Building High-Performance Language Implementations With Low Effort

Building High-Performance
Language Implementations
With Low Effort
Stefan Marr
FOSDEM 2015, Brussels, Belgium
January 31st, 2015
@smarr
http://stefan-marr.de

Why should you care about how
Programming Languages work?
2
SMBC: http://www.smbc-comics.com/?id=2088

3
SMBC: http://www.smbc-comics.com/?id=2088
Why should you care about how
Programming Languages work?
• Performance isn’t magic
• Domain-specific languages
• More concise
• More productive
• It’s easier than it looks
• Often open source
• Contributions welcome

What’s “High-Performance”?
4
Based on latest data from http://benchmarksgame.alioth.debian.org/
Geometric mean over available benchmarks.
Disclaimer: Not indicate for application performance!
Competitively Fast!
0
3
5
8
10
13
15
18
Java V8 C# Dart Python Lua PHP Ruby

Small and
Manageable
16
260
525
562
1 10 100 1000
What’s “Low Effort”?
5
KLOC: 1000 Lines of Code, without blank lines and comments
V8 JavaScript
HotSpot
Java Virtual Machine
Dart VM
Lua 5.3 interp.

Language Implementation Approaches
6
Source
Program
Interpreter
Run TimeDevelopment
Time
Input
Output
Source
Program
Compiler Binary
Input
Output
Run TimeDevelopment
Time
Simple, but often slow More complex, but often faster
Not ideal for all languages.

Modern Virtual Machines
7
Source
Program
Interpreter
Run TimeDevelopment Time
Input
Output
Binary
Runtime Info
Compiler
Virtual Machine
with
Just-In-Time
Compilation

VMs are Highly Complex
8
Interpreter
Input
Output
Compiler Optimizer
Garbage
Collector
CodeGen
Foreign
Function
Interface
Threads
and
Memory
Model
How to reuse most parts
for a new language?
Debugging
Profiling
…
Easily
500 KLOC

How to reuse most parts
for a new language?
9
Input
Output
Make Interpreters Replaceable Components!
Interpreter
Compiler Optimizer
Garbage
Collector
CodeGen
Foreign
Function
Interface
Threads
and
Memory
Model
Garbage
Collector
…
Interpreter
Interpreter
…

Interpreter-based Approaches
Truffle + Graal
with Partial Evaluation
Oracle Labs
RPython
with Meta-Tracing
[3] Würthinger et al., One VM to Rule Them All, Onward!
2013, ACM, pp. 187-204.
[2] Bolz et al., Tracing the Meta-level: PyPy's Tracing JIT
Compiler, ICOOOLPS Workshop 2009, ACM, pp. 18-25.

SELF-OPTIMIZING TREES
A Simple Technique for Language Implementation and Optimization
[1] Würthinger, T.; Wöß, A.; Stadler, L.; Duboscq, G.; Simon, D. & Wimmer, C. (2012), Self-
Optimizing AST Interpreters, in 'Proc. of the 8th Dynamic Languages Symposium' , pp. 73-82.

Code Convention
12
Python-ish
Interpreter Code
Java-ish
Application Code

A Simple
Abstract Syntax Tree Interpreter
13
root_node = parse(file)
root_node.execute(Frame())
if (condition) {
cnt := cnt + 1;
} else {
cnt := 0;
}
cnt
1
+
cnt:
=
if
cnt:
=
0
cond
root_node

Implementing AST Nodes
14
if (condition) {
cnt := cnt + 1;
} else {
cnt := 0;
}
class Literal(ASTNode):
final value
def execute(frame):
return value
class VarWrite(ASTNode):
child sub_expr
final idx
def execute(frame):
val := sub_expr.execute(frame)
frame.local_obj[idx]:= val
return val
class VarRead(ASTNode):
final idx
def execute(frame):
return frame.local_obj[idx]
cnt
1
+
cnt:
=
if
cnt:
=
0
cond

