PR c/6677 and al.

[Eric Botcazou]

> >  j = ((char) (i << 1)) / 2;
> [...]
> > The proposed fix is to disallow the second transformation when the type
> > is signed and the global operation not a MULT_EXPR.
>
> This isn't the best thing, since we _are_ allowed to produce the screwey
> result with "int" instead of char.

Do you mean that [int being 32-bit]

  int i = 0x7fffffff;
  int j;

  j = (i << 1) / 2;

is allowed to return j = 0x7fffffff ?

Note that gcc 2.91.66, gcc 2.95.3 and Borland C++ 5.5.1 return -1, while gcc
2.96 and gcc 3.1 return 0x7fffffff.

> And indeed, it would appear that your other problem
>
> >  j = ((char) (i * 2)) / 2;
>
> is the same root problem.  We can't look through the narrowing cast as
> readily as we're doing.

The problem is that we don't look through the narrowing cast in the fist
case: (i<<1):QI is really transformed into (i:QI) << (1:QI), which is valid
per se.

--
Eric Botcazou
ebotcazou@multimania.com