Incorrect stack alignment of 16byte aligned aggregates on linux

[MrSmith33]

When struct is passed on the stack and alignment of struct is 16 bytes compiler does not align the struct properly. Here is a comparison between ldc, dmd and gcc. Both ldc and dmd miss the alignment.

https://godbolt.org/z/xKc8WK

D code:

align(16)
struct s { long a; }
extern(C) void foo(int, int, int, int, int, int, s, s);
void bar() {
    // push    10
    // push     7
    // call    foo@PLT
    foo(1,2,3,4,5,6, s(7),s(10));
}

C code:

struct __attribute__((aligned(16))) s { long a; };
void foo(int, int, int, int, int, int, s, s);
void bar() {
    // push     0
    // push    10
    // push     0
    // push     7
    foo(1,2,3,4,5,6, s{7},s{10});
}

However alignment starts to work if you replace 16 with 32.

https://godbolt.org/z/6r7GbW

dmd issue: https://issues.dlang.org/show_bug.cgi?id=21230

[kinke]

This is... problematic. The problem is https://github.com/dlang/dmd/blob/f3a588f41c63a1655fcb733c06ef08e152da0125/src/dmd/argtypes_sysv_x64.d#L399-L404, which I've analyzed back then with such a struct being passed as the first arg (i.e., in registers), where it only occupies one register, despite its padded size of 16 bytes. If it is spilled to the stack, it's pushed as full 16-bytes struct with appropriate alignment. So the toArgTypes machinery doesn't work well in this case, as it doesn't account for such special cases, where the passed payload size depends on the parameter position.