Clarification of rules for flattening structs containing arrays of empty records

[asb]

The hard FP calling convention currently states "Fields containing empty structs or unions are ignored while flattening, even in C++, unless they have nontrivial copy constructors or destructors." It appears that g++ doesn't ignore an array of empty records in a struct. If this is indeed the desired behaviour, it would be good to document it.

$ cat t.c
struct empty { struct { struct { } e; }; };
struct s { struct empty e[1]; float f; };

float test_s(struct s a) {
  return a.f+1.0;
}

In the above example, the float is passed in a0 with g++ but fa0 with gcc.

[asb]

However zero-sized arrays of empty records are ignored by both g++ and gcc for flattening (thanks @luismarques for the suggestion to check that case!). I'm attempting to align Clang/LLVM with this behaviour in https://reviews.llvm.org/D142327

[nick-knight]

My interpretation is that s::e is, in the language of our spec, a "field containing an empty struct". So I would expect g++ to pass the float in fa0. Whether or not my interpretation is correct, I agree a clarification would be useful.

I'm not sure it is helpful to this discussion, but I believe that empty structs are a GNU C extension, while being (standard) C++; additionally, zero-length arrays are a GNU C/C++ extension. It's not clear to me what our psABI is expected to say about language extensions.

[kito-cheng]

ack, I plan to dig more deeper on lang spec in next few days, and seems sizeof (struct s) has different value in C++ or C mode for both GCC and clang, and study how other ABI handle this.