On Wed, Aug 2, 2023 at 1:25 AM Jakub Jelinek via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Wed, Aug 02, 2023 at 10:04:26AM +0200, Richard Biener via Gcc-patches
> wrote:
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -1157,8 +1157,9 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > >
> > > /* Simplify ~X & X as zero. */
> > > (simplify
> > > - (bit_and:c (convert? @0) (convert? (bit_not @0)))
> > > - { build_zero_cst (type); })
> > > + (bit_and (convert? @0) (convert? @1))
> > > + (if (bitwise_inverted_equal_p (@0, @1))
> > > + { build_zero_cst (type); }))
>
> I wonder if the above isn't incorrect.
> Without the possibility of widening converts it would be ok,
> but for widening conversions it is significant not just that
> the bits of @0 and @1 are inverted, but also that they are either
> both signed or both unsigned and so the MS bit (which is guaranteed
> to be different) extends to 0s in one case and to all 1s in the other
> one, so that even the upper bits are inverted.
> But that isn't the case here. Something like (untested):
> long long
> foo (unsigned int x)
> {
> int y = x;
> y = ~y;
> return ((long long) x) & y;
> }
> Actually maybe for this pattern it happens to be ok, because while
> the upper bits in this case might not be inverted between the extended
> operands (if x has msb set), it will be 0 & 0 in the upper bits.
>
> > >
> > > /* PR71636: Transform x & ((1U << b) - 1) -> x & ~(~0U << b); */
> > > (simplify
> > > @@ -1395,8 +1396,9 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > > /* ~x ^ x -> -1 */
> > > (for op (bit_ior bit_xor)
> > > (simplify
> > > - (op:c (convert? @0) (convert? (bit_not @0)))
> > > - (convert { build_all_ones_cst (TREE_TYPE (@0)); })))
> > > + (op (convert? @0) (convert? @1))
> > > + (if (bitwise_inverted_equal_p (@0, @1))
> > > + (convert { build_all_ones_cst (TREE_TYPE (@0)); }))))
>
> But not here.
> long long
> bar (unsigned int x)
> {
> int y = x;
> y = ~y;
> return ((long long) x) ^ y;
> }
>
> long long
> baz (unsigned int x)
> {
> int y = x;
> y = ~y;
> return y ^ ((long long) x);
> }
> You pick TREE_TYPE (@0), but that is a random signedness if the two
> operands have different signedness.
Oh you are correct, I am testing a patch which adds the test to make
sure the types of @0 and @1 match which brings us back to basically
was done beforehand and still provides the benefit of using
bitwise_inverted_equal_p for the comparisons.
Thanks,
Andrew
>
> Jakub
>
