ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

One VM to Rule Them All

?Download as PPTX, PDF?
?
T
Tianxiang Xiong

The document discusses creating a high-performance virtual machine (VM) architecture that can easily implement VMs for new languages. It introduces Truffle, a language implementation framework, and Graal, a compilation infrastructure. With Truffle and Graal, language implementers can write interpreters for their languages and reusable services will provide a high-performance VM without requiring extensive work in compiler optimization. The guest language interpreter is partially evaluated and optimized based on profiling information during execution.

1 of 13
Download to read offline
One VM to Rule Them All Tianxiang Xiong
Background ¡ñ A small number of languages have high-perf VMs ¡ð E.g. Java, C# ¡ñ Many languages have low-perf VMs ¡ð Ruby, Python, R, MATLAB, PHP¡­ ¡ñ High-perf VMs are currently: ¡ð Highly complex ¡ð Biased towards a single language or family of languages
Big Idea ¡ñ Create architecture for easily implementing high-perf VMs for new languages 1. Language implementer writes interpreter 2. *A miracle occurs* 3. High-performance VM
Truffle and Graal ¡ñ Truffle--language implementation framework ¡ñ Graal--compilation infrastructure ¡ñ VMs for many guest languages implemented in a managed host language ¡ð Only guest-specific parts are written by language implementer ¡ð Reusable host services are provided by framework ¡ñ Guest language (usually) dynamically typed
One VM to Rule Them All
AST and Optimization ¡ñ Guest language implements AST interpreter ¡ñ Node rewriting--during interpretation, a node can replace itself w/ another, based on profiling feedback: ¡ð Dynamic type info ¡ð Execution freqs ¡ð Rate of node rewrites ¡ñ When compiler is evoked, it uses profiling info to partially evaluate the AST interpreter
AST and Deoptimization ¡ñ Parts of interpreter responsible for node rewriting are omitted from compilation ¡ð Branches that perform rewriting cause deoptimization ¡ñ Deoptimization--when type specialization fails, optimized code is discarded and execution reverts to interpreter ¡ñ More info is gathered during interpretation, followed by another round of partial evaluation, and so on
One VM to Rule Them All
Node Rewriting Conditions ¡ñ Guest language developer is responsible for fulfilling conditions of node rewriting: ¡ð Completeness--node provides rewrites for all cases it does not handle ¡ð Finiteness--After finite # of node rewrites, must end up in state that handles full language semantics ¡ð Locality--a node can only rewrite itself and its subtree
Branch probabilities and loop freqs ¡ñ Guest lang interpreter author is also (optionally) responsible for providing: ¡ð Branch probabilities ¡ð Loop frequencies ¡ñ Interpreter can do profiling for optimizing compiler ¡ð If not, values are derived from host system during partial evaluation
Related Work ¡ñ PyPy ¡ð Big idea: write interpreter in high-level (less dynamic) lang, then generate custom VM, tracing dynamic compiler form interpreter ¡ð Python interpreter, VM written in RPython -> translated to C ¡ñ Self-optimizing interpreters ¡ð Compilers can perform optimizations beyond reach of interpreters ¡ñ Compile to IR (e.g. LLVM), e.g. Rubinius
TruffleClojure ¡ñ TruffleClojure ¡ð Master¡¯s thesis, not ready for use ¡ñ Features: ¡ð Implements TCO (by avoiding Java bytecode) ¡ð Perf improvements via type specialization (e.g. of contains?)
One VM to Rule Them All

More Related Content

One VM to Rule Them All

  • 1. One VM to Rule Them All Tianxiang Xiong
  • 2. Background ¡ñ A small number of languages have high-perf VMs ¡ð E.g. Java, C# ¡ñ Many languages have low-perf VMs ¡ð Ruby, Python, R, MATLAB, PHP¡­ ¡ñ High-perf VMs are currently: ¡ð Highly complex ¡ð Biased towards a single language or family of languages
  • 3. Big Idea ¡ñ Create architecture for easily implementing high-perf VMs for new languages 1. Language implementer writes interpreter 2. *A miracle occurs* 3. High-performance VM
  • 4. Truffle and Graal ¡ñ Truffle--language implementation framework ¡ñ Graal--compilation infrastructure ¡ñ VMs for many guest languages implemented in a managed host language ¡ð Only guest-specific parts are written by language implementer ¡ð Reusable host services are provided by framework ¡ñ Guest language (usually) dynamically typed
  • 6. AST and Optimization ¡ñ Guest language implements AST interpreter ¡ñ Node rewriting--during interpretation, a node can replace itself w/ another, based on profiling feedback: ¡ð Dynamic type info ¡ð Execution freqs ¡ð Rate of node rewrites ¡ñ When compiler is evoked, it uses profiling info to partially evaluate the AST interpreter
  • 7. AST and Deoptimization ¡ñ Parts of interpreter responsible for node rewriting are omitted from compilation ¡ð Branches that perform rewriting cause deoptimization ¡ñ Deoptimization--when type specialization fails, optimized code is discarded and execution reverts to interpreter ¡ñ More info is gathered during interpretation, followed by another round of partial evaluation, and so on
  • 9. Node Rewriting Conditions ¡ñ Guest language developer is responsible for fulfilling conditions of node rewriting: ¡ð Completeness--node provides rewrites for all cases it does not handle ¡ð Finiteness--After finite # of node rewrites, must end up in state that handles full language semantics ¡ð Locality--a node can only rewrite itself and its subtree
  • 10. Branch probabilities and loop freqs ¡ñ Guest lang interpreter author is also (optionally) responsible for providing: ¡ð Branch probabilities ¡ð Loop frequencies ¡ñ Interpreter can do profiling for optimizing compiler ¡ð If not, values are derived from host system during partial evaluation
  • 11. Related Work ¡ñ PyPy ¡ð Big idea: write interpreter in high-level (less dynamic) lang, then generate custom VM, tracing dynamic compiler form interpreter ¡ð Python interpreter, VM written in RPython -> translated to C ¡ñ Self-optimizing interpreters ¡ð Compilers can perform optimizations beyond reach of interpreters ¡ñ Compile to IR (e.g. LLVM), e.g. Rubinius
  • 12. TruffleClojure ¡ñ TruffleClojure ¡ð Master¡¯s thesis, not ready for use ¡ñ Features: ¡ð Implements TCO (by avoiding Java bytecode) ¡ð Perf improvements via type specialization (e.g. of contains?)