Re: [boost] Operating System API Error Reporting Guidelines

From: Dave Harris (brangdon_at_[hidden])
Date: 2006-02-21 10:07:20

In-Reply-To: <dta3ga$o0b$1_at_[hidden]>
bdawes_at_[hidden] (Beman Dawes) wrote (abridged):
> Fresh draft up at www.esva.net/~beman/error_reporting_guidelines.html.

At first sight this involves writing functions twice, one version that 
throws system_error and another version that returns 
system_error_code_type. Is that burdensome in practice?

Some code examples might help. Is this the kind of thing you have in mind?

    int my_close( int fd ) {
        system_error_code_type error;
        int result = my_close( fd, error );
        if (error != 0)
            throw_exception( system_error( "close", error ) );
        return result;
    }

Looking at this, it seems to me it might be better if the caller sets 
system_error_code_type to 0, rather than the callee. It's not nice for the 
caller to have an uninitialised variable, and there's no need for them 
both to do it.

Is 0 a reasonable "non-error" value for all platforms? It is tempting to 
use a more elaborate type here, eg one that adds a bool to flag the error 
case. Is the code already so unportable that this isn't an issue?

Another reason for using a user-defined type for system_error_code_type is 
better type safety. I am slightly bothered that, as it stands, we are 
overloading based on an int argument. Isn't that error-prone? It means we 
can't have function sets like:
     int func( int );
     int func( int, int );
     int func( int, system_error_code_type & );
     int func( int, int, system_error_code_type & );

even if that is what the user might naively expect. 

Do you have a scheme for generating the throwing function automatically, 
via macros or templates? I'm not advocating such a scheme. I am wondering 
what your thoughts are.

-- Dave Harris, Nottingham, UK.