llvm dev - Oct 2020 - Using polly to change loop schedules in an llvm pass

Hello All

I am looking into using polly, in conjunction with a set of passes that I am
writing, to modify loop schedules/transform loops. I was wondering if there is a
way to interface with polly directly through an LLVM pass? It seems by looking
at the documentation, that the way to modify loop schedules is to generate the
jscop files, modify them, and then reload them with polly, which is not a very
“friendly” mechanism to do automatically in a pass. Is my understanding of the
usage of polly correct here, or am I missing something?

Thanks!
-Adel Ejjeh
--
Adel Ejjeh
PhD Candidate | Computer Science
University of Illinois at Urbana Champaign
Email: aejjeh at illinois.edu<mailto:aejjeh at illinois.edu> | adel.ejjeh
at gmail.com<mailto:adel.ejjeh at gmail.com>
Zoom: https://illinois.zoom.us/my/aejjeh

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201022/137b18ff/attachment.html>

Hey Adel,

You don't have to go through json per-se but Polly's design also
wasn't
for the use case you describe.
I cc'ed Michael who has some extensions (I think) that allow the 
frontend to decide how to modify the schedule.
You should be able to piggy back on that either through IR annotations 
or direct calls of the appropriate functions
in Polly that modify the schedule tree.

Best of luck :)

~ Johannes



On 10/22/20 11:18 AM, Ejjeh, Adel via llvm-dev wrote:
> Hello All
>
> I am looking into using polly, in conjunction with a set of passes that I
am writing, to modify loop schedules/transform loops. I was wondering if there
is a way to interface with polly directly through an LLVM pass? It seems by
looking at the documentation, that the way to modify loop schedules is to
generate the jscop files, modify them, and then reload them with polly, which is
not a very “friendly” mechanism to do automatically in a pass. Is my
understanding of the usage of polly correct here, or am I missing something?
>
> Thanks!
> -Adel Ejjeh
> --
> Adel Ejjeh
> PhD Candidate | Computer Science
> University of Illinois at Urbana Champaign
> Email: aejjeh at illinois.edu<mailto:aejjeh at illinois.edu> |
adel.ejjeh at gmail.com<mailto:adel.ejjeh at gmail.com>
> Zoom: https://illinois.zoom.us/my/aejjeh
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Hi,

for passes that apply to SCoPs I would suggest to write a ScopPass
within polly. Importing/exporting a jscop file would be more something
for external tools. The reason is that from a LLVM loop- or
function-pass it is non-obvious what a SCoP (which is a loop nest that
fulfils the requirements for Polly's analysis to work) will be. A
ScopPass will just be invoked on previously detected SCoPs, can modify
them, and Polly will re-generate LLVM-IR again when all ScopPasses
have been applied.

Michael



Am Do., 22. Okt. 2020 um 11:18 Uhr schrieb Ejjeh, Adel via llvm-dev
<llvm-dev at lists.llvm.org>:
>
> Hello All
>
>
>
> I am looking into using polly, in conjunction with a set of passes that I
am writing, to modify loop schedules/transform loops. I was wondering if there
is a way to interface with polly directly through an LLVM pass? It seems by
looking at the documentation, that the way to modify loop schedules is to
generate the jscop files, modify them, and then reload them with polly, which is
not a very “friendly” mechanism to do automatically in a pass. Is my
understanding of the usage of polly correct here, or am I missing something?
>
>
>
> Thanks!
>
> -Adel Ejjeh
>
> --
>
> Adel Ejjeh
>
> PhD Candidate | Computer Science
>
> University of Illinois at Urbana Champaign
>
> Email: aejjeh at illinois.edu | adel.ejjeh at gmail.com
>
> Zoom: https://illinois.zoom.us/my/aejjeh
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

I see. Thank you Michael, I will look into that. I'll reach out again if I
have any other questions.

Regards
-Adel

-----Original Message-----
From: Michael Kruse <llvmdev at meinersbur.de> 
Sent: Thursday, October 22, 2020 1:15 PM
To: Ejjeh, Adel <aejjeh at illinois.edu>
Cc: LLVM Dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] Using polly to change loop schedules in an llvm pass

Hi,

for passes that apply to SCoPs I would suggest to write a ScopPass within polly.
Importing/exporting a jscop file would be more something for external tools. The
reason is that from a LLVM loop- or function-pass it is non-obvious what a SCoP
(which is a loop nest that fulfils the requirements for Polly's analysis to
work) will be. A ScopPass will just be invoked on previously detected SCoPs, can
modify them, and Polly will re-generate LLVM-IR again when all ScopPasses have
been applied.

Michael



Am Do., 22. Okt. 2020 um 11:18 Uhr schrieb Ejjeh, Adel via llvm-dev
<llvm-dev at lists.llvm.org>:
>
> Hello All
>
>
>
> I am looking into using polly, in conjunction with a set of passes that I
am writing, to modify loop schedules/transform loops. I was wondering if there
is a way to interface with polly directly through an LLVM pass? It seems by
looking at the documentation, that the way to modify loop schedules is to
generate the jscop files, modify them, and then reload them with polly, which is
not a very “friendly” mechanism to do automatically in a pass. Is my
understanding of the usage of polly correct here, or am I missing something?
>
>
>
> Thanks!
>
> -Adel Ejjeh
>
> --
>
> Adel Ejjeh
>
> PhD Candidate | Computer Science
>
> University of Illinois at Urbana Champaign
>
> Email: aejjeh at illinois.edu | adel.ejjeh at gmail.com
>
> Zoom: https://illinois.zoom.us/my/aejjeh
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Hello Michael

Is there any good resources you can point me to for a better understanding of
the polly internals, something like LLVM’s “Programmer’s Guide” for instance?

Thanks
-Adel

From: Michael Kruse <llvmdev at meinersbur.de>
Date: Thursday, October 22, 2020 at 1:16 PM
To: Ejjeh, Adel <aejjeh at illinois.edu>
Cc: LLVM Dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] Using polly to change loop schedules in an llvm pass
Hi,

for passes that apply to SCoPs I would suggest to write a ScopPass
within polly. Importing/exporting a jscop file would be more something
for external tools. The reason is that from a LLVM loop- or
function-pass it is non-obvious what a SCoP (which is a loop nest that
fulfils the requirements for Polly's analysis to work) will be. A
ScopPass will just be invoked on previously detected SCoPs, can modify
them, and Polly will re-generate LLVM-IR again when all ScopPasses
have been applied.

Michael



Am Do., 22. Okt. 2020 um 11:18 Uhr schrieb Ejjeh, Adel via llvm-dev
<llvm-dev at lists.llvm.org>:
>
> Hello All
>
>
>
> I am looking into using polly, in conjunction with a set of passes that I
am writing, to modify loop schedules/transform loops. I was wondering if there
is a way to interface with polly directly through an LLVM pass? It seems by
looking at the documentation, that the way to modify loop schedules is to
generate the jscop files, modify them, and then reload them with polly, which is
not a very “friendly” mechanism to do automatically in a pass. Is my
understanding of the usage of polly correct here, or am I missing something?
>
>
>
> Thanks!
>
> -Adel Ejjeh
>
> --
>
> Adel Ejjeh
>
> PhD Candidate | Computer Science
>
> University of Illinois at Urbana Champaign
>
> Email: aejjeh at illinois.edu | adel.ejjeh at gmail.com
>
> Zoom: https://illinois.zoom.us/my/aejjeh
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201023/7f3e8e9d/attachment.html>