From: Jeff Law <jeffreyalaw@gmail.com>
To: Vineet Gupta <vineetg@rivosinc.com>
Cc: gcc@gcc.gnu.org, Kito Cheng <kito.cheng@gmail.com>,
Palmer Dabbelt <palmer@rivosinc.com>,
gnu-toolchain <gnu-toolchain@rivosinc.com>,
pinskia@gmail.com, GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: Followup on PR/109279: large constants on RISCV
Date: Wed, 7 Jun 2023 17:44:00 -0600 [thread overview]
Message-ID: <80218276-fc3c-032b-eaa7-0a4b0e8a859f@gmail.com> (raw)
In-Reply-To: <09f944a2-1123-75e3-ce2f-2080df94c964@rivosinc.com>
On 6/1/23 20:38, Vineet Gupta wrote:
> Hi Jeff,
>
> I finally got around to collecting various observations on PR/109279 -
> more importantly the state of large constants in RV backend, apologies
> in advance for the long email.
>
> It seems the various commits in area have improved the original test
> case of 0x1010101_01010101
>
> Before 2e886eef7f2b | With 2e886eef7f2b | With
> 0530254413f8 | With c104ef4b5eb1
Right. The handling of that constant shows a nice progression. On our
architecture the latter two versions are probably equivalent from a
latency standpoint, but the last is obviously best as it's smaller and
probably better on in-order architectures as well.
>
> But same commits seem to have regressed Andrew's test from same PR
> (which is the theme of this email).
> The seemingly contrived test turned out to be much more than I'd hoped for.
>
> long long f(void)
> {
> unsigned t = 0x101_0101;
> long long t1 = t;
> long long t2 = ((unsigned long long )t) << 32;
> asm("":"+r"(t1));
> return t1 | t2;
> }
[ ... ]
It may be more instructions, but I suspect they end up being the same
performance for us across all three varaints. Fusion and out-of-order
execution save the day. But I realize there may be targets where the
first is going to be preferred.
>
> Before 2e886eef7f2b | With 2e886eef7f2b | With 0530254413f8
> (ideal code) | define_insn_and_split | "splitter relaxed new
> | | pseudos"
> li a0,0x1010000 | li a5,0x1010000 | li a0,0x101_0000
> addi a0,a0,0x101 | addi a5,a5,0x101 | addi a0,a0,0x101
> slli a5,a0,32 | mv a0,a5 | li a5,0x101_0000
> or a0,a0,a5 | slli a5,a5,32 | slli a0,a0,32
> ret | or a0,a0,a5 | addi a5,a5,0x101
> | ret | or a0,a5,a0
> | ret
>
> As a baseline, RTL just before cse1 (in 260r.dfinit) in all of above is:
[ ... ]
Right. Standard looking synthesis.
>
>
> Prior to 2e886eef7f2b, cse1 could do its job: finding oldest equivalent
> registers for the fragments of const and reusing the reg.
Right. That's what I would expect.
[ ... ]
>
>
> With 2e886eef7f2b, define_insn_and_split "*mvconst_internal" recog()
> kicks in during cse1, eliding insns for a const_int.
>
> (insn 7 6 8 2 (set (reg:DI 137)
> (const_int [0x1010101])) {*mvconst_internal}
> (expr_list:REG_EQUAL (const_int [0x1010101])))
> [...]
>
> (insn 11 10 12 2 (set (reg:DI 140)
> (const_int [0x1010101_00000000])) {*mvconst_internal}
> (expr_list:REG_EQUAL (const_int [0x1010101_00000000]) ))
Understood. Not ideal, but we generally don't have good ways to limit
patterns to being available at different times during the optimization
phase. One thing you might want to try (which I thought we used at one
point) was make the pattern conditional on cse_not_expected. The goal
would be to avoid exposing the pattern until a later point in the
optimizer pipeline. It may have been the case that we dropped that over
time during development. It's all getting fuzzy at this point.
>
> Eventually split1 breaks it up using same mvconst_internal splitter, but
> the cse opportunity has been lost.
Right. I'd have to look at the pass definitions, but I suspect the
splitting pass where this happens is after the last standard CSE pass.
So we don't get a chance to CSE the constant synthesis.
> *This is a now a baseline for large consts handling for RV backend which
> we all need to be aware of*.
