Google Data

With new year upon us, I thought I would take a minute to update everyone on the great progress made in 2009 by the LLVM Project. For those not familiar with LLVM project, it's a cross platform complier infrastructure. We had two successful releases of LLVM, our first release of Clang, a compiler front end for various C languages, and we held our third annual Developers' meeting on October 2, 2009. This meeting is an opportunity for both LLVM and Clang developers to have a face to face meeting, exchange ideas, share their experiences and work together on LLVM or Clang.

This Developers' Meeting was the largest to date! We had 170 attendees with a huge range of different academic and company affiliations. The Developers' meeting was structured to have both a general overview about the major LLVM subsystems and also applications of LLVM for various projects. The LLVM subsystems talks included Clang, scalar evolution and loop optimization, the future of the LLVM register allocator, and a tutorial on building a backend for LLVM. We had many talks on applications of LLVM or Clang, such as OpenCL, Unladen Swallow (a Google sponsored project), Rubinius, and so much more! Community members gave a total of 17 technical presentations on LLVM and its applications, and you can view all the slides and videos from the talks on the Developers' Meeting website.

This event is not possible without the support of our sponsors. Google generously helped fund several students and active members of the LLVM community to attend the meeting and present their LLVM-related work. I'd like to briefly summarize the work by some of these developers and students:

Anton Korobeynikov

Anton is a long time developer for the LLVM project, LLVM's project administrator for Google Summer of Code™ (GSoC) and also an LLVM code owner. For his day job, he is a Ph.D student in applied statistics at Saint Petersburg State University, Russia.

Anton presented an invaluable Tutorial on Building a Backend in 24 Hours. His tutorial overviews the various code generation phases, such as SelectionDAG, Register Allocation, and post register allocation. He goes into the different pieces of the backend that one will need to implement such as the target, subtarget, lowering, register set, instruction selection, and the calling convention. If you have ever wanted to write a new backend for LLVM, this is the talk that you will want to see. (Slides, Video)

Bruno Cardoso Lopes

Bruno is a multi-year participant with the GSoC project, active LLVM contributor, and Ph.D. student at University of Campinas, Brazil. This year, he presented Object Code Emission and llvm-mc. His talk gave a high level overview of the LLVM Machine Code Emitter and focused on the emission of object files. The motivation behind direct object code emission is to bypass the external assembler, and speed up compile time.

His work is a part of the LLVM Machine Code (MC) Toolkit project. The project aims to build better tools for dealing with machine code, object file formats, etc. The idea is to generate most of the target specific assemblers and disassemblers from existing LLVM target .td files and to build an infrastructure for reading and writing common object file formats. His talk goes into the details regarding the design, current implementation status, and future directions. (Slides, Video)

Duncan Sands

Duncan is also a long time developer for the LLVM project and an LLVM Code Owner. His presentation was titled Reimplementing llvm-gcc as a gcc plugin. His project, DragonEgg, aims to replace gcc's optimizers and code generators with those in LLVM without modifying gcc at all. This is done using a plugin via mainline gcc's new ability to load additional logic and passes at runtime via a plug-in mechanism.

The plugin is a shared library that is loaded by gcc-4.5 at runtime and is currently under development. If you are interested in helping with this project, please see the DragonEgg website for more information. Take a look at the slides and video from Duncan's presentation to see where you can get started.

Santosh Nagarakatte

Santosh is currently a Ph.D. student at the University of Pennsylvania. He presented SoftBound, one of his current research projects. SoftBound is a compile-time transformation for enforcing spatial safety of C. It works by recording base and bound information for every pointer as disjoint metadata. It is a software-only approach and performs metadata manipulation only when loading or storing pointer values. The advantage of this approach is that it provides spatial safety without requiring changes to C source code. His talk provides a brief description of the formal proof and LLVM implementation. (Slides, Video)

If you are interested in attending one of our future Developers' meetings, please join the LLVM-announce mailing list. Many thanks again to Google's Open Source Programs' Office for making this event possible!

By The LLVM Team