

0x02 使用Cython编译python脚本

$ pip install cython
$ cython -3 --embed pystone.py
$ gcc -pthread -fPIC -fwrapv -O2 -Wall -fno-strict-aliasing -I/usr/include/python3.7 -l:libpython3.7m.so -o pystone pystone.c
$ ls -l pystone
-rwxrwxrwx 1 drunkdream drunkdream 178928 Sep  6 15:42 pystone
$ readelf -d pystone
Dynamic section at offset 0x1fd08 contains 26 entries:
  Tag        Type                         Name/Value
 0x0000000000000001 (NEEDED)             Shared library: [libpython3.7m.so.1.0]
 0x0000000000000001 (NEEDED)             Shared library: [libpthread.so.0]
 0x0000000000000001 (NEEDED)             Shared library: [libc.so.6]
 0x000000000000000c (INIT)               0x403000
 0x000000000000000d (FINI)               0x41b514
 0x0000000000000019 (INIT_ARRAY)         0x420cf8
 0x000000000000001b (INIT_ARRAYSZ)       8 (bytes)
 0x000000000000001a (FINI_ARRAY)         0x420d00
 0x000000000000001c (FINI_ARRAYSZ)       8 (bytes)
 0x000000006ffffef5 (GNU_HASH)           0x400308
 0x0000000000000005 (STRTAB)             0x401078
 0x0000000000000006 (SYMTAB)             0x400328
 0x000000000000000a (STRSZ)              2404 (bytes)
 0x000000000000000b (SYMENT)             24 (bytes)
 0x0000000000000015 (DEBUG)              0x0
 0x0000000000000003 (PLTGOT)             0x421000
 0x0000000000000002 (PLTRELSZ)           2592 (bytes)
 0x0000000000000014 (PLTREL)             RELA
 0x0000000000000017 (JMPREL)             0x401e30
 0x0000000000000007 (RELA)               0x401b18
 0x0000000000000008 (RELASZ)             792 (bytes)
 0x0000000000000009 (RELAENT)            24 (bytes)
 0x000000006ffffffe (VERNEED)            0x401af8
 0x000000006fffffff (VERNEEDNUM)         1
 0x000000006ffffff0 (VERSYM)             0x4019dc
 0x0000000000000000 (NULL)               0x0
$ ./pystone
Pystone(1.1) time for 50000 passes = 0.171947
This machine benchmarks at 290787 pystones/second
  COPY

可以看出，编译成二进制文件后，性能上略有提升，并且需要依赖libpython3.7m.so才能运行。下面是pystone.c文件的部分代码：

  /* "pystone.py":73
 *     return Proc0(loops)
 * 
 * IntGlob = 0             # <<<<<<<<<<<<<<
 * BoolGlob = FALSE
 * Char1Glob = '\0'
 */
  if (PyDict_SetItem(__pyx_d, __pyx_n_s_IntGlob, __pyx_int_0) < 0) __PYX_ERR(0, 73, __pyx_L1_error)
  /* "pystone.py":74
 * 
 * IntGlob = 0
 * BoolGlob = FALSE             # <<<<<<<<<<<<<<
 * Char1Glob = '\0'
 * Char2Glob = '\0'
 */
  __Pyx_GetModuleGlobalName(__pyx_t_7, __pyx_n_s_FALSE); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 74, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_7);
  if (PyDict_SetItem(__pyx_d, __pyx_n_s_BoolGlob, __pyx_t_7) < 0) __PYX_ERR(0, 74, __pyx_L1_error)
  __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0;
  /* "pystone.py":75
 * IntGlob = 0
 * BoolGlob = FALSE
 * Char1Glob = '\0'             # <<<<<<<<<<<<<<
 * Char2Glob = '\0'
 * Array1Glob = [0]*51
 */
  if (PyDict_SetItem(__pyx_d, __pyx_n_s_Char1Glob, __pyx_kp_u__12) < 0) __PYX_ERR(0, 75, __pyx_L1_error)
  /* "pystone.py":76
 * BoolGlob = FALSE
 * Char1Glob = '\0'
 * Char2Glob = '\0'             # <<<<<<<<<<<<<<
 * Array1Glob = [0]*51
 * Array2Glob = list(map(lambda x: x[:], [Array1Glob]*51))
 */
  if (PyDict_SetItem(__pyx_d, __pyx_n_s_Char2Glob, __pyx_kp_u__12) < 0) __PYX_ERR(0, 76, __pyx_L1_error)
  /* "pystone.py":77
 * Char1Glob = '\0'
 * Char2Glob = '\0'
 * Array1Glob = [0]*51             # <<<<<<<<<<<<<<
 * Array2Glob = list(map(lambda x: x[:], [Array1Glob]*51))
 * PtrGlb = None
 */
  __pyx_t_7 = PyList_New(1 * 51); if (unlikely(!__pyx_t_7)) __PYX_ERR(0, 77, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_7);
  { Py_ssize_t __pyx_temp;
    for (__pyx_temp=0; __pyx_temp < 51; __pyx_temp++) {
      __Pyx_INCREF(__pyx_int_0);
      __Pyx_GIVEREF(__pyx_int_0);
      PyList_SET_ITEM(__pyx_t_7, __pyx_temp, __pyx_int_0);
    }
  }
  if (PyDict_SetItem(__pyx_d, __pyx_n_s_Array1Glob, __pyx_t_7) < 0) __PYX_ERR(0, 77, __pyx_L1_error)
  __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0;
  COPY

