Re: FreeSec crypt()

[Rich Felker <dalias@aerifal.cx>]

On Wed, Jun 13, 2012 at 08:45:46PM +0400, Solar Designer wrote:
> On Wed, Jun 13, 2012 at 10:56:03AM -0400, Rich Felker wrote:
> > On Wed, Jun 13, 2012 at 05:18:07PM +0400, Solar Designer wrote:
> > > > Note that even if the behavior were defined, this code seems to have
> > > > different behavior for high bytes depending on the signedness of char.
> ....
> > > Why would signedness of char matter
> > > if the behavior of the signed char overflowing left shift were defined?
> > 
> > Well if char is signed, (char)0x80 << 1 is -256. If char is unsigned,
> > (char)0x80 << 1 is 256.
> 
> Sure, but we had:
> 
> 	const char *key;
> 	u_char *q;
> 	*q++ = *key << 1;
> 
> so while *key << 1 is either -256 or 256 (promoted to int or unsigned
> int), those high bits get dropped on the assignment to *q anyway,
> resulting in the same value there either way.  No?

You're right on that. Ideally the functions should just take arguments
of type unsigned char *, and the crypt/crypt_r wrapper should cast the
original char * to unsigned char *. This is the same way all the
standard string functions (like strcmp) are required to work.

Casting/converting the _value_ after reading it also happens to work,
and is sufficient for musl's purposes (we assume, per POSIX, that
CHAR_BIT is 8, but also that signed types are twos complement), but
only reinterpreting by casting the pointer before reading it is 100%
portable. On a non-twos-complement machine, reading a signed char is
lossy (it can only obtain 255 possible values, not 256).

Rich