On Thu, 16 Jun 2022, Tamar Christina wrote:
> Hi All,
>
> This optimizes the following sequence
>
> ((a < b) & c) | ((a >= b) & d)
>
> into
>
> (a < b ? c : d) & 1
>
> for scalar. On vector we can omit the & 1.
>
> This changes the code generation from
>
> zoo2:
> cmp w0, w1
> cset w0, lt
> cset w1, ge
> and w0, w0, w2
> and w1, w1, w3
> orr w0, w0, w1
> ret
>
> into
>
> cmp w0, w1
> csel w0, w2, w3, lt
> and w0, w0, 1
> ret
>
> and significantly reduces the number of selects we have to do in the vector
> code.
>
> Bootstrapped Regtested on aarch64-none-linux-gnu,
> x86_64-pc-linux-gnu and no issues.
>
> Ok for master?
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> * fold-const.cc (inverse_conditions_p): Traverse if SSA_NAME.
> * match.pd: Add new rule.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/aarch64/if-compare_1.c: New test.
> * gcc.target/aarch64/if-compare_2.c: New test.
>
> --- inline copy of patch --
> diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc
> index
> 39a5a52958d87497f301826e706886b290771a2d..f180599b90150acd3ed895a64280aa3255061256
> 100644
> --- a/gcc/fold-const.cc
> +++ b/gcc/fold-const.cc
> @@ -2833,15 +2833,38 @@ compcode_to_comparison (enum comparison_code code)
> bool
> inverse_conditions_p (const_tree cond1, const_tree cond2)
> {
> - return (COMPARISON_CLASS_P (cond1)
> - && COMPARISON_CLASS_P (cond2)
> - && (invert_tree_comparison
> - (TREE_CODE (cond1),
> - HONOR_NANS (TREE_OPERAND (cond1, 0))) == TREE_CODE (cond2))
> - && operand_equal_p (TREE_OPERAND (cond1, 0),
> - TREE_OPERAND (cond2, 0), 0)
> - && operand_equal_p (TREE_OPERAND (cond1, 1),
> - TREE_OPERAND (cond2, 1), 0));
> + if (COMPARISON_CLASS_P (cond1)
> + && COMPARISON_CLASS_P (cond2)
> + && (invert_tree_comparison
> + (TREE_CODE (cond1),
> + HONOR_NANS (TREE_OPERAND (cond1, 0))) == TREE_CODE (cond2))
> + && operand_equal_p (TREE_OPERAND (cond1, 0),
> + TREE_OPERAND (cond2, 0), 0)
> + && operand_equal_p (TREE_OPERAND (cond1, 1),
> + TREE_OPERAND (cond2, 1), 0))
> + return true;
> +
> + if (TREE_CODE (cond1) == SSA_NAME
> + && TREE_CODE (cond2) == SSA_NAME)
> + {
> + gimple *gcond1 = SSA_NAME_DEF_STMT (cond1);
> + gimple *gcond2 = SSA_NAME_DEF_STMT (cond2);
> + if (!is_gimple_assign (gcond1) || !is_gimple_assign (gcond2))
> + return false;
> +
> + tree_code code1 = gimple_assign_rhs_code (gcond1);
> + tree_code code2 = gimple_assign_rhs_code (gcond2);
> + return TREE_CODE_CLASS (code1) == tcc_comparison
> + && TREE_CODE_CLASS (code2) == tcc_comparison
> + && invert_tree_comparison (code1,
> + HONOR_NANS (gimple_arg (gcond1, 0))) == code2
> + && operand_equal_p (gimple_arg (gcond1, 0),
> + gimple_arg (gcond2, 0), 0)
> + && operand_equal_p (gimple_arg (gcond1, 1),
> + gimple_arg (gcond2, 1), 0);
> + }
> +
> + return false;
if we do extend inverse_condition_p please add an overload like
bool
inverse_condition_p (enum tree_code, tree op00, tree op01,
enum tree_code, tree op10, tree op11)
so you can avoid some code duplication here.
> }
>
> /* Return a tree for the comparison which is the combination of
> diff --git a/gcc/match.pd b/gcc/match.pd
> index
> 6d691d302b339c0e4556b40af158b5208c12d08f..bad49dd348add751d9ec1e3023e34d9ac123194f
> 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -1160,6 +1160,32 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (convert (bit_and (negate (convert:utype { pmop[0]; }))
> (convert:utype @1)))))))
>
> +/* Fold (((a < b) & c) | ((a >= b) & d)) into (a < b ? c : d) & 1. */
> +(simplify
> + (bit_ior
> + (bit_and:c (convert? @0) @2)
> + (bit_and:c (convert? @1) @3))
in case the comparison returns a signed bool this might turn out wrong.
