On Tue, 20 Oct 2009, Diego Novillo wrote:
> On my nightly tester I force libstdc++ to be built with -flto enabled using
>
> $ make check RUNTESTFLAGS=--target_board=unix/"\{-flto/-O2,-fwhopr/-O2\}"
>
> Recently, this started showing 17 regressions that we don't see in
> other testsuites. It would be useful if we would run libstdc++ with
> -flto as well as the regular options. This will make testing
> libstdc++ longer, but it is useful for LTO as several of these tests
> are significant.
>
> Would this be OK?
libstdc++ is a very slow testsuite to run (compared with g++, for
example), at least if your target board is slow. There are infinitely
many combinations of tests you could run; that a combination is useful
does not mean it is appropriate to increase default testing time for
everyone with it. Instead, different people should run regression testers
with different options and report bugs found to Bugzilla; just as some
people run the testsuites with -fpic/-fPIC, or have processes for
generating testcases likely to trigger bugs and then report those bugs,
you could do run such a tester with LTO options and report the bugs you
find. Diversity of testers, and people actually going through the results
and ensuring there are PRs for all existing or new failures and that
regressions are marked as such, is more worthwhile than duplication of
similarly configured testers.
--
Joseph S. Myers
jos...@codesourcery.com
