https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101062
--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Richard Biener from comment #7) > Now, it looks to me this is rather an issue that the access is larger than > the object and thus a general bug - at least I don't see how it should only > manifest with bitfields in unions? > > Note we do > > if (TREE_CODE (to) == COMPONENT_REF > && DECL_BIT_FIELD_TYPE (TREE_OPERAND (to, 1))) > get_bit_range (&bitregion_start, &bitregion_end, to, &bitpos, > &offset); > /* The C++ memory model naturally applies to byte-aligned fields. > However, if we do not have a DECL_BIT_FIELD_TYPE but BITPOS or > BITSIZE are not byte-aligned, there is no need to limit the range > we can access. This can occur with packed structures in Ada. */ > else if (maybe_gt (bitsize, 0) > && multiple_p (bitsize, BITS_PER_UNIT) > && multiple_p (bitpos, BITS_PER_UNIT)) > { > bitregion_start = bitpos; > bitregion_end = bitpos + bitsize - 1; > } > > but if we assume that for DECL_BIT_FIELD_TYPE there's a representative > then we miss the else if, so - maybe get_bit_range should return whether > it handled things or the else if part should be done unconditionally > in case bitregion_start/end is not {0,0}? This wouldn't help us, bitsize is > 0, but not a multiple of BITS_PER_UNIT in this case. Furthermore, even if we add there bitregion_start/end for the base variable if any as further fallthrough, I think most C/C++ programmers will expect that with union U { int a; int b : 5; } u[64]; u[4].b = 1; can be done safely in one thread and u[5].a = 2; in another one. My patch fixes that (or another possibility would be to compute the representative even in UNION_TYPE (no idea about QUAL_UNION_TYPE) - could be as simple as removing the early out and instead of doing prev = field; in the loop do if (TREE_CODE (t) != RECORD_TYPE) { finish_bitfield_representative (repr, field); repr = NULL_TREE; } else prev = field; and in finish_bitfield_representative override nextf to NULL_TREE). Improving expand_assignment can be done too, sure, but independently to this.