https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102211
--- Comment #10 from Jim Wilson <wilson at gcc dot gnu.org> --- I attached a patch which is my proposed solution to the RISC-V backend. It adds a new f_register_operand predicate and modifies patterns that use the f constraint to use it instead of register_operand. This was tested with an default language gcc build, glibc build, and glibc check on an unmatched running OpenEmbedded. And an all language gcc build, glibc build, and glibc check on an unleashed running Fedora with an old 4.15 kernel. Both succeeded as well as expected. I'll be trying gcc check next. Meanwhile, the validate_subregs patch was reverted, so we don't immediately need this to fix build errors. But it still might be useful if validate_subregs changes again. Or if it happens to give better code, though I think it won't do anything if validate_subregs is rejecting the subregs we are checking for in f_register_operand.
