Hi! The expansion relies on lhs being kept for calls that return addressable types. On the following testcase (which is a GNU extension, pedantically we error out on zero sized arrays) we return TREE_ADDRESSABLE zero_sized_type and optimize away the lhs which we need later on.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, during the bootstrap+regtests it only made a difference on the newly added testcase. Ok for trunk? 2017-09-27 Jakub Jelinek <ja...@redhat.com> PR c++/82159 * gimplify.c (gimplify_modify_expr): Don't optimize away zero sized lhs from calls if the lhs has addressable type. * g++.dg/opt/pr82159.C: New test. --- gcc/gimplify.c.jj 2017-09-01 09:25:34.000000000 +0200 +++ gcc/gimplify.c 2017-09-26 13:03:11.614726601 +0200 @@ -5479,7 +5479,12 @@ gimplify_modify_expr (tree *expr_p, gimp side as statements and throw away the assignment. Do this after gimplify_modify_expr_rhs so we handle TARGET_EXPRs of addressable types properly. */ - if (zero_sized_type (TREE_TYPE (*from_p)) && !want_value) + if (zero_sized_type (TREE_TYPE (*from_p)) + && !want_value + /* Don't do this for calls that return addressable types, expand_call + relies on those having a lhs. */ + && !(TREE_ADDRESSABLE (TREE_TYPE (*from_p)) + && TREE_CODE (*from_p) == CALL_EXPR)) { gimplify_stmt (from_p, pre_p); gimplify_stmt (to_p, pre_p); --- gcc/testsuite/g++.dg/opt/pr82159.C.jj 2017-09-26 13:04:08.711027279 +0200 +++ gcc/testsuite/g++.dg/opt/pr82159.C 2017-09-26 14:20:01.519361945 +0200 @@ -0,0 +1,18 @@ +// PR c++/82159 +// { dg-do compile } +// { dg-options "" } + +template<int N> +struct S +{ + ~S () {} + template<int M> S<M> foo () { return S<M> (); } + unsigned char data[N]; +}; + +int +main () +{ + S<16> d; + S<0> t = d.foo<0> (); +} Jakub