On Mon, 12 May 2025, Icen Zeyada wrote:
> Generalize existing scalar gimple_fold rules to apply the same
> bitwise comparison simplifications to vector types. Previously, an
> expression like
>
> (x < y) && (x > y)
>
> would fold to `false` if x and y are scalars, but equivalent vector
> comparisons were left untouched. This patch enables folding of
> patterns of the form
>
> (cmp x y) bit_and (cmp x y)
> (cmp x y) bit_ior (cmp x y)
>
> for vector operands as well, ensuring consistent optimization across
> all data types.
>
> PR tree-optimization/119196
>
> gcc/ChangeLog:
>
> * match.pd: Allow scalar optimizations with bitwise AND/OR to apply to
> vectors.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/vector-compare-5.c: Add new test for vector
> compare simplification.
>
> Signed-off-by: Icen Zeyada <icen.zeya...@arm.com>
> ---
> gcc/match.pd | 19 +++++++--
> .../gcc.target/aarch64/vector-compare-5.c | 41 +++++++++++++++++++
> 2 files changed, 56 insertions(+), 4 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/vector-compare-5.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index ab496d923cc0..a8a2e01e5e64 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -3620,7 +3620,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (bit_and:c (code1:c@3 @0 @1) (code2:c@4 (convert?@c0 @0) @2))
> (if ((TREE_CODE (@1) == INTEGER_CST
> && TREE_CODE (@2) == INTEGER_CST)
> - || ((INTEGRAL_TYPE_P (TREE_TYPE (@1))
> + || ((ANY_INTEGRAL_TYPE_P (TREE_TYPE (@1))
> || POINTER_TYPE_P (TREE_TYPE (@1)))
> && bitwise_equal_p (@1, @2)))
> (with
> @@ -3697,7 +3697,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (bit_and (code1:c@3 @0 @1) (code2:c@4 @0 @2))
> (if ((TREE_CODE (@1) == INTEGER_CST
> && TREE_CODE (@2) == INTEGER_CST)
> - || ((INTEGRAL_TYPE_P (TREE_TYPE (@1))
> + || ((ANY_INTEGRAL_TYPE_P (TREE_TYPE (@1))
> || POINTER_TYPE_P (TREE_TYPE (@1)))
> && operand_equal_p (@1, @2)))
> (with
> @@ -3747,7 +3747,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (bit_ior:c (code1:c@3 @0 @1) (code2:c@4 (convert?@c0 @0) @2))
> (if ((TREE_CODE (@1) == INTEGER_CST
> && TREE_CODE (@2) == INTEGER_CST)
> - || ((INTEGRAL_TYPE_P (TREE_TYPE (@1))
> + || ((ANY_INTEGRAL_TYPE_P (TREE_TYPE (@1))
> || POINTER_TYPE_P (TREE_TYPE (@1)))
> && bitwise_equal_p (@1, @2)))
> (with
> @@ -3880,7 +3880,8 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> rcmp (eq gt le eq ge lt)
> (simplify
> (eq:c (cmp1:c @0 @1) (cmp2 @0 @1))
> - (if (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
> + (if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
> + || POINTER_TYPE_P (TREE_TYPE (@0)))
> (rcmp @0 @1))))
For all of the above you need to ensure that we can
expand the vector comparison via expand_vec_cmp_expr_p.
> /* (type)([0,1]@a != 0) -> (type)a
> @@ -6510,6 +6511,16 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> { build_int_cst (integer_type_node, prec - 1);}))))))
> #endif
>
> +(for op1 (simple_comparison)
> + (for op2 (simple_comparison)
> + (for lop (bit_and bit_ior)
> + (simplify
> + (lop
> + (vec_cond (op1 @0 @1) integer_minus_onep@2 integer_zerop@3)
> + (vec_cond (op2 @0 @1) integer_minus_onep@2 integer_zerop@3))
> + (with { tree op_type = truth_type_for (TREE_TYPE (@0)); }
> + (vec_cond (lop:op_type (op1 @0 @1) (op2 @0 @1)) @2 @3))))))
> +
Likewise here, also for the vec_cond via expand_vec_cond_expr_p. Why
do you compute a "new" op_type here and not re-use that of the
(op1 @0 @1) compares and the compares themselves? I'd have expected
> + (simplify
> + (lop
> + (vec_cond (op1@00 @0 @1) integer_minus_onep@2 integer_zerop@3)
> + (vec_cond (op2@01 @0 @1) integer_minus_onep@2 integer_zerop@3))
(vec_cond (lop @00 @01) @2 @3)
and why match a comparison at all? Why not
(simplify
(lop
(vec_cond @0 integer_minus_onep@2 integer_zerop@3)
(vec_cond @1 @2 @3))
(vec_cond (lop @0 @1) @2 @3))
?
