[Fwd: Re: [Discuss] Interpreters vs Compilers]

[David Bronaugh]

John Blomfield wrote:
> pw wrote:
>> John Blomfield wrote:
>>>
>>> I don't have anything more authoritative to add than I did last 
>>> night, except to say again, that Python and Java for example compile 
>>> their source code into "bytecode" which is slightly optimized code 
>>> but it is not machine code. A byte code file still has to be 
>>> interpreted, which means translated into machine code instructions 
>>> line by line by the so called "Virtual Machine".  Each platform 
>>> needs the Virtual Machine running to run a bytecode file.  In the 
>>> case of a loop, the loop is in bytecode and must be interpreted line 
>>> by line every time through the loop, into possibly many many machine 
>>> instructions.  In the case of a compiled C program the binary is in 
>>> machine instructions that the CPU understands without 
>>> interpretation, hence the speed.  The optimization in C is very very 
>>> good, so good that it is almost impossible for the average 
>>> programmer to do better using Assembler.  I once did some 
>>> programming of micro-controllers, which are quite slow compared to 
>>> the typical PC, in both Microchip Assembler and Microchip C and the 
>>> C compiler always won.  This one of the reasons most hardware is 
>>> programmed in C, including Linux and Python I think.  Python 
>>> certainly uses a lot of built in mathematical functions that are 
>>> written in C to help speed its execution.
>>>
>>> Its all a question of horses for courses!  Scripting languages like 
>>> Perl and Python are often referred to as the "glue" that sticks the 
>>> C and C++ modules together.  So if you want Zoom Zoom go C!
>>>
>>> John Blomfield
>>
>> byte code is a type of pseudo machine code.
>> This code is run in a virtual machine much like
>> emulating the operation of a microprocessor.
>>
>> The java virtual machine, for example, is merely a
>> microprocessor/toolkit  architecture that has been defined
>> to use a machine language of java byte code. The virtual machine
>> is compiled for various computers to provide java system emulation.
>>
>> The speed of this emulation is related to the optimizations built
>> into the virtual machine as compiled for your particular system.
>>
>> (ie: system meaning I86 or PPC or sparc or etc..)
>>
>> Hence, the byte code process is not exactly akin to
>> 'interpretation'. The interpretation has already been
>> done in the compilation from source code to byte code.
>>
>> In theory byte code processing can be as fast as
>> machine code since the virtual machine can directly
>> transfer the values of byte code to applicable machine
>> level instructions prior to running the application,
>> thus running the application in 'native' machine code.
>>
> Are you saying that the Python Virtual Machine reads all the pyc 
> bytecode file and converts it to binary and then runs it?  If so this 
> is contrary to my understanding.  About Java I am not sure.
The Python interpreter compiles Python code to bytecode, and then the VM 
runs it.

FWIW, the Java VM (1.3, at least) is about 4x slower than C/C++ when I 
tested it. The Ocaml bytecode interpreter is about 18x slower. I haven't 
done any playing around with Python bytecode, but I would assume it'd 
fall somewhere in between those two speeds.

The only really fair way to test how much better a bytecode interpreter 
is, is to test it on a short bit of code that takes very little time to 
convert to bytecode, but takes a proportionally much longer time to run. 
Then you see the full advantages without confounding effects of the time 
required to convert to bytecode.

It's true what pw says; at least in theory, it should be possible for 
bytecode to run as fast as native. But in practice it doesn't tend to 
work out that way.

David