On Wed, May 27, 2026 at 8:48 AM Andrew Pinski <
[email protected]> wrote:

> On Tue, May 26, 2026 at 7:48 PM Andrew Pinski
> <[email protected]> wrote:
> >
> > On Mon, May 25, 2026 at 8:45 PM Shivam Gupta <[email protected]>
> wrote:
> > >
> > > At -O1, GCC lowers (~a & 1) == (~b & 1) into two different GIMPLE
> > > forms depending on whether the boolean values come from a function
> > > return boundary or an inline expression:
> > >
> > >   1. Via function return (e.g. inlined is_even), GCC preserves the
> > >      eq/ne form because the function boundary hides the internal
> > >      structure from the gimplifier:
> > >        ((a&1)==0) == ((b&1)==0)   outer node is eq/ne
> > >
> > >   2. Via inline expression (directly written), GCC sees the full
> > >      boolean expression tree at once and folds to bit_xor during
> > >      gimplification:
> > >        ((a&1)!=0) ^ ((b&1)==0)   outer node is bit_xor
> > >
> > > Add a generalized match.pd rule handling:
> > >   1. ((x & c) == 0) OP ((y & c) == 0)
> > >   2. ((x & c) != 0) OP ((y & c) == 0)
> > >   3. ((x & c) == 0) OP (y & c)
> >
> > Can you leave out 3 for now? We will need to deal with it later I think.
> > It is needed due to:
> > ```
> > /* (type)([0,1]@a != 0) -> (type)a
> >    (type)([0,1]@a == 1) -> (type)a
> >    (type)([0,1]@a == 0) -> a ^ 1
> >    (type)([0,1]@a != 1) -> a ^ 1.  */
> > (for eqne (eq ne)
> >  (simplify
> >   (convert (eqne zero_one_valued_p@0 INTEGER_CST@1))
> >   (if ((integer_zerop (@1) || integer_onep (@1)))
> >    (if ((eqne == EQ_EXPR) ^ integer_zerop (@1))
> >     (convert @0)
> >     /* Only do this if the types match as (type)(a == 0) is
> >        canonical form normally, while `a ^ 1` is canonical when
> >        there is no type change. */
> >     (if (types_match (type, TREE_TYPE (@0)))
> >      (bit_xor @0 { build_one_cst (type); } ))))))
> > ```
> > Which was added by
> >
> https://inbox.sourceware.org/gcc-patches/[email protected]/
> > .
> >
> > And then we don't support (t1_7 is zero_one_valued_p here):
> >   _1 = t1_7 == 0;
> >   _2 = (int) _1;
> >
> > Into:
> > _t = (t1_7^1)
> > _2 = (int)_t;
> >
> > As types_match is not true. That was not done .
> >
> > I didn't realize this came into play for the `&1` part until now :).
> >
> > You can keep the testcases for the ones which will match 3 but xfail
> them.
>
> That is because we should handle it differently and separately as that
> only matters with `&1` there.
>
> The pattern there will be something like
> (ne zero_one_valued_p@0 (convert? (eq zero_one_valued_p@1 integer_zerop)))
> simplified to:
> (eq (convert:TREE_TYPE(@1) @0) @1)
>
> With that pattern, this:
>   t1_5 = a_4(D) & 1;
>   t2_7 = b_6(D) & 1; // t2_7 is zero_one_valued_p
>   _1 = t1_5 == 0;
>   _2 = (int) _1; // _2 is zero_one_valued_p
>   _3 = (int) t2_7; // _3 is zero_one_valued_p
>   _8 = _2 != _3;
>   return _8;
>
> Will gets turned into:
> _8 = t2_7 == t1_5
>
>
OK, I can try that in some time as a different patch.

Thanks,
Shivam

Thanks,
> Drea
>
> >
> > Thanks,
> > Andrea
> >
> >
> >
> > >
> > > for OP in { ==, !=, ^ } and pow2 masks. The third form is required
> > > because GCC canonicalize: ((x & c) != 0) into (x & c) before this
> > > rule runs.
> > >
> > > Fold these directly into the canonical xor-mask form:
> > >   ((x ^ y) & c) == 0
> > >   ((x ^ y) & c) != 0
> > >
> > > Bootstrapped and tested on aarch64-linux-gnu.
