The LLVM Compiler Infrastructure
Site Map:
Download!
Search this Site


Useful Links
Release Emails
19.1.1: Oct 2024
19.1.0: Sep 2024
18.1.8: Jun 2024
18.1.7: Jun 2024
18.1.6: May 2024
18.1.5: May 2024
18.1.4: Apr 2024
18.1.3: Apr 2024
18.1.2: Mar 2024
18.1.1: Mar 2024
18.1.0: Mar 2024
17.0.6: Nov 2023
17.0.5: Nov 2023
17.0.4: Oct 2023
17.0.3: Oct 2023
17.0.2: Oct 2023
17.0.1: Sep 2023
All Announcements

Maintained by the
llvm-admin team
LLVM Performance Workshop at CGO
  • What: LLVM Performance Workshop at CGO
  • When: Saturday February 4th, 2017
  • Where: Austin, Texas, USA

An LLVM Performance Workshop will be held at CGO 2017. The workshop is co-located with CC, HPCA, and PPoPP. If you are interested in attending the workshop, please register at the CGO website.

Program

The workshop takes place at the Hilton Hotel in downtown Austin (500 East 4th St).

Update: If you indicated this morning that you wanted to join us for dinner, here's the location of the restaurant: Manuel's Downtown, 310 Congress Avenue, Austin, TX 78701. We have a reservation at 5pm (dinner is at your own expense). The restaurant is within walking distance from the hotel.

Time Room Speaker Title  
7:30-8:30 616AB Breakfast
    Session 1: Parallel Code Generation
8:30am 400/402 Johannes Doerfert (Saarland University) Polyhedral "Driven" Optimizations on Real Codes [Abstract] [Slides]
9:00am 400/402 Tobias Grosser (ETH Zurich) Polly-ACC - Accelerator support with Polly-ACC [Abstract] [Slides]
9:30am 400/402 Tao Schardl and William Moses (MIT) The Tapir Extension to LLVM's Intermediate Representation for Fork-Join Parallelism [Abstract] [Slides]
10:00-10:30 616AB Break
    Session 2: Performance in Libraries and Languages
10:30am 400/402 Hal Finkel (Argonne National Laboratory) Modeling restrict-qualified pointers in LLVM [Abstract] [Slides]
11am 400/402 Pranav Bhandarkar, Anshuman Dasgupta, Ron Lieberman, Dan Palermo (Qualcomm Innovation Center) Dillon Sharlet and Andrew Adams (Google) Halide for Hexagon DSP with Hexagon Vector eXtensions (HVX) using LLVM [Abstract] [Slides]
11:30am 400/402 Aditya Kumar and Sebastian Pop (Samsung Austin R&D Center) Performance analysis of libcxx [Abstract] [Slides]
12:00-1:30   Lunch
    Session 3: Whole-application performance tuning
1:30pm 400/402 Brian Railing (CMU) Improving LLVM Instrumentation Overheads [Abstract] [Slides]
2pm 400/402 Sergei Larin, Harsha Jagasia and Tobias Edler von Koch (Qualcomm Innovation Center) Impact of the current LLVM inlining strategy on complex embedded application memory utilization and performance [Abstract] [Slides]
2:30pm 400/402 Mehdi Amini (Apple) LTO/ThinLTO BoF [Abstract]
3:00-3:30 616AB Break
    Session 4: Backend optimizations
3:30pm 400/402 Krzysztof Parzyszek (Qualcomm Innovation Center) Register Data Flow framework [Abstract] [Slides]
4pm 400/402 Evandro Menezes, Sebastian Pop and Aditya Kumar (Samsung Austin R&D Center) Efficient clustering of case statements for indirect branch predictors [Abstract] [Slides]
4:30pm   Workshop ends.

Abstracts

  • Krzysztof Parzyszek: Register Data Flow framework

    Register Data Flow is a framework implemented in LLVM that enables data-flow optimizations on machine IR after register allocation. While most of the data-flow optimizations on machine IR take place during the SSA phase, when virtual registers obey the static single assignment form, passes like pseudo-instruction expansion or frame index replacement may expose opportunities for further optimizations. At the same time, data-flow analysis is much more complicated after register allocation, and implementing compiler passes that require it may not seem like a worthwhile investment. The intent of RDF is to abstract this analysis and provide access to it through a familiar and convenient interface.

