Re: deprecated symbol warnings

[Ken Raeburn]

On May 14, 2005, at 08:40, Neil Jerram wrote:
> This looks very useful.  I'm not an expert in this kind of thing, but 
> here are some comments.

Thanks...

> > #elif defined _WIN32
>
> Does the __declspec syntax work for all Windows compilers?  If it's 
> actually specific to MSVC (which is the only compiler I'm familiar 
> with), you should use _MSC_VER.  If it's for all Windows compilers, 
> I'm surprised by the underscore - i.e. I'd normally use WIN32; are 
> WIN32 and _WIN32 equivalent?

I have no idea.  I do a small amount of work with MSVC, and the other 
compiler I'd be concerned about would be GCC, which is checked for 
first (though, actually, only for gcc 3 and later, should fix that); I 
have no idea what other Windows compilers would do, though I would 
expect for maximum interoperability they would implement or at least 
ignore the __declspec annotations.  In the code I deal with at work, we 
generally check _WIN32.

I'm hoping some Windows developer can help out here... we've got some 
at work, maybe I can ping them about it.

> > /* N.B.: Application code will sometimes test whether one of these
> >     macros is defined, to figure out if the version of Guile in use
> >     predates the creation of the macro.  So at deprecation time, we
> >     still want the macro to be visible.  ANSI C says "#define foo foo"
> >     is okay, but if we get complaints about it, try switching the
> >     non-macro name to "foo_" or "foo_deprecated" or something; make it
> >     a simple concatenation so that we can make the other macros
> >     continue to be simple.  */
>
> I think we should assume in advance that we'll hit trouble with this 
> on some platforms.  Otherwise it's just another hiccup to push people 
> away from trying Guile out.

*sigh*  I was afraid of that.  So when do we start requiring "real" 
ANSI C support? :-(

Doing this means the compile-time messages will give the wrong symbol 
names.  They'll be close to the names used by the application, but not 
the same.  Still, getting messages that are close is probably better 
than explaining to people why this strange use of the preprocessor is 
actually valid and it's their compiler that's broken.  I wonder if it's 
really a problem these days, though, a decade and a half after the spec 
was published....

> Nice.  Slight extra overhead, but well worth it IMO for the extra 
> function.

Yeah, consuming a bit more space in the library is slightly annoying, 
but the run-time performance for deprecated functions probably isn't 
anything to worry about.

> > The link-time warnings are a bit messier.
>
> What is the extra benefit of link-time warnings over compile-time?  
> Are there any cases where the user will see a link-time warning 
> without a corresponding compile-time one?

If the application refers to Guile functions without including 
libguile.h, yes.

Or if an older version of gcc is used that can generate ELF sections 
but doesn't support the "deprecated" attribute.  (I *think* "section" 
came first.)

Or if we add support for another compiler that lets us use 'asm' to 
emit warning sections or symbols or whatever the tool chain supports, 
but doesn't have a compile-time equivalent to the "deprecated" 
attribute, but we could treat that case separately.

Or...

>   Would link-time warnings show up when linking against a 3rd party 
> library that chose to use deprecated symbols?

Yes.  Hmm, at least for a static library.  I'm not sure about a shared 
library, if it would happen when the library was built, or used, or 
both.  I would assume they would be printed when the library is built, 
at least.  I'm guessing that this would probably be the most 
interesting of the cases.

Or, in the real world, if all the non-fatal compile-time warnings have 
all scrolled off the screen and the trailing bits of the "make" output 
still visible, and the only bits likely to get looked at unless 
something breaks, include only the final link step.


If it's not that interesting after all, I could drop it (as I 
indicated, it's the more complicated approach, though not terribly so) 
and just go with the compile-time warnings based on the predefined 
macros.  Turning macros into variables and/or functions is needed even 
without link-time warnings, as GCC can't attach attributes to the 
macros, only to the variable and function symbols, so it doesn't 
simplify things all that much.

> > scm_c_issue_deprecation_warning (though I don't know if I need to 
> > worry about i18n interactions there).
>
> I don't believe it's yet been suggested that we would translate the 
> Scheme names of primitives, so I doubt there would be a problem here. 
> (Out of interest, do any other scripting languages do this?)

I was thinking of the "is deprecated, please use" part.  The symbol 
names we should display as they are, but possibly the recommendation to 
change the application code should be given in Swahili or Klingon or 
whatever.  That we could do at run time, and gcc could do it at compile 
time, but I don't think we can do anything about it at link time.

Ken