~T::t':1 (urY 2

THE DE:SIG N AND IMPLEMENTAnON

OF PROGRAMMING LANGUAGES

John Hughes

Technical Monograph PRG-40

July 1983
(published as a monograph September 19&4)

Oxford University Computing Laboratory

Programming Research Group
g-ll Keble Road
Oxford OXI 3QD
 6l

~
i
0
~ l?":"4'l? 0
0> Q' ~ .... 'A""
... = ... ~
..Q.i"lSQ. ~

"ClET ...
l}SC: 0
~e· ~"" S· e. go
S ~ - ~IQ ~
'00­
~S ~.
.0 ~ '" il. f.
~
...Eo _.­ c ~~
... "8.
~
:z:
co 0
~IT'"
c:~
S-~ "
O'"
..
IT
II
o~ "=-.
e. G­
.. CO
II Q. ..
"'" "t"'
!:oI:
... ~

o II .g.
pi
"'0 ;;
S .... ~

... ...
0
~

1>
AbAtr.CI

The Design end Imptementanon 01 Progremmlng lAnguegea

R. J, M. Hughe5

Implementation ~trategje~ for purely funcl/onal language. .r. r• .,/e.e. end

a new one u~;ng ~super-combjnator~~ proposed. An ettlc/ent e/gorlthm for

compIlatIon to ~uper-combjna,or~ I~ described. and re.lIsetlon~ 01 ,he

elgor1thm are pfluented in imperetive. 'uncrlonal and logic progremmlng

lengueges. The new method I~ compared wllh Turner'~ comb/n.fors b~ .n

e.perlmental compefl~on end by a rheoretlcal ana/y~is.

The ob~erved Inability of lunctlonal program~ to m.lte ettlc/en' u~e of flore

I~ Invest/golfed. and It Is ~hown that this I~ due to the sequentl.' nature of

fhe underlying eb~rrec' m.chlne~. l.nguage e.ten~'on. fO Incorpora'e

per.Uen~m Bre Introduced and their adequacy l~ demon~tre'.d In ~e~erel

••• mples.

Garbage collecrlon J~ d'scus~ed end referenCe counting •• /ected a~ rhe

mo~' promising ~'rareg,. An e.'en~lon to reterence counting '0 enable II to

collect cJrculer strucrure~ I~ de~crlb.d.

ActIlOWtedtemenls

18m deeply Indebled 10 Bernard Sufrln. my supervlaor. Without his conslanl

encouregement this thesis would never have been wrillen. I am also very

grltelul to Peler Henderson and Tony Hoare: aach 01 them has Influenced
my work proroundly. Chris Dollln IIrst Inleresled me In luncllonal language~.

end Devld WI,e and Jack Dennis between Ihem Inleresled me In reference
counting garbege collection. Geralnl Jone, listened patiently 10 many 01 my

IdlU belore they were even half-baked. and I have had many Inleresllng
Ind enloyeble discussIons with Simon Jones and David Turner. Meny thanks
to them ell. Finally. I am very grateful 10 my wile. Mary. end my parents
lor Ihelr conslent moral support. and the Science end Engineering Research
Council 01 Gre.t Brllaln for their IInenclal support.
 THE DESIGN AND IMPLEMENTATION OF PROGRAMMING LANGUAOES

1. InlrOduCllOn

2 FunCllonal Languages 3
2 , InlroduCllon 3
22 Functlonal Fealuro5 e
2.3 Formal Spectllca!ion
•
3 ImplemenlalJonS '0
3.1 Inlroduction '0
3.2 The ~-calculus 10
3.3
3e
Represenlalion

Tne SECD machine I.

Ie

3.
3 .•
CAF reduction

The Sk.I machine I.

18

e Super-comblnators 23
el IntrOducllon

e.2 The basic melhod

"
e.3 Parameter order
"
27

e.e Optimising conditionals

e.5 Oraphical comblnators

"
e.8 Eltperlmental re5ult5
"
"
5 Super-comblmuor compilers .0
5.1 InlroducUon '0
5.'
5.3
rh. algorithm

An ImperaUwe compiler
A funcllonal complier
..
.0

5.'
5.5 A logic compiler
"
55
5.6 Conclusion 59

6 Analysis 01 efficiency 61
6.1 Introduction 61

6.' Turner's combinalors 62

6.3 Super-comblnalors 63

6.' AccounUng Irees 66

6.5 Union-lind algorithms 70
6.6 Conclusion 72

,. Ev.luelion order 75

,..
1. , InlroduCllon 75
Lazy evaluallon 17
1.3 The need lor parallelism 6.
1.' PAR tor more parallelism 66
1.5 S't'NCH lor less parallelism 66
16 Qulcksorl 91
11 Pipes 9.
1.6 Dls.avenleges 96
1.• Conclusion 96
8. Garbage COll8cUon 99
81 Introduction 99
8.2 Garbage collection sirategtes 100
8.3 CIrcularity 101
8.' 'he sHIUe program graph 103

8.' The graph In moUon 107

8.6 ExamInation 01 COSIS 113
8.7 Brownbridge's method 11.
88 ConclusIon 11.

9 AelalecJ ..ork 11.

'.1 lnlroductlon 11.
9.2 Siring reduction 11.
93 Graph reduclion at the source 119

••• SEeD machInes 121

9.' Turner's combinalor$ 12.

••• Super-comblnalOr approaches 12.

8.7 Oala/lo.. 127

10. ConclusIon 129

References
 CHAPTER 1

INTRODUCTION

The Ulle 01 this the!lIs. "The Design and Implemenlatlon 0' Programming
languages". was chosen before any of Ihe work rep0rled was envisaged
to saUsty University regulations. Nevertheless. /I Is an BpI. If Cheeky.
descrIption of rne conlents. II 15 Cheeky beceuse we arB actually concerned
only wllh functional programming languages. II 15 apt because our first advice
10 a language designer would be to make his language functional. We b&gln
In chapler 2 by JusUfylng lhls advice.

Tnereatter we ere concerned maInly with Implemenlauon. We have been

Implemen'/ng functional languages on and off since 1979 land Lisp for even

longer) and we bellevL! that Ihls experience has given us e gOOd gen&ral
underslandlng 0' the principles Involved. In Chapter 3 we expleln our
viewpOint and Introduce some terminOlogy lor use laler.

We have long been en admirer 01 Turner's combinalor Implemenlatlon

melhod. and much of this lhests 15 concerned with our own Improvement
on II using ·super-comblnalors". Our melhod Is explained Bnd analysed In
tne lollowlng three chapters. First. In chapter 4. we explain the principles
of our method and a number of enhancements 10 It. Here we are only
concerned with determining whel code a program should be complied InlO.
 2

In lhe next chapter we demonslrale Ihat an efficient compll8r can be wrUlen

to Ganerate this code by exhibiting Ihree dltterenl one8. w"nen In the

Imper..'I"'" tunellanel and logic programming styles. This alao provides an

opportunity to compare the three dln8,anl kinds of language In aclIon. finally.

In Chllpte, e we eveluate our nllw melhod by a theoretical comparison of

11, BlIlclency with Turner', method.

We have observed e.perlmenlally Ihat functional programs sometimes require

unreasonably large amounts of storage. and thai even qulle simple programs

can gradually clog the memory with useless garbage. We argue In chapter

7 Ih81 there 8,e deep-seated reasons 'or Ihls behaviOur. and Ih81 It can

only be evolded by .. fundamental change 10 parallel abstraci machlruts.

Wa propose an extension to functional lenguages 10 control parallelism and

justlly our choice by showing how II can be used In seyeral examples.

fln8l1y. In chapter B we lurn our ellenUon to garbage coJlecUon. We belleye

Ihat there are compelling reasons tor usIng relerence counllng garbage

collection. bul hlther10 1hls has been awkward because r.r.rence counting
garbage collectors had dllflcully wllh circular structures. We propose an

eliliension that enables reference counUng 10 be used with any kind 01

alructure.

This completes an oyerylew 01 our thesl5. We wish 10 nole one other polnl.

The reader will find Ihat the words "theorem· and ·prool" occur yery rarely.

II .. all. We make no apologies lor Ihls. On the contrary. we l!Ire following

normal methemetlcal praclice In prelerrlng a Conylnclng argumenl to a lormal

demonSlralion. We hope thai lhe reader will gain " better underslandlng

'rom our Informal elilplanalions and "prOOfs by elemple" Ihan he would haye

done Irom lhe pages ot symbols they replace.

 "
CHAPTER 2

FUNCTIONAL LANGUAOES

2.1. INTRODUC TION

II Is olten clalmeCl thai funcllonal programming will revolutionise Ihe SollwSre

Industry by making programs an order 01 magnitude Basler to write. end

yel functional programming 15 usually dellned as ·programmlng '1ldthoul

85slgnmen'", It 15 very dlUlcult to see why omitting Ihe asslgnmenl statements

trom one's programs will bring such benefliS. 50 we leel II Is worthwhile

10 examine this pOint In mont detatl. We shall try 10 answer the quesHon

·why Is function al programming so good".

II 15 helpful to recall the history of anothe' software reVOlution: struclured

programming. Very sImilar claims W8rB made lor slruclUred programming

in lis early days, and yel II was often dell ned as 'programmlng withoullhe

gOlO·, II is dUflcult 10 see why avoiding gOlo slalemenls shOuld make

programming easIer. In fact. 01 course, Ihls negaUve delinlUon of struclvred

programming betrays a gross mtsunderslandlng 01 the whole Idea. II Is nol

the omission 01 gOlos that makes structured programming easler, It Is the

InclusIon 01 new slruclured programming construels such lIS Ihe while loop

and 1l-1hen-el:se In new languages like Pascal, One can omll gotos In

FORTRAN IV unlll one Is blue In the face withoul makIng programming any

easier.
 •

In Ih. same way tile negaUve dellnllion falls IO'IIIly to cepture the splrll

of functional programming. Its SUCC8U Is nOl due 10 the omission 01

aaslgnment. but to the Inclusion of new. powerlul le.lures In new functional

proor.mmlng languages. It would even be possible 10 design • functional

proorammlng language Including lIulgnmen1. In lhe slime _ey 8S Pasce'
Includes the goto; but. like the Pescal gOlD. nslgnmenl would be nellher

useful nor necessary. In section 2 01 Ihls chapter we shall ende ....our 10

Identify the Important features or funcllonai programming languages. Some
01 these tealUres are also found In non-functional t.n~uBoes luch 8S

Gedanken IAeynolds70J. Scheme 1518818781 and Ml IGordon79]. We do not

118181 thai Ihls weakens our ClUe at all. In ract 11 slrengthens II. becausa
.ach or Ih.s. languag.s Is nlc. pr.cls.ly b.cause It Includes sam. Of the

'••Iur.s we ar. applaUding. None of Ih.m Includ.s all the Important f.alur.s
we ,"II d.scrlbe, howe.... r.

Anal".r slmllerlty between structured and lunctlonel programming Ie thai

progrlm. h..... nlc.r formal prop.rtl.s Ihan Ih.lr unstructur.d or
non-luncllon.1 count.rparts (single .nlry .nd .lllt of blocks In structur.d
prooraml. lind r.'ar.nUal transparency In functlonel programs). This Is

Important. bur w. do not f.al that II Is a malar lactor In making programs

.asl,r 10 wrlta. Ralhar. It opens up the possibility of formal program
de....lopm.nt. In secllon 3 w. dlscU55 Ihe clos. relallonshlp b.tween

funciionai programming and lormal speclflcallon lU It Is practls.d at O)(lord .

• nd lI.scrlba our ... I.w of how programs could b. d.....lop.d In the fUlur•.

t.t. FUNCTlONAl. FEATURES

We h..... said Ihat wa b.lla ...e functional languag.s m.lI. progremmlng •• sle,
DeCau.. th.y Includa naw, powerful featur.,. But what ,.aclly are lh.se

".Iur.,? In Ihls s.cllon w. Id.nllfy thO•• we f•• 1 .r. the rnosl Important.
 5

Onlt reason why the functional programmer does less work 15 thai the
functional syslem does more work. The system essumes responslblllly In 8
number 01 areas. Ireelng the programmer Irom hevlng 10 Ihlnk aboul them

lit all. One Importsnt area Is srorege allocetlon. The functional programmer

Is not reQuired to decide the Illetlmes of the oblects he creates; Instead

lhe 'yscem deletes them and reuses the space !hey occupied when tl1ere

ere no references to them lel1. This avoids the risk 01 bugs caused by

delellng objects too early. or excessive storage use caused by delellng them
100 late (e "dumb" garbage collector or bad vlrluel machine design can

stili cause oblc-cl, 10 be deleted 100 lale). The mosi Imporlsnt advantage.
however. Is thai the programmer does not need 10 think aboul It.

AnOlher area In which (he system assumes responslbllltv Is Ihal 01 deC/cling

evaluallon order. In imperative languages. Ihe evaluation order Is ellpllcll

In the program. but In funcllonal languages the system Is free 10 use any

sullable order; Ihis gives rise 10 Ihe possJbllily 01 several different slrategles

tor delermlnlng order. InCluding lazy evaluation. This Is even more Imporlent

than automatic storage allocetlon. beceuse It means that Ihe structure o(

a program need no longer be determined by the desIred evaluation or~er.

Oreal s/mplllications can result: lor example. problems lhat are usually solved

by backtracking 10 lind lhe IIrsl solution can Instead be solved by wriling

functions thai return a IIsl 01 aU solullons. II only the IIrSI solullon In the

tlsl Is used. Ihen only that solullon will be compuled: on the olher hand.

II Ihat solullon proves unsatisfactory and Ihe nexl solution Is used Inslead

lhen Ihe syslem will pick up where II lell olf IIhel Is. we can gel backlracilng

behavIour wllhoul writing backtracking programs). Many olher problems ~an

be solved most naturally by programs whose struclure does not rellecl the

desIred ell'aluallon order: lor a larger exemple. see section 5....

 6

Another consequence 01 breakIng lhls connection Is thet we cen progrllm

using Inflnllll dala-structures. provided we never require the aystem to

compute all thalr components. Non-fermlnaUng progreme can ollen be

expressed wry nSIUy III functions on Infinite datil-structures. TheIr use elln

also Improve the mOd\llllrllY of lermlnallng programs: for example. a

numerical lIatallon Clln be expressed as one tunellan Ihlll compules an
Infinite list 01 IIpprOltlm"Uons and errOr bounds. lind anal her Ihal selects

II parllcular elemenl 01 thai lin This separeles the concerns of computing

the approximations. and deciding. using whatever crl1erlon seems
appropriate. which IIpproxlmallon 10 setlle for.

Inllnlte datil-structures lire often applied fa Input end output. Typically a

progrllm Is pused an Infinite IIsl of Inputs and returns .n Infinite list of

outputs. and the system chooses an execution order In whIch the outpuls

ere computed 85 the Inputs are consumed. Progremmlng In lerms of the

enTire Input or output 01 a program allows progrems 10 be comblnecr very

eetlly for exemple. funcllonal operallng systems will nol need specie' ·plpes·
In order lor one program 10 read another', output

Turning to more specillc lellures. higher-order luncUona are one 01 the

most valuable. These ara luncllons thll lake olher lunellons as ergumenls
or return them as results. Using higher-order functions we can etlecUvely

extend our language with whalever control structures we desire. We can use

this to Improve the mOdularity 01 our programs by IntrOducing specie I ~urpose

control structures as part of abstract types (cl. lIerators In CLU ILlskov79JI.

For e:remple. the well known map funcllon 15 en approprlale conlrol structure

lOr list processing. It enebles us to write certain 'unctlons on IIsls without

knowing the details of their Internal alructure.

 7

The recursion squallon style 01 funCtion definition Is 81150 .,ery valuable. In

essence this Is lusl another way of writing certain condilionais. lis advantages
are thai equallona are usually mora readable than conCIIlIonals. en(l that
It allows quite complex conditions to be expressed very simply. ThIs Is

particularly Irue of funcllons 01 many arguments. where e sequence a! tesls

can become very unperspicuous.

The expressIve power of functional languages Is consIderably Increased by

Incorporallng a ·sel-e.presslon" nolallon. as Turner does In KRC lTumeral]

1"1" nOtation seems Ideal as 8 functIonal Uerallan constrUC!. II brings the

same Cll!lrlty 10 luncf/onal programs 8S the lor-loop brought 10 Imperallve

programs.

Finally. a Yaluahle leature which IS not confined to lunctlonal langulges.

but WhiCh Ills e~slly Inlo theIr conCeptual framework. Is the provIsIon 01

abs1racl data-types, Implemenlauon is easy: alt that is necessary Is an

ebstracUon function that stamps objects with a type. Its Inverse. and a type

testing lunctlon, lhls simple extension allows the programmer to make hIs

Inlenllon far clearer. and allows the system to catch lar more erron. lWtlh

this simple kind of abstract data-lype the programmer Is able to express

the conceplual difforence between two objects with the same represenl~lIon

Dy stamping them with dillersni types. and so the system Is able to clleck

that the usage 0' these objects Is consislent with Ihe programmer's Intenl/on.

Carerul use 01 scope Is necessary to guaranlee thaI only a partiCular sel

01 operallons Is applied to Objects Of a particular Iype. so this Is not qlJlfe

an Implementation 01 "abstract data-types" as the term Is usually understood),

 •
We hive described Ihe parUcular features we teel are most Important In

mailing functlona' lengueges powerlul end easy 10 use. Equally ImpOrlanl.

however. II the fact thlt the various Illnguage elements cen be combined

In .n, wly. There Is no has 1 of speCla' cases end ,eslrlcllons In a functlonel

lengulOI. In contresl 10 Fortran. Pascal end Ada; each construct may be

descrIbed limply end used wherever Ihe programmer chooses. This slmpllcl1y

end Ir••dom Illows one 10 become conversant with the language qUiCkly.

end to u.. II conlldently.

2.3. FORMAL SPECIFICATtQN

Itt, ",ry compelling r•• son lor I.vourlng functlonsl programming Is thaI II
fils ""y wen with forme' program development We shell Illustrets Ihls tty

d.scrlblng how programs are developed formally 'It Oxford lind ellpl81nlng

Why IIllng luncUonal programming I.ngueges can simplify lhe process greauJ'.

To begin wllh. a 'ormal speCIfication of the 1ask 10 be per'ol"med Is written.

Thll conltlta 0' a number 0' deflnlUons 01 types end runctlons on them.

Thl.1 deflnilloni are constructed In lerms Of the b.slc objects of

malhamaUcs: leis. nalural numbers. retallons. etc. Oeflnltlons may be

con,tructive. e"ecUYely giVing • melhod for computing lh. cbJect defined.

or non-construcllve. For ellample. Ihe squaring 'unclion Clln be defined

con,lruCtively by

.qua:re: N ..... N

.qua:re - ~n:N. n.n

 9

and the square raol luncl!on can be defined non-conSlTucUvely by

N _ N

"". square
-,

Having written the lor mal speclllcallon. the programmer lranslorms II InlO

an entirely conslrucllve one. by provIding conslrucllve dellnillons for 811 hi!

Junellans and proving the new deflnltions equIvalent 10 the old ones. AI the

end of thls slage. the specilicallon not only dellnas wnat Is 10 be computed.

II describes 8 reasonable way of computing II.

The IInal stage 01 program development conalsls 01 laking the Iinel

specification lind 1ransl811ng If Inlo Pascal. The uanslatlon Is done Informally.

and Is nol proved Corree!. Indeed. the resulting program usually lalls to

mael 1he speClflcallon because It has tmpiemenl8Uon limits built Inlo II.

Thl' last stage Is a greal Inconvenience to lormal program developen II

cannot be proved COrrect. because In general It Isn'l correct. Moremar.

It InvDlves a 101 of work at a very IDW level, Drganlslng slorage managemen1.

execution order BtC. In lact, It Is 10lally unnecessary. The IInal speclflcatlon.

being completely cDnstructlve. Is already a functlDnal prDgram. There Is no

need 10 go any furlher. Slopping al this stage nOI only ellmlnales Ihe InlDrmal

pari Df program development - It also saves the programmer B lot or work.

'0 summarise. formal specillcalion oilers luncllonal prDgrammers a way Df

develDplng prDvably correcl p'Dgrams. Functional prDgrammlng offers formal

speClllers a way 01 ImplemenUng their speclllcallons correctly and easily.

Tooether. lhey will be powerful Indeed.

 10

CHAPTER 3

tWPLEMEHTAllONS

3.1. INTRODUCTION

In this chapter ..e discuss the main Implementallon techniques used lor

functional languages. and Investigate their advantages and disadvantages.

Beginning with the obser"'allon that all functional languages can be translated

Into (he l.-calculus pius constants. we lake the l.-calculus as the canonical
luncllonal lenguage. Aller 8 descrIption of reduction, which 15 an 8)(ecutlon

parldlgm appropriate to the 1. -calculus. we go on to develOp specific

Implement. lions. We consider machine representations. and explain why we
choose • gr.ph. We Identify the maJor Inefficiency of a graph-reduction

l-reducer. •nd show how different anempts to cur. It can lead to the SEeD
machine end Turne", combinalor,.

3.2. THE l-GALCULUS

We begin with. brie' description ot the ~-c.lculus ICurry58J. 8doptln.Q Ihe

following .bslr.ct synl'. for e.presslons:

B ,,- C I V I (B Bl I 'V.B

wllere C r.nges over • sel or const.nts .nd Y r.nges avel'" • ,el 01 V81'"18ble
 11

names. (E 1 E I) represents E 1 applied 10 E2 and ). Y.E represents the

function 01 Y thai E Is. C Includes numbers. boolean, etc .. and alao tHule

functions such as Cons. ). Y.E is said 10 bind the variable V Inside the body.

and unbound variables are said 10 be free.

The malnlng 01 an 8llpreu1on Is defined by reduction rules 01 the form

EJ. red El. which means Ihlll wherever an 8llpresslon of Ihe form El

appears. It may be replaced by E J. Each basic 'unellan has lis own ruill.
lor ell8mple

head (cons a b) red a

+ m n red m+n

provldad thai m and n life natural numbers. (Here we adopl Ihe convention

that application is leU associative and drop unnecessary brackets), The elleel

01 applying 8 ).,-Sllpreulon Is defined by the ~-rute

(Sl OV.EI E. red ElE./V]

whare the right hand aide stands for E with all free occurrences of II

replaced by Eo.

However. the p-rule Is onty applicable when Eo has no free variables. This

Is more restrIctive than Ihe p-rule I!lS usually stated. but we C1on·t care

because we are Interested In reducing complete programs. whIch a "priori

have no Iree variables. ThIs means that Ihe top-level application of •

program must have arguments with no free variables either. and so the

reslrlcteCl p-rule Is applicable. If the lop-leva I 01 a program ever becomes

a ), -expression. then wa are content nof 10 perform any reduc(lons on the

 12

body or Ihls ),-exprsssion. because we regard reductions Inside funcllons

.. program transformation. not program execullon. The advantage we oaln

by lhl' Is thai we ne-.'ef need to Invoke the expensive a-rule (which renames
yallables to avollS Inadvertent blncllngl.

An Implementation 01 the i-calculus must oper.le on en expression by

applying feducllon rules fa II unUl no more are applicable. The expression

I, then said 10 be In normal form. and this norma' form Is the result of
the compulsUon.

TI'le choice of the reducllons 10 be performed 81 any particular moment Is

• maner 0' Implemenla1l0n stralegy. II Is constrained by (he faCI that cerlaln

functions (for 811ample. •) are ,I,ICI. lhet Is can only be applied 10 arguments

In normll form. and by Ihe tact that reductions 0' the top-level of the

program shOuld alweys be per'ormed since they lead most CSlrecUy to

Ploductlon of the answer. WIthin lhese constraints we may choose:

(1) never to reduce an application unless the argument Is In

normal lorm, Thll gives • "slflct" or ·call-by-valueA lemantlca

which we argued against In chapter 2.

(2) 10 reduce all reducIble ellpresllons In perallel. This gives

"eager· evalualion, which may waste resources as some

reduclions might ntwer hive been necessary.

(3) to perlorm only reductIons a( the top-Ieve' of the program,

or Ihose directly necessary to enable a strict top-Ieve' reducllon

10 lake place, This gives "lazy evaluaUon", This approaCh can run

In(o some very sublle problems (aee chapter 1).

(4) some combination of (2) and (3), This I. the approach we

adyocaie. anCS will eluclCSale In chapter 1.

In ordef to run ...eal 'unctlonal programming languages on a lo-calculus

machine. we mus' Ifanslate them Into Iha lo-calculus. We define tfanslatlon

rules 01 Ihe form E 1 Irans E:z, whiCh when applied repeetedly will translate

• p"'og ...am Inl0 the equivalent lo-expfesslon. A complete set 0' translation
rules would be too long-winded 10 reproduce here. but we give " lew

examples:

E1 + E:z trane + E1 E:z

let I - El in E:z tra.n8 (lol.E:z) El

Jl El then E:z else EJ trans IF El E2 EJ

"here IF Is dellned by

IF true E1 E:z red E1

IF false El E:z red E2

Recursive decla ...allons are translated by the rule

letrec I - El In E:z trans let 1 - I (lol.El) in E:z

uSing Ihe basic funcllon Y which we have not seen belore. The reduction

rule lOr V Is

Y fred f (I f)

which uses Y I again on the right hand sIde. This means Ihat Y I has no

normal lorm, because 11 ...educes to (f (f (l ... m. II Is Ihls pOlentlally Intlnlle

behaviour tha1 allows Y 10 Implement recursIon. V Is very Imporiant and we

.hall relurn 10 Its Implementation In the next secUon.

