Re: [boost] Review request for Boost.Align

Subject: Re: [boost] Review request for Boost.Align
From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2014-02-20 05:07:04

On Thu, Feb 20, 2014 at 1:55 PM, Andrey Semashev
<andrey.semashev_at_[hidden]> wrote:
> On Thu, Feb 20, 2014 at 1:21 PM, Glen Fernandes
> <glen.fernandes_at_[hidden]> wrote:
>>
>>> Can this class be extended so that I can specify the alignment
>>> explicitly in the template parameters?
>>
>> Definitely. Instead of specifying the alignment value as a template
>> parameter, maybe a preferable design is:
>>
>>     template<class T, template<class> class A = boost::alignment_of>
>>     class aligned_allocator;
>>
>> ...where 'A' could supply different alignment values for different
>> rebound types.
>
> I think it complicates things too much. For example, if I just want to
> align memory at cache line boundary (64 bytes for Intel CPUs), I would
> have to create a metafunction just to provide this single constant to
> the allocator.
>
> I can understand why you would want to use a metafunction here - you
> want rebound allocator to use the proper alignment. But realistically,
> no one ever wants to align memory weaker than malloc/new, and it is
> already guaranteed to align properly for any builtin types. So you
> would specify a stronger alignment than malloc/new provides, which
> means it should already be strong enough for any T. That's why I think
> a simple constant would be enough.
>
>   template< typename T, unsigned int Alignment = 0 >
>   class aligned_allocator
>   {
>     static constexpr unsigned int alignment = Alignment > 0 ?
> Alignment : alignment_of< T >::value;
>     static_assert(alignment >= alignment_of< T >::value, "Specified
> alignment is too weak");
>   };

Or even:

   template< typename T, unsigned int Alignment = 0 >
   class aligned_allocator
   {
     static constexpr unsigned int alignment =
       Alignment > alignment_of< T >::value ? Alignment :
alignment_of< T >::value;
   };

This would make it "just work".

> But if you decide to go with the metafunction, then at least provide a
> boilerplate so that it is easy to specify a constant.