Maxine A JVM in Java - JUG Saxony3 The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into

<Insert Picture Here>

Maxine: A JVM Written in Java

Michael HauptOracle LabsPotsdam, Germany

3

The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, and timing of any features or functionality described for Oracle’s products remains at the sole discretion of Oracle.

Speaker's Background

4

Technische Universität Darmstadt, 2001–2006

Doctoral research, Steamloom (virtual machine support for aspect-oriented programming)


4



Hasso-Plattner-Institut Potsdam, 2006–2011

Postdoc, virtual machine architecture disentangling, SOM family, NXTalk


4



Hasso-Plattner-Institut Potsdam, 2006–2011

Postdoc, virtual machine architecture disentangling, SOM family, NXTalk

Oracle Labs, Potsdam,since 2011

Maxine team, virtual machine architecture, multi-language virtual machines, code cache management, JSR292

Who Needs a JVM?

5

JVM

Who Needs a JVM?

5

application developers, end users ✔

JVM

Java

Applications

Who Needs a JVM?

5


researchers ✔

JVM

Java

Applications

Dynamic Compiler Garbage Collector Object Layout Runtime

Optimisations

Target Architecture Parallel

Concurrent Language Specifics Thread Migration

Concurrency AbstractionsTagging

Who Needs a JVM?

5


researchers ✔

programming language implementers ✔

JVM

Java

Applications


Optimisations




Java, Ruby, Python, Smalltalk, ...

Who Needs a JVM?

5


researchers ✔


JVM

Java

Applications


Optimisations





! ! ! !

Who Needs a JVM?

5


researchers ✔


JVM

Java

Applications


Optimisations





J?VM

! ! ! !

Maxine

6

almost 100 % Java

use unmodified JDK

meta-circularity studies

Home page: http://wikis.oracle.com/display/MaxineVM/Home

Source (GPLv2): https://hg.kenai.com/hg/maxine~maxine

Mailing list via: http://kenai.com/projects/maxine

Maxine

6

almost 100 % Java

use unmodified JDK


supported on 64-bit x86

Solaris, Mac OS X, Linux

Virtual Edition: Xen




Maxine

6

almost 100 % Java

use unmodified JDK


full IDE support

(Eclipse, NetBeans)

tooling: Inspector







Maxine

6

almost 100 % Java

use unmodified JDK


full IDE support

(Eclipse, NetBeans)

tooling: Inspector







dynamic compilation only

semi-space GC

modular architecture

VM Research Group at Oracle Labs

7




DougSimon

MarioWolczko

Director

ThomasWürthinger

MichaelHaupt

ChristianWimmer

LaurentDaynès

MickJordan

MichaelVan De Vanter

Why Java?

8

language• reflection• annotations

JDK• rich and powerful API• inclusion in VM

ecosystem• IDE support• tooling

VM architecture• uniform representation• same compiler chain

Bootstrapping Maxine

9

Maxine Source Code


9

Maxine Source Code

Maxine Bytecodes

javacJust-in-Time Compilers

Memory Manager

... others ...

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes

Boot Image Generator

javac

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes



Memory Manager

... others ...

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes



Memory Manager

... others ...

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes



Memory Manager

... others ...

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes


MaxineBoot Image


Memory Manager

... others ...

Host VM (HotSpot)


9

Maxine Source Code

Maxine Bytecodes


Boot Loader

MaxineBoot Image


Memory Manager

... others ...

Maxine Scheme Abstractions

10

Maxine VM

Dynamic Compilers Garbage Collector Object Layout Runtime


10

Maxine VM


interface LayoutScheme


10

Maxine VM


interface LayoutScheme interface ReferenceScheme


10

Maxine VM


interface LayoutScheme interface ReferenceSchemeinterface HeapScheme


10

Maxine VM



interface MonitorScheme


10

Maxine VM



interface MonitorScheme interface RunScheme


10

Maxine VM



interface MonitorScheme interface RunSchemeinterface CompilationBroker

Example: ReferenceScheme

11

What you use in Maxine to access a field:

int x = Reference.fromJava(obj).readInt(24); field offset (in bytes) is computed by LayoutScheme;we assume a constant offset in this example


11


int x = Reference.fromJava(obj).readInt(24);

Actual call chain after forced inlining via @INLINE or @FOLD:

Reference ref = MaxineVM.vm.config.referenceScheme.toReference(obj);Pointer ptr = MaxineVM.vm.config.referenceScheme.toOrigin(ref);int x = ptr.readInt(24);

field offset (in bytes) is computed by LayoutScheme;we assume a constant offset in this example


11





All fields are marked @CONSTANT, so the compiler knows theconcrete scheme implementation used:

class DirectReferenceScheme implements ReferenceScheme { @INTRINSIC(UNSAFE_CAST) native Reference toReference(Object origin); @INTRINSIC(UNSAFE_CAST) native Word toOrigin(Reference ref);}



