[Bug target/53513] [SH] Add support for fschg and fpchg insns and improve fenv support

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53513

--- Comment #26 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Kazumoto Kojima from comment #25)
> (In reply to Oleg Endo from comment #23)
> > Kaz, could you please have an early look at it?
> 
> The idea looks OK to me.  Build fails on sh4-linux with the patch, though.
> Maybe a wrong version?
> There is a typo divdf3 expand.  s/gen_divdf3 /gen_divdf3_i /
> 
> -      expand_df_binop (&gen_divdf3_i, operands);
> +      emit_insn (gen_divdf3 (operands[0], operands[1], operands[2]));
> 
> With fixing it, yet segfaults during compiling libgcc2.c __powidf2 at

No, that was the correct version unfortunately.  And that was my
infinite-memory problem.  Thanks for spotting it.

> 
> #0  0x089706b4 in sh_adjust_cost (insn=0xb7f767e0, link=0xb7f79380, 
>     dep_insn=0xb7f762d0, cost=25) at trunk/gcc/config/sh/sh.c:10908
> 10908                                         SET_SRC (use_pat)))
> (gdb) l
> 10903                cycle earlier.  */
> 10904             else if (reload_completed
> 10905                      && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES
> 10906                      && (use_pat = single_set (insn))
> 10907                      && ! regno_use_in (REGNO (SET_DEST (single_set
> (dep_insn))),
> 10908                                         SET_SRC (use_pat)))
> 
> It looks
> 
> -   (use (match_operand:PSI 3 "fpscr_operand" "c"))]
> +   (set (reg:SI FPSCR_STAT_REG) (const_int 0))
> +   (use (reg:SI FPSCR_MODES_REG))]
> 
> breaks the assumption of sh_adjust_cost which divdf3_i has only one
> set insn.

Yep, single_set now returns null for basically all fp insns and it's not
checked there.  The default implementation of single_set is also the reason why
delay slot stuffing is not working on fp insns anymore.