https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103534
--- Comment #4 from Martin Sebor <msebor at gcc dot gnu.org> --- Yes, the warning does disappear when malloc() and free() are used instead of operator new and delete. foo() also ends up much better optimized, even at -O1: __attribute__((abi_tag ("cxx11"))) struct string foo () { struct string & _7(D); char * _69; <bb 2> [local count: 1073741824]: _69 = __builtin_malloc (17); __builtin_memcpy (_69, "1234567890123456", 16); MEM[(char_type &)_69 + 16] = 0; MEM[(struct basic_string *)_7(D)]._M_dataplus._M_p = _69; MEM[(struct basic_string *)_7(D)].D.33183._M_allocated_capacity = 16; MEM[(struct basic_string *)_7(D)]._M_string_length = 16; return _7(D); } The -O1 dump in comment #1 doesn't look right, I may have messed something up. The same issue happens at -O2 where the dump is as follows: =========== BB 3 ============ Imports: n_5(D) Exports: _1 n_5(D) _1 : n_5(D)(I) n_5(D) int [-INF, -1][1, +INF] <bb 3> [local count: 536870913]: _1 = (sizetype) n_5(D); if (_1 == 1) goto <bb 4>; [51.12%] else goto <bb 5>; [48.88%] _1 : sizetype [1, 2147483647][18446744071562067968, +INF] 3->4 (T) _1 : sizetype [1, 1] 3->4 (T) n_5(D) : int [1, 1] 3->5 (F) _1 : sizetype [2, 2147483647][18446744071562067968, +INF] 3->5 (F) n_5(D) : int [-INF, -1][2, +INF] =========== BB 4 ============ <bb 4> [local count: 274448412]: MEM[(char *)&b] = 0; goto <bb 6>; [100.00%] =========== BB 5 ============ <bb 5> [local count: 262422500]: __builtin_memcpy (&b, &a, _1);