Hi!

With -fno-tree-dce the scalar MASK_LOAD isn't removed from the IL and we ICE
on it during expansion (as we support only the vector loads, if those aren't
supported, MASK_LOAD is either not created by if-conversion at all, or
vectorization refuses to vectorize the loop and thus it is cfg cleaned up
away.

The following patch replaces the scalar MASK_LOAD manually with load of
zero, similarly how we do it for vectorizable calls.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2014-03-18  Jakub Jelinek  <ja...@redhat.com>

        PR tree-optimization/60559
        * vectorizable_mask_load_store): Replace scalar MASK_LOAD
        with build_zero_cst assignment.

        * g++.dg/vect/pr60559.cc: New test.

--- gcc/tree-vect-stmts.c.jj    2014-03-03 08:24:33.000000000 +0100
+++ gcc/tree-vect-stmts.c       2014-03-18 14:01:40.969657763 +0100
@@ -2038,6 +2038,15 @@ vectorizable_mask_load_store (gimple stm
            STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
          prev_stmt_info = vinfo_for_stmt (new_stmt);
        }
+
+      /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed
+        from the IL.  */
+      tree lhs = gimple_call_lhs (stmt);
+      new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
+      set_vinfo_for_stmt (new_stmt, stmt_info);
+      set_vinfo_for_stmt (stmt, NULL);
+      STMT_VINFO_STMT (stmt_info) = new_stmt;
+      gsi_replace (gsi, new_stmt, true);
       return true;
     }
   else if (is_store)
@@ -2149,6 +2158,18 @@ vectorizable_mask_load_store (gimple stm
        }
     }
 
+  if (!is_store)
+    {
+      /* Ensure that even with -fno-tree-dce the scalar MASK_LOAD is removed
+        from the IL.  */
+      tree lhs = gimple_call_lhs (stmt);
+      new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
+      set_vinfo_for_stmt (new_stmt, stmt_info);
+      set_vinfo_for_stmt (stmt, NULL);
+      STMT_VINFO_STMT (stmt_info) = new_stmt;
+      gsi_replace (gsi, new_stmt, true);
+    }
+
   return true;
 }
 
--- gcc/testsuite/g++.dg/vect/pr60559.cc.jj     2014-03-18 14:04:55.173449250 
+0100
+++ gcc/testsuite/g++.dg/vect/pr60559.cc        2014-03-18 14:05:26.610273088 
+0100
@@ -0,0 +1,8 @@
+// PR tree-optimization/60559
+// { dg-do compile }
+// { dg-additional-options "-O3 -std=c++11 -fnon-call-exceptions 
-fno-tree-dce" }
+// { dg-additional-options "-mavx2" { target { i?86-*-* x86_64-*-* } } }
+
+#include "pr60023.cc"
+
+// { dg-final { cleanup-tree-dump "vect" } }

        Jakub

Reply via email to