Re: IRIX6 Makefiles

["Kaveh R. Ghazi" <ghazi at caip dot rutgers dot edu>]

 > From: Mark Mitchell <mark@codesourcery.com>
 > 
 > Gavin --
 > 
 >   There's something distinctly odd about the configure process on
 > IRIX6.
 > 
 >   Because of historical bugs in IRIX cc, we have in the toplevel
 > config/mh-irix6:
 > 
 >   # Specify the ABI, to ensure that all Irix 6 systems will behave the same.
 >   # Also, using -32 avoids bugs that exist in the n32/n64 support in some
 >   # versions of the SGI compiler.
 >   CC = cc -32
 > 
 > Correspondingly, in gcc/config/mips/x-iris6, we have:
 > 
 >   # We force the use of the O32 ABI for two reasons.
 >   # 1) For consistency, because some versions of Irix 6 default to the O32 ABI
 >   #    and some versions default to the N64 ABI.
 >   # 2) To avoid SGI compiler bugs.  The v6.x and v7.0 compilers from SGI have
 >   #    bugs that cause gcc to be miscompiled when the N32 or N64 ABIs are used.
 >   #    The O32 ABI is known to be OK.
 >   CC = $(OLDCC)
 >   OLDCC = cc -32
 > 
 > OK, so far so good.  The point of all this is to try to force the
 > build system to use `cc -32' for building stage1.
 > 
 >   But, if you have GCC in your path, the toplevel configure finds it.
 > And because its sed replacements happen *after* the host fragment is
 > inserted in the Makefile, the top-level Makefile ends up with:
 > 
 >   CC = gcc
 > 
 > even though the GCC Makefile has the `CC = $(OLDCC); OLDCC = cc -32'
 > bit.
 > 
 >   Now, everything is sort-of OK if you do a bootstrap because the
 > CC=gcc is passed down to the GCC Makefile, meaning that you build
 > everything with GCC.  But heaven help you if you go back to the GCC
 > directory and say make after making a change; you get the `cc -32'
 > stuff, and since your initial build was done with GCC, which defaults
 > to the N32 object model, not the o32 object model, you're hosed.
 > 
 >   I think that we could just lose the O32 override bits at this point.
 > Or, we could make the top-level configure leave teh attempted override
 > of CC alone.  Or, we could have the top-level configure notice whether
 > or not you're using GCC, and old add -32 in that case.
 > 
 >   I think the GCC-specific override should go in any case; that's just
 > inviting the kind of skew that's present between the top-level and the
 > GCC subdirectory.
 > 
 >   Thoughts?

If it ever happened that the "cc -32" stuff worked, you couldn't
bootstrap with cc for stage1.  Consider, the top level libiberty would
be compiled in 032 mode.  Since gcc defaults to N32, anything linked
in stage2 against libiberty would choke.

As I mentioned to you in private email many weeks ago, I had suggested
in a patch from 1998 that we rebuild libiberty in the gcc dir at each
stage to make sure the ABIs match.  This was rejected at the time in
favor of forcing one to set CC='cc -n32' for stage1 on Irix6.  I've
been doing this myself ever since, however anyone else bootstrapping
with cc on Irix must figure this out for themselves.  Anyone with the
so-called "buggy" 6.0 or 7.0 compilers simply cannot 3-stage
bootstrap.  They have to single stage a gcc and use that to recompile
the top level libibery and then 3-stage bootstrap.

Needless to say, I agree the current situation is distinctly odd.

		--Kaveh
--
Kaveh R. Ghazi			Engagement Manager / Project Services
ghazi@caip.rutgers.edu		Qwest Internet Solutions