Maybe simply use zero_one_valued_p@0 here instead of (convert? @0)?
> + (if (inverse_conditions_p (@0, @1)
> + /* The scalar version has to be canonicalized after vectorization
> + because it makes unconditional loads conditional ones, which
> + means we lose vectorization because the loads may trap. */
> + && canonicalize_math_after_vectorization_p ())
> + (bit_and (cond @0 @2 @3) { build_each_one_cst (type); })))
I think you should restrict this to INTEGRAL_TYPE_P and use
build_one_cst (type) (also see below).
you can do inverse_onditions_p with lock-step for over
tcc_comparison and inverted_tcc_comparison{,_with_nans} (see existing
examples).
> +(simplify
> + (bit_ior
> + (bit_and:c (convert? (vec_cond:s @0 @4 integer_zerop)) @2)
> + (bit_and:c (convert? (vec_cond:s @1 @4 integer_zerop)) @3))
> + (if (inverse_conditions_p (@0, @1)
> + && integer_onep (@4))
> + (bit_and (vec_cond @0 @2 @3) @4)))
> +/* Fold (((a < b) & c) | ((a >= b) & d)) into a < b ? c : d. */
> +(simplify
> + (bit_ior
> + (bit_and:c (convert? (vec_cond:s @0 integer_minus_onep integer_zerop)) @2)
> + (bit_and:c (convert? (vec_cond:s @1 integer_minus_onep integer_zerop)) @3))
> + (if (inverse_conditions_p (@0, @1))
> + (vec_cond @0 @2 @3)))
I think the duplication is pre-mature optimization - we should get the
(bit_and (..) integer_minus_onep) simplified. Also didn't we have
(vec_cond @0 -1 0) -> (view_convert @0) when the modes match?
We might want to add (match zero_minus_one_valued_p) or use
truth_valued_p (with appropriate vector semantics, plus extend it).
Why do you need (convert? ...) for the vector case btw?
I guess the integral type and vector type cases are similar enough
that the patterns can be merged with conditionalizing only the
replacement.
> +
> /* X % Y is smaller than Y. */
> (for cmp (lt ge)
> (simplify
> diff --git a/gcc/testsuite/gcc.target/aarch64/if-compare_1.c
> b/gcc/testsuite/gcc.target/aarch64/if-compare_1.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..05a1292fa90c70b14a7985122f43711f55d047ea
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/if-compare_1.c
> @@ -0,0 +1,16 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O" } */
> +/* { dg-final { check-function-bodies "**" "" "" { target { le } } } } */
> +
> +/*
> +**zoo2:
> +** cmp w0, w1
> +** csel w0, w2, w3, lt
> +** and w0, w0, 1
> +** ret
> +*/
> +int zoo2 (int a, int b, int c, int d)
> +{
> + return ((a < b) & c) | ((a >= b) & d);
> +}
> +
> diff --git a/gcc/testsuite/gcc.target/aarch64/if-compare_2.c
> b/gcc/testsuite/gcc.target/aarch64/if-compare_2.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..34bc65f5db10eae81b8dee3316dfb7d12bf471c8
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/if-compare_2.c
> @@ -0,0 +1,32 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O3 -std=c99" } */
> +/* { dg-final { check-function-bodies "**" "" "" { target { le } } } } */
> +
> +typedef int v4si __attribute__ ((vector_size (16)));
> +
> +/*
> +**foo:
> +** cmgt v0.4s, v1.4s, v0.4s
> +** bsl v0.16b, v2.16b, v3.16b
> +** ret
> +*/
> +v4si foo (v4si a, v4si b, v4si c, v4si d) {
> + return ((a < b) & c) | ((a >= b) & d);
> +}
> +
> +
> +/**
> +**bar:
> +**...
> +** cmge v[0-9]+.4s, v[0-9]+.4s, v[0-9]+.4s
> +** bsl v[0-9]+.16b, v[0-9]+.16b, v[0-9]+.16b
> +** and v[0-9]+.16b, v[0-9]+.16b, v[0-9]+.16b
> +**...
> +*/
> +void bar (int * restrict a, int * restrict b, int * restrict c,
> + int * restrict d, int * restrict res, int n)
> +{
> + for (int i = 0; i < (n & -4); i++)
> + res[i] = ((a[i] < b[i]) & c[i]) | ((a[i] >= b[i]) & d[i]);
> +}
> +
>
>
>
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstraße 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
