Re: [PATCH] Disable loop2_invariant for -Os

[Richard Guenther <richard.guenther@gmail.com>]

On Wed, Jun 27, 2012 at 10:40 AM, Zhenqiang Chen <zhenqiang.chen@arm.com> wrote:
> Hi,
>
> In general, invariant motion itself can not reduce code size.

It can expose CSE opportunities across loops though.

> But it will
> change the liverange of the invariant, which might lead to more spilling.

"might" - indeed.  I wonder what the trade-off is here ... but given that you
leave tree loop invariant motion enabled it might not make much of a difference.

Still as this is mostly a spilling issue it looks odd to do that generally.  In
fact you could improve things by only disabling motion when that increases
register lifetime - it can after all reduce overall register lifetime:

for (;;)
  inv = inv1 + inv2;
  ... use inv;

to

inv = inv1 + inv2;
for (;;)
  ... use inv;

has register lifetime reduced.

Or at least like I suggest below.

> The patch disables loop2_invariant when optimizing for size.
>
> I measured the code size benefit for four targets based on CSiBE benchmark:
>
> ARM: 0.33%
> MIPS: 1.15%
> PPC: 0.24%
> X86: 0.45%
>
> Is it OK for trunk?
>
> Thanks!
> -Zhenqiang
>
> ChangeLog:
> 2012-06-27  Zhenqiang Chen <zhenqiang.chen@arm.com>
>
>        * loop-init.c (gate_rtl_move_loop_invariants): Disable
> loop2_invariant
>        when optimizing function for size.
>
> diff --git a/gcc/loop-init.c b/gcc/loop-init.c
> index 03f8f61..5d8cf73 100644
> --- a/gcc/loop-init.c
> +++ b/gcc/loop-init.c
> @@ -273,6 +273,12 @@ struct rtl_opt_pass pass_rtl_loop_done =
>  static bool
>  gate_rtl_move_loop_invariants (void)
>  {
> +  /* In general, invariant motion can not reduce code size. But it will
> +     change the liverange of the invariant, which increases the register
> +     pressure and might lead to more spilling.  */
> +  if (optimize_function_for_size_p (cfun))
> +    return false;
> +

Can you do this per loop instead?  Using optimize_loop_nest_for_size_p?

Thanks,
Richard.

>   return flag_move_loop_invariants;
>  }
>
>
>