JVM for Dummies - OSCON 2011

JVM for Dummies
(and for the rest of you, as well)

Wednesday, July 27, 2011

Intro
• Charles Oliver Nutter
• “JRuby Guy”
• Sun Microsystems 2006-2009
• Engine Yard 2009-
• Primarily responsible for compiler, perf
• Lots of bytecode generation
• Lots of JIT monitoring


Today
• JVM Bytecode
• Inspection
• Generation
• How it works
• JVM JIT
• How it works
• Monitoring
• Assembly (don’t be scared!)


Today
• JVM Bytecode
•
•
•
Inspection
Generation } For Dummies

}
How it works
• JVM JIT
• How it works For people who want
• Monitoring to feel like Dummies
• Assembly


Part One:
Bytecode


Bytecode Deﬁnition

• “... instruction sets designed for efﬁcient
execution by a software interpreter ...”
• “... suitable for further compilation into
machine code.


Byte Code

• One-byte instructions
• 256 possible “opcodes”
• 200 in use on current JVMs
• Room for more :-)
• Little variation since Java 1.0


Microsoft’s CLR

• Stack-based, but not interpreted
• Two-byte “Wordcodes”
• Similar operations to JVM


Why Learn It
• Know your platform
• Full understanding from top to bottom
• Bytecode generation is fun and easy
• Build your own language?
• May need to read bytecode someday
• Many libraries generate bytecode

Hello World

public class HelloWorld {
public static void main(String[] args) {
System.out.println("Hello, world");
}
}


javap
• Java class ﬁle disassembler
• Basic operation shows class structure
• Methods, superclasses, interface, etc
• -c ﬂag includes bytecode
• -public, -private, -protected
• -verbose for stack size, locals, args

javap

~/projects/bytecode_for_dummies ➔ javap HelloWorld
Compiled from "HelloWorld.java"
public class HelloWorld extends java.lang.Object{
public HelloWorld();
public static void main(java.lang.String[]);
}


javap -c
~/projects/bytecode_for_dummies ➔ javap -c HelloWorld
public class HelloWorld extends java.lang.Object{
Code:
0:! aload_0
1:! invokespecial!#1; //Method java/lang/Object."<init>":()V
4:! return

Code:
0:! getstatic! #2; //Field java/lang/System.out:Ljava/io/PrintStream;
3:! ldc!#3; //String Hello, world
5:! invokevirtual!#4; //Method java/io/PrintStream.println:
(Ljava/lang/String;)V
8:! return

}


javap -verbose
~/projects/bytecode_for_dummies ➔ javap -c -verbose HelloWorld
public class HelloWorld extends java.lang.Object
SourceFile: "HelloWorld.java"
minor version: 0
major version: 50
Constant pool:
const #1 = Method! #6.#15;!// java/lang/Object."<init>":()V
const #2 = Field!#16.#17;! / java/lang/System.out:Ljava/io/PrintStream;
/
const #3 = String! #18;! / Hello, world
/
const #4 = Method! #19.#20;! / java/io/PrintStream.println:(Ljava/lang/String;)V
/
const #5 = class!#21;! / HelloWorld
/
...

{


javap -verbose

...
Code:
Stack=1, Locals=1, Args_size=1
0:!aload_0
1:!invokespecial! #1; //Method java/lang/Object."<init>":()V
4:!return
LineNumberTable:
line 1: 0


javap -verbose
Code:
Stack=2, Locals=1, Args_size=1
0:! getstatic!#2; //Field java/lang/System.out:Ljava/io/PrintStream;
3:! ldc!
#3; //String Hello, world
5:! invokevirtual!
#4; //Method java/io/PrintStream.println:
(Ljava/lang/String;)V
8:! return
LineNumberTable:
line 3: 0
line 4: 8

}


Thank you!


Thank you!
(Just Kidding)


Let’s try something a
little easier...