 ,.

Mutual recursion Is • IUlle Irlckler. The syntax we use I.

letrec I1-El and I~-E~ and .•• in Eo

Wt need 10 Introduce two new rules 10 translale It. firstly

11-£1 and J2-E~ trans (11.13) - cone £1 £2

This If.nslates the mutually recursive dellnilions Inlo It single recursive

"structured" definition. which declares several names to be the components

of • •tructure. This single deflnlUon can be translated further by OUf earUer

rvles. producing a ),-expresslon binding a ·structured .... rlable·. We translale

thIs kind of )..-exprsulon Into slmpl. )..-e.presslons as follows:

),,(11.12)£ trane U (l.Ill.I3.E)

ullng an explicit unpacking function U defined by

u t x red t (head xl (taU x)

W. hope Ihal these Bll8mptes h....e convinced Ihe feader that translallon

,nto the ),-calculus Is not an onerous IUk.

1.3. REPRESENTATION

A crlUcal choice In the design 0' a reduction machine Is lhe choice of

the represent.llon 01 e.presslons. There are two essentially different

posallMllllea: as a s,rlng. or .s • graph.
 15

In !llrlnl1 reduction machines the program Is stored as 8 sequence 01

symbols. much 85 II appears on paper. During luncl/on appllcBllon lhe
maChine actually subslllutas one 8}(presslon Into another. and musl shuffle
the symbol' arovnd 10 make room. Thl' Is potenllally 8 costly operation.
Stili worse. If the argument of II function Is nOI In normal lorm then lhe

subslltutlon may greatly Increase the amounl of work to be done. slnc~ It

may create several copies. each 01 which must be reduced Independently.

FOr clfample. in the case

(~x.+ x x) expensive ~ed + expensive expensive

then the amount of work Is doubled. assuming thaI reducing "expenS"B"

Is considerably more cosily Ihan an addition or III function appllC8l1on. for

Ihls reason 51rlng redvc!lon 15 really only sult/lble 'or strict languages. which

guarantea thai arguments aut In normal form before substltuUon. Mago's

Atrlng reduction machine. lor example. uses FFP ....hlch Is Indeed e strict

language IMag079l. Since we believe non-striCtness to be e vllal

programmln~ tool we relact slrln~ reduction.

In contrasl. a graph reduction mllchlne represents every expreUlon as II

cell which contains poinlers 10 lis SUb-expressIons. Now when en ergument

Is subslUuled Into a funCllon bOdy II Is nOl necessary to copy Ihe IIrgument:

InSlead a pointer to the argument Clln be subsflluled. This makes luncllon

application much more ellicienl. Not only this. II also means thai there Is

only ever one copy 0' the argument The IIrSI lime II Is reduced It will be

replaced by Its normal lorm. and all polntars to lhe argument will now

aUlometlcally refer 10 lis normal form. ThIs means that no melter when l~e

ar~ument Is flrsl reduced. II will not be reduced more then once. Ora,h

reduction maChines therefore provlda good support 'or non-strict languages..

 16

They elsa permit an ellielani Implementation of Y. Recall Ihet 'he reduction

ruft for Y Is

y t red t (Y !)

On II graph reductlon machine •• can lake adventage of the 'acl Ihel III
copy of Ihe left hanes aide appears on the rlghl by cons.ructlng II circular
result

y t red i.J
n II clear Ihal Ihls ,eduellan rule Is equivalent 10 the 0(.0In8' one. but Is

much mo'e efllelent since It reduce, CY D completely In onl~ one step.

Ttl' dl'.d....Ol.08 01 graph reduction machine, Is thai. lust as there Is only

....., one copy 01 en arGument. so there Is only ever one copy of • funcllon.

This me.ns that substitution during lunctlon application must not be

destructive - lor It It 'Iller. then the funcUon WOuld be corrupted and could

nol be called ageln. Therelor•. Inaleed 01 copyIng the argument when we

apply a 'unCllon. we must copy lhe function bOdY. This Is the lunda mental
Inelllclency of graph-reduCtion: Ihe teChniques we shall discuss In the resl
01 Ihls chapter seek 10 ellmlnale or reduce II.

1.4. THE SECD MACHINE

A, we remarked above. In a graph reducllon architecture funclion IIppllcaUon

requires copying the lunctlon bOdy. and this can become wery e.penslve.

Con,lder lor e.ample

lalblc. E
 17

When this funclton Is applied. E would have 10 be copied three time!. Mea

10 subsiliule lor a. once 10 sUbstitute lor b. and once to subsillute lor c.

In general. an expression would be copied once for every variable In sc~pe

In 11. Since there are often many variable, In scope 8t • lime thle Is

prohIbitively expensive.

One 01 the anampls 10 8void Ihls Is the SEeD machIne ILandln64). The

fundamental Idea is 10 delay subsUtuUons until the very lasl minute. so Ihat

when we are finally lorced 10 perform them we can do several together.

copyIng the expreSSion only once. To do this we need Iwo new Clala types:

an "environment". which Is a mapping Irom Identlllers 10 values. and a

·closure" which consists 01 an expression and 8n environment 01 delayed

subsllltJllons to apply to It. Writing a closure as E(pJ. where p Is (he

environment. we Change the reduction rules 01 fhe machine to

(El E:a) (p) red EJ(p} E:a{p}

«11.El){pl}) (E:a{p:a) red El(pl e (ll-+E~{p~))

C (p) red C
V (pI red p(V)

11 It clear from this descrlpllon Ihat lhe SEeD machine Is stili a graph

redtJcllon machine. Like other graph reducllon approaches. It provIdes good

support for non-slrict languages. and for the same reason: the environments

contain only pointers to expreSsIons. and so each expression Is only reduced

once.

However. real SEeD maChines use Jour reglslers. Ihe steck. environment.

COntrol and dump to avoid constructing some 01 the Inlermedlale clo,ures

Uhls I, Ihe reason for lhe name). ThIs Optlmlsallon. coupled wllh the more

complex reduCllon rules. makes the behaviour 01 an SEeD machine quIte

 ,.

dllllcutt 10 r8e50n aboul. Because of this we prefer 10 work wllh simpler

.pproIlChes. Some of our results elnnOl be applied to SEeD mllchlnes 81

.11.

Ttl. sEeD machine does Ind8sd ....old doing any subslltullons. bul It replaces

lhem wllh enylronment lookups which elln proys very 8xIJenslye when there

.r. m.ny names In the environment. Turner found thet his SEeD

Implementation spent mosl of Its time lookIng up nllmes In the envIronment

CTurner791. This hIlS led to other 811empts to reduce the cosl 01 subsl/1utlon.

~.5. CAF REDUCTION

SInce substitution Is Ihe meJor Inefficiency In 8 gr.ph reduction machine.

lind subsllluUon only occurs when ~-expresSlons ere .ppll&d. II Is nalurel

10 try 10 eliminate ~-expresslons lind "'lIrI.bles altogether. To this end we

D.flne the language 01 con.!lanl appllcer/v9 lorm.! (cals) whose syntax Is

E ::. C I (E E)

This language Is particularly simple and easy to Implemant: however. lor

II to be useful we must have a way 01 ,ranslallng ~-expresslons Inlo /I.

To do lhls we have to Inlroduce a new Class 0' Conslsnt. Ihe combinator.

These have the syntax

C : : - X I 1 .•. h.B
B ::. C I I I (B B)

The reduction rule lor combln8l0rs Is

(,I(ll ..• h.B) E1 .•• l':n red B[E1/Il, . . _,E,,/I,,)

 19

A combinator Is very like a ),-expresslon. but there are two Imporlant

dlNerences; firstly, 8 combinator takes several arguments at once. and

secondly. 8 comtllnetor must have no free varlablea. Taken together. these

mean thai an expreSSIon can never be substituted Into twice. since the onl~

expressions thai can be subslUuted Inlo are combinalor bodies. end the
resutl of !lubsUlullon Is nOI a combinator body. So we have eliminated the

malor Inefficiency 01 grapl'l re(JucUon.

We may \rBnslate the ),-calculus Inlo cals as follows: If the tree variables

of ), V.E are IL.ln. Ihen

),V.E trans (11::11 ••• 1" V.E) J1 ••• I"

Repeated appllcsllons 01 this rule will eliminate all the ),-expresslons In Ihe
program. and II Is clear 11"/81 Iha resuillng comblnators will hav& no free

variables. as requIred. ThiS approach has been L1sed by Johnsson In his

ML complier IJohnsson8:lJ.

1.8. THE SKI MACHINE

II has long been known that ),-expresslons can be lranslafed Inlo cafs which

Lise only three dlfferenl comblnators.

S - Kabc. a c (b c)
K - Kab. a
1 - Ka. a
 20

by the 1,8nel8tlon rules

lV.V tran. I
lV.VI trane K VI
lV.E! E2 trans S (lV.EI) (lV.E41)

Howe....'. applying these ',anslallon rules to .lin expression of eny size yields
an enormously large result. Turner observed lTurner191 thai using 8n
optlml.atlon rule

S (K a) (K b) trane K (a b)

dr••llceUy reduces the size of the result. He used this rule. together with

• lew other comblnetors which abbre",late common lorms. to make 8 prlCllcal

cal-reduction Implement.tlon 01 SASL He lound that his Implementation

performed conslder"bly beller (han hIs SEeD-machine Implementation of lhe

..me language. apperently because the ellcessl"'9 cost of environment

~ups .e. a"'olded.

He round thai the use of his opumlS8uon rule had another Interesllng

consequence. which we may Ulustra1e by considering the elC"mple

lX. + 1 2

or course. e...ery lime this function Is applied II returns the answer 3. We

.rl Inlere.1ed In whether Ihe eddttlon II performed on every cell. or once

only. Ullng Ihe 1-reductlon rules. we find lh.1

(1"'. + 1 2) • red + 1 2 red 3

 21

The addition Is performed on every call, since the function body Is COpl,,<l

during applicatIon.

Translating to comblnators. we gel

~x. + 1 2 trans S(S(K +)(K l»(K 2)

and when this function Is applied. we rind

S(S(K +)(X l»(K 2) a red seX +)(K l)a(X 2 a)

red K + a (K 1 a) 2
red ... 1 2 red 3

So. as In the case of direci ). -reducllon, the e.pression (+ 1 2) Is

constructed and reduced on every call.

But. using Turne,'s opllmlsallon rule. we find that

). x. ... 1 2 tr ana K (+ 1 2)

and

X (+ 1 2) a red + 1 2 red 3

In Ihls cue. the r8!lult 01 applying K 1+ 1 2) Is a polntsr 10 t('s argument.

(+ 1 2), When fhls Sll'prasslon Is reduced 10 3. K's ergumenl becomes 3.

$0. aller the IIrst call the lunctlon Is slmpty (K 3). end II returns 3 wllhou,
performing any addl1lons at all.
 22

We have Iusl demonstrated thai Turner's opllmlsallon rUle guaranless thl!ll

constant a_pression, are only eVl!lluated once. II, effect 18 mor. far-reaching

tt1ln thhl. though. because It Ireals any e:olpre"lon Insloe l!l lunellan thai

-'olll!l nOI Involve the bOund variable In the ume way. It gU8ranl88S that
any !Iuch expression Is evaluated only once. and thl!lt II, value will be used
direCtly lherealter. Thus II subsumes such optlmlsallons as constant fOlding

and mOve-out from lOOps In Imperlillve languaOas.

We may summarise our conclusions 8S

every expression I, evalualed 81 mosl once aller the variables

In It heve been bound.

WII call this properly fully lazy 8v8/uation. 11 I, IInalogou5 to IlIzy evalul!ltlon,

but I, more general ,Ince the latler slales only thai e..,ery argument of a.

,.Uterlon Is e~aluated al most once.

 2J

CHAPTER 4

SUPER-COM8INATORS

4.1. INTRODUCTION

We haye described two dllle,snt approaChes 10 Improving the efficiency 01

graph-reduClion Implementations of the l.-Ctllculus. The combinator

technique or sactlon 3.5 Is simple. and has the advantage thai compllaUCJn
Into those comblnators Is easy. Turne,', method. on the other hand. requires

8 more complex complier and breaks the execution down Inlo very sm811

steps: an applies lion of S. fOr 8Jl8mple. does nOI achleY8 very much
compared to an appllcallon 01 some larger comblnators. Howeller. It brings

with It the very Imparlant advantage of fUlly lazy 8v8luallon.

In this chapter we shall develOp II methOd that combines the besl fealurtlS

01 both approaChes. Evaluation wIll be lully lazy. bul IndivIdual comblnators

will accomplish much. The basic method will be IntrOduced In secllon .,2.

Subsequent sections present various Improvements that can bEl made. and

the Iinal section presents the results of an experimental comparison of' our

auper-comblnalors with Turner's approach.

 2.

4.'. THE BASIC METHOD

OUr purpose Is 10 modify the approach of aecUon 3.5 so that It avoids

unnecessary repeated evaluation of expressions thai .re Independent of some
bound variable,. We IIrsl of all give these expressions 8 name: they afe

the 'r•• eJtpre3310ns 01 i- •• presslon,. by analogy wUh free ".rlables. In lact.

Iree vsrlables .re the minimal Iree expressions 01 " i-expression. We call
I". e.presslons thai ar. not perl of any larger free expression m:ulmlll
'rea expreSSions (mI.,), For aramp1e. the i-expression ),y.(+ I- x x) (. y
yll haa many free expressions. IncludIng .... -. and (- J( Jd since none 01
masa eJlpr.aslons Involv8S y. bUI U ha. only two maxim.' free a.presslon,.
being (+ (- x x)) and

W. can now dellne II new Iran.lallon scheme. Consider the l-eJ(presslon

lV.E. and suppose thai El ... En are 115 ma~mal free expressions. let 11

'n be Identifiers nol used In the l-expresslon. Then

lV.E tl'an.
{ldl. .. InV.E[ll/El, ...• In/En) El ... En

By the definition of combinator application. It Is clear 'hat both sides 01

thle aquatlon ara equl"alent. and therefore the translation 15 correct. Also.

tha comblnetor we ha"a produced cannot ha"e any tree YBrlables. because
.ny frea ,,8rleble 15 8 tree expression. and would theretore be enelosed

In one ot the maximal Iree expressions El En. Theretore the Iranslatlon

prOduce. oenuln. comblnetors. Ae In ncllon 3.5. re"•• ted application 01
... f'ule wfll translale a whola program 10 comblnators.
 25

We have yet 10 show how ltlls method gives fUlly tliZY 8Yelullllon, We shall

flT!l' or all consider our prevIous Bumpls. h.+ 1 2. This Is 'ransleted 8S

l.Jl:.+ 12 tt:Bna (<<Ax.e) (+ 1 2)

-X(+12)

8XIlc::t1y the same code as prOduced by Turner's approach. When this funcllon

Is applied II returns II pointer 10 the argument (+ 1 2>. and so when this

Is reduced 10 3 the lunCTlon actually becomes (K 3>. No furlher 8<1<1111on,

ere performed.

To lake a more realistic 8ll:ample. we consider the lunctlon Ihel selecls Ihe

nih elemenl of e sequence.

81 II YP.e!lnlS.
IF (- n 1) (hd 0) (01 (- n 1) (tlo»)

This function can be parllelly paramelerlsed. liS In (el 2>. to give II 'unellan

Ih81 alway, selecl, II particular slemen! or II sequence. Fully lazy eV8lu~tlon

In Ihls cllse would mil en lhet such II pllrlUllly parameterised function would

need to do no arithmetic on n 10 selecl the right element

looking 1II1 Ihe Innermost ),-ell:pression first. we see thllt II hilS mllll:lmlli Iree

elilpressions (IF (= n H) lind (el (- n H). It Is Iransillted Into

• (IF (- n 1) (81 (- n 1»

where a - kabe. II (hd a) (b (tl e»

 26

When we apply the same process 10 the other ).-expresslons too. we end
up with

el II Y .,
., IQ lCe1. IJ el
IJ '" Kel n. a (IF ( ... n 1») (el (- n 1»)
a a lea b 8. a (hd 8) (b (tl e»

Now. wllan /he partially pBramelerl:\oed function (01 2) 1& IIrst applied 10 a

sequence, It wltl be reduced as lollows:

el 2 red Y .,. 2
red.,e12
red IJ e1 2
red a (IF (- 2 1) ) (el (- 2 1)

red a (If false) (01 1)

red a (IF false) (a (IF true) (01 0»

All 01 these (eClucUons will be performed on the ItrSI call, trenslormlng (el

2) InlO a luncllon Ihal seJects the second element 01 e sequence without

doIng any arithmetIc.

Noltee Ihal we did noc oother to abstract out constant free expressions In

Ihls example. In lacl. It is unnecessary 10 do so In order 10 guararHee thai

the comblnators generatea have no Iree ..,8rlables, It Is convenlenl. both

on paper and In ImplementatIons. to allow comblnators 10 InclUde con slant

expressions. but wllh the understanding lhal the lutl laziness property stili

applies. We mOdify our understanding 01 COmblnalor appllC8110n SO that lIca.+

1 21 Is equivalent to Cub.a) (+ 1 2).
So. this basiC method gives liS a fully lazy evaluation of Our original program.
_l1houl breaking execution down InlO small sleps - Indeed, the code

produced contelns only one combinator for each source l-expres~lon.

Experiments show that It Is already more efficient than Turner's method.

In subsequent sections we wtll see how It can be Improved sUit lurther.

4.3. PARAMETER ORDER

The basic ~uper-comblnalor method does not eJellne the comblnelors

uniquely. sInce we allowed the parameters Of the combinator to appear In
any order. We should ask ourselves whether one order Is likely to be Miler
than anOth6r. 0"- whether all orders are equally good.

One faclor Which might Influence a choice 01 parameter order Is Ihe eftslre

to eliminate redundanl paramelers and comblnalors. An eJiUlmple 01 •

redundant combinator Is lCa.aa. which Is eQulvalent to a. Using a Insleed

Is more efficIent. sInce It eliminates an unnecessary combinator appllclUon.

Parameter order has a bearing on thle In a case such as ,S8ICab.abe, which

Ie not equlval&nl to a. II a Is a combinatOr IntrodUCed by lhe complier.
and the complier has the choice 01 the order In which the parameters 01

a appear. then II can choose the Olher oreJer, making ,SQlCab.aab which
Is equlvalenl to a Thus a correCI chotC& 01 parameter order can help make
more comblnators redundanl.

An example 01 a redundant parameler Is b In f1Cab.alio a llb. which Is

equivalent to 1C8.a(+ a D. The advantage or eliminating b In this csse Is

thai thiS will permit the Combinator 10 be applied when fewer arguments
are available. and hence will allow the result 01 the applicallon 10 be ~hered

more Widely. ThIs will reduc& the total number 01 combinator appllCUlons
 2.

necesssry. Once egaln. El correct choice 01 perameter order for a Is

essential. since alhenll'lse IJ would be -:ab.ab(+ 8 1l In whIch no parameters

are redundant

Anotller leclor Influencing choice 0' parameter order Is Ihe desire 10 make

maximal free elllpressions 85 large as possIble. II. by rearrangIng the

parameters 01 one combinator. we can make several mles. 01 en enclosing

~-e.pre55Ion combine InlO one larger one. then we have Improved the
elliciency 01 the enclosing combinator. This Is analogous to trying 10 move

8S much work 8S possible oul of e loop. For example. consider

~n. a (hd 8) (+ n 1) (tl 9)

whiCh hEls mfes (0 (hd 5» and (II s) This )..-eJilpresslon will be translaled

1010

(lI:abn. a (+ n 1) b) (0 (hd 9» (tIs)

However. If the parameters 01 a had been arranged as lollows:

l..n. " (hd a) (tl a) (+ n 1)

then lhe ).. -ell:pression would have been replaced by

(xan. a (+ n 1)) (a (hd 3) (tl s)'

The laller form Is more elficlen!. bolh because Ihe combinator has lewer

arguments and a simpler body. and so Is more efficient to appty. and

because the large ell:presslon (a <hd s) (II sl) need only be conslructed
once and can then be shared between all calls 01 lhe function.
We derive a general rula Irom this example. The combinator a Is derived
'rom a k-ellpruslon with mfes (hd s), (II 15) and (i- n n. We have se,"

lhat. when choosing a paramaler order lor a. those mles which are also
free e_presslons 01 the nelll enclosing k-ellpresslon should ap~ar before
those which are not. Suppose a has paramelers E 1 En, so that the call
of a will appear as

aEl •• _En

There should be some J such that. lor all I less than or equal 10 J. EI Is
a Iree allpresslon 01 the ned enclosIng k-axpresslon. and fOr all k grealer
than J. Ek Is not. This guarantees that (a E1 E~ Is a free expression
0' tha nellt enclosing k-ell'presslon.

Now. c~nslder the k-ellpresslon enclosing that. To maximise the size of tis
mfes In Ihe same manner, all the EI which are free In II should appear

before the Ek which are not, and so on and so forth, In general. the optimal
Ordering under this criterion can be established as follows. Every El Is a

Iree ellpresslon 01 one or more enclosing k-ellpresslons. Call the Innermost

k-expresslon In whIch Ei Is not free Its natl~ k-ellpresslon. This Is Ihe
Innermost k-ellpresslon whICh bJnas a ...arlable In EI. II the nau"'e
k-expression 01 E. encloses Ihe natl...e ),-expresslon ot EI. then El precedes
El In tha optimal ordarlng. ThIs does not deline the optimal ordering uniquely.

because ellpress10ns wllh the same natl'i8 ),-ellpresslon can appear In any
order. This doesn't maner. because any ordering satisfying our condition
Is as optimal as any olher.

Nonce fhat an ellpresslon has no meaning outside Its nallve k-elilpres,lon,

bitCause the bound lIarlable 01 Its native k-ellpresslon appears In It

lomewhere. Also. constenl ellpresslons ha...e no naUve k-ellpresslons at 'II,

 30

because they ara free In alt A-expressions. For the 58ke of uniformity 1hey
.re regarded 8S being native 10 some nollona' )..-ellpresslon enclosing the
whole program.

We hay. deduced en opllrnal ordering to ml!lJllrnlse the s 128 or mf.s. lei

ua return 10 the subject 01 redundant parameters. First we observe thai the
complier can only choose the order 01 parameters 01 comblnalor!, and so
lhe only case In which II elln help make parameters redundanl Is when

one combinator Is defined directly as .. call 01 anOlher. For example. suppose

6 Is defined by

6 • kpq~8. a ...•...

No parameters are redundanl unt.u s Is. but is was the bound variable

or the A-expression 6 wu derived from. Therelor•. s was the bound variable

of the A-e.presslon Immediately enclosing a. If the parameters of a have
been or<lered opllmally l!l5 dellned above. then all perameters InVOlving s

coma al the end of Its parameter IIsl. It Ihere Is only one such parameter.

• "<1 II la simply s. then s Is a redundant para meier and can be eliminated:

otherwise s Is not redundent end nor are any of the other parameters. II

e I, redundant then the call 01 a musl lake the lorm

a £1 ••. Eft e

where s does nOI occur In Ella En. Therefore each Ej Is frae In 11. and
hence so Is (a El ... En). If n Is non-zero. Ihen 11 would actually be defined

by

11 • kpe. p e
 31

where p corresponds 10 (Q E1 En). II n Is zero. then

~.kB.«a

In the IIrsl case IJ Is equal to I lind we may ellmlnal. It using (a E 1

En) Instead of (Il (a EJ. En)). In the second case P Is equlvah!!tnl to a.

lind so elln be eliminated. So we see Ihat the opumat orderIng we "av8

defined also guarantees that parameters and comblnators will be made

redundant II possible. lind moreover, detection Is reduced 10 looking lor

two simple cases.

We shall demonstrate the Importance 01 Ihls opllmlutlon by eXhibiting a

(rather pathological> example _here It makes lin enormous difference 10 the

size 01 code prOduced. Consider the function Fn thai applies II function 0

10 lis n arguments In raverse Order.

Fn III k I 1 . . . l. In. G I n . . . I 1

The malllmal free 8Kprllsslons 0' the Innermost k-9xpresslon are 11 ... In-"
so. II we do not choose the opllmal order. _e could translate

Un. G In ... 11

Inlo

an In-I ... 11

where "n • Kh·l ... 111n. G h ... 11

where no parameters ara redundant. Bul no_

Pn _ ),11. •. kln-1. an In-' •.. 11

 32

which Is In Ihe same form 8S II was In orIginally. II *e continue the

tranalatlon In this way. we will define lIn-l .. lI 1. where

III. KII-I . . . I1Ji. ai+l IH .•• 11

, ... til

Here Cll Is of size om. 8nd so we will Ir8nslale the ),-e:lCpresslon of size

O(n) Info code 01 size Ofn 1 ). This Is bad news Indeed.

