Open64: A Framework for High performance Compiler

1
Open64 A Framework for High performance Compiler
March 2007
2
Presenters

• Shin-Ming Liu, HP
• Guang Gao, University of Deleware
• Chunhua Liao, University of Houston
• Robert Hundt, Google
• Soo-Mook Moon, Seoul National University
• Wei Chen, University of California, Berkeley
• Tong Chen, IBM Research
• Pen-Chung Yew, University of Minnisota
• Sun Chan, Simplight Nanoelectronics
• Dong Yuan, Tsinghua University

3
Outline

• Open64 History
• Osprey Project
• Research Activities
• Retargetability

4
Open64 History

• Derived from SGI MIPSpro compiler suite
• SGI did the initial port to Itanium
• SGI open sourced the MIPSpro compilers in Summer
  2000 as the Pro64 compiler suite
• Gnu-front-end based C, C compilers
• Cray-FORTRAN front-end based F95 compiler
• University of Delaware took over maintenance of
  Pro64 compiler suite
• New distribution re-named to be Open64 compiler
  suite
• Retargeted to several machines
• Mips, IA64, X86-64, Ceva, Tensilica, XScale
  (ARM),

5
Open64 History

• Intel enhanced the Open64 compiler code generator
  with advanced optimizations for Itanium
• Itanium-enhanced Open64 compiler suite released
  as the Open Research Compiler (ORC) today to
  facilitate Itanium-related research efforts in
  academia
• Pathscale retargeted Open64 compilers for x86-64
  and seeded new repository to enable further
  development by open source compiler developers

6
Open64 History

• Thus, Open64 compiler technology
• is gnu-compatible since it uses the gcc, g
  front-ends
• incorporates back-end support for Itanium
• is based on mature compiler technology developed
  by SGI with wide production exposure
• Multiple OS bringup, compiler self build,
  numerous large applications Oracle, Pro/E,
  Abacus, Morgan-Stanley, Maya,
• is suitable for computer architecture, compiler
  architecture and compiler research

7
Open64 in the Commercial Space

• STMicro
• Equator Technologies (acquired by Pixelworks)
• Cognigine (acquired by Huawei)
• Tensilica
• CEVA
• Pathscale (acquired by QLogic)
• Qualcomm (Unconfirmed)
• Absoft
• Nvidia
• These are based on Pathscale branch

8
Osprey Project Objectives

• Rejuvenate Compiler Research for Itanium
  Processor
• Memory Hierarchy Optimizations
• Exploit Multi-core features
• Exploit ILP to the full extend
• As GCC with high performance alternative back end
  
• Productize Open64 for Itanium Processor
• For performance critical applications wedded to
  open source compiler
• Build Distributed development model
• Centralized Source Repository bug tracking
• Using svn and Bugzilla
• Hosted in www.open64.net
• Provide tight integration platform for activities
  from research institutes
• Participation from Universities
• University of Delaware, Tsinghua University,
  CAS-ICT
• Coordinated from HP
• Program management and testing infrastructure
• Working toward third party product level support
• Working toward steering committee for open64
  community

9
Comparison between ORC and Osprey
Osprey ORC
As a Compiler Research Platform Yes Yes
Promote leveraging research results across Universities/Institutes Yes No
Provide Source Code Repository Yes No
One Bug data base for reporting problems Yes No
Production Quality as a goal Yes No
Key Take Away Osprey Project wants to build a
research and development community around Open64
Technology
10
Osprey Project 2006 Achievements

• Recreated source code repository
• Use subversion hosted on http//www.open64.net
• Includes all Open64 past releases in kept in
  Sourceforge
• Incorporated changes made in Pathscale
• Form extended engineering team, by engaging key
  academic and research institutions
• Released Opencc 3.0
• Adopted Pathscale 2.1 tree for GCC3.x front-end
  updates
• Incorporate latest ORC Itanium Code Generator
• Implement IA64 C ABI compatibility
• Ensure gcc command option compatibility
• Booted Linux Kernel with few source work-arounds

11
Osprey Project 2007 Goals

• Merge Pathscale 3.0 changes
• Catch up with GCC4.2 front ends
• Full GCC ABI compatibility
• New Inter Procedural Analysis/Optimizations
• Memory Hierarchy Optimizations
• New Itanium Specific Optimizations
• Advanced Prefetch
• Exploit Multi-core Trend
• Loop Parallelization, Automatic or Program Advise
• Thread Level Parallelization

12
Contributions from Pathscale (Qlogic)

• Osprey Project leverages effort made by Pathscale
  heavily
• Pathscale contributions (a Partial List)
• Brought up x8664 target support to Open64
• Option level compatibility with GCC/G
• Upgraded the GNU 3.x and GNU 4.x front end
• Added support for C Exception Handling in
  bridge
• Introduced new abstraction between GNU front end
  and Open64 framework
• Fixed many bugs in the front end /back end bridge
• Improved quality of middle end optimizations
• Updated Fortran front-end to support 2003
  standard
• Updated IPA linker to binutils-2.16.1
• My encouragement to support Pathscale effort

13
Research Areas My Wish List

• All dimensions of Multi-core / Multi-thread
• Break the Memory Hierarchy Barrier
• Better Integration with Java and C
• Mix-mode programming paradigm
• Trade-off between Static and Dynamic Compilation
  model
• Debugging and Performance Analysis/Advisory Tools
• Productivity and Security Check Tools
• Scalable/Incremental Inter Procedural Analysis
  Framework for Huge Applications
• Instruction Level Parallelism
• Is it all done? Clearly Not!

14
Open64 Research Activities

• UPC UC Berkeley
• OpenMP Tools U of Houston
• EPS, a global scheduler SNU
• Static analysis / performance advisory Google
• Cyclops University of Delaware
• Speculative Parallel Threading UMN

15
(No Transcript)