On Tue, Feb 21, 2023 at 05:23:52PM +0100, Laszlo Ersek wrote:
On 2/20/23 21:38, Eric Blake wrote:
> On Mon, Feb 20, 2023 at 06:03:13PM +0100, Laszlo Ersek wrote:
>> On 2/15/23 21:27, Eric Blake wrote:
>>> On Wed, Feb 15, 2023 at 03:11:34PM +0100, Laszlo Ersek wrote:
>>>> The "name##_iter" function is used 11 times in libnbd; in all
those cases,
>>>> "name" is "string_vector", and the "f"
callback is "free":
>>>>
>>>> string_vector_iter (..., (void *) free);
>>>>
>>>> Casting "free" to (void*) is ugly. (Well-defined by POSIX, but
still.)
>>>
>>> Tangentially related: casting function pointers in general may get
>>> harder as more compilers move towards C23 and its newer rules (see for
>>> example
>>>
https://lists.gnu.org/archive/html/bug-gnulib/2023-02/msg00055.html or
>>> this gcc 13 bug
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108694
>>> which highlights some of the warnings that newer compilers will start
>>> warning us about). While this patch focuses on avoiding casts between
>>> fn(*)(type*) and void*, I don't know off-hand if C23 will permit
>>> fn(*)(void*) as a compatible function pointer with fn(*)(type).
>>
>> My understanding is that, per C99 at least, ret_type(*)(void*) is
>> compatible with ret_type(*)(type*) precisely if void* is compatible with
>> type* (6.7.5.3p15).
>>
>> Whether void* is compatible with type* depends on... ugh, that's hard to
>> deduce from the standard. 6.7.5.1p2 says, "For two pointer types to be
>> compatible, both shall be identically qualified and both shall
>> be pointers to compatible types". I don't think "void" (as a
type in
>> itself) is compatible with any type!
>>
>> Now, there is one particular statement on void* -- 6.2.5p27 says, "A
>> pointer to void shall have the same representation and alignment
>> requirements as a pointer to a character type."
>>
>> (I think the statements about *converting* void* to type* and vice versa
>> do not apply here; AFAICT "compatibility" is about reinterpreting the
>> bit patterns, not converting values.)
>>
>> In Annex I (Common warnings, "informative"), the following is listed:
>> "An implicit narrowing conversion is encountered, such as the assignment
>> of [...] a pointer to void to a pointer to any type other than a
>> character type".
>>
>> All in all I don't think ret_type(*)(type*) is compatible with
>> ret_type(*)(void*) in the general case, and that's why in this patch I
>> didn't want to go more general than I absolutely needed to.
>
> Thanks for at least trying to find something definitive in the
> standard. Now you know why I skipped researching this particular
> issue - it's not straightforward to figure out when function pointers
> with differing parameter types are compatible.
Let's be honest: it's staggeringly difficult to collect whatever
"compatible type" means, in the standard.
>
>>>
>>> Thinking higher-level now, your new macro is something where we have
>>> to do a two-step declaration of macro types where we want the new
>>> function. Would it make more sense to change the signature of the
>>> DEFINE_VECTOR_TYPE() macro to take a third argument containing the
>>> function name to call on cleanup paths, with the ability to easily
>>> write/reuse a no-op function for vectors that don't need to call
>>> free(), where we can then unconditionally declare name##_empty() that
>>> will work with all vector types? That is, should we consider instead
>>> doing something like:
>>>
>>> DEFINE_VECTOR_TYPE (string_vector, char *, free);
>>>
>>> DEFINE_VECTOR_TYPE (int_vector, int, noop);
>>
>> My counter-arguments:
>>
>> - this requires updates to all existent DEFINE_VECTOR_TYPE macro
>> invocations,
>>
>> - with "noop" passed to _reset, _reset and _empty become effectively
the
>> same, so we get (at least partially) duplicate APIs,
>>
>> - this would be a step towards combinatorial explosion
>>
>> - if "noop" does nothing, then why call it on each element of the
vector
>> anyway? It's not only the function call that becomes superfluous in the
>> loop bodym with the function being "noop", but the loop *itself*
becomes
>> superfluous. So then we might want to compare the function pointer
>> against "noop" outside of the loop... and that way we get a bunch of
>> complications :)
>>
>> I chose this approach because it is purely additive and precisely as
>> generic/specific as it needs to be. We already have 11 use cases, so I
>> don't think it's *too* specific.
>
> We may still want some division of:
>
> DEFINE_VECTOR_TYPE (int_vector, int);
> DEFINE_POINTER_VECTOR_TYPE (string_vector, char *, free);
>
> where under the hood, DEFINE_POINTER_VECTOR_TYPE(type, base, fn)
> invokes both DEFINE_VECTOR_TYPE(type, base) and
> DEFINE_VECTOR_EMPTY(type, fn), or whatever we name the second
> function.
I quite like Eric's suggestion, but it's probably too much complexity
for this patch series.
This is doable, but I hope it's not expected that
DEFINE_POINTER_VECTOR_TYPE() *enforce* that the element type be a pointer :)
You might ignore this for a first draft, but it is apparently possible
to statically detect this (at least, if using GCC/clang):
https://stackoverflow.com/questions/19255148/check-if-a-macro-argument-is...
> ADD_VECTOR_EMPTY_METHOD() instead of DEFINE_VECTOR_EMPTY() works
for
> me.
OK, ADD_VECTOR_EMPTY_METHOD() can work with the above.
This sounds fine for now, and since these are implementation details
we can always revisit them in future.
Rich.
--
Richard Jones, Virtualization Group, Red Hat
http://people.redhat.com/~rjones
Read my programming and virtualization blog:
http://rwmj.wordpress.com
virt-builder quickly builds VMs from scratch
http://libguestfs.org/virt-builder.1.html