Vectorization for Standard ML in MLton
Anton Bachin (abachin AT andrew)
Anand Subramanian (asubrama AT andrew)
Description
We are investigating the benefits of vectorization for Standard ML. Our original goal was to consider the interaction between CPS form and auto-vectorization, in the example of the MLRISC back end of the SML/NJ compiler. However, we have found development for SML/NJ to be difficult. The SML/NJ project seems to be limited to maintenance releases. Instead, we are focusing on the MLton compiler, which uses intermediate representations that are basically equivalent to regular SSA form.
Status
We have analyzed large portions of the MLton code, including the major intermediate representations, the back end, and the x86-64 code generator. We are currently modifying each of these to support SIMD intrinsics and SSE instructions. This task is nearly complete.
Links
Milestone report in PDF format (18 April)
Project proposal in PDF format
MLton website (down at the time of posting)
MLton benchmark suite
MLRISC project website
SML/NJ home page
References
Naishlos, D. Autovectorization in GCC
Nuzman, D., Zaks, A. Autovectorization in GCC - two years later
Nuzman, D., et. al. Vapor SIMD: Auto-Vectorize Once, Run Everywhere
Anderson, S., Hudak, P. Compilation of Haskell Array Comprehensions for Scientific Computing
Shivers, O. Control-Flow Analysis of Higher-Order Languages, or Taming Lambda
Appel, A. Loop Headers in Lambda Calculus or CPS
Kelsey, R. A Correspondence Between Continuation Passing Style and Static Single Assignment Form
Reppy, J. Optimizing Nested Loops Using Local CPS Conversion
George L., Guillame F. A Portable and Optimizing Back End for the SML/NJ Compiler
Bernard, A., Harper, R., Lee, P. How Generic is a Generic Back End? Using MLRISC as a Back End for the TIL Compiler
Pucella, R. Notes on Programming Standard ML of New Jersey