Self-Optimization by Node Specialization
15
cnt := cnt + 1
def UninitVarWrite.execute(frame):
val := sub_expr.execute(frame)
return specialize(val).
execute_evaluated(frame, val)
uninitialized
variable write
cnt
1
+
cnt:
=
cnt:
=
def UninitVarWrite.specialize(val):
if val instanceof int:
return replace(IntVarWrite(sub_expr))
elif …:
…
else:
return replace(GenericVarWrite(sub_expr))
specialized

Self-Optimization by Node Specialization
16
cnt := cnt + 1
def IntVarWrite.execute(frame):
try:
val := sub_expr.execute_int(frame)
return execute_eval_int(frame, val)
except ResultExp, e:
return respecialize(e.result).
execute_evaluated(frame, e.result)
def IntVarWrite.execute_eval_int(frame, anInt):
frame.local_int[idx] := anInt
return anInt
int
variable write
cnt
1
+
cnt:
=

Some Possible Self-Optimizations
• Type profiling and specialization
• Value caching
• Inline caching
• Operation inlining
• Library Lowering
17

Library Lowering for Array class
createSomeArray() { return Array.new(1000, ‘fast fast fast’); }
18
class Array {
static new(size, lambda) {
return new(size).setAll(lambda);
}
setAll(lambda) {
forEach((i, v) -> { this[i] = lambda.eval(); });
}
}
class Object {
eval() { return this; }
}

Optimizing for Object Values
19
.new
Array
global lookup
method
invocation
1000
int literal
‘fast’
string literal
Object, but not a lambda
Optimization
potential

Specialized new(size, lambda)
def UninitArrNew.execute(frame):
size := size_expr.execute(frame)
val := val_expr.execute(frame)
return specialize(size, val).
execute_evaluated(frame, size, val)
20
def UninitArrNew.specialize(size, val):
if val instanceof Lambda:
return replace(StdMethodInvocation())
else:
return replace(ArrNewWithValue())

Specialized new(size, lambda)
def ArrNewWithValue.execute_evaluated(frame, size,
val):
return Array([val] * 1000)
21
1 specialized node vs. 1000x `this[i] = lambda.eval()`
1000x `eval() { return this; }`
.new
Array
global lookup
1000
int literal
‘fast’
string literal
specialized

JUST-IN-TIME COMPILATION FOR
INTERPRETERS
Generating Efficient Native Code
22

How to Get Fast Program Execution?
23
VarWrite.execute(frame)
IntVarWrite.execute(frame)
VarRead.execute(frame)
Literal.execute(frame)
ArrayNewWithValue.execute(frame)
..VW_execute() # bin
..IVW_execute() # bin
..VR_execute() # bin
..L_execute() # bin
..ANWV_execute() # bin
Standard Compilation: 1 node at a time
Minimal Optimization Potential

Problems with Node-by-Node Compilation
24
cnt
1
+
cnt:
=
Slow Polymorphic Dispatches
def IntVarWrite.execute(frame):
try:
val := sub_expr.execute_int(frame)
return execute_eval_int(frame, val)
except ResultExp, e:
return respecialize(e.result).
execute_evaluated(frame, e.result)
cnt:
=
Runtime checks in general

Compilation Unit based on User Program
Meta-Tracing Partial Evaluation
Guided By AST
25
cnt
1
+
cnt:
=
if
cnt:
=
0
cnt
1
+
cnt:
=if cnt:
=
0
[3] Würthinger et al., One VM to Rule Them
All, Onward! 2013, ACM, pp. 187-204.
[2] Bolz et al., Tracing the Meta-level: PyPy's
Tracing JIT Compiler, ICOOOLPS Workshop
2009, ACM, pp. 18-25.

RPython
Just-in-Time Compilation with
Meta Tracing

RPython
• Subset of Python
– Type-inferenced
• Generates VMs
27
Interpreter
source
RPython
Toolchain
Meta-Tracing
JIT Compiler
Interpreter
http://rpython.readthedocs.org/
Garbage
Collector
…

Meta-Tracing of an Interpreter
28
cnt
1
+cnt:=
if
cnt:= 0
[2] Bolz et al., Tracing the Meta-level: PyPy's Tracing JIT Compiler, ICOOOLPS Workshop
2009, ACM, pp. 18-25.

