Re: [gfortran] add SIGNAL and ALARM intrinsics

[FX Coudert <fxcoudert at gmail dot com>]

> +#ifdef HAVE_SIGNAL
> +  if (status != NULL)
> +    *status = (int) signal (*number, handler);
> +  else
> +    signal (*number, handler);
> +#else
>
> (twice in that file) can't be right.  signal returns a function pointer
> or SIG_ERR.  Casting a pointer to int is wrong.

Well, I'm just doing things the way g77 did. The g77 docs explicitly say:

---------------
Due to the side effects performed by this intrinsic, the function form 
is not recommended.

Warning: If the returned value is stored in an INTEGER(KIND=1) (default 
INTEGER) argument, truncation of the original return value occurs on 
some systems (such as Alphas, which have 64-bit pointers but 32-bit 
default integers), with no warning issued by g77 under normal circumstances.

Therefore, the following code fragment might silently fail on some systems:

     INTEGER RTN
     EXTERNAL MYHNDL
     RTN = SIGNAL(signum, MYHNDL)
     ...
     ! Restore original handler:
     RTN = SIGNAL(signum, RTN)

The reason for the failure is that `RTN' might not hold all the 
information on the original handler for the signal, thus restoring an 
invalid handler. This bug could manifest itself as a spurious run-time 
failure at an arbitrary point later during the program's execution, for 
example.
---------------

I think it's clear enough. Though, if people think it's better to change 
this behaviour, we could do it.

FX