https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88007
--- Comment #2 from rguenther at suse dot de <rguenther at suse dot de> --- On Tue, 13 Nov 2018, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88007 > > Jakub Jelinek <jakub at gcc dot gnu.org> changed: > > What |Removed |Added > ---------------------------------------------------------------------------- > CC| |jakub at gcc dot gnu.org > > --- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > This is because parse_optimize_options has: > r = q = (char *) ggc_alloc_atomic (len2 + 3); > and constructs in there the whole option, like "-falign-functions=16" , and > then options processing comes and points arg into that string (so "16"). > As gcc_options is marked GTY, the const char * fields in there thus need to be > either not GC memory at all, or they need to point at the start of a GC chunk, > or into the middle of STRING_CST, everything else is disallowed. > > So, either we need to pass down to the option processing info that the options > are GC allocated and if we'd have arg pointing into the middle of such a > string, ggc_strdup it, or maybe c-common.c should use XOBNEWVEC > (&opts_obstack, > char, len2 + 3); instead of (char *) ggc_alloc_atomic (len2 + 3) ? Using the opts_obstack sounds reasonable to me.
