On Mon, Nov 17, 2014 at 05:46:55PM +0000, Joseph Myers wrote:
> On Mon, 17 Nov 2014, Jakub Jelinek wrote:
>
> > > If it is true that a type satisfying TYPE_USER_ALIGN will never be
> > > allocated at an address less-aligned than its TYPE_ALIGN, even if that's
> > > greater than BIGGEST_ALIGNMENT, then the change seems correct for C11
> > > _Alignof.
> >
> > I think it depends on which target and where.
> > In structs (unless packed) the user aligned fields should be properly
> > aligned with respect to start of struct and the struct should have user
> > alignment in that case, automatic vars these days use alloca with
> > realignment if not handled better by the target, so should be fine too.
> > For data section vars and for common vars I think it really depends on the
> > target. Perhaps for TYPE_USER_ALIGN use minimum of the TYPE_ALIGN and
> > MAX_OFILE_ALIGNMENT ?
> > For heap objects, it really depends on how it has been allocated, but if
> > allocated through malloc, the extra alignment is never guaranteed.
> > So, it really depends in malloc counts or not.
>
> The question, for both _Alignas and ubsan, is the alignment guaranteed *in
> valid programs*.
>
> malloc only provides sufficient alignment for types with fundamental
> alignment requirements (although there are various problems with the C11
> wording; see DR#445). So if you use malloc to allocate a type with an
> extended alignment requirement (without doing extra realignment on the
> result of malloc), that's not a valid program. And if an alignment is
> larger than MAX_OFILE_ALIGNMENT, you get an error for declaring non-stack
> variables requiring that alignment. So I don't think there's any need to
> check MAX_OFILE_ALIGNMENT here.
If so, then Yuri's original patch (the one changing min_align_of_type)
should be fine, right?
Jakub
