[LLVMdev] LLVM Parallel IR

[Herbert Jordan]

Hi William,

as has been mentioned by Kevin, the idea of building an explicit 
parallel IR and associated tools has its merits. With the Insieme 
Project (http://insieme-compiler.org) we went along that road and built 
an explicit parallel high-level IR for analysing, manipulating and 
tuning parallel codes. In particular we managed to design a compact IR 
unifying various different parallel language extensions and APIs, 
including OpenMP, Cilk, OpenCL and -- to some extend -- MPI.

As part of this work we got to believe that in the general case the 
compiler supported coordination and tuning of parallelism in programs is 
beyond the scope of conventional low-level IRs. If you are interested in 
our experiences, concepts, design and results, an overview can be found 
in [1] and a rather extensive description and formalization in the 
associated thesis [2].

Maybe those can provide you some insights regarding potential 
alternative approaches for your work.

Best,
Herbert

[1] http://dl.acm.org/citation.cfm?id=2523721.2523727
[2] http://www.dps.uibk.ac.at/~csaf7445/pub/phd_thesis_jordan.pdf - in 
particular Chapter 3


On 2015-03-17 01:12, Dounia Khaldi wrote:
> Hi William,
>
> Currently, LLVM lowers parallel constructs of OpenMP at the front-end 
> level to runtime calls; although simple, this ad-hoc solution prevents 
> high-level reasoning about parallel languages while possibly 
> introducing semantic inconsistencies in existing sequential program 
> analyses. Worse, users implicitly rely upon the lack of 
> interprocedural analyses within many compilers to preserve the 
> semantics of programs even in the presence of code optimizations.
>
> As part of my PhD work, I worked on introducing a parallel 
> intermediate representation extension for parallel compilers which I 
> called /SPIRE/. SPIRE was successfully applied to automatically 
> parallelize sequential programs within the PIPS source to source 
> compiler. (Here is a link to my PhD thesis: 
> https://tel.archives-ouvertes.fr/pastel-00935483/document).
>
> As part of my Postdoc work at University of Houston, I am currently 
> applying SPIRE to LLVM in order to optimize PGAS languages, 
> specifically OpenSHMEM. The idea is to be able to recognize OpenSHMEM 
> constructs and represent them explicitly in the LLVM IR using SPIRE to 
> take advantage of sequential optimizations offered by LLVM. The 
> ultimate goal is to develop new parallel optimizations based on 
> SPIRE(LLVM IR). The tricky part here is to prove that the enabled 
> optimizations will generate correct codes. I am using SPIRE formal 
> operational semantics I developed to prove the correctness of 
> optimizations performed on parallel programs.
>
> I can send you a submitted paper regarding this work on LLVM if you 
> are interested.
>
> Best,
>
>
> -- 
> Dounia KHALDI
> Postdoctoral Fellow
> University of Houston
> Department of Computer Science
> 501 Philip G. Hoffman Hall
> Houston, TX  77204-3010
>
> On Mon, Mar 9, 2015 at 10:58 AM, William Moses <wmoses at csail.mit.edu 
> <mailto:wmoses at csail.mit.edu>> wrote:
>
>     I'm part of a research group at MIT looking to create an extension
>     of LLVM that inherently allows one to nicely code a parallel loop.
>
>     Most parallel frameworks tend to take the body of a parallel loop
>     and stick it inside of a function for the parallel runtime to call
>     when appropriate. However, this makes optimizations significantly
>     more difficult as most compiler optimizations tend to be
>     intraprocedural. The goal of this parallel IR is to allow LLVM to
>     more easily optimize parallel code -- which often gets the short
>     end of the optimization stick.
>
>     Long story short: I was wondering if anyone in the LLVM community
>     has already begun a similar project or knows of something that
>     begins to accomplish this.
>
>     Additionally, does anyone who has worked with parallelism in LLVM
>     have any specific comments or suggestions regarding this project.
>
>     Sincerely,
>     William Moses
>     MIT Computer Science and Artificial Intelligence Laboratory
>
>
>     _______________________________________________
>     LLVM Developers mailing list
>     LLVMdev at cs.uiuc.edu <mailto:LLVMdev at cs.uiuc.edu>
>     http://llvm.cs.uiuc.edu
>     http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
>
> -- 
> Dounia KHALDI
> Postdoctoral Fellow
> University of Houston
> Department of Computer Science
> 501 Philip G. Hoffman Hall
> Houston, TX  77204-3010
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150409/97f01964/attachment.html>