BiteScript

• (J)Ruby DSL for emitting JVM bytecode
• Internal DSL
• Primitive “macro” support
• Reads like javap -c (but nicer)
• http://github.com/headius/bitescript


Installation
• Download JRuby from http://jruby.org
• Unpack, optionally add bin/ to PATH
• Ahead of PATH if you have Ruby already
• [bin/]jruby -S gem install bitescript
• `bite myfile.bs` to run myfile.bs file
• `bitec myfile.bs` to compile myfile.bs file

BiteScript Users

• Mirah
• Ruby-like language for writing Java code
• BiteScript for JVM bytecode backend
• BrainF*ck implementation
• Other miscellaneous bytecode experiments


BiteScript

main do
getstatic java.lang.System, "out",
java.io.PrintStream
ldc "Hello, world!"
invokevirtual java.io.PrintStream, "println",
[java.lang.Void::TYPE, java.lang.Object]
returnvoid
end


BiteScript

import java.lang.System JRuby’s “import”
import java.io.PrintStream for Java classes
main do
getstatic System, "out", PrintStream
ldc "Hello, world!"
invokevirtual PrintStream, "println", [void, object]
returnvoid
end
Shortcuts for
void, int, string,
object, etc

BiteScript

main do
ldc "Hello, world!"
aprintln
returnvoid A BiteScript “macro”
end


BiteScript

macro :aprintln do
swap
invokevirtual PrintStream, "println",
[void, object]
end


The Basics

• Stack machine
• Basic operations
• Flow control
• Class structures
• Exception handling


Stack Machine
• The “operand stack” holds operands
• Operations push and/or pop stack values
• Exceptions: nop, wide, goto, jsr/ret
• Stack must be consistent
• Largest part of bytecode veriﬁer
• Stack is explicitly sized per method

The JVM Stack
Depth Value
import java.lang.System
import java.io.PrintStream 0

main do
1
ldc "Hello, world!"
2
[void, object]
returnvoid 3
end
4


The JVM Stack
Depth Value
import java.io.PrintStream 0 out (a PS)

main do
1
ldc "Hello, world!"
2
[void, object]
returnvoid 3
end
4


The JVM Stack
Depth Value
import java.io.PrintStream 0 “Hello, world!”

main do
1 out (a PS)
ldc "Hello, world!"
2
[void, object]
returnvoid 3
end
4


Basic Operations

• Stack manipulation
• Local variables
• Math
• Boolean


Stack Operations
0x00 nop Do nothing.
0x57 pop Discard top value from stack
0x58 pop2 Discard top two values
0x59 dup Duplicate and push top value again
0x5A dup_x1 Dup and push top value below second value
0x5B dup_x2 Dup and push top value below third value
0x5C dup2 Dup top two values and push
0x5D dup2_x1 ...below second value
0x5E dup2_x2 ...below third value
0x5F swap Swap top two values


Stack Juggling
Depth Value
dup
0 value_0
pop
swap 1 value_1
dup_x1
2
dup2_x2
3

4


Stack Juggling
Depth Value
dup
0 value_0
pop
swap 1 value_0
dup_x1
2 value_1
dup2_x2
3

4


Stack Juggling
Depth Value
dup
0 value_1
pop
swap 1 value_0
dup_x1
2
dup2_x2
3

4


Stack Juggling
Depth Value
dup
0 value_1
pop
swap 1 value_0
dup_x1
2 value_1
dup2_x2
3

4


Stack Juggling
Depth Value
dup
0 value_1
pop
swap 1 value_0
dup_x1
2 value_1
dup2_x2
3 value_1

4 value_0


Typed Opcodes
<type><operation>
b byte Constant values
s short Local vars (load, store)
c char
Array operations (aload, astore)
i int
Math ops (add, sub, mul, div)
l long
Boolean and bitwise
f ﬂoat
d double Comparisons
a reference Conversions


Where’s boolean?

