Øredev 2011 - JVM JIT for Dummies (What the JVM Does With Your Bytecode When You're Not Looking)

What the JVM Does
With Your Bytecode
When Nobody’s Looking

JVM JIT for Dummies
And the rest of you, too.

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

What We Won’t

• GC tuning
• GC monitoring with VisualVM
• Google ‘visualgc’, it’s awesome

What We Will Lean

• How the JVM’s JIT works
• Monitoring the JIT
• Finding problems
• Dumping assembly (don’t be scared!)

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?

The Golden Rule of
Optimization

Don’t do unnecessary work.

Optimization
• Method inlining
• Loop unrolling
• Lock coarsening/eliding
• Dead code elimination
• Duplicate code elimination
• Escape analysis

Inlining?

• Combine caller and callee into one unit
• e.g. based on proﬁle
• Perhaps with a guard/test
• Optimize as a whole
• More code means better visibility

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 the call
}

Loop Unrolling

• Works for small, constant loops
• Avoid tests, branching
• Allow inlining a single call as many

Loop Unrolling
private static final String[] options =
{ "yes", "no", "maybe"};
public void looper() {
for (String option : options) {
process(option);
}
}
Small loop, constant stride,
constant size

Loop Unrolling
private static final String[] options =
{ "yes", "no", "maybe"};
public void looper() {
process(options[0]);
process(options[1]); Unrolled!
process(options[2]);
}

Lock Coarsening
public void needsLocks() {
for (option : options) {
process(option);
} Repeatedly locking
}

private synchronized String process(String option) {
// some wacky thread-unsafe code
}

Lock Coarsening
public void needsLocks() { Lock once
synchronized (this) {
// some wacky thread-unsafe code
}
}
}

Lock Eliding
public void overCautious() { Synchronize on
List l = new ArrayList();
synchronized (l) {
new Object
l.add(process(option));
}
}
}
But we know it
never escapes this
thread...

Lock Eliding
public void overCautious() {
List l = new ArrayList();
l.add(
/* process()’s code */);
}
}
No need to lock

Escape Analysis
private static class Foo {
public String a;
public String b;
Foo(String a, String b) {
this.a = a;
this.b = b;
}
}

Escape Analysis
public void bar() {
Foo f = new Foo("Hello", "Øredev");
baz(f);
}

public void baz(Foo f) { Same object all
System.out.print(f.a);
System.out.print(", "); the way through
quux(f);
}
Never “escapes”
public void quux(Foo f) { these methods
System.out.print(f.b);
System.out.println('!');
}

Escape Analysis

public secret awesome inlinedBarBazQuux() {
System.out.print("Hello");
System.out.print(", ");
System.out.print("Øredev");
System.out.println('!');
}
Don’t bother allocating
Foo object!

Perf Sinks
• Memory accesses
• By far the biggest expense
• Calls
• Memory ref + branch kills pipeline
• Call stack, register juggling costs
• Locks and volatile writes

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

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) inlines aggressively
• Based on richer runtime proﬁling
• We’ll focus on this
• Tiered mode combines them
• -XX:+TieredCompilation

C2 “server” Inlining
• Profile to find “hot spots”
• Call sites
• Branch statistics
• Profile until 10k calls
• Inline mono/bimorphic calls
• Other mechanisms for polymorphic calls

Tuning Inlining
• -XX:+MaxInlineSize=35
• Largest inlinable method (bytecode)
• -XX:+InlineSmallCode=#
• Largest inlinable compiled method
• -XX:+FreqInlineSize=#
• Largest frequently-called method...

