Re: [patch] [RFC] Fix PR 37385 - ICE in set_mem_alias_set with the vectorizer

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Wed, Sep 10, 2008 at 5:21 AM, Ira Rosen <IRAR@il.ibm.com> wrote:
>
>
> gcc-patches-owner@gcc.gnu.org wrote on 10/09/2008 10:18:10:
>
>>
>>
>> "Richard Guenther" <richard.guenther@gmail.com> wrote on 09/09/2008
>> 12:14:53:
>>
>> > On Tue, Sep 9, 2008 at 10:20 AM, Ira Rosen <IRAR@il.ibm.com> wrote:
>> > >
>> > > The fix of PR 36493
>> > > (http://gcc.gnu.org/ml/gcc-patches/2008-06/msg01023.html) verifies
> that
>> the
>> > > accesses through the vector pointer and the original scalar memory
>> > > reference alias. In addition before that patch, the type of vector
>> store
>> > > statement was determined by the RHS of the scalar statement, and now
> it
>> is
>> > > determined by the type of LHS, which seems to cause the ICE in
>> > > set_mem_alias_set, at emit-rtl.c:1789.
>> > >
>> > > The attached patch reverts the second part of the PR 36493 patch.
> This
>> > > fixes the problem (without hurting PR 36493). I am not sure that the
>> fix is
>> > > correct, since the types of RHS and LHS are compatible anyway.
>> > > I am bootstrapping and testing the patch on x86_64-linux now.
>> >
>> > I think it's reasonable.  This may be papering over a real problem
>> somewhere
>> > else as we are not actually verifying if the types in our statements
>> > are correct.
>>
>> Committed revision 140195.
>>
>> Ira
>>
>>
>
> I realized, that now, after changing the vectype to be the type of the rhs,
> the vectype being checked in vectorizable_store for alias purposes is not
> the correct type (but the type of the lhs). This patch fixes that.
>
> Bootstrapped and now testing on x86_64-linux. O.K. to apply once the
> testing completes?
>
> Thanks,
> Ira
>
> Index: tree-vect-transform.c
> ===================================================================
> --- tree-vect-transform.c       (revision 140195)
> +++ tree-vect-transform.c       (working copy)
> @@ -5180,6 +5180,9 @@ vectorizable_store (gimple stmt, gimple_
>       return false;
>     }
>
> +  /* The type of the vector store is determined by the rhs.  */
> +  vectype = get_vectype_for_scalar_type (TREE_TYPE (op));
> +
>   /* If accesses through a pointer to vectype do not alias the original
>      memory reference we have a problem.  */
>   if (get_alias_set (vectype) != get_alias_set (TREE_TYPE (scalar_dest))
>
>

Hi,

This patch caused:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37491


-- 
H.J.