[PATCH] Simple fix to enhance outer-loop vectorization.

[Yuri Rumyantsev]

Richard,

First of all, I don't think that it is possible to write out test for
outer-loop vectorization with zero-step reference because of possible
loop-carried dependencies and run-time aliasing is not supported for
outer-loop. If there are no such dependencies pre or pdse does
hoisting (sinking) of such invariant references. So I add a check on
it to accept zero-step references for outer loop marked with
forc-vectorize flag to guarantee absence of loop-carried dependencies
between inner-loop iterations.
I included run-time test that checks vectorization correctness.

Update patch is attached.
Yuri..

2015-05-28 14:39 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
> On Thu, May 28, 2015 at 1:00 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>> Hi All,
>>
>> Here is a simple patch which removes restriction on outer-loop
>> vectorization -  allow references in inner-loop with zero step. This
>> case was found in one important benchmark.
>>
>> Bootstrap and regression testing did not show any new failures.
>> Is it OK for trunk.
>>
>> ChangeLog:
>> 2015-05-28  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>
>> * tree-vect-data-refs.c (vect_analyze_data_ref_access): Allow
>> consecutive accesses within outer-loop vectorization for references
>> with zero step in inner-loop.
>>
>> gcc/testsuite/ChangeLog:
>> * gcc.dg/vect/fast-math-vect-outer-1.c: New test.
>
> Can you please add a non-omp-simd testcase that triggers this as well and that
> is a runtime testcase verifying the transform is correct?
>
> Also please don't add to the strange testcase-name machinery but just
> use { dg-additional-options "-ffast-math" }
>
> Index: tree-vect-data-refs.c
> ===================================================================
> --- tree-vect-data-refs.c       (revision 223653)
> +++ tree-vect-data-refs.c       (working copy)
> @@ -2261,7 +2261,6 @@
>    return true;
>  }
>
> -
>  /* Analyze the access pattern of the data-reference DR.
>     In case of non-consecutive accesses call vect_analyze_group_access() to
>     analyze groups of accesses.  */
>
> spurious white-space change
>
>
> @@ -2291,14 +2290,8 @@
>    if (loop_vinfo && integer_zerop (step))
>
> Surely the comment before this needs updating now.
>
>      {
>        GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
> -      if (nested_in_vect_loop_p (loop, stmt))
> -       {
> -         if (dump_enabled_p ())
> -           dump_printf_loc (MSG_NOTE, vect_location,
> -                            "zero step in inner loop of nest\n");
> -         return false;
> -       }
> -      return DR_IS_READ (dr);
> +      if (!nested_in_vect_loop_p (loop, stmt))
> +       return DR_IS_READ (dr);
>      }
>
>    if (loop && nested_in_vect_loop_p (loop, stmt))
>
> so what happens after the patch?  It would be nice to have a comment
> explaining what happens in the nested_in_vect_loop_p case for
> the case when the outer-loop step is zero and when it is not zero.
>
> In particular as you don't need any code generation changes - this hints
> at that you may miss something ;)
>
> Otherwise of course the patch is ok - lifting restrictions is good.
>
> Thanks,
> Richard.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: patch.1
Type: application/octet-stream
Size: 3186 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20150528/9cd1eb86/attachment.obj>