Index ¦ Archives ¦ Atom

Compilation and Interpretation in C and CPython

It is common knowledge that programs written in high level languages have to be translated into a low level language using programs referred to as translators. This low level language is either in a native form, in the sense that it is understood by the operating system itself, or in an intermediate form which is understood by an intermediate program such as the bytecode interpreter. It is also known that C is a compiled language, whereas CPython is both first compiled and then interpreted.

In this article, I will try to illustrate the difference between the two languages by carrying out some simple experiments (on Linux).

Consider a program, helloworld.c:

# include<stdio.h>

int main(int argc, char **argv)
{
   printf("Hello World!\n");
   return 0;
}

We can compile and execute this program as follows:

$ gcc -o helloworld helloworld.c
$ ./helloworld
Hello World

The executable file helloworld is the low level language equivalent of the high level language program, helloworld.c. This is what the operating system on your computer understands and hence when it is executed, it prints Hello World on the screen. This process of converting helloworld.c to helloworld represents the translation process. In the case of C, this translation process is performed by the compiler, gcc (this translation process is really a process pipeline and involves two other processes preprocessing and linking, which are carried out by separate programs, automatically invoked by gcc). Nevertheless, compilation is at the core of the translation process of C programs and is responsible for converting a high level language to its low level equivalent - a version readily executable by the operating system on your computer. It is important to note that this executable file is composed of the instructions you wrote in your C program, along with a other details which are necessary to execute your program. These details are specific to the architecture and operating system you created the executable file on and hence if you copy a executable file you created on a computer with an Intel processor, it will not work at all on a computer which has an ARM processor, for example. Hence, you will have to recompile the program on the new computer before you can execute it. Note that your C program still remains the same, but the low level machine language equivalent is different and gcc takes care of this translation.

Now, let us consider our first CPython program:

# Print Hello world
if __name__=='__main__':
    print 'Hello World!'

You executed this program as follows

$ python helloworld.py
Hello World!

Unlike C, where you compiled the program first to get a separate executable file and then executed it (a two step process), here you executed the program in a single step - your program is directly executed on-the-fly. This is how traditionally interpreters (interpreted languages) worked. However, modern day interpreted languages like CPython (and others) also involve a compilation step. Your program helloworld.py is first converted to an intermediate representation which is a low level equivalent of your high level language program. The difference from C is that the instructions in these low level language equivalents are not meant to be executed by a real computer, but a process virtual machine [1]. In the case of CPython, the intermediate representation is known as bytecodes and the virtual machine referred to as the bytecode interpreter or the CPython virtual machine. Hence, the CPython code is first converted into its bytecode equivalent which is then executed by the bytecode interpreter. When you run a CPython program using python helloworld.py, both these steps happen in the background.

While discussing the section on C compilation I mentioned that the executable you create on an Intel computer will not run on an ARM computer, because of the architecture specific instructions embedded into the executable required for executing the program. In the case of CPython the bytecodes (result of compilation of the CPython program) are executed by the CPython virtual machine, instead of the real computer. This extra layer of abstraction allows you to execute the same bytecodes (without recompiling) on an Intel computer and an ARM computer, for example.

Let us understand this better with a real example. I will use two computers for the experiments: System1 and System2, with both running Fedora Linux. However, System1's instruction set architecture is x86_64 (Intel) where as System2 is a RaspberryPi with an armv6l (ARM) instruction set.

C

First, I will the consider the helloworld.c program. I will compile this program on System1:

$ arch
x86_64
$ gcc -o helloworld helloworld.c
$ file ./helloworld
./helloworld: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.32,
BuildID[sha1]=0xc50d74290927cb25ef9e34055af6c437e89ed5eb, not stripped

The file command shows the type of a file [2] and from the above output, the key information for us is that the file helloworld is a ELF 64-bit LSB executable, x86-64, version 1 (SYSV). You can of course execute the program as we have done earlier using ./helloworld.

