dynamic interpretation for dynamic scripting languagesjasonmc.net/old/papers/cgo10_pres.pdf ·...

Dynamic Interpretation for Dynamic Scripting Languages

Dynamic Interpretation for DynamicScripting Languages

Kevin Williams, Jason McCandless, David Gregg

Trinity College Dublin

April 28, 2010

1 /42


Outline

Introduction

� Novel intermediate representation for scriptinglanguages that explicitly encodes types ofvariables.

2 /42


Outline

1 Motivation

2 Background

3 Motivating Example

4 Dynamic Intermediate Representation

5 Experimental Evaluation

6 Scope for Optimization

7 General Applicability

8 Conclusions3 /42


Motivation

Motivation

� Scripting languages growing really fast!� Run slower than statically typed languages (type

checks)� Lua: a popular portable scripting language amongst

many developers

4 /42


Motivation

Motivation

� Standard interpretation loop often used to executearray of opcodes

� Run-time type checks a significant portion ofprogram execution

� Research shows advantages of type specializationin JIT, little research on specialization of interpreters

5 /42


Motivation

Our Idea

� A Dynamic Intermediate Representation thatexplicitly encodes types of variables at eachexecution point

� Dynamic IR is a flow graph, each edge directsprogram flow based on control and type changes inthe program

� Therefore specialized path in the graph for everysequence of control and type changes

6 /42


Motivation

This Paper

� Present the initial development of our prototypeimplementation in the Lua VM

� Design of dynamic representation as well astechniques used for its efficient interpretation

7 /42


Background

Background

� High level languages have type systems. Theyenable detection and prevention of invalidoperations...

� Dynamically typed languages – each time anoperation is applied to a variable during programexecution, the type of the variable must be checked

� Goal of much research to reduce or eliminatedynamic type checks in compilers or virtualmachines

8 /42


Background

Background

� Lua VM has nine variable types: {nil, boolean,lightuserdata, number, string, table, function,thread, userdata}

� Register based machine� Thirty-eight instructions� Interpreter accesses array of instructions through

program counter

9 /42


Motivating Example

Static IR

Static IR

1 local N = 1002 local flags = {}3 for i = 2, N do4 if not flags[i] then5 for k = i+i, N, i do6 flags[k] = true7 end8 end9 end

10 /42


Motivating Example

Static IR

Static IR

1 loadk r0 k0 ; reg0 = constant02 newtable r1 0 0 ; reg1 = new table (0 ,0)3 loadk r2 k1 ; reg2 = constant14 move r3 r0 ; reg3 = reg05 loadk r4 k2 ; reg4 = constant26 forprep r2 L15 ; perform forloop prep , goto[15]7 gettable r6 r1 r5 ; reg6 = reg1[reg5]8 test r6 L15 ; i f reg6 , goto[9] else goto[15]9 add r6 r5 r5 ; reg6 = reg5 + reg5

10 move r7 r0 ; reg7 = reg011 move r8 r5 ; reg8 = reg512 forprep r6 r1 ; perform forloop prep , goto[14]13 settable r1 r9 k3 ; reg9[reg1] = constant314 forloop r6 L13 ; i f loop , goto[13] else goto[15]15 forloop r2 L7 ; i f loop , goto[7] else [end]

11 /42


Motivating Example

Static IR

‘Imaginary’ Static IR graph(a) 1 loadk

(b) 2 newtable

(c) 3 loadk

(d) 4 move

(e) 5 loadk

(f) 6 forprep

(o) 15 forloop

(g) 7 gettable

(h) 8 test

(i) 9 add

(j) 10 move

(k) 11 move

(l) 12 forprep

(n) 14 forloop

(m) 13 settable

An ‘imaginary’ control flow graph for the staticinterpretation of the Sieve of Eratosthenes algorithm.