可以看出，C代码本身就已经很难读懂了，编译后的二进制文件基本是不可能还原出原始的python代码的。

不过，目前这种方式有个问题，就是只能编译单个文件。但很多时候，我们是想将多个包编译成一个独立的可执行文件。

0x03 使用Nuitka编译Python脚本

$ pip install nuitka
$ nuitka pystone.py
Nuitka-Options:INFO: Used command line options: pystone.py
Nuitka-Options:WARNING: You did not specify to follow or include anything but main program. Check options
Nuitka-Options:WARNING: and make sure that is intended.                                                     
Nuitka:WARNING: Using very slow fallback for ordered sets, please install 'orderedset' PyPI package for best
Nuitka:WARNING: Python compile time performance.                                                            
Nuitka:INFO: Starting Python compilation with Nuitka '1.1.8' on Python '3.7' commercial grade 'not installed'.
Nuitka:INFO: Completed Python level compilation and optimization.
Nuitka:INFO: Generating source code for C backend compiler.
Nuitka:INFO: Running data composer tool for optimal constant value handling.
Nuitka:INFO: Running C compilation via Scons.
Nuitka-Scons:INFO: Backend C compiler: gcc (gcc).
Nuitka-Scons:INFO: Backend linking program with 9 files (no progress information available).
Nuitka-Scons:WARNING: You are not using ccache.
Nuitka:INFO: Keeping build directory 'pystone.build'.                                                       
Nuitka:INFO: Successfully created 'pystone.bin'.
$ ls -l pystone.bin
-rwxrwxrwx 1 drunkdream drunkdream 268440 Sep  6 20:57 pystone.bin
$ ./pystone.bin
Pystone(1.1) time for 50000 passes = 0.12965
This machine benchmarks at 385654 pystones/second
  COPY

可以看到使用nuitka性能上明显比原生的Python要高出许多。

本来想在Python 3.11下测试下性能，不过发现目前最新版本的nuitka还没适配Python 3.11，编译会有报错。

nuitka还有些可选参数，比较重要的有：

• -o FILENAME: 指定要生成的文件名
• --standalone: 将依赖库都编译到一个文件中，不过对于依赖的动态链接库，还是会以多个文件的形式存在
• --onefile: 这个参数可以解决--standalone参数会有多个文件的问题，保证最终生成的是一个可执行文件
• --nofollow-imports: 不编译import进来的第三方库
• --clang: 强制使用clang作为编译后端
• --static-libpython=yes: 静态链接libpython
• --show-scons: 显示编译C代码过程中的详细日志

通过观察可以发现，nuitka也是通过将python代码转换成C代码，然后编译成最终的可执行文件。使用--static-libpython=yes参数可以静态链接libpython库，使用的命令行如下：

$ gcc -o pystone.bin -fuse-linker-plugin -flto=8 -fpartial-inlining -freorder-functions -O2 -s -z noexecstack -Wl,-R,'/usr/lib' -Wl,--disable-new-dtags -Wl,-b -Wl,binary -Wl,./__constants.bin -Wl,-b -Wl,elf64-x86-64 -Wl,-defsym -Wl,constant_bin_data=_binary_____constants_bin_start @"./@link_input.txt" -L/usr/lib -ldl -lm /usr/lib/libpython3.7m.a
  COPY

不过在实际执行时会有报错，原因是命令行中没有包含-lz -lpthread -lexpat -lutil等参数，针对这个问题有一个专门的issue。

0x04 结论

相比于py2exe、pyinstaller等方案，Cython和Nuitka采用了先生成C代码，再进行编译的方案，相对来说安全性和性能上都优于前两种方案。

而Nuitka相比Cython，可以同时编译多个Python脚本，功能上更加强大一些，性能也提升了不少。不过Nuitka使用--onefile参数生成的可执行文件大小会远大于pyinstaller生成的文件大小。