On Mon, 25 May 2026 at 12:41 AM, Andrew Pinski <
[email protected]> wrote:

> On Sat, May 16, 2026 at 11:50 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 two match.pd rules to handle both forms:
> >
> >   1. (cmp (eq (x&1) 0) (eq (y&1) 0)) -> (cmp (x&1) (y&1))
> >   2. (bit_xor (eq  (x&1) 0) (eq (y&1) 0)) -> (ne (x&1) (y&1))
> >      (bit_xor (ne  (x&1) 0) (eq (y&1) 0)) -> (eq (x&1) (y&1))
> >
> > Subsequent passes further simplify these to (x^y)&1 == 0.
> >
> > Bootstrapped and tested on aarch64-linux-gnu with
> > RUNTESTFLAGS="tree-ssa.exp".
> >
> > PR tree-optimization/112533
> >
> > gcc/ChangeLog:
> >         * match.pd: Simplify ((x&1)==0)==((y&1)==0) to (x&1)==(y&1)
> >         and ((x&1) cmp 0)^((y&1)==0) to (x&1) icmp (y&1).
> >
> > gcc/testsuite/ChangeLog:
> >         * gcc.dg/tree-ssa/bool-eq-simplify-O1.c: New test.
> >
> > Signed-off-by: Shivam Gupta <[email protected]>
> > ---
> >  gcc/match.pd                                  | 27 +++++++++++
> >  .../gcc.dg/tree-ssa/bool-eq-simplify-O1.c     | 47 +++++++++++++++++++
> >  2 files changed, 74 insertions(+)
> >  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/bool-eq-simplify-O1.c
> >
> > diff --git a/gcc/match.pd b/gcc/match.pd
> > index ff13a07ea94..7afa3b0d546 100644
> > --- a/gcc/match.pd
> > +++ b/gcc/match.pd
> > @@ -1464,6 +1464,33 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> >    (if (TYPE_UNSIGNED (type))
> >      (bit_and @0 (bit_not (lshift { build_all_ones_cst (type); } @1)))))
> >
> > +/* ((x & 1) == 0) == ((y & 1) == 0) -> (x & 1) == (y & 1)  */
> > +/* ((x & 1) == 0) != ((y & 1) == 0) -> (x & 1) != (y & 1)  */
> > +(for cmp (eq ne)
> > + (simplify
> > +  (cmp
> > +   (eq (bit_and:c @0 INTEGER_CST@1) integer_zerop)
> > +   (eq (bit_and:c @2 @1) integer_zerop))
> > +  (if (integer_onep (@1)
>
> This works for all values of powers of 2 rather than just 1.
> So INTEGER_CST should be integer_pow2p. You can use integer_pow2p as a
> predicate on the capture too (integer_onep could be done the same
> way).
> Also with both bit_and you don't need `:c` since constants are always
> placed on the 2nd operand due to canonicalization.
>

Thanks for the review. Addressed in v2 of the patch.

Thanks,
Shivam

>
> > +       && types_match (@0, @2))
> > +   (cmp
> > +    (bit_and @0 @1)
> > +    (bit_and @2 @1)))))
> > +
> > +/* ((x & 1) != 0) ^ ((y & 1) == 0) ->  (x & 1) == (y & 1)  */
> > +/* ((x & 1) == 0) ^ ((y & 1) == 0) ->  (x & 1) != (y & 1)  */
> > +(for cmp  (eq ne)
> > +     icmp (ne eq)
> > + (simplify
> > +  (bit_xor:c
> > +   (cmp (bit_and:c @0 INTEGER_CST@1) integer_zerop)
> > +   (eq   (bit_and:c @2 @1) integer_zerop))
> > +  (if (integer_onep (@1)
> > +       && types_match (@0, @2))
>
> Same story here.
> You can combine these two by having a resulting operator.
>
> Thanks,
> Andrea
>
>
> > +   (icmp
> > +    (bit_and @0 @1)
> > +    (bit_and @2 @1)))))
> > +
> >  (for bitop (bit_and bit_ior)
> >       cmp (eq ne)
> >   /* PR35691: Transform
> > diff --git a/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-simplify-O1.c
> b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-simplify-O1.c
> > new file mode 100644
> > index 00000000000..3d4f5328c6e
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/tree-ssa/bool-eq-simplify-O1.c
> > @@ -0,0 +1,47 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O1 -fdump-tree-optimized" } */
> > +
> > +/* Verify the GIMPLE-level fix for PR112533:
> > +   (~A & 1) == (~B & 1) is lowered by GCC to one of two GIMPLE forms
> > +   which should both be simplified to (A ^ B) & 1 == 0.  */
> > +
> > +typedef unsigned int u32;
> > +
> > +/* Source-level form from the bug report.  */
> > +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);
> > +}
> > +
> > +/* GIMPLE-level form.  */
> > +_Bool
> > +same_evenness_o1 (u32 a, u32 b)
> > +{
> > +  u32 t1 = a & 1;
> > +  u32 t2 = b & 1;
> > +  return (t1 == 0) == (t2 == 0);
> > +}
> > +
> > +_Bool
> > +diff_evenness_o1 (u32 a, u32 b)
> > +{
> > +  u32 t1 = a & 1;
> > +  u32 t2 = b & 1;
> > +  return (t1 == 0) != (t2 == 0);
> > +}
> > +
> > +/* Verify all four functions are optimized to the canonical XOR form.
> */
> > +/* { dg-final { scan-tree-dump-times "\\\^" 4 "optimized" } } */
> > --
> > 2.34.1
> >
>

Reply via email to