Re: [tlug] Alan Cox's remark at Fosdem

[ijw@example.com]

> On Tue, 1 Mar 2005 00:42:04 -0000 (GMT), ijw@example.com
> <ijw@example.com> wrote:
>> > That is what the assembler code should handle. In a modern kernel, if
>> > it is in C, it should not be tied to any one architecture.
>> >
>> > -Josh
>>
>> That's not really fair.  For instance, x86 architectures have low memory
>> and high memory, a 2 level VM tree, small and large pages (sometimes);
>> anything written to support these memory structures is going to be x86
>> specific.  That doesn't mean that you should be writing it in assembler.
>
> Perhaps I'm missing your point and digressing, but I think that Josh's
> statement is a fair statement. The compiler, linker and assembler
> should handle the architecture not the code per se. This is where GNU
> compilers go wrong by being all things to all architectures. In
> comparision see how it was handled by BL:
>
> http://plan9.bell-labs.com/sys/doc/comp.html
> http://plan9.bell-labs.com/sys/doc/compiler.html

I was just being pedantic about the word "architecture" - the problem with
code is not that it's specific to a processor - which is quite reasonable
when you're trying to write a driver for the MMU, for instance - but when
you write code that is dependent on the compiler treating it in exactly
the right way.

GCC having a generic retargettable backend is another issue, and there are
advantages to it as well as disadvantages. It's a compromise, like any
other design decision.

-- 
Ian.