On .he other h8nd. It we use the optimal parameter ordering. we will be

'oreed to define

(In _ KI1. •• I". G In .•• 11

glvlOQ

," • 1. 11 .•• J n-l. an J.1 '" J n-1

All the olh8r 1Il -..111 be rsOundenl. 80 we will fInally define f'"l"I_an. The code
will be 01 size Ofn). The tremendous Improvement In this example leUs
U!I to expect 8 slgnUlclinJ Improvement In practice. In chapter e we will show

tholt using the optImal parameter order leads (0 8n Improvement In the

complexity 01 lhe code size.

4.4. OPTIMISIHO CONDlTtONAlS

When the method described above Is used In prac11ce .a find thaI. on the
whole, II performs well, but In some clrcumstanca5 performs WOrse lhan

Turner's method. An example where Ihls happens Is the naive Fibonacci

'unction
 33

fib III Y (lUbln. IF « n 2)

n
(+ (fib (- n 1»
(fib (- n 2»»

Alter Iranal8110n Into super-comblnalors. fib looks like

t lb II Y a
II • I(f n. If « n 2) n

(+ (f (- n 1» (f (- n 2»)

Because at the dlv'Cle-lIInd-conquer nalure 01 the function. lib Is called with

argument 0 or T dlsproporllonalely more allen then wllh other values. In

these cases II Is reduced l!lS follows

f 1b 1 red a t ib 1
red IF « 1 2) 1
(+ (fib (­ 1 1» (fib (- 1 2»)
reel tF t.rue 1
(0 (fib (­ 1 1» (fib (- 12»)
red 1

end we see tlla! the large expression (+ (fib (- 1 1)) (lib (- 1 2))) Is

conSlrucled end never evaluated. Ttle cost or construCting Ihls expression

I, likely 10 oU1weigh lhe rest 01 the cost 01 computln~ <lib ll. Since Ihe

expression I, never actually required. this effort I. entirely wasled.

In fect, Ihla kind of sUuaUon cen occur whenever we wrlle a condJtlanal

expreSSion. In the expresa10n (IF fa E:a EI). 11 Is certain that only one

of E:3 and E:II will actually be required. end so to construci both Is wasteful.
 ,.

Turne,', approaCh actually avoids conSlrucling the unnecessary one, since

the Construction Is done only on demand. In elleel. Turner Ireets (h. IF

El El Ell IS O.X. IF EJ Hh.E1J xl (().. X.El) x)). Since bOth 'uncllons ).x.El
and h.E3 Bre Independent of x. the only expressions that need 10 be
constructed on eech call arB the applications 01 these functions 10 x.

Fonunalsly. the same technique works for super-comblnators. When

compiling an expression

IF EJ £:1 £1

with the enClosing 1.-axpr8ulon binding 1(, we can treat It 85

IF EJ «l.X.El)ll) «l.z..EJ)x)

In this case h.E1 and h.El will be free expressions of the enClosln(l

l.-eltpresslon. and the combinator bOdy will contain only

IP El (a ll) (b K)

wnere a and b are parameter nemes corresponding 10 E]: and EJ. Only
two cetls will be allocaled towards E:z and EJ when x Is bound. Aller an

alternalNe hes been chosen Ihe selected brench will be construcled and

e~.lueled. or course. II one or both Of El: and EJ doe, not In~ol~e x lhen

h· will nOI need lhls lreatment.

Apptylng Ihls opUmlsallon 10 Ihe Fibonacci example. the new code Is

tib • Y (.rib. a tib (II tib)

II • d ib an. I' « n 2) n (a n)
II • .rib n. + (fib (- n 1)) (fib (- n 2»
Counting the number of application celie allocaled during a call at (fit! 1),

we find that 11 has decreased from 13 to 6. while the number ot cells

.lIocated per recursion for n greater than 1 haa only Increased from 13
to 104. This rapresents a greal Improvement In eHlclency.

We have discussed thls opllmlsatlon In the con led 01 conditionals. Howeyer.

It Is worlh replacing any large expression E Ihal may well never be evahlated
by (h.E) x. where I( Is the bound variable of the enclosing l.-el(presSlon.
To do thiS to every ellpresslon. though. would be to creale comblnators
almost as small as Turner's ones. and so 10 throwaway the main advantage
of super-comblnators. Forlunalely, mOsl el(presslons Ihat the programmer
wrtles are eventually evaluated. and so It Is only In cases like the condilional

Ihal It Is necessary 10 use this transformation.

4.5. GRAPHICAL COM81NATORS

The method we have described starts froth a translation ot th" source

program Into th" l. -calculus and prOduces comblnators trOth thaI. This Illds
to sO'"e Inefflclency In the treatment of declarations: for example .... :I '"

1 In 1+2 Is lass efficIent than 1+2. because the former Is translaled Into

u..x,x+2J 1. and thenca InlO (n.J(+2) 1. and so reqUires a combinator

application during Its evaluation. II would be nice It these two equlvllent
programs were equally elfjClenl.

We Cln achieve lhls II we elClend the comblnalor language slightly to Inchide

decl'ratlons. In this case let lC"'l In x+2 Is a pertectly valid elCpresslon In

Ihe oblect language. In tacl we can Interpret auch expressions dlrecUy as

 3.

graphs. For example. let x:::1+2 In cons l( Jl can be Interpreted as the graph

(cone tP
~<. 1 2)

end Ietrec 111 cons 1 11 In x can be Interpreted 8S

c<con. 1 r
Since 11'8 ere already worlling wllh graph ,educllon, the Inlroducllon of

comblnators with more complex graphs as bodies represents no reat

6xllnslon at aU. Under this graphical Interpretation of declarallons the two

e.amples 11'8 began with actually represent the same graph. and so ar.

certainly equelly elflclent.

We now have comblnators with general graphs as bodle,. rather than trees.

An ...ample 0' lhls Illnd of combinator I' Y. which we can define as

y • ~l. letrec x - t x in ~

since (y n I, reduced 10

l
d
In order to perform super-comblnalor abstraction on 8 program con1alnlng

decl,raUons, 11'8 first of .n floal the declarallons outwards as tar as possible.

uslno Ihe lact thet

~x.let J-E1 in EI let 1-£1 ln h .. El

 37

provIded x does not occur In E 1. This equallon holds because both sides

represent the same graph. and we are simply Choosing the mosl convenient

.rlnen representation of II. H....lng done this. we simply perform ordinary

super-combinator ebsirecllon. remembering tha' 8 varIable defined In e

declaralJon Is free In II l-expresslon If the express/on n represents Is Uhls

requires some care In Ihe case of tetrec deciarallons). The result will be

graph structured combinator code lor the program

FOr example. starting from

letrec fib - In. IF ( n 2)

n
«'n. + (fib (- n 1))
(fib (- n 2»)
n)

1n f lb

we del"lve

letrec fib - lCn. IF « n 2) n (a n)

and • - .n. + (fib (- n 1») (fib (- n 2)
1n f lb

80th comblnalors .re simpler and more efficient Ihen Ihelr earlier
counterparts. Tile Improvement In efficiency Is perllcularly great In the elISe

of mUlueuy recursIve declarel/ons. which under the old melhod were

".n,lelad Inlo the construcllon and subsequent destrucllon Of • IIsl of lhe

values being declared.

 3.

We will not pur!5ue this parllcular extension any further. A5 we have remarked

_boY8, Ihe presence of comblnators wUh graph-structured bodlas does 1'101

'nlerfere wl1h other 8Specls of graph reducllon. and so we can simplify 1he

resl of this thasls by discussing tree-strUclured comblnators only. Where

rellt'l8nt we will remark on the extensions necessary 10 cope wllh the more

gener.' cese.

4,1!I. EXPERIMENTAL RESULTS

To lest the super-combinator melhod In pracllce we compared II agalnsl

Turner's method. His method was chosen lor the 811perlment becausa lhere
Is already consIderable evidence thet II Is more effldenl than • fUy SEeD

machine lTurner791 lPeylon-Jones82J. A complier 10r 8 high-level lunclIona'

langulige was wrlnen. whiCh translalad InlO Ihe ),-calculus and COuld Ihen
Qenltr81e either kind of comblnators. Only the methOds of secilons 4.2 and
4.3 were used. The code wes run on a BCPl Interpreter which conlalned
precompiled deflnlUons 01 Turner's combinalors end could load

luper-combinator definitions. complied InlO mechlne code. If necessery. The

complier and reducer m.de a number of measurements. Including code size,
number of reductions performed during execution. lOla I number of cells
claimed. and run-time.

Ten ,mall programs were wr!lIen and benchmerked, rangIng from Ackerman',
Junellon 10 a unillcailon algorithm. and the resull, are summarised In the

table below.
Program Size ACode AReductione ACelle ATime
SiZe Claimed

1 26 36 -13 39 o
2 36 -10 -48 46 o
3 49 o -42 10 -12
4 H -5 -47 32 o
5 75 9 -36 9 -3
6 93 -9 -42 -16 o
7 106 -9 -23 -19 -21
8 115 -13 -30 -8 -11
9 307 -7 -39 21 -17
10 317 -31 -60 -35 -45

The figures in the table are the percentage

change in .oving to euper-comblnatore.

These resul1s do not demonstrate an ewesome superiority. Howeyer. It shOuld

be borne In mind thai all the exemple were very smell. and that the
.dyanleges 0' super-comblnators should become more pronounced 'or larger

programs (since small programs tend to compUe to small super-comblnators.

the advantage of wlarge executIon steps· Is 10sO. The lab'e Is arranged In

order of Increasing program size, end there Is a ylslble Improyement lIS

we look down the co'umns. The anomalies In the ·cells claimed" column

are prObably due to the fact Ihal we did nol use Ihe teChnique of secHon
<4.<4. which affects particularly programs ~. <4 and 9. We are heartened lhat

no progrsm ren more slowly whan compUed to super-comblnators. Therefore

. . are ressonably con'ldent the I the use of suPer-comblnators will prOduce

a slgnlflcsnt Improyemen1 In the per'ormance of real progrems. rot'
......_ e;f - - Uf&,i_.... S4L iI.a. At,..........·
 40

CHAPTER 5

SUPER-COMBINATOR COMPILERS

5.1. INTRODUCTION

In Chapler .. we described how functional programs can be Ir,nslated Into

euper-comblne'ars. but we did nol give an algorithm for C"ls translation.

We have not yet demonstrated that It can be done reasonably efficiently.

In section 5.2 we gl.... an Inlormel dascrlpUon 01 such en algorithm. which

converls a ~-8.presslon Into super-comblnators In II single pass. and In

the loll owing three eectlons we oulIIns three dlltersnl Implemenl811onsol this

a'gOfllhm In Imperative. funenan.' end logtc languages. Finally. In sBellon

5.7 d will review our e.perlence of Implemenllng • reasonably complex

algorllhm In lhe three different styles.

5.2. THE ALGORITHM

We rlrsl describe an algorithm thai Incorporates the techniques 01 secUon$

4.2 tnd ".3. The optlmlsaUon 01 $ecllon ..... Is aaslly added. and the

genelallon of greph-slruclUred comblnators hU already been discussed..

Since the algorllhm Is going 10 order combinator parameters optimally. It

'lIrt11 need to work with the netlve ,l-axpresslons of parts 0' the program.

We beQln by observing that the native ).-expresalon of any expression can

be ldenllfled by a single number. This 15 because the ).-exprenlons

 41

enclosing a particular program polnl lorm II sequence. and '0 we mey reler

to them by thalr position In this sequence. Thus. 1 refers to the oulermosl

l-expresslon. 2 refers to the nexl one In etc. We may therelore .,sDellle

• number Idenlltylng Us native l-expresslon with every expreSSion In the
program. We call thiS number an expression" lexical lelfel. or somellmes
lust It' 1"...,. II Is conll'8nlenl 10 assIgn constant expressions a level or zero.

• 'nce this corresponds to II notional l-expresslon enClosing the whole

program.

We can compute the texical level of every expression In the program In

• III1g18 recursive pass. as follows. We assign con,!"n! expressions 8 lEvel

01 zero, and we deal wlltt Identifiers using an environment which maps lhem

10 their lelllcal level (the number corresponding 10 the l-ellpresslon bindIng

them).

Wa assign applications Ihe maximum of the lexical levels of the func~)on

and argument. This Is jusllfled becausa the maximum corresponds 10 the

Innermo" of Ihe two naUve l-expresslons. II Is clear Ihal Ihe application

la not free In this l-expreu/on. but Is free In all Inner l-ellpresslons. and

10 this rule correctly Identifies Ihe application's native l-expresslon.

DeterminIng the lexical level of a l-expreSSlon Is more complex. We nollce.

Ihough. that Ihe problem can be avoided II we can replace l-expresslons

by corresponding applicatlve forms ·on Ihe lIy', In this clISe we can 8can

Ihe body of a l-expresslon. replace II by the corresponding

luper-combinator appllCIIUon. and then compute the lexical level of this

application using the rules above. This Is the approach we will take.

We m,y summarIse our conclusions so lar .s follows. The program synllllx

Irea Is scannad In depth-IIrsl order. As soon as any sub-Iree has been

scanned. lis lelllclIl level Is delermlned, and II Is converled Into the equlvalenl
 42

combinator 'orm. When the scan termlnatss, the whole progr.m will helve

been complied Inlo super-comb Inators.

9.Ior8 811plelnlng how ),-expresslons are complied on the fly. we consIder

the prOblem 01 Identifying maximal free 8wpresslons. We .re InlereSisd. nOl

Jus' In Identltylng the mles of lhe ne.rest enclosing ),-express1on. bUI In

Idenltrylng all the expressions In the program whiCh ere mlas 01 any

).-811'pre'llon. Fortunately this cen be done using only the level numbers.

For In expression 10 be maxima' 'ree It must first be free In some

).-elfpreSSlon. Ie Its level number must be less than that 01 the nearest

ancloslng l-ellpress1on. and secondly It must be maximal. This means that

It, level number must diller from the level number of the Immediately
ancloslng expression (or It must be Ihe enllrety 01 the body of e

l-ell:prenlon). Otherwise. both II and Ihe Immediately enClosing expression

would ba Iree In ell Ihe seme }. -expressions and n could not be mallimat
'ree In any. We can even Identify the }.-expresslon It will be mexlma' Iree
In; II ha. 10 be one In which It Is free. bUI In which the enclosing expression

Is nol IrH. Therefore 11 Is maximal free In the naUve }.-expresslon of the

ancloslng expression.

We may summarise lhese rules by: an expression 15 an mle If II 15 the

body of a }. -expression and lis lexical le...el 15 less than the le...el of Ihe

l-exp1esalon. or It It Is a function or argument and lis lexlcal level Is less

than the level of the application II lorms pari 01. The expression Is an mle

0' the }.-expresslon It Is the body of In the Ilrsl case. and Ihe native

l-t:lpreIlIOn of the eppllcatlon In Iha second case.

 43

Therefore. once we have scanned the body 01 II ).-elllpressIOn W8 will have

Idenlltled all lis maxImal Iree elfpresslons. This lells us what the parameters

at the corresponding super-combinator will be. We need 10 sort Ihem Into

the opllma' order. but this Is easy because the order depends on wwtllch

nallve ),-expresslons enclose which others. and thIs can be determined by

comparing level numbers. We can order the parame'ers opllmally by sorting
them Inlo order of Increasing lellcel level.

Now ... can replace the ).-express1on by II super-combinator 8ppll80 to

the mfes, and construct the combinator by replaCing the m,., In the

).-••pr."'on', bOdy by approprlale paramet.r name, (nllmbers In real

ImplememaUons). This completes the description 01 our algorithm.

For the most part. this algorUhm Is lalrly Simple. However, It Is quIte triCky
10 replace mfes by argument names. as we gaily said In the last pllregr~ph.

In an efficient way. The next threlJ sections outline lmplementetlons of Itlls

algorllhm In Imperative. tuncllonal and logical styles. and diller prlmarlll In
the solutions adopted to this prOblem.

5.3. AN IMPERATrvE COMPILER

We shall describe our Imperative complier by giving II Pascal-llke skel610n

and leaving ttle reader to flit In the details. We begin by giVing the tr'pe

01 tha synlax tree nodes:

 44

type node - record level: integer;

cue kind of
constant: ( );
variable: ( );
application: (tn. 4rg: node);
lambda: (bvar, body: node);
argument: (argnum: integer);
combinator: (numargs: integer;
body: node)
end;

Sine. the complier operates by physIcally transforming Itle original syntax

tre. Into the combinator version there Is provision 'or storing 8 level number

In Ihe node and there are 811ern811...e5 to represent comblnators.

Nollce thai we refer to combinator arguments by Integers. These correspond

10 Black offsels during e.ecullon. and their precalculatlon means thaI

combinator arguments do nol haye to be looked up In an environment 81

run-lime, This also meens that we can represent 8 super-combInator by

Its body and a count of Its arguments.

We wrlle node conslrucllon functIons In upper case by convenllon. end we

use three of them: APPLY to conslruct an application node. SUPER 10

construct a combinalor node. and ARG to construct a com blnator argument

name.
 45

1Ivt&
The complier malnlslns ....... Imporlant global yarlables, whose declaraUons

.re:

var cl: integer :- 0;

mfes: array [1 .. ] of

"'I.' . . eequence of address of node;

"~ [I..] of s~cc of Mel",. of nOlA..;
cI Is the currenl level, Inl1l811y zero. It Is Incremented when the compiler

begIns 10 scan is ). -expression snd decremented afterwards. so It always

holds the nestIng depth of the nee rest enclosing ).-ellpresslon. mfes Is used
to alore maxlmsl frea expressions .15 lhey are found. II Is 8n array with

one element per ).-ellpresslon enclosing (he current node. The element holds
8 sequence 01 all the mles 01 that ).-ellpresslon found so far. In 'acl. II

Is the addresses 01 the mIss which are held In the sequence; this permits

the replacement of mles by argumenl names alluded 10 above. bvs \.o\d.s

~ ............o.s ~ \\ o"",,-,,CIS .5 " _ ......!&s... +t.-.t
....~ __ Ioc. " .....1HrOII. ~ b.
The complier Iisell Is the 'orrowing procedure:
 '6

procedure compile (VAr n:node; e: env);

it laconstant (n) then n.level :- 0
elit lsvariable (n) then
n.level ,- lookup (n, eli ~.......t(""'s(n.k....I].QAcl"'U ,,)
elif lsapplication (n) then
compile (n.fn, eli compile (n.arg, eli
n.level max (n.tn.level, n.a:r:g.level);
for cpt 1n (tn. arg) do
if n.cpt.level < n.level then
augment (mfes [n.level] t

address n.cpt)
f1
ocl
eUt lelambda (n) then
el ,- el + 1; mC . . [elJ ,- 0; b-ts(c.l]:~<)i
compile (n.body, bind (n.bvar, cl, e»;
aortmfes (mfe. [ell);
..1­
n :- SUPER (length (mces [elk n.body);
tor: 1 in 1. . length (mfee [ell) do
n :- APPLY (n, derl!Jt mfes (el] (1]);

n.levsl :- n.arg.level;
it n.tn.level < n.level then
augment (mfes (n.level], address n. tn)
t1:
dare! mles [ell [1] :- "-HG (1)

ocl;~~ , In I.,,~~(.... lc.l)) "­

el ,- el - 1 ...--f ",,,,["1(iJ :­
f1 Milo (1~+("fU(d)).i)·
end. .el.
 41

Noles: n Is 8 ver parameter 50 Ihal the complier cen overwrite It witt! the

complied "'",alon 01 the node. The environment Is manIpulated with bind.

which returns lin envIronment equal 10 the given one except thet the glv,,"

variable Is mappet' 10 'he given Inleger. and lookUp which returns the level

01 a verlable. augment adds an element to the end of • sequence. eddress

returns the address 01 II variable. and derel permits thai variable to be

relld or updated vie lis a"Or8n. sorlmles sorlS II sequence 01 mfs addresses

Into optimal order. according 10 the crllerlon explaIned above.

The essential feature of thIs complier Is the way Ihal II stores mfe addresses

so that II can both letch the mle to construct the replacement applle8tlve

lorm. and update the reference to II In lhe function body to reter to an

.rgument name Insleed.

5.4. A FUNCTIONAL COMPILER

As In the lesl section. we shall begin by describing Ihe types we use. There

Ire several types In the functional complier. since we cennot use assignment

10 slore ell the Intormallon In the same nOde. We use • synt•• Ilk. Ihal

01 HOPE (Burstell801 to describe them.

The Input and output from the complier are NODEs;

data NODE - VAR( ••. )

CONST( ..• )
APPLY (HODE.NODE) I
LAMBDA (NODE. NODE) I
ARG(NUKBER) I
SUPER(NUKBER.NODE)
 .0

The recursive compile luncllon Itsel' mu,' h.8ve ,everal resul1s. since we
expecl II 10 return lhe level of en expre55lon. find Its mfes. and return a
ver.lon of the e-presslon wllh ell mles replaced by AR!3(n} nodes. We

enc.paul.te them In the type EXPA:

data EXPR - EXPR(NUMBER,EXPR-LIST,NODE)

Nollee Ih.8t II Is the complied form of the mle, thai Is relurned.

Slnct we expect lhe complier 10 replace mfes Dy AAGCn} nodes. end since
Ihe n.mes to be used depend on lhe context In which the expression being

complied occurs. we must pass Ihe complier an argumen! givIng the names
10 use tor e.Ch mle found. We use the NAME type to pBSS the neme of
.n mle. and ttle names of .11 I,s suD-mfes logelher.

data NAME - NAME(NUMBER,NAME-LIST)

FlntUy. environments heve type ENV which Is NODE-toNUMBERENUMBEA.

Given e VAA-NODE they return e level number .nd .n argument number

10 be used as a replacement.

The type 01 Ihe complier Is

compile: ENV-toNAME-LIST-.NODE-toEXPR

Since Ihe NAME-LIST gives the nemes 01 the mfes found. It will always have

.n Isomorphic struclure to the EXPA-liST perl 01 the result. This Isomorphism

will elltend to all levels 01 lhe ,Iructures.
 49

We shell describe the compile funcllon In a SASl-like nolaUon lTurner761.

wllh • lew elliensions. Its skeleton Is:

letrec compile env namee exp •

case exp of

VAR( •.. ) -+ ...

CONST ( ••• ) --+

APPLY(f.a) -+
LAMBDA (v , e) -. ...
esac

Compiling variables and conslanis Is easy: the relevant paris of compile afe

VAR( ... ) -+ let lev, num - env exp In

EXPR(lev, lJ, ARG(num»
CONST( . . . ) -+ EXPR(O, (J, exp)

Appllcallons are harder 10 deal wllh since the complier must decide whether

either the function or the argument Is an mIa, 8nd If &0 replace 11 ~y lis

name. GelUng Ihe NAME-LIST parameter right to the recursive calls 01

compUe Is also tricky. We shalt first show the necessary program and lhen
BlIplsln 11.
 50

APPLY(f,a) -+
letrlltc EXPR(flev,fmfllts,fnew) ..
compile env fnames f
and EXPR(alev,amfee,anew) ..
compile env anames a
and result, fnamee, anames ­
flev-alev -+
(EXPR(flev,fmfes++amfes,
APPLY(fnew,anew»,

,
take(£fmfee)namee,
drop(~fee)names) ;
let HAME(num,eubnamee):names' - names
In
f lev<alev -+
(EXPR(alev,
EXPR(flev,fmfes,fnew)~amfee,

APPLY(ARG(num),anew»,
eubnamee, names');
e lev)alev -+
(EXPR (f lev,
EXPR( alev, ameee, anew) : fmL ee,
APPLY(fnew,ARC(num»,
namee'r eubnamee)
in result
 51

Noles: The unusual functions used In this expression are .s !ollows: ++ Is

append. : Is cons. f: Is length 01 l!l Ust. and lake and drop return the first
n elements and all bul the Ilrsl n elements of • list respecUvely.

Notice that compile computes !names end enames from the results of the

recursive cells. even though they are themselves parameters 10 lhose celts.
11 guarantees thai II the original call had 8 names pare meier Isomorphic

to US EXPA-UST resull. then Ihe recursive calls will have 100. This circular

style 01 programming Is only possible because of lazy evaluellon, lind ey!n

SQ. one must convince oneself that lhe lunctlon Is IIctually defined. II Is

easy lor funeUans wrillen In this way to 1.11 10 termlnale. In the case or

compl/e. one can argue that all level numben are obviously defined. end
so all mles are 100. Finally. this Implies Ihal all names and new ellpresslons

are also defined end so compile always termlnales.

 52

"-...pres,lons sre complied by the expression below.

