Re: [PATCH] manual: remove an obsolete requirement on aligned_alloc() usage

[Martin Sebor <msebor@gmail.com>]

On 10/25/21 7:50 PM, Paul Eggert wrote:
> On 10/25/21 08:45, Martin Sebor via Libc-alpha wrote:
>> Based on my code inspection of _mid_memalign in malloc.c I think
>> a change that would more accurately reflect the implementation
>> is one that described that when alignment is not a power of two
>> it's bumped up to next larger power of two, and adjusted
>> the EINVAL condition appropriately.
> 
> Let's not document this implementation detail, as it might tie us down 
> unnecessarily in the future.
> 
>> I believe the DR460 change went into C17.  Here's an early C2x
>> draft from 2018 publicly available on the WG14 site that should
>> still be very close to C17:
>>
>>   http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2310.pdf
> 
> This says that aligned_alloc "shall fail" if the alignment is not a 
> valid alignment supported by the implementation. This means the current 
> glibc alloc_aligned behavior does not conform to C17 and should be fixed 
> so that it does. Something like the attached (untested) patch, say.

I agree.  I took the term "valid alignment" here to mean one
specifically accepted by aligned_alloc() but 6.2.8 is explicit
that

   "Every valid alignment value shall be a nonnegative integral
   power of two."

so although implementations get to decide on the set of valid
extended alignments they're not allowed to be accept ones that
are in conflict with the blanket requirement above.

Martin