Jeff Law wrote:
> On 09/09/2017 02:51 AM, Eric Botcazou wrote:
> >> No, the stack never gets misaligned - my patch doesn't change that at all.
> >
> > Yes, it does. Dynamic allocation works like this: the amount to be
> > allocated
> > is added to VIRTUAL_STACK_DYNAMIC_REGNUM and the result is then dynamically
> > aligned. Your patch assumes that VIRTUAL_STACK_DYNAMIC_REGNUM is as
> > aligned
> > as the stack pointer, i.e. STACK_BOUNDARY, but that's wrong for 32-bit
> > SPARC
> > at least (that's why I put the ??? note at line 5746 in emit-rtl.c).
> This seems like a SPARC target problem to me -- essentially it's
> claiming a higher STACK_BOUNDARY than it really has.
>
> Presumably there's a good reason for this and some kind of hack may be
> needed to deal with it in dynamically allocated space. But it does not
> seem like we should be forcing all targets to allocate unnecessary space
> to deal with this.
It's not just STACK_BOUNDARY, the outgoing argument offset is incorrect too.
These snippets of code from PR78468 (comment 20) look very wrong:
sub %sp, %g2, %sp
add %sp, 108, %g3 ; g3 = fp - 28 (x)
sub %sp, %g2, %sp
add %sp, 108, %g2 ; g2 = fp - 44 (d)
sub %sp, %g1, %sp
add %sp, 112, %g1 ; g1 = fp - 56 (e)
There are several different outgoing argument offsets used here (108 and 112).
This not only results in alloca blocks being unaligned (when they should be at
least
aligned to STACK_BOUNDARY, ideally PREFERRED_STACK_BOUNDARY),
but this also means you end up with different alloca blocks overlapping and
corrupting
each other in non-trivial ways...
Wilco
