Suggested warning: "negating an expression of unsigned type does not yield a negative value"

[Falk Hueffner]

Joe Buck <jbuck@synopsys.com> writes:

> On Mon, Oct 06, 2003 at 02:46:43PM +0200, Falk Hueffner wrote:
> > I just found yet another bug of the kind:
> > 
> > int f (int *p, unsigned x) { return p[-x]; }
> > 
> > which only manifests on 64 bit platforms, because most (all?)
> > platforms have wrapping address arithmetic.
> 
> The C and C++ standards require that unsigned values obey modulo 2**N
> arithmetic, so the value of -x is rigorously defined.

Sure it is. But it is not what is intended. Example: x = 5, then
-x=4294967291, i.e., p will be advanced by 4294967291 bytes, which is
way beyond the legal range of p, but happens to work anyway on 32 bit
architectures (but not on 64 bit architectures).

> >   char *namestart;
> >   size_t namelen;
> >   [...]
> >   for (namelen = 1; !ISSPACE (namestart[-namelen]); namelen++)
> > 
> > This looks actually invalid to me, although it will probably work
> > everywhere.
> 
> It's valid everywhere.

I'm pretty sure it's not. -namelen is, again, something like
4294967291 (or 18446744073709551611), which is not a legal array
index.

> > In fold_const.c, there's
> > 
> > case RSHIFT_EXPR:
> >       int2l = -int2l;
> > 
> > also "invalid but works" since it's later passed to a function taking
> > int.
> 
> Again, this is valid everywhere.

No, this produces an unsigned value which cannot be represented in a
signed value of same width, but is converted to signed, which is
undefined according to the standard.

> > Then there's everybody's favourite idiom "x &= -x", but it can be
> > expressed clearer as "x &= ~x + 1".
> 
> Again, it's fine as is.

I agree with that.

-- 
	Falk