https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119733
Krister Walfridsson <kristerw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Resolution|INVALID |FIXED
--- Comment #4 from Krister Walfridsson <kristerw at gcc dot gnu.org> ---
The struct is 64-bit aligned in the C semantics, but the GIMPLE semantics is
more permissive (see the recent discussions on the mailing list [1][2]), and
several passes create misaligned pointers. For example, ivopts decides to
calculate &u.buf[off] as:
union
{
char buf[31];
long long int align_int;
long double align_fp;
} * _13;
_14 = (sizetype) off_4(D);
_13 = &u + _14;
when compiling the function
static union {
char buf[31];
long long align_int;
long double align_fp;
} u;
void check (int off)
{
char *q;
int i;
q = u.buf;
for (i = 0; i < off; i++, q++)
if (*q != 'a')
__builtin_abort ();
}
which makes _13 misaligned when off is not a multiple of 8.
So I claim the store-merging pass is not allowed to use the alignment of the
pointer to do the optimization. But it may of course be the case that it's
store-merging that is correct, and ivopts, etc. that are wrong. In that case, I
need some additional information about what the correct GIMPLE semantics is, so
I can report the real issues...
References:
1. https://gcc.gnu.org/pipermail/gcc/2025-March/245763.html
2. https://gcc.gnu.org/pipermail/gcc/2025-April/245870.html
