[Bug bootstrap/61078] [5 Regression] ESA mode bootstrap failure since r209897

["krebbel at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61078

--- Comment #4 from Andreas Krebbel <krebbel at gcc dot gnu.org> ---
While the patch fixes this particular case the problem appears to be more
fundamental. The S/390 backend only considers GPR register pairs starting with
an even register to be valid (REGNO_PAIR_OK in s390.h). However, thanks to our
ABI we might encounter invalid pairs with hard regs during parameter passing.
This used to be resolved by the call preparation code which loads the parameter
into a pseudo first. That's why our move patterns are capable of dealing with
overlapping register pairs. However, the other patterns aren't. In that case it
was the negdi pattern but there might be similar issues with add, sub, ... .

In this example the invalid register pair in the negdi pattern comes from
combining the move which has been generated during call preparation (insn 18)
with the negdi pattern:

before combine:

(insn 17 16 18 5 (parallel [
            (set (reg:DI 49 [ D.1409 ])
                (neg:DI (reg:DI 44 [ D.1409 ])))
            (clobber (reg:CC 33 %cc))
        ]) t.c:20 491 {*negdi2_31}
     (expr_list:REG_DEAD (reg:DI 44 [ D.1409 ])
        (expr_list:REG_UNUSED (reg:CC 33 %cc)
            (nil))))
(insn 18 17 19 5 (set (reg:DI 3 %r3)
        (reg:DI 49 [ D.1409 ])) t.c:20 64 {*movdi_31}
     (expr_list:REG_DEAD (reg:DI 49 [ D.1409 ])
        (nil)))

after combine:

(insn 18 17 19 5 (parallel [
            (set (reg:DI 3 %r3)
                (neg:DI (reg:DI 44 [ D.1409 ])))
            (clobber (reg:CC 33 %cc))
        ]) t.c:20 491 {*negdi2_31}
     (expr_list:REG_UNUSED (reg:CC 33 %cc)
        (expr_list:REG_DEAD (reg:DI 44 [ D.1409 ])
            (nil))))

r44 then is assigned to hard reg r2 creating the problematic overlap:

(insn 18 20 19 5 (parallel [
            (set (reg:DI 3 %r3)
                (neg:DI (reg:DI 2 %r2 [orig:44 D.1409 ] [44])))
            (clobber (reg:CC 33 %cc))
        ]) t.c:20 491 {*negdi2_31}
     (nil))

I'm not sure what the best way is to fix this. Instead of implementing
splitters for all these cases the easiest might to mark the destination
operands as early clobber in the split patterns. That way reload is forced to
repair the situation.