From: "juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai>
To: rguenther <rguenther@suse.de>
Cc: gcc-patches <gcc-patches@gcc.gnu.org>,
jeffreyalaw <jeffreyalaw@gmail.com>
Subject: Re: Re: [PATCH] LCM: Export 2 helpful functions as global for VSETVL PASS use in RISC-V backend
Date: Mon, 21 Aug 2023 15:24:24 +0800 [thread overview]
Message-ID: <E80DBDAEE20277B9+202308211524232457307@rivai.ai> (raw)
In-Reply-To: <nycvar.YFH.7.77.849.2308210707100.12935@jbgna.fhfr.qr>
[-- Attachment #1: Type: text/plain, Size: 5921 bytes --]
Hi, Richi.
This patch is how I use LCM functions:
https://gcc.gnu.org/pipermail/gcc-patches/2023-August/627953.html
>> they are just necessary
>> to compute pre_edge_lcm_avs which I see you already do.
In Phase 4 I use pre_edge_lcm_av to PRE fo current VSETVL cfg information.
However, it's not enough since I need phase 3 fuse VSETVL information to get better codegen.
The is how I use the functions:
+ /* Compute global availability. */
compute_available (m_vector_manager->vector_comp,
m_vector_manager->vector_kill,
m_vector_manager->vector_avout,
m_vector_manager->vector_avin);
- changed_p |= cleanup_illegal_dirty_blocks ();
+ /* Compute global anticipatability. */
+ compute_antinout_edge (m_vector_manager->vector_antic,
+ m_vector_manager->vector_transp,
+ m_vector_manager->vector_antin,
+ m_vector_manager->vector_antout);
+ /* Compute earliestness. */
+ compute_earliest (m_vector_manager->vector_edge_list,
+ m_vector_manager->vector_exprs.length (),
+ m_vector_manager->vector_antin,
+ m_vector_manager->vector_antout,
+ m_vector_manager->vector_avout,
+ m_vector_manager->vector_kill,
+ m_vector_manager->vector_earliest);
+ changed_p |= earliest_fusion ();
You can see I explicitly call 'compute_earliest' followed by 'earliest_fusion'I need the result from 'compute_earliest' to do the VSETVL fusion that's the information 'pre_edge_lcm_av' didn't give us.
>> Just to say
>> you are possibly going to blow up compile-time complexity of your
>> VSETVL dataflow problem?
No, I export 'compute_earliest' as global because 'pre_edge_lcm_av' didn't give us the 'earliest' result that I need.
Thanks.
juzhe.zhong@rivai.ai
From: Richard Biener
Date: 2023-08-21 15:09
To: Juzhe-Zhong
CC: gcc-patches; jeffreyalaw
Subject: Re: [PATCH] LCM: Export 2 helpful functions as global for VSETVL PASS use in RISC-V backend
On Mon, 21 Aug 2023, Juzhe-Zhong wrote:
> This patch exports 'compute_antinout_edge' and 'compute_earliest' as global scope
> which is going to be used in VSETVL PASS of RISC-V backend.
>
> The demand fusion is the fusion of VSETVL information to emit VSETVL which dominate and pre-config for most
> of the RVV instructions in order to elide redundant VSETVLs.
>
> For exmaple:
>
> for
> for
> for
> if (cond}
> VSETVL demand 1: SEW/LMUL = 16 and TU policy
> else
> VSETVL demand 2: SEW = 32
>
> VSETVL pass should be able to fuse demand 1 and demand 2 into new demand: SEW = 32, LMUL = M2, TU policy.
> Then emit such VSETVL at the outmost of the for loop to get the most optimal codegen and run-time execution.
>
> Currenty the VSETVL PASS Phase 3 (demand fusion) is really messy and un-reliable as well as un-maintainable.
> And, I recently read dragon book and morgan's book again, I found there "earliest" can allow us to do the
> demand fusion in a very reliable and optimal way.
>
> So, this patch exports these 2 functions which are very helpful for VSETVL pass.
