vmkit : a substrate for virtual machinesthomas_g/research/biblio/2009/...llvm posix threads dwarf...

VMKit : a Substrate for Virtual Machines Nicolas Geoffray, Gaël Thomas, Charles Clément, Bertil Folliot and Gilles Muller Design of VMKit Virtual Machines optimizations require a huge amount of time. Although they share some common principles, such as a Just In Time Compiler or a Garbage Collector, this opportunity for sharing has not been exploited in current VMs. VMKit is a first attempt to build a common substrate that eases the development of high-level VMs by reusing existing projects: LLVM, GNU Classpath, Mono Library, Boehm GC, GCC, Posix Threads. JnJVM GNU Classpath Virtual Machine Substrate LLVM Posix Threads Dwarf Boehm or Home made GC Glibc Java Initial Class Loader Class Type Verifier .class → LLVM IR Translator N3 Lisp VM Mono Library Assembly Loader Class Type Verifier Ass.→ LLVM IR Translator Home made Initial methods Lisp Lexer/Parser (build AST) Type inference AST → LLVM IR Translator Applicative Memory Not included in locs ∼ 20’000 locs ∼ 20’000 locs ∼ 5’000 locs VMKit has been successfully used to build three VMs: a Java Virtual Machine (JnJVM), a Common Language Runtime (N3) and a Lisp-like runtime with type inference. Our high level VMs are only 20,000 lines of code, it took one of the author a month to develop a Common Language Runtime and implementing new ideas in the VMs was remarkably easy [IJVM - DSN09]. VM Library Just In Time Compiler or Interpreter Thread Exception Memory Manager Input/Ouput Intermediate Representation Loader Intermediate Representation Verifier Load Compile Native Applicative Methods Method Stub (load/compile) Generated Code 1 Generate Use Manage Invoke main Generated Code 2 Generated Code 3 Applicative Objects add, call create, join read socket throw , catch gcmalloc Unified API Invoke Invoke Invoke Invoke Invoke Intermediate Representation Translator Virtual Machine Substrate High Level Virtual Machine Implementation of three VMs 0 0,2 0,4 0,6 0,8 1 1,2 1,4 1,6 N3 Mono 25,8 4,56 3,79 3,7 3,5 2,53 2,43 1,06 10 Instruc7on Selec7on Live Variable Analysis Linear Scan Register Allocator Live Interval Analysis Predicate Simpliﬁer Java bytecode to LLVM IR Global Value Numbering X86 Machine Code EmiPer Other 0 1 2 3 4 5 6 JnJVM IBM OpenJDK Jikes RVM Cacao Performance (Linux 2.6/1.5GHz/512Mo) + LLVM excellent for CPU intensive Java and .Net applications + Fast VM development + Ahead of Time Compiler − Ongoing work for a baseline JIT − Ongoing work to use MMTk − Ongoing work to optimize exceptions Results Contact [email protected] [email protected] Homepage http://vmkit.llvm.org Compilation overhead (Tomcat 6.0.16) : OpenJDK 8s, JnJVM 75s Scimark (C#) JavaGrande Forum Sec 3, SpecJVM98 and SpecJBB05 GC GC GC GC Exceptions Exceptions Array Bounds Check

Upload: others

Post on 30-Jan-2021

0 views

Category:

Documents

0 download

Report

Download

Embed Size (px):

TRANSCRIPT

VMKit : a Substrate for Virtual Machines Nicolas Geoffray, Gaël Thomas, Charles Clément, Bertil Folliot and Gilles Muller

Design of VMKit Virtual Machines optimizations require a huge amount of time. Although they share some common principles, such as a Just In Time Compiler or a Garbage Collector, this opportunity for sharing has not been exploited in current VMs.

VMKit is a first attempt to build a common substrate that eases the development of high-level VMs by reusing existing projects: LLVM, GNU Classpath, Mono Library, Boehm GC, GCC, Posix Threads.

JnJVM

GNU Classpath

Virtual Machine Substrate

LLVM Posix Threads Dwarf Boehm or Home made GC Glibc

Java Initial Class Loader

Class Type Verifier

.class → LLVM IR Translator

N3 Lisp VM

Mono Library

Assembly Loader

Class Type Verifier

Ass.→ LLVM IR Translator

Home made Initial methods

Lisp Lexer/Parser (build AST)

Type inference

AST → LLVM IR Translator

Applicative Memory

Not included in locs

∼ 20’000 locs ∼ 20’000 locs ∼ 5’000 locs VMKit has been successfully used to build three VMs: a Java Virtual Machine (JnJVM), a Common Language Runtime (N3) and a Lisp-like runtime with type inference. Our high level VMs are only 20,000 lines of code, it took one of the author a month to develop a Common Language Runtime and implementing new ideas in the VMs was remarkably easy [IJVM - DSN09].

VM Library

Just In Time Compiler or Interpreter

Thread Exception Memory Manager Input/Ouput

Intermediate Representation

Loader

Intermediate Representation

Verifier

Load

Compile

Native Applicative

Methods

Method Stub (load/compile)

Generated Code 1

Generate Use

Manage

Invoke main

Generated Code 2

Generated Code 3

Applicative Objects

add, call

create, join

read socket

throw, catch

gcmalloc

Unified API

Invoke Invoke Invoke

Invoke

Invoke

Intermediate Representation

Translator

Virtual Machine Substrate

High Level Virtual Machine

Implementation of three VMs 

0 

0,2 

0,4 

0,6 

0,8 

1 

1,2 

1,4 

1,6 

N3 

Mono 

25,8 

4,56 3,79 

3,7 

3,5 

2,53 

2,43 

1,06 

10  Instruc7on Selec7on 

Live Variable Analysis 

Linear Scan Register Allocator 

Live Interval Analysis 

Predicate Simplifier 

Java bytecode to LLVM IR 

Global Value Numbering 

X86 Machine Code EmiPer 

Other 

0 

1 

2 

3 

4 

5 

6 

JnJVM 

IBM 

OpenJDK 

Jikes RVM 

Cacao 

Performance (Linux 2.6/1.5GHz/512Mo) 

+  LLVM excellent for CPU intensive Java and .Net applications

+  Fast VM development +  Ahead of Time Compiler

−  Ongoing work for a baseline JIT −  Ongoing work to use MMTk −  Ongoing work to optimize exceptions

Results 

Contact [email protected]

[email protected]

Homepage 

http://vmkit.llvm.org

Compilation overhead (Tomcat 6.0.16) : OpenJDK 8s, JnJVM 75s

Scimark (C#)

JavaGrande Forum Sec 3, SpecJVM98 and SpecJBB05

GC GC GC

GC

Exceptions

Exceptions

Array Bounds Check