Hi! Related to the PR69207 changes, the problem there has been that fold_convertible_p allowed conversion from a vector type to same sized integral type, and fold_convert used NOP_EXPR in that case. But, such an operation is much better handled using VCE.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2016-01-11 Jakub Jelinek <ja...@redhat.com> * fold-const.c (fold_convertible_p): Don't return true for conversion of VECTOR_TYPE to same sized integral type. (fold_convert_loc): Fix up formatting. Fold conversion of VECTOR_TYPE to same sized integral type using VIEW_CONVERT_EXPR instead of NOP_EXPR. --- gcc/fold-const.c.jj 2016-01-09 08:36:15.000000000 +0100 +++ gcc/fold-const.c 2016-01-11 11:53:41.470765831 +0100 @@ -2182,11 +2182,8 @@ fold_convertible_p (const_tree type, con case INTEGER_TYPE: case ENUMERAL_TYPE: case BOOLEAN_TYPE: case POINTER_TYPE: case REFERENCE_TYPE: case OFFSET_TYPE: - if (INTEGRAL_TYPE_P (orig) || POINTER_TYPE_P (orig) - || TREE_CODE (orig) == OFFSET_TYPE) - return true; - return (TREE_CODE (orig) == VECTOR_TYPE - && tree_int_cst_equal (TYPE_SIZE (type), TYPE_SIZE (orig))); + return (INTEGRAL_TYPE_P (orig) || POINTER_TYPE_P (orig) + || TREE_CODE (orig) == OFFSET_TYPE); case REAL_TYPE: case FIXED_POINT_TYPE: @@ -2241,11 +2238,11 @@ fold_convert_loc (location_t loc, tree t return fold_build1_loc (loc, NOP_EXPR, type, arg); if (TREE_CODE (orig) == COMPLEX_TYPE) return fold_convert_loc (loc, type, - fold_build1_loc (loc, REALPART_EXPR, - TREE_TYPE (orig), arg)); + fold_build1_loc (loc, REALPART_EXPR, + TREE_TYPE (orig), arg)); gcc_assert (TREE_CODE (orig) == VECTOR_TYPE && tree_int_cst_equal (TYPE_SIZE (type), TYPE_SIZE (orig))); - return fold_build1_loc (loc, NOP_EXPR, type, arg); + return fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, arg); case REAL_TYPE: if (TREE_CODE (arg) == INTEGER_CST) Jakub