RE: [RFC, vectorizer] Allow single element vector types for vector reduction operations

[Richard Biener]

On Tue, 5 Sep 2017, Tamar Christina wrote:

> 
> 
> > -----Original Message-----
> > From: Richard Biener [mailto:rguent...@suse.de]
> > Sent: 05 September 2017 13:51
> > To: Tamar Christina
> > Cc: Andrew Pinski; Andreas Schwab; Jon Beniston; gcc-patches@gcc.gnu.org;
> > nd
> > Subject: RE: [RFC, vectorizer] Allow single element vector types for vector
> > reduction operations
> > 
> > On Tue, 5 Sep 2017, Richard Biener wrote:
> > 
> > > On Tue, 5 Sep 2017, Tamar Christina wrote:
> > >
> > > > Hi Richard,
> > > >
> > > > That was an really interesting analysis, thanks for the details!
> > > >
> > > > Would you be submitting the patch you proposed at the end as a fix?
> > >
> > > I'm testing it currently.
> > 
> > Unfortunately it breaks some required lowering.  I'll have to more closely
> > look at this.
> 
> Ah, ok. In the meantime, can this patch be reverted? It's currently breaking 
> spec for us so we're
> Not able to get any benchmarking numbers.

Testing the following instead:

Index: gcc/tree-vect-generic.c
===================================================================
--- gcc/tree-vect-generic.c     (revision 251642)
+++ gcc/tree-vect-generic.c     (working copy)
@@ -1640,7 +1640,7 @@ expand_vector_operations_1 (gimple_stmt_
       || code == VEC_UNPACK_FLOAT_LO_EXPR)
     type = TREE_TYPE (rhs1);
 
-  /* For widening/narrowing vector operations, the relevant type is of 
the
+  /* For widening vector operations, the relevant type is of the
      arguments, not the widened result.  VEC_UNPACK_FLOAT_*_EXPR is
      calculated in the same way above.  */
   if (code == WIDEN_SUM_EXPR
@@ -1650,9 +1650,6 @@ expand_vector_operations_1 (gimple_stmt_
       || code == VEC_WIDEN_MULT_ODD_EXPR
       || code == VEC_UNPACK_HI_EXPR
       || code == VEC_UNPACK_LO_EXPR
-      || code == VEC_PACK_TRUNC_EXPR
-      || code == VEC_PACK_SAT_EXPR
-      || code == VEC_PACK_FIX_TRUNC_EXPR
       || code == VEC_WIDEN_LSHIFT_HI_EXPR
       || code == VEC_WIDEN_LSHIFT_LO_EXPR)
     type = TREE_TYPE (rhs1);


also fix for a bug uncovered by the previous one:

Index: gcc/gimple-ssa-strength-reduction.c
===================================================================
--- gcc/gimple-ssa-strength-reduction.c (revision 251710)
+++ gcc/gimple-ssa-strength-reduction.c (working copy)
@@ -1742,8 +1742,7 @@ find_candidates_dom_walker::before_dom_c
        slsr_process_ref (gs);
 
       else if (is_gimple_assign (gs)
-              && SCALAR_INT_MODE_P
-                   (TYPE_MODE (TREE_TYPE (gimple_assign_lhs (gs)))))
+              && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (gs))))
        {
          tree rhs1 = NULL_TREE, rhs2 = NULL_TREE;