• Boolean is generally int 0 or 1
• Boolean operations push int 0 or 1
• Boolean branches expect 0 or nonzero
• To set a boolean...use int 0 or 1


Constant Values
0x01 aconst_null Push null on stack
0x02-0x08 iload_[m1-5] Push integer [-1 to 5] on stack
0x09-0x0A lconst_[0,1] Push long [0 or 1] on stack
0x0B-0x0D fconst_[0,1,2] Push ﬂoat [0.0, 1.0, 2.0] on stack
0x0E-0x0F dconst_[0,1] Push double [0.0, 1.0] on stack
0x10 bipush Push byte value to stack as integer
0x11 sipush Push short value to stack as integer
0x12 ldc Push 32-bit constant to stack (int, ﬂoat, string)
0x14 ldc2_w Push 64-bit constant to stack (long, double)


Why So Many?
• Reducing bytecode size
• Special iconst_0 and friends take no args
• bipush, sipush: only 8, 16 bits arguments
• Pre-optimizing JVM
• Specialized instructions can be optimized
• Doesn’t matter at all now

Constant Values
Depth Value

ldc "hello" 0
dconst_1
1
aconst_null
bipush 4 2
ldc_float 2.0
3
4
5


Constant Values
Depth Value

ldc "hello" 0 “hello”
dconst_1
1
aconst_null
bipush 4 2
ldc_float 2.0
3
4
5


Constant Values
Depth Value

ldc "hello" 0
dconst_1
1.0d
1
aconst_null
bipush 4 2 “hello”
ldc_float 2.0
3
4
5


Woah, Two Slots?

• JVM stack slots (and local vars) are 32-bit
• 64-bit values take up two slots
• “wide” before or “w” sufﬁx
• 64-bit ﬁeld updates not atomic!
• Mind those concurrent longs/doubles!


Constant Values
Depth Value

ldc "hello" 0 null
dconst_1
1
aconst_null 1.0d
bipush 4 2
ldc_float 2.0
3 “hello”
4
5


Constant Values
Depth Value

ldc "hello" 0 4
dconst_1
1 null
aconst_null
bipush 4 2
ldc_float 2.0 1.0d
3
4 “hello”
5


Constant Values
Depth Value

ldc "hello" 0 2.0f
dconst_1
1 4
aconst_null
bipush 4 2 null
ldc_float 2.0
3
1.0
4
5 “hello”


Local Variable Table

• Local variables numbered from 0
• Instance methods have “this” at 0
• Separate table maps numbers to names
• Explicitly sized in method deﬁnition


Local Variables
0x15 iload Load integer from local variable onto stack
0x16 lload ...long...
0x17 fload ...float...
0x18 dload ...double...
0x19 aload ...reference...
0x1A-0x2D Packed loads iload_0, aload_3, etc
0x36 istore Store integer from stack into local variable
0x37 lstore ...long...
0x38 fstore ...float...
0x39 dstore ...double...
0x3A astore ...reference...
0x3B-0x4E Packed stores fstore_2, dstore_0, etc
0x84 iinc Add given amount to int local variable

Local Variables
Var Value Depth Value
ldc "hello"
0 bipush 4 0
istore 3
1 dconst_0 1
dstore 1
2 astore 0 2
aload 0
3 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 bipush 4 0 “hello”
istore 3
1 dconst_0 1
dstore 1
2 astore 0 2
aload 0
3 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 bipush 4 0 4
istore 3
1 dconst_0 1 “hello”
dstore 1
2 astore 0 2
aload 0
3 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
istore 3
1 dconst_0 1
dstore 1
2 astore 0 2
aload 0
3 4 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 bipush 4 0
istore 3 0.0
1 dconst_0 1
dstore 1
2 astore 0 2 “hello”
aload 0
3 4 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
istore 3
1 dconst_0 1
0.0 dstore 1
2 astore 0 2
aload 0
3 4 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 “hello” bipush 4 0
istore 3
1 dconst_0 1
0.0 dstore 1
2 astore 0 2
aload 0
3 4 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 “hello” bipush 4 0 “hello”
istore 3
1 dconst_0 1
0.0 dstore 1
2 astore 0 2
aload 0
3 4 iinc 3, 5
3

4 4


Local Variables
ldc "hello"
0 “hello” bipush 4 0 “hello”
istore 3
1 dconst_0 1
0.0 dstore 1
2 astore 0 2
aload 0
3 9 iinc 3, 5
3

4 4


Arrays
0x2E-0x35 [i,l,f,d,a,b,c,d]aload Load [int, long, ...] from array (on stack) to stack

