https://gcc.gnu.org/bugzilla/show_bug.cgi?id=51982
--- Comment #5 from Segher Boessenkool <segher at gcc dot gnu.org> --- Yes, the only two calls in the resulting machine code are both to unicode_eq. Splitting the ranges (and then assigning volatile regs to the first half) hurts if it doesn't end up helping shrink-wrapping. There also is the problem that some pseudos are already coalesced earlier (at expand time even) while in fact they are independent (some will naturally end up in r3, the return register, while others need to be non-volatile).
