On Wed, 17 Jan 2024, Monk Chiang wrote:

> This allows the backend to generate movcc instructions, if target
> machine has movcc pattern.
> 
> branchless-cond.c needs to be updated since some target machines have
> conditional move instructions, and the experssion will not change to
> branchless expression.

While I agree this pattern should possibly be applied during RTL
expansion or instruction selection on x86 which also has movcc
the multiplication is cheaper.  So I don't think this isn't the way to go.

I'd rather revert the change than trying to "fix" it this way?

Thanks,
Richard.

> gcc/ChangeLog:
>       PR target/113095
>       * match.pd (`(zero_one == 0) ? y : z <op> y`,
>       `(zero_one != 0) ? z <op> y : y`): Do not match to branchless
>       expression, if target machine has conditional move pattern.
> 
> gcc/testsuite/ChangeLog:
>       * gcc.dg/tree-ssa/branchless-cond.c: Update testcase.
> ---
>  gcc/match.pd                                  | 30 +++++++++++++++++--
>  .../gcc.dg/tree-ssa/branchless-cond.c         |  6 ++--
>  2 files changed, 31 insertions(+), 5 deletions(-)
> 
> diff --git a/gcc/match.pd b/gcc/match.pd
> index e42ecaf9ec7..a1f90b1cd41 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -4231,7 +4231,20 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    (if (INTEGRAL_TYPE_P (type)
>         && TYPE_PRECISION (type) > 1
>         && (INTEGRAL_TYPE_P (TREE_TYPE (@0))))
> -       (op (mult (convert:type @0) @2) @1))))
> +   (with {
> +      bool can_movecc_p = false;
> +      if (can_conditionally_move_p (TYPE_MODE (type)))
> +     can_movecc_p = true;
> +
> +      /* Some target only support word_mode for movcc pattern, if type can
> +      extend to word_mode then use conditional move. Even if there is a
> +      extend instruction, the cost is lower than branchless.  */
> +      if (can_extend_p (word_mode, TYPE_MODE (type), TYPE_UNSIGNED (type))
> +       && can_conditionally_move_p (word_mode))
> +     can_movecc_p = true;
> +    }
> +    (if (!can_movecc_p)
> +     (op (mult (convert:type @0) @2) @1))))))
>  
>  /* (zero_one != 0) ? z <op> y : y -> ((typeof(y))zero_one * z) <op> y */
>  (for op (bit_xor bit_ior plus)
> @@ -4243,7 +4256,20 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    (if (INTEGRAL_TYPE_P (type)
>         && TYPE_PRECISION (type) > 1
>         && (INTEGRAL_TYPE_P (TREE_TYPE (@0))))
> -       (op (mult (convert:type @0) @2) @1))))
> +   (with {
> +      bool can_movecc_p = false;
> +      if (can_conditionally_move_p (TYPE_MODE (type)))
> +     can_movecc_p = true;
> +
> +      /* Some target only support word_mode for movcc pattern, if type can
> +      extend to word_mode then use conditional move. Even if there is a
> +      extend instruction, the cost is lower than branchless.  */
> +      if (can_extend_p (word_mode, TYPE_MODE (type), TYPE_UNSIGNED (type))
> +       && can_conditionally_move_p (word_mode))
> +     can_movecc_p = true;
> +    }
> +    (if (!can_movecc_p)
> +     (op (mult (convert:type @0) @2) @1))))))
>  
>  /* ?: Value replacement. */
>  /* a == 0 ? b : b + a  -> b + a */
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c 
> b/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> index e063dc4bb5f..c002ed97364 100644
> --- a/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> @@ -21,6 +21,6 @@ int f4(unsigned int x, unsigned int y, unsigned int z)
>    return ((x & 1) != 0) ? z | y : y;
>  }
>  
> -/* { dg-final { scan-tree-dump-times " \\\*" 4 "optimized" } } */
> -/* { dg-final { scan-tree-dump-times " & " 4 "optimized" } } */
> -/* { dg-final { scan-tree-dump-not "if " "optimized" } } */
> +/* { dg-final { scan-tree-dump-times " \\\*" 4 "optimized" { xfail { 
> "aarch64*-*-* alpha*-*-* bfin*-*-* epiphany-*-* i?86-*-* x86_64-*-* 
> nds32*-*-*" } } } } */
> +/* { dg-final { scan-tree-dump-times " & " 4 "optimized" { xfail { 
> "aarch64*-*-* alpha*-*-* bfin*-*-* epiphany-*-* i?86-*-* x86_64-*-* 
> nds32*-*-*" } } } } */
> +/* { dg-final { scan-tree-dump-not "if " "optimized" { xfail { "aarch64*-*-* 
> alpha*-*-* bfin*-*-* epiphany-*-* i?86-*-* x86_64-*-* nds32*-*-*" } } } } */
> 

-- 
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)

Reply via email to