Re: FYI: SIGBUS with world built by clang

On Thu, Jul 05, 2012 at 12:32:01AM +0900, Taku YAMAMOTO wrote:
> On Wed, 04 Jul 2012 17:08:30 +0200
> Dimitry Andric <dim_at_FreeBSD.org> wrote:
> 
> > On 2012-07-04 16:33, Taku YAMAMOTO wrote:
> > > For people having SIGBUS with clang-build world + gcc-build binaries,
> > > 
> > > 
> > > In short words, for any libraries (and never forget about rtld-elf!)
> > > which are potentially called from arbitrary binaries,
> > > compile them with either -mstackrealign or -mstack-alignment=8!
> > > 
> > > The detail is as follows.
> > > 
> > > I've observed that clang carelessly expects the stack being aligned at
> > > 16 byte boundary.
> > 
> > Eh, this is a requirement of the amd64 ABI.  Any compiler that *doesn't*
> > align the stack on 16-byte boundaries is basically broken.  Or are you
> > experiencing this on i386?  Even there, 16-byte alignment would be much
> > better in combination with SSE instructions (which arent' enabled by
> > default, btw).
> 
> Oops, I had to be clear about that!
> Yes, the experiment was took on i386 (actually -march=pentium4).
> 
> > Note that you would get the same issue with newer versions of gcc, which
> > will also assume this alignment.
> 
> Interesting, but the base gcc we currently have won't on i386, I think.
> (I occationally get bitten by similar problem when using -ftree-vectorize)
As far as I understand the rules, $esp % 16 must be zero before call
instruction is executed.

i386 csu explicitely aligns the stack before calling into C land, everything
else should be the C compiler own offence :).