Understood. Though it's not as bad as you might think :-) You can
spend an inordinate amount of time improving constant synthesis,
generate code that looks really good, but in the end it may not make a
bit of different in real performance. Been there, done that. I'm not
saying we give up, but we need to keep in mind that we're often better
off trading a bit on the constant synthesis if doing so helps code where
those constants get used.
>
>
> (2) Now on to the nuances as to why things get progressively worse after
> commit 0530254413f8.
>
> It all seems to get down to register allocation passes:
>
> sched1 before 0530254413f8
>
> ;; 0--> b 0: i 22 r140=0x1010000 :alu
> ;; 1--> b 0: i 20 r137=0x1010000 :alu
> ;; 2--> b 0: i 23 r140=r140+0x101 :alu
> ;; 3--> b 0: i 21 r137=r137+0x101 :alu
> ;; 4--> b 0: i 24 r140=r140<<0x20 :alu
> ;; 5--> b 0: i 25 r136=r137 :alu
> ;; 6--> b 0: i 8 r136=asm_operands :nothing
> ;; 7--> b 0: i 17 a0=r136|r140 :alu
> ;; 8--> b 0: i 18 use a0 :nothing
>
> sched1 with 0530254413f8
>
> ;; 0--> b 0: i 22 r144=0x1010000 :alu
> ;; 1--> b 0: i 20 r143=0x1010000 :alu
> ;; 2--> b 0: i 23 r145=r144+0x101 :alu
> ;; 3--> b 0: i 21 r137=r143+0x101 :alu
> ;; 4--> b 0: i 24 r140=r145<<0x20 :alu
> ;; 5--> b 0: i 25 r136=r137 :alu
> ;; 6--> b 0: i 8 r136=asm_operands :nothing
> ;; 7--> b 0: i 17 a0=r136|r140 :alu
> ;; 8--> b 0: i 18 use a0 :nothing
>
> The insn stream is same, only differences being registers reuse (due to
> splitter restriction) vs. not.
>
> Next IRA, for reasons I don't understand (and not brave enough yet to
> dive into) decides to regenerate const_int.
Sure. It's pretty standard practice. When it finds a register that has
a known re-synthesizable value it will often replace the register with
the value. It can help in cases where register pressure it excessively
high by reducing the range of the register holding the value.
> And my guess is, this being so late in the game that it gets
> rematerialized as is in the end, causing the duplicity.
Yup. Though there is a post-reload CSE pass. It's pretty limited in
what it can do and register assignments often make it impossible to do
anything, but it's worth looking at to see if why it's not helping here.
> FWIW, IRA for latter case only, emits additional REG_EQUIV notes which
> could also be playing a role.
REG_EQUAL notes get promoted to REG_EQUIV notes in some cases. And when
other equivalences are discovered it may create a REG_EQUIV note out of
thin air.
The REG_EQUIV note essentially means that everywhere the register occurs
you can validly (from a program semantics standpoint) replace the
register with the value. It might require reloading, but it's a valid
semantic transformation which may reduce register pressure -- especially
for constants that were subject to LICM.
Contrast to REG_EQUAL which creates an equivalence at a particular point
in the IL, but the equivalence may not hold elsewhere in the IL.
>
> I naively tried to gate mvconst_internal to !reload_completed, but that
> triggered some ICE.
I wouldn't expect that to help here.
I would first start to see if using cse_not_expected in the splitter
pattern. I would also look at reload_cse_regs which should give us
some chance at seeing the value reuse if/when IRA/LRA muck things up.
jeff
next prev parent reply other threads:[~2023-06-07 23:44 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-06-02 2:38 Vineet Gupta
2023-06-07 23:44 ` Jeff Law [this message]
2023-06-12 19:32 ` Vineet Gupta
2023-06-17 22:13 ` Jeff Law
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=80218276-fc3c-032b-eaa7-0a4b0e8a859f@gmail.com \
--to=jeffreyalaw@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=gcc@gcc.gnu.org \
--cc=gnu-toolchain@rivosinc.com \
--cc=kito.cheng@gmail.com \
--cc=palmer@rivosinc.com \
--cc=pinskia@gmail.com \
--cc=vineetg@rivosinc.com \
/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).