LAMBDA(v,e) -+
I_tree vlev. env' - bind v vnum env
and em!ee. enew ­
I.tree EXPR(elev' ,emtes' renew') ­
compile env' enames' e
and enamee' ­
vlev-elev' -+enames; l!lubnamee
and INAME(num,eUbnames)] - enames
In v!ev-elev' -+
em!ee', enew';
[EXPR(e!ev I. emfee' , ene",'}],
ARG(num)
and orderedmtes, permutation
let II ­

sort (~(7.EXPR(alevr7.1»

~ (7 .EXPR(blev, 7.7».
alev)blev)
(zip (1. .£emfee] emfee)
1n map (~( 1 • e) . e) II, map ( .. ( 1 , e) . 1) l!!I

and tf!lf!lult. orderednamel!l, vnum ­

mkap comb orderedmfee names
and comb - EXPR(O, [], SUPER(vnum,enew})

and enamea ­
a.p O.(l,n).n)

(eort ('ll,7)'lj,7). l<j)

(zip permutation
orderednam.ee) )
in result
 53

Flnt, we add Ihe bound variable 10 Ihe envirOnment using bind. which we

assume also assigns and returns Ihe new level number. Then we compile

the body 01 the l,-expresslon IlInd take account 01 Ihe lacl Ihal Ihe whole

bOdy might be an mre (If vlev Is greater Ihan eI8\1'). Nollee thai the

declaration of num and subnames can only be executed when enames has

precisely one elemen!. This Is perleetly acceptable. since In olher cases

nellher nurn or sUbnamas Is used and so Ihe declaration does not need

to be executed.

We go on 10 sort Ihe mles found Inio optimal order, and record Ihe

permul8!Jon used so thai Ihe corresponding names cen be restored 10 Ihe

orIginal order laler. The question mark used 8S 8 variable name mBans Ihal

the variable Is nOI aClUally requIred. sort lakes a parUal order and a Itst

and SOrts Ihe IIsl Inlo Increasing order. zip lakes IwO lists and returns the

corresponding list of pairs.

We can then use the subSidIary funcllon mkap. which lakes a comblnalOr

and a list 01 arguments lin reverse order) end conslructs the application

of the combine tor to the argumenls. It also needs to be passed the names

corresponding to the mfes of fhe result. or course. and It returns a list

of names to be used to replece the arguments and Ihe nexl available

ergument number

Finally. we can construct Ihe super-comblnelor itself and sorl the names

to replece the arguments Into the original argument order.

 54

II on I, remains 10 eJellnB mkap We do 50 below with no further 6)(planellon

since It 15 very similar 10 the APPLY case of compile

letrec mkap f [] (] .. (f, I J, 1)

and mkap f
(EXPR(alev,amfes,anew);args)
rnames '"
letrec EXPR(flev,fmfes,fnew),
anames' ,
nextarg ..
mkap f args fnames
and NAHE(num,subnames):rnamee' ­
rnames
and r, (names, anames ­
Clev=alev __

(EXPR(alev,
amfes++fmfes,

APPLY(fnew, anew»,
drop (£amfes) rnames,
take (£amfes) rnames);
Clev<alev ~

(EXPR(alev,
EXPR(flev,fmfes,fnew):amfes,
APPLY(ARG(num) , anew»,
subnames,
rnames' )
1n (r,
NAME (nextarg. anames) : anamee' I

nextarg'H)
5.5. A 1OO1C COMPILER

We present our logtc programming complier In Prolog IKowelskl79). Th9

dalatypes used correspond to the one:3 In 'he luncllonal ...erslon. so we shllil
nol describe them further. The compile predicate lakes the form

compi1e(env ,node,expr,namee)

where env Is an environment. noeJe Is Ihe orlgln81 ellpresslon. expr Is a

structure wllh funclor EXPA containing the level. mles and new form 01 the

elpresslon. and names Is a list of nemes Isomorphic to Ihe mre componenl

of elpr.

compile Is dellned by lour clauses. lhe Ilrsl lwo 01 which compile variables
and conSlsnls. They are:

compile(env,VAR( ... ) ,EXPR(l, [] ,ARG(n», ()

looKup(env, VAR( ... ) ,l,n).
comp11e(_,CONST( .•. ),EXPR(O,[] ,CONST( ..• )) ,[]).

We have assumed Ihal lOOkup finds lhe level and corresponding argument
number 01 a variable Irom the envlronmenl. and that Ihe underline Is an
anonymous variable.

Applications are complied by Ihe clause

comp11e(env,APPLY(f,a),expr,namee)
compile(env,f,fexpr,fnamee),
comp11e(env,a,lIexpr,lInameB),
apply(fexpr,aexpr,fnameB,anames,expr,names).
 56

The luncllon and argument are compiled firs!. and then apply Is used to

combine the results. II the luncUon and argumenl have Ihe same level. Ihen

epply simply applies one 10 Ihe other

apply(EXPR(l,fmfes,fnew),EXPR(l,amfeB.anew),
(names,anames,
EXPR(l,mfeB,APPLY(fnew,anew» ,names)
append(fmfes,amfes.mfes),
append(fnamee,anames,names) .

Otherwise Ihe one with Ihe lowest level Is made Into an mle

apply(EXPR(fl,fmfes,fnew),EXPR(al,amfes,anew).
(names I an ames r

EXPR(al,EXPR(fl,fmfes,fnew).8mfes,
APPLY(ARG(n) ,anew»,
NAME(n,[names).anames) :- [Hal.

apply(EXPR (f L fmt es, fnew) , EXPR( air amf es, anew) ,

fnameB,8names.
EXPR( f 1, EXPR(al, amfee, anew) . fmf es.

APPLY(fnew,ARG(n»)J.
NAME(n, anames). fnames) : - ! l>al.
),-ellpresslons are complied by the clause:

comp11e(env,LAHBDA(v,el,expr,namee)
b1nd(env,v,nextarg,env',lev),
compl1e(env',e,eexp',enarnee'),
lambody(lev,eexp',enames',
EXPR(elev,emfes,enew},enamee),
Bortmfes(emfee,enames,Bortedmfee,eortedname3),

mkap(EXPR(O,[],SUPER(nextarg,enew»,
aortedmfes,Bortednamee,

expr,namee,nextarg).

Flrsl we bind v Into the environment. gIving It argument number nextsrg.

We assume 'hal bind computes both the new environment env' and the new

level number lev, 85 In the functional V8rslon. Then we compile the body.
and take account 01 (he lacl that II might lorm 8n mfe by IbeIF (In lambody).
aortmfes sorts the mres and names Into optimal order, and finally mkap

Is used to construct the replacement expression. assign names and compute

Itle number of arguments 01 the combinator.

lambody Is dallned by

lambody(lev,EXPR(lev,mfes,new),namee,

EXPR(lev,mfea,new),namee).

lambody(lev,EXPR(l,mfes,new),namea,

EXPR(lev,(EXPR(l,mfee,new)],ARG(n)),
[NAME(n,namea))) ;- l<lev.

which Just makes the bOdy 01 the ),-expresslon Into an mle II It doesn't
con'aln the bound variable.
 58

mkllp applies lhe combinator 10 Us arguments and assigns Ihelr names. Its

form Is

mkap(comb,mfee,namee,expr,enames,nextarg)

where comb Is the combinator to be applied. mfas Is lis arguments. names

Ihe names 10 be essig ned 10 those argumenls. expr the resulting expression.
en ames the names at mIss of that expression. and nexlarg Ihe next Iree
argument number. II Is defined by the clauses:

altap ( f • r)• I ) • f • r 1• 1) •
akap(f,arg.mfee,NAME(nextarg',anames) .names,
e,enames,nextarg)
mkap(f,mfes,names,e',enames',nextarg'),
apply(e',arg,enames',anamee,e,enames),
nextarg 19 nextarg'+l.

We have made greal use of the Prolog variable In this program to dls1rlbute
(he computation 01 name structures and argument numbers to convenient

places. We feel that this program demonstrates the power of Prolog rather
well: such things as sorllng mles and names proved considerably simpler
(han In the luncUonal equIvalent. We have also used thla style 10 Include
an opUmlS8uon In lhe complier. which removes repeated mles from the

par_meter list. This can be done by the predIcate

optlmlee(mfes,names,mfes' ,names')
whiCh lakes lin IJnopUmlsed mle Iisl In mfes .nd names. and computes the
optimised equlvalenl In mles' and names', II Is defined a&

optimise ([ J, [J, [J , []) •

opti~iee(mfe.mtea.name.namea.

mfe.mfee' ,name.namee')
not(member(mfe,mtee».
optimiee(mfe8,namea.mfee',namee').

optimise(mte.mfee,name.namea,mfee',namee')
element(mfea,i,mte), element(namea,i.name),
optimiae(mfea.namea,mtea'.names').

where elemenHllsl.el.lnded Is Irue It el Is lhe Indexth element of IIsl. lhe

corresponding opllmlsallon In lhe functional complier WIIS 100 compllcliled
to Include,

5.1. CONCLUSION

The super-combinator abstraclion algorIthm described In 'his chapter Is

reesonably complex, and so II Is lnleresllng to compare our three different
Implemenlallon5. Curiously, the Imperallve program Is the shorlest 8nd
requires least explanation, ThIs Is partly because we IISsumed that mlny
facilities 01 functional languages were available. which Is not usually the
case. However, lhe Imperative solullon represenls all Its Inlormetlon In Dna

com pie. data-structure. ralher Ihan breaking It down Into simpler unlls. and
It depends crucially on slde-ellecls happenIng at the rlghl times. We contend

Ihat these make It dlfllcult to understand, even Ihough II Is relal!vely short.

Of course. they also render II tOlally unsuitable lor a parallel Implementalion.
 60

The funcUonll1 and logic programs lire very sImilar. However. the

Isomorphism between name slruclUres and mle structures (which Is crucial

10 both 01 them) Is concealed In the luncUonal version. sInce the neme

.'ructures appear 8S argumenls and the mfe structures appear 8S results.

This renders Ihe structure 0' the functional version more obscure. We

consider this strong evidence thai the need 10 specify dlrecllon of Information

110" can occasionally lead 10 badly structured programs. Excep1 for this point

we found the funcllonal style more expressive than Prolog. Both the functional

and logic versions are divide-and-conquer programs. and so both are

suitable for ImplementaUon on 8 parallel machine.

We would have liked 10 claim thai the experience of Impleme"tI"g the same

complex algorithm In three dlllerant languages demonstrated conclusively lhe

superiority of functional and logiC programming over Imperative programming.

UnlOrlunalely Ihls example Is nol terribly conClusive. The Pascal hacker may

well claim thai the Impera!lve ImplementaUon Is easier to understa"d because

II's shorter. but lhls Is because he Is used to lhlnfllng In lerms 01

,lda-effec1s and time. However. our own underslandlng 01 the abslractlon

Ilgorlthm was advanced considerably by the ellperlence of wrlttng lhe

lunctlonal and logiC versIons. while writing the Imperative progrem served

only to contuse us. In this respect we fael that Ihe aumple has demonstrated

lhe superiority of declarative programming. 11 only 10 ourselves.

 CHAPTER 8

ANALYSIS OF EFFICIENCY

8.1. INTOOOUCTION

In this chapler we allempl II theorlltlCll1 analysIs 01 the ellJclency 01

supllr-comblnlltors. Although II Is dlfflcult to oblilin concrete results In this

IIreli. we hllve two results which we think slgnlflcllnt.

We consldllr transilition of II progrllm of 51le n Into super-comblnillors and

Inlo Turner's combinators. lind we lInd the Order Of 1he slle of code

produced. Burton has IIlrelidy shown that the comblnillor code need be no

larger than O(nlogn) (BUrlon82): we will show thet. on IIverllge. It Is Indeed

thIs Ilirge. We will also show thllt Ihe super-combinator code Clln be no

Ilirger Ihan O<nlogn). bul thllt II mlly sometimes be thIs Ilirge, lind we will

offer some evlden ce thllt II Is usulilly smlliler.

Thl5 problem Is Interesting for two rellsons. Firstly. In ellrly ImplementeUClns

using Turner's comblnlllor5. excessive code slle WIIS II serious problem.

Some smllll programs complied Inlo code so Ilirge Ihlll 11 could not posSibly

be run. We wouM like to demonslrllte thllt Ihls Cllnnot hllppen using

auper-comblnillors. Secondly. III lell51 In ·stralght line" programs. coda Aile

Is an approltlmale measure of execution speed. We use this 10 matle a hery

vllgue) comparison with the SECD mllchlne.

 .2

•. ~. TURNER'S CQMBINATOAS

We begin by quoting the appropriate results for Turne,'. comblnators.

kennaw.y haa shown that the translallon method Turner describes can lead

10 Ihe producUon 01 code 01 size DIn 2) (Kennaway82l. Burlon has given

an Improved method and ,.hown Ih", It produces code 01 size O(nlognJ In

the worst case IBurton82J. It Is BUy 10 ase that this result Is opllmel. end

we give • proof here

We take the slZ9 of an a.prosslon to be the number of nodes In lis synlax

tree. since Ihl!. 15 consistent wllh the gTeph reduction applications we have

In mind. This Is equl'f'elenl to the number of symbols In lis ..rlllen

TepresenlaUon. nOI counllng brackets. In this deflnlllon "'8 diller Irom Burton.

whO counts the lengths 01 IdenUlIers In his sIze measure. (However • •e heve

1ranslaled his result. glyen above. Into our terms).

Now. since Turner usee a llxed sel 01 comblnators. 5 dlfferenl ones SBy.

11 I, ~lear thel there ere al mosl 5" dlfterenl ~ombln8tor expressIons Of

size n. However. sln~e a l-expresslon may conlaln Din} different symbols.

there may be O(n") different l-expresslons of size n. In lact. •e can exhibit

n" non-lnlerconverUble ). -expressions 01 size OCnl. being

lVI ..• Vn. CONS W1 ( .•. (CONS w" NIL) ••• )

where all the YI are dlrrerenl Identifiers. and each wi Is one 01 the VI. These

l-exprestlons are cleerly non-lnlerconvertlble since they all yield dltterenl

resuns when epplled to n different .rgumenIS.
Now. each 01 the nn k-expre5510ns we have given mU5t compile 10 ..
dltlerent combinator expression. so If N Is the wOrst case cOde size then
there must be al least nil different combinator expreSSions 01 alze no more
than N. We muse therefore have

nn ( gN

and so. takIng logarllhms.

nlogn ( NlogS

That Is. N Is at IBa51 OlnlognJ. We can actually derive a stronger result

from this argument" no mailer how gOOd the complier. almost all

).-expresslons must compile to code 01 size at least Ocnlogn). Burton's resull

Iherefore lells us that the sverSQe code size of a ).-expres510n aller
translallon 10 comblnators 15 OCnlognJ.

'.S. 8UPEA-COMBINATORS

An analysis of the complexity 01 Ihe code size of 5uper-comblnators Is much

more dllllcull. We wlll begin by observing thai transJallon 10

super-comblnatOrs does jncreese (he 511.8 01 the program. and by Identllylng

the source of the Increase.

Consider 8 single ).-expresslon h.E. and suppose It has mle5 E 1 tel En.
It will be translated Inlo

(..:11 •.. 111%. E(II/EI]) El ..• En

 6.

" we consider the syntax tree of the body of the original ~-8Xpresslon, we
••• thai It Ie broken up Inlo the mfes. and the non-mre part. The parts

811 reappeer In the code. the mfel as 8uper-comblnelor arguments. end

Ihe non-mle part 85 the super-combinator body. The diagram beloW'

IIluelrltt. thill, using dot1ed lines to dellmll the perts of the tree .

•y

IC:l11:1y ,,
=
,,
,,
/1
"'
• J( "z\:.J

I'

Since luper-combinator ergumenls are actually represented by Inlegers It

It onty nec8asary 10 store the IIrlty 01 II super-comblnalor 81 liS heed. not

th' nemes of ItS .rguments. I(ll .. Ir«,E would therelore be represented lIS

«n,E. Ind so we count the rand argumsnl nllmes as one node. the same

I' Ax.
h appears thai the only nodes In the cOde which do nof correspond directly
'0 nodes In Ihe original !lource are thoae usad to apply the combinator
10 Its arguments. There Is one 01 lhese lor each mle. and !l0 we may

conclude that when a program Is translated Inlo super-comblnators Us size

Increases by the lolal number 01 mles found durIng translallon.

Since accounllng for mfes Is so cenlral to our problem II I!I worth studying
them further. We haya already pointed oul Ihat an mle Is an mle of Ihe

naUye ~-expresslon 01 lis Immediately enClosing ellpresslon. To see why.

suppose Eo Is an mle. and lis Immediately enClosing expression Is El.
Cleatly Eo and E 1 haye different natl'te l. -ellpreaslons. and the naUye
l.-expresslon 01 Eo encloses Ihal of El. Since El Is Itee In all ),-expresslons
enclo!lad by Its naU'te one. Eo cannOI tie maximal Iree In any 01 Ih,m.
Howeyer. since E.1 Is not Iree In Its na!lye l.-expresslon. and Eo Is. Inen

Eo Is maximal free In II. In Implementallon letms this corresponds to

Observing thet. since E1 will be passed 'tom lis naUye l. -ellpresslon 10 lis
proper locallon. E II will be carried along Inside It and need nol be pnssd
uparalely.

Moreover. each mrs will only be an mle 01 ona ),-exprasslon. That Is. II

will not be an m'e of the expression prOduced when the l. -expression II

Is en mte of la complied. This Is becausa. when Ihal ),-ellpresslon Is

translated Into e combinator eppllcallon. lhe comblnetot parameters are
ordered by the optimal orderIng gl'ten In seclion 4.3. Consider EI In

CI E I _.. EI •.. Ell

By the optlmallty crltetlon. ell Ihe EJ to the leU 01 EI haye nallye

l.-expresslons enClosing or equal to the natlye l.-elpresslon of EI. and so

tCi E 1.. El) has the same natl'te ),-expresslon a. EI. This means thet EI Is

not an m'e of eny other ),-expresslon. allltouglt hI EI ... EI) may be.
 55

So. now we know that we need only count mfes, each or which will only
Occur once. HOW8\1er. the slluatlon 15 com plica led by Ihe fact Ihat 8lCpressions

Introduced during Ir8n5111tlon molly themselves turn oul 10 be mfes 01 rurther

),-expresslons. For example. i1 e ).,-ellpresslon has Iwo mfes E 1 end E.

wllh dlllerenl native ),-expresslons. Ihen It wlll be replaced by (<I El Ell,

and (II Ell wilt I\sell be an mfe 01 the nlilive ),-expresslon of El. NOlonly

tnls, (lJ El EJ) me)' be an mle of the next enclosIng ),-express1on, EIther
of thes8 new mles may then cause enClosing ),-expte:nlons to have slill

further new mfes. The problem 01 lInalyslng Ihe code StZ8 01

lupar-comblnatora 15 accounting lor lhese mfes generated during translation,

.! well aa lhose originally present 11'1 Ihe progrtlm.

8.4. ACCOUNTING TREES

In order 10 keep lrack of ttle genera led mfes we Introduce the notion 01
an accounflnQ tree. An Iccountlng free lor a program Is a Iree whose

verllces are ttle l. -expressions ot ttle program. and whose edges are
constraIned to connect a l.-expresslon 10 an enclosing one. By con'tentlon
we draw accounting trees growing upwards wl\tI the notional ouiermOs1

l.-t1apreulon II 'tie boltom. so. lor example. the diagram below shows two
possible accoul'l1lng Irees lor

l.a. pb.l.c.b) (ld.a)

 67

r '\lid
'V
d

'a

In our
root

diagram. edges connecl k-expresslons

1 above to enclosing

l-expr85stons below.

For any program. thefe 15 8 ',allesl" accounting Iree which we call the Initial

accounting Iree. In which all ).,-e-.:pre,slons are connected '0 their

Immediately enclosing >.-expresslons. The Iree on the left above Is the Inillal

8ccounrlng tree for the given program.

Slarllng Irom I!I program's lolllal accounting tree. we will use the mles

originally present In the program, one by one. to transform the accounting

Iree so thai we can discover Irom II Which lurlher mfas will be genera led

as 8 consequence of the original ones. The number of addillonal mles due

10 an orIginal ana will also be 8 reasonable measure 01 lhe 'cosI" 01 the

aS80cieled transformallon 10 the trea. This will allow us to translate Ihe

problem 01 super-combInator code size Inlo an equlvalenl one concernJng

the cosl 01 a Iree-lranslormallon algorllhm.

 68

During the translormation prac8es we will pr8S8r'l'8 the lollewlng tnvarlant:

••ch l-ellpresslon will be connected to the nallve >..-ellprees1on 01 Its most

recently Cll.covered mfa. original or generated. unless none 01 Its mfas have

been dleco.... red yet. In which cese It will be connected 10 lis Immediately

encloSing >.. -expression. The order in which original mfas are used 10

transform the tree will be conSlralned b'J' the rOllowlng rule: If Eland E:2

are mIn such that the neUve >.. -e;tpresslon 01 E.1 encloses (15 nearer the

root 01 the tree than) the naUve A-expression of E 1. then we wilt use E 1

belors we UI!I8 E:a. ThIs constraint means thai we wHl always dIscover the

mfes 01 any parllcular )., -a_pression In reverse order. although not

nec.ts.flly one alter another. ihla property Ie crucial 10 the prOOf.

No... consider processing an original mfe E 01 A-expression A. Let N be

E'e native ).,-ellpresslon. We must at leasl remove Ihe edge leading

downwards from A and conn eel A 10 N Instead. In order to preserve our

Invlrlanl. Suppose B Is the ).,-ellpress1on Immediately below A before this

mOdification. Then we know that A will be replaced by a comblnlilor

expression

It .•• E E'

_h.re E' has neUve ).,-ellpresslon B. or that A will be replaceo by

It ••• E

Ind B Is ..... Immedlalely enclosing A-expression. We know this because

of the optimal ordering 01 mles. end because mles are discovered In reverse

order as noled 81:10ve. In either cue.

(a .. _ E)

will be • genarated mfe 01 8. So. 8 must .1150 be reconnected to N. and

the sama argumenl applies 10 C. tha ),-expresslon Imme(llately below B.

and so on. There must be a chain or ),-expresslons B.C.D .... leading

downwards Irom A Ihrough the accounting tree. each ot which should be
reconnected dlrec lIy 10 ..... and each of which has a generated mle with

naUve ),-axpresslon N. ThiS chaIn must termlnale somewhere. and cannol

termlnale at the root 01 Iha accounting tree since the root nollonal

~-expresslon cannot possibly have any mfes at all. II must termlnale Instead

.1 N. Strictly apealdng. It tarmlnates al a ),-expresslon JUSI above N. M say.

since M already has an mfa with nallve ), -exprenlon N and sO no new

mla Is ectually generated. The Iransformatlon that resulls Is Illustraled In
the diagram below.

\/
I
\ A

\/
\
~

.:.....: ...
- \/. . vvv
\/
i
 70

So. If the path length from /II. 10 N Is n. then the presence of 8n original

mi. 01 /II. with native L -exprenlon N Implies the presence or n-3 generated

mi... We l!ISsoclala a ·cosl· or n with performing the assccllllied

Iranslormatlon 01 the accounting tree. so the lolal Increase In size or 8

program when translated 10 super-comblnators will be lese Ihan Ihe total

cost of the transformations applied to Its accounting trea.

We may there lore conclude thai when a program Is translated Inio

super-comblnators the Increase In Its size I, less than Ihe maximum lolal

cost of applying as many "!lanen" opera lions as the prognllm originally

contaIns mfas to 8 tree 01 8S many nOdes u the program conlalns

l-axpresslons. where a f1allen operation consists 01 selecting a palh In the

tree leedlng towards tha rOOI and reconnecting every olhe r node on the

path direCtly 10 (he and near.llS1 lhe root. end where the cost of a flallen

operation Is Ihe lenglh of Iha palh.

1.5. UNION-FINO AU30AITHMS

We break off from enalyslng super-comblnators now to (lescrlbe the

unIon-lind problem end algorIthms lor Its solullon. II will lurn oul lhat

previous analyses of these algorllhms will enable us 10 complele our analysis

01 euper-comblnalor code size.

Tha UNION-FINO problem concerns the menlpulallon of a number of dlsJolnl

sets which partlllon a universe. tnlll811y the universe Is pertltloned Inlo

slnglelons. and therealter two kinds 01 operallon may be applied:

UNIONCA.B.C) which unites sels A and 8 Into a new set called C. and FINo(x)

which detarmlnes which set the element • belongs to.

The basic UNION-FINO algorithm represents the sets as Irees 01 elemenlS

wUh set nllmes 81 the rools. UNIONs are performed by making the rool 0'

'he tree represenUng one set point at the root of the free represenllng Ihe
other. FINO' are performed by following polnlers unlll one reaches a rool.
UNION I" therefore 01 constanl cost, and FINO 01 cost proportional 10 the
langth of Ihe peth lollowed. Two optimisations can be applied to Ihls
elgorllhm:

The Collap~/ng Rule: when a FINO Is performed. all nodes on the palh 10

1he root are made to poInt dlreclly lit Ihe root. speedIng up subsequent

FINOs.

Tha We/ghtlng Rule: a UNION operation alweys makes Ihe set containIng

lewer elements poInt at the set conlalnlng more. This will elso lend fo speed

up subsequent FINDs.

n turns out thel. lor the purposes of analysIng these algorithms. we elln
consider all UNION" to occur first. building a tree. followed by partial FINDs.
which start at an element and follow a path part of the way to the rool.
H should be clear thai the algorithm using only the COllapsing rule performs

exactly the same kind 01 trllns'ormatlons as our menlpulallons of accounllng

Irees In the last section. end moreover the cost measures are the same

In bolh cases.

