[Bug target/125992] [16/17 Regression] veusz miscompilation on s390x with -march=z13 -O2 since r16-3301

[stefansf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125992

--- Comment #3 from Stefan Schulze Frielinghaus <stefansf at gcc dot gnu.org> ---
Yikes, this is an unfortunate one. As you already said, we make use of the
pre-RA splitter in order to enforce that operands 2 and 3 get allocated a
register pair. What is unfortunate here is that the solution depends on the
order of the clobbers in order to "mark" the split such that we don't run into
an infinite loop which looks a bit brittle to me. However, at the moment I do
not have a prettier solution. This may be something where dependent register
filters could become handy.

I can confirm that this restores the veusz build.  Bootstrap+regtest are
running for z900,z10,z13,z14,z17.  I will come back to you once they finished.