Re: Once more cross compilation

[Fred Kiefer]

Nicola Pero wrote:
> > With my crocsscompiling attempt I did get much further after changing 
> > the intalled common.make to use:
> >
> > ifeq ($(GNUSTEP_FLATTENED),)
> >   WHICH_LIB_SCRIPT \
> >         = 
> > $(GNUSTEP_MAKEFILES)/$(GNUSTEP_BUILD_CPU)/$(GNUSTEP_BUILD_OS)/which_lib
> > else
> >   WHICH_LIB_SCRIPT = $(GNUSTEP_MAKEFILES)/which_lib
> > endif
>
>
> which_lib should be in the GNUSTEP_HOST_* keyed directory (so the existing
> code is correct) ... but if that doesn't work because which_lib is in the
> GNUSTEP_BUILD_* directory, then it's probably gnustep-make's own
> GNUmakefile's fault for installing it in the wrong place, so you probably
> want to fix GNUmakefile.in

Lets take up this discussion again. As the results from my tests seem to 
show, GNUstep should stick with the standard definitions of the three 
machine environment variables BUILD, HOST and TARGET (as used by 
autoconf) and change the documentation and usage to that. That is:

BUILD, the machine, where the configuration and compilation takes place.
HOST, the machine, where the resulting  application will be run.
TARGET, the machine, where the output of the application will be used. 
Mostly relevant when compiling a crosscompiler. GNUstep may savely 
ignore this in most cases.

That is GNUstep components will need to be configured for 
crosscompilation like this:
./configure --disable-flattened --host=arm-linux --build=ix86-linux

The only added complexity is that gnustep-make needs to be compiled 
twice, one additional time with HOST set to BUILD, as it is needed for 
the build process of the other GNUstep components. But this remark and 
that a non-flattend structure should be used is already in the 
documentation.

After this change in the documentation the above change is needed (As 
you surely will agree) to get the correct which_lib selected. The fine 
thing is that for environments where BUILD=HOST=TARGET this will make no 
difference also, in the flattend case. So the only environment where 
this may cause trouble could be the MINGW on Cygwin case, with the 
current build instructions, so I would suggest that we change them as 
well to use BUILD and HOST instead of TARGET.

I don't have any hard opinion on what we should do to the usage of 
GNUSTEP_HOST_XXX in the GNUstep.sh. They seem to be self contained and 
should not cause any trouble, so we could just keep them, although the 
names would now be inconsistent.

If you still disagree, I would like to get instructions on how to run a 
cross-compilation of GNUstep. Everything you suggested up to now just 
doesn't work and this suggested shema is consistent with the autoconf 
manual I have lying here on my desk. The only other idea I have still 
left is that we are using a rather strange way to sort out HOST and 
TARGET in the configure.ac file. The automake manual suggests to use 
AC_CANONICAL_SYSTEM whereas we use AC_CANONICAL_TARGET. But I don't 
expect that to make much of a change.

Fred