[PATCH] PR18327: -Wsystem-headers introduces build errors

[Richard Smith]

On Thu, Jan 9, 2014 at 2:39 PM, Marshall Clow <mclow.lists at gmail.com> wrote:

> On Jan 9, 2014, at 2:20 PM, Richard Smith <richard at metafoo.co.uk> wrote:
>
> On Thu, Jan 9, 2014 at 7:25 AM, Alp Toker <alp at nuanti.com> wrote:
>
>>
>
> It's not fair on the user or the system header maintainers to bring up
>> unexpected errors with this flag.
>>
>
> I don't think this is clear. If the code in the system header is
> ill-formed, I would expect the system header maintainers to be *grateful*
> that we flag it as an error and not just as a warning. But I'll wait to
> hear what they have to say about that.
>
>
> A few points, in no particular order:
>
> * If we have ill-formed code in system headers, I would expect it to fail
> to compile whether the user specifies -Wsystem-headers or not.
>

And yet in some cases we don't, because we deliberately support some
flavors of broken system headers. Given that, would you want
-Wsystem-headers to make such code produce an error or only a warning?

* If a user has a clean build, and then rebuilds with -Wsystem-headers, I
> would expect to get warnings - not errors. [ Isn’t that what PR18327 is
> all about? ]
>

Do you expect this because the flag starts with -W? (Maybe that's the
problem here -- it's *not* a warning flag in the sense that the other -W
flags are. But that's not unprecedented -- nor is -Werror.)

* There are some “interesting” language features which are only enabled for
> system headers, and cause warnings if used in user code.
> [ User-defined suffixes that do not start with an underscore, for example.
> ]
>

This is a really great example, thanks. So it seems there are at least
three different classes of errors that we might think about producing in
system headers:

 1) warnings that the user has turned into errors with -Werror or
-Werror=foo
 2) errors for ill-formed code that we suppress by default in system
headers as a workaround for a system header bug
 3) errors for code that is ill-formed outside system headers but valid
within system headers (using reserved names, adding names to namespace std,
that sort of thing)

With -Wsystem-headers enabled, I think (1) should be an error, (3) should
remain suppressed (not even a warning), and (2) should be either a warning
or an error (and your first two bullets don't give me a clear idea of which
way these cases should go).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140109/15da7cef/attachment.html>