On Wed, Feb 12, 2014 at 10:30:01AM +0100, Richard Biener wrote:
> Bah. I am testing the following.
But then there is no guarantee that ptr is aligned after the call.
char buf[32] __attribute__((aligned (32)));
int
foo (void)
{
void *ptr = buf + 1;
posix_memalign (&ptr, 32, -1);
/* Assume posix_memalign has failed. */
return ((__UINTPTR_TYPE__)ptr) & 31;
}
This should return 1, but supposedly doesn't with the optimization.
So, either we need to revert the lowering, or perhaps do it only if
the original posix_memalign has a lhs and do it like:
void *tmp;
int res = posix_memalign (&tmp, align, size);
if (!res)
ptr = __builtin_assume_aligned (tmp, align);
or so (no need to initialize tmp and copy it back for the failure case,
but perhaps it would result in better code).
Jakub
