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

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

[Falk Hueffner]

Hi,

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.

So I was wondering about a general warning about negating unsigned
values, since I couldn't really think of a legitimate application. A
quick check with the gcc source turned up:

gengtype-lex.l:
  char *namestart;
  size_t namelen;
  [...]
  for (namelen = 1; !ISSPACE (namestart[-namelen]); namelen++)

This looks actually invalid to me, although it will probably work
everywhere.

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.

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

Then there's constant folding in neg_double. Hm. Damn. I can't think
of any reformulation which does not obscure the code. So this warning
should probably not be turned on by -W. But it seems generally useful.

Any opinions?

-- 
	Falk

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

[Joe Buck]

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.
 
> So I was wondering about a general warning about negating unsigned
> values, since I couldn't really think of a legitimate application. 

There are legitimate applications, and I've used them in my code.

> quick check with the gcc source turned up:
> 
> gengtype-lex.l:
>   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.
 
> 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.
 
> 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.  Just the fact that your proposed warning will
turn on at least four complaints against correct usage in gcc shows
that it is not a good idea.

Re: 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

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

[Jamie Lokier]

Falk Hueffner wrote:
> > 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).

Good point.  That's elusive enough that I saw ptr[-x] and thought it
was was fine; I am now humbly corrected.

-- Jamie

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

[Joe Buck]

On Mon, Oct 06, 2003 at 06:11:05PM +0200, Falk Hueffner wrote:
> Joe Buck <jbuck@synopsys.com> writes:
> > 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).

Good catch.  It seems that the real problem is not negation of unsigned
values, but negation followed by extension to a potentially larger word size.
If you can figure out how to warn specifically about that, it might
be worth adding.  But it would be tricky.

 

Re: 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 06:11:05PM +0200, Falk Hueffner wrote:
> > Joe Buck <jbuck@synopsys.com> writes:
> > > 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).
> 
> Good catch.  It seems that the real problem is not negation of
> unsigned values, but negation followed by extension to a potentially
> larger word size.  If you can figure out how to warn specifically
> about that, it might be worth adding.  But it would be tricky.

I'll give it a try, but I'm not sure how to do that right now, warning
about it every time was a lot easier :)

-- 
	Falk