[SYCL] Use C-arrays as storage in sycl::vec

[uditagarwal97]

#14130 caused sycl::vec to use std::array as its underlying storage. However, operations on std::array may emit debug-mode-only functions, on which the device compiler may fail.

This PR replaces std::array with C-style arrays as storage in sycl::vec.

[steffenlarsen]

Although this does not change the size of the vec type, I worry that this is breaking ABI as the vec type crosses the library boundary. To understand why this could be an issue, imagine the following scenario:

Let this change be part of a version Y of the runtime and headers, which came after a version X without it. By ABI/API guarantees we make, programs compiled with X should be runable with Y. However, if such a program passes vec across the library boundary, the runtime library of Y will think the data representation inside the type is a C-style array, while it is actually a std::array.

Granted, it's likely that C-style arrays and std::array are the exact same layout, but making such assumptions is also potentially playing with fire.

[uditagarwal97]

However, if such a program passes vec across the library boundary, the runtime library of Y will think the data representation inside the type is a C-style array, while it is actually a std::array.

Granted, it's likely that C-style arrays and std::array are the exact same layout, but making such assumptions is also potentially playing with fire.

Doesn't the spec guarantees that std::array and C-style arrays will have the same memory layout? C++ spec says (https://en.cppreference.com/w/cpp/container/array):

This (as in std::array) container is an aggregate type with the same semantics as a struct holding a C-style array T[N] as its only non-static data member.

If that's the case and a program passes sycl::vec across boundaries, what's the worse that can happen: we might treat std::array as a C-style array. But, looking at usages of m_Data in sycl::vec, we either access m_Data using [] operator, bitcast to vector_t, or memcpy into m_Data (change made in this PR). If std::array and C-style array share the same memory layout always, I think the above-mentioned operations on m_Data will produce the same results, even if we treat std::array as a C-style array.

[steffenlarsen]

I agree, by the definition given by CPPReference it should (depending on how we should read "semantics") always have the same layout. That said, it gets awfully close to breaking strict-aliasing rules in a way you don't see every day.

[uditagarwal97]

I see. To be on a safer side, I've put these changes under preview-breaking flag.