> >> Interesting. I'd never thought about the generation/use angle to prove > >> a pointer was non-null. ISTM we could use that same logic to infer that > >> more pointers are non-null in extract_range_from_binary_expr. > >> > >> Interested in tackling that improvement, obviously as an independent patch? > >> > > > > I'm not familiar with vrp code, but.. something like this? > > > > Index: tree-vrp.c > > =================================================================== > > --- tree-vrp.c (revision 173703) > > +++ tree-vrp.c (working copy) > > @@ -2273,7 +2273,12 @@ extract_range_from_binary_expr (value_ra > > { > > /* For pointer types, we are really only interested in asserting > > whether the expression evaluates to non-NULL. */ > > - if (range_is_nonnull (&vr0) || range_is_nonnull (&vr1)) > > + if (flag_delete_null_pointer_checks && nowrap_type_p (expr_type)) > > the latter would always return true > > Btw, I guess you'll "miscompile" a load of code that is strictly > undefined. So I'm not sure we want to do this against our users ...
Probably not, at least unless the user explicitly asks for it -- for example, we could have -fdelete-null-pointer-checks=2. In fact, it might be a good idea to implement this flag anyway, since some current uses of flag_delete_null_pointer_checks can lead to "miscompilations" when user makes an error in their code and would probably appreciate more having their program crash. > Oh, and of course it's even wrong. I thing it needs && > !range_includes_zero (&vr1) (which we probably don't have). The > offset may be 0 and NULL + 0 > is still NULL. actually, the result of NULL + 0 is undefined (pointer arithmetics is only defined for pointers to actual objects, and NULL cannot point to one). Zdenek