Re: [PATCH] libstdc++: Check [ptr, end) and [ptr, ptr+n) ranges with _GLIBCXX_ASSERTIONS

[Jonathan Wakely <jwakely@redhat.com>]

On Thu, 14 Oct 2021 at 18:11, François Dumont <frs.dumont@gmail.com> wrote:
>
> Hi
>
>      On a related subject I am waiting for some feedback on:
>
> https://gcc.gnu.org/pipermail/libstdc++/2021-August/053005.html

I'm concerned that this adds too much overhead for the
_GLIBCXX_ASSERTIONS case. It adds function calls which are not
necessarily inlined, and which perform arithmetic and comparisons on
the arguments. That has a runtime cost which is non-zero.

The patches I sent in this thread have zero runtime cost, because they
use the compiler built-in which compiles away to nothing if the sizes
aren't known.

>
> On 11/10/21 6:49 pm, Jonathan Wakely wrote:
> > This enables lightweight checks for the __glibcxx_requires_valid_range
> > and __glibcxx_requires_string_len macros  when _GLIBCXX_ASSERTIONS is
> > defined.  By using __builtin_object_size we can check whether the end of
> > the range is part of the same object as the start of the range, and
> > detect problems like in PR 89927.
> >
> > libstdc++-v3/ChangeLog:
> >
> >       * include/debug/debug.h (__valid_range_p, __valid_range_n): New
> >       inline functions using __builtin_object_size to check ranges
> >       delimited by pointers.
> >       [_GLIBCXX_ASSERTIONS] (__glibcxx_requires_valid_range): Use
> >       __valid_range_p.
> >       [_GLIBCXX_ASSERTIONS] (__glibcxx_requires_string_len): Use
> >       __valid_range_n.
> >
> >
> > The first patch allows us to detect bugs like string("foo", "bar"),
> > like in PR 89927. Debug mode cannot currently detect this. The new
> > check uses the compiler built-in to detect when the two arguments are
> > not part of the same object. This assumes we're optimizing and the
> > compiler knows the values of the pointers. If it doesn't, then the
> > function just returns true and should inline to nothing.
>
> I see, it does not detect that input pointers are unrelated but as they
> are the computed size is >= __sz.
>
> Isn't it UB to compare unrelated pointers ?

Yes, and my patch doesn't compare any pointers, does it?