Tarlan hes analysed these algorllhms ITarJan75] and hIs results are Ihal.

If m)n partial FINOs IIrll performed on a tree 01 n nodes (hen the lotal

cosl. Hm.n>. will SIlUSI)'

t(a,n) - O(m.maz(1,10g(nJ/m}/log(2m/n»)
 72

If only Ihe collapsing rule I, usad. and

t(.,n) - O(m.a(m,n»

II the CO".pslng and weighting rules are both used, where a Is relaled 10
the lnyer,s of Ackerman's luncllon and grows so slowly (hal II Is bounded
by • ,men constsnl for ell practical purposes.

Fischer has given en ,xample rFIscher72J where

t(n,n) CI nlogn

using only the collapsing rUle. showing the former upper bound tight In the
cas. m=n.

e.e. CONCLUSION

We may thereto'. conclude thai the Increase In the alze of 8 program

conUllnlng n "-expressions and m>n orIginal mles Is bounded by

O(m.~az(1.log(n)/m)/log(2m/n»)

LeWng «=m/n be Ihe average number of origInal ml.s per },-expresslon.

Ihl, bound can be resllpressed 85

O(_.max(l,loQ(n/a)/log 2a»

It Is cle.r Ih81, for a>1. IOO(n/alllog 2a decreues .s a Increases, 8nd

11 Is equ81 10 1 for a= -/(nI2). Therefore, lor a> -/(n/2l the bound Is linear
In m, and there'ore linear In the size of lhEt program. For 1«(I( v(n/2J. the

bound 15 less than the value at (1"'1. O(nlogn). In 'act. logln/(I)/Iog 211(k

when (I>(k.,vC2n»/2, 50 lor lhEtsEt values 0' a Ihe bound Is km. For 0(1­

we observe that decreasing thEt numbEtr or FINDs cannot possibly Increase

the overall cost. which Is lherelorll sUlI bounded by O(nlogn).

Theretore IhEt code size 01 II progrllm of size N when lranslilled 10

super-comblnillors Is bounclEtd by OINlogNl. Flschllr's example shows Ihat

there are programs whose code Is this large; howevEtr. his IIXlimple Is highly

symme1rlc lind Is nOl II likely strUClure tor II relll progrllm. For programs

where lhe avEtrage number 01 mlEts per ~-lIxpresslon 15 large enough fhen

the code size Is linear In the program size, For progrllms "blllllnced' In

the sense that their InlUal IIccounllng treEts could be constructed by ..

seqUEtncll of UNIONs satisfying 1he weighting rule II much Ilghler. IIlmost

IInEtlir. bound applies Lastly. Ihese are worst-CllSe bounds. lind we have

been told that In praC11cEt Etven Ihe COllapsing rute 15 sufficient to maKe

II UNION-FIND algorllhm run In nEtarly-lInelir timEt, This suggests thll! InEt

average bEthllvlOur Is beller than OlnlognL lind thllt In prllcllce

luper-combinator code size will be nEtarly linear In the size 0' the program.

We believe Ihal the Code size of combinator and super-combinator

Imp lemen III lions rellecls a ·stow-down factor", Whereby Ihe slime expression

may lake 10ngEtr to execule II II Is part 01 a IlirgEt program thlln II It 15

pllrt of a small one, We believe code size Is a gOOd measure of Ihls

slow-down laclor because. In each kind 01 Impiemeniatlon. Ihe time to IIpply

a combinator Is proportional 10 Its size; Ihere arll no "InstrucUons· which

may Ilike a 't'arlable amount 01 Ume to execute. It might appellr thllt this

slow-down Is presenl only In these ImplEtmEtnllitlons. but In 'act It 15 11150

present In thEt SEeD machine. and In conventional Implementations of

languages like Pascal, In IhEtse cases II appears lIS the time ror an

environment lookup. which varies with the number 01 names In Ihe

 74

environment. Because there Is no dIrect wey 01 relallng the Bverage number

or names In scope wllh lhe size 01 II program II Is dllllcult 10 estimate the

slow-aown 'aclor of the SEeD mechlne. btll we observe lh81 II Is pOSSible

to wrlla programs 01 size n with 8 slowdown 'acior 01 Din) lor Henderson's

SEeD machine (HendersonBOl. and It seerns reasonable 10 us that the

average number 01 names In scope should Increase at le8s1 with the
logarUhm 01 the program size. givIng 8 slow-down lactor of 81 leasl O<logn).
Since the mall/mum possible slow-down faClor 'or 8 super-combinator
ImplementatIOn I~ DClogn). this suggests super-comblnators may be an
Inherenlly more elflclent Implement.Uon method Ih8n lhe SEeD machine.

w. conjeclure that any Implementation method for the l, -cl!llculus (or any
other languaga .,.llh Algol-like scope rules) must have a slow-down ,aclor.
ot ......t
and thai In Ih. worsi case thl, 'aclor will be A OUogn).
 I CHAPTER 1

EVALUATION ORDER

1.1. INTRODUCTION

In Chapter 2 we argued thai one 01 the mosi Imporlanl advantages 01

functional programming languages Is thai they relieve the programmer of

the burden of expressing a desired eveluetlon order through the struclure

of his program. In (his chapler we exhibit cues whe~e a partlculer evalullllon

order Is critical 10 Ihe efficiency of the prognm. and we suggesl

structure-Independenl ways of controlling II.

let us first con sider how conirol over evaluation order Is used to ImJjrove

efllclency In Imperative languages. We take lIS an example a progrllm lhat

reads e file and prlnlS the number 01 capllal As In II. One possible evaluillon

order Is to read Ihe whole lIIe Into memory IIrsl. and then counl all lhe

As In It and print the resull. This would be programmed as:

read the f1le;

count the Aa

(~
 76

Alternlll\'8ly. we could read the Ille one ctulrllcl.r at 1!I time and update

Ihe count as we go. This would be progr.. mmed .5;

open the fIle;

zero the count;
until end or rile do
read a character;
if it'e an A ~en increment the count f1
04,
claee the file;
print the count

Thl' laUe' program Is much less modular than 1he former: the two logically

Independent ope,etlons of reading the file and counting the As have had
10 be programmed loge the,. However. It Is much more efficient beceuse
tI needs only • constant amount of spece 10 run In. The former method
may requires en amounl of space which depends on the size of the file.
Thl. I. such • gross dillerence that we cannot aflord to Ignore II.

So Bv.lu.Ung e program In the correct order can be crlUca' 10 space

elllC:lency. Since we cannot (and do nOI wanl to) express an order through
the structure 01 a luncllonal program. we must ask ourselves: IIrslly. Is lazy

e~.luellon (or eny other slrategy already In use> the correct choice enyway,

.nd If nol. how can we besl cause the correCI order 10 be used? In secllon
7.'1 w. will eJlhlbll eJlamples WhiCh demonstral8 (hat lazy evaluaUon- Is

sorn.tlrnes e poor choice. In secUon 7.3 we will give an eJlample that no

..qu.nll.' .....Iuallon order can evaluate efficiently. We conclude thai a

per.llel eb.lrecl machln. Is a prerequisite for space-elflclenl evaluallon. and

we ...ooe.. eJlpllcl! structure-Independent ways 01 controlling parallel

"',t.•
Now let us move on 10 a simple lellt processing funC110n tp. Ip allows

.bbrevlallons 10 be defined end sUbsUMed Into the lext. For simplicity. we

usume thar only one abbrevlallon need be defined at III Ume_ An ebbrevlallon
definitiOn appears In the tellt u Ihe abbrevlalton enclosed In brackets. and
subsequent ellclamallon merks .re replaced by Ihe lext between Ihe braclr;ets.
So. lor ellample. tp would convert

(abbreviation)An I definition appears in the text

as the ! enclosed in brackete.

InlO

An abbreviation definition appears in the text as

the abbreviation enclosed in brackets.

tp lakes ,,",0 argumenls. the telll 10 be processed (a list 01 characters) and

the Inlllal abbreviation value, end returns lhe IIsl 01 characters In the result.

" Is defined by

tp IJ ab . [J
tp (-I-:X) ab • ab ++ tp x ab
tp (- (.: x) ab .. tp (after ")" xl (before .). x)

tp (c :1) ab - c: tp x ab

where belore and afler are 8$sumed to return the characters In the second
OW4rtU\tL
.rgument up to and beyond the first I II 1 II or the IIr" argument

respectively.
 7.

This observation moUvates us to accumul!lle the sum differently. so Ihat

partial sums lire reducible earlier. We Introduce an accumulating parameter.

\hus:

length 1 - length' 1 0
length' (] n - n
length' (a:x) n - length' x (n+1)

Now leng1h 11.2.3} can be reduced as follow,:

length [1,2,3) red length' [1.2.3] 0

red length' (2,3] (0+1)
red length' [2,3J 1
red length' [3 J (1+1)
red length' [3J 2
red length I [1 (2+1)
red length' [1 3
red 3

u,lng only constant space. Howellsr. thl, conslant space versIon depends
on Ihe second parame1er of length' being reduced belore length' Is applied

lise". This would nol be so In a completely lazy system: Inslead the sum
wculd be accumulated unevalua'ed, as the 811presslon (((0+1)+11+1), and

would only be reduced when returned as the re~nJlI of lenglh. So III .llllzy

lValuator "oull! make this version no more efflclenl than the lormer. We

will ,ho" Illter how we can force 8n otherwise IlIzy evaluator to execute
this function elllclenUy.
evaluallon In sectlons 7.4 and 7.5. In seclIons 7.8 and 7.7 we tackle some

Inlerestlng examples using Our prlmll!ves. and ehow thai they allow efficIent

1I0lutlons 10 be .rlllen. In section 7.8 we repOrt the disadvantages 01 our

prlmillves. and In secUon 7.9 we draw Our conclusions.

7.2. I.AZ1 EVALUATiON

In this section we will explain 8 lew examples which show thai simple lazy

evaluatlon can sometimes be grossly Inelflclent. Flrsl 01 all. we cOf1slder

the 'unction that compules the length 01 a list. The simplest dellnltlon we

could use Is the IOllowjng:

length [J - 0
length (a:x) • 1 + length x

Although we WOUI~ expect 10 computa Ihe length of a IIsl In constant space.

Ihls formulallon will Clearly require space proporllonal 10 'he lenglh 01 lis
argument since. for ellample. leng1h 11.2.31 Is reduced lIa lollows:

length (1,2,3) rlld 1 + length (2,31

red 1 • (1 • length [3])
rod 1 • (1 • (1' length [ll)
red 1 + (l + (l + O})

red 3

The chain of addWons which Is buill up Is nOl reducible unlll II Is complele.

When the last element of the list Is scanned by length. then the enllre chilin

must be present In store and so will consume space proportional to the

length of the argument.

 60

One would expect thai the only space requ1rttd by tp 10 process a file would
be the space required to slore the curren I abbreviation. Alas. this Is nOI
10. Consider the InpUl 11,1 ·la)bcdelgh1jklmnop.. • Ip 0' this IIsl will be

reduced by " laly evaluator as foHows:

tp ·(.)be ... • 1 red tp (after .). -a)bc ...• )

(betore .). ·a)bc ... ·)
red tp ·be •.. • (before .). -.)be .. 0·)
red -b-:tp ·cd ••• • (before . ) .
-.)be .. 0·)

So. since the value of the abbreviation Is not reqUired unlll en exclamation
mark occurs In the Inpul. It will aClually be stored In the unevalualed lorm

(balora "). ·albe....). It will retaIn II pOinter 10 almost all the Input lIIe.

preventing II trom being reclaimed by the garbage collectOr. Thus Ip. like
lenglh. mey reqUire an .rbllr.ry amount 01 space to run In.

In bOth these ceses a large expreSSion which reduces 10 a small value Is

Ielt unevaluated far • long time. representing a gross wasle 01 storage. In
ordar 10 lorce eerller evaluation we Inlroduce an optional slrlCI luncllon call.
Wa define a comtiinator VAL by

VAL ! x red t x

but we Insl" lhal II Is evaluated t1elore VAL Is applied. We will usually use

an 'nllx notation. writing I «".1 x) Ins lead ot VAL I ll. and we will leel free
(0 use the nOlation (yal E) In any conle~ where E Is e 'unCtion ergument,

whether Ihls 15 Implicit or ellpllcll. For ellample. we will hepplly write lei
x • .,al E In... 10 Oenote e strict dectarellon. and we will write la. yal b,
cl to denote that b Is evaluated before the luple Is constructed. While a
and c are left unevaluated until required.

Now we can write

length' (a:x) n - length' x (val n+l)

10 fOrce evaluallon of length to be In Ihe elllcleni order.

We can use VAL to define a "sequential evaluellon" operator WhIch evaluates

lis flrsl argument and then relurns Its second:

a ; b II (lX. b) val a

.nd we can dellne olher functions which control evalueUon order, for eJlample

force -- an IdentIty lunctlon on lists that relurns Its result only after all
the elements 01 the list heve been BV8lueled.

torce 1 - ca.. 1 at
[} --+ IJ
(a:b) -+ a ; torce b

esac ; 1

Finally we can make tp store abbreviations elllclenll~ by writing

tp (-(-:x) ab - tp (atter .). x)

(val force (before . ) . x»

 ·2

All lheea .....Iuatlon conlrol methods ere In usa In the lIspkll 80ltwar8 tieIng

.rlhln by Henderson. Jonel and Jones IHenderson83J.

In Ihl' SBction we exhlblled two funcllons which 8r8 evalualed Inefficiently

by III lazy evaluator and W8 showed how they can be made efflclenl using
VAL We regard this as III good solullon because It doas nol necessltale
chenglng program structure. II seems thal we can wrlle our programs wlthoul
consIdering executIon order 81 IIrst. and then 8nn0181e them with VAL and

otha, conlrol funellan! such as force to make Iham elflclent. UnfOrlunalely

maner. are nol qulle thai sImple. 85 we will seB In Ihe nex1 section.

7.5. THE NEED FOR PARAUEUSM

In lhls section we will consIder III very simple paning problem. We wanl

• l~nctlon split which lekes II list 01 characters and relurns a pelr whos(

flr.1 componenl Is the IIrst Une 01 characters. and whose second component

I, the rest 01 the argumenl list. split could be defined by:

split I • [beCole n I l . aCtel nl IJ

where nl Is the newline character. Na1Urally. we would like spill 10 run In

cons tent space II possible. For example. a program lIke

program 1 • let [a, bJ - split 1 1n

{length a. length b)

(8n Obviously be executed In constant space.

However. the ,'mple dellnltlon gIven above does nOl have this property The
reason for this Is (Informally) thai bolh components of the resUI1 Of SP~II

contain pointers 10 lhe entire Input list. and so whichever Is evaluated IIrst.
the other will prevent garbage collection of any of the Input In the simple
program above this will lead to al least the firS! Une 01 Input beIng present
In memory In Its enUrely. SInce the IIrst line may be arbItrarily long. this
Is an Intolerable OVErrhead. Ot mIght ,eErm that unnecessary buffering 01 one

line Is InsignifIcant. II should be remembered that this Is an extremely simple

exam pte. and that the ,arne behaviour can also arIse In rnuch rnore ,evere
forms. FOr examplEr. reading thEr first illEr olf a magnErllc tape Is e~aclty

.. nalogous to the case we arEr dIscussing. but the unnecessary bulferlng

0' an enUre lIIe Is much more serious).

We might try to address this problem In the same way a. we solved the
problems In the last section. by rewrlung split In .. different way and using
VAL to control the execuUon order, However, this example does nOl yield
to thl, treatment. We will prove Informally thai no sErquentlel evaluetor can
eJ(ecule any verslan of ,pllt elllclently.

First 01 all we cladly Ihe noUon of a ·,equentlal evalualor". By this we mean

that. once the evaluator has begun 10 reduce en expression E, II will only

reduca E and olher expressions 'hal E demands until E has been cornpletely
reduced.

Now. we assume I"a' s Is a vErrslon of ,pili which will run In constent ,pace.
provided It Is used In the right context. We Irnaglne thet e Is applied to
the InUnlte Ust 01 characters from Ihe keyboard. and thaI characters are

typed relallvely slowly. We consider the second component of s's result Ihe'
pari of the Input IIsl atter Ihe flrst newline. This Is a pOinter Inlo Ihe Input
 84

II., etler Ihe IIrsl newline characler, and so C8nnOI poSSibly be com puled

unlll an the IIrst line has been typed, Therefore. If the conleX1 demands

Iht.. value before II has consumed all the Itrst line. then II 'Will be 5uspended

until lhl Ilrst newline Is typed. and arbitrarily much Of lhe line may need

10 be 'bun,red up' 10 be consumed leter. We assume, theretore. that Ihe

coniallt consumes all 01 the rtrst line belore demanding Ihe rest of Ihe InpUI

lei K be Ihe maximum number 01 characters of Inpul which are held In

memory ,Imultaneously. We tlnow thai K exists. since otherwise Ihe program

woule reqUire mOre than constant space. Lei I be Ihe Inpul list. and lei

I" be lhe part of Ihe Input list aller Ihe IIrsl n characters. Since is must

retu,n Its two fasulls to Ihe conlexl belore U"le context can consume any

characters al all. we know that s rellJrns belore K+l characters have been

typed. Therelore the expreSSIon E denoting the "rest of the Input- must be

ere_led Defore K+ 1 Characters have been typed_ SInce E evaluates to a

pointer Inlo Ihe Inpul list Us unevaluated lorm must also coniain a pointer

Inio the Input list ThIs unevatuated form Is created before K+l characters

havi been typed, end '0 the pointer musl point before 'to. If this pointer

remains unchenged throughout the consumption 0' the 'Irst line Ihen It wllt

cause mosl of the Inpul to be relelned In memory. consuming en arbItrary

am..unt of space.

We must therelore assume Ihal lhe relerence from E 10 lie Is via some other

ellpresslon e', WhiCh Is demanded, and so reduced. by the consumpllon

01 (he flrs( line. This reduction musl eliminate Ihe pointer to Ito. We have

alreedy ahown that the consumption 01 the first line cannot demand E. so

ttlerefote e and E' are not equal. We must have InSl8ad

E - P E'

'or aome Junellon F. In order to sells1y our assump1l0n eboul K, E' must
be reduced aftel" al most 2K charaClers have been typed, and consequenlly

the result 01 reducing e' must contain a poInter Inlo the Input Itsf belore

I., Once agaIn. thIs pointer. II relalned throughout the consumption 01 Ihe

IIrsl line. would leed 10 an arbUrary amounl of Input being bullered In Sl{)re.

end so II must be via some third expression e" which Is demanded by

the consumpllon 01 the first line, We must have

E' - F' E"

By continuing Ihls argumenl. II lollows Ihat. If the firS! line Is n characters

long. a chain of el least n/K expressions will be buill up. Since none of

lhe expressions ere equal. 8t leasl n/K cells will be used. and 8

non-constant amount of space will be consumed This proves our 8ss8r1l0n:

no sequential evaluator can execute any versIon of SpIlt In con slant sp8ce.

Lei us consider the significance 01 this result. We have demonstrated only

that one parllcular usefUl function cannot be run elllclenlly on sequential

absl(ect machines: however. II Is Intuitively clear that a similar prOblem 'Irises

whenever there Is more than one ·Con,umer· lor a value: If the sequenllal

nalure 01 the abstract machine lorces one consumer to run first then the

other will retain the value and delay garbage collecllon. (In the case 01

spHI. thare are two consumers. They are Us IwO results>. Even " II were

possible 10 combine Ihe IwO consumers Inlo one Ihls would be a bad solullon

since It would destroy Ihe modularlly of the program_ We have shown l.hal

even Ihls bad solution is nOI pOSSible In general on a sequenllal machIne.

We are IherelOl"e compelled 10 Introduce parallelism Into our abstrscl

machine.
 86

Having done so the prOblem Is nol yet solved: we musl address lhe problem

of scheduling ,aduellons so as 10 minImIse the space used. We can take

two approaChes here: eUher scheduling Is ImpllcUly performed by the

Implementallon. or II Is expllcilly controlled by the programmer. Some
progrus has been made with the lormer approach. Wadler has shown In

[Wadlllr8:l1 thai If 8 program can be 8lC8cuied In consleni space then all

scheclullng can be done al compile time. and he has given an algorithm

lor Doing so. However. he has also shown that If 8 program cannot be

execuled In conslant space. then the problem 01 schedulIng liS execution

to minimise the space u.sed Is NP-complelB IWadler831. Although tmptlcU

SCheduling seems the more desirable solullon In Ihe long run. we consider
lhal Lhese dllilculiles malle It Impracllcal al pre5enl. We prefer 10 look lor

51mple ~ays In whIch the programmer can control 5chedullng e.pllcllly.

7.... PAR FOR MORE PARAlLELISM

We begin by prO~ldlng an elCpllclI mechanism for slarllng perallel e~aluat\on.

By analogy wUh the VAL IlJncllOn which we IntrodlJced In sec lion 7.2. we

denne a combinalor PAR by

PAR t x :red t x

wllh Ihe addillonal property that • Slarl5 e.eclJling In parallel before the

reduction Is performed Since PAR d085 nol walt for • to finish II Is trUly

11\ ICienUty function (VAL 15 not because II Is slrlCl In .l, We will use en

Il'Ifi. not.lIon for PAR and detlne control funCllons lor use wIth II. IUSl 85

we did lor VAL.

First *e note 1hat the spilt prOblem can be solved by dellnlng spill by

split 1 • [before n I l . par after n I l )

so that. as the conlelct consumes the flrSI line. the second component Is
al.o ell:ecut.lng and s*allo*s characters as they arrive. (par II, the Inflll:
version of PAR. Ind the notation In this enmple means lhat Cafter nl n
starts ell:ecullng Immediately the resull or split Is conSlructed).

Now. " *e dellne

a lib. (~x. b) par a

whiCh sfarts a execullng In parallel and relurns b. then *e can 'IlIrlte

par 1 1st 1 - caBe 1 of

[1 ~ [J
(a:b) ---to a II parllet b
eBac II 1

parllsl I relurns I. bUI 51arts parallej evaluallon 01 all I's components. We

could use parllst In tha dallnUJon or spill thus:

spIlt 1 • (par parlht (before n I l ) .

par atter n I l )

....hlch 15 betler than the former In lhal. even \I the contell:l does not refer
10 the first line until long aller Il has consumed the rest of the Input the
 BB

first Uti. will be compu1ed and will nOI relaln II pointer 10 the entire Input

list. So Ihe programmer Is able to ensure thai spill runs efllclenlly by

.nnOI.,lng II wllh par and control 'uncllons.

7.5. SYNCH FOA lESS PARAUELISM

We hlye shown hoW' the programmer can deliberately Introduce parallelism

10 make split run elllelsnlly when applied to the keyboard, but we haye so

fer O't'8rlooked the need 10 reduce parallelism by synchronisatIon. To see

Why lhls can be naeenary. reconsider Our last definition of spill:

split 1 - [p.r parlist (before nIl) ,

par after nl 1J

It spill I, not epplled 10 the keyboard. bul to some computed 1151. then there

Is • dlnger th81 the expreSSion <par stier nr n will cause I to be eyaluated

'.,ter than !he contelC1 Is able 10 consume the Ilrsl line, This may result

In much 01 the IIrsl line being buffered while the context cetches up. Wa

neta to synchronise the consumption of I In Ihe two expreSSions.

We recall at this llage that "lazy evaluation" sCheduling 0' graph reduction

IS .chleved by propagallon of demand. OrigInally the result of the whole

program Is demanded. end there alter demand Is propagated by strict

operations through the graph. Only paris 01 the graph at which demand has

.rllyed are reduced. PAR modifies the propagation of demand by propagallng

It 10 two nodes slmulteneously. We now need 10 synchronise executIon by

restraining the propagation of demand.

 89

We Introduce a new function SYNCH 10 aChieve this. SYNCH I~ defined by

SYNCH e - [e, e]

However, the two caple, 01 e whIch Bre relurned arB actually dlllerenl' call
Ihem B.l and 81. No demand Is propagated from 81 or 81 10 e unlll both

have been demanded. Thus. II we write

let [a, hJ .. Si'NCH (1 + 2) in

{par factorial a, pA~ fibonacci b)

we cen be sure thel 1 and 2 will not be added unlll both (ac10rlsl and

lIbonaccl are ready 10 USB the answer, SYNCH I, actually ,. dangerous

funCUOn In that II can cause deadlock II onB 01 81 and 81 Is never

demanded. Nevertheless. It Is an ImpOrlant control mechanism.

In the spill example. we need 10 ensure that two separate processes consume
• list at the sarna rate. We will define a funcllon SYNCHUST thet takes a

lis: Cind relurns two versions of It. In such a way thai both versions must

be consumed 81 the same rate.