Tuning Inlining

• -XX:+MaxInlineLevel=9
• How deep does the rabbit hole go?
• -XX:+MaxRecursiveInlineLevel=#
• Recursive inlining

Monitoring the JIT

• Dozens of ﬂags
• Reams of output
• Always evolving
• How can you understand it?

public class Accumulator {
public static void main(String[] args) {
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;
}
}

$ java -version
openjdk version "1.7.0-b147"
OpenJDK Runtime Environment (build 1.7.0-
b147-20110927)
OpenJDK 64-Bit Server VM (build 21.0-b17, mixed mode)

$ javac Accumulator.java

$ java Accumulator 1000
499500

Print Compilation

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

$ java -XX:+PrintCompilation Accumulator 1000
53 1 java.lang.String::hashCode (67 bytes)
499500

499500

Where’s our code?!?

499500

Where’s our code?!?
Remember...10k calls before JIT

10k loop, 10k calls to add

64 2 Accumulator::add (4 bytes)
49995000

Hooray!

But what’s this?

64 2 Accumulator::add (4 bytes)
49995000

Class loading, security logic, other stuff...

Dear god...there’s zombies in my code?!?
1401 70 java.util.concurrent.ConcurrentHashMap::hash (49 bytes)
1412 71 java.lang.String::indexOf (7 bytes)
1420 72 ! java.io.BufferedReader::readLine (304 bytes)
1420 73 sun.nio.cs.UTF_8$Decoder::decodeArrayLoop (543 bytes)
1422 42 java.util.zip.ZipCoder::getBytes (192 bytes) made not entrant
1435 74 n java.lang.Object::hashCode (0 bytes)
1443 29 ! sun.misc.URLClassPath$JarLoader::getResource (91 bytes) made zombie
1443 25 sun.misc.URLClassPath::getResource (74 bytes) made zombie
1443 36 sun.misc.URLClassPath::getResource (74 bytes) made not entrant
1443 43 java.util.zip.ZipCoder::encoder (35 bytes) made not entrant
1449 75 java.lang.String::endsWith (15 bytes)
1631 1 % sun.misc.URLClassPath::getResource @ 39 (74 bytes)
1665 76 java.lang.ClassLoader::checkName (43 bytes)

Dear god...there’s zombies in my code?!?

Not entrant? What the heck?

Optimistic Compilers

• 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 new calls enter
• “zombie” - on its way to deadness

No JIT At All?

• Code is too big
• Code isn’t called enough

That looks exciting!


Exception handling in here (boring!)


What’s this “n” all about?


This method is native...maybe “intrinsic”


We’ll come back to that...

And this one?


Method has been replaced while running (OSR)

Millis from JVM start

Sequence number of compilation

Print Inlining

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

$ java -XX:+UnlockDiagnosticVMOptions
> -XX:+PrintInlining
> Accumulator 10000
49995000

> Accumulator 10000
49995000
Um...I don’t see anything inlining

}

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

return a + b;
}
}

}

int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum;
Called 10k times
}

return a + b;
}
}

}

int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum;
Called 10k times
}

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

}

int accum = 0;
for (int i = 0; i < max; i++) {
}
return accum;
Called 10k times
}

static int add(int a, int b) { JITs as expected
return a + b;
}
} But makes no calls!

static double addAllSqrts(int max) {
double accum = 0;
for (int i = 0; i < max; i++) {
accum = addSqrt(accum, i);
}
return accum;
}

static int addSqrt(double a, int b) {
return a + sqrt(b);
}

static double sqrt(int a) {
return Math.sqrt(b);
}

> -XX:+PrintCompilation
> Accumulator 10000
65 2 Accumulator::addSqrt (7 bytes)
@ 3 Accumulator::sqrt (6 bytes) inline (hot)
@ 2 java.lang.Math::sqrt (5 bytes) (intrinsic)
65 3 Accumulator::sqrt (6 bytes)
666616.4591971082

