Hi! If va_list is one-entry array of structs, those RECORD_TYPEs don't have TYPE_BINFO.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? I've made the testcase compile instead of run, because while it with the patch works both on x86_64-linux and i686-linux, not sure if it really should be well defined - at least in C one can copy va_list only through va_copy, by throwing it and catching it is copied in some other way. 2016-08-08 Jakub Jelinek <ja...@redhat.com> PR c++/72809 * rtti.c (get_pseudo_ti_index): Return TK_CLASS_TYPE for builtin aggregate types without TYPE_BINFO. * g++.dg/eh/stdarg1.C: New test. --- gcc/cp/rtti.c.jj 2016-04-22 18:21:27.000000000 +0200 +++ gcc/cp/rtti.c 2016-08-08 16:21:32.937053462 +0200 @@ -1293,7 +1293,8 @@ get_pseudo_ti_index (tree type) ix = TK_CLASS_TYPE; break; } - else if (!BINFO_N_BASE_BINFOS (TYPE_BINFO (type))) + else if (!TYPE_BINFO (type) + || !BINFO_N_BASE_BINFOS (TYPE_BINFO (type))) { ix = TK_CLASS_TYPE; break; --- gcc/testsuite/g++.dg/eh/stdarg1.C.jj 2016-08-08 16:31:52.553510463 +0200 +++ gcc/testsuite/g++.dg/eh/stdarg1.C 2016-08-08 16:30:20.000000000 +0200 @@ -0,0 +1,30 @@ +// PR c++/72809 +// { dg-do compile } + +#include <stdarg.h> + +int +foo (int a, ...) +{ + va_list ap; + int r = 0; + va_start (ap, a); + try + { + if (a == 1) + throw (ap); + } + catch (va_list b) + { + r = va_arg (b, int); + } + va_end (ap); + return r; +} + +int +main () +{ + if (foo (0) != 0 || foo (1, 7) != 7) + __builtin_abort (); +} Jakub