Compiler Research at the Indian Institute of Science Bangalore, India

1
Compiler Research at the Indian Institute of
ScienceBangalore, India

• Y.N. Srikant
• Professor and Chairman
• Department of Computer Science and Automation
• Indian Institute of Science
• srikant_at_csa.iisc.ernet.in

2
Past Research

• Automatic generation of incremental parsers and
  semantic analyzers
• Optimizer generator
• Machine code generation with top-down tree
  pattern matching
• New algorithm for partial redundancy elimination
• Automatic parallelization for distributed memory
  machines
• Automatic data partitioning based on genetic
  algorithms (for DMMs)

3
Current Research Areas

• Code Generator Generators
• Instruction Scheduling
• Power-aware compilation
• Just-In-Time compilation

4
Code Generator Generators

• Emphasis on generating efficient code generators
  from machine descriptions
• Uses new tree pattern matching alogorithm based
  on LR-parsing (in-house)
• Integrated into TRIMARAN and is based on LCODE
  intermediate code
• Tried on DSP (Texas Instr.) and Itanium
• Faculty member Priti Shankar

5
Instruction Scheduling

• Faculty members
• R. Govindarajan, Priti Shankar and Y.N. Srikant
• New approaches for VLIW and Superscalar
  architectures
• Integer linear programming
• Co-scheduling hardware and software pipelines

6
Instruction Scheduling (contd.)

• Instruction scheduling for clustered DSP
  architectures
• Integrated scheduling combining clustering and
  scheduling
• Graph matching based scheduling
• Integrated into SUIF
• Experiments with TI 32064X

7
Power Aware Compilation

• Transition-aware scheduling
• Reduces the number of transitions from active to
  low-power and vice-versa
• Saves energy
• With XScale, XTREM, and gcc
• Experiments with heterogeneous interconnects in
  clustered VLIW architectures
• Two interconnects, one with high speed and high
  power requirement, another with lower speed and
  lower power requirement
• Experiments with TRIMARAN (suitably modified)

8
Power Aware Compilation (contd.)

• Experiments with Object cache
• Separate cache for objects
• Combination saves energy and is as fast as a
  general purpose cache
• Use of static analysis and profiling to classify
  objects as long living and short living
• A Framework for writing power aware applications
  with ECOS
• Hooks into the OS
• ??

9
JIT Compilation

• An efficient JIT compiler for .NET CLI on ROTOR
• Profile-guided optimizations
• Cluster and Collect concurrent garbage
  collector
• Detects long-living clusters of objects by
  profiling and static analysis
• Allocates them in a separate mature object space
• Stack allocation for scoped objects
• Architectural innovations to speedup profile
  collection
• Low profiling speed is the bottleneck with
  profile-guided optimizations
• Most common types of profiling can be carried out

10
Thank You

• Questions?