https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122900

--- Comment #4 from Jan Žižka <jan.zizka at nokia dot com> ---
(In reply to Richard Biener from comment #3)
> So just to confirm.  You have a set of TUs TA that are built with
> -fprofile-generate, lined into some application A and profile is generated
> for those.
> Then you have application B that re-uses object files from TA plus adds
> TUs from a set TB, built without -fprofile-generate, and when linking those
> together with LTO and -fprofile-use this ICEs.

Yes that's like that. Only the application B is actually application A with new
features added with new TUs.

> But when compiling the TB TUs with -fno-profile-use but still using
> -fprofile-use at link time(?) this avoids the ICE?

Yes indeed this avoids ICE in our case.

> In principle there's nothing wrong with using -fprofile-use for TUs without
> any training data.

Ok good, we were not sure if this is something what might not be actually
supported.

> If you can share libxxx.so.ltrans13.ltrans.o (the LTRANS IL) this might
> be a reproducer people can work with if you are on an unpatched GCC 15.2
> and for example on x86_64-linux.

We are able to reproduce this with unpatched GCC 15.2 and on x84_64-linux. We
cannot yet share the libxxx.so as it contains proprietary application code :-(
but we are working on reproducer with content we will be allowed to share.

Reply via email to