> > >
> > > Changes since v1:
> > > * v3: Collapse both previous simplify rules into one generalized
> > >       matcher covering all combinations of outer {eq, ne, bit_xor}
> > >       and inner {eq, ne} comparisons.
> > >       Add missing mixed-polarity test cases.
> > >
> > > * v2: Generalize mask from integer_onep to integer_pow2p.
> > >       Combine both patterns into single one.
> > >       Split test into bool-eq-evenness.c, bool-eq-bitxor.c,
> > >       bool-eq-pow2.c.
> > >
> > >         PR tree-optimization/112533
> > > gcc/ChangeLog:
> > >         * match.pd: Canonicalize boolean comparisons of masked pow2
> > >         bits into xor-mask tests.
> > >
> > > gcc/testsuite/ChangeLog:
> > >         * gcc.dg/tree-ssa/bool-eq-bitxor.c: New test.
> > >         * gcc.dg/tree-ssa/bool-eq-evenness.c: New test.
> > >         * gcc.dg/tree-ssa/bool-eq-pow2.c: New test.
> > >
> > > Signed-off-by: Shivam Gupta <[email protected]>
> > > ---
> > >  gcc/match.pd                                  | 38 ++++++++
> > >  .../gcc.dg/tree-ssa/bool-eq-bitxor.c          | 39 ++++++++
> > >  .../gcc.dg/tree-ssa/bool-eq-evenness.c        | 43 +++++++++
> > >  gcc/testsuite/gcc.dg/tree-ssa/bool-eq-pow2.c  | 89 +++++++++++++++++++
> > >  4 files changed, 209 insertions(+)
> > >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/bool-eq-bitxor.c
> > >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/bool-eq-evenness.c
> > >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/bool-eq-pow2.c
> > >
> > > diff --git a/gcc/match.pd b/gcc/match.pd
> > > index ff13a07ea94..d41a2b2f329 100644
> > > --- a/gcc/match.pd
> > > +++ b/gcc/match.pd
> > > @@ -1464,6 +1464,44 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > >    (if (TYPE_UNSIGNED (type))
> > >      (bit_and @0 (bit_not (lshift { build_all_ones_cst (type); }
> @1)))))
> > >
> > > +/* PR112533: Canonicalize boolean comparisons of masked pow2 bits into
> > > +   xor-mask tests.
> > > +   ((x & c) == 0) == ((y & c) == 0) -> (((x ^ y) & c) == 0)
> > > +   ((x & c) == 0) != ((y & c) == 0) -> (((x ^ y) & c) != 0)
> > > +   and equivalent xor forms.  */
> > > +(for ocmp (eq ne bit_xor)
> > > + (for icmp0 (eq eq ne ne)
> > > +      icmp1 (eq ne eq ne)
> > > +  (simplify
> > > +   (ocmp
> > > +    (icmp0 (bit_and @0 integer_pow2p@1) integer_zerop@3)
> > > +    (icmp1 (bit_and @2 @1) @3))
> > > +   (if (types_match (@0, @2))
> > > +    (with {
> > > +      bool innersame
> > > +        = (icmp0 == EQ_EXPR) == (icmp1 == EQ_EXPR);
> > > +    }
> > > +     (if (innersame == (ocmp == EQ_EXPR))
> > > +      (eq (bit_and (bit_xor @0 @2) @1) @3)
> > > +      (ne (bit_and (bit_xor @0 @2) @1) @3)))))
> > > +
> > > +  /* Canonicalized: ((x & c) == 0) OP (y & c)
> > > +     because ((y & c) != 0) may fold into (y & c).  */
> > > +  (simplify
> > > +   (ocmp
> > > +    (convert?@4
> > > +     (icmp0 (bit_and @0 integer_pow2p@1) integer_zerop@3))
> > > +    (convert?@5
> > > +     (bit_and @2 @1)))
> > > +   (if (types_match (@0, @2))
> > > +    (with {
> > > +      bool innersame
> > > +        = (icmp0 == EQ_EXPR) == false;
> > > +    }
> > > +     (if (innersame == (ocmp == EQ_EXPR))
> > > +      (eq (bit_and (bit_xor @0 @2) @1) @3)
> > > +      (ne (bit_and (bit_xor @0 @2) @1) @3)))))))
> > > +
> > >  (for bitop (bit_and bit_ior)
> > >       cmp (eq ne)
> > >   /* PR35691: Transform
> > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-bitxor.c
> b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-bitxor.c
> > > new file mode 100644
> > > index 00000000000..05cee9696cf
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-bitxor.c
> > > @@ -0,0 +1,39 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O1 -fdump-tree-optimized" } */
> > > +
> > > +typedef unsigned int u32;
> > > +
> > > +_Bool
