README.md

In duncantl/R2llvm: LLVM-based compiler for R code

R2llvm

R2llvm is the next generation of RLLVMCompile package to provide a compiler for R code that generates machine code using LLVM via the Rllvm package. This differs from RLLVMCompile as it uses a control flow graph computed via the rstatic package and type information computed from that control flow graph via the RTypeInference package.

Goals

The goal of this is to provide a customizable and extensible tool that 1) people can use to compile "R-like" code, and 2) tool developers can experiment with to explore different approaches to making code fast. People can use this for implementing different compilation strategies for aspects of the R language, for targetting different platforms, e.g. GPUs, and also for compiling different DSLs within R, e.g., hierarchical models.

The idea is that we can translate R-like code to LLVM instructions, create those instructions using the Rllvm package and then generate native/compiled code and invoke it from the R session, serialize it for use in other sessions or even applications (e.g. JavaScript, Python).

We might compile the same code in different ways for different contexts. As a result, a single one-size-fits-all approach to compiling R is probably too restrictive. Furthermore, we want to be able to explore new approaches easily w/o having to recompile all of R or learn a new compilation framework tha is specific to R.

This approach is very different from building a compiler that requires changing the code of the R interpreter, or developing a different implementation of R, i.e. separate from GNU R. This sits on top of the existing GNU R implementation as a package and does not require any changes to GNU R.

One of the strengths of LLVM is that it embeddable and extensible and provides a user-level API. A compiler for R should also do the same. We have learned that a centralized code source that requires a core group to make all "official" and "distributed" changes limits innovation (but does improve stability).

LLVM also generates code for different backends, e.g., different hardware, GPUs, etc. and emscripten can generate JavaScript from the IR code we generate with LLVM.

Getting Started

To compile an R function into machine code, use

compileFunction()

Ideally you specify the R function and the function takes care of the rest. This will attempt to infer the types for this function. It will also attempt to determine the signature - parameter and return types - for other functions this function calls. This returns a reference to the LLVM Function. You can then invoke this using the .llvm() function from the Rllvm package. The following shows how to compute 42 slowly!

library(R2llvm)
f = function() 42L
cf = compileFunction(f, optimize = FALSE) #XXX
.llvm(cf)

History

Vince Buffalo and I started the RLLVMCompile package several years ago (late 2010) after the development of the Rllvm package. Unfortunately, I had other committments (2 books, directing the Davis Data Science Initiative). Nick Ulle has been working on type inference starting in 2015/2016.

duncantl/R2llvm documentation built on May 14, 2019, 9:42 a.m.