[Bug target/65505] [5 Regression][SH] ICE in sh_disp_addr_displacement

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65505

--- Comment #1 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Created attachment 35084
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35084&action=edit
possible patch

The problem here is that sh_disp_addr_displacement gets an address that looks
like:
(plus:SI (subreg:SI (reg:DI ...) ..) (const_int ...)

The predicates displacement_mem_operand and simple_mem_operand use
arith_reg_operand to match the register operand.  This also matches sign_extend
and subreg, which is not handled by constraints such as Sdd.

This is a latent problem, although it seems it never triggered on 4.8 and 4.9.

If we allow subregs for the register operand of address expressions, things get
a bit hairy.  Additional checks need to be added everywhere to make sure that
the subreg is indeed a register operand and not e.g. subreg (mem ...). 
Disallowing subregs in address expressions seems easier in this case.

The attached patch fixes the problem.