I've just hit into a bug in gcc 4.x, it sucks to have a bug in your toolchain when that's probably the last thing you would ever suspect.

The code goes like this:

#include

struct d {
unsigned int lo, hi;
};

static inline unsigned long long r(struct d x)
{
return *(unsigned long long *)&x;
}

void
g(unsigned long long x)
{
printf("%llx\n", x);
}

int
main(void)
{
struct d c;
c.lo = 0xdeadbeef;
c.hi = 0x2;

g(r(c));
return(0);
}

Compile with gcc -O2 -m32 ...

You'd expect to see 0x2deadbeef on a 32-bit little-endian machine, right? No ... the top 32 bits get garbled.

It all works without -fstrict-aliasing (implicitly set with -O2).

On -Os, you get a big fat zero.

On a 64-bit machine ... you get a zero too.

The key, apparently is the inline. If it's not inlined, then it all works .... sigh.