Re: [boost] Review request for Boost.Align

On Thu, Feb 20, 2014 at 12:52 PM, Andrey Semashev wrote:

On Thu, Feb 20, 2014 at 12:27 PM, Glen Fernandes wrote:

...

Source: http://github.com/glenfe/align

Documentation: http://glenfe.com/align/doc/

Motivation: Tiny library to be used for all common alignment related chores found in a few Boost libraries. (e.g. we have some in Boost.Smart_Ptr).

Provides: 1. Function boost::align, a pointer alignment function, for implementations that do not yet provide std::align. (Uses std::align if available)

2. Functions boost::aligned_alloc and boost::aligned_free, for aligned allocation and deallocation. (Use platform specific functions, such _aligned_malloc or posix_memalign, or otherwise uses std::malloc with the above boost::align)

3. Class template boost::aligned_allocator as a replacement for std::allocator that respects over-aligned types. (Up to the extent the implementation supports over-alignment)

Many thanks to Peter Dimov for his feedback on the design and implementation of this library.

Notes: Small enough that it might satisfy the requirements for a fast-track review.

First, nice to see this submission, I've been missing and reimplementing this functionality from time to time.

Second, you may see this:

https://github.com/boostorg/log/blob/master/include/boost/log/detail/malloc_...

to improve portability of aligned_alloc. As you can see, posix_memalign is not always available, even if advertised by POSIX macros. Note also that posix_memalign imposes specific requirements on the alignment that are not described in your docs. And _aligned_malloc/_aligned_free are available on compilers other than MSVC.

One other note. I can see that aligned_allocator does not allow to specify the alignment. I assume it uses the type T alignment? Can this class be extended so that I can specify the alignment explicitly in the template parameters? This would be very useful for cases when the elements of type T must have stronger alignment, for example, when element processing needs to be vectorized.