https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103168
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> --- Created attachment 51847 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51847&action=edit patch Looks like I have to exclude summary->global_memory_read since otherwise a trivial testcase like the following is miscompiled (the access tree does not have the access for 'p'). unsigned p; unsigned __attribute__((noinline)) test (int) { return p; } int main() { p = 1; if (test (0) != 1) __builtin_abort (); p = 2; if (test (0) != 2) __builtin_abort (); return 0; }