[llvm-dev] [GSoC] Improve inlining using inter-procedural analysis or profile

Mon Feb 8 12:22:17 PST 2021

On Mon, Feb 8, 2021 at 12:10 PM Stefanos Baziotis via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Hi Liqiang,
>
> Jakub's recommendations are great and they reminded me of one of the talks
> we gave in the last LLVM conf: https://youtu.be/I4Iv-HefknA?t=1081
> In this part about inlining, in the end I have a "further reading" section
> that you might find interesting (note that slides are available in the YT
> comments).
> Some of the papers that Jakub mentioned are there and some more that you
> may find useful.
>
> Hopefully, the rest of the part will help you understand LLVM's inlining a
> little better too. Unfortunately, when I was preparing this part, a lot of
> content had to be thrown
> out as the part ended up being a talk by itself. The reason I'm mentioning
> that is to motivate you to please ask questions if you want! There is a lot
> of content
> about inlining in LLVM that has not been presented anywhere and although
> I'm no expert on inlining, I'll try to help. And I hope the rest of the
> community will do too.
>
> Best,
> Stefanos
>
> Στις Δευ, 8 Φεβ 2021 στις 9:58 μ.μ., ο/η Jakub (Kuba) Kuderski via
> llvm-dev <llvm-dev at lists.llvm.org> έγραψε:
>
>> Hi Liqiang,
>>
>> Here are some papers that may be loosely related to the problem you
>> described, without any particular order:
>>
>> 1. Automatic construction of inlining heuristics using machine learning,
>> Sameer Kulkarni et al., 2013
>> https://dl.acm.org/doi/10.1109/CGO.2013.6495004
>> 2. Adaptive online context-sensitive inlining, Kim Hazelwood, David
>> Grove, 2003
>> http://www.cs.cmu.edu/afs/cs/academic/class/15745-s07/www/papers/hazelwood-cgo03.pdf
>> 3. An Adaptive Strategy for Inline Substitution, Keith Cooper et al.,
>> 2008  https://link.springer.com/chapter/10.1007/978-3-540-78791-4_5
>> 4. Automatic Tuning of Inlining Heuristics, John Cavazos, Michael F.P.
>> O’Boyle, 2005  https://www.eecis.udel.edu/~cavazos/cpc-2006.pdf
>> 5. Towards Better Inlining Decisions Using Inlining Trials, Jeffrey Dean,
>> Craig Chambers, 1994
>> https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.50.1043&rep=rep1&type=pdf
>> 6. Exploiting Hardware Performance Counters with Flow and Context
>> Sensitive Profiling, Glenn Ammons et al., 1997
>> http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.45.2010
>> 7. Boosting Inlining Heuristics, Nuno Lopes, 2010
>> https://web.ist.utl.pt/nuno.lopes/pubs/inline-stats10.pdf
>>
>> -Jakub
>>
>> On Mon, Feb 8, 2021 at 8:26 AM Liqiang Tao via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>> To Whom it May Concern,
>>>
>>> I am Liqiang Tao, a graduate student in the Graduate School of Arts and
>>> Sciences of the University of Tokyo.
>>>
>>> First, I would like to talk about my interests here. When I saw the Open
>>> Projects page, I was attracted by this area.
>>>
>>> > Moreover, we want to explore the case that is helpful to group
>>> strongly related functions together and optimize them as a group instead of
>>> displaying them individually. I consider whether it is possible to inline
>>> some related methods to improve performance.
>>>
>>> Then I found a paper[1]. In this paper, the methods are inlined by
>>> clusters rather than individuals. The profile and IR size are applied to
>>> decide inline or not. I also noticed that the profile directed inlining, as
>>> a todo item, also appears in the Open Projects page. It might be a nice
>>> start, right?
>>>
>>
(Not sure I follow) could you link the Open Projects page? Asking because
the LLVM inliner is profile - driven (hence my confusion)

>
>>> Besides, if possible, I am also interested in applying inter-procedural
>>> analysis or static analysis to guide inlining. Unfortunately, I didn't find
>>> any paper about this idea. Could you please provide me with some
>>> suggestions about this? Is there any process related to these ideas in LLVM?
>>>
>>
In some sense, today's inliner does perform some inter-procedural analysis
(other than the caller/callee of the call site being probed) - see the
"shouldInline" method in lib/Analysis/InlineAdvisor.cpp. One possibility
would be to extend the subgraph analyzed there.

>>> As for my background relevant to this project:
>>> * I am participating in porting V8 to RISC-V from July, 2020.
>>> https://github.com/v8-riscv/v8
>>> * Recently, I began learning the inliner code in LLVM. And I also tried
>>> to create my own pass to be familiar with LLVM infrastructure.
>>>
>>> Looking forward to more discussions,
>>>
>>> Best Wishes,
>>> Liqiang Tao
>>>
>>> [1] An Optimization-Driven Incremental Inline Substitution Algorithm for
>>> Just-In-Time Compilers
>>> https://dl.acm.org/doi/10.5555/3314872.3314893
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>>
>>
>> --
>> Jakub Kuderski
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
> _______________________________________________
> 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/20210208/498e8f21/attachment.html>