Re: [patch -v2] flat: fix data sections alignment

[Johannes Weiner <jw@emlix.com>]

On Wed, Mar 04, 2009 at 03:00:25PM -0500, Mike Frysinger wrote:
> On Wed, Mar 4, 2009 at 14:33, Johannes Weiner wrote:
> > On Wed, Mar 04, 2009 at 01:04:00PM -0500, Mike Frysinger wrote:
> >> On Wed, Mar 4, 2009 at 08:51, Johannes Weiner wrote:
> >> > The flat loader uses an architecture's flat_stack_align() to align the
> >> > stack but assumes word-alignment is enough for the data sections.
> >> >
> >> > However, on the Xtensa S6000 we have registers up to 128bit width
> >> > which can be used from userspace and therefor need userspace stack and
> >> > data-section alignment of at least this size.
> >>
> >> could this perhaps be a gcc problem ?  x86 has a similar problem with
> >> sse and they addressed it with a function attribute.  after all, just
> >> because your stack started out 128bit aligned doesnt mean gcc will
> >> keep it that way when calling other functions.  so having the stack
> >> start out aligned would only "fix" the stack for the application's
> >> entry point right (which would in practice bubble up to main()) ?  so
> >> you'd be right back where you started ...
> >
> > gcc generates sp changes only ever in multiples of 16 deltas, I just
> > checked it again with various amounts of stack variables.
> >
> > The stack frames allocate themselves with an ENTRY instruction and the
> > generated code I read here allocates stack frames of n * 16 bytes.
> >
> > So we are good to go as long as the initial stack frame is properly
> > aligned.
> 
> throwing a few random cases at gcc isnt really a good way to validate.
>  this would have worked for x86 too with older versions.  only when
> common code in later gcc versions got more aggressive with stack
> packing did people notice the issue.
> 
> so, lets look at the authoritative place: the gcc source code for xtensa
> 
> $ grep define.*STACK_BOUNDARY -B 2 gcc/config/xtensa/*.h
> xtensa.h-/* Align stack frames on 128 bits for Xtensa.  This is necessary for
> xtensa.h-   128-bit datatypes defined in TIE (e.g., for Vectra).  */
> xtensa.h:#define STACK_BOUNDARY 128
> 
> ok, now i believe that forcing a stack alignment of 128bits in the
> kernel is correct ;)

Now I do too.  Heh.

Seriously, thanks for fishing this out.  Is this an Ack? ;)

> -mike

	Hannes