SYNCHL I ST 1 ..
let [sl, e2J - S~NCHLIST (tail 1) in
let [11, 12J - CAse 1 of
[J --> [[I, [I)
(x:I') --i' [x:eI, x:s2)
eSAc in

let [wI. w21 - S~NCH 1 in

[wI: 11, w2 ; 12)
 90

SYNCHUSI works 85 follows' when It Is ftrst called none of the declara!lons

lira t!lvelusled 8nd It relurn, lis two results Immedlalely. Evenlually. demand

will arrive 8t one of them, say (w1; In. and so wl will be demanded. Since

.1 Is 8 result from SYNCH the computation wUl b8 suspended 81 this point

Leier another parallsl prOcess will demand (w2; 12J. and so demand will
arrive 8t w2. Now the conditions lor SYNCH 10 propagate demand are
satisfied. and so SYNCH's argumenl I will be evaluated and returned 8s

the value 01 wl and w2. ThIs wi" allow 11 snd 12 10 begin 8.lCeCUllng. and

so 8ach result 01 SYNCHUST will be computed. giving )(:<another synChronised

IISI). nLlS. as required. SYNCHlIST conSlralns the consumers of Its two

resulb 10 work 81 Ihe same rate.

Now If we define

split 1 - let [11. 12] - SYNCHLIST 1 1n

[before 01 11. par after nl 12]

Ihen we ere essured thet characlers will be consumed by both processes

as tney lire <lemended by the consumar. In thIs case It would be

Inepproprlate 10 epply parllSI 10 Ihe IIrs! result since we wenl evaluation

to be driven by Ihe consumer 01 Ihe firSI line. (Technical pOint: for lhe

putpases of SYNCH we regard demand 10 have arrived at an e.llpresslon

when 11 Is reclaimed by the garbage collector Wlthoul Ihls the e.llample above

would deadlock el the end 01 Ihe IIrsl line because no more demands woul.d

IIrrlve al 11).

So by IInnOlatlng split with e.pllcll perallellsm and synChronisation funcllons

we can make It run In constant space In a variety of conte.ll!s. This enno\aUon

<lOIS nol represent a Change In the structure of the program. and so we

 91

really can write Ihe prognm first without regard 10 e..,aluallon order. and
annoIale II alterwards. However. we musl leke InlO account the Intended use
of /I function when we annolate It and we may need dltlersnUy annotsted
"s"lons for use In dlflerenl conlexts. In the remainder of Ihls chapter we

ifill gl...e a lew more examples 01 the application of PAR and SYNCH.

7.e. QUICKSORT

Quicksort can be expressed yery elegantly In 1:1 funclional lenOL/age. II can

be dellned by the equallons

aot:t l] - II
80:rt (a:xl - BO:rt (b ....... ; b<a) ++
(al ++
so:rt (b4-tl:; b)a)

using Turner's ZF nolaUon [Turner811 Ho.....; b(8) means Ihe list of elements
b of lI: whIch ere less than s). The ImperaU..,e Quicksort sorts 8 IIsl of n

elements In space OCn) and time 0(n2) In Ihe worst case. Olnlogo) on
a"'elage. We shall examine the complexlly of the functional ....,slon.

Fil,t of all we consider the lime laken 10 sort e list 01 n elements. Tin).
On average. Ihe IW'O recursi...e calls to sort will be on IIsls 01 lenglh nl2.
and ~o we ha... e

T(O) - a constant
T(n) - 2T(n/2) + the rBet of the coet
 92

Since the Test of Ihe Cosi 01 a calt of sari Is proporllonal '0 the length

of lhe aroument, we have

T(n) - 2Tln/2) + O(n)

WhiCh has lhe saluUon

T(n) • 0(n10gn)

However. In the worst case Ihe Iisl will be spill InlO an empty list and 8

list of length n-1. and 80 we will have

T(n) - T(n-l) + O(n)

with the solution

T(n) - 0(n 1 )

So Wf obeerve with salls faction thai the functlona' version of QuIcksort has

Ihe eame Ume complexIty 85 Ihe lmperallv. one.

How. lei us consider the space complexUy. SCn). Clsarly SCn) ( Ten). We
Consider the amount 01 'pace In use Just aller computing the lirst element

of Ihe ,orled list. Since this element could origInally be anywhere In the

.rgument to sort II Is clear thai sort must force the complete eveluatlon

01 lis argument belore II can compute lhls element. So. considering the

_econd equellon lor ,art. we 'ee that by thl, stage the expression (b .........;
b<1I mu,t be completely evaluated. and '0 wIll not ,hare any cell, wIth

the orlglnel argumenl. On a lazy evaluator. the second recursive call 0' sort.
sari Ib.f--tl:; b~a) wUl nClt have been evaluated at aiL and so will contain a

relerence 10 the original argument x Therefore Ihe space In use will be

approximately the space required lor x plus the space required 10 sort tb~.

b(a} In Ihe average case, then.

5(0) • n + 5(n/2)

and :<;0

Sen) • O(n)

In Ihe worst case, however,

5(0) • n + S(n-l}

and so

5(n) ·O(n 2 )

So Ihe worsl case space complexity 01 the luncllonal Quicksort Is as bad

as lis lIme compleldty, lar worse than the Imperative equivalent Even In

casas thai deviate slightly Irom lhe average. this means a worse space
behavIour,

Of course. Ihls Is JUSl the kind 01 prOblem Ihat our primitives were lnlended

10 solve. We shalt use them to lorce bolh recursive sorts 10 consume the

argument IIsl at lhe same rale, This can be done by definIng

lSort (a:x) .. let [xl, xZ) .. SYNCHLIST x 1n

par lSort {b........:l; b(a} ++
pa..r [a] ++
pa..r sort (b........:Z; b)a)
 .4

To find lhe apace complewlty 01 this funcllon, we observe that. since access

10 Ihe argument list Is synchronised. II Is consumed as II Is computed. No

storage Is wasted In buUsrlng. Therefore. each recursive Invacsllon of sari
requires only 8 constant amount of space. Also. the number of parallel
InvOC8tlo~s of sort Is bounded by twice the number 01 elements to be sorled.

since an element Is consumed when two parallel InvQcalions 8re Sl8rted.

" follows thai this sort 'unction requires only O(n) space 10 sort n elements.

as gOOd 8 rasul1 8S the Imperative one. Moreover. on a machine wllh enough

parallelism. Ihls version of sOrt will sari n etements In O(n) time using (on

average) O(logn) processors.

ThiS erample shows lhat our prlmlllves can be used not only 10 make cenaln

programs run In constant space. but can also prOvide substantial

Impr0¥8menls In Ihe space complexity 01 programs Ihal do not.

7.7. PIPES

The UNIX operating system provides a lacUlty whereby two (or more)

pr~lams may be run In parallel connected by 8 "pIpe". so thai one program

receives as Input Ihe output of the other. This cen be a convenient way

of 'll"rlling multi-pass compilers. lor exam pte. so lhal each pass receives
the OUlPUt of the prevIous one. and all passes can run In parallel and exploit

Ihe capabilities 01 mUl1i-processor machines. The programs are loosely

synchronised In that each prOducer may run ahead of Its consumer. bUl

on~ by 8 IImlled amounl.

Pi~es are naturally Incorporated Inlo a functional operating system. Assuming

1hal programs are luncllons from a IIsl 01 Inpuls 10 a list 01 outputs. we

may connect two programs using the luncUon

connect prog1 prog2 input - prog1 (prog2 input)

which Just compotes the lwo programs. Running (connect progl prog2)

elfectively fUns prog 1 and prog2 In parallel. connecled by a pipe. However.

In this case Ihe programs are synchronised ellaclly. since prog2 will only

run when prog 1 demands an InplJl. This Is undesirable If the underlying

machine Is capable 01 real parallelism. because II could lead to processOrs

slanding idle while Ihere Is real work 10 be done. We can easily correcl

Ihls by defining

connect prog1 prog2 input ­

prog1 par pipe N (prog2 input)
tot ... some number

using the lunctlon pIpe which takes a number and a !lst and relurns a copy

01 the list. but starts Ihe evaluation 01 the N'4-I'th list elemenl as soon as

the l'th Is demanded. pipe Is defined by

pipe n 1 .. par1ist (take n 1) It

pipe' (drop n 1) 1
pipe' [} 1 "" 1
pipe' (a:f) (b:1) - a II (b:pipe' [ 1)

Now when (connect prog 1 prog21 is run. prog2 Is started Immedlalely and

Its IIrst N outputs are demanded Therealter. whenever progl consumes one

01 lis Inpuls. another output of prog2 starts to be compuled. prog' and prog2

tun in parallel ana prog2 Is allowed to get up to N elemenls ahead. This

Is exactly the behaviour 01 two programs connected by a pipe In UNIX

 96

The method we have used here Is actually more general than lhe UNIX pIpe.

Allhough we have applied II 10 lists. we could 85 88511y wrne III function

which works on trees. or any other dal.a-5tructure. We have presented III

gener., It'ay 01 starling the evalualJon 01 dat8 shortly before It Is required:

It can be used In any conlexl where this Is desIrable. f~r example to starl

disc translers shortly berore the data on the disc Is needed. or Just (0

Inere.5e the available parallelism In III controlled way.

7.8. DtSADYANTAGES

Since PAR and SYNCH are a control mechanism. nol a panacea. it Is

pOssIble 10 use them 10 cause undesirable consequences_ There are 1wO

main categories of such errors: deadlock and rampanl parallelism. Deadlock

occur! II SYNCH Is used wlthoul a corresponding PAR. For elCample. II ~ort

were defined by

Bart (a:x) - let [xl, x2) - SYNCHLIST x 1n

Bort (b~l; b(a) ++
[a] tt

Salt {b....-,r;2; b)aJ

then 11 would never produce any results because demand would never arrive

at 12. On Ihe olher hanlf. rampant parallelism occurs II PAR Is used too

ofl,n wlthoul corresponding SYNCHs. For eumple. If pipe were defined by

pipe n 1 - paIliBt 1

"'," II would try 10 compute all the outpul 01 Ih, prOducer process In

plr.Uel. possibly cloggIng lhe system.

 97

Thesa problems lire worrying because II good understanding 01 the

Implementation Is required 10 avoid them. The nalye user la unlikely to see

anything wrong with the erroneous dellnltlons In this section. In laci. II Is

necessary 10 know lriCky delalls 01 the Implementalton In order to predict

how some programs will behave For example.

let [a, b] - SYNCH e 1n II + b

will be 8'118lualed correctly If the Implementation e..,aluates the arguments

01 • In parallel. bul will deadlock If they are evaluated In seQuence. Another

example Is the funcllon ellists

exists p - p. \ function p
exists p - exiets (p true) or exists (p falee)

whICh lakes II prl!tdlc8te. II curried function expecting seyeral boolean 'falues

and returning 8 boolean. and returns true II lis argument Is not Identically

'alse, If the arguments 01 or are evaluated In sequence then eldsls pertorms

a space efficient sequentlal search. but If they are e",aluated In parallel then

rampant parallelism will result which will clog the system completely.

Therefore PAR and SYNCH. allhough flexible. requIre considerable expertise

In theIr use. This may slgnlly thai lhey should be replaced by a more

structured equlvalenl: bul al the moment we ere unable 10 suggest 'what

form It might 1ake.

 98

7.'. CONCLUSKJN

In thIs chapter we have demonstrated beyond doubt thai parallel abstract

machines are a prerequisite for elflclent Impiement8IJons of functional

languages. and that correct scheduling Is vllal II this efficiency Is 10 be

achieved. IT I:!I debalable which 15 the besl way 10 ensure correct sCheduling.

but we lavour expltcH conlrol by the programmer USing simple prlmllives

We hays defined two such primitives and wOrked enough 8.ll8mples 10

engender confidence thai they are sulflclenlly powerfUl lor most pr8cllcel

problema. Our prImitives have the ImpOrlant advantage Ihal they are

Slructure-Independent. so that the programmer does nOI have to take them

Inlo consldarallon when first deslgntng his program.

 '9

CHAPTER 8

GARBAGE COLLECTiON

8.1. WTROOUCTION

We have argued In Chapter 2 thai on8 of the Important advantages of

functional languages Is the provision of a garbage collecla,. FOllowing Dennis

IOe'1'11581) we 81'0 believe that the besl way to provide 1118slor8, and

detabases Is within a very large garbage-cOllected vlrtua' memory. 111 Ihls

chllpter W8 concarn oursalv8!1 with garbage collection methods slJltabh!t 'or
'unctlonal languages running In very larg8 virtual memories.

In secllon 8.2 we will discuss the general siralegies available and argue
thai feference counung Is the most promising lor our needs. In seellOl') 8.3
we will 81lamlne how circular structures, the bane of reference counllnl1

ga,bege collectors, arIse. In sectiOns 8.4 and 8,5 we will p'esenl our
eJltenslon 10 reference counting fo, managIng circular structures. flnall~, In

5ectlon 8.e we will examine the costs or our method, In section 8.7 we will

refer briefly to Brownbrldge's method. and In section 8.8 we will presenl

our conclusIons.
 100

1.2. GARBAGE COLLECTION STRATEOIES

There 119 two main 51r81egles lor deleting objects no longer required. which

we shall reler to as mark-scan garbage collection and relerence counllng

garbage collection. Mark-scan algorllhms delermlne which objects can be

deletea by vis lUng all obJec1s accessible trom any machine regIster. and
lhen delellng the others. Relerence counllng algorithms siore a count of

the number 01 references 10 an object wHh the object and delele 'he oblect

when Ihis count becomes zero. Both categories cover III wide varlel"/ 01

algorllhms. and. In parltcuhH. eUher strategy may be used In parallel wllh

Ihe mlln compuilillon IDII!l;slra791 (Hudak82l rOTt1811.

In 11'1, conlext we hllve described. however, reference countIng seems to

have a marked advantage Flrslty vlslUng every accessible Object as

mark-scan algorllhms do. Is very ellpenslve when there are a very large

number of objects. Our vary large virtual memory will conlaln a very large

number of accessible objects, Relerence counllng algorllhms. on 1he 01her

hand. visit only objecls as Ihey are being processed.

Secondly. merk-scan algorllhms may nol delele Inaccessible obJecls until

long alter they become Inaccessible. while reference counllng algorllhms.

delelo them Immediately This Is partlCulerly Imp0rlanl In a virtual memory

Syslem, becauso il means lhal. wilh reference counting. shari-lived obJecls

can all share the same locations. These locallons wIll therefore form part

Of the compulallon's working set. Since appllcallve language execullon

generates a very large number 01 short-lived oblects. we do nol believe

'hit It very large vlrlual memory Is vIable il Ihey are not deleted promplly
 101

Thirdly. mark-Scan algorithms operate In several phases. and some kind

01 synchronlsaliOn Is necessary when the phase changes. This Is. 81 Ihe

least. an Inconvenience. In conlrast. reference counllng Involves only vsry
local Information and requires no global synchronlsallon.

For Ihese reasons we CIa nol believe Ihal mark-scan garbage collection Is
viable In largo Ylrtual memory parallel appllcaUve systems. However.

reference counllng also suliers hom a serious disadvantage - objects which

are accessible Irom themselves. and so form pari oj cIrcular Slructures.

are nOI deleted at aiL Since even (he Ubiquitous recursive 'unelian is usually
represented by a circular struclure. 11'1115 presents a very serious problem

which could negate all the other advantages of reference counting.

8.3. CIRCULARITY

It Is worth examinIng how circular struclUres arise during execution. Some

are created becausa Ihe programmer has dellberalely dellned a circular

data-slructure. 'or example by

let x - cons (1, x)

The majority are the representations of recursIve functions. For exeomple.

Ihe factorial lunctlon might be represenled by

~n . if n - a then 1
else n * fact (n-1)
vhere fact ­

where the circular pOinter Is shown by the arrow.

 102

II has Ileen suggested thai II Is not necessary 10 use circular structures

Olher thin recursive lunctlons (which are vUat since naralla" Is expressed

8a recursion In appllcatlve languages!. Dennis has proposed 8 non-circular

represerllallon tor recursive lunctlons In IOenn1982J. end Friedman and Wise

have oherved In IFriedman791 (hal reference counting will work for recursIve

functions On their representation) 11 the circular pointer Is nol counted.

However. neither 01 Ihese schemes allows the conslruclion 01 genuinely

circular (11118-slruclUres. Our own 8lCparlence 01 using KRC lTurner811 which

does ~ol allow circular structures 10 be constructed al run-time. suggests

thai ,~ey are almosl always unnecessary: but occasionally a program cannOI

be wIllIen elllclenlly and conveniently without them.

An 81ample 01 such a program Is a p'-ogrammlng language Interpreter. The

p,-oblem here Is rhat, since luncllons In the language being Inlerpreted must

be represenled by dala-structures In Ihe underlying language. recursive

'unCllons would nlliurally be ,-epresenled by circular dala-slructures. We have

encountered slmller problems In wrUlng II synchronous process simulator

and In the IIlgorlthm for compllallon 10 super-comblnators given In section

5.04. In these cases il seems 10 be ellher very dlfflcul1 or extremely Inefflclenl

to program withouT circulllr slrUClures.

Bobrow has described IBobrowBOJ a method thai manages circular

daliJ-structuros uSing inlormatlon supplied by the programmer. His met~od

hal some strong sImilarities to our own: however. Changes 10 Ihe

daill-structures can Invalldale Ihe InformatIon the programmer supplied.

le.dlng 10 delay or outright failure In reclaimIng some oblects. For this

rBlISon Its applicability Is limited

 103

We believe thaI the extension of relerence counllng 10 deel with circular

structures In all their generality 15 essential.

8.... THE STATIC PAQORAM ORAPH

We restrict ourselves to super-combinator Implementations of functional

languagn. and lor Ihe lime being we assume thai the graphical comblnators

Introduced In section •. 5 are nOI used. We begin our extension of reference

counting by consIdering a snapshot of the progrem graph at 8 particular

In"lanl. and looking for 8 way 10 describe Its circulatilles. We recall two

detlnllIons Irom graph theory.

Definition: A greph Is "'tongly connected II, for any two nodes A and B.

there Is 8 path from A 10 B and vice versa. For example. the lell hand

graph below Is strongly connected. but the right hand graph 15 nol.

6 I\~ ... • r -------.

 104

Definition: A IJtronQJy connected compon."t 01 8 grl!lph Is 8 ma.lmal strongly

connected sUbgraph.

tI tollows from Ihese definitions Ihat any graph elln be decomposed Into
disJoint ,trongly connected componenls The !lgure below shOWS a sample

Df.ph Ind lis decomposition.

Th' 'Irongly connected components 01 8 program graph are Ihe unlls of

circularity; for, II one node In a component Is accessible. then the whole

component I. accessible; conversely. It one nCX1e In 8 component cen be

Cl8kled. then ell Ihe nodes In Ihe component can. II 101l0W5 thai It Is more

.pproprl8le 10 consIder garbage collecllon of strongly connected camponanis

Ih.1'I 01 Indlvldul!ll nodes.
To this end we daflne the derived Qraph G' 0' II graptl 0 10 be (he result

0' coalescing all nodes In Ihe same strongly connected component. More
precisely, the nodes of 0' are the strongly connected components 01 1],

end there Is an edge 'rom node A to B 01 G' II there Is an edge In I]

'rom a node In A 10 a node In B. The nexl diagram shows an example

grapn 8nd 115 derived graph.

The derived graph Is always acyclic. We can see thts because. If It _were

not then there would be two distinct strongly connected components A and

B with paths both hom A 10 B and tram B 10 A. Bul this means thaI. In

the original graph. (here must have been paths from a node In A to a nOd~

In B and vice versa. Because A and Bare bOlh strongly connected. lhls
 106

Implies lh!lt there must have been 8 path Irom any node In A 10 any node

In B end vice vers!!; so A !Ind B would both be pari 01 the S8me strongly

connected component This vlolal8s our assumpllon thai A and B were

dlslolnl.

Since the derived graph Is acyclic. II can be garbage collected safely using

re'erence counllng. ThIs mollvates us to modify the program graph by adding

• field 10 each node which points to 8 shared reference count. All the nodes

In the same Sirongly connected component point to the same shared

reference counl, which conlalns the number or references 10 the whole

component Irom other components. An ellsmple 01 8 program graph wllh

IIhared reference counts drawn as boxed numbers Is

[D .

I!J
·······lIJ

CONS 1
 107

We can now delete a whole strongly connected component when the shared

reference count becomes zero. We can also perform a lasl tast to discover
whether two nodes are In the same strongly connected componen!. by

comparing their shared reference count pointers. ThIs will be Imporlanllater.

As well as (he shared reference counls. W8 relaln the reference counts of

each Indhlldual node. whIch we refer to as Ihe local reference counts.

8.5. lHE GRAPH IN MOTION

We can now use reference counting to reclaim circular SlruclUres provIded

we can keep the shared reference count structure up 10 dele. In this secllon

we show how the machine can update the ,hared reference counts during
reduction.

Since our machine language Is appllcallye. all the reducllon rules obey e

very Imporlent property: all the nodes accessible from the root ot a reduction

after the reduction are either newly claImed. or were already accessIble
before the reduclion. Because of thIs. two dlsllnct strongly connected
components will always remain so. slnca no reducllon can make one
accessible tram the other If /1 was not already so. The only ways In whiCh
a strongly conneCled componenl can change are by growing. /e having newly
created nodes added 10 It. or by splitting InlO several smaller ones.

For Iho purposes of expostlion. we consider each reduction 10 happen 'In

two slages. In the IIrsl slage. new cells are claimed and pOinters 10 Ihe

'esult are added 10 the node being reduced. This may cause slrongly

connected components to grow. but cannol cause them 10 split. In the

second stage. the old polnlers from the rool are deleted. This may cause
 108

Slrongt, connected components 10 split. but cannot cause them 10 grow.

In be~een Ihe two stages there will be more than IwO pointers Irom the

root This will C8use no p'0blems In the Implementation because some 01

Ihe pointers will actually be held In registers.

1.5.1. ADDINO NODES AND POINTERS

When adc:Ung nodes and poInters 10 the program graph we must decIde which

,1 rang')' connected component the new nodes belong 10. Having decIded

this Irlo local and shared reference counts may be adjusted according I),.

We Il'iU51'818 how this may be done using the combInator S. whose elise I

Ie aMwn below.

s A
Arter Ihe first stage In Ihe application 01 S. Ihe graph will appear as shown
below. wllh foul'" pointers from the rool. T"e nodes marked • are new and

must be allocaled 10 a strongly connected component.

We nollce I"e following property of strongly connected components: If a node

Is pari of a non-trivial component (one conslsllng of more than one node)
then al least one of \he nodes poinllng to It Is In lI'le same component.

Since the starred nodes form a tree. t"ere Is only one node polntln·g at
each one: the node directly above It In the tree. II follows thai If any slarred
node Is In a non-trivIal component 1hen the node dlreclly above It 15 In
 110

the S8rrte component. and. ultimately. so Is the root. Therelore. every slarred

node ,lltler forms 8 trivial strongly connected component by Itself. or Is

In the same component as the rool.

In order 10 decide Which 01 these cases applies. we noUce another properly

of strongly connected components: If 8 node Is part 01 a non-Irlvlal

component then al leBst one 01 the nodes II points 10 Is part or the same

component. A starred node In 8 non-trivial component must therelore point

elthel to an argument 01 the reducUon In the same component. or 10 another

slamd component wllh the same properly. II lallows Ihal all starred nodes

In non-trivial components lie on paths from the root to an argument In the

same component.

Con'lersely. any node on a path bSlween two nodes In the same component
Is also In that component. We may lherelore allocate new nodes 10
components as follows: when a redUCtion Is performed lhen any new nodes
on I palh from lhe root to an argument In lhe same strongly conneCled

component as the toot should be added to Ihat component. Others should

lorm new, trivial componenls by themselves. ThIs Is a simple and complete
rute lor keeping our data-structures correct

Even lhe '( opera lor Is covered by this rule. It adds no new nodes 10 the

graph. but adds a circular poInter. This pointer cannot aller lhe strongly
connected component struClure or the graph: II only atlers the local reference

count 01 the rool.

 111

········GJ CD
=

x F

8.5.2. DELETING POINTERS

When a pointe,. Is deleted there are two casas to consider. If the pointer
connects nodes In two dUlerent strongly connected components then It Is
only necessary to update the local l!IInd shared reference counts 01 Ihe lergs!.

and delel8 the large I II necessary. All Ihe nodes at 8 strongly connected
component can be deleted at once because they BrB all accessible from
any node In Ihe component.

II the polnler connecls two nodes In Ihe same slrongly connecled componen!.
Ihen II Is possible that Ihe component may split Into many smaller ones.

For example. aelellng Ihe marked poinler In Ihe diagram below spillS Ihe

strongly connceled componenl up as shown.

 112

,"
, '
•••
(~''-'-'
I "
.......
I "
I
,I
I

.•'.'
I

•
We hive to find the strongly connected components of lhe component beIng

Spill. and mark them 8S new strongly connected components of the program

greph. There Is an elf/clenl algorithm due 10 Tarl8n (18118n721 for finding

Ihe ~lrongly connected components or 8 graph In one scan over It. We can

apply this elgorllhm 10 the component being Split. and II will run quickly

beca~se we lUll only epplylng II 10 8 small pari at (he program graph. AI

the ~ame time we can compute the shared reference counts 01 the new

components by adding the local reference counlS of their consUluent nodes

logether and subtracting the number of Internal pointers. Any component

who!8 shared reference count Is zero Is deleted.

