[Bug sanitizer/82046] [7/8 Regression] Bogus -fsanitize=undefined error with -O2 -Wall

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82046

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I suppose fancy threading gets in the way...

  <bb 10> [84.09%]:
  # p_4 = PHI <p_55(6), p_14(9)>
  if (p_4 == 0B)
    goto <bb 11>; [0.04%]
  else
    goto <bb 20>; [99.96%]

  <bb 11> [0.03%]:
  __builtin___ubsan_handle_nonnull_arg (&*.Lubsan_data0);
  __builtin_sprintf (0B, "%lu", last_count_13);
  __builtin___ubsan_handle_nonnull_arg (&*.Lubsan_data2);

  <bb 12> [84.09%]:
  _1 = __builtin_strlen (p_4);

threaded the two nonnull checks on p for sprintf and strlen.

Not sure what we can do about these cases (besides forcing no-warning
on all stmts in duplicated blocks by threading).  Less stupid
sanitizing maybe, or earlier simplifying.  Or not exposing the
NULL checks early but combine them with the sanitizer builtins.

__builtin___ubsan_handle_nunull_arg (p_4, &*.Lubsan_data0);

etc.  this way we can even "CSE" dominating ones (I suppose getting
the redundant ubsan warning on strlen isn't important).  We'd then
lower the control flow at sanopt time for example or in an extra pass.