[Bug rtl-optimization/126276] [17 Regression] Recent change breaks cris-elf

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126276

--- Comment #7 from Hans-Peter Nilsson <hp at gcc dot gnu.org> ---
Note mostly to self: Revisiting this, I can't help thinking this should instead
have changed the mode of those non-SImode (match_dup N) RHS of zero_extract in
the define_insn_and_splits generated split-ups to SImode, *or* drop the mode
size on RHS of the zero_extract of "*btst<mode>", instead of making the
"*btst<mode>" have sub-mode variants, as it's a single-size-mode insn.  (The
modes alluded to in the names are CCmodes.)  Something about having
iterator-generated variants not sitting well.  I need to chew on this.