 113

IUS. EXANINAnON OF COSTS

ThiS melhod requires more storage Ihen IradUional relerence counting.

because each node mUSl have 8 shared reference count pointer slored In
It. Since, In general. I reference count requires about 85 many clls ItS 8

polnler. It would seem thi!ll adding e shared reference count polnler 10 8

node already containIng 8 local reference count. e head pointer. and I lell

polnler. represents about l!l 33'11:1 Increase In storage reqUirement tn practice

we expect the Increase to be smaller than thiS. because nodes wlll also
need space lor scheduling Inlormatlon and the like (nOdes In ALICE

IDartinglon811 occupy 8 lola I of 32 bytes).

very little extra siore need be used 10 hold the shared reference counts

themselves. II one 01 lhe nodes of each slrongly connecled component Is

distinguished In some wey. then the shared relerence count can be stored
In Its shared reference count pOinter lIeld. and Ihe other nodes In lhe
component cen pain! 10 It. This requires only one bit per node. UsIng this
scheme non-circular par1s 01 the greph. whIch consist 01 s!rangly connectad
components 01 one node each. can be stored compactly and have their
reference counlS updated quiCkly. ThIs Is very Importanl because the bulk
01 the graph Is of lhis lorm.

Our method is also slighlty slower than ordinary relerence counllng. pE!rtly

because 01 the extra cos I 01 updaUng shared re'erence counls. but mainly

because 01 the cosl 01 splllllng up slrongly connected components.

Fortunetely. we do nol believe slrongly Connected components will be spill

very olten. This Is because clrculer slructures ere usuelly buill and then

used several limes: lor example recursive functions are usually called many
 114

lime!. Since 5lrono'y connected componenis need to be split only when Ihey

change shape. lhl:s will happen whUe they lire beIng buill. bul nol while

they are In use. Therefore. for example, calling II recursive funcllon should
Involta no splitting.

The amounl 01 splllting could be reduced slill lurlher by using graphical

comtlinalors. 8S suggested In secllon 4.5. This would allow many CIrcular

slructures 10 be buill In one slep. rather than by applying Y followed by

'ree comblnators. and so would eliminate splilllng In these cases. Provided

Ihallhe combinator body contains Information on lis own sirongly connected

component structure. the structure 01 the result 01 applicatIon Is easy 10

delermlne. In II similar way 10 that used above.

OUt methotl shares Ihe advantages of ordinary reference countIng. namely

thai reference counllng acllvlUes can be tlone In parallel wllh Ihe main
computation. anti that only local Inlormatlon Is requlretl. We lherefore expect

It to be suitable lor vlrlual memory apptlcatlve systems. anti reasonably

elflclenl.

8.1. BAOWNBAIDGE'S METHOD

Our method timers Irom the other approaChes we discussed in section 8.3

In being a complele solullon 10 Ihe garbage collecllon prOblem tor gr~ph

reduction.> Implementallons: thai Is. no additional resirlcilons need be placed

on the appllcatlve programmer 10 ensure thai garbage colfecllon works.

8rownbrldge will shortly publish an extension of relerence counting Ihat Is

 115

completely general IBrownbrldgeB31. His method dlsUngulShtU ·strong" 8nd

'wasko palnlars. and relerence counls the two kind, separately. When lhe

1851 suong pol nler 10 III node I, dele led lhe garbage collector aHer, lhe
,Ialus of enough other pointers to ensure thai lhe whole graph I! spanned

by an acyclic graph Of strong pointers. and deletes parts 01 lhe graph that

are no longer referenced. On occasion this process may need 10 scan a

large pari of the graph.

The greal advantage of Brownbrldge', melhod over our own Is thel It reqUires

no sssumplJons about Iha kind 01 usa 10 which II " put whereas our method
can only be used wJlh graph rsductlon Implementations Of appllcaUve

languages. However. since lhe two algorithms are so dUleren1. It Is very

dlHlcult to predict whIch Is the more elllclent lor eny particular use to which

they are both applicable. We are unable to compare them In any more derail

sInce we do nol understand Brownbrldge's algorithm yel.

8,8, CONCLUSION

We expect that computer syslems In Ihe future will be hIghly parallel. and

will have large ....Irtual memories. Garbage collection will be e55enllal. but

mark-scan garbage collection will be ImpracUcal In the past. relerence

counUng garbage collection has been unable 10 collect general circular

.truclures, allhough particular cases have been covered We belle....e that

such general circular slruclures are a ....lIal programming 1001. We have shown

how, al leasl In a graph-reductton machine executing appllcatlve program!,

relerence counting can be extended to handle any structure al all. Our

method Is slightly more costly than ordinary relerence counHng. but. we

believe, vastly Cheaper than excluding circular structures altogether.

 "6

CHAPTER g

RELATED WORK

•. 1, INTRODUCTION

In 1~ls chapter we survey other work In lhe area of the ellicleni

ImplementatIon 01 functional languages. The lIeld I, very active. thanks partly

10 Me software engIneering l!IcJvanlages 01 lunctlonal langueges. 8nd partly

to Ihe promise thai they can be used 10 exploit highly parallel arChitectures.

As • consequence. 81mOsi 811 possIble allenU9' are being Invesllgaled. In

aec~on 9.2 .e will descrIbe 1""0 string reduction architectures. In sec lion

g.a we .111 consider architectures thai perform graph reduction on the

original program. or some not-vary-heavily complied version or If. Section

9.4 describes SByeral approaches based on lhe SEeD machine. In secllon

9.5 we wIll dlscu!n attempts 10 e.-ploil Turner'S SKI approach lalrly directly.

end In section 9.6 we ""III describe lmplemenlallons basad on

sup,r-comblnal0rs or very similar Ideas. Finally. sec lion 9.7 concerns 11sell

_Ith de'eflow.

••1. STRING AEDUCTtON

In (his secllon _e will describe two string reduction a,.chl1ectures. one

dl!lIlgned by Berkllng and one by Mago.

 117

Berkllng's architecture (BerkllngJ I, designed for reducing l,-expresslons. It

consists 01 a processor and two stllcks. II lell slack and II right SlIlek Un

'acl U1e processor contains several olher ,lacks for holding expressions

temporarily). The mllchlne sl8rls wIth Iha Input expression In ellner Ihe left

or Ihe right slack and transIent II back and fOrlh between them performing

fJ reductions until II Is no longer reducible. The final expression Is then

outpul. A prototype machIne wes actually conslrucled.

Allhough Ihls Is perhaps Ine most nalural way 10 construct II reduCllon

machine. It ha, .a number 01 dl.S8dvenlages. Firstly. since no allempt Is made

to select appropriate expressIons for reducllon, Ihe machine Clin wilsie II

greal deal 01 lime on reductIons which lire nOI eClually necessary. This Is

compounded by 1he substltullon ot parllelly reduced IIrgumen1s Inlo funcUon

bodies: If the argumenl Is used more Ihan once then any subsequent

reductions ot It mUSI be duplicated. Secondly. operstlons on large

data-structures are expensive sInce the whole deta-structure must be

Icanned tor every operallon. However, the srchltecture can eeslly be

extended to exploit seversl processors. and Ihls mlghl psr!lally Outweigh lhe

dlsadvantsges.

Mago's architecture Is entirely dlrrerent. II IS desIgned to execute Beckus'

FP IBsckus7B) ralher Ihan Ihe 1,-calculus. The mschlne consIsts of a large

number 01 cells. 8I!Ich 01 one 01 IWO kinds. One kind 01 cell Is arrsnged

In a long linear sequence whiCh holds the symbOls or the FP progrsm bell)g

reduced. and Ihe Olher kind Is arranged ItS a binary tree connected to the

flrsl kind al the leaves.

 118

The machlna operates In cycles. during whIch 811 reducible expressions are
IOCllld, mIcrocode broadcast through the tree 10 the cells that need 11. emply

spac!t!l r8arranged 10 make room for expansion of the program. and finally.

rtle reductions per-lormed. 1"e machIne Is therelore able 10 reduce all

reducible tlJL:presslons in the program at the same time. provided there 15

Spice 10 hOld all Ihe resuns.

Mago's machine Circumvents the flrst flaw In Berklfng's. An expression Is

onl~ considered redUCible II all liS sub-expr8551ons arB completely reduced.

TI'111 ensures lhat no reducible 8:olpresslon Is ever copied. and so work Is

ne~8r duplicated. The prIce. howoyer, Is Ih81 all funcllons ere strict. and

so the programmer cannot teke advantage 01 tazy avalualion. Since we regard

lar, avaluallon as one 0' the mosl Important advlllrliages 01 lunctlonal

programming. we consider this a serious deflclency.

The second flaw In Berkllng's machine Is also present In Mago's: since

.r~umenjs .re copied from place 10 place, II Is very ell:penslve 10 manipulate'
large dala-structures. Mago has proposed a partial solution to Ihls problem
 "9

IMego81J. whereby the machIne can be made to leave 8 large data-structure

where It I, and moyo other arguments end results past II. This Is not 8

completely general solution. however. II remelns to be seen 10 whel axlant

the adyantages 01 parellellsm In this machine .,,111 be olfset by Ihe cosl 01

dal8 movement.

•. 3. GRAPH REDUCTION OF THE SOURCE

Several people have designed Implementations whIch (more or less) perform

graph reduction on the original source of the program. We say "more or

'ess' because many 01 lhesa Implementations perform some Irlvlel

compilation. but nOlhlng so rna/or 8S compiling to Turne,'s comblna1ors or

10 SEeD machine code.

Turner has used greph reducllon of the saurce In his KRC Implemenl8110n

ITurner811 This Is an InterpretIve Implementallon on conventional mechlnes.

whIch uses the kAC program. a flat sel 01 reCursion equations. as the
reduction rules fOr the Inlerpreler. The advanlage on a conventional machine

Is thai It Is able to produce errOr messages lhet are extremely Intelligible

to Ihe progremmer. since they conlaln references to nemes and expressions

In Ihe original source.

keller. Lindstrom and Palll use graph reducllon or source programs In their

AMPS (appllcatlve muillprocessing system> {Keller791. In Ihelr case Ihe

machine Is prog rammed In a "functiOnal graph language" (FOU. a dialect

or LISP. end the machine operates by graph reducllon of FOL programs.

LIke Mego's machine. their design consists or a large number or pllrallel

processors connecled 10 the leaves or a binary Iree or dlfferenl processors.

However. In their case the Indlvlduel proceuofll are each capable or

,Ignlflcanl computlltlons. and the Iree serves only lor communication lind
 120

Ioed b,Ianclng. The communlcallon tree supports 8 global address space.

ao any process may be run on any prOCessor. Processes are created by
INVOKE Instructions. analogous 10 subroutine calls. The AMPS does nol try
10 exploit parallelism on 8 very small scale twlthln subroullnes}. The INVOKE

InalruCllon place, the neWly crealed process on an INVOKE-list In Ihe

execuUng processor. end when Ihe processor Ilnishes (or suspends) the

current process 11 IIIk8' the n8Kl one Irom lhe INVOKE-list and proceeds
with 11181 Inslead. Periodically the communlcaUon Iree oblalns the lenglh of

the I"'VOKE-Iis! In eech prOcessor and !ranslers processes from heavily

I08de12 processors 10 nearby lightly loaded ones. In this way work 15

distributed through the machine.

The AMPS communication tree 15 able 10 support local communIcatIons very

eflIC'Elntly. since they need only pa55 through a small number 01 node5.
This Is both /15 strength and Its weakness. 115 deslgner5 belle". thai executing
proorems will exhlb/l 5ufflclant locality of reference that almost all
communlcetlons will be local and therefore 'a51. However. since all
communlcetlon& tre"ellino a dlslance 01 hall Ihe machine or more must go
through the root Of Ihe Iree. lhen If only a 5mell percentage 01
cOJllmunlceUons lurn out to be trUly long-dl5lance the rool will become the
bolUeneck 0' the enllre system.

AUCE lOarllngtonBlI Is another de5ign using graph reducllon of lhe source.

In thl5 cese the procenors and memorle5 are separate. and communicate

Itiroullh e packel switching ne1work. Thus alt proceno,s are the same
distance from all memorle5. and locality 01 reference 15 not an Issue. The

program Ilreph conslsl5 01 applications 01 funcllons whose dellnlUons are

held In the microcode sfore. Thl5 store must be loaded wllh defInitions 01

"I funcllon5 In Ihe user's prooram before execution begins. ALICE al50

IUpports ~-8xpreulons. WhiCh It Implement5 by copying Ihe funcllon body

 121

SubsUlullng lor the arguments. so nesled ).-expresslons are nOI very ellicient.

Only reference counting garbage collecTion Is performed by the hardware.

so additional garbage collecllon must be programmed elllpllclUy II circular

structures are 10 be used. ReducIble nodes are held In 8 pool which

circulates among the proces90rs. and so when 8 processor becomes Idle

II need only grab one 01 Ihese nodes 10 COntInue 9xecullng. The grain of

parallelism Is very line. since all the operalions In a function may be

performed In parallel. Parallelism can be conlrOlied explicitly. and. In

parllcular, both PAR and SYNCH (chapter 7) can easily be Implemented

ALICE contain!! some messy lealUres. lor example. uncurrled luncllons with

81 mosl three arguments are much more erUclenl lhan olher kInds. We

suspecl Ihat ellons 10 exploll such fealUres will lead 10 complex end unWieldy

sohware Nevertheless. the overall design accords very closely wllh our own

views. and we are conridenl Ihal all the resulls of our work would prove

easily applicable 10 ALICE.

9.4. SEeD MACHINES

The SEeD machine (described In secllon 3.4) has also been used as the

basis lor 1uncUonal language Implemenlattons. Henderson uses Ihls approach

In hIs Implementations of lIspklt a very simple dlalecl 01 LIsp IHenderson801.

He complies lIspkU Inlo an (almos\) linear machine code with 21 Inslruc!lons.

and then executes the machine code wllh an Inlerpreler. Inlerprelers have

been wrillen lor a wide varlely 01 machines <ranging from a ZBO 10 a VAX)

and Henderson Is now consIdering a hardware Implemenlatlon. A

considerable bOdy 01 Uspkll sollware has been wrll1en by Henderson. Jones

and Jones IHendersonB31. InclUding a complele sell-hosting programming

environment.
 122

Then Implementations perform fairly w811, but have it lendency to run oul

of 510re during long computations. Thl, 15 pSrlly due 10 Ihe prOblems

dlsc~ssed In chaptsr 7. since the Implementations are completely sequenUal.

It 15 BI!lD due to Ihe use of en"lronmenI5. which by fhelr very nalure are

wBlllelul of slora. since an en"lronmenl may be relalned because one value

In It Is needed. and may thus cause the retention of many Objects which

COuld actually be thrown away. (This problem CQuid be amellorl!lled by using

'ree variable lists rather than environments. wl'1tch contain only Ihe pari 01

the anvlronmenl that is l:lClUally needed. However. many more Iree variable

IIsls than environments would need to be constructed. and so the gains

'rom thIs technique might well be outweIghed by the extra cos I 01

construcllon). Ne'lertheless lIspkll Is a pracllcal and pro 'len programming

SYllll!m.

Steele and Sussman also used the SECD machine as the basis for their

VLSI lisP Interpreter. SCHEME-79 ISteele80). Their ptOCessor Interprels a

dlal8cl of Usp called SCHEME ISteele781. which Is similar 10 L1spklt Lisp.
The most Imporlanl dlllerences are lhat SCHEME Is not purely functional,

and that It doas nOl support lazy e'laluatlon. A '1ery simple compiler con'lerts
standard funcllons In SCHEME programs Into Indexes Into lhe microcode.
and '1arlable names Into Indexes Into the en'llronment. The SCHEME-79

processor Is able 10 Interpret lisp al aboul the same speed as e PDP-l0

KA·IO. and Sleele and Sussman expecl an Impro'led '1erslon 10 be an order

of magnttude laster.

Uh Henderson's lIspkU Implementation. SCHEME makes no aHempt to exploit

parallel processors. It Is also subject to lhe same crillclsm that en"lronmenls

are wasleful 01 storage. allhough this mailers less In the absence of lazy

8'1aluallon since suspended expressIons. the main sources 0' references 10

 123

envirOnments, do not occur lis adV8nlage Is thaI It Involves 'alrly

weU-understood technology. and so If Is more practical In the shari lerm

than some of the other approaChes In this chapter.

The SEeD machine Is also s( the heart of a project Involving Friedman.

WIS8. Johnson and Kohlstaedt. whose aim Is 10 design a parallel apptlcallve

language processor. Their work Is based on .8 language called Daisy

(KohlslaedtBl) which runs on a virtual machine called oSI IJohnsonBlI. OSI

provides lazy evaluation. a non-deterministic mullJ-set constructor lor

InWatlng parallelism IFriedman801. and allows some circular Slruclures to

be crealed. An extensIon of relerence counting garbage colleellon due 10

Friedman and Wise Is used which Is able 10 recover the permissible Circular

structures IFr/edman791. At present oSI Is Implemented bV an Interpreter

running on a convenlional machfne. but II Is expected thai a parallel

Implemenlallon In hardware will be construCled eventually. ThIs

Implementallon will make use Of many processors and many memories

connected by a switChing network described In IWIseB1L an archlteclure Quite

similar to ALICE.

Sinca oSI Is an SECD machIne the sarna criticisms of the Inelllclency 01

using environments apply to It However. DaIsy Includes some leahsres

designed to minimise the unnecessary relenllon 01 envlronmeOls. and OSI

does permit parallelism. so Ihe problem may be much less severe in this

coni ext. Parallelism is controlled In a verV dillereni way Irom Our own

suggestions. and II Is nol Clear whelher the programmer can control slDre

use In Daisy as etlectlvely as he can using PAR and SYNCH.

 12'

8.S. TURNER'S COMBrNATORS

We ha"6 already given a 'lory brief 811lplanallon 01 Turner's comblnalor

lmpementallon technique lTurner791 In section a.e. We remarked thero thai

In Ills Implemenlallon of SASL. Turner used some addlllonal comblnators

We will nol present Ihem In delalL but we will explain theIr flavour Recall

thai S Is defined by

s - ~abc. a c (b c)

.nd Introduced Into complied programs by the rute

lV.El El trans S (lV. EI) P.V. El)

S "ay be regarded as a "dlrecto," combinator C1Kennawey82) thai lakes

lis Ihlrd .rgument M and directs II lowards both Eland E J. " elthor E 1

or EJ does not require V then II must reJecl It (using the K combinator),

Tumer Improved the quality 01 the code considerably by IntrOducing selecllve
director combineiors: Ihat IS. combinalors Ihal direct the argument only

wl'1tre It Is required.

One of the allraCtions of Turner's comblnalors Is Ihal (here ara lew enough
of them to form Ihe machine code 01 a compuler. This has been exploued

by Clarke. Gladstone. MacLean and Norman In the SKIM machine IClarke801,

Tl'1ls Is a bllsllce unl-processor _llh a cell-structured memory

microprogrammed to exeeule Turner'. comblnators. SKIM Is programmed In

SNAll. a SASl-Uke appllcaUve language. end complies the SMAll compiler

IIhtif wrlUen In SMALU In aboul II quarter 01 an hour. An ImproYed version

Ie expected to offer comparable performance 10 a 88000 run nlng conventional

'enguages.
Jones ana Muchnlk have lakan a rather dllferent approech [Jones8~). The

compile each combInator InlO more prlmfHve oparallons ot a stack machine

code and then optimise the code produced. This Is It hybrid between

conventlonal compHallon and comblnalor reduction. The euthors have nOt

yet compared lis elliclency to direct combinator Interprelallon. or 10 SKIM.

The view 01 comblnators es "directors" has been taken turther by ~ennaway

and Sleep IKennawayB2J. They represenl a director as one 0' ,. I. \. and

- Indica ling thai an argumenT Is 10 be dlrecled to both. the lell. the rIght.

or neither 01 Ihe branches of a lunctlon eppllcatlon. They use strings 01

such directors as comb/nators. Indicating the appropriate directors lor

successive arguments. DirectOr strtngs are attractive as comblnators because

Ihey can be represented very conCisely. using only two bits per dlreclOr.

These dlreclor strings will be used In Burton and Sleep's ZAPP (lero

Asslgnmenl Parallel Processor) [BUrlonBJJ. ZAPP conslsfs of a large number

of processors. each wllh considerable local memory. The processors are

connected In a cycliC netWOrk thai allows each processor 10 see an Inllnlle

virtual binary tree 01 olher processors. As In AMPS. each processor maintains

a list 01 pending processes. When a processor becomes lightly loaeled It

may steal a penellng proCes' tram one 01 Its neighbours. with Ihe reslrlcllon

that a stolen process may never be re-stolen. This restriction guarantees

thai a process never migrates furthar than one network communication Irom

115 origin. and therelore that Its result need only be transmllled one nelwork

step once It Is computed. Since eaCh processor sees 8n Infinite binary 1ree.

work can spread through the network very quickly. lAPP does not support

8 global address space: Instead all deta a process may need 15 slolen along

wllh Ihe process. Herein lies the weakness 01 the design. tor data can be'

,tolen before It I, computed. One unevalualed datum can be atolen many

 126

times bf many cJlllerenl processors. forcing them all 10 evaluate II and

mUltiplying Ihe amount 01 work 10 be done. Thus the attempt 10 ensure

locality 01 relerence risks enormously dupllc!Ued computation. II remains to

be seen whether this will occur In pracuce.

ZAPP ~I$O Incorporales an elegent device tor controlling parallelism

8utomallcally. Burlon and Sleep observe thai breadth-Ilrst computation leads

10 lois 01 parallsllsm. but also lOis 01 space utilisation; on tha olher hand.

depth-llr'l computation Is very space ellicleni but gives rise 10 no

parallelism. ZAPP compromises by assigning breadth-first processes II lower

priority than depth-lIrsl onas. and executing higher prlorlly processes by

pr8Ier,nce. This leads 10 breadth-first computation unlll all processors are

In use. followed by deplh-lInt computation. resulllng In a space requirement

proporllonai 10 the number 01 processors.

8.8. SUPER-COMBINATOR APPROACHES

We hElve not yet made a realistic Implementallon of a functional language

using super-comblnators. and nor. 10 our knowledge. has anyone else.

However. Johnsson has Independently developed a very similar teChnIque

whlcft he uses In his Ml compiler IJohnsson93) (actually he complies a

purely Juncllonal dialect 01 Ml with lazy evaluallon) Johnsson generales

Combinalors from Ml programs by a process he calls "lambda IIftln~·. which

Is ,nalogous to generating super-comblnalors usIng only the methods 01

aectlons 3.5 and 4.5. The comblnators are then compiled Into machine cotje

for tile G-machlne, an abstract machine with a slack In addition 10 the graph

atorage. which perlorms the assoclaled graph Iransformalion. The G-machlne

code Is translated Into VAX machine code and called from a graph reducllon

Interpreter. However. Johnsson goes 10 8 great deal of trouble to optimIse

 127

the O-machlne code so thel /I avOids constructing pieces of graph which

can Ufe" be reduced Immediately: tor eXllmple. the 'unction auec dellned

by

suec n • n + 1

Is complied Inlo O-machlne code Which never constructs the expression n+ 1.

but just com pules Ifs value sireighl IIwlly. This. together with other efforts

10 optimise the code produced. makes Johnsson's lezy Ml run liS lest 85
more conven,lonal lengu_ges. Johnsson Is considering other Implemenillfions

at the a-machine. Including olin Implementllflon In herdwere end l!l parallel

version.

In lu1<111100. FaIrbairn will shortly be using super-comblnelors In an

ImplementaliOn 01 hIs \angu8g8 Ponder IFllirbelrn82J for the Molorolll 88000.

like Johns,on. he will make grellt efforts to "lIecule each combinator more

elflclenlly (han by simple graph Iranslormatlon.

1.7. DATAFLOW

A very dlfterenl view ot htncllonal language Impiemeniatlon Is embodlad In

detaflow designs. The origInal Idea ot dalaflow was thai the program woula

be represented by a network 01 Inslruellons through which date would flow.

and out of which answars would emerge. Since many paris of lhe graph

might be actlye at the same Ume the approllch often good prospacts lor

parallelism. DennIs et al. designed a machine based on Ihese principles

described In IOennl5791. which consisted or eight processors conneclad by

• switChing network like Ihe one used In ALICE and DSI. Its disadvantage

I. Inherent In the o"glnal conception: since the program Is slaUe. II Cllnnol

coniain Invocations or recursive or higher-order functions.

 128

Walson ana Gu," proposed 8 partial solullon 10 this problem In (W'IlS0n79).

whereby d818 Is ·coloured" wltl1 the partlcula, function Invocation It belongs

10. Thll 8110ws mUltiple concurrent activations 01 the same piece of progrem

graph. and so recursion Is possible. but Ihls scheme slill does nol support
higher-order funcllons n8lurally. Walson and Gurd have built 8 single

prOCQ!;sor machine 10 their design. which achieves high speed by parallelism

In81de Ihe processor. Extension 10 8 mu/llprocessor version would present

no problems. An Imporlant point aboul their design Is Ih8\ one processor

execuBs several processes slmulteneously. and so II Is not held up while

• communication. such as • memory access. Is In progress. ThIs means
the, their communlcellon network. while It musl have high Ihroughpul. need

nOI n.cessarUy complete Individual fransacUonl quickly.

