https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87763
Wilco <wilco at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |wilco at gcc dot gnu.org
--- Comment #3 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Segher Boessenkool from comment #2)
> The two relevant insns are
>
> (insn 35 4 40 2 (set (reg:SI 33 v1 [99])
> (reg:SI 0 x0 [ a ])) "87763.c":2:1 46 {*movsi_aarch64}
> (nil))
>
> (insn 7 10 8 2 (set (reg:SF 33 v1 [orig:95 c ] [95])
> (float:SF (reg:SI 33 v1 [99]))) "87763.c":2:14 921 {floatsisf2}
> (nil))
>
> (this is the LRA dump; IRA already allocated those registers, LRA didn't
> change anything).
>
> 99 was assigned a floating point register:
> Choosing alt 12 in insn 35: (0) w (1) rZ {*movsi_aarch64}
> Choosing alt 0 in insn 7: (0) =w (1) w {floatsisf2}
>
> an integer register would be alt 0 (r<-r), resp. alt 1 (w<-?r). So
> apparently
> the "?" was costed as more strongly than the cost of the removable register
> move, or IRA didn't see that move was unnecessary.
>
> Confirmed, btw.
IRA costing doesn't consider the possibility of a simple move being removeable.
Neither does it consider that some variants are more expensive than others -
use of '?' is the only way to mark more expensive variants.
The example shown is fixed when I change the w=rZ variant to ?w=rZ in
movsi_aarch64.
