[PATCH] vect: Move suggested_unroll_factor applying [PR105940]

[PATCH] vect: Move suggested_unroll_factor applying [PR105940]

[Kewen.Lin]

Hi,

As PR105940 shown, when rs6000 port tries to assign
m_suggested_unroll_factor by 4 or so, there will be ICE
on below statement:

  exact_div (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
             loop_vinfo->suggested_unroll_factor);

In function vect_analyze_loop_2, the current place of
suggested_unroll_factor applying can't guarantee it's
applied for all cases.  As the case shows, vectorizer
could retry with SLP forced off, the vf is reset by
saved_vectorization_factor which isn't applied with
suggested_unroll_factor before.  It means it can end
up with one vf which neglects suggested_unroll_factor.
I think it's off design, we should move the applying
of suggested_unroll_factor after start_over.

Bootstrapped and regtested on x86_64-redhat-linux,
aarch64-linux-gnu and powerpc64{,le}-linux-gnu.

Is it ok for trunk?

BR,
Kewen
-----
	PR tree-optimization/105940

gcc/ChangeLog:

	* tree-vect-loop.cc (vect_analyze_loop_2): Move the place of
	applying suggested_unroll_factor after start_over.
---
 gcc/tree-vect-loop.cc | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 896218f23ea..af955d26f8d 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -2393,15 +2393,15 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal,
      set of rgroups.  */
   gcc_assert (LOOP_VINFO_MASKS (loop_vinfo).is_empty ());

+  /* This is the point where we can re-start analysis with SLP forced off.  */
+start_over:
+
   /* Apply the suggested unrolling factor, this was determined by the backend
      during finish_cost the first time we ran the analyzis for this
      vector mode.  */
   if (loop_vinfo->suggested_unroll_factor > 1)
     LOOP_VINFO_VECT_FACTOR (loop_vinfo) *= loop_vinfo->suggested_unroll_factor;

-  /* This is the point where we can re-start analysis with SLP forced off.  */
-start_over:
-
   /* Now the vectorization factor is final.  */
   poly_uint64 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
   gcc_assert (known_ne (vectorization_factor, 0U));
--
2.27.0

Re: [PATCH] vect: Move suggested_unroll_factor applying [PR105940]

[Richard Biener]

On Mon, Jun 13, 2022 at 12:02 PM Kewen.Lin <linkw@linux.ibm.com> wrote:
>
> Hi,
>
> As PR105940 shown, when rs6000 port tries to assign
> m_suggested_unroll_factor by 4 or so, there will be ICE
> on below statement:
>
>   exact_div (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
>              loop_vinfo->suggested_unroll_factor);
>
> In function vect_analyze_loop_2, the current place of
> suggested_unroll_factor applying can't guarantee it's
> applied for all cases.  As the case shows, vectorizer
> could retry with SLP forced off, the vf is reset by
> saved_vectorization_factor which isn't applied with
> suggested_unroll_factor before.  It means it can end
> up with one vf which neglects suggested_unroll_factor.
> I think it's off design, we should move the applying
> of suggested_unroll_factor after start_over.
>
> Bootstrapped and regtested on x86_64-redhat-linux,
> aarch64-linux-gnu and powerpc64{,le}-linux-gnu.
>
> Is it ok for trunk?

OK (I think the GCC 12 branch is also affected).

Richard.

>
> BR,
> Kewen
> -----
>         PR tree-optimization/105940
>
> gcc/ChangeLog:
>
>         * tree-vect-loop.cc (vect_analyze_loop_2): Move the place of
>         applying suggested_unroll_factor after start_over.
> ---
>  gcc/tree-vect-loop.cc | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> index 896218f23ea..af955d26f8d 100644
> --- a/gcc/tree-vect-loop.cc
> +++ b/gcc/tree-vect-loop.cc
> @@ -2393,15 +2393,15 @@ vect_analyze_loop_2 (loop_vec_info loop_vinfo, bool &fatal,
>       set of rgroups.  */
>    gcc_assert (LOOP_VINFO_MASKS (loop_vinfo).is_empty ());
>
> +  /* This is the point where we can re-start analysis with SLP forced off.  */
> +start_over:
> +
>    /* Apply the suggested unrolling factor, this was determined by the backend
>       during finish_cost the first time we ran the analyzis for this
>       vector mode.  */
>    if (loop_vinfo->suggested_unroll_factor > 1)
>      LOOP_VINFO_VECT_FACTOR (loop_vinfo) *= loop_vinfo->suggested_unroll_factor;
>
> -  /* This is the point where we can re-start analysis with SLP forced off.  */
> -start_over:
> -
>    /* Now the vectorization factor is final.  */
>    poly_uint64 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
>    gcc_assert (known_ne (vectorization_factor, 0U));
> --
> 2.27.0

Re: [PATCH] vect: Move suggested_unroll_factor applying [PR105940]

[Kewen.Lin]

on 2022/6/13 19:38, Richard Biener wrote:
> On Mon, Jun 13, 2022 at 12:02 PM Kewen.Lin <linkw@linux.ibm.com> wrote:
>>
>> Hi,
>>
>> As PR105940 shown, when rs6000 port tries to assign
>> m_suggested_unroll_factor by 4 or so, there will be ICE
>> on below statement:
>>
>>   exact_div (LOOP_VINFO_VECT_FACTOR (loop_vinfo),
>>              loop_vinfo->suggested_unroll_factor);
>>
>> In function vect_analyze_loop_2, the current place of
>> suggested_unroll_factor applying can't guarantee it's
>> applied for all cases.  As the case shows, vectorizer
>> could retry with SLP forced off, the vf is reset by
>> saved_vectorization_factor which isn't applied with
>> suggested_unroll_factor before.  It means it can end
>> up with one vf which neglects suggested_unroll_factor.
>> I think it's off design, we should move the applying
>> of suggested_unroll_factor after start_over.
>>
>> Bootstrapped and regtested on x86_64-redhat-linux,
>> aarch64-linux-gnu and powerpc64{,le}-linux-gnu.
>>
>> Is it ok for trunk?
> 
> OK (I think the GCC 12 branch is also affected).
> 

Thanks Richi!  Committed as r13-1083.

I'll backport this to GCC12 branch in a week if it goes well in trunk.

BR,
Kewen