On 04/06/2018 03:33 PM, Jeff King wrote:
On Fri, Apr 06, 2018 at 02:06:50PM -0400, Jonathan Primrose wrote:
A while ago, the configure test for FREAD_READS_DIRECTORIES was changed
to (attempt to) check for a NULL result from fopen. Unfortunately, the
test will always fail - because it won't compile. The code snippet in
configure.ac translates to:
return f);
Either there's an extra ) or a missing (. A cast to int wouldn't hurt
either.
Oops. This is due to my 3adf9fdecf (configure.ac: loosen
FREAD_READS_DIRECTORIES test program, 2017-06-14).
Neither I nor the original tester noticed, because we're on Linux, which
needs that set.
I'm also running Linux, but somehow I ended up needing to check
config.log and saw the error. I don't remember why I had to check.
(I've been fixing the problem locally for the last few releases,
figuring someone else would notice. I also remove the 'rm configure'
from the distclean target in Makefile, just in case I decide to
rebuild later.)
I'd supply a patch to fix this, but I'm not sure what the results of
suddenly fixing the test would be. It seems to work well on my
machine, but I don't stress git much here, and it's just one machine.
I think it should be fixed (and agreed on the "int" thing; the return
value should be "f != NULL" or similar).
Ironically, most of the time I check config.log is because of the
missing cast to int. A few years ago I set this machine up as a
tri-arch system (x86, x86_64, and x32). Because of subtle errors
that appear when converting between int and pointer on x32, I
tend to use -Werror=int-conversion on all my builds, and I wrap
configure in a script that (among other things) checks config.log
for that particular error. Even though x32 isn't as promising
anymore (thanks to a few projects rejecting support patches),
there are very few packages that still trigger the error.
I don't know autoconf very well, but is there a way to invoke it that
will let us know if a test-program fails to compile at all? Obviously
for probing the system compler/includes, "does not compile" is an
expected possible outcome. But for tests like this it's really a
tristate: yes, no, or something went horribly wrong.
I don't know if that's supported. From what I've seen in the docs I've
found online, it looks like the closest is a case for "I couldn't run
the program because I'm cross-compiling." You might be able to
determine that the compile and/or link failed by looking at some of
autoconf's internal variables, but it would be fragile.
-Peff
Jonathan
