Re: What if?

[J.A. Magallon]

On 2004.12.02, Theodore Ts'o wrote:
> On Thu, Dec 02, 2004 at 05:34:08AM +0530, Imanpreet Singh Arora wrote:
> > 
> >    I realize most of the unhappiness lies with C++ compilers being 
> > slow. Also the fact that a lot of Hackers around here are a lot more 
> > familiar with C, rather than C++. However other than that what are the  
> > _implementation_  issues that you hackers might need to consider if it 
> > were to be implemented in C++. 
> 
> The suckitude of C++ compilers is only part of the issues.
> 
> > My question is regarding how will kernel 
> > deal with C++ doing too much behind the back, Calling constructors, 
> > templates exceptions and other. What are the possible issues of such an 
> > approach while writing device drivers?  What sort of modifications do 
> > you reckon might be needed if such a move were to be made?
> 
> The way the kernel will deal with C++ language being a complete
> disaster (where something as simple as "a = b + c + d +e" could
> involve a dozen or more memory allocations, implicit type conversions,
> and overloaded operators) is to not use it.  Think about the words of
> wisdom from the movie Wargames: "The only way to win is not to play
> the game".
> 

Don't ge silly. I have written C++ code to deal with SSE operations
on vectors, and there you really need to control the assembler produced,
and with the correct const and & on correct places the code is as
efficient as C. Or more. You can control everything. No more temporal
objects, no more copies, but still the type checking. I can send you,
for example, the code for a Vector class (no __asm in it),
and the assembler that g++ spits for a thing like a = b+c.
You would do not better in handcoded asm.

It is as all things, you need to know it deeply to use it well.
There are a ton of myths around C++.

For the kernel, you don't need exceptions nor iostreams, so don't use
the C++ runtime library.

Constructor/destuctor is needed, and also virtual single inheritance.
And those two things are already being done by hand with function
pointers inside structs in the kernel. The cost of (single and multiple)
inheritance in C++ is also just an indexed jump, the difference is that
in the pseudo object oriented things done in the kernel you have to
always initializa the funtions pointers, and C++ will do it for you.
How many bugs have you seen because somebody wrote a bad
.method = the wrong_func in an initializer ?

The kernel is full of things like
if (__builtin_constant_pointer(xxxx) and others to check if an argument
is constant and optimize some path, and in C++ you could write
class T {
	T& f();
	const T& f() const;
}
and the compiler will do it at compile time also.

In short, with C++ you can generate code as efficient as C or asm.
You just have to know how.

But C++ is not supported in the kernel. I can live with it.

--
J.A. Magallon <jamagallon()able!es>     \               Software is like sex:
werewolf!able!es                         \         It's better when it's free
Mandrakelinux release 10.2 (Cooker) for i586
Linux 2.6.10-rc2-jam4 (gcc 3.4.1 (Mandrakelinux 10.1 3.4.1-4mdk)) #2


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/