Re: [PATCH v2] gimple-match.pd Add more optimization for gimple_cond

[Andrew Pinski <pinskia@gmail.com>]

On Mon, Nov 27, 2023 at 6:56 PM Feng Wang <wangfeng@eswincomputing.com> wrote:
>
> The link of PATCH v1: https://www.mail-archive.com/gcc-patches@gcc.gnu.org/msg326661.html
> This patch add another condition for gimple-cond optimization. Refer to
> the following test case.
> int foo1 (int data, int res)
> {
>   res = data & 0xf;
>   res |= res << 4;
>   if (res < 0x22)
>     return 0x22;
>   return res;
> }
> with the compilation flag "-O2",
> before this patch the log info of phiopt2 pass is
>   <bb 2> [local count: 1073741824]:
>   res_5 = data_1(D) & 15;
>   _6 = (unsigned int) res_5;
>   _7 = _6 * 17;
>   res_8 = (int) _7;
>   if (_7 <= 33)
>     goto <bb 3>; [21.72%]
>   else
>     goto <bb 4>; [78.28%]
>
>   <bb 3> [local count: 233216728]:
>
>   <bb 4> [local count: 1073741824]:
>   # _9 = PHI <res_8(2), 34(3)>
>   return _9;
> after this patch the the log info of phiopt2 pass is
>   <bb 2> [local count: 1073741824]:
>   res_5 = data_1(D) & 15;
>   _6 = (unsigned int) res_5;
>   _7 = _6 * 17;
>   res_8 = (int) _7;
>   _10 = MAX_EXPR <_7, 34>;
>   _3 = (int) _10;
>   return _3;
> This patch optimizes the phi node to generate "MAX_EXPR".
> The root cause of minmax replacement failure is the type of "_7"
> is unsigned, but the type of const_int "34" is signed. It makes
> types_match (c2_type, from_type) return false. So I add another
> condition to process this scenario.
>
> gcc/ChangeLog:
>
>         * match.pd: Add another condition to process type mismatch.

This should most likely be:
 ((cond (cmp (convert1? x) c1) (convert2? x) c2) pattern): Also allow
conversions that only change the sign.

>
> gcc/testsuite/ChangeLog:
>
>         * gcc.dg/tree-ssa/phi-opt-41.c: New test.
> ---
>  gcc/match.pd                               |  5 ++++-
>  gcc/testsuite/gcc.dg/tree-ssa/phi-opt-41.c | 24 ++++++++++++++++++++++
>  2 files changed, 28 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/phi-opt-41.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 95225e4ca5f..e864845bfa9 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -5419,7 +5419,10 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>          && (types_match (c2_type, from_type)
>              || (TYPE_PRECISION (c2_type) > TYPE_PRECISION (from_type)
>                  && (TYPE_UNSIGNED (from_type)
> -                    || TYPE_SIGN (c2_type) == TYPE_SIGN (from_type)))))
> +                    || TYPE_SIGN (c2_type) == TYPE_SIGN (from_type)))
> +             || (TYPE_UNSIGNED (from_type) != TYPE_UNSIGNED (c2_type)
> +                 && TYPE_PRECISION (c2_type) == TYPE_PRECISION (from_type)
> +                 && !TYPE_OVERFLOW_WRAPS (c2_type))))

What is the need for TYPE_OVERFLOW_WRAPS here? Also I think you just
need the check for TYPE_PRECISION instead of the rest.
Maybe instead of types_match here, tree_nop_conversion_p could be used
instead. I am not 100% sure though.

I also suspect you should add a few other testcases that don't depend
on VRP changing things. Maybe a runtime test too.

Thanks,
Andrew

>         {
>          if (cmp != EQ_EXPR)
>            code = minmax_from_comparison (cmp, @1, @3, @1, @2);
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-41.c b/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-41.c
> new file mode 100644
> index 00000000000..d1101c2f9f7
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-41.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-phiopt2" } */
> +
> +int foo1 (int data, int res)
> +{
> +  res = data & 0xf;
> +  res |= res << 4;
> +  if (res < 0x22)
> +    return 0x22;
> +  return res;
> +}
> +
> +int foo2 (int data, int res)
> +{
> +  res = data & 0xf;
> +  unsigned int r = res;
> +  r*=17;
> +  res = r;
> +  if (r < 0x22)
> +    return 0x22;
> +  return res;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "MAX_EXPR" 2 "phiopt2" } } */
> \ No newline at end of file
> --
> 2.17.1
>