0x4F-0x56 [i,l,f,d,a,b,c,d]astore Store [int, long, ...] from stack to array (on stack)

0xBC newarray Construct new primitive array

0xBD anewarray Construct new reference array

0xBE arraylength Get array length

0xC5 multianewarray Create multi-dimensional array


Arrays
Depth Value
iconst_2
newarray int 0
dup
iconst_0 1
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 2
dup
iconst_0 1
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 int[2] {0,0}
dup
iconst_0 1
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 int[2] {0,0}
dup
iconst_0 1 int[2] {0,0}
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 0
dup
iconst_0 1 int[2] {0,0}
iconst_m1
2 int[2] {0,0}
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 -1
dup
iconst_0 1 0
iconst_m1
2 int[2] {0,0}
iastore
iconst_0 3 int[2] {0,0}
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 int[2] {-1, 0}
dup
iconst_0 1
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 0
dup
iconst_0 1 int[2] {-1, 0}
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Arrays
Depth Value
iconst_2
newarray int 0 -1
dup
iconst_0 1
iconst_m1
2
iastore
iconst_0 3
iaload
4
5


Math Operations
add subtract multiply divide remainder negate
+ - * / % -()

int iadd isub imul idiv irem ineg

long ladd lsub lmul ldiv lrem lneg

ﬂoat fadd fsub fmul fdiv frem fneg

double dadd dsub dmul ddiv drem dneg


Boolean and Bitwise

unsigned
shift left shift right and or xor
shift right

int ishl ishr iushr iand ior ixor


Conversions
To:
int long ﬂoat double byte char short

int - i2l i2f i2d i2b i2c i2s
From:

long l2i - l2f l2d - - -

ﬂoat f2i f2l - f2d - - -

double d2i d2l d2f - - - -


Comparisons
0x94 lcmp Compare two longs, push int -1, 0, 1

0x95 fcmpl Compare two ﬂoats, push in -1, 0, 1 (-1 for NaN)

0x96 fcmpg Compare two ﬂoats, push in -1, 0, 1 (1 for NaN)

0x97 dcmpl Compare two doubles, push in -1, 0, 1 (-1 for NaN)

0x98 dcmpg Compare two doubles, push in -1, 0, 1 (1 for NaN)


Flow Control

• Inspect stack and branch
• Or just branch, via goto
• Labels mark branch targets
• Wide variety of tests


Flow Control
0x99 ifeq If zero on stack, branch
0x9A ifne If nonzero on stack, branch
0x9B iﬂt If stack value is less than zero, branch
0x9C ifge If stack value is greater than or equal to zero, branch
0x9D ifgt If stack value is greater than zero, branch
0x9E iﬂe If stack value is less than or equal to zero, branch
0x9F if_icmpeq If two integers on stack are eq, branch
0xA0 if_icmpne If two integers on stack are ne, branch
0xA1 if_icmplt If two integers on stack are lt, branch
0xA2 if_icmpge If two integers on stack are ge, branch
0xA3 if_icmpgt If tw
If two integers on stack are gt, branch
0xA4 if_icmple If two integers on stack are le, branch
0xA5 if_acmpeq If two references on stack are the same, branch
0xA6 if_acmpne If two references on stack are different, branch
0xA7 goto GOTO!


Other Flow Control
0xA8 jsr Jump to subroutine (deprecated)

0xA9 ret Return from subroutine (deprecated)

0xAA tableswitch Branch using an indexed table of jump offsets

0xAB lookupswitch Branch using a lookup-based table of jump offsets

0xAC-0xB0 [i,l,f,d,a]return Return (int, long, ﬂoat, double, reference) value

0xB1 return Void return (exit method, return nothing)

0xC6 ifnull If reference on stack is null

0xC7 ifnonnull If reference on stack is not null


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch" String[]
invokevirtual string, "equals", 0 {“branch”}
[boolean, object]
ifne :branch
1
ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0 0
[boolean, object]
ifne :branch
1 String[]{“branch”}

ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0 “branch”
[boolean, object]
ifne :branch
1
ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0 “branch”
[boolean, object]
ifne :branch
1 “branch”

ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0 1
[boolean, object]
ifne :branch
1
ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0
[boolean, object]
ifne :branch
1
ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Flow Control
aload 0
ldc 0
aaload Depth Value
ldc "branch"
invokevirtual string, "equals", 0 “Equal!”
[boolean, object]
ifne :branch
1
ldc "Not equal!"
2
aprintln
goto :end 3
label :branch
ldc "Equal!" 4
aprintln
label :end 5
returnvoid


Classes and Types

• Signatures!!!
• Probably the most painful part
• ...but not a big deal if you understand


Using Classes
0xB2 getstatic Fetch static field from class
0xB3 putstatic Set static field in class
0xB4 getfield Get instance field from object
0xB5 setfield Set instance field in object
0xB6 invokevirtual Invoke instance method on object
0xB7 invokespecial Invoke constructor or “super” on object
0xB8 invokestatic Invoke static method on class
0xB9 invokeinterface Invoke interface method on object
0xBA invokedynamic Invoke method dynamically on object (Java 7)
0xBB new Construct new instance of object
0xC0 checkcast Attempt to cast object to type
0xC1 instanceof Push nonzero if object is instanceof specified type


Using Classes
new ArrayList
dup
invokespecial ArrayList, '<init>', Depth Value
[void]
an ArrayList
checkcast Collection 0 (uninitialized)
dup
ldc "first element" 1
invokeinterface Collection, 'add',
[boolean, object] 2
pop
checkcast ArrayList 3
ldc 0
invokevirtual ArrayList, 'get', 4
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
an ArrayList
checkcast Collection 0 (uninitialized)
dup
an ArrayList
ldc "first element" 1 (uninitialized)
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0 an ArrayList
dup
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0 a Collection
dup
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0 a Collection
dup
ldc "first element" 1 a Collection
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
“ﬁrst
checkcast Collection 0 element”
dup
[boolean, object] 2 a Collection
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0 1 (true)
dup
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0 0
dup
ldc "first element" 1 an ArrayList
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
“ﬁrst
checkcast Collection 0 element”
dup
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Using Classes
new ArrayList
dup
[void]
checkcast Collection 0
dup
[boolean, object] 2
pop
ldc 0
[object, int]
aprintln 5
returnvoid


Exceptions and
Synchronization
Table structure for a method indicating start/end of
- trycatch
try/catch and logic to run on exception

0xC2 monitorenter Enter synchronized block against object on stack

0xC3 monitorexit Exit synchronized block (against same object)


More Examples

• A simple loop
• Fibonacci


A Simple Loop
main do
aload 0
push_int 0
aaload
label :top
dup
aprintln
goto :top
returnvoid
end

Fibonacci
public_static_method "fib", [], int, int do
iload 0
ldc 2
if_icmpge :recurse
iload 0
ireturn
label :recurse
iload 0
ldc 1
isub
invokestatic this, "fib", [int, int]
iload 0
ldc 2
isub
iadd
ireturn
end


main do
load_times
istore 1
Fibonacci
ldc "Raw bytecode fib(45) performance:"
aprintln

label :top
iload 1
ifeq :done
iinc 1, -1

start_timing 2
ldc 45
pop
end_timing 2

ldc "Time: "
aprintln
lprintln 2
goto :top

label :done
returnvoid
end


main do
load_times
istore 1
Fibonacci
ldc "Raw bytecode fib(45) performance:"
aprintln

label :top
iload 1
ifeq :done
iinc 1, -1

start_timing 2
Macros
ldc 45
pop
end_timing 2

ldc "Time: "
aprintln
lprintln 2
goto :top

label :done
returnvoid
end


Fibonacci

macro :load_times do
aload 0
ldc 0
aaload # number of times
invokestatic JInteger, 'parseInt',
[int, string]
end


Fibonacci

macro :start_timing do |i|
load_time
lstore i
end


Fibonacci

macro :load_time do
invokestatic System, "currentTimeMillis", long
end


Fibonacci
macro :end_timing do |i|
load_time
lload i
lsub
lstore i
end


Fibonacci
macro :lprintln do |i|
lload i
[void, long]
end


ASM

• “All purpose bytecode manipulation and
analysis framework.”
• De facto standard bytecode library
• http://asm.ow2.org