IMO this pattern should be in a separate patch.
> (for cnd (cond vec_cond)
> /* (a != b) ? (a - b) : 0 -> (a - b) */
> (simplify
> diff --git a/gcc/testsuite/gcc.target/aarch64/vector-compare-5.c
> b/gcc/testsuite/gcc.target/aarch64/vector-compare-5.c
> new file mode 100644
> index 000000000000..c4b95a21996d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/vector-compare-5.c
> @@ -0,0 +1,41 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +/* { dg-additional-options "-fdump-tree-original-all" } */
> +
> +typedef int v4i __attribute__((vector_size(4*sizeof(int))));
> +
> +/* Ensure we can simplify `VEC_COND_EXPR(a OP1 b) OP2 VEC_COND_EXPR(a OP3 b)`
> + * into `VEC_COND_EXPR(a OP4 b)`
> + */
> +
> +void use (v4i const *z);
> +
> +void
> +g (v4i *x, v4i const *y, v4i *z, v4i *t)
> +{
> + *z = *x > *y | *x == *y; // expect >=
> + *t = *x > *y | *x <= *y; // expect true
> +}
> +
> +void
> +h (v4i *x, v4i const *y, v4i *z, v4i *t)
> +{
> + *z = *x <= *y & *x >= *y; // expect x == y
> + *t = *x <= *y & *x != *y; // expect x<y
> +}
> +
> +void
> +i (v4i *x, v4i const *y, v4i *z, v4i *t)
> +{
> + *z = *x == *y | *x != *y; // expect true
> + *t = *x == *y & *x != *y; // expect false
> +}
> +
> +/* { dg-final { scan-tree-dump
> ".*\\*zD\\.\\d+\\s*=\\s*VEC_COND_EXPR\\s*<\\s*\\*xD\\.\\d+\\s*>=\\s*VIEW_CONVERT_EXPR<v4iD\\.\\d+>\\(\\*yD\\.\\d+\\)\\s*,\\s*\\{\\s*-1(,\\s*-1){3}\\s*\\}\\s*,\\s*\\{\\s*0(,\\s*0){3}\\s*\\}\\s*>\\s*;"
> "original" } } */
> +/* { dg-final { scan-tree-dump
> ".*\\*tD\\.\\d+\\s*=\\s*\\{\\s*-1(,\\s*-1){3}\\s*\\}\\s*;" "original" } } */
> +/* { dg-final { scan-tree-dump
> ".*\\*zD\\.\\d+\\s*=\\s*VEC_COND_EXPR\\s*<\\s*\\*xD\\.\\d+\\s*==\\s*VIEW_CONVERT_EXPR<v4iD\\.\\d+>\\(\\*yD\\.\\d+\\)\\s*,\\s*\\{\\s*-1(,\\s*-1){3}\\s*\\}\\s*,\\s*\\{\\s*0(,\\s*0){3}\\s*\\}\\s*>\\s*;"
> "original" } } */
> +/* { dg-final { scan-tree-dump
> ".*\\*tD\\.\\d+\\s*=\\s*VEC_COND_EXPR\\s*<\\s*\\*xD\\.\\d+\\s*<\\s*VIEW_CONVERT_EXPR<v4iD\\.\\d+>\\(\\*yD\\.\\d+\\)\\s*,\\s*\\{\\s*-1(,\\s*-1){3}\\s*\\}\\s*,\\s*\\{\\s*0(,\\s*0){3}\\s*\\}\\s*>\\s*;"
> "original" } } */
> +/* { dg-final { scan-tree-dump
> ".*\\*zD\\.\\d+\\s*=\\s*\\{\\s*-1(,\\s*-1){3}\\s*\\}\\s*;" "original" } } */
> +/* { dg-final { scan-tree-dump
> ".*\\*tD\\.\\d+\\s*=\\s*\\{\\s*0(,\\s*0){3}\\s*\\}\\s*;" "original" } } */
> +
> +
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
