Re: Minimal machine description file

[Alexandre Courbot <Alexandre dot Courbot at lifl dot fr>]

> I have managed to modify my configure scripts so that gcc recognizes my new
> target (called "npe"). Thanks to Alex Courbot for his help here.
> My next challenge is to write my machine description files. Scrolling
> through the gcc archives, I realised that many people have been looking for
> the minimal characteristics to achieve a functioning port of gcc. I was
> wondering if anybody had achieved this and if so where would I find it
> documented?

The GCC Internals manual has a detailled description of the macros that take 
place in target.h chapter 10. It is said when a macro is optional, so I'd 
advice you to go through this chapter (or rather, go through the entire 
manual to get a good overview of the compiler first) and start implementing 
the macros you need and those which are mandatory. Then, chapter 9 has a 
description of all the instructions names you can use to match your RTL 
templates. The most obviously needed being moving instructions, some basic 
arithmetic, extension, jumping, calling. Chapter 8 and 9 are *very* important 
for this part.

> If not, would anybody recommend some target files for a similar
> architecture to mine. y intention would then be to modify these files to
> accomodate my architecture.

In my experience I have started my port completely from scratch. Having a look 
at some description files of architectures I knew were useful, but I haven't 
started with a modified version of some other arch. My architecture is quite 
particular however, so your mileage may vary. Anyway, starting a first port 
is something I've found very difficult, but perseverance do pay. Now that you 
have the configure stuff working, don't dive into coding. Read the manual 
carefully, make sure you understand the Parse tree/RTL generation/RTL 
matching stuff. Then, start writing your target.h file the following way: 
order your macros in the same order as they are in the manual, and copy their 
documentation into comments above each of them. Then write a basic target.md 
and target.c. You won't be able to compile gcc before you get these files 
written in their minimal form. So your goal should be to get something that 
compiles, then something that runs without segfaulting even if it doesn't 
produce something useful. Once you have a basic skeleton, start writing your 
macros the right way. I don't know whether a skeleton architecture (something 
that compiles and run, but doesn't produce anything), heavily commented, 
could be doable/useful considering how much architectures are different. 
Like, putting in gcc/config a 'dummy' directory which doesn't compile 
anything useful but makes a sane basis for newporters. Maybe I can submit it 
in this case, if I can find my earlier work. I remember I had needed one 
whole month (plus several stupid questions here) to get my first successful 
compilation.

Alex.