https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106904
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2022-12-07
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Status|UNCONFIRMED |ASSIGNED
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note we diagnose
MEM <unsigned char[8]> [(char * {ref-all})vectp.4_10] = MEM <unsigned char[8]>
[(char * {ref-all})&wp];
where vectp.4_10 == &ps_5(D)->mp.hwnd;
that happens because SLP vectorization produces
vectp.4_10 = &ps_5(D)->wp.hwnd;
vect__1.5_11 = MEM[(int *)vectp.4_10];
vectp.4_12 = vectp.4_10 + 4;
vectp.4_14 = vectp.4_10 + 8;
vect__1.7_15 = MEM[(int *)vectp.4_14];
and we then CSE the memcpy address in the following code to vectp.4_10:
_3 = &ps_5(D)->wp;
__builtin_memcpy (_3, &wp, 8);
the access diagnostics have the issue that they mis-interpret addresses
as more than just pointer arithmetic. Eventually part of this could be
avoided by not introducing any non-invariant ADDR_EXPRs at least but
use POINTER_PLUS_EXPR where possible (like in the above case). Alternatively
we could strip zero-offset components at these points.
