[AAPCS64] Clarify how __bf16 affects HFAs

[ostannard]

Currently, it's not completely clear whether a struct containing a mixture of different 16-bit floating point types is a Homogeneous Aggregate or not:

struct foo {
__fp16 a;
__bf16 b;
};

I think that the best interpretation of the current text is that this is an HFA, because there is only one entry in the "Fundamental Data Types" table, which is the only thing considered in the definition of a Homogeneous Aggregate.

However, the table in the "Mapping of C & C++ built-in data types" section used the formats in a column which otherwise contains Fundamental Data Types, which could be read to imply that the above struct is not an HFA. I don't think this is correct, so I've moved the formats to the notes column, making it clear that the two source types map to the same Fundamental Data Type.

Current compiler behaviour:

• Clang considers this struct to be an HFA, matches the behaviour described in this ABI patch.
• GCC does not consider it to be an HFA, but it also does not think that a struct only containing __bf16 is an HFA either. That's unambiguously incorrect both before and after this patch, so it's just a special case of a broader GCC bug.

[smithp35]

LGTM.

I think this is the best option given the existing text. While it does seem strange to consider a mix of __fp16 and __bf16 as homogenous, in terms of making good use of the registers in the calling convention it makes sense.

Would like to see if there are any objections on the GNU side?

[smithp35]

Agreed to merge this PR with Arm's GNU team. While it is not ideal to have __bf16 and __fp16 mapping to the same underlying natural type, changing it risks breaking too much existing code.

[rsandifo-arm]

For the record, I've filed https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120007 to track this on the GCC side.