Re: Re: [PATCH] RISC-V: Avoid unnecessary slideup in compress pattern of vec_perm
Address comment: [PATCH V2] RISC-V: Avoid unnecessary slideup in compress
pattern of vec_perm (gnu.org)
juzhe.zh...@rivai.ai
From: Jeff Law
Date: 2023-09-10 21:34
To: Juzhe-Zhong; gcc-patches
CC: kito.cheng; kito.cheng; rdapp.gcc
Subject: Re: [PATCH] RISC-V: Avoid unnecessary slideup in compress pattern of
vec_perm
On 9/9/23 21:55, Juzhe-Zhong wrote:
> If a const vector all elements are same, the slide up is unnecessary.
>
> gcc/ChangeLog:
>
> * config/riscv/riscv-v.cc (shuffle_compress_patterns): Avoid unnecessary
> slideup.
>
> ---
> gcc/config/riscv/riscv-v.cc | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
> index bee60de1d26..7ef884907b8 100644
> --- a/gcc/config/riscv/riscv-v.cc
> +++ b/gcc/config/riscv/riscv-v.cc
> @@ -2697,7 +2697,7 @@ shuffle_compress_patterns (struct expand_vec_perm_d *d)
> rtx mask = force_reg (mask_mode, builder.build ());
>
> rtx merge = d->op1;
> - if (need_slideup_p)
> + if (need_slideup_p && !const_vec_duplicate_p (d->op1))
> {
> int slideup_cnt = vlen - (d->perm[vlen - 1].to_constant () % vlen) -
> 1;
> rtx ops[] = {d->target, d->op1, gen_int_mode (slideup_cnt, Pmode)};
Would it be better to adjust how we compute need_slidup_p to check
!const_vec_duplicate_p (d->op1) instead of doing it here?
That way the name "need_slideup_p" stays consistent with the intent of
the code. It would also mean we wouldn't need to duplicate the
additional check if we wanted to model the use of slideup in the cost
calculations.
Jeff
Re: [PATCH] RISC-V: Avoid unnecessary slideup in compress pattern of vec_perm
On 9/9/23 21:55, Juzhe-Zhong wrote:
If a const vector all elements are same, the slide up is unnecessary.
gcc/ChangeLog:
* config/riscv/riscv-v.cc (shuffle_compress_patterns): Avoid
unnecessary slideup.
---
gcc/config/riscv/riscv-v.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index bee60de1d26..7ef884907b8 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -2697,7 +2697,7 @@ shuffle_compress_patterns (struct expand_vec_perm_d *d)
rtx mask = force_reg (mask_mode, builder.build ());
rtx merge = d->op1;
- if (need_slideup_p)
+ if (need_slideup_p && !const_vec_duplicate_p (d->op1))
{
int slideup_cnt = vlen - (d->perm[vlen - 1].to_constant () % vlen) - 1;
rtx ops[] = {d->target, d->op1, gen_int_mode (slideup_cnt, Pmode)};
Would it be better to adjust how we compute need_slidup_p to check
!const_vec_duplicate_p (d->op1) instead of doing it here?
That way the name "need_slideup_p" stays consistent with the intent of
the code. It would also mean we wouldn't need to duplicate the
additional check if we wanted to model the use of slideup in the cost
calculations.
Jeff
[PATCH] RISC-V: Avoid unnecessary slideup in compress pattern of vec_perm
If a const vector all elements are same, the slide up is unnecessary.
gcc/ChangeLog:
* config/riscv/riscv-v.cc (shuffle_compress_patterns): Avoid
unnecessary slideup.
---
gcc/config/riscv/riscv-v.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index bee60de1d26..7ef884907b8 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -2697,7 +2697,7 @@ shuffle_compress_patterns (struct expand_vec_perm_d *d)
rtx mask = force_reg (mask_mode, builder.build ());
rtx merge = d->op1;
- if (need_slideup_p)
+ if (need_slideup_p && !const_vec_duplicate_p (d->op1))
{
int slideup_cnt = vlen - (d->perm[vlen - 1].to_constant () % vlen) - 1;
rtx ops[] = {d->target, d->op1, gen_int_mode (slideup_cnt, Pmode)};
--
2.36.3
