https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66804
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
CCP computes this precisely:
Visiting PHI node: # PT = { D.1842 } (nonlocal, escaped)
p_3 = PHI <&MEM[(void *)&xxx + 2444B](2), p_4(3)>
Argument #0 (2 -> 4 executable)
&MEM[(void *)&xxx + 2444B] Value: CONSTANT &MEM[(void *)&xxx +
2444B]
Argument #1 (3 -> 4 executable)
p_4 Value: CONSTANT 0x0 (0x0ffffffffffffffff)
PHI node value: CONSTANT 0x0 (0x0ffffffffffffffff)
but we don't record the alignment (because it looks useless to know - it's
byte alignment). So the alignment on the access (invalid due to your use
of a bogus aligned type) prevails.
