Simon Josefsson wrote: > my patch invokes valgrind with -q so that it does not > output anything, unless there are errors, of course
Oh, I see. Indeed. But it spends a lot of time in "make check", and does not exploit the full information that valgrind could provide. For a test, I ran "make check" in gettext/gettext-tools/tests, * With --enable-shared and libtool wrappers: - without valgrind: 105 sec. - with valgrind -q for just the sh that runs the test: 1191 sec. - with valgrind -q for just the sh that runs the libtool wrappers: 6457 sec. * Without libtool wrappers: - without valgrind: 42 sec. - with valgrind -q for just the programs: 2147 sec. And it spends that much time for only a particular use of valgrind: no checking for memory leaks. Given this, that a test suite that was designed to take 10 minutes on an older hardware (10 years ago) and 1 minute on nowadays hardware now takes an hour with valgrind, I would certainly not enable valgrind by default in a tarball. But I do agree with you that maintainers should run the tests with valgrind regularly. Bruno
