llvm dev - Feb 2021 - Fwd: Superoptimization for RISC-V: What is the state of the art now?

Hi all,

I appreciate it if you can drop some information/papers/open source project
urls that are related to superoptimization. Although I am targeting RISC-V
ISA, there might be some other research works / open source projects
available in the LLVM community.

(Sorry for the duplicate email.)

---------- Forwarded message ---------
From: Wei Wu (吴伟) <lazyparser at gmail.com>
Date: Wed, Feb 17, 2021 at 12:26 PM
Subject: Superoptimization for RISC-V: What is the state of the art now?
To: RISC-V SW Dev <sw-dev at groups.riscv.org>


Hi all,

In the last code optimization meeting (formly code-size-reduction and
code-speed-opt), the idea of using superoptimization[4] for RISC-V had been
discussed. I'm interested in this area. After a quick search I found very
few results, though. GNU/Embecosm had a superopt [1] but I am not sure it
still works either for GCC 10 or RISC-V backend. Google open sourced a
LLVM-IR level tool named souper[2] would help. STOKE[3] is yet another
optimizer which targets x86 only.

I appreciate it if you can drop some information/papers/open source project
urls that are related to superoptimization. I am going to stand on the
basis of gnu-superopt[1], and not sure it is the right/effective way to go.

Thanks.

[1] https://github.com/embecosm/gnu-superopt
[2] https://github.com/google/souper
[3] https://github.com/StanfordPL/stoke
[4] https://en.wikipedia.org/wiki/Superoptimization

-- 
Best wishes,
Wei Wu (吴伟)


-- 
Best wishes,
Wei Wu (吴伟)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210217/f4205734/attachment.html>

STOKE & Souper are the most well supported & maintained
superoptimization tools out there. They are very different: STOKE does genetic
search, while Souper uses only SMT solvers for (symbolic) search.

There’s also Rosette, which does a hybrid of concrete & symbolic search
(mixing concrete search with other techniques seems to be the way to go).

 

I would also look at the SyGuS competition (syntax-guided synthesis). These
tools, such as CVC4, can be used for superoptimization as well.

 

In terms of tools, I think that’s it. In terms of papers, there are plenty of
research groups working on synthesis in general.

 

Nuno

 

 

From: Wei Wu

Sent: 17 February 2021 09:54
To: llvm-dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] Fwd: Superoptimization for RISC-V: What is the state of the
art now?

 

Hi all,

 

I appreciate it if you can drop some information/papers/open source project urls
that are related to superoptimization. Although I am targeting RISC-V ISA, there
might be some other research works / open source projects available in the LLVM
community.

 

(Sorry for the duplicate email.)

 

---------- Forwarded message ---------
From: Wei Wu (吴伟) <lazyparser at gmail.com <mailto:lazyparser at
gmail.com> >
Date: Wed, Feb 17, 2021 at 12:26 PM
Subject: Superoptimization for RISC-V: What is the state of the art now?
To: RISC-V SW Dev <sw-dev at groups.riscv.org <mailto:sw-dev at
groups.riscv.org> >

 

Hi all,

 

In the last code optimization meeting (formly code-size-reduction and
code-speed-opt), the idea of using superoptimization[4] for RISC-V had been
discussed. I'm interested in this area. After a quick search I found very
few results, though. GNU/Embecosm had a superopt [1] but I am not sure it still
works either for GCC 10 or RISC-V backend. Google open sourced a LLVM-IR level
tool named souper[2] would help. STOKE[3] is yet another optimizer which targets
x86 only.

 

I appreciate it if you can drop some information/papers/open source project urls
that are related to superoptimization. I am going to stand on the basis of
gnu-superopt[1], and not sure it is the right/effective way to go.

 

Thanks.

 

[1] https://github.com/embecosm/gnu-superopt

[2] https://github.com/google/souper

[3] https://github.com/StanfordPL/stoke

[4] https://en.wikipedia.org/wiki/Superoptimization

 

-- 

Best wishes,
Wei Wu (吴伟)

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210217/5e125df0/attachment.html>

Hi Wei,

Back in 2018 I worked in a commercial superoptimizer for RISCV - S10,
which I presented as a poster in the RISCV Workshop in Barcelona:
https://youtu.be/ylTA63vPHYU?t=1969
https://linki.tools/s10.html

To be fair, this was mostly used for consultancy into specific hot code
blocks. Since then I moved on to other things but still hold a keen
interest in Superoptimization.

S10 was built upon greenthumb:
https://github.com/mangpo/greenthumb

I have been meaning to open source S10 but writing the documentation,
cleaning up the code, etc is taking quite a bit of time although it's
one of my priorities for 2021.

In any case, if you want to know more or if I can help any further, feel
free to drop me an email.

--
Paulo Matos

Wei Wu (吴伟) via llvm-dev writes:
> Hi all,
>
> I appreciate it if you can drop some information/papers/open source project
> urls that are related to superoptimization. Although I am targeting RISC-V
> ISA, there might be some other research works / open source projects
> available in the LLVM community.
>
> (Sorry for the duplicate email.)
>
> ---------- Forwarded message ---------
> From: Wei Wu (吴伟) <lazyparser at gmail.com>
> Date: Wed, Feb 17, 2021 at 12:26 PM
> Subject: Superoptimization for RISC-V: What is the state of the art now?
> To: RISC-V SW Dev <sw-dev at groups.riscv.org>
>
>
> Hi all,
>
> In the last code optimization meeting (formly code-size-reduction and
> code-speed-opt), the idea of using superoptimization[4] for RISC-V had been
> discussed. I'm interested in this area. After a quick search I found
very
> few results, though. GNU/Embecosm had a superopt [1] but I am not sure it
> still works either for GCC 10 or RISC-V backend. Google open sourced a
> LLVM-IR level tool named souper[2] would help. STOKE[3] is yet another
> optimizer which targets x86 only.
>
> I appreciate it if you can drop some information/papers/open source project
> urls that are related to superoptimization. I am going to stand on the
> basis of gnu-superopt[1], and not sure it is the right/effective way to go.
>
> Thanks.
>
> [1] https://github.com/embecosm/gnu-superopt
> [2] https://github.com/google/souper
> [3] https://github.com/StanfordPL/stoke
> [4] https://en.wikipedia.org/wiki/Superoptimization
>
> -- 
> Best wishes,
> Wei Wu (吴伟)

-- 
Paulo Matos