Re: [Qemu-devel] [PATCH v4 07/15] target-ppc: implement branch-less divw[o][.]

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 5056 bytes --]

On Wed, Jul 27, 2016 at 12:11:08PM +0530, Nikunj A Dadhania wrote:
> David Gibson <david@gibson.dropbear.id.au> writes:
> 
> > [ Unknown signature status ]
> > On Wed, Jul 27, 2016 at 11:47:15AM +0530, Nikunj A Dadhania wrote:
> >> David Gibson <david@gibson.dropbear.id.au> writes:
> >> 
> >> > [ Unknown signature status ]
> >> > On Tue, Jul 26, 2016 at 05:28:30PM +0530, Nikunj A Dadhania wrote:
> >> >> While implementing modulo instructions figured out that the
> >> >> implementation uses many branches. Change the logic to achieve the
> >> >> branch-less code. Undefined value is set to dividend in case of invalid
> >> >> input.
> >> >> 
> >> >> Signed-off-by: Nikunj A Dadhania <nikunj@linux.vnet.ibm.com>
> >> >> ---
> >> >>  target-ppc/translate.c | 48 +++++++++++++++++++++++-------------------------
> >> >>  1 file changed, 23 insertions(+), 25 deletions(-)
> >> >> 
> >> >> diff --git a/target-ppc/translate.c b/target-ppc/translate.c
> >> >> index 7c7328f..69d9ae0 100644
> >> >> --- a/target-ppc/translate.c
> >> >> +++ b/target-ppc/translate.c
> >> >> @@ -1049,41 +1049,39 @@ static void gen_addpcis(DisasContext *ctx)
> >> >>  static inline void gen_op_arith_divw(DisasContext *ctx, TCGv ret, TCGv arg1,
> >> >>                                       TCGv arg2, int sign, int compute_ov)
> >> >>  {
> >> >> -    TCGLabel *l1 = gen_new_label();
> >> >> -    TCGLabel *l2 = gen_new_label();
> >> >> -    TCGv_i32 t0 = tcg_temp_local_new_i32();
> >> >> -    TCGv_i32 t1 = tcg_temp_local_new_i32();
> >> >> +    TCGv_i32 t0 = tcg_temp_new_i32();
> >> >> +    TCGv_i32 t1 = tcg_temp_new_i32();
> >> >> +    TCGv_i32 t2 = tcg_temp_new_i32();
> >> >> +    TCGv_i32 t3 = tcg_temp_new_i32();
> >> >>  
> >> >>      tcg_gen_trunc_tl_i32(t0, arg1);
> >> >>      tcg_gen_trunc_tl_i32(t1, arg2);
> >> >> -    tcg_gen_brcondi_i32(TCG_COND_EQ, t1, 0, l1);
> >> >> -    if (sign) {
> >> >> -        TCGLabel *l3 = gen_new_label();
> >> >> -        tcg_gen_brcondi_i32(TCG_COND_NE, t1, -1, l3);
> >> >> -        tcg_gen_brcondi_i32(TCG_COND_EQ, t0, INT32_MIN, l1);
> >> >> -        gen_set_label(l3);
> >> >> -        tcg_gen_div_i32(t0, t0, t1);
> >> >> -    } else {
> >> >> -        tcg_gen_divu_i32(t0, t0, t1);
> >> >> -    }
> >> >> -    if (compute_ov) {
> >> >> -        tcg_gen_movi_tl(cpu_ov, 0);
> >> >> -    }
> >> >> -    tcg_gen_br(l2);
> >> >> -    gen_set_label(l1);
> >> >>      if (sign) {
> >> >> -        tcg_gen_sari_i32(t0, t0, 31);
> >> >> +        tcg_gen_setcondi_i32(TCG_COND_EQ, t2, t0, INT_MIN);
> >> >> +        tcg_gen_setcondi_i32(TCG_COND_EQ, t3, t1, -1);
> >> >> +        tcg_gen_and_i32(t2, t2, t3);
> >> >> +        tcg_gen_setcondi_i32(TCG_COND_EQ, t3, t1, 0);
> >> >> +        tcg_gen_or_i32(t2, t2, t3);
> >> >> +        tcg_gen_movi_i32(t3, 0);
> >> >> +        tcg_gen_movcond_i32(TCG_COND_NE, t1, t2, t3, t2, t1);
> >> >> +        tcg_gen_div_i32(t3, t0, t1);
> >> >> +        tcg_gen_extu_i32_tl(ret, t3);
> >> >
> >> > Should this be a signed extend, given it's a signed divide?
> >> 
> >> Don't think so, as the instruction is 32-bit, caller will only look at
> >> the 32bit and div_i32 is signed, it will take care of extending sign
> >> till 32-boundary.
> >
> > Hrm.  I thought most 32-bit arithmetic operations on Power actually
> > set the underlying 64-bit registers to a sign extended version of the
> > 32-bit result.
> 
> I think, when I want to operate on it as 64-bit, i will need signed
> extension. rth can give more info on this.
> 
> Retained the behaviour as per the previous code as well:

Ah, yes RT[0:31] <- undefined according to the ISA, so I guess this is
ok (though ideally we'd double check what the actual hardware does and
match that, just in case something relies on it).

> 
>     tcg_gen_trunc_tl_i32(t0, arg1);
>     tcg_gen_trunc_tl_i32(t1, arg2);
>     tcg_gen_brcondi_i32(TCG_COND_EQ, t1, 0, l1);
>     if (sign) {
>         TCGLabel *l3 = gen_new_label();
>         tcg_gen_brcondi_i32(TCG_COND_NE, t1, -1, l3);
>         tcg_gen_brcondi_i32(TCG_COND_EQ, t0, INT32_MIN, l1);
>         gen_set_label(l3);
>         tcg_gen_div_i32(t0, t0, t1);
>     } else {
>         tcg_gen_divu_i32(t0, t0, t1);
>     }
>     if (compute_ov) {
>         tcg_gen_movi_tl(cpu_ov, 0);
>     }
>     tcg_gen_br(l2);
>     gen_set_label(l1);
>     if (sign) {
>         tcg_gen_sari_i32(t0, t0, 31);
>     } else {
>         tcg_gen_movi_i32(t0, 0);
>     }
>     if (compute_ov) {
>         tcg_gen_movi_tl(cpu_ov, 1);
>         tcg_gen_movi_tl(cpu_so, 1);
>     }
>     gen_set_label(l2);
>     tcg_gen_extu_i32_tl(ret, t0);                     <<<<<<<<<<<<<<<<<<<<<<<
>     tcg_temp_free_i32(t0);
>     tcg_temp_free_i32(t1);
> 
> IMO, thats correct.
> 
> Regards
> Nikunj
> 

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]