Re: test-poll vs ioctl

[Simon Josefsson]

Bruno Haible <address@hidden> writes:

> Simon Josefsson wrote:
>> The current situation breaks simple test-cases like this:
>> 
>> rm -rf m;gnulib-tool --create-testdir --with-tests --dir m poll; cd m; 
>> ./configure --host=i586-mingw32msvc --build=i686-pc-linux-gnu; make check; 
>> cd ..
>> 
>> Regardless of the solution, I think the above command should work.
>
> The above command does work.

No, it fails with latest gnulib master:

i586-mingw32msvc-gcc  -g -O2   -o test-perror.exe test-perror.o 
../gllib/libgnu.a 
i586-mingw32msvc-gcc -DHAVE_CONFIG_H -I.  -I. -I. -I.. -I./.. -I../gllib 
-I./../gllib   -g -O2 -MT test-poll.o -MD -MP -MF .deps/test-poll.Tpo -c -o 
test-poll.o test-poll.c
test-poll.c:31:23: error: sys/ioctl.h: No such file or directory
make[3]: *** [test-poll.o] Error 1
make[3]: Leaving directory `/home/jas/src/gnulib/m/gltests'
make[2]: *** [check-recursive] Error 1
make[2]: Leaving directory `/home/jas/src/gnulib/m/gltests'
make[1]: *** [check] Error 2
make[1]: Leaving directory `/home/jas/src/gnulib/m/gltests'
make: *** [check-recursive] Error 1
address@hidden:~/src/gnulib master$

However, I now realize this is a different problem than when --import is
used.

It seems that dependencies of test modules aren't respected by
--create-testdir --with-tests, there is no sys_ioctl.in.h file:

address@hidden:~/src/gnulib master$ find m |grep ioct
m/glm4/sys_ioctl_h.m4
address@hidden:~/src/gnulib master$ 

>> The problem is caused by:
>> 
>> 1) One gnulib module (e.g., shutdown) pulls in one file (winsock.c) in
>>    the gl/ directory
>> 
>> 2) A test module (e.g., poll-tests) depends on a module (e.g., ioctl)
>>    that change how winsock.c behaves (i.e., causes it to #include
>>    additional headers such as sys/ioctl.h)
>> 
>> I guess this happens whenever you re-use the same *.c file for more than
>> one module
>
> Correct. The aforementioned move of files from lib/ to tests/ assumes that
> every source file is compiled only for one module.
>
> I think what gnulib-tool does is basically right, and it's not worth the
> effort to make gnulib-tool understand which files may be compilation units
> in which modules. Rather, I see two solutions, that both introduce separate
> compilation units for each sockets module:
>
>   a) Rename winsock.c to w32sock.h.
>      Create a file connect.c containing
>
>        #define GNULIB_CONNECT 1
>        #include "w32sock.h"
>
>      and similarly for the other modules that make up winsock.c.
>
>      Remove the gl_MODULE_INDICATOR([$1]) line from the definition of
>      gl_SYS_SOCKET_MODULE_INDICATOR.
>
>   b) Split winsock.c into accept.c, connect.c, etc. - like we are doing
>      in gnulib for most other functionalities. The inline functions
>      (FD_TO_SOCKET, SOCKET_TO_FD, set_winsock_errno) can go to a private
>      header file, let's say, w32sock.h or socket-internal.h.
>
> I would prefer b), because it's leads to a more understandable code structure.

I don't care strongly, but b) seems easier to understand so I would
prefer it too.  Since Paolo didn't object to either proposal, I am
preparing patches to implement b) now.  Will push and e-mail them
separately.

/Simon