tmlen tmlen - 5 months ago 29
C++ Question

Alignment and size of C++ primitive types

In C++, it seems that for all integer types (

long long int
, ...) and for the floating point types, it is always
sizeof(T) == alignof(T)

Is this compiler/platform-specific, or guaranteed to be true? Could there be a platform where
s do not need to be aligned on a 32 bit boundary (as long as they don't overlap)?


For fundamental types, alignof(T) == sizeof(T) is true for most modern ABIs, but it is not guaranteed to be true. The C++ standard leaves alignment mostly implementation-defined, subject only to these constraints (see [basic.align] and [basic.fundamental] in C++11):

  • alignof(T) <= sizeof(T). This is not explicit in the standard, but there is never any padding between the elements of an array, so regardless of the alignment of the first element of T t[2], the second element cannot have alignment greater than sizeof(T), and therefore that is the maximum alignment for type T. Note that the alignment of a variable can be greater, for instance if alignas is used on it.

  • alignof(T) <= alignof(std::max_align_t), except possibly in the presence of alignas. Whether alignas can cause "over-alignment" (to a greater granularity than max_align_t) is implementation-defined.

  • char, signed char, and unsigned char all have the same alignment requirement (namely 1, because sizeof(char) == 1 by definition).

  • All unsigned integer types have the same alignment requirement as the corresponding signed type.

  • wchar_t, char16_t, and char32_t have the same alignment requirements as their "underlying" integer types.

There have been historical ABIs where fundamental types were not aligned to their size. One well-known example is double and long double in the original System V ABI for the 80386, which were only aligned to 4-byte granularity despite being 8 and 12 bytes wide respectively. This was because the stack pointer was only guaranteed to be aligned to 4-byte granularity, and it's a pain to align things within an activation record to a greater granularity than the stack pointer.1

Nowadays, this same issue of stack alignment may come up with vector types; for instance, on x86-64 the stack pointer is guaranteed to be aligned to a 16-byte boundary, but the hardware now supports up to 512-bit (64-byte) vectors.

Those are the only counterexamples I personally know about. It would not surprise me, however, to learn that an ABI for a memory-constrained embedded environment specified no alignment for anything (that is, alignof(T) == 1 for all T). It's not nearly as hard for the CPU to handle that as it used to be, especially if there's no memory virtualization involved.

1 Fun fact: the C++ standard does not require the implementation to have a function-call stack. It requires support for recursion, and there's something called "stack unwinding" that happens when exceptions are thrown, but everything is carefully written so that you could implement it using something other than the linear stacks we're all familiar with.