On 03/04/2016 01:37 PM, Jakub Jelinek wrote:
Hi!As mentioned in the PR, on the following testcase we emit invalid error on s390* in the va_arg ((ap), int) case, va_arg (ap, int) is fine (and the former is fine in -std=c++11 or -std=c++98 too). The bug only triggers in constructors (or destructors), and happens because build_va_arg creates ADDR_EXPR that has slightly different type, and has INDIRECT_REF inside of it. copy_tree_body_r notices this and cancels the two, but unlike e.g. build_fold_addr_expr_with_type_loc if (TREE_CODE (t) == INDIRECT_REF) { t = TREE_OPERAND (t, 0); if (TREE_TYPE (t) != ptrtype) t = build1_loc (loc, NOP_EXPR, ptrtype, t); <<<<<<---------- this } it doesn't attempt to handle the case where ADDR_EXPR contained different type from the type of INDIRECT_REF's operand. Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, bootstrap/regtest on s390{,x}-linux pending. Ok for trunk if it passes even there? 2016-03-04 Jakub Jelinek <ja...@redhat.com> PR c++/70084 * tree-inline.c (copy_tree_body_r): When cancelling ADDR_EXPR of INDIRECT_REF and ADDR_EXPR changed type, fold_convert it to the right type. * g++.dg/expr/stdarg3.C: New test.
OK. jeff