12 /42


Motivating Example

Dynamic IR

Dynamic IR graph(a) 1 loadk_number

(b) 2 newtable

(c) 3 loadk_number

(d) 4 move_number

(e) 5 loadk_number

(f) 6 forprep_number_number_number

(g) 7 gettable_table_number

nil

(h) 8 test_nil_1

(i) 9 add_number_number

(j) 10 move_number

(k) 11 move_number

(l) 15 forloop_lt_number_number_number

loop branch

(m) 12 forprep_number_number_number (n) 12 forprep_number_number_number

(o) 13 settable_table_number

(p) 14 forloop_lt_number_number_number

loop branch

(q) 14 forloop_lt_number_number_number_number

loop branch

(r) 7 gettable_table_number

nil

boolean

(s) 7 gettable_table_number

nil

boolean

(t) 8 test_nil_1

cond branch

(u) 8 test_boolean_1

cond branch

(v) 9 add_number_number

(w) 10 move_number_number

(x) 11 move_number_number

(y) 15 forloop_lt_number_number_number_number

loop branch

(z) 15 forloop_lt_number_number_number_number

loop branch

Outer loop later iterations

Outer loop first iteration

Initialization

13 /42


Dynamic Intermediate Representation

Dynamic IR

� Dynamic IR a flow graph� Nodes represent specialized instructions� Edges represent either control-flow or type-flow

� Every type change results in a new path -> allvariable types are known at every point inexecution

� Construction of the dynamic flow graph guided bythe Static IR

14 /42



Standard Node

Standard Node

int opcode;char* type;Node* next;

� Always result in same control and type flow

15 /42



Conditional Node

Conditional Node

int opcode;char* type;Node* fallthrou;Node* target;

� Instructions that result in two paths of control flowbut no type changes

16 /42



Type-Directed Node

Type-Directed Node

int opcode;char* type;Node* target[9];

� Dispatch the next node based on the type of theoperation’s result variable

17 /42



Call Node

Call Node

int opcode;char* type;Table<Instr*,Node*> call_table;Node* mru_call_node;Instr* mru_call_key;Table<Node*,Node*>return_table;Node* mru_return_node;Node* mru_return_key;

