https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78821
--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #13)
> (In reply to Uroš Bizjak from comment #12)
>
> > --cut here--
> > struct s { char a; char b; char c; char d; };
> >
> > void foo (struct s *__restrict a, struct s *__restrict b)
> > {
> > a->a = b->a;
> > a->b = b->b;
> > a->c = ~b->c;
> > a->d = b->d;
> > }
> > --cut here--
> >
> > This testcase can be optimized by inserting xorl mask between load and
> > store, as suggested above.
>
> Also,
>
> a->a = 0;
> a->b = 0;
> a->c = b->c;
> a->d = 0;
>
> could use andl mask, and similar
>
> a->a = 0xff;
> a->b = 0xff;
> a->c = b->c;
> a->d = 0xff;
>
> could use orl mask.
I'm not entirely sure if we can do this last thing, because the original just
reads from b->c, if b->d or b->{a,b} could trap while b->c doesn't (such as for
the 32-bit load not being aligned). At least for the BIT_NOT_EXPR vs. missing
BIT_NOT_EXPR cases with some effort supporting it wouldn't be that difficult,
we'd need to replace the optional BIT_NOT_EXPR with BIT_XOR_EXPR computed
bitmasks based on what stores have them and what don't (in any of the 3 spots
with bit_not_p).
Trying to support something else, like
a->a = b->a | 123;
a->b = b->b & 12;
a->c = b->c ^ 14;
would be harder, but in theory possible.
In any case, none of this needs the bswap infrastructure, while some of the
earlier testcases do need.