Basic Process

• Construct a ClassWriter
• Visit structure
• Annotations, methods, ﬁelds, inner classes
• Write out bytes


Blah.java
public class Blah implements Cloneable {
private final String fieldName;

public Blah() {
fieldName = "hello";
}

public static Blah makeBlah() {
return new Blah();
}
}


ClassWriter

ClassWriter cv = new ClassWriter(
ClassWriter.COMPUTE_MAXS |
ClassWriter.COMPUTE_FRAMES);


COMPUTE...what?
• COMPUTE_MAXS
• ASM will calculate max stack/local vars
• COMPUTE_FRAMES
• ASM will calculate Java 6 stack map
• Hints to veriﬁer that we’ve pre-validated
stack contents (sort of)


Visit Class
cv.visit(
Opcodes.V1_6,
Opcodes.ACC_PUBLIC,
"Blah",
null,
"java/lang/Object",
new String[] {"java/lang/Cloneable"});


Opcodes

• Interface full of constants
• Bytecodes
• Visibility modiﬁers
• Java versions
• Other stuff


ACC_*

• Some you know
• ACC_PUBLIC, ACC_ABSTRACT, etc
• Some you don’t
• ACC_BRIDGE, ACC_SYNTHETIC


Java Version

• V1_1 through V1_7
• Sorry 1.0!


Class Names
"java/lang/Object"

packageClass.replaceAll('.', '/')


Visit Source

cv.visitSource(
"Blah.java",
"JSR-45 source map here");


Visit Annotation

AnnotationVisitor av = cv.visitAnnotation("some/Annotation", true);
av.visitArray("name1", ...);
av.visitEnum("name2", ...);
av.visitEnd();


Blah.java

private final String fieldName;


Visit Field
FieldVisitor fv = cv.visitField(
Opcodes.ACC_PRIVATE | Opcodes.ACC_FINAL,
"fieldName",
"Ljava.lang.String;",
null);
fv.visitAnnotation(...);
fv.visitAttribute(...);
fv.visitEnd();


Descriptor
"Ljava.lang.String;"
"(IF[JLjava.lang.Object;)V"

• Primitive types
• B,C,S,I,J,F,D,Z,V
• Reference types
• Lsome/Class;
• Array
• Preﬁx with [

Blah.java

public Blah() {
...
}

...
}


Visit Method
MethodVisitor construct = cv.visitMethod(
Opcodes.ACC_PUBLIC,
"<init>",
"()V",
null,
null);
MethodVisitor makeBlah = cv.visitMethod(
Opcodes.ACC_PUBLIC, ACC_STATIC,
"makeBlah",
"()LBlah;",
null,
null);


Special Methods

• <init>
• Constructor
• <clinit>
• Static initializer


MethodVisitor

• Visit annotation stuff
• Visit code
• Bytecodes, frames, local vars, line nums
• Visit maxs
• Pass bogus values if COMPUTE_MAXS


Blah.java

public Blah() {
fieldName = "hello";
}

return new Blah();
}


Visit Method Body
construct.visitCode();
construct.visitVarInsn(ALOAD, 0);
construct.visitMethodInsn(INVOKESPECIAL,
"java/lang/Object",
"<init>",
"()V");
construct.visitVarInsn(ALOAD, 0);
construct.visitLdcInsn("hello");
construct.visitFieldInsn(PUTFIELD,
"Blah",
"fieldName",
"Ljava/lang/String;");
construct.visitInsn(RETURN);
construct.visitMaxs(2, 1);
construct.visitEnd();


ASMiﬁerClassVisitor

• Dump ASM visitor calls from .class ﬁle
• Very raw, but very useful


~/oscon ➔ java -cp asm-3.3.1.jar:asm-util-3.3.1.jar
org.objectweb.asm.util.ASMifierClassVisitor
Blah.class
import java.util.*;
import org.objectweb.asm.*;
import org.objectweb.asm.attrs.*;
public class BlahDump implements Opcodes {
public static byte[] dump () throws Exception {