Meta Tracers need to know the Loops
class WhileNode(ASTNode):
child cond_expr
child body_expr
def execute(frame):
while True:
jit_merge_point(node=self)
cond = cond_expr.execute_bool(frame)
if not cond:
break
body_expr.execute(frame)
29
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
guard(cond_expr == Const(IntLessThan))

Tracing Records one Concrete Execution
class IntLessThan(ASTNode):
child left_expr
child right_expr
def execute_bool(frame):
try:
left = left_expr.execute_int()
except UnexpectedResult r:
...
try:
right = right_expr.execute_int()
expect UnexpectedResult r:
...
return left < right
30
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
guard(left_expr == Const(IntVarRead))

class IntVarRead(ASTNode):
final idx
def execute_int(frame):
if frame.is_int(idx):
return frame.local_int[idx]
else:
new_node = respecialize()
raise UnexpectedResult(new_node.execute())
31
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
i1 := left_expr.idx # Const(1)

final idx
else:
32
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
guard(i2 == Const(F_INT))

final idx
else:
33
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
a2 := frame.local_int
i4 := a2[i3]

child left_expr
child right_expr
try:
...
try:
...
return left < right
34
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
guard_no_exception(Const(UnexpectedResult)

child left_expr
child right_expr
try:
...
try:
...
return left < right
35
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
guard(right_expr == Const(IntLiteral))

child left_expr
child right_expr
try:
...
try:
...
return left < right
36
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
i5 := right_expr.value # Const(100)

child left_expr
child right_expr
try:
...
try:
...
return left < right
37
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]

child left_expr
child right_expr
try:
...
try:
...
return left < right
38
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
b1 := i4 < i5

child cond_expr
child body_expr
def execute(frame):
while True:
if not cond:
break
39
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
b1 := i4 < i5
guard_true(b1)

child cond_expr
child body_expr
def execute(frame):
while True:
if not cond:
break
40
while (cnt < 100) {
cnt := cnt + 1;
}
Trace
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
b1 := i4 < i5
guard_true(b1)
...

Traces are Ideal for Optimization
guard(cond_expr ==
Const(IntLessThan))
guard(left_expr ==
Const(IntVarRead))
a1 := frame.layout
i2 := a1[i1]
i4 := a2[i3]
guard_no_exception(
Const(UnexpectedResult))
guard(right_expr ==
Const(IntLiteral))
guard_no_exception(
Const(UnexpectedResult))
b1 := i4 < i5
guard_true(b1)
...
a1 := frame.layout
i1 := a1[Const(1)]
i4 := a2[i3]
b1 := i2 < i5
guard_true(b1)
...
a1 := frame.layout
i1 := a1[1]
guard(i1 == F_INT)
i2 := a2[1]
b1 := i2 < 100
guard_true(b1)
...

Truffle + Graal
Just-in-Time Compilation with
Partial Evaluation
Oracle Labs

Truffle+Graal
• Java framework
– AST interpreters
• Based on HotSpot
JVM
43
Interpreter
Graal Compiler +
Truffle Partial Evaluator
http://www.ssw.uni-linz.ac.at/Research/Projects/JVM/Truffle.html
http://www.oracle.com/technetwork/oracle-labs/program-languages/overview/index-2301583.html
Garbage
Collector
…
+ Truffle
Framework
HotSpot JVM

Partial Evaluation Guided By AST
44
cnt
1
+cnt:=
if
cnt:= 0
[3] Würthinger et al., One VM to Rule Them All, Onward! 2013, ACM, pp. 187-204.