11





All fields are marked @CONSTANT, so the compiler knows theconcrete scheme implementation used:

class DirectReferenceScheme implements ReferenceScheme { @INTRINSIC(UNSAFE_CAST) native Reference toReference(Object origin); @INTRINSIC(UNSAFE_CAST) native Word toOrigin(Reference ref);}

Final machine code:

movl [rax + 24], rbx


Baseline Compiler: T1X

12

job: assemble pre-generated bytecode templates• written in Java• compiled by optimising compiler at VM build time• very small amount of glue assembly


12

job: assemble pre-generated bytecode templates• written in Java• compiled by optimising compiler at VM build time• very small amount of glue assembly

template code for integer array load (IALOAD)@T1X_TEMPLATE(IALOAD)static int iaload(@Slot(1) Object array, @Slot(0) int index) { ArrayAccess.checkIndex(array, index); return result = ArrayAccess.getInt(array, index);}@INLINEstatic void checkIndex(Object array, int index) { int length = Layout.readArrayLength(Reference.fromJava(array)); if (UnsignedMath.aboveOrEqual(index, length)) { throw Throw.arrayIndexOutOfBoundsException(array, index); }}

@INLINEstatic int getInt(Object array, int index) { return Layout.getInt(Reference.fromJava(array), index);}


13

instantiated template code for integer array load

mov rdi, [rsp + 16] mov esi, [rsp] mov rax, rdi movsxd rax, [rax + 16] cmp esi, eax jb L1 call Throw.indexOutOfBoundsException() nop mov rdi, rax call MaxRuntimeCalls.runtimeUnwindException nop L1: movsxd rsi, esi movsxd rax, rdi[rsi * 4 + 24] addq rsp, 0x10 mov [rsp], eax

prologue(loading of operand stack slots)

compiled template

epilogue(storing of operand stack slots)

8 bytes hub (class pointer)8 bytes misc (locking, GC)4 bytes array length4 bytes alignment

Inspector

14

Optimising Compilers: C1X and Graal

15

HotSpot VMRuntime

InterpreterObject Locking

C1 / C2

Inline Assembly

Garbage Collector Type Information

Compiler Interface

C++


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface

Graal VMRuntime


Graal

Snippets


Compiler Interface


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface

Graal VMRuntime


Graal

Snippets


Compiler Interface


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface

Graal VMRuntime


Graal

Snippets


Compiler Interface

Maxine VM with GraalRuntime

T1XObject Locking

Graal

Snippets


Compiler Interface


15

HotSpot VMRuntime


C1 / C2

Inline Assembly


Compiler Interface

Java

C++

Maxine VMRuntime

T1XObject Locking

C1X

XIR


Compiler Interface

Graal VMRuntime


Graal

Snippets


Compiler Interface

Maxine VM with GraalRuntime

T1XObject Locking

Graal

Snippets


Compiler Interface

Graal

16

Home page: http://openjdk.java.net/projects/graal/

Source (GPLv2): http://hg.openjdk.java.net/graal/graal/

Mailing list: [email protected]

Vision Statement

Create an extensible, modular, dynamic, and aggressive compiler using

object-oriented and reflective Java programming,

a graph-based and visualizable intermediate representation,

and Java snippets.

-- Thomas Würthinger

Collaboration with Johannes-Kepler-Universität Linz, Austria

Lukas Stadler, Gilles Duboscq, Christian Häubl, Prof. Hanspeter MössenböckJKU

Graal

17

Architecture Lowering

Memory Optimisations

VM Lowering

Loop Optimisations

Escape Analysis

Type Inference

Method Inlining

Java Bytecode Parser

Peephole Optimisations

Register Allocation

Code Generation

Java Bytecodes

Machine Code

Graal

17



VM Lowering

Loop Optimisations

Escape Analysis

Type Inference

Method Inlining



Register Allocation

Code Generation

Java Bytecodes

Machine Code

Graal

17



VM Lowering

Loop Optimisations

Escape Analysis

Type Inference

Method Inlining



Register Allocation

Code Generation

Java Bytecodes

Machine Code

Language dependent

Language independentVM independent

Architecture independent

VM dependentArchitecture independent

VM dependentArchitecture dependent

Performance: Maxine/C1X

18

2-socket dual core AMD Opteron 2214, 2.2 GHz, 4 total cores4 GByte main memory

Oracle Enterprise Linux, version 2.6.18-238.9.1.0.1.e15

71% 64%

58% 54%

76% 74%

151%

119%

81%

157%

191%

138%

0%

20%

40%

60%

80%

100%

120%

140%

160%

180%

200%

SPECjvm2008 DaCapo peak DaCapo startup SPECjbb2005 SciMark 2.0 JavaGrande

Client Maxine/C1X Server

Performance: GraalVM

19

Intel(R) Core(TM) i5-750 @ 2.67GHz (4 cores)8 GByte main memory

Ubuntu Linux 11.10, kernel 3.0.0-12

72%

48%

66%

77%

61%

74%

0%

20%

40%

60%

80%

100%

120%

DaCapo Scala-DaCapo SPECjvm2008

Client GraalVM Server

Maxine VM

Maxine: Code Cache Management

20

T1X

C1X

Maxine VMT1X Cache C1X Cache


20

m()

n()

o()