� Represent a call site and known set of parametertypes

18 /42



Return Node

Return Node

� Represent a return instruction and a set of returntypes

19 /42



Instruction Specialization

Instruction Specialization

1 Register Loads2 Arithmetic Operations3 Table Access4 Conditional Branches

20 /42


Experimental Evaluation

Benchmarks

� No real standard for scripting language benchmarks

21 /42



Micros

� We use a set of kernel benchmarks taken fromCLBG and GWCLS

22 /42



Experimental Setup

Experimental Setup

� Intel Xeon dual core 2.13GHz

23 /42



Performance

PerformanceStaticDynamic

0

2

4

6

8

10

12

14bi

nary

−tr

ees

fann

kuch

fast

a

k−nu

cleo

tide

man

delb

rot

n−bo

dy

nsie

ve

nsie

ve−

bits

part

ial−

sum

s

recu

rsiv

e

spec

tral

−no

rm

reve

rse−

com

p

thre

ad−

ring

acke

rman

n

arra

y−A

cces

s

hash

−A

cces

s

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

s

rand

omnu

mge

n

strin

g−co

ncat

mea

n

Exe

cutio

n tim

e in

sec

onds

24 /42



Performance

PerformanceStaticDynamic

0x

0.5x

1x

1.5x

2x

2.5xbi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gA

cker

man

nA

rray

−A

cces

sH

ash−

Acc

ess

Fib

onac

ciH

eaps

ort

Mat

rix−

Mul

Nes

ted−

Loop

sR

ando

m−

Num

−G

enS

trin

g−C

onca

tm

ean

Per

form

ance

Spe

edup

25 /42



Machine Instructions Issued

x86 Machine Instructions IssuedStaticDynamic

0%

20%

40%

60%

80%

100%

120%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Inst

ruct

ions

issu

ed

26 /42



Branch Instructions

Conditional Branch InstructionsStaticDynamic

0%

20%

40%

60%

80%

100%

120%

140%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Con

ditio

nal b

ranc

h in

stru

ctio

ns

27 /42



Cache Access

L1 Data CacheStaticDynamic

0%

20%

40%

60%

80%

100%

120%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Leve

l 1 d

ata

cach

e ac

cess

es

28 /42



Cache Access

L1 Instruction CacheStaticDynamic

0%

20%

40%

60%

80%

100%

120%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Leve

l 1 in

stru

ctio

n ca

che

acce

sses

29 /42



Processor Stalls

Cycles Stalled On Any ResourceStaticDynamic

0%

20%

40%

60%

80%

100%

120%

140%

160%

180%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Cyc

les

stal

led

on a

ny r

esou

rce

30 /42



Memory

Memory UseStaticDynamic

0%

50%

100%

150%

200%

250%

300%

350%bi

nary

−tr

ees

fann

kuch

fast

ak−

nucl

eotid

em

ande

lbro

tn−

body

nsie

vens

ieve

−bi

tspa

rtia

l−su

ms

recu

rsiv

esp

ectr

al−

norm

reve

rse−

com

pth

read

−rin

gac

kerm

ann

arra

y−A

cces

sha

sh−

Acc

ess

fibon

acci

heap

sort

mat

rix−

mul

nest

ed−

loop

sra

ndom

num

gen

strin

g−co

ncat

mea

n

Mem

ory

Use

31 /42


Scope for Optimization

Interpreter Dispatch

� Switch based currently used in Lua. Othertechniques equally applicable to our representation

� As specialization reduces the bottleneck of typechecks, instruction dispatch becomes moreimportant to overall performance

32 /42



Register Caching

� Popular among stack based virtual machines

33 /42



Super Nodes

� Concatenate commons pairs of instructions to forma single instruction

34 /42



Loop Optimizations

� Leverage the type knowledge of variables insideloops to eliminate array bounds checking and movememory allocation outside loop body

� Specialize loop counters to integers when loopbound is constant

35 /42



Dead Node Removal

� Specialization can lead to redundant nodes.� Can be removed and all entry nodes directed to

target exit node

36 /42



Dynamic IR graph(a) 1 loadk_number

(b) 2 newtable

(c) 3 loadk_number

(d) 4 move_number

(e) 5 loadk_number

(f) 6 forprep_number_number_number

(g) 7 gettable_table_number

nil

(h) 8 test_nil_1

(i) 9 add_number_number

(j) 10 move_number

(k) 11 move_number

(l) 15 forloop_lt_number_number_number

loop branch

(m) 12 forprep_number_number_number (n) 12 forprep_number_number_number

(o) 13 settable_table_number

(p) 14 forloop_lt_number_number_number

loop branch

(q) 14 forloop_lt_number_number_number_number

loop branch

(r) 7 gettable_table_number

nil

boolean

(s) 7 gettable_table_number

nil

boolean

(t) 8 test_nil_1

cond branch

(u) 8 test_boolean_1

cond branch

(v) 9 add_number_number

(w) 10 move_number_number

(x) 11 move_number_number

(y) 15 forloop_lt_number_number_number_number

loop branch

(z) 15 forloop_lt_number_number_number_number

loop branch

Outer loop later iterations

Outer loop first iteration

Initialization

37 /42



Library Nodes

� Stack incrementing and decrementing required forexecuting standard library operations

38 /42



JIT Compilation

� Mixed-mode� Dynamic IR will improve current model of static

interpretation followed by speculated specialization

39 /42


General Applicability

General Applicability

� DIR implemented in Lua� Applicable to other scripting languages

40 /42


Conclusions

Conclusions

� Presented a novel approach to scripting languagespecialization

� First stage of a project to bring specialization to alllevels of interpretation

� Not addressed some potential scaling issues withrepresentation

41 /42


Conclusions

Thanks

Questions?

42 /42

dynamic interpretation for dynamic scripting languagesjasonmc.net/old/papers/cgo10_pres.pdf ·...

Documents