Partial Evaluation inlines
based on Runtime Constants
child cond_expr
child body_expr
def execute(frame):
while True:
if not cond:
break
45
while (cnt < 100) {
cnt := cnt + 1;
}

child cond_expr
child body_expr
def execute(frame):
while True:
if not cond:
break
46
while (cnt < 100) {
cnt := cnt + 1;
}
child left_expr
child right_expr
try:
...
try:
...
return left < right

child cond_expr
child body_expr
def execute(frame):
while True:
try:
left = cond_expr.left_expr.execute_int()
...
try:
right = cond_expr.right_expr.execute_int()
...
cond = left < right
if not cond:
break
47
while (cnt < 100) {
cnt := cnt + 1;
}

child cond_expr
child body_expr
def execute(frame):
while True:
try:
left = cond_expr.left_expr.execute_int()
...
try:
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}
final idx
else:
raise UnexpectedResult(new_node.ex

child cond_expr
child body_expr
def execute(frame):
while True:
try:
if frame.is_int(1):
left = frame.local_int[1]
else:
...
try:
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}

Optimize Optimistically
child cond_expr
child body_expr
def execute(frame):
while True:
try:
if frame.is_int(1):
else:
...
try:
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}

Optimize Optimistically
child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
__deopt_return_to_interp()
try:
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}

child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
try:
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}
class IntLiteral(ASTNode):
final value
return value

child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
try:
right = 100
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}
final value
return value

Classic Optimizations:
Dead Code Elimination
child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
try:
right = 100
...
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}
final value
return value

Constant Propagation
child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
right = 100
cond = left < right
if not cond:
break
while (cnt < 100) {
cnt := cnt + 1;
}
final value
return value

Loop Invariant Code Motion
child cond_expr
child body_expr
def execute(frame):
while True:
if frame.is_int(1):
else:
if not (left < 100):
break
while (cnt < 100) {
cnt := cnt + 1;
}

child cond_expr
child body_expr
def execute(frame):
if not frame.is_int(1):
while True:
if not (frame.local_int[1] < 100):
break
while (cnt < 100) {
cnt := cnt + 1;
}
Loop Invariant Code Motion

Compilation Unit based on User Program
Meta-Tracing Partial Evaluation
Guided by AST
58
cnt
1
+
cnt:
=
if
cnt:
=
0
cnt
1
+
cnt:
=if cnt:
=
0
[3] Würthinger et al., One VM to Rule Them
All, Onward! 2013, ACM, pp. 187-204.
[2] Bolz et al., Tracing the Meta-level: PyPy's
Tracing JIT Compiler, ICOOOLPS Workshop
2009, ACM, pp. 18-25.

WHAT’S POSSIBLE FOR A SIMPLE
INTERPRETER?
Results
59

Designed for Teaching:
• Simple
• Conceptual Clarity
• An Interpreter family
– in C, C++, Java, JavaScript,
RPython, Smalltalk
Used in the past by:
http://som-st.github.io
60

Self-Optimizing SOMs
61
SOMME
RTruffleSOM
Meta-Tracing
RPython
SOMPE
TruffleSOM
Partial Evaluation +
Graal Compiler
on the HotSpot JVM
JIT Compiled JIT Compiled
github.com/SOM-st/TruffleSOMgithub.com/SOM-st/RTruffleSOM

Java 8 -server vs. SOM+JIT
JIT-compiled Peak Performance
62
3.5x slower
(min. 1.6x, max. 6.3x)
RPython
2.8x slower
(min. 3%, max. 5x)
Truffle+Graal
Compiled
SOMMT
Compiled
SOMPE
●●●
●●●
●●●●●●●●●●
●
●●●●●●
●●●●
●●
●●
●
●●●●●●
●●●●●●●●●●●
●●●
●●●●●●●
●
●●
●
●●●
●●●●
●
●●●●●●●●
●
●●●●
●●●
●●●●●●●●●●●
●
●
●
●●●
●●●●●●
●
●●●●●●●
●
●●●●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●●●●●●●●
●
●
●
●●●●●●●●●●
●
●
●
●●
1
4
8
Bounce
BubbleSort
DeltaBlue
Fannkuch
Mandelbrot
NBody
Permute
Queens
QuickSort
Richards
Sieve
Storage
Towers
Bounce
BubbleSort
DeltaBlue
Fannkuch
Mandelbrot
NBody
Permute
Queens
QuickSort
Richards
Sieve
Storage
Towers
Runtimenormalizedto
Java(compiledorinterpreted)

Implementation: Smaller Than Lua
63
Meta-Tracing
SOMMT (RTruffleSOM)
Partial Evaluation
SOMPE (TruffleSOM)
KLOC: 1000 Lines of Code, without blank lines and comments
4.2
9.8
16
260
525
562
1 10 100 1000
V8 JavaScript
HotSpot
Java Virtual Machine
Dart VM
Lua 5.3 interp.

