[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20140428132645.GH12324@port70.net>
Date: Mon, 28 Apr 2014 15:26:45 +0200
From: Szabolcs Nagy <nsz@...t70.net>
To: musl@...ts.openwall.com
Subject: Re: [PATCH 3/3] stddef: Define max_align_t
* Jens Gustedt <jens.gustedt@...ia.fr> [2014-04-28 14:22:44 +0200]:
> Am Montag, den 28.04.2014, 12:11 +0200 schrieb Szabolcs Nagy:
> > i think i386 abi is non-conforming to the c11 alignment requirements now:
> > long long has 8 byte alignment, but in a struct/union it has only 4
> > (this is why the attrs are needed above)
> >
> > long long x; // _Alignof(x) == 8
> > struct {long long x;} y; // _Alignof(y.x) == 4
>
> I don't think that it is non-conforming
>
> _Alignof of a type only tells you on what alignments the programmer
> may place objects of the corresponding type (if he deals with this
> manually) and gives no guarantee what the implementation itself choses
> under all circumstances
>
> this holds especially if a type has "extended alignment", I think
the standard says
"An object type imposes an alignment requirement on every object of that
type: stricter alignment can be requested using the _Alignas keyword."
"The _Alignof operator yields the alignment requirement of its operand type."
to me this means that all long long objects should have the same
alignment requirement and _Alignof should return this consistently
(unless _Alignas imposes further stricter alignment requirements,
but it never gets weaker)
in my example x and y.x are long long objects but have different
alignment requirements
there is no exception here for types with extended alignments or
"alignment support contexts"
> > i think the standard requires that all (addressable) long long objects
> > should have the same alignment (or stricter) than _Alignof(x)
(it seems that _Alignof is specified for objects with register storage
class too, so even non-addressable long longs should have the same
alignment requirement whatever that means)
> > max_align_t is defined to be the "greatest alignment supported in all
> > contexts", i don't know why it is not just
>
> This only concerns "fundamental alignments". Also, this sentence has
> an implicit "minimum" operator that comes from the "in all contexts"
>
> if the alignment restriction is 8 in some context and 4 in others, the
> result is 4 and not 8.
>
> Perhaps all of this can be made consistent on i386 by having
> _Alignof(max_align_t) to be 4 and declare 8 byte and 16 byte aligned
> types as having "extended alignment"
>
no, long long (or any scalar type) cannot be an "over-aligned" type
"A type having an extended alignment requirement is an over-aligned type."
"Every over-aligned type is, or contains, a structure or union type with
a member to which an extended alignment has been applied."
> I am not sure that I remember correctly, but it seems to me that i386
> allows for 4 byte alignment of all types, only that this results in
> suboptimal code
if long long has an alignment requirement of 4 byte then _Alignof should say so
> > typedef char max_align_t __attribute__((aligned(__BIGGEST_ALIGNMENT__)));
> >
> > which gives 16 byte alignment on i386 gcc, i thought it was supported
> > in all contexts
>
> I think this just not necessary and even counter productive.
depends on what is the semantic meaning of max_align_t
currently it means "guaranteed to be supported in all contexts"
> > if gcc and clang went with the same definition we should follow, but
> > this makes the type less meaningful
>
> By looking at the page that you linked to, my impression is that they
> got it wrong. I think basically
>
> typedef union max_align_t max_align_t;
>
> union max_align_t {
> long double a;
> uintmax_t b;
> void* c;
> max_align_t* d;
> };
>
> should do the trick. All other alignments should be considered as
> extended alignments.
the implementation may support greater alignments
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
