[PATCH] Define alignof_slot using _Alignof when using C11 or newer

[PATCH] Define alignof_slot using _Alignof when using C11 or newer

[Khem Raj]

WG14 N2350 made very clear that it is an UB having type definitions
within "offsetof" [1]. This patch enhances the implementation of macro
alignof_slot to use builtin "_Alignof" to avoid undefined behavior on
when using std=c11 or newer

clang 16+ has started to flag this [2]

Fixes build when using -std >= gnu11 and using clang16+

[1] https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2350.htm
[2] https://reviews.llvm.org/D133574

Signed-off-by: Khem Raj <raj.khem@gmail.com>
---
 ChangeLog     | 5 +++++
 lib/alignof.h | 2 ++
 2 files changed, 7 insertions(+)

diff --git a/ChangeLog b/ChangeLog
index fb467a3c14..71d7f7b7f6 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,8 @@
+2023-01-15  Khem Raj  <raj.khem@gmail.com>
+
+	* lib/alignof.h (alignof_slot): Use _Alignof when using C11 or newer
+	standard.
+
 2023-01-14  Bruno Haible  <bruno@clisp.org>
 
 	error, verror tests: Fix link error when the package uses libintl.
diff --git a/lib/alignof.h b/lib/alignof.h
index 505ad97aa4..66a9496518 100644
--- a/lib/alignof.h
+++ b/lib/alignof.h
@@ -28,6 +28,8 @@
 #if defined __cplusplus
   template <class type> struct alignof_helper { char __slot1; type __slot2; };
 # define alignof_slot(type) offsetof (alignof_helper<type>, __slot2)
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__cplusplus)
+# define alignof_slot(type) _Alignof(type)
 #else
 # define alignof_slot(type) offsetof (struct { char __slot1; type __slot2; }, __slot2)
 #endif
-- 
2.39.0

Re: [PATCH] Define alignof_slot using _Alignof when using C11 or newer

[Bruno Haible]

Khem Raj wrote:
> [2] https://reviews.llvm.org/D133574

Thanks for the heads-up.

> WG14 N2350 made very clear that it is an UB having type definitions
> within "offsetof" [1].
> [1] https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2350.htm

This document is not normative; it is merely a discussion document.

What's (quasi) normative is

  * n3047.pdf = C23, which says in 7.21.(3)
    "If the specified type defines a new type or if the specified member is
     a bit-field, the behavior is undefined."

  * n2176.pdf = C18, which in 7.19.(3) merely says
    "If the specified member is a bit-field, the behavior is undefined."

So, only C23 and higher require to avoid offsetof with a new type.

Since gnulib/lib/stdalign.h mentions an _Alignof bug in older versions of
GCC and clang and some of these older versions (e.g. GCC 4.8.5) support C11,
it is better to enable the workaround only for C23 and higher, not for
C11 and higher.

> +#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__cplusplus)

The '!defined(__cplusplus)' part is redundant here.

Bruno

Re: [PATCH] Define alignof_slot using _Alignof when using C11 or newer

[Khem Raj]

On Sat, Jan 14, 2023 at 4:17 PM Bruno Haible <bruno@clisp.org> wrote:
>
> Khem Raj wrote:
> > [2] https://reviews.llvm.org/D133574
>
> Thanks for the heads-up.
>
> > WG14 N2350 made very clear that it is an UB having type definitions
> > within "offsetof" [1].
> > [1] https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2350.htm
>
> This document is not normative; it is merely a discussion document.
>
> What's (quasi) normative is
>
>   * n3047.pdf = C23, which says in 7.21.(3)
>     "If the specified type defines a new type or if the specified member is
>      a bit-field, the behavior is undefined."
>
>   * n2176.pdf = C18, which in 7.19.(3) merely says
>     "If the specified member is a bit-field, the behavior is undefined."
>
> So, only C23 and higher require to avoid offsetof with a new type.
>
> Since gnulib/lib/stdalign.h mentions an _Alignof bug in older versions of
> GCC and clang and some of these older versions (e.g. GCC 4.8.5) support C11,
> it is better to enable the workaround only for C23 and higher, not for
> C11 and higher.
>

I think if its enabled only for c2x and newer, clang will still errors out  for
__STDC_VERSION__= 201710L
I have sent out a patch which replicates the logic from
gnulib/lib/stdalign.h.in which should
address this as well.

> > +#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__cplusplus)
>
> The '!defined(__cplusplus)' part is redundant here.
>

right. Dropped in v2.

> Bruno
>
>
>

Re: [PATCH] Define alignof_slot using _Alignof when using C11 or newer

[Paul Eggert]

On 2023-01-14 16:17, Bruno Haible wrote:
> What's (quasi) normative is
> 
>    * n3047.pdf = C23, which says in 7.21.(3)
>      "If the specified type defines a new type or if the specified member is
>       a bit-field, the behavior is undefined."

I'm afraid this wording in n3047 is unclear. Ordinarily one would look 
at n3047's section "Type definitions" to interpret the phrase "defines a 
new type", but as that section merely talks about typedefs it doesn't 
seem to be relevant here. I can't find any part of the standard that 
talks about "defining a new type".

All the standardizers wanted here is for the TYPE to not contain a 
top-level comma in 'offsetof (TYPE, MEMBER)' because that would screw up 
macro argument identification.

I hope Clang 16 is not reporting an error for 'offsetof (struct {char a; 
int b;}, b)'; if so, it's gone off the deep end. A nameless struct does 
not "define a new type" for some reasonable definitions of the term 
"define a new type", and since the standard doesn't define that term the 
Clang developers should be cautious about diagnostics in this messy area 
when a lot of code out there uses offsetof in precisely this way.