]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
Fix possible wrong-code with masked store-lanes
authorRichard Biener <rguenther@suse.de>
Thu, 10 Oct 2024 12:00:11 +0000 (14:00 +0200)
committerRichard Biener <rguenth@gcc.gnu.org>
Fri, 11 Oct 2024 05:59:27 +0000 (07:59 +0200)
When we're doing masked store-lanes one mask element applies to all
loads of one struct element.  This requires uniform masks for all
of the SLP lanes, something we already compute into STMT_VINFO_SLP_VECT_ONLY
but fail to check when doing SLP store-lanes.  The following corrects
this.  The following also adjusts the store-lane heuristic to properly
check for masked or non-masked optab support.

* tree-vect-slp.cc (vect_slp_prefer_store_lanes_p): Allow
passing in of vectype, pass in whether the stores are masked
and query the correct optab.
(vect_build_slp_instance): Guard store-lanes query with
! STMT_VINFO_SLP_VECT_ONLY, guaranteeing an uniform mask.

gcc/tree-vect-slp.cc

index 96f1992cfbff4290c5a6ccd4ad7c9e4d5764b812..3024b87a1f835b59d3964b2723c2233b7d5d6df1 100644 (file)
@@ -3504,17 +3504,22 @@ vect_match_slp_patterns (slp_instance instance, vec_info *vinfo,
 }
 
 /* STMT_INFO is a store group of size GROUP_SIZE that we are considering
-   splitting into two, with the first split group having size NEW_GROUP_SIZE.
+   vectorizing with VECTYPE that might be NULL.  MASKED_P indicates whether
+   the stores are masked.
    Return true if we could use IFN_STORE_LANES instead and if that appears
    to be the better approach.  */
 
 static bool
 vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
+                              tree vectype, bool masked_p,
                               unsigned int group_size,
                               unsigned int new_group_size)
 {
-  tree scalar_type = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
-  tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
+  if (!vectype)
+    {
+      tree scalar_type = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
+      vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
+    }
   if (!vectype)
     return false;
   /* Allow the split if one of the two new groups would operate on full
@@ -3528,7 +3533,7 @@ vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
   if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
       || multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
     return false;
-  return vect_store_lanes_supported (vectype, group_size, false) != IFN_LAST;
+  return vect_store_lanes_supported (vectype, group_size, masked_p) != IFN_LAST;
 }
 
 /* Analyze an SLP instance starting from a group of grouped stores.  Call
@@ -3973,6 +3978,10 @@ vect_build_slp_instance (vec_info *vinfo,
       else if (is_a <loop_vec_info> (vinfo)
               && (group_size != 1 && i < group_size))
        {
+         gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
+         bool masked_p = call
+             && gimple_call_internal_p (call)
+             && internal_fn_mask_index (gimple_call_internal_fn (call)) != -1;
          /* There are targets that cannot do even/odd interleaving schemes
             so they absolutely need to use load/store-lanes.  For now
             force single-lane SLP for them - they would be happy with
@@ -3987,9 +3996,10 @@ vect_build_slp_instance (vec_info *vinfo,
          bool want_store_lanes
            = (! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
               && ! STMT_VINFO_STRIDED_P (stmt_info)
+              && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
               && compare_step_with_zero (vinfo, stmt_info) > 0
-              && vect_slp_prefer_store_lanes_p (vinfo, stmt_info,
-                                                group_size, 1));
+              && vect_slp_prefer_store_lanes_p (vinfo, stmt_info, NULL_TREE,
+                                                masked_p, group_size, 1));
          if (want_store_lanes || force_single_lane)
            i = 1;
 
@@ -4074,14 +4084,14 @@ vect_build_slp_instance (vec_info *vinfo,
 
          /* Now re-assess whether we want store lanes in case the
             discovery ended up producing all single-lane RHSs.  */
-         if (rhs_common_nlanes == 1
+         if (! want_store_lanes
+             && rhs_common_nlanes == 1
              && ! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
              && ! STMT_VINFO_STRIDED_P (stmt_info)
+             && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
              && compare_step_with_zero (vinfo, stmt_info) > 0
              && (vect_store_lanes_supported (SLP_TREE_VECTYPE (rhs_nodes[0]),
-                                             group_size,
-                                             SLP_TREE_CHILDREN
-                                               (rhs_nodes[0]).length () != 1)
+                                             group_size, masked_p)
                  != IFN_LAST))
            want_store_lanes = true;