    The central concept in RDF is a data-flow graph, which emulates SSA. In contrast to the SSA-based optimization phase where SSA is a part of the program representation, the RDF data-flow graph is a separate, auxiliary structure. It can be built on demand and it does not require any modifications to the program. Traversal of the graph can provide information about reaching definitions of any given register access, as well as reached definitions and reached uses for register definitions. The graph provides connections for easily locating the corresponding elements of the machine IR. A utility class that recalculates basic block live-in information is implemented to make writing whole-function optimizations easier. In this talk, I will give an overview of RDF and its use in the Hexagon backend.

  • Tao Schardl and William Moses: The Tapir Extension to LLVM's Intermediate Representation for Fork-Join Parallelism

    This talk explores how fork-join parallelism, as supported by dynamic-multithreading concurrency platforms such as Cilk and OpenMP, can be embedded into a compiler's intermediate representation (IR). Mainstream compilers typically treat parallel linguistic constructs as syntactic sugar for function calls into a parallel runtime. These calls prevent the compiler from performing optimizations across parallel control flow. Remedying this situation, however, is generally thought to require an extensive reworking of compiler analyses and code transformations to handle parallel semantics.

    Tapir is a compiler IR that represents logically parallel tasks asymmetrically in the program's control flow graph. Tapir allows the compiler to optimize across parallel control flow with only minor changes to its existing analyses and code transformations. To prototype Tapir in the LLVM compiler, for example, we added or modified approximately 5000 lines of LLVM's approximately 3-million-line codebase. Tapir enables many traditional compiler optimizations for serial code, including loop-invariant-code motion, common-subexpression elimination, and tail-recursion elimination, to optimize across parallel control flow, as well as purely parallel optimizations.

    This work was conducted in collaboration with Charles E. Leiserson. The proposal is a preliminary copy of our paper on Tapir, which will appear at PPoPP 2017. This talk will focus on the technical details of implementing Tapir in LLVM.

  • Aditya Kumar, Sebastian Pop, and Laxman Sole: Performance analysis of libcxx

    We will discuss the improvements and future work on libcxx. This includes the improvements on standard library algorithms like string::find and basic_streambuf::xsgetn. These algorithms were suboptimal and we got huge improvements after optimizing them. Similarly, we enabled the inlining of constructor and destructor of std::string. We will present a systematic analysis of function attributes in libc++ and the places where we added missing attributes. We will present a comparative analysis of clang-libc++ vs. gcc-libstdc++ on representative benchmarks. Finally we will talk about our contributions to google-benchmark, which comes with libc++, to help keep track of performance regressions.

  • Hal Finkel: Modeling restrict-qualified pointers in LLVM

    It is not always possible for a compiler to statically determine enough about the pointer-aliasing properties of a program, especially for functions which need to be considered in isolation, to generate the highest-performance code possible. Multiversioning can be employed but its effectiveness is limited by the combinatorially-large number of potential configurations. To address these practical problems, the C standard introduced the restrict keyword which can adorn pointer variables. The restrict keyword can be used by the programmer to convey pointer-aliasing information to the optimizer. Often, this is information that is difficult or impossible for the optimizer to deduce on its own.

    The semantics of restrict, however, are subtle and rely on source-level constructs that are not generally represented within LLVM's IR. Maximally maintaining the aliasing information correctly in the face of function inlining and other code-motion transformations, without interfering with those transformations, is not trivial. While LLVM has long used strict-qualified pointers that are function arguments, and an initial phase of this work provided a way to preserve this information in the face of function inlining, I'll describe a new scheme in LLVM that allows the representation of aliasing information from block-local restrict-qualified pointers as well. This more-general class of restrict-qualified pointers is widely used in scientific code.

    In this talk, I'll cover the use cases for restrict-qualified pointers, the difficulties in representing their semantics at the IR level, why the existing aliasing metadata cannot represent restrict-qualified pointers effectively, how the proposed representation allows the preservation of these semantics with minimal impact to the optimizer, and how the optimizer can use this information to generate higher-performance code. I'll also discuss how this scheme relates to others related to pointer variables (e.g. TBAA and alignment assumptions).

  • Mehdi Amini: LTO/ThinLTO BoF

    LTO is an important technique for getting the maximum performance from the compiler. We presented the ThinLTO model and implementation in LLVM at the last LLVM Dev Meeting. This provided the audience with a good overview of the high-level flow of ThinLTO and the 3-phases split involved.

    The proposal for this BoF is to gather and discuss the existing user-experience, the current limitations and what features folks are expecting the most out of ThinLTO. We can go over the current optimizations currently in development upstream.