It would be nice to put these internal functions into a class or a
namespace given their non LCM name. I don't see how you are going
to use these intermediate DF functions - they are just necessary
to compute pre_edge_lcm_avs which I see you already do. Just to say
you are possibly going to blow up compile-time complexity of your
VSETVL dataflow problem?
> gcc/ChangeLog:
>
> * lcm.cc (compute_antinout_edge): Export as global use.
> (compute_earliest): Ditto.
> (compute_rev_insert_delete): Ditto.
> * lcm.h (compute_antinout_edge): Ditto.
> (compute_earliest): Ditto.
>
> ---
> gcc/lcm.cc | 7 ++-----
> gcc/lcm.h | 3 +++
> 2 files changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/gcc/lcm.cc b/gcc/lcm.cc
> index 94a3ed43aea..03421e490e4 100644
> --- a/gcc/lcm.cc
> +++ b/gcc/lcm.cc
> @@ -56,9 +56,6 @@ along with GCC; see the file COPYING3. If not see
> #include "lcm.h"
>
> /* Edge based LCM routines. */
> -static void compute_antinout_edge (sbitmap *, sbitmap *, sbitmap *, sbitmap *);
> -static void compute_earliest (struct edge_list *, int, sbitmap *, sbitmap *,
> - sbitmap *, sbitmap *, sbitmap *);
> static void compute_laterin (struct edge_list *, sbitmap *, sbitmap *,
> sbitmap *, sbitmap *);
> static void compute_insert_delete (struct edge_list *edge_list, sbitmap *,
> @@ -79,7 +76,7 @@ static void compute_rev_insert_delete (struct edge_list *edge_list, sbitmap *,
> This is done based on the flow graph, and not on the pred-succ lists.
> Other than that, its pretty much identical to compute_antinout. */
>
> -static void
> +void
> compute_antinout_edge (sbitmap *antloc, sbitmap *transp, sbitmap *antin,
> sbitmap *antout)
> {
> @@ -170,7 +167,7 @@ compute_antinout_edge (sbitmap *antloc, sbitmap *transp, sbitmap *antin,
>
> /* Compute the earliest vector for edge based lcm. */
>
> -static void
> +void
> compute_earliest (struct edge_list *edge_list, int n_exprs, sbitmap *antin,
> sbitmap *antout, sbitmap *avout, sbitmap *kill,
> sbitmap *earliest)
> diff --git a/gcc/lcm.h b/gcc/lcm.h
> index e08339352e0..7145d6fc46d 100644
> --- a/gcc/lcm.h
> +++ b/gcc/lcm.h
> @@ -31,4 +31,7 @@ extern struct edge_list *pre_edge_rev_lcm (int, sbitmap *,
> sbitmap *, sbitmap *,
> sbitmap *, sbitmap **,
> sbitmap **);
> +extern void compute_antinout_edge (sbitmap *, sbitmap *, sbitmap *, sbitmap *);
> +extern void compute_earliest (struct edge_list *, int, sbitmap *, sbitmap *,
> + sbitmap *, sbitmap *, sbitmap *);
> #endif /* GCC_LCM_H */
>
--
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
next prev parent reply other threads:[~2023-08-21 7:24 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-08-21 1:04 Juzhe-Zhong
2023-08-21 7:09 ` Richard Biener
2023-08-21 7:24 ` juzhe.zhong [this message]
2023-08-21 7:46 ` juzhe.zhong
2023-08-21 7:50 ` Richard Biener
2023-08-21 8:06 ` juzhe.zhong
[not found] ` <202308211606108820614@rivai.ai>
2023-08-21 8:38 ` juzhe.zhong
2023-08-21 9:12 ` Richard Biener
2023-08-21 9:32 ` Lehua Ding
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=E80DBDAEE20277B9+202308211524232457307@rivai.ai \
--to=juzhe.zhong@rivai.ai \
--cc=gcc-patches@gcc.gnu.org \
--cc=jeffreyalaw@gmail.com \
--cc=rguenther@suse.de \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).