<clinit>

T1X

C1X



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

T1X

C1X



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X

T1X Cache with Semi-Space Memory Management Scheme

to-space from-space



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X


to-space from-space

m() n()<clinit>



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X


to-space from-space

m() n()<clinit> o()



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X


to-space from-space

m() n()<clinit>



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X


to-space from-spaceto-spacefrom-space

m() n()<clinit>



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X



m() n()<clinit> n()<clinit>



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X



n()<clinit> n()<clinit> m()



20

m()

n()

o()

<clinit>

m()

o()

m()

o()

<clinit>

T1X

C1X



n()<clinit> n()<clinit> m() o()

Code Cache Management: Which Methods are Live?

21

t1 call stack

baseline (T1X) method optimised (C1X) method ✔ live method

o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()


21

t1 call stack


o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()

✔ ✔ ✔


21

t1 call stack


o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()

k()

✔ ✔ ✔

✔


21

t1 call stack


o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()

k()

j()

✔ ✔ ✔ ✔

✔


21

t1 call stack


o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()

k()

j()

✔ ✔ ✔ ✔

✔


21

t1 call stack


o()

T1X Cache

to-space from-space

p()

f()

g()

h()

g()

f() g() h() i() j()

k() l()

k()

j()

live methods• methods on stack• direct callees of on-stack methods

others to protect from eviction• freshly compiled, not yet installed• invocation counter within optimisation threshold• existing type profile

✔ ✔ ✔ ✔

✔

Code Cache Management: Eviction

22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

i()

l()

j()

k()

f()✔ g()✔ h()✔ j()✔

k()✔


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

i()

l()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)

f()✔ g()✔ h()✔ j()✔

k()✔


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)

f()✔ g()✔ h()✔ j()✔

k()✔


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)2. move marked methods (update vtable entries, direct calls)

f()✔ g()✔ h()✔ j()✔

k()✔


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)2. move marked methods (update vtable entries, direct calls)

f() g() h() j() k()


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)2. move marked methods (update vtable entries, direct calls)3. patch return addresses

f() g() h() j() k()


22

t1 call stack


o()

T1X Cache

from-space to-space

p()

f()

g()

h()

g()

j()

k() course of action1. invalidate unmarked methods (and vtable entries, direct calls)2. move marked methods (update vtable entries, direct calls)3. patch return addresses

f() g() h() j() k()

Where Do Pointers to Machine Code Occur,and How Are They Handled?

23

vtables/itables ✓directly accessible via meta-level API (actors, hubs); invalidate (replace with trampoline calls) or relocate

direct calls ✓directly accessible via target method API (safepoints); invalidate (replace with trampoline calls) or relocate

return addresses ✓ accessible during stack walking; relocate accordingly

local variables ✖ These might have to be relocated—how to determine which Address is a code pointer?member variables ✖

These might have to be relocated—how to determine which Address is a code pointer?


23







23







23







23






Tagged Code Pointers

24

Word

Address Offset

Pointer Size

CodePointer

Object


24

Word

Address Offset

Pointer Size

CodePointer

Object

1 tag bit63 bits payload (offset from base code address)


24

class CodePointer { ... @INLINE public static CodePointer from(long value) { if (isHosted()) { return new CodePointer(tag(value)); } return UnsafeCast.asCodePointer(tag(value)); } ...}

Word

Address Offset

Pointer Size

CodePointer

Object



24

class CodePointer { ... @INLINE public static CodePointer from(long value) { if (isHosted()) { return new CodePointer(tag(value)); } return UnsafeCast.asCodePointer(tag(value)); } ...}

@INTRINSIC(UNSAFE_CAST)public static CodePointer asCodePointer(long value) { return CodePointer.from(value);}

Word

Address Offset

Pointer Size

CodePointer

Object


Stack Reference Maps and Tagged Code Pointers

25

static void m(int i, MyClass o) { int j = 2 * i; o.f(i);}

0: iconst_21: iload_02: imul3: istore_24: aload_15: iload_26: invokevirtual <f:(I)V>9: return


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address

idealised stack frame layout


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5

stack reference map(at instruction 6)


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5


void x(...) { ... MyClass o = ...; CodePointer cp = ...; ...}


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5



0 1 2 3 4 5


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5



0 1 2 3 4 5

Maxine VM

GC code eviction

stack reference map builder


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5



0 1 2 3 4 5

Maxine VM

GC code eviction


ignore tagged refs,scan others


25



i

o

j

arguments

locals

operands

1

2

3

4

5

<r>0 return address


0 1 2 3 4 5



0 1 2 3 4 5

Maxine VM

GC code eviction


ignore tagged refs,scan others

regard and relocateonly tagged refs

<Insert

26

Thank you for your attention.

Acknowledgements: these slides are joint work of the VM research group.




Maxine A JVM in Java - JUG Saxony3 The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into

Documents