Now, copy the file, helloworld to System2, and try to execute the object file:

$ arch
armv6l
$ file helloworld
helloworld: ELF 64-bit LSB executable, x86-64, version 1 (SYSV),
dynamically linked (uses shared libs), for GNU/Linux 2.6.32,
BuildID[sha1]=0xc50d74290927cb25ef9e34055af6c437e89ed5eb, not stripped
$ ./helloworld
-bash: ./helloworld: cannot execute binary file

It is clear from the above error message, that helloworld could not be executed on System2. Now, transfer the helloworld.c file to System2 and compile and execute the file as on System1:

$ gcc -o helloworld helloworld.c
$ file helloworld
helloworld: ELF 32-bit LSB executable, ARM, version 1 (SYSV),
dynamically linked (uses shared libs), for GNU/Linux 2.6.32,
BuildID[sha1]=0xba57691af19ff94f894645398e66e263c8f57a9b, not stripped
$ ./helloworld
Hello World!

As you can see, the file format of helloworld is different on System2 as expected and hence it had to be recreated to execute it.

CPython

On System1, create the compiled version of helloworld.py using the following code [3]:

$ python -c "import py_compile;py_compile.compile('helloworld.py')"

Or, the cleaner version: $ python -m py_compile helloworld.py. This will create a helloworld.pyc file in your directory. Once again, we can use the file command to see the file type of helloworld.pyc:

$ file helloworld.pyc
helloworld.pyc: python 2.7 byte-compiled

To execute the compiled file, simply invoke the python interpreter with the helloworld.pyc file as an argument, rather than the source file: python helloworld.pyc.

Now, copy the file helloworld.pyc to System2 and try to execute it:

$ arch
armv6l
$ file helloworld.pyc
helloworld.pyc: python 2.7 byte-compiled
$ python helloworld.pyc
Hello World!

To summarize, the compiled helloworld.pyc could be executed without being re-created from its source file, helloworld.py on two systems with different instruction set architecture. This was made possible by the python bytecode interpreter on the two systems, which created an abstraction between the bytecodes and the native instruction set architecture [4]. I should mention here that if your CPython application has anything to do beyond pure CPython code (C extension, for example), the results of the experiments here will not be applicable.

python

The CPython executable, python is nothing but a ELF file (similar to your helloworld but obviously created from a more complicated set of C source files). The almost magical behavior of CPython bytecodes that we saw in the previous section is made possible by python taking care of the steps necessary to execute the bytecodes on systems with different instruction set architecture. To understand this better, consider the following two commands, the first on System1 and the second on System2:

$ file /usr/bin/python2.7
/usr/bin/python2.7: ELF 64-bit LSB executable, x86-64, version 1
(SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.32,
BuildID[sha1]=0x9d8a414b778ff11ec075995248c43cdf5b67f17a, stripped

$ file /usr/bin/python2.7
/usr/bin/python2.7: ELF 32-bit LSB executable, ARM, version 1 (SYSV),
dynamically linked (uses shared libs), for GNU/Linux 2.6.32,
BuildID[sha1]=0x63fd81d3591769d6be0619b7273935ab9521010c, stripped

As is clear from the above output, the file /usr/bin/python2.7 (/usr/bin/python is symlinked to /usr/bin/python2, which is in turn symlinked to /usr/bin/python2.7 in reality), is an ELF executable and it has obviously been compiled separately on both these systems (thus showing the different ELF file formats).

Conclusion

The above experiments have hopefully shed some light on C being a compiled language and CPython being a compiled and interpreted language - this design leads to its interoperability between different architectures.

However, it is important that I mention a language is not compiled or interpreted. That is, it is not technically 100% accurate to say that C is a compiled language. A language implementation, rather than the language is compiled or interpreted. There are interpreters for the C language which interpret your C programs and there are CPython implementations which are compiled (The water gets murkier in the case of CPython, and the boundary between being compiled and interpreted not always clear).

Footnotes

[1]Note that, there are two kinds of virtual machines that can be implemented in software: system virtual machine and process virtual machine. Here, I am referring to the process virtual machine. See the Wikipedia article on Virtual Machine to learn more.
[2]http://linux.die.net/man/1/file
[3]The py_compile module can be used to compile a CPython program into its bytecode equivalent. This is the version of your program that is executed the CPython bytecode interpreter. See: http://docs.python.org/2/library/py_compile.html.
[4]Actually, to be more accurate, the python executable takes care of the interfacing with the operating system kernel (Linux Kernel), which is once again different on systems with different instruction set architecture.

Share on: Diaspora*TwitterFacebookGoogle+Email

© Amit Saha. Built using Pelican. Theme by Giulio Fidente on github.