On 12/17/2015 02:01 PM, Corinna Vinschen wrote:

>> Here's what happens:
>>
>> One of the Gnulib modules includes sys/types.h, which includes sys/select.h
>> because of the recent changes.  This brings in Gnulib's sys/select.h, which
>> includes signal.h.  We then get the errors I posted because we haven't yet
>> finished including sys/types.h.

Gnulib has been taught to work around early inclusion problems before;
sounds like this will be another case where gnulib has to make sure the
system header is complete before its own replacements kick in.

>>
>> All the build errors disappear if I remove '#include <sys/select.h>' from
>> sys/types.h.  You said above that the macros related to select don't really
>> belong in sys/types.h.  So why does the latter include sys/select.h?
> 
> Because it's done exactly the same way on FreeBSD and OpenBSD:
> 
>   # if    __BSD_VISIBLE
>   #include <sys/select.h>
>   [...]
> 
> Gnulib should allow to work with this to be portable.  So why does gnulib
> provide its own sys/select.h?  Is it configurable?

I'll raise this on the gnulib list, and we can come up with workarounds.
 But bypassing gnulib's sys/select.h probably won't work in general
(gnulib tends to always provide wrapper headers that use #include_next
to the system header, even if nothing is replaced, because it is easier
than trying to figure out when a wrapper header is needed).  Of course,
fixing it in gnulib won't help the myriad of projects that will be
bitten by this until they do a new release based on updated gnulib, so
maybe we can also find a way to make cygwin play nice without a gnulib
update required.  I'm still investigating.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org