• [RFC][PIR] Parallel IR, Stage 0: IR extensions
• Preview
• 1 Introduction
• 1.1 Motivation
• 1.2 Advantages of fork-join parallelism
• 1.2.1 Minimal IR extension
• 1.2.2 Simple semantics
• 1.2.3 Simple and quick implementation
• 1.2.4 Incremental integration into LLVM
• 1.2.5 Low-level representation of parallelism
• Implementing divide-and-conquer algorithms
• Implementing parallel for loops
• Implementing map
• Implementing reduction
• Implementing scan
• 1.2.6 Representation of static parallelism
• 1.2.7 Representative of existing languages
• 1.2.8 Existence of prototypes
• 1.2.9 Better optimization opportunities
• 1.2.10 Good tooling support
• Performance models
• Scalability profilers
• Race detectors
• 1.2 Scope
• 2 Design Goals
• 3 Instruction Specification
• 3.1 fork Terminator
• 3.2 halt Terminator
• 3.3 join Terminator
• 4 Examples
• 4.1 Simplest case (2 parallel threads)
• Cilk
• OpenMP
• PIR
• 4.2 More than 2 threads
• Cilk
• OpenMP
• PIR
• 4.3 A Parallel loop
• Cilk
• OpenMP
• PIR
• 4.4 Nested parallel loops
• Cilk
• OpenMP
• PIR
• 5 Outlook
• 6 References