ClassWriter cw = new ClassWriter(0);
FieldVisitor fv;
MethodVisitor mv;
AnnotationVisitor av0;

cw.visit(V1_6, ACC_PUBLIC + ACC_SUPER, "Blah", null, "java/lang/Object", new String[] { "java/lang/Cloneable" });
{

fv = cw.visitField(ACC_PRIVATE + ACC_FINAL, "fieldName", "Ljava/lang/String;", null, null);
fv.visitEnd();
}
{

mv = cw.visitMethod(ACC_PUBLIC, "<init>", "()V", null, null);
mv.visitCode();
mv.visitVarInsn(ALOAD, 0);
mv.visitMethodInsn(INVOKESPECIAL, "java/lang/Object", "<init>", "()V");
mv.visitVarInsn(ALOAD, 0);
mv.visitLdcInsn("hello");
mv.visitFieldInsn(PUTFIELD, "Blah", "fieldName", "Ljava/lang/String;");
mv.visitInsn(RETURN);
mv.visitMaxs(2, 1);
mv.visitEnd();
}
{

mv = cw.visitMethod(ACC_PUBLIC + ACC_STATIC, "makeBlah", "()LBlah;", null, null);
mv.visitCode();
mv.visitTypeInsn(NEW, "Blah");
mv.visitInsn(DUP);
mv.visitMethodInsn(INVOKESPECIAL, "Blah", "<init>", "()V");
mv.visitInsn(ARETURN);
mv.visitMaxs(2, 0);
mv.visitEnd();
}
cw.visitEnd();

return cw.toByteArray();
}
}


Real-world Cases
• Reﬂection-free invocation
• JRuby, Groovy, other languages
• Bytecoded data objects
• Hibernate, other data layers
• java.lang.reﬂect.Proxy and others
• Language compilers

Part Two:
JVM JIT


JIT

• Just-In-Time compilation
• Compiled when needed
• Maybe immediately before execution
• ...or when we decide it’s important
• ...or never?


Mixed-Mode
• Interpreted
• Bytecode-walking
• Artiﬁcial stack
• Compiled
• Direct native operations
• Native registers, memory, etc

Proﬁling

• Gather data about code while interpreting
• Invariants (types, constants, nulls)
• Statistics (branches, calls)
• Use that information to optimize
• Educated guess?


Optimization

• Loop unrolling
• Lock coarsening
• Method inlining
• Dead code elimination
• Duplicate code elimination


The Golden Rule
of Optimization

Don’t do unnecessary work.


Perf Sinks
• Memory accesses
• By far the biggest expense
• Calls
• Opaque memory ref + branch
• Locks, volatile writes
• Kills multi-cpu perf

Volatile?
• Each CPU maintains a memory cache
• Caches may be out of sync
• If it doesn’t matter, no problem
• If it does matter, threads disagree!
• Volatile forces synchronization of cache
• Across cores and to main memory

Inlining?

• Combine caller and callee into one unit
• e.g. based on proﬁle
• Perhaps with a sanity check
• Optimize as a whole


Inlining
int addAll(int max) {
int accum = 0;
for (int i = 0; i < max; i++) {
accum = add(accum, i);
}
return accum;
}

int add(int a, int b) {
return a + b;
}


Inlining
int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum;
} Only one target is ever seen

int add(int a, int b) {
return a + b;
}


Inlining

int accum = 0;
for (int i = 0; i < max; i++) {
accum = accum + i;
}
return accum;
} Don’t bother making a call


Call Site
• The place where you make a call
• Monomorphic (“one shape”)
• Single target class
• Bimorphic (“two shapes”)
• Polymorphic (“many shapes”)
• Megamorphic (“you’re screwed”)

Blah.java
System.currentTimeMillis(); // static, monomorphic

List list1 = new ArrayList(); // constructor, monomorphic
List list2 = new LinkedList();

for (List list : new List[]{ list1, list2 }) {
list.add("hello"); // bimorphic
}

for (Object obj : new Object[]{ 'foo', list1, new Object() }) {
obj.toString(); // polymorphic
}


Hotspot
• -client mode (C1) inlines, less aggressive
• Fewer opportunities to optimize
• -server mode (C2) proﬁles, inlines
• We’ll focus on this
• Tiered mode combines them
• -XX:+TieredCompilation

Hotspot Inlining

• Profile to find “hot spots”
• Largely focused around call sites
• Profile until 10k calls
• Inline mono/bimorphic calls
• Other mechanisms for polymorphic calls


Now it gets fun!


