bug#24630: guile-2.0.12: Comparison is always true warning - may cause problems.

[Andy Wingo <wingo@pobox.com>]

Hi,

On Thu 06 Oct 2016 22:49, Jens Bauer <jens-guile-bug@plustv.dk> writes:

> I get the following warnings, when building on Mac OS X.
> (It should show up for all platforms, though):
>
> In file included from /Users/jens/open-source/Source/guile-2.0.12/libguile/numbers.c:9731:
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c: In function 'scm_to_int8':
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c:94: warning: comparison is always true due to limited range of data type
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c:94: warning: comparison is always true due to limited range of data type

These are not really bugs.  I mean, we shouldn't produce warnings, but
GCC doesn't warn on these, so clearly there is a heuristic which clang
has set differently; but the actual code is fine.

In your investigations below there are some errors.  I include a couple
of inline comments for your enjoyment.

> In file included from /Users/jens/open-source/Source/guile-2.0.12/libguile/numbers.c:9747:
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c: In function 'scm_to_int16':
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c:94: warning: comparison is always true due to limited range of data type
> /Users/jens/open-source/Source/guile-2.0.12/libguile/conv-integer.i.c:94: warning: comparison is always true due to limited range of data type
>
> Notice that it's only from line 94, which reads...
> 	  if (n >= TYPE_MIN && n <= TYPE_MAX)
>
> ... looking at the top of the file, it says: "It is only for signed types", so I look in ...
>
>   numbers.c:9731
>   numbers.c:9747
> ... which is int8 and int16 (signed integers); this should be as intended.
>
> The variable 'n' is declared as scm_t_signed_bits, which is a scm_t_intptr, which is an intptr_t, which is just a 'long'.
>
> So my guess is that the problem must be with TYPE_MIN and TYPE_MAX.
>
> In numbers.c, line 9742, they're defined as follows:
> #define TYPE                     scm_t_int16
> #define TYPE_MIN                 SCM_T_INT16_MIN
> #define TYPE_MAX                 SCM_T_INT16_MAX
>
> ... looks good to me, but where's the definition of SCM_T_INT16_MIN and SCM_T_INT16_MAX ?
> -It seems to be in __scm.h:
>
> #define SCM_I_UTYPE_MAX(type)      ((type)-1)
> #define SCM_I_TYPE_MAX(type,umax)  ((type)((umax)/2))
> #define SCM_I_TYPE_MIN(type,umax)  (-((type)((umax)/2))-1)
>
> #define SCM_T_UINT8_MAX   SCM_I_UTYPE_MAX(scm_t_uint8)
> #define SCM_T_INT8_MIN    SCM_I_TYPE_MIN(scm_t_int8,SCM_T_UINT8_MAX)
> #define SCM_T_INT8_MAX    SCM_I_TYPE_MAX(scm_t_int8,SCM_T_UINT8_MAX)
>
> #define SCM_T_UINT16_MAX  SCM_I_UTYPE_MAX(scm_t_uint16)
> #define SCM_T_INT16_MIN   SCM_I_TYPE_MIN(scm_t_int16,SCM_T_UINT16_MAX)
> #define SCM_T_INT16_MAX   SCM_I_TYPE_MAX(scm_t_int16,SCM_T_UINT16_MAX)
>
> Now, this is where things get interesting. The macros are cool, but I think the use seems to be incorrect.
>
> Let's try an example (SCM_T_INT16_MIN):
> SCM_T_INT16_MIN = SCM_I_TYPE_MIN(scm_t_int16,SCM_T_UINT16_MAX)
> Expands to ...
> SCM_T_INT16_MIN = (-((scm_t_int16)((-1)/2))-1)

SCM_T_UINT16_MAX expands to ((scm_t_uint16)-1) which expands to the
uint16_t value 0xffff.  (These intermediate expansions have type in
addition to value.)  SCM_T_INT16_MIN is -(0xffff/2)-1, which is
(int16_t)-0x8000.

> ... which can be cleaned up ...
>
> SCM_T_INT16_MIN = (-(((-1)/2))-1)
>
> A signed integer of value -1 divided by 2, is the same as bitshifting to the right by using ASR; the result will be -1.
>
> SCM_T_INT16_MIN = (-(((-1)))-1)
> SCM_T_INT16_MIN = (-((-1))-1)
> SCM_T_INT16_MIN = (-(-1)-1)
> SCM_T_INT16_MIN = (+1-1)
> SCM_T_INT16_MIN = (0)
>
> ... Ehm ... Did I do something wrong ?
> I expected the value -32768, but got 0.
>
> Wouldn't it be correct to typecast as scm_t_uint16 instead of scm_t_int16 (and thus scm_t_uint8 instead of scm_t_int8) ?
>

Happy hacking,

Andy