FIXNUM_OVERFLOW_P on amd64

FIXNUM_OVERFLOW_P on amd64

[Francesco Potorti`]

On gcc (GCC) 4.1.2 20061028 (prerelease) (Debian 4.1.1-19)

I get this:

gcc -c -D_BSD_SOURCE   -Demacs -DHAVE_CONFIG_H   -I. -I/home/pot/gnu/emacs-22.0.91/src -D_BSD_SOURCE  -g -O2 -Wno-pointer-sign  editfns.c
editfns.c: In function 'Fuser_uid':
editfns.c:1317: warning: comparison is always false due to limited range of data type
editfns.c:1317: warning: comparison is always false due to limited range of data type

The relevant line is:
  return make_fixnum_or_float (geteuid ());

which expands to:
  return (((long)(geteuid ()) > (((long) 1 << ((64 - 3) - 1)) - 1)
	   || (long) (geteuid ()) < - ((long) 1 << ((64 - 3) - 1)))
	  ? make_float (geteuid ())
	  : (((long) ((long)(geteuid ()))) << 3));

in lisp.h:
#define make_fixnum_or_float(val) \
   (FIXNUM_OVERFLOW_P (val) \
    ? make_float (val) \
    : make_number ((EMACS_INT)(val)))

#define FIXNUM_OVERFLOW_P(i) \
  ((EMACS_INT)(i) > MOST_POSITIVE_FIXNUM \
   || (EMACS_INT) (i) < MOST_NEGATIVE_FIXNUM)

#define MOST_POSITIVE_FIXNUM	(((EMACS_INT) 1 << (VALBITS - 1)) - 1)


So the problem apparently is that gcc realises that taking geteuid(),
stretching it to long and then comparing it with something bigger that
what geteuid() possibly can be is a no-op.  This is what is intended, in
fact.  To remove the warning, I tried to change the definition of
FIXNUM_OVERFLOW_P in lisp.h by adding a precondition like this:

#define FIXNUM_OVERFLOW_P(i) \
  (sizeof(i) >= sizeof(EMACS_INT) \
   && ((EMACS_INT)(i) > MOST_POSITIVE_FIXNUM \
       || (EMACS_INT)(i) < MOST_NEGATIVE_FIXNUM))

But nothing changed: the warning is still there.  While trying to find a
way to solve the problem, I realised that the make_fixnum_or_float is a
macro doing multiple evaluations of its argument, and editfns.c calls it
with a function call as an argument, which is generally not a good thing.

Re: FIXNUM_OVERFLOW_P on amd64

[Stefan Monnier]

> gcc -c -D_BSD_SOURCE   -Demacs -DHAVE_CONFIG_H   -I. -I/home/pot/gnu/emacs-22.0.91/src -D_BSD_SOURCE  -g -O2 -Wno-pointer-sign  editfns.c
> editfns.c: In function 'Fuser_uid':
> editfns.c:1317: warning: comparison is always false due to limited range of data type
> editfns.c:1317: warning: comparison is always false due to limited range of data type
[...]
> So the problem apparently is that gcc realises that taking geteuid(),
> stretching it to long and then comparing it with something bigger that
> what geteuid() possibly can be is a no-op.  This is what is intended, in
> fact.  To remove the warning, I tried to change the definition of
> FIXNUM_OVERFLOW_P in lisp.h by adding a precondition like this:

This warning is clearly wrong.  The same problem happens at various places
with SINGLE_BYTE_P which is sometimes called from a macro at spots where we
happen to always know that the char is < 256 (for example).

The problem is that such constant expressions are sometimes an indication of
a programming bug, but they can also be the result of very good coding
practices, where you want to let the compiler do the optimization.


        Stefan

Re: FIXNUM_OVERFLOW_P on amd64

[Francesco Potorti`]

>
>> gcc -c -D_BSD_SOURCE   -Demacs -DHAVE_CONFIG_H   -I. -I/home/pot/gnu/emacs-22.0.91/src -D_BSD_SOURCE  -g -O2 -Wno-pointer-sign  editfns.c
>> editfns.c: In function 'Fuser_uid':
>> editfns.c:1317: warning: comparison is always false due to limited range of data type
>> editfns.c:1317: warning: comparison is always false due to limited range of data type
>[...]
>> So the problem apparently is that gcc realises that taking geteuid(),
>> stretching it to long and then comparing it with something bigger that
>> what geteuid() possibly can be is a no-op.  This is what is intended, in
>> fact.  To remove the warning, I tried to change the definition of
>> FIXNUM_OVERFLOW_P in lisp.h by adding a precondition like this:
>
>This warning is clearly wrong.  The same problem happens at various places
>with SINGLE_BYTE_P which is sometimes called from a macro at spots where we
>happen to always know that the char is < 256 (for example).

I also have many others.  Most of them are related to the same macro in lisp.h.

>The problem is that such constant expressions are sometimes an indication of
>a programming bug, but they can also be the result of very good coding
>practices, where you want to let the compiler do the optimization.

Yes.  Can we do anything for preventing gcc from emitting those warnings?