On Fri, 02 Aug 2019 at 11:41:05 +0100, Iain Lane wrote:
> For the record: at Debconf doko suggested to me that a way to start
> debugging this from the GCC end would be to produce one working and one
> "broken" build of the same version of glib2.0, and then copy object
> files from the broken to the working one until that too becomes broken.
> That sounded long / tedious / fiddly so we didn't manage to get round to
> doing it while we were both there, unfortunately.
This is not straightforward in the Meson/Ninja build system, but what I
could do was to hack build.ninja to compile different files with different
compilers, which narrowed it down to gtestutils.c (?!).
Playing with #pragma GCC optimize ("O1") and #pragma GCC optimize ("O2")
I was able to narrow it down further, to this:
If you compile test_run_seed() with -O1, and the rest of gtestutils.c
with -O2, the clutter test hangs.
(But I need to test that minimal change in a clean build to make sure
that's really true and not being affected by other hacks.)
smcv
