On Fri, Jun 21, 2024 at 9:12 AM Eikansh Gupta <quic_eikag...@quicinc.com> wrote:
>
> We can optimize (vec_cond eq/ne vec_cond) when vec_cond is a
> result of (vec CMP vec). The optimization is because of the
> observation that in vec_cond, (-1 != 0) is true. So, we can
> generate vec_cond of xor of vec resulting in a single
> VEC_COND_EXPR instead of 3.
>
> The patch adds match pattern for vec a, b:
> (a ? -1 : 0) != (b ? -1 : 0) --> (a^b) ? -1 : 0
> (a ? -1 : 0) == (b ? -1 : 0) --> ~(a^b) ? -1 : 0
Why should this only work for uniform -1 and 0 vectors?
It seems to me it's valid for arbitrary values, thus
(a ? x : y) != (b ? x : y) -> a^b ? x : y
(a ? x : y) == (b ? x : y) -> a^b ? y : x
no?
> PR tree-optimization/111150
>
> gcc/ChangeLog:
>
> * match.pd: Optimization for above mentioned pattern.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/tree-ssa/pr111150.c: New test.
>
> Signed-off-by: Eikansh Gupta <quic_eikag...@quicinc.com>
> ---
> gcc/match.pd | 18 ++++++++++++++++++
> gcc/testsuite/gcc.dg/tree-ssa/pr111150.c | 19 +++++++++++++++++++
> 2 files changed, 37 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr111150.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 3d0689c9312..5cb78bd7ff9 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -5522,6 +5522,24 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (vec_cond (bit_and (bit_not @0) @1) @2 @3)))
> #endif
>
> +/* (a ? -1 : 0) != (b ? -1 : 0) --> (a^b) ? -1 : 0 */
> +/* (a ? -1 : 0) == (b ? -1 : 0) --> ~(a^b) ? -1 : 0 */
> +(for eqne (eq ne)
> + (simplify
> + (eqne:c (vec_cond @0 uniform_integer_cst_p@2 uniform_integer_cst_p@3)
> + (vec_cond @1 @2 @3))
> + (with
> + {
> + tree newop1 = @2;
> + tree newop2 = @3;
> + if (eqne == NE_EXPR)
> + std::swap (newop1, newop2);
> + }
> + (if (integer_all_onesp (@2) && integer_zerop (@3))
> + (vec_cond (bit_xor @0 @1) {newop1;} {newop2;})
> + (if (integer_all_onesp (@3) && integer_zerop (@2))
> + (vec_cond (bit_xor @0 @1) {newop2;} {newop1;}))))))
> +
> /* Canonicalize mask ? { 0, ... } : { -1, ...} to ~mask if the mask
> types are compatible. */
> (simplify
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr111150.c
> b/gcc/testsuite/gcc.dg/tree-ssa/pr111150.c
> new file mode 100644
> index 00000000000..d10564fd722
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/pr111150.c
> @@ -0,0 +1,19 @@
> +/* PR tree-optimization/111150 */
> +/* { dg-do compile } */
> +/* { dg-options "-O1 -fdump-tree-forwprop1" } */
> +
> +typedef int v4si __attribute((__vector_size__(4 * sizeof(int))));
> +
> +v4si f1_(v4si a, v4si b, v4si c, v4si d) {
> + v4si X = a == b;
> + v4si Y = c == d;
> + return (X != Y);
> +}
> +
> +v4si f2_(v4si a, v4si b, v4si c, v4si d) {
> + v4si X = a == b;
> + v4si Y = c == d;
> + return (X == Y);
> +}
> +
> +/* { dg-final { scan-tree-dump-times " VEC_COND_EXPR " 2 "forwprop1" } } */
> --
> 2.17.1
>
