copyprop_hardreg_forward removes division

[Jan Hubicka]

> I have a problem with a dissapearing division instruction.  Before
> copyprop_hardreg_forward in regrename.c is called, the situation looks
> like this:
> 
>     (insn 47 39 19 (set (reg:SI 3 3)
>             (reg:SI 3 3)) -1 (nil)
>         (nil))
>     
>     ;; This instruction divides (reg:SI 3) by (const_int 3), and puts
>     ;; the quotient in (reg:SI 3) and the remainder in the next register,
>     ;; (reg:SI 4).
>     (insn 19 47 24 (parallel[ 
>                 (set (subreg:SI (reg:DI 3 3 [30]) 0)
>                     (div:SI (reg:SI 3 3)
>                         (const_int 3 [0x3])))
>                 (set (subreg:SI (reg:DI 3 3 [30]) 4)
>                     (mod:SI (reg:SI 3 3)
>                         (const_int 3 [0x3])))
>             ] ) 107 {IDIV} (insn_list 18 (nil))
>         (expr_list:REG_UNUSED (reg:SI 4 4)
>             (nil)))

It looks like you are having same problem as I had. copyprop_hardreg misses
the stores to subregs.  I've commited fix to that so hope it is OK in the
tree (see kill_value).

Honza
>     
>     (note 24 19 27 NOTE_INSN_FUNCTION_END)
>     
>     (insn 27 24 30 (set (reg/i:SI 1 1)
>             (reg:SI 3 3 [29])) 74 {*movsi_cache} (insn_list 19 (nil))
>         (expr_list:REG_DEAD (reg:SI 3 3 [29])
>             (nil)))
> 
> However, when copyprop_hardreg_forward_1 is called from
> copyprop_hardreg_forward, it will change (reg:SI 3) in insns 47 and 27
> to (reg:SI 2), but not in insn 19:
> 
>     (insn 47 39 19 (set (reg:SI 3 3)
>             (reg:SI 2 2 [3])) 74 {*movsi_cache} (nil)
>         (nil))
>     
>     (insn 19 47 24 (parallel[ 
>                 (set (subreg:SI (reg:DI 3 3 [30]) 0)
>                     (div:SI (reg:SI 3 3)
>                         (const_int 3 [0x3])))
>                 (set (subreg:SI (reg:DI 3 3 [30]) 4)
>                     (mod:SI (reg:SI 3 3)
>                         (const_int 3 [0x3])))
>             ] ) 107 {IDIV} (insn_list 18 (nil))
>         (expr_list:REG_UNUSED (reg:SI 4 4)
>             (nil)))
>     
>     (note 24 19 27 NOTE_INSN_FUNCTION_END)
>     
>     (insn 27 24 30 (set (reg/i:SI 1 1)
>             (reg:SI 2 2 [29])) 74 {*movsi_cache} (insn_list 19 (nil))
>         (expr_list:REG_DEAD (reg:SI 3 3 [29])
>             (nil)))
> 
> update_life_info will then happily remove the unnecessary insn 19.
> 
> How can I avoid this?
> 
> -- 
> Lars Brinkhoff          http://lars.nocrew.org/     Linux, GCC, PDP-10,
> Brinkhoff Consulting    http://www.brinkhoff.se/    HTTP programming