> -XX:+PrintInlining HOT HOT HOT!
> Accumulator 10000
666616.4591971082

LogCompilation

• -XX:+LogCompilation
• Worst. XML. Evar.
• <JDK>/hotspot/src/share/tools/LogCompilation
• or http://github.com/headius/logc

scopes_pcs_offset='1384' dependencies_offset='1576' handler_table_offset='1592' nul_chk_table_offset='1736'
oops_offset='992' method='org/jruby/lexer/yacc/ByteArrayLexerSource$ByteArrayCursor read ()I' bytes='49'
count='5296' backedge_count='1' iicount='10296' stamp='0.412'/>
<writer thread='4425007104'/>
<nmethod compile_id='21' compiler='C2' entry='4345862528' size='1152' address='4345862160'
relocation_offset='288' insts_offset='368' stub_offset='688' scopes_data_offset='840' scopes_pcs_offset='904'
dependencies_offset='1016' handler_table_offset='1032' oops_offset='784' method='org/jruby/lexer/yacc/
ByteArrayLexerSource forward (I)I' bytes='111' count='5296' backedge_count='1' iicount='10296' stamp='0.412'/>
<task_queued compile_id='22' method='org/jruby/lexer/yacc/ByteArrayLexerSource read ()I' bytes='10'
count='5000' backedge_count='1' iicount='10000' stamp='0.433' comment='count' hot_count='10000'/>
relocation_offset='288' insts_offset='400' stub_offset='912' scopes_data_offset='1104'
scopes_pcs_offset='1496' dependencies_offset='1704' handler_table_offset='1720' nul_chk_table_offset='1864'
oops_offset='1024' method='org/jruby/lexer/yacc/ByteArrayLexerSource read ()I' bytes='10' count='5044'
backedge_count='1' iicount='10044' stamp='0.435'/>
<task_queued compile_id='23' method='java/util/HashMap hash (I)I' bytes='23' count='5000' backedge_count='1'
iicount='10000' stamp='0.442' comment='count' hot_count='10000'/>
relocation_offset='288' insts_offset='304' stub_offset='368' scopes_data_offset='392' scopes_pcs_offset='400'
dependencies_offset='432' method='java/util/HashMap hash (I)I' bytes='23' count='5039' backedge_count='1'
iicount='10039' stamp='0.442'/>
<dependency_failed type='abstract_with_unique_concrete_subtype' ctxk='org/jruby/lexer/yacc/LexerSource'
x='org/jruby/lexer/yacc/ByteArrayLexerSource' witness='org/jruby/lexer/yacc/InputStreamLexerSource'
stamp='0.456'/>
stamp='0.456'/>
stamp='0.456'/>
stamp='0.456'/>

$ java -jar logc.jar hotspot.log
1 java.lang.String::hashCode (67 bytes)
2 Accumulator::addSqrt (7 bytes)
3 Accumulator::sqrt (6 bytes)

$ java -jar logc.jar hotspot.log
1 java.lang.String::hashCode (67 bytes)
2 Accumulator::addSqrt (7 bytes)
@ 2 Accumulator::sqrt (6 bytes) (end time: 0.0660 nodes: 36)
@ 2 java.lang.Math::sqrt (5 bytes)
3 Accumulator::sqrt (6 bytes)
@ 2 java.lang.Math::sqrt (5 bytes)

8 sun.nio.cs.UTF_8$Encoder::encode (361 bytes)
6 uncommon trap null_check make_not_entrant
@8 java/lang/String equals (Ljava/lang/Object;)Z
6 make_not_entrant
9 java.lang.String::equals (88 bytes)
10 java.util.LinkedList::indexOf (73 bytes)

Hotspot sees it’s 100% String
10 java.util.LinkedList::indexOf (73 bytes)
@ 52 java.lang.Object::equals (11 bytes)
type profile java/lang/Object -> java/lang/String (100%)
@ 52 java.lang.String::equals (88 bytes)
11 java.lang.String::indexOf (87 bytes)
@ 83 java.lang.String::indexOfSupplementary too big

Too big to inline! Could be bad?

Intrinsic?

