This is based on Jakub's patch* which is used with many distributions – and is has
to be maintained by all of them; otherwise issues like lp #1878760 might creep in,
as discussed in #gcc yesterday. - As I am a huge fan of reducing code duplication
and local patches, I propose to add it to GCC proper behind a configure flag.

The patch idea is: offloading support is configured for one or multiple offload devices
but compilation silently ignores a target if its offload compiler is not found at
compile time and also dlopen errors of the associated libgomp plugin is also
silently ignored such that also the system where the program is run do not show
a run-time error.

This makes a lot of sense for distribution compilers – as it permits to configure
GCC to support offloading to one/multiple offload targets, but the actual
offload compiler and the run-time support can be put into separate, optional packages.

(When -foffload=$TARGET is specified explicitly, it still gives
an error if the offload compiler is not installed; this does not not affect the
libgomp/run-time side.)

However, for nondistro builds, compilers are usually installed "as is" and not
split into different packages – thus, compilation or run-time issues should give
an error.

Hence, the distro behaviour is only active when configured with
--enable-offload-defaulted.

Comments? OK for mainline?

Tobias
* https://src.fedoraproject.org/rpms/gcc/blob/rawhide/f/gcc11-foffload-default.patch

-----------------
Mentor Graphics (Deutschland) GmbH, Arnulfstrasse 201, 80634 München Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Frank Thürauf