Re: RFC: should we use -Werror? (& sample patch to do it)

[Zack Weinberg <zack at codesourcery dot com>]

On Wed, Sep 05, 2001 at 01:59:43PM -0400, Kaveh R. Ghazi wrote:
>  > > Anyway, another problem is we have to get to zero warnings before
>  > > using -Werror.  I propose doing it one module at a time.  The patch
>  > > below is a candidate which does this for rtl.o to start with.  If that
>  > > lasts for a week without major pandemonium, we can do more files.
>  > 
>  > I think this is backward.  First we should get to zero warnings
>  > throughout the compiler, _then_ we talk about -Werror.  If only
>  > because if we do it your way, we'll be adding explicit compile rules
>  > to module after module.  I don't think there's enough marginal benefit
>  > from -Werror applied to a few modules, to eat that kind of complexity
>  > burden in the Makefile which is already too complicated.
> 
> The problem with getting to zero is each platform is likely to have
> its own special warnings stemming either from the config directory
> files or from platform specific wierdness. 

Yes.  When I said zero, I meant zero machine-independent warnings.  We
aren't even there yet, and I think it's simply because the remaining
MI warnings are such a pain to deal with - combine.c comparing mode
bitsizes to rtx integer slots, e.g.

Or there's the cp/error.c mess, which the only sane thing to do with
it is finish rewriting it in terms of diagnostic.c, and *that*'s a
lack-of-time issue; it's been on my list for about a year now.

> I thought starting with one file would ease people in slowly.  Once we
> got the whole compiler I would take out the individual rules and put
> it in the .c.o rule or into GCC_WARN_CFLAGS or something.  So whatever
> extra complexity added would be temporary.

See, that's nice in theory, but I bet you "temporary" is measured in
years.

(There are cleaner ways to do the same thing, e.g. look at the hack
used to avoid -pedantic for the non-C front ends.)

>  > Next biggest offender is
>  > 
>  >        7 string length `???' is greater than the length `???' ISO C89
>  >          compilers are required to support
>  > 
>  > These are more or less intractable: they come from big hairy spec
>  > strings, e.g.
> 
> IIRC, these string warnings are from -pedantic, which isn't affected
> by -Werror.  To make those hard errors, I think you have to use
> -pedantic-errors.  (Again, IIRC.)

Not true.  -Werror turns _all_ warnings into errors.

zw