• Known to the JIT
• Don’t inline bytecode
• Do insert “best” native code
• e.g. kernel-level memory operation
• e.g. optimized sqrt in machine code

> Accumulator 10000
666616.4591971082

Calls treated specially by JIT

Common Intrinsics
• String#equals
• Most (all?) Math methods
• System.arraycopy
• Object#hashCode
• Object#getClass
• sun.misc.Unsafe methods

Did someone say
MACHINE CODE?!

The Red Pill

• Knowing code compiles is good
• Knowing code inlines is better
• Seeing the actual assembly is best!

Caveat

• I don’t really know assembly.
• But I fake it really well.

Print Assembly

• -XX:+PrintAssembly
• Google “hotspot printassembly”
• http://wikis.sun.com/display/
HotSpotInternals/PrintAssembly
• Assembly-dumping plugin for Hotspot

Alternative

• -XX:+PrintOptoAssembly
• Only in debug/fastdebug builds
• Not as pretty

Wednesday, July 27, 2011

~/oscon ! java -XX:+UnlockDiagnosticVMOptions
> -XX:+PrintAssembly
> Accumulator 10000
OpenJDK 64-Bit Server VM warning: PrintAssembly is enabled;
turning on DebugNonSafepoints to gain additional output
Loaded disassembler from hsdis-amd64.dylib
...

Decoding compiled method 11343cbd0:
Code:
[Disassembling for mach='i386:x86-64']
[Entry Point]
[Verified Entry Point]
[Constants]
# {method} 'add' '(II)I' in 'Accumulator'
# parm0: rsi = int
# parm1: rdx = int
# [sp+0x20] (sp of caller)
11343cd00: push %rbp
11343cd01: sub $0x10,%rsp
11343cd05: nop ;*synchronization entry
; - Accumulator::add@-1 (line 16)
11343cd06: mov %esi,%eax
11343cd08: add %edx,%eax ;*iadd
; - Accumulator::add@2 (line 16)
11343cd0a: add $0x10,%rsp
11343cd0e: pop %rbp
11343cd0f: test %eax,-0x1303fd15(%rip) # 1003fd000
; {poll_return}
11343cd15: retq

x86_64 Assembly 101
add Two’s complement add
sub ...subtract
mov* Move data from a to b
jmp goto
je, jne, jl, jge, ... Jump if ==, !=, <, >=, ...
push, pop Call stack operations
call*, ret* Call, return from subroutine
eax, ebx, esi, ... 32-bit registers
rax, rbx, rsi, ... 64-bit registers

Register Machine

• Instead of stack moves, we have “slots”
• Move data into slots
• Trigger operations that manipulate data
• Get new data out of slots
• JVM stack, locals end up as register ops

Stack?

• Native code has a stack too
• Maintains registers from call to call
• Various calling conventions
• Caller saves registers?
• Callee saves registers?

Decoding compiled method 11343cbd0: <= address of new compiled code
Code:
[Disassembling for mach='i386:x86-64'] <= architecture
[Entry Point]
[Verified Entry Point]
[Constants]
# {method} 'add' '(II)I' in 'Accumulator' <= method, signature, class
# parm0: rsi = int <= first parm to method goes in rsi
# parm1: rdx = int <= second parm goes in rdx
# [sp+0x20] (sp of caller) <= caller’s pointer into native stack

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

rbp points at current stack frame, so we save it off.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Two args, so we bump stack pointer by 0x10.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Do nothing, e.g. to memory-align code.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

At the “-1” instruction of our add() method...
i.e. here we go!

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Move parm1 into eax.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Add parm0 and parm1, store result in eax.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

How nice, Hotspot shows us this is our “iadd” op!

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Put stack pointer back where it was.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Restore rbp from stack.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

Poll a “safepoint”...give JVM a chance to GC, etc.

11343cd01: sub $0x10,%rsp
11343cd0e: pop %rbp
; {poll_return}
11343cd15: retq

All done!

