https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82445
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |wrong-code
Target|ARM |arm
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Does -fno-store-merging fix it?
store-merging currently does
unsigned HOST_WIDE_INT align = split_store->align;
tree offset_type = get_alias_type_for_stmts (split_store->orig_stmts);
location_t loc = get_location_for_stmts (split_store->orig_stmts);
tree int_type = build_nonstandard_integer_type (try_size, UNSIGNED);
int_type = build_aligned_type (int_type, align);
tree dest = fold_build2 (MEM_REF, int_type, addr,
build_int_cst (offset_type, try_pos));
and expects the target to cope with this. But IIRC RTL expansion only
ever "honors" this if either it has a movmisalign optab for the mode
or it is SLOW_UNALIGNED_ACCESS (ok, arm seems to be SLOW_UNALIGNED_ACCESS).
So this means we go the bitfield store way (in the end this means store-merging
lacks proper costing for strict-align targets). That may end up using
all sorts of patterns, like insv and friends. Not sure what happens in the
end.
