Use SCALAR_TYPE_MODE in vect_create_epilog_for_reduction

This follows on from similar changes a couple of months ago and
is needed when general modes have variable size.

2017-10-23  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* tree-vect-loop.c (vect_create_epilog_for_reduction): Use
	SCALAR_TYPE_MODE instead of TYPE_MODE.

From-SVN: r254002

gcc/ChangeLog

 2017-10-23  Richard Sandiford  <richard.sandiford@linaro.org>
+
+	* tree-vect-loop.c (vect_create_epilog_for_reduction): Use
+	SCALAR_TYPE_MODE instead of TYPE_MODE.
+
+2017-10-23  Richard Sandiford  <richard.sandiford@linaro.org>
 	    Alan Hayward  <alan.hayward@arm.com>
 	    David Sherwood  <david.sherwood@arm.com>
 

gcc/tree-vect-loop.c

@@ -4487,7 +4487,7 @@ vect_create_epilog_for_reduction (vec<tree> vect_defs, gimple *stmt,
       gcc_assert (gimple_assign_rhs_code (vec_stmt) == VEC_COND_EXPR);
 
       int scalar_precision
-	= GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (vectype)));
+	= GET_MODE_PRECISION (SCALAR_TYPE_MODE (TREE_TYPE (vectype)));
       tree cr_index_scalar_type = make_unsigned_type (scalar_precision);
       tree cr_index_vector_type = build_vector_type
 	(cr_index_scalar_type, TYPE_VECTOR_SUBPARTS (vectype));