Things to Watch For

• CALL operations
• Indicates something failed to inline
• LOCK operations
• Cache-busting, e.g. volatility

CALL
1134858f5: xchg %ax,%ax
1134858f7: callq 113414aa0 ; OopMap{off=316}
;*invokespecial addAsBignum
; - org.jruby.RubyFixnum::addFixnum@29 (line 348)
; {optimized virtual_call}
1134858fc: jmpq 11348586d

Ruby integer adds might overﬂow into Bignum, leading to
addAsBignum call. In this case, it’s never called, so Hotspot
emits callq assuming we won’t hit it.

LOCK
Code from a RubyBasicObject’s default constructor.
11345d823: mov 0x70(%r8),%r9d ;*getstatic NULL_OBJECT_ARRAY
; - org.jruby.RubyBasicObject::<init>@5 (line 76)
; - org.jruby.RubyObject::<init>@2 (line 118)
; - org.jruby.RubyNumeric::<init>@2 (line 111)
; - org.jruby.RubyInteger::<init>@2 (line 95)
; - org.jruby.RubyFixnum::<init>@5 (line 112)
; - org.jruby.RubyFixnum::newFixnum@25 (line 173)
11345d827: mov %r9d,0x14(%rax)
11345d82b: lock addl $0x0,(%rsp) ;*putfield varTable
; - org.jruby.RubyBasicObject::<init>@8 (line 76)
; - org.jruby.RubyObject::<init>@2 (line 118)
; - org.jruby.RubyNumeric::<init>@2 (line 111)
; - org.jruby.RubyInteger::<init>@2 (line 95)
; - org.jruby.RubyFixnum::<init>@5 (line 112)
; - org.jruby.RubyFixnum::newFixnum@25 (line 173)

Why are we doing a volatile write in the constructor?

LOCK
public class RubyBasicObject ... {
private static final boolean DEBUG = false;
private static final Object[] NULL_OBJECT_ARRAY = new Object[0];

// The class of this object
protected transient RubyClass metaClass;

// zeroed by jvm
protected int flags;

// variable table, lazily allocated as needed (if needed)
private volatile Object[] varTable = NULL_OBJECT_ARRAY;

Maybe it’s not such a good idea to pre-init a volatile?

LOCK

~/projects/jruby ! git log 2f935de1e40bfd8b29b3a74eaed699e519571046 -1 | cat
commit 2f935de1e40bfd8b29b3a74eaed699e519571046
Author: Charles Oliver Nutter <headius@headius.com>
Date: Tue Jun 14 02:59:41 2011 -0500

Do not eagerly initialize volatile varTable field in RubyBasicObject;
speeds object creation significantly.

LEVEL UP!

invokedynamic?

• Largely, it works the same
• MethodHandles optimize to x86 asm
• Inlining as normal
• Performance nearly the same as static!
• And that’s exactly the point!

What Have We Learned?

• How Hotspot’s JIT works
• How to monitor the JIT
• How to find problems
• How to fix problems we find

What We Missed

• Tuning GC settings in JVM
• Monitoring GC with VisualVM
• Google ‘visualgc’...it’s awesome

You’re no dummy now!
;-)

Thank you!

• headius@headius.com, @headius
• http://blog.headius.com
• “java virtual machine speciﬁcation”
• “jvm opcodes”

Øredev 2011 - JVM JIT for Dummies (What the JVM Does With Your Bytecode When You're Not Looking)

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (10)

Similar to Øredev 2011 - JVM JIT for Dummies (What the JVM Does With Your Bytecode When You're Not Looking)

Similar to Øredev 2011 - JVM JIT for Dummies (What the JVM Does With Your Bytecode When You're Not Looking) (20)

More from Charles Nutter

More from Charles Nutter (20)

Recently uploaded

Recently uploaded (20)

Øredev 2011 - JVM JIT for Dummies (What the JVM Does With Your Bytecode When You're Not Looking)

Editor's Notes