Simple and Fast Interpreters are Possible!
• Self-optimizing AST interpreters
• RPython or Truffle for JIT Compilation
65
[1] Würthinger et al., Self-Optimizing AST Interpreters, Proc. of the 8th Dynamic Languages Symposium, 2012, pp.
73-82.
[3] Würthinger et al., One VM to Rule Them All, Onward! 2013, ACM, pp. 187-204.
[4] Marr et al., Are We There Yet? Simple Language Implementation Techniques for the 21st Century. IEEE Software
31(5):60—67, 2014
[2] Bolz et al., Tracing the Meta-level: PyPy's Tracing JIT Compiler, ICOOOLPS Workshop 2009, ACM, pp. 18-25.
Literature on the ideas:

RPython
• #pypy on irc.freenode.net
• rpython.readthedocs.org
• Kermit Example interpreter
https://bitbucket.org/pypy/example-interpreter
• A Tutorial
http://morepypy.blogspot.be/2011/04/tutorial-
writing-interpreter-with-pypy.html
• Language implementations
https://www.evernote.com/shard/s130/sh/4d42
a591-c540-4516-9911-
c5684334bd45/d391564875442656a514f7ece5
602210
Truffle
• http://mail.openjdk.java.net/
mailman/listinfo/graal-dev
• SimpleLanguage interpreter
https://github.com/OracleLabs/GraalVM/tree/mast
er/graal/com.oracle.truffle.sl/src/com/oracle/truffle
/sl
• A Tutorial
http://cesquivias.github.io/blog/2014/10/13/writin
g-a-language-in-truffle-part-1-a-simple-slow-
interpreter/
• Project
– http://www.ssw.uni-
linz.ac.at/Research/Projects/JVM/Truffle.html
– http://www.oracle.com/technetwork/oracle-
labs/program-languages/overview/index-
2301583.html 66
Big Thank You!
to both communities,
for help, answering questions, debugging support, etc…!!!

Languages: Small, Elegant, and Fast!
67
cn
t
1
+
cnt:
=
if
cnt:
=
0
cnt
1
+cnt:=
if
cnt:= 0
Compiled
SOMMT
Compiled
SOMPE
●●●
●●●
●●●●●●●●●●
●
●●●●●●
●●●●
●●
●●
●
●●●●●●
●●●●●●●●●●●
●●●
●●●●●●●
●
●●
●
●●●
●●●●
●
●●●●●●●●
●
●●●●
●●●
●●●●●●●●●●●
●
●
●
●●●
●●●●●●
●
●●●●●●●
●
●●●●●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●●●●●●●●
●
●
●
●●●●●●●●●●
●
●
●
●●
1
4
8 Bounce
BubbleSort
DeltaBlue
Fannkuch
Mandelbrot
NBody
Permute
Queens
QuickSort
Richards
Sieve
Storage
Towers
Bounce
BubbleSort
DeltaBlue
Fannkuch
Mandelbrot
NBody
Permute
Queens
QuickSort
Richards
Sieve
Storage
Towers
Runtimenormalizedto
Java(compiledorinterpreted)
3.5x slower
(min. 1.6x, max. 6.3x)
4.2 KLOC
RPython
2.8x slower
(min. 3%, max. 5x)
9.8 KLOC
Truffle+Graal
@smarr | http://stefan-marr.de

Building High-Performance Language Implementations With Low Effort

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to Building High-Performance Language Implementations With Low Effort (20)

More from Stefan Marr (19)

Recently uploaded (20)

Building High-Performance Language Implementations With Low Effort

Editor's Notes