Quoting Eric Botcazou <ebotca...@adacore.com>:
This fixes the problem on the Epiphany, but I haven't tested if it also
fixes it on the SH - AAUI you'd have to back out a workaround first to
properly test it.
The fix belongs in the caller instead - it should make sure that what it is
doing makes sense at all.
set_unique_reg note already makes a number of checks so that its
multitude of callers doesn't have to. E.g. it checks that there is
indeed only one set,
not only one set of a live register.
As you can see in the PR, there are different pieces in the compiler that
end up in the same bogosity as they add notes with the wrong mode of data.
Having the check & fix in one place makes this better maintainable.