Re: compile librump/compat.o failed on amd64

[Antti Kantee <pooka%NetBSD.org@localhost>]

On Sat, Jul 24, 2010 at 06:35:31PM +0200, Jean-Yves Migeon wrote:
> On 24.07.2010 17:16, Nicolas Joly wrote:
> > On Sat, Jul 24, 2010 at 03:37:10PM +0200, Jean-Yves Migeon wrote:
> >> On 24.07.2010 09:00, KS wrote:
> >>>
> >>> Hi,
> >>>
> >>> after cvs updateing -current some minutes ago, build.sh ... distribution
> >>> for amd64 fails:
> >>> [snip]
> >>> and so on and so on...
> >>
> >> Clean build.sh rumptest for i386 and amd64 did not raise that issue on
> >> my side.
> >>
> >> Currently investigating.
> > 
> > This breaks while compiling compat stuff ... 32bit rump libraries.
> > Well amd64 should be pulling i386 pte.h for 32bit compilation.
> > 
> > The following patch should make the build goes further.

Yea, the #ifdef game is the "right" way to fix these issues.

> Alright, got it.
> 
> I hardly see how my commit breaks the compat stuff though, I did not
> touch headers much with my patch (especially pte.h), most of them were
> already there with Xen's PAE.

An #include <machine/pte.h> in cpu.h can have quite a few side
effects.  But of course the rule is to do a _full_ build every time
you touch a header, since anything can be affected ...

That aside, I've been thinking of disabling rump in the compat
build.  It doesn't seem particularly useful since none of the driver
components (fs,net,dev,kern) get a compat build.  Either that or
add the drivers to the compat build too.  I'm not sure if the amd64
compat32 ABI is even the same as the i386 kernel ABI, i.e. could
you use i386 kernel modules on amd64 that way.  Comments?