Hm, trying this with my native GCC compiler it does behave as you claim. I
think I may have misdiagnosed the issue I was running into. I apologize for
taking up your time.

On Thu, Jun 28, 2018 at 10:41 AM Rich Felker <dalias@libc.org> wrote:

> On Thu, Jun 28, 2018 at 10:20:28AM -0700, Mark Winterrowd wrote:
> > Hi all,
> >
> > I believe I have found an out of bounds memory read in vfprintf.c
> >
> > On line 509 in src/stdio/vfprintf.c in the current source tree head, you
> > can observe the following snippet of code:
> >
> > /* Format specifier state machine */
> > st=0;
> > do {
> > if (OOB(*s)) goto inval;
> > ps=st;
> > st=states[st]S(*s++);
> > } while (st-1<STOP);
> > if (!st) goto inval;
> >
> > Note that on line 99 the OOB macro expands to the following test whether
> > the argument falls outside of 'A' and 'z', written to use a single
> compare:
> >
> > #define OOB(x) ((unsigned)(x)-'A' > 'z'-'A')
> > Unfortunately, the cast to unsigned binds tighter than the subtract
>
> For this idiom, it's intentional that it bind higher. Here since x is
> small (char-range) anyway it doesn't matter, but in general the
> pattern (x-'A') could overflow, producing UB, if x weren't alreaady
> unsigned.
>
> > from 'A', so if x is less than 'A',
> > OOB will return false. This is common in the case of space, which has
> > an ascii value of 32
>
> No, the result of (unsigned)(x)-'A' is unsigned, and in the case
> x<'A', it's a value larger than INT_MAX which is much larger than
> 'z'-'A'.
>
> > compared to 'A' 's value of 65.
> >
> > This causes us to index into states with a negative value for its
> > second dimension, causing us to
> > index to an unpredictable location in states, possibly even off the
> beginning.
>
> Did you test this? It's possible there's another mistake we're not
> seeing, but the above isn't one. Also note that passing an invalid
> format string is UB already, so any graceful handling of that is just
> hardening, not correctness.
>
> Rich
>