[Bug translation/96159] atomic creates incorrect code for possible isaligned struct

[uecker at eecs dot berkeley.edu]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96159

--- Comment #1 from Martin Uecker <uecker at eecs dot berkeley.edu> ---



On x86-64 the following struct has alignment 4 but gcc
creates a single mov instruction which according to my
understanding may fail to be atomic when it crosses a
cache line boundary. 

Documentation seems to imply that __atomic_load is
supposed to work for all types. It also compiles
without warning.


typedef struct { float re; float im; } foo_t;

int foo_align(void)
{
        foo_t x;
        return _Alignof(x);
}

foo_t foo_load(foo_t* x)
{
        foo_t r;
        __atomic_load(x, &r, __ATOMIC_SEQ_CST);
        return r;
}


assembler:

foo_align:
        mov     eax, 4
        ret
foo_load:
        movq    xmm0, QWORD PTR [rdi]
        ret