  • Johannes Doerfert: Polyhedral "Driven" Optimizations on Real Codes

    In this talk I will present polyhedral "driven" optimizations on real codes. The term polyhedral "driven" is used as there are two flavors of optimization I want to discuss (depending on my progress and the duration of the talk).

    The first follows the classical approach applied by LLVM/Polly but with special consideration of general benchmarks like SPEC. I will show how LLVM/Polly can be used to perform beneficial optimizations in (at least) libquantum, hmmer, lbm and bzip2. I will also discuss what I think is needed to identify such optimization opportunities automatically.

    The second polyhedral driven optimization I want to present is a conceptual follow-up of the "Polyhedral Info" GSoC project. This project was the first try to augment LLVM analysis and transformation passes with polyhedral information. While the project was build on top of LLVM/Polly, I will present an alternative approach. First I will introduce a modular, demand driven and caching polyhedral program analysis that natively integrates into the existing LLVM pipeline. Then I will show how to utilize this analysis in existing LLVM optimizations to improve performance. Finally, I will use the polyhedral analysis to derive new, complex control flow optimizations that are not, or only in a simpler form, present in LLVM.

  • Tobias Grosser: Polly-ACC - Accelerator support with Polly-ACC

    Programming today's increasingly complex heterogeneous hardware is difficult, as it commonly requires the use of data-parallel languages, pragma annotations, specialized libraries, or DSL compilers. Adding explicit accelerator support into a larger code base is not only costly, but also introduces additional complexity that hinders long-term maintenance. We propose a new heterogeneous compiler that brings us closer to the dream of automatic accelerator mapping. Starting from a sequential compiler IR, we automatically generate a hybrid executable that - in combination with a new data management system - transparently offloads suitable code regions. Our approach is almost regression free for a wide range of applications while improving a range of compute kernels as well as two full SPEC CPU applications. We expect our work to reduce the initial cost of accelerator usage and to free developer time to investigate algorithmic changes.

  • Brian Railing: Improving LLVM Instrumentation Overheads

    The behavior and structure of a shared-memory parallel program can be characterized by a task graph that encodes the instructions, memory accesses, and dependencies of each piece of parallel work. The task graph representation can encode the actions of any threading library and is agnostic to the target architecture. Contech [1] is an LLVM-based tool that generates a task graph representation, by instrumenting the program when it is compiled such that it ultimately outputs a task graph when executed. This paper describes several approaches to improving the overhead of Contech's instrumentation by augmenting the static compiler analysis.

    The additional analyses are able to first determine similar memory address calculations in the LLVM intermediate representation and elide them from the instrumentation to reduce the data recorded, an approach only previously attempted with dynamic binary instrumentation based on common registers [2] [3]. Second, this analysis is supplemented by performing tail duplication which increases the memory operations in a single basic block and therefore may provide further opportunities to elide instrumentation, without compromising the accuracy or detail of the data recorded. These optimizations reduce the data recorded by 22%, which has a proportionate decrease in overhead from 3.7x to 3.3x for PARSEC benchmarks.

    [1] B. P. Railing, E. R. Hein, and T. M. Conte. "Contech: Efficiently Generating Dynamic Task Graphs for Arbitrary Parallel Programs". In: ACM Trans. Archit. Code Optim. 12.2 (July 2015), 25:1-25:24.

    [2] Q. Zhao, I. Cutcutache, and W.-F. Wong. "Pipa: Pipelined Profiling and Analysis on Multi-core Systems". In: Proceedings of the 6th Annual IEEE/ACM International Symposium on Code Generation and Optimization. CGO '08. Boston, MA, USA: ACM, 2008, pp. 185-194.

    [3] K. Jee et al. "ShadowReplica: Efficient Parallelization of Dynamic Data Flow Tracking". In: Proceedings of the 2013 ACM SIGSAC Conference on Computer & Communications Security. CCS '13. Berlin, Germany: ACM, 2013, pp. 235-246.

  • Evandro Menezes, Sebastian Pop, and Aditya Kumar: Efficient clustering of case statements for indirect branch predictors

    We present an O(nlogn) algorithm as implemented in LLVM to compile a switch statement into jump tables. To generate jump tables that can be efficiently predicted by current hardware branch predictors, we added an upper bound on the number of entries in each generated jump table. This modification of the previously best known algorithm reduces the complexity from O(n^2) to O(nlogn). We illustrate the performance achieved by the improved algorithm on the Samsung Exynos-M1 processor running several benchmarks.