Monitoring the JIT
• Dozens of ﬂags
• PrintCompilation
• PrintInlining
• LogCompilation
• PrintAssembly
• Some in product, some in debug...

public class Accumulator {
int max = Integer.parseInt(args[0]);
System.out.println(addAll(max));
}

static int addAll(int max) {
int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum;
}

static int add(int a, int b) {
return a + b;
}
}


~/oscon ➔ java -version
openjdk version "1.7.0-internal"
OpenJDK Runtime Environment (build 1.7.0-internal-b00)
OpenJDK 64-Bit Server VM (build 21.0-b17, mixed mode)

~/oscon ➔ javac Accumulator.java

~/oscon ➔ java Accumulator 1000
499500


PrintCompilation

• -XX:+PrintCompilation
• Print methods as they are jitted
• Class + name + size


~/oscon ➔ java -XX:+PrintCompilation Accumulator 1000
1 java.lang.String::hashCode (64 bytes)
2 java.math.BigInteger::mulAdd (81 bytes)
3 java.math.BigInteger::multiplyToLen (219 bytes)
4 java.math.BigInteger::addOne (77 bytes)
5 java.math.BigInteger::squareToLen (172 bytes)
6 java.math.BigInteger::primitiveLeftShift (79 bytes)
7 java.math.BigInteger::montReduce (99 bytes)
8 sun.security.provider.SHA::implCompress (491 bytes)
9 java.lang.String::charAt (33 bytes)
499500


499500
Where’s our methods?!


499500
Where’s our methods?!
...remember...10k calls


10k loop, 10k calls to add

10 Accumulator::add (4 bytes)
49995000

Hooray!


What’s this stuff?


What’s this stuff?

Class loading, security, other boot logic.


What if you see this...
~/oscon ➔ java -client -XX:+PrintCompilation Accumulator 1000
5 made not entrant java.math.BigInteger::squareToLen (172 bytes)
499500


Optimistic Compiler

• Assume proﬁle is accurate
• Aggressively optimize based on proﬁle
• Bail out if we’re wrong
• And hope that we’re usually right


Deoptimization

• Bail out of running code
• Monitoring ﬂags describe process
• “uncommon trap” - we were wrong
• “not entrant” - don’t let anyone enter
• “zombie” - on its way to deadness


--- n java.lang.System::arraycopy (static)
25 java.lang.String$CaseInsensitiveComparator::compare (115
bytes)
26 java.lang.Character::toLowerCase (162 bytes)


--- n java.lang.System::arraycopy (static)
25 java.lang.String$CaseInsensitiveComparator::compare (115
bytes)
26 java.lang.Character::toLowerCase (162 bytes)

Native calls don’t compile, may be
intrinsic. We’ll come back to that.


PrintInlining

• -XX:+UnlockDiagnosticVMOptions
-XX:+PrintInlining
• Display hierarchy of inlined methods
• Include reasons for not inlining
• More, better output on OpenJDK 7


~/oscon ➔ java -XX:+UnlockDiagnosticVMOptions
> -XX:+PrintInlining
> Accumulator 10000
49995000


~/oscon ➔ java -XX:+UnlockDiagnosticVMOptions
> -XX:+PrintInlining
> Accumulator 10000
49995000

Um...I don’t see anything inlining...


public class Accumulator {
int max = Integer.parseInt(args[0]);
System.out.println(addAll(max));
}
Called once
static int addAll(int max) {
int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum; Called 10k times...
}

static int add(int a, int b) {
return a + b;
} ...but no calls to inline!
}


JVM for Dummies - OSCON 2011

More Related Content

JVM for Dummies - OSCON 2011

Editor's Notes