On Thu, May 15, 2025 at 10:29 AM MCC CS <mc...@gmx.com> wrote:
>
> Dear all,
>
> Here's my patch for PR120265. Bootstrapped and tested on aarch64 that it
> causes no regressions. I also added a testcase. I'd be grateful
> if you could commit it.
>
> Otherwise, feedback to improve it is welcome.
>
> Many thanks
> MCCCS
>
>
> From 1e901c3fa5c8cc3e55d4f1715b4aae4ae3d66714 Mon Sep 17 00:00:00 2001
> From: MCCCS <mc...@gmx.com>
> Date: Thu, 15 May 2025 09:16:49 +0100
> Subject: [PATCH] tree-optimization/120265 - Optimize modular counters
>
> This PR is about replacing trunc_mod with with a
> simpler expression given the bounds of variables.
>
> PR tree-optimization/120265
> * match.pd:
> X % M -> X for X in 0 to M-1
> X % M -> (X == M) ? 0 : X for X in 0 to M
> X % M -> (X >= M) ? (X - M) : X for X in 0 to 2*M-1.
While the first case looks profitable the 2nd and third might not be
when optimizing for size and they are definitely not canonicalizations
and thus might interfere with association with division? So I wonder
whether those are better suited to be performed at RTL expansion time?
Like on AVR when type is bigger than word_mode the transform looks bad.
When the target does not have a conditional move instruction the generated
branch might be also difficult to predict?
I would have expected we do the first pattern in VRPs simplification
step, but I might misremeber.
Richard.
> * gcc.dg/pr120265.c. New testcase.
> ---
> gcc/match.pd | 27 ++++++++++++++++++++
> gcc/testsuite/gcc.dg/pr120265.c | 44 +++++++++++++++++++++++++++++++++
> 2 files changed, 71 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/pr120265.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 79485f9678a..bd8950b4e10 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -5602,6 +5602,33 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> optab_vector)))
> (eq (trunc_mod @0 @1) { build_zero_cst (TREE_TYPE (@0)); })))
>
> +#if GIMPLE
> +/* X % M -> X for X in 0 to M-1. */
> +/* X % M -> (X == M) ? 0 : X for X in 0 to M. */
> +/* X % M -> (X >= M) ? (X - M) : X for X in 0 to 2*M-1. */
> +(simplify
> + (trunc_mod @0 @1)
> + (with { int_range_max vr0, vr1; }
> + (if (get_range_query (cfun)->range_of_expr (vr0, @0)
> + && get_range_query (cfun)->range_of_expr (vr1, @1)
> + && !vr0.undefined_p ()
> + && !vr1.undefined_p ()
> + && !integer_zerop (@1)
> + && (TYPE_UNSIGNED (type)
> + || (vr0.nonnegative_p () && vr1.nonnegative_p ())))
> + (with
> + { wide_int twice = 2 * vr1.lower_bound (); }
> + (switch
> + (if (wi::gtu_p (vr1.lower_bound (), vr0.upper_bound ()))
> + @0)
> + (if (wi::geu_p (vr1.lower_bound (), vr0.upper_bound ()))
> + (cond (eq @0 @1)
> + { build_zero_cst (type); }
> + @0))
> + (if (wi::gtu_p (twice, vr0.upper_bound ()))
> + (cond (ge @0 @1) (minus @0 @1) @0)))))))
> +#endif
> +
> /* ((X /[ex] C1) +- C2) * (C1 * C3) --> (X * C3) +- (C1 * C2 * C3). */
> (for op (plus minus)
> (simplify
> diff --git a/gcc/testsuite/gcc.dg/pr120265.c b/gcc/testsuite/gcc.dg/pr120265.c
> new file mode 100644
> index 00000000000..2634af36226
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr120265.c
> @@ -0,0 +1,44 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O -fdump-tree-optimized" } */
> +__attribute__((noipa)) void g(int r)
> +{
> + (void) r;
> +}
> +
> +int x;
> +
> +void a(void)
> +{
> + unsigned m = 0;
> + for(int i = 0; i < 300; i++)
> + {
> + m++;
> + m %= 600;
> + g(m);
> + }
> +}
> +
> +void b(void)
> +{
> + unsigned m = 0;
> + for(int i = 0; i < x; i++)
> + {
> + m++;
> + m %= 600;
> + g(m);
> + }
> +}
> +
> +void c(void)
> +{
> + unsigned m = 0;
> + for(int i = 0; i < x; i++)
> + {
> + m += 7;
> + m %= 600;
> + g(m);
> + }
> +}
> +
> +/* { dg-final { scan-tree-dump-not "% 600" "optimized" } } */
> +
> --
> 2.45.2
