https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79058
--- Comment #14 from Dominik Vogt <vogt at linux dot vnet.ibm.com> --- Isn't this more or less the same problem as the Avr issue? https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78883 On Avr, the register allocator would allow r31:HI if the expression is a paradoxical subreg of memory. This results in an invalid register pair, r31/r32. On Arm, the register allocator allows r0:SI for (subreg:DI (reg:SI)). Again we get an invalid register pair, r-1/r0. (Note that the cause of the crash is a "-1" in "regno"; comment 8.) => Bug in the register allocator?