Freebsd-7.4 + std gcc 4.2.1 fails to honour -march=i586

[Julian H. Stacey]

Hi John & all,

> On Wednesday, July 20, 2011 8:34:57 am Julian H. Stacey wrote:
> > Hi,
> > Reference:
> > > From:		John Baldwin <jhb at freebsd.org> 
> > > Date:		Wed, 20 Jul 2011 07:48:05 -0400 
> > > Message-id:	<201107200748.05786.jhb at freebsd.org> 
> > 
> > John Baldwin wrote:
> > > On Wednesday, July 20, 2011 7:33:26 am Julian H. Stacey wrote:
> > > > What should FreeBSD do ? 
> > > > 	Add a comment to man gcc ... that -march=i586 is not
> > > > 	enough, & feed the comment back to gcc project & see how
> > > > 	they want to handle it ?
> > > 
> > > No, this is not a GCC bug.  If you want to use a single build machine that 
> > > will compile programs for other machines on a network to use, it must use the 
> > > lowest common denominator for its CPUTYPE in /etc/make.conf.
> > > 
> > > The out-of-the-box crt files from a FreeBSD install will work fine on a 486
> > > and above, so can only get yourself into this quandry by building a new world 
> > > with a CPUTYPE (or similar CFLAGS) setting in /etc/make.conf that violates 
> > > this rule.
> > > 
> > > You can fix your machine by fixing the CPUTYPE in your build machine and 
> > > building and installing a new world (do use NO_CLEAN=yes for your build, do a 
> > > full build).  You will also want to rebuild any other binaries on this machine 
> > > that you share with other machines.
> > > 
> > > -- 
> > > John Baldwin
> > 
> > 
> > Hi John, 
> > Yes I realise all that now thanks.
> > The point now is man gcc is misleading/ incomplete:
> > 	"-march=cpu-type
> >            Generate instructions for the machine type cpu-type."
> > Those instructions will not make a program that runs on a lesser CPU type
> > without alternate crt for lesse CPU.
> > 
> > CFLAGS in make.conf is fine for single developer/ host debugging,
> > but is not appropriate for multi user multi make builds os src/ ports/ other
> > 
> > Its seems just that gcc -march fails to select alternate crt.
> > 
> > One gross ugly solution needing root would be to create chroots,
> > but Too ugly.
> > 
> > I suggest gcc via -march should also be used as a selection for sub
> > dirs for crt ? & we should co-operate with gcc project on that,
> > *BSD & *linux must all face same phenomena so lets not develop
> > a non standard solution.
> 
> I think this is a harder problem than you expect.

Yes, Needs to be thought through, that's why I posted 13:33:26 +0200 :

> > Question: with more complex programs than who, might there 
> > be sensitivity to other libs too ?


Per Andriy's :

> Talk to GCC people now?

I'm not clear what's best, but yes, when its clear, then talk to gcc people.


> It is not just the crt files
> that matter, but every library.  You would need CPU-specific versions of every
> static library on the build system,

Would be preferable.
Though as we default to link dynamic, not as essential as crt.o .
We could also have some ifdef to avoid by default generating for
all of [3-6]86 that would save space & time,
If the directory was not filled, a user of gcc --march=i586 would
then fail to link from /usr/lib/i586/ & could receive a good error,
instead of a 686 crt.o with no error.


> and possibly you would want to do this for
> all shared libraries too.

I guess we could skip those, Presuming a target system should have a
matched set for the right CPU & release etc.


I'm not sure about the recent plethora of lib stuff
	/lib 
		Seems to be all .so so I guess not affected.
	/libexec 
		ld-elf.so.1 ld-elf32.so.1	# no man ld-elf
	/usr/lib 
		This is where I'd suggest sub dirs of [3-6]86
		with default machine native either
			 in top dir 
			or perhaps a native/ sub dir,
			or sym linked
		whatever gcc & out & other projects find best ?
	/usr/lib32 
		... Sigh ! same as above ? ..
	/usr/libdata /ldscripts & /lint
		?
	/usr/libexec
		Presumably can ignore this as real programs.

	& ignore inderminent localy variant dirs shown by
		cd /usr/local ; find . -type d -name \*lib\* -print

Cheers,
Julian
-- 
Julian Stacey, BSD Unix Linux C Sys Eng Consultants Munich http://berklix.com
 Reply below, not above;  Indent with "> ";  Cumulative like a play script.
 Format: Plain text. Not HTML, multipart/alternative, base64, quoted-printable.