http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52838
--- Comment #2 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2012-04-04
21:41:02 UTC ---
> This looks like a combine problem:
>
> (insn 8 6 9 2 (set (reg/f:SI 59 [ D.1705 ])
> (subreg/s/u:SI (reg:DI 60) 0)) pr52838.c:6 64 {*movsi_internal}
> (expr_list:REG_DEAD (reg:DI 60)
> (nil)))
>
> (insn 9 8 10 2 (set (reg/f:DI 62 [ D.1705 ])
> (zero_extend:DI (reg/f:SI 59 [ D.1705 ]))) pr52838.c:6 112
> {*zero_extendsidi2_rex64}
> (expr_list:REG_DEAD (reg/f:SI 59 [ D.1705 ])
> (nil)))
>
> This gets transformed by combine pass:
>
>
> Trying 8 -> 9:
> Successfully matched this instruction:
> (set (reg/f:DI 62 [ D.1705 ])
> (reg:DI 60))
> deferring deletion of insn with uid = 8.
> modifying insn i3 9 r62:DI=r60:DI
> REG_DEAD: r60:DI
>
> into:
>
> (note 8 6 9 2 NOTE_INSN_DELETED)
>
> (insn 9 8 10 2 (set (reg/f:DI 62 [ D.1705 ])
> (reg:DI 60)) pr52838.c:6 62 {*movdi_internal_rex64}
> (expr_list:REG_DEAD (reg:DI 60)
> (nil)))
>
> But, but ... high 32 bits are not cleared anymore!
The existence of (subreg/s/u:SI (reg:DI 60) 0) means that (reg:DI 60) is known
to be a zero-extended 32-bit value, so the optimization is valid (and useful).