  • Pranav Bhandarkar, Anshuman Dasgupta, Ron Lieberman, Dan Palermo, Dillon Sharlet, and Andrew Adams: Halide for Hexagon DSP with Hexagon Vector eXtensions (HVX) using LLVM

    Halide is a domain specific language that endeavors to make it easier to construct large and composite image processing applications. Halide is unique in its design approach to decoupling the algorithm from the organization (schedule) of the computation. Algorithms once written and tested for correctness can then be continually tuned for performance as Halide allows for easily changing the schedule - tiling, parallelizing, prefetching or vectorizing different dimensions of the loop nest that form the structure of the algorithm.

    Halide programs are transformed into the Halide Intermediate Representation (IR) by the Halide compiler. This IR is analyzed and optimized before generating LLVM bitcode for the target requested. Halide links with the LLVM optimizer and codegen libraries for supported targets, and uses these to generate object code.

    In this workshop, we will present our work on retargeting Halide to the Hexagon DSP with focus on the Hexagon Vector eXtensions (HVX).

    Our workshop will present the halide constructs used in a simple blur 5x5, the corresponding Halide IR, and a few of the important LLVM Hexagon passes which generate HVX vector instructions.

    We will demonstrate compilation using LLVM.org and Halide.org tools, and execution of the blur 5x5 pipeline on a Snapdragon 820 development board using the Halide Hexagon offloader. In particular we will demonstrate the various improvements which can be realized with scheduling changes.

  • Sergei Larin, Harsha Jagasia and Tobias Edler von Koch: Impact of the current LLVM inlining strategy on complex embedded application memory utilization and performance

    Sophisticated embedded applications with extensive and fine degree of memory management are presenting a unique challenge to contemporary tool chains. Like many open source projects LLVM optimizes its core optimization tradeoffs for common cases and a set of common architectures. Even with back end specific hooks, it is not always possible to exert appropriate degree of control over some key optimizations. We propose a case study on "in-depth" analysis of LLVM PGO assisted inlining in a complex embedded application.

    The program in question is a large scale embedded networking application designed to be custom tuned for a variety of actual embedded platforms with a range of memory and performance constrains. It makes a high use of linker scripts to configure and fine tune memory assignment to ultimately guarantee optimal performance in constrained memory environment while being extremely power conscious.

    The moment a tool chain is addressing a non-uniform memory model, "one size fits all" approach to optimizations like inlining stops being optimal. For instance, based on section assignment, completely unknown to the compiler, inlining takes place in areas that are facing different cost/benefit tradeoffs. The content of L1 and L2 Icache should not be "enlarged" even if performance can theoretically improve. Inlining across such section boundaries are also ill-advisable, since control flow exchange (jump) between sections destined to different levels of memory hierarchy can produce unexpected performance implications. Finally, tightly budgeted low-level and high-performance memories might swell beyond their physical limits.

    The current state of LLVM inline is somewhat transitional in anticipation of structural updates to the pass manager, and as such it still strongly relies on heuristic + PGO based inline cost computation. In such situation the introduction of back-end hooks might allow targets to fine-tune inlining decisions to some degree but they still fall far short to the degree of control needed by the above-described systems. Additional challenge is posed by high degree of complexity to capture actual system run-time behavior, and even collecting appropriate traces to generate meaningful PGO data. Battery powered embedded chips rarely have sophisticated tracing capabilities, yet present extremely complex run time environments.

Call for Speakers

We invite speakers from academia and industry to present their work on the following list of topics (including and not limited to:)

  • improving performance and size of code generated by LLVM,
  • improving performance of LLVM's runtime libraries,
  • tools developed with LLVM for performance analysis of compiler generated code,
  • bots and trackers of performance over time,
  • improving the security of generated code,
  • any other topic related to improving and maintaining the performance and quality of LLVM generated code.
While the primary focus of the workshop is on these topics, we welcome any submission related to the LLVM compiler infrastructure, its sub-projects (Clang, Linker, libraries), and its use in industry and academia.

We are looking for:

  • keynote speakers,
  • technical presentations: 30 minutes plus questions and discussion,
  • tutorials,
  • BOFs.

Proposals should provide enough information for the review committee to be able to judge the quality of the submission. Proposals can be submitted under the form of an extended abstract, full paper, or slides. Proposals should be submitted to Easychair LLVM-CGO 2017. The deadline for receiving submissions is December 1st, 2016. Speakers will be notified of acceptance or rejection by December 15.

Workshop organization: Sebastian Pop, Aditya Kumar, Tobias Edler von Koch, and Tanya Lattner.