> > > +xor_ne_eq (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) ^ (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +xor_eq_eq (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) ^ (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +xor_eq_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) ^ (t2 != 0);
> > > +}
> > > +
> > > +_Bool
> > > +xor_ne_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) ^ (t2 != 0);
> > > +}
> > > +
> > > +/* Verify all functions canonicalize to xor-mask tests.  */
> > > +/* { dg-final { scan-tree-dump-times "& 1" 4 "optimized" } } */
> > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-evenness.c
> b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-evenness.c
> > > new file mode 100644
> > > index 00000000000..75df70dc6a4
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-evenness.c
> > > @@ -0,0 +1,43 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O1 -fdump-tree-optimized" } */
> > > +
> > > +/* Verify the GIMPLE-level fix for PR112533. */
> > > +
> > > +typedef unsigned int u32;
> > > +
> > > +static _Bool
> > > +is_even (u32 a)
> > > +{
> > > +  return a % 2 == 0;
> > > +}
> > > +
> > > +_Bool
> > > +same_evenness (u32 a, u32 b)
> > > +{
> > > +  return is_even (a) == is_even (b);
> > > +}
> > > +
> > > +_Bool
> > > +diff_evenness (u32 a, u32 b)
> > > +{
> > > +  return is_even (a) != is_even (b);
> > > +}
> > > +
> > > +_Bool
> > > +same_evenness_cmp_zero (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) == (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +diff_evenness_cmp_zero (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) != (t2 == 0);
> > > +}
> > > +
> > > +/* Verify all functions canonicalize to xor forms.  */
> > > +/* { dg-final { scan-tree-dump-times "\\^" 4 "optimized" } } */
> > > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-pow2.c
> b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-pow2.c
> > > new file mode 100644
> > > index 00000000000..109c5f03610
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-pow2.c
> > > @@ -0,0 +1,89 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-O1 -fdump-tree-optimized" } */
> > > +
> > > +typedef unsigned int u32;
> > > +
> > > +_Bool
> > > +eq_eq_outer_eq_mask2 (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 2;
> > > +  u32 t2 = b & 2;
> > > +  return (t1 == 0) == (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +eq_eq_outer_eq_mask4 (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 4;
> > > +  u32 t2 = b & 4;
> > > +  return (t1 == 0) == (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +eq_eq_outer_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) != (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +ne_eq_outer_eq (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) == (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +ne_eq_outer_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) != (t2 == 0);
> > > +}
> > > +
> > > +_Bool
> > > +eq_ne_outer_eq (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) == (t2 != 0);
> > > +}
> > > +
> > > +_Bool
> > > +eq_ne_outer_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 == 0) != (t2 != 0);
> > > +}
> > > +
> > > +_Bool
> > > +ne_ne_outer_eq (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) == (t2 != 0);
> > > +}
> > > +
> > > +_Bool
> > > +ne_ne_outer_ne (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 1;
> > > +  u32 t2 = b & 1;
> > > +  return (t1 != 0) != (t2 != 0);
> > > +}
> > > +
> > > +/* Non-power-of-2: should not be simplified.  */
> > > +_Bool
> > > +eq_eq_outer_eq_nonpow2 (u32 a, u32 b)
> > > +{
> > > +  u32 t1 = a & 3;
> > > +  u32 t2 = b & 3;
> > > +  return (t1 == 0) == (t2 == 0);
> > > +}
> > > +
> > > +/* Verify power-of-2 masks are simplified, non-power-of-2 masks are
> not.  */
> > > +/* { dg-final { scan-tree-dump-times "a_\[0-9\]+\\(D\\) \\^
> b_\[0-9\]+\\(D\\)" 9 "optimized" } } */
> > > +/* { dg-final { scan-tree-dump-times "& 3" 2 "optimized" } } */
> > > --
> > > 2.34.1
> > >
>

Reply via email to