PR 84723 is about an ICE caused by GCC accepting the mutually
exclusive attributes target_clones and naked during parsing,
only to then later crash while assuming such conflicts don't
happen.
The attached patch adds an exclusion for the two attributes
to detect when they're both specified on declarations of the
same function and avoid the ICE that way (there are other
problems with target_clones that are unrelated to other
interaction with other attributes).
As this is the first instance of an exclusion between a pair
of a front/middle-end and back-end attributes, testing the
straightforward solution of just adding an exclusion for them
exposed a restriction in the implementation that required more
extensive changes to make things work. The current (unpatched)
code requires exclusions to be symmetric (e.g., when there is
an exclusion entry for target_clones and naked, there must also
be another exclusion entry the other way around). This makes
sense for attributes in the same table (e.g., all C/C++
attributes, or all i386 back-end attributes) but not between
attributes in different tables. The patch makes additional
changes to allow asymmetric exclusions without compromising
the detection of attribute conflicts. This additional
enhancement allowed me to remove the self tests.
Since I started working on this last night Jakub has already
posted an alternate fix that doesn't make use of the exclusion
framework (plus it fixes the additional bugs/ICEs I referred
to above). This patch is thus meant to apply on top of Jakub's
(the tests will need tweaking to adjust the expected diagnostics).
If my solution is considered too invasive at this stage I'm fine
holding off on it and re-posting it in stage 1. While working
on it I noticed other opportunities for improvements in this
area that I would like to make then.
Martin