C++ SysV ABI incompatibility wrt. passing non-PODs by value

[kinke]

While checking assembly and clang-IR wrt. DIP 1014, I stumbled on what looks like a huge ABI incompatibility, at least for the SysV ABI, and probably the 32-bit x86 ABIs too (Win64 should be safe though).

As soon as a C++ struct/class has a copy or move ctor, pass-by-value semantics seem to match ExplicitByvalRewrite (pass a pointer to a caller-allocated copy). LDC and DMD instead pass the aggregate on the stack (i.e., like a big POD struct). More information to follow.

[kinke]

To play around with C++: https://godbolt.org/g/XV8Pem (try commenting out the copy ctor and see what happens)

[kinke]

GDC, at least v5.2 as latest version on godbolt, is affected too. See this D snippet.

D additionally doesn't call the C++ copy ctor when passing an lvalue by value, I thought that'd be implemented already. That can be worked around with a D postblit ctor though.

Then there's the different destruction rules; D moves the arg (or arg copy for lvalues) to the callee's memory params stack and lets the callee destruct it, C++ passes a ref to the rvalue (or lvalue copy) to the callee and the calleR destructs it after the call. In frontend terms, a C++ non-POD by-value param is more like an exotic ref, accepting rvalue args and requiring a copy for lvalue args.

Anyway, this explains why clang has no problems calling a move ctor, while LDC would probably have a hard time for DIP1014 with the current ABI (final address of callee param in IR only exposed to callee via implicit byval pointer param; original address of moved-from arg only known by caller).

[kinke]

Non-PODs use the IndirectByvalRewrite now, and the C++ destruction rules are applied now too since v2.096 (https://issues.dlang.org/show_bug.cgi?id=20235).