In IDlII1nl,81j DennIs proposes a new kind 01 dataflow archlleclure In whIch

the f)'ogrem graph chenges dynamically. It differs Irom previous dalaflow

dellg!s In that a datallow greph represenls an Object. which reduces 10

e velu,. rather than a function through whIch dala musl tlow. " Is eble

10 support higher-order funcllons and so on. and In lacl. II Ie reSilly a graph

reducllon erchlleclure 01 lhe kInd we have been discussing In this thesis.

The major dlllerence Is thai evaluation Is dala-drlven rather lhan

demand-drIven. fhal Is. all reducible 8llpresslons In the program are

evellliled In parallel. whether Ihelr results are really needed or not. We

suspect thaI some method of restraining parallelism will prove necessary.

whereupon dalallow archlloclures and other graph reducllon architectures

will have converged almost completely.

 CHAPTER 10

CONCLUSION

Our IIrsl conclusion Is Ihat super-comblnators provide a reasonable and

efficient Implementation method, superior III least to Turner's comblnators.

We hllve demonslrated 'his experlmenlal!~ and theoretically, and other work

(seCtion 9.6) 5UPP0rlS our conclusion.

Of particular Interest is the discovery 01 the ·slo....-do....n faciO'" (chapter

6>. In which we refused to believe at llrsl. We originally lhought thll no

"Q1IW-down lactor was the mark Of II good Implemenla1l0n melhod. and did

not believe Ulat Turner's comblnators suffered Irom one unlll we saw

Kennawa~'s proot. We 'houghl that super-comblnators had no slow-down

lactor unlll _e saw Fischer's example. Only after being wrong twice did we

te-examlne other approaches and observe ,he same behaviour In them all

We conclude fr<lm Ihls Ihal II Is dangerous to design Implementalion

stralegies sale Iy 10 make Individual operauons last. When a new strategy

I, proposed II sh<luld be accompanied by a carelul analysis 01 115 slow-down

behaviour.

We believe we have demonstrated conCluslyely thai luncllonal abstract'

machInes must suppOrl parallelism. lind thet there is strong 8'Wldence Ihal

Ihls parallelism must be controlled explicitly b~ the programmer. This can

 130

be d"ne adequately using our PAR/SYNCH construcllons. whiCh have lhe

ImpManl property thai they can be IHlded 10 a workIng program: they need

not be considered durlng (he desIgn. However. a more structured al1ernallve

might prove eesler to lise.

In '~e more distant luture mark-scan garbage collecllon wlll become

Impractical due to the size 01 virtual memories thai wilt need 10 be garbage

collected. Our IInal conclusIon Is that an 8ltlension 01 reference counting

can be used Inslead. at a moderate additional cost

 REFERENCES

IBackus781 J. Backus, 'Can programmIng be IIberatad lrom Ihs von Neumann

slyle? A functional 51yle and lis algebra 01 programs'. CommunlClllons of

the ACM. \/'01. 21, No. 0

IBerkllngl K. J Berkling. 'A Complele AbslraCI lambda Calculus "'achlne".

tBobrow801 D G. Bobrow. 'Managing reentrant struclures USing relerence

counIS·. ACM Toplas \/'01 2. No.3. JUly.

lilrownbridgeB31 O. Brownbridge. PrIvate communication.

IBurslall80l A. M Burslall, 0 B. MacQueen. D. T. SaneUa. ·HOPE. an

ellpenmen1al apphcalive language'. Proceedings Of Ihe ACM lisp Con~rence.

IBurlonSli F. W. Burlon, .... A. Sleep. "Executing funCtiona' programs on a

virtual Irell of processors'. Proceedings 01 Ihe ACM Conference on Funcllonal

Programming Languages and Com puler Archlleclura.

IButlonS21 F W. BUrian, "A linear space Iranslalion 01 luncllonal programs

10 Turner comblnalors',

(ClarkeOOI T. J, W. Clarke. P. J. S. Gladstone. C. O. "'aclean. A. C. Norman.

'SKIM - Ihe S. K. I reduction machine', Proceedings 01 Ihe 1980 lisp

Conlerence.
(Curry58] H. B. Curry, R. Feys. ·Comblnatory logic·, North-Holland Publishing

Com pan,. Amsterdam

IOarlinglon811 J Darlington, M. Reeve, "ALICE - a mUlti-processor reduclion

machine lor Ihe parallel e..-aluallon 01 applicallve langueges", Proceedings

of Ihe 1981 ACM Conlerence on Functional Programming Languages and

CompUler Architecture.

LDennl579J J. B. Dennis, "Tile ..-arielles 01 daia-llow computers·, Proceedings

ot Ihe lsi Internallonal Conlerence on DIstrIbuted Compuung.

(Dennis8lJ J. B. Dennis. "Data should nOI Change - a model 'Or a computer

syslem', SymposIum on FuncUonal Languages and Com pUler Arclliteclure.

IDennll821 J. B. DenniS, Pri..-ate communlcallon

IDI,kslra791 E W. DIJkstra, L. Lamport. A. J. Martin. C. S. Schollen. E f.

M. Sleflens, "On the lIy garbage collecllon: an exercise In cooperaUon".

Comm~nlcaUons 01 the ACM Vol. 21 No. 11.

IFairbalrn821 J. f-airbalrn, ·Ponder and lis Iype syslem", unIversity 01

C.mblldge Computer laboratory Technlcel Repon No. 31

lFlscter721 M. J. Flscller, "Elflclency 01 equivalence algorUhms', CompleJi'y

of Canpuler Computations. Plenum Press. New York.

(Frledman791 D. P. Friedman, D. S. Wise, "Relerence counting can manage

Ihe circular In..-lronmenls 01 mulual recursion". Information Processing Lellers,

Vol. 8, No.
lFriedman80J O. P. Friedman. D. S. WIse. "An indetermlnale constrUClor lor

appllcallve programming". 7th Annual Symposium on PrincIples 01

ProgrammIng languages.

IGordon79l M. Gordon. A. Milner, C. P. Wadsworlh, "Edinburgh lCF~.

Springer-Verlag leclure NOles In Computer Science VOl. 78.

IOr1(81) D. H. Onl. R. l. Page, "Deleling Irrelevanl tesks In an

expressIon-orIented mulU-processor system", ACM Toples Vol. 3, No.1.

a
IHenderson76J P. Henderson. J. H. MOrris, "A lazy evaluator , Proceedings

01 the 3rd Annual SIGACT-SIGPLAN Symposium on Principles 01 Programming

languages. Allanla.

IHenderson801 P. Henderson. "Funclional programming: appllcaUon and

ImplemenlaliOn R
• Prentice-Hall.

IHenderson83) P. Henderson, G. A. Jones, S B. Jones. "The Ilspkil manual",

Oxlord University Programming Research Group Technical Monograph

PAO-32.

lHudak82) P. Hudak. R. M. Keller, "Garbage coliecUon and task deleUon In

diSlrlbuted applicaUve processing sySlems", Proceedings 01 the ACM

SymposIum on LISP and FuncUonal Programming.

IHughes82l R. J. M. Hughes, "Super-comblnators: a new ImplemenlllUon

method lor applicaUve languages~, Proceedings 0' lhe ACM Symposium on

LISP and Funclionai Programming.

(rtughes82J R. J M Hughes, "Graph-reduction wIth super-comblnaIOrs".

Olillon:! Ul1ll1erslty Programming Research Group Technical Monograph

PRG-28.

tJohns()n81) S D Johnson, A. T Kohlslaedl. "DSI program description".

Indiana Unillersily Computer SCience Departmenl Technical Report No 120

LJOhnssonB31 T Juhnsson. "The G-machine· an abstract machine lor graph

reducllon". JUint SERC/Chalmers Unillersily Declaralille Programming

WorksllOp, UnllltHSlly COllege London. May

{Jone~a2J N D. Jonos. S S. Mucnnlk, "A fixed program machine tor

COmbinator expression ellaluahon', ACM SympoSIum on LISP and Functional

ProgfilmmlOg

IKarl~sonfi Ij I'\. I'\.arlsson. "An oul/tne at the SKY reduclion mactllne",

Symposium 011 Functional Languages and Compul8r Archlleclure

lI<eller791 R. M keller, G Llndslrom, S PaUl, "A loosely coupled apphcahve

mulll-processHlg ~ystem", ProceedIngs 01 1he 1979 AFIPS Conference

IKeuIlaway£l21 J. R Kennaway, M. A. Sleep. "Director strings as comblnators".

Unlyerslty of Ea~l Angha.

IKennawayB2J J. A I'\.ennaway. "The complexlly 01 a translallon 01 ),-cafculus

10 combinators".

IKohlslaedlBli A. T Kohlslaedl, "Daisy 10 reference manual", IndIana

Unl'8rslly Computer SCience Departmenl Technical R9pOr. No 119.

Ikowalski79J A. A. kowalskI, "logiC for problem solvlno". Nor1h-Holland.

Ilandln6041 P. J. lan(lin. "The mechaniCal evalualion 01 a,-prassions".

Compuler Journal. January.

(llskov791 B. Liskol/. A. Alkinson. T. Bloom. E. Moss. C. Schaller!. B. Schelfler,

A. Snyder, ·ClU relerence manual". Massachussels Institute 01 Technology

labOratory of Compuler SClance Technical Ropori MlT/lCStTR-225.

IMag0791 G A. Mago. ·A network 01 mIcroprocessors 10 execute raductlon

languages·, two parts, International Journal of Com puler and Information

SclenCRs. vol. IL Nos 5. 6.

IMag081J G, A. Mago, 'Copylng operands ...ersus copying resulls: It solution

10 the problem 01 large operandS In FFPs·, Proceedings 01 tha 1981 ACM

Conlerence on FunClional Programming languages and Computer

Archileciure

IPaylon-Jones62J S l. Peyton-Jones. "An in...e.sligalion ot the relative

elllciencies 01 comOlnalors and ),-eltpresslons", Proceedings 01 the ACM

Symposium on lisp and Funclional Programming.

IReynolds701 J. C Reynolds. 'OEDANkEN - a simple type less languaoa based

on the principle 01 completeness and the relerence concept".

Communlcallons 01 Ihe ACM. Vol. '3. No.5. May.

ISIeele781 O. L. Steele Jr" G. J. Sussman, "The revised reporl on SCHEME:

• (Ilalecl 01 LISP", Massachussels Inslllute ot Technology Artificial Intelligence.

labOratory Memo 045.2,

 IS18el8801 G, L, Steele Jr.. 0, J, Sussman, 'Deslgn 01 8 LISP based

microprocessor', Communications 01 Ihe ACM, Vol. 23, No 11

ISutrin8JI B, A Sulnn, "Readmg formal specifications', Oxlord Unlverslly

Programming Research Group Technical Monograph PRG-2-4,

1T8rJ~n72J R. E. larlan. "Depth !lrSl search and Imear graph algorithms',

SIAM Journal of Compuling Vol, I, NO 2

n aqan75J A E Taqan, "Elhctency Of a gOad but not linear sel union

algorlttJm", Journal 01 the ACM, Vol. 22. No 2. April

lTurner161 D. A Turner. "SASl language manual'. 51, Andrews UnIversity,

(Turner1Sl D A, Turner. "A new lmplemenlatlon lechnique lor appllcallve

langulges', Sollware, Practice and Expenence, Vol. 9,

lTurnar191 D, A, Turner, "Another algorithm for bracket abslracUon", Journal

at SymbOliC logIc, \/01. 44, No_ 2, June,

lTurner811 D A, Turner, "RecurSIon equallons as e programming language',

Newcastle Summer School In FuncUonal Programming.

(Wad18r83J P, l Wadle" "llsllessness Is beller lhan laziness", Ph,D, Thesis,

CarnegIe-Mellon Universlly

IWacller831 P l Wadler. prlvale communlcaHon.

~
fWadsworU17lJ C. P. WadswOrth. "Semantics .nd pragmatics 0' Ihe

l--calculus·, OxlOnJ University D. Phil. U1esls.

[Walson]i1 I. Watson, J. Gurd. "A prototype dataflow computer with tok;en

labelling", Proceedings of Ihe AFIPS Conference

[WiseBl] 0 S. WIse, 'Compaci tayouls 01 b&nyan/FFT nlt(works", VlSI Syslems

and Computations, Compuler Science Press. Rockville. Maryland.

 Appendix
A.l Introduction
A.2 Nose: the language and examples
A.3 The Nose compiler
A.4 A Lispkit compiler in Prolog
A.S The interpreter
A.6 Experimental results
A.7 References
A.l Introduction
Thc.sc appendices contain details of the experiments referred to briefly
In section 4.6. The experiments were performed during i981, and were
intended to compare the performance of an SKI implemenlaion with a
super-combinator implementation. Since I did not originally intend to
include my more than the results of these experiments in my thesis I
did not preserve listings of the programs used or sample!! of their
outpuL Instead, these appendices consist of an explanation of the
cxperimentJ in sufficient dctail to allow the reader to repeat them.

We begin in section A.2 with a brief description of the source

language we compiled, and some sample programs. In section A.3 we
will describe: the compiler used to generate SKI and super-combinator
code. Sinee we cannol reproduce a listing of it. we include in sec lion
A.4 a lilting of a working compHer written in Prolog, which compiles
Lispkil Lisp into super-combinators. In section A.5 we discuss the
abstract Inachine-code interpreter and the slatistics we gathered. Finally,
in sectio~ A.6 we present the e~perimental results in more detail.
A.2 Nose: the language and example.
The functional language which we compiled in our experiments was
Nose (an acronym for NO Side Effects). Nose was designed at
Cambridge in 1980 as part of a Diploma in Computer Science [Hughes80].
It draws heavily on the applicative subset of PAL [PAL ler] for its syntax,
and so should appear familiar to anyone knowing PAL. ISWlM (Landin66] or
SASL [Turnu76). Like PAL, Nose has fixed length luplcs rather than lists.
and these must be chained together for list processing. Unlike PAL,
Nose is purely functional and has a lazy semantics. Nose type
checking is performed at run-lime.

Rather than give a precise definition of Nose, we wiJI indicate its

flavour by giving seYeral example programs. The first of these is a
program to compute Ackerman's function. This program was actually
used in our experiments.

let ree IClt m n •

ir m::rO then n ..1
t1ir n=O then IClt (m-I) I
else aell: (In-I) (IClt 1"11 (n-I))
r­
inad.2.]

This is a curried version of Ackerman's function. When programming in

Nose we tended to write non-curried functions instead, using Nose
tuples as argument lists. The follOWing program is a non-curried version
of At.kerman's function, and was also used as an example in our
experiments.

let rec aell: hn.n) •

Ir m.O tben n'"
em n-o 'ben lell: [m-IJ.)
else cd: (1II-J..ac.ll:(m.n-ID
n
in ad: [2.lJ
As an example of list processing in Nose, we include the follOWing
program for computing primes.

.how [20.a,eYc(rtOIll 2.)]

where fie
( rro. n • (nJrom(n+I))
also dna (P...) • (p,Ilne(rIIter l»)
wbere tee filler [n,.z) •
Jr n~p=O tben niter J:
else [n,liher lJ
n
allO tee 'boW [n)) •
Jr n-O \ben ._­
elM [111.••• ,Ibo'll{n-lJ.aD
n
)

Tbil pros ram. whicb Usel Eratostbene:.' lieve, was adapted from one in
[Turnu79a~ It illustrates the use or Nose's indexing operator (!) to access
componcfltJ or I. pair. This program was also used in our experiments.

One or our most complex examples was a runctional uniCicalion

algorithrn. The program we actually used has been lost, bUl a similar
Nose definition might be:

let ree unify [a..,b,env)

•• a and b au: uprusions 10 be unified. tnv is tither 'empty' (or Iype

void), Dr it is a tuple of bind'lngs of type [nrin.,any}"'. unify rcturns

•• a neW environmcnl whieh II an c.llcnsion of cnv in which a and b

arc equal (aflcr substilution of l'aluu for vJriablr.s), or 'empty' if

• no sueh cnvironment nists. Variablu are string I beginning with ~ .. ~.

kind env in
empty:l'oid ~) empty
bindings:[suinl,any]'" .)
if a=b thcn bindinls
eli' isl'ar a thcn
if bound[a,bindinll] then
unif y[l ook up(&,bi ndi nls], b,bi ndings J
cl.e [a,b) prc bindinRI
r;
elif isl'ar b t\:en unify[b,a,bindinlsJ
else kind [a,b] in
[aLup:anY"',btup:anyll] ~)
if lcngth atup=lenllh btup tbcn
uu[l,bindinll]
where ree uu[i,env'] =
if i>lenlth i:.tup thcn cnv
ehe unify[&tup!i,blup!i,
uu[i+l,cnv'll
r;
else empty
r;
I otber:any -) empty
dnik
r;
dnik
where bound [a,eM] • (bound'
where tee bound' i ..
if i)lenltb env then false
dK let (a',v'] ,. enl'!i In
if a=a' then true else bound' 0+1) n
fi)
also lookUp [a,env] • (lookup' I
wbuc rtc lookup' i ..
leI [.',v'] II enl'!l. in
if .e.' Ihen y' tlK lookUp' (i+1) fi
also in.r ...
kind. in
S:lltrinl -) Item s II ~II"

I olhu::any -) falK
dnik

This example demonstrates Nose "kind-expressions", They are the only

type-testinl mechanism Nose provides,
These examples should be enough to Sltisry the reader's euriosity. We
do not claim that Nose is a partieularly lood language. Several aspects
or ill deslSft appear wrong in reCrospeet - ror eumple, the USe or
filed lengtb tuples instead or lists. and the unwieldy kind-expressions.
It wu only an experimental tool, and it served its purpose.
A.3 The Nose compiler
The Nose compiler wu originally written at Cambridge to investigate
optlmisati<ln methods for lazy languages [HUlh~80]. It was written in
GEDANlCEN and run on the CPU GEDANKEN system, developed at
Cambridge between 1976 and 1980. It was ported to Odord by Writing
an abstrael machine-code interpreter for GEDANKEN code.

The compiler wu extended to compile to either SKI combinator! or

super-col'l1binalors. First a common pass transformed declarations into
lambda-cIpressions and syntactic constructs into calls of standard
functions (see section 3.2). The output of this pass was an expression
in lhe pue lambda-calculus with constants. Since everything up to this
stage was common to the SKI and super-combinator compilers, no
eXtraneoUI factors influenced our experiments.

The lambda-expressions were translated into SKI combinalors using the

methods of [Turner79a] and (Turner79b]. That is, Turner's tater "optimising"
combinators (S', B' and C') Were used. The lambda-e:,:pressions Were
translated into super-combinator! using the imperative algorithm of
section B. The super-combinators were then compiled into Dept cod~
that, whm execut~d, built the result of applying the combinator. Th~
SepL code wu called by the combinator interpreter at run-time. An
earller ~rsion interpreted super-combinator definitions: compiling to Bept
improved speed by I~%. The improvements of sections 4.4 and 4.5
were nol used, and in some cases this has biased our results against
luper-conbinators. This is most noticeable in the case of Ackerman's
function (see section A.6).

A deficiency of our compiler was that Nose tuples were compiled as

lisn. Since list processing programs in Nose chain tupLes together this
introduced an unnecessary overhead into the code produced by both
compilers. which will have distorted our results somewhat. Differences
In efficiency which we have observed would be more pronounced of
[hi. had been corrected.

Secondly, our compiler was very slow due to the use of a sept
Interpreler for GEDANI:EN abstract machine-code. This, coupled with the
.man unount of store available (about 25K words) limited· the
.ize of example we could try sever~ly. Our results may therefore be
distorted by "end-effeeu". We have interpreted them by looking for
trend. as program size increues.
A.4 A Lispkil compiler in Prolog
The Note compiler described in section A.3 was rar too large to
include as an appendix. So that the reader my run a super-combinator
compiler if he so wishes, we have included a version written in
ProloK;_ This compiler iI the model for section ,.,. It compiles Lispkit
Lisp rather than Nose. but is otherwise very similar.

COMPILE.PRO

/- Tbis rue contains the stcerln. prolram or the Lisplr:il compiler. Consult it. Ind then
In~olr:c tbe compiler by. for cn.mple, compilt{'frolo.lso','holo-sup').
-,
1- consul1(lilu). /_ standud li.t-prCKa.inJ (unctions _/
1- consult(lisplr:h).
1- consul1(.up).
1- con.ult(optimlse).
7- conlult(writecode).
7· conluh(cn'l').

colnpllt{From.To) :­
w,itc('USP Compiler to Supcr·comblnatou'). n1,
wrlte(From)• •rite{' ---) '), wriu:(To). nl,
(docomplle(Prom,To). ~t'rlte('Compllltion .ucuedJ·). nl;
wrlle('Com pllalion reUed'). nil.
docomplle(From.To) :­
IU(from). read(Prol). ItCn,
wrlte('Syntu. Analy.l. CODlpltt". nI. I.
preproc(Prol.X). lI.pklt(X.Lam).
wdte(·Con..crttd to Ilmbda ealclIhu'). nI. I.
sup(nlle.Lam .upr(O,[J.5I1P....J.m.
wrlte('Com pUed to Juper-c:om blnatou·). nl. I.
oplimlR(Sup.Opt),
wrhe('Optim Itation com plete·). nI. I.
ldl(To). wrilecode(Opt). nl. told. l
L1SPICIT.PRO

r. This file derma the predicate Iisptit(A.B), which c.onyerLJ I lispkit pralum
into lambda-ealtulus ready (or compilation.
'/

Iispklt(Name,quotc(NIDlc)) :- ttandlrd(Namc).
Ilspkil(lnt,quotc(constlnl(lnt))) :- irnclu(lnt).
hspkit(Namc,Namc) :- IlomCNlmc).
hiP Iti 1([quotc,El,q yale{canstanICE»).
1Isptil«(lam bd"',[],Bodyl,Body!) ;- lisptil(Body.Bodyl).
lisplr:it«(b,m bd.,Nam toNlm CI.8ody 1.11mbda( N I mc,Body I» :­
lisp ki t([I_ m bd.,N 1m es,Bodyl.Body I).
lisp kll«(1 cl.Bod y leehJ.E) :­
decom pose(Defs,Nlm eI. Val,),
Iispti 1(( II m bd.,N 1m es,.BodyJ.EI),
I PPIisl( Et.V IIs,E),
hsptit«(Ict rec"Body IDd11•• Pply (Body 2,Ipply(quote:( fi ll).Vab2»)) :­
decom pose(OcfI,NI m d. Vals).
Iispkit(Body,Body!),
lam Iist( N I mes.Body I,Body 2),
mlkeIiSl(Vab.VahU,
lam Iist(Names,Valsl,Vals2).
li1pkit((F],F1) ;.- lispkil(F,FI).
lilpkil((F,Al,lpply(FI,AI)) :- Iispkil(F,FI), lispkil(A,AI).
l\t;Jkil«(F,AlRl,E) :- R\o:(l lispkil([(F,AJIR1,E).

appHst(E'(l,E).
appllst(E,V.vs.EI) ;- Iispkit(V.VI), appliSl(apply(E,VI),Vs,El).

lalll Ilst([],F,apply(quotc(k),F».
bmllst(A.B.F,apply(quotc(unpal;k)Jlmbda(A,A») ;- Ilmlist(B.F,FI).

m Ikclist([],quotc( ni I».
DIlkdi.t(A.B,apply(lpply(quotc(cons),AI.BI» ;-.
matelill(B,BI). Iilpkit(A,AI).

d""m pOK([]!U]).
dcmm po:ae((A.BlClJAIN1,(OrvlJ ~ dceomposc(C,N.V).

allndard(add).
Itlndard(sub).
ItIJdard(di'r-).
Ilildard(mul).
11I.ldard(rem ).
tl~rrd.rd(eq).
Italdard(lcss).
•taldard(.reater).
tlardard(cons).
IUlrdard(c.ar).
Itardard(tdr).
Itamard(lf).
ltanlard(wbicb).

preproc(A,A) :- atomlc{A).
preproc(A.B,ALBI) ;- preproc(A,AI), preproc(B,BI).
prcproc(A.C) :- A-..e. prcproc(B,C).
SUP.PRO
/_ TbiI Is a compiler to luper-combinaton written In Prolol. Input syntax:
atom quole(conn) IppIJ«n,ar1l la.bet.Od.e)
Output syntu:

., aupu(nul',body) .,,(n)

/. .up(EnY.ln.Out,Nlmes) takes In enyironmenl and an Input expression, and

computes an output eJIprcuion of the ror. eJIpr(levd,mres.object.source).

., Names I. bOlllorphle to mfes.

luP(EnY,ld.apr(LJJ•• fa(N),ld),(]) :- Itom(ld). lookup(Eny.Id,L.N), L

lupL..... quolc(ConJ[).eJIpr(O.(].quotc(COrat).quotc(COnst»).~.: :- L
.up(EnY,lpply(Fn,.Af.).E:lpr ,Names) :­
sup(EnY.Fn.FnExp,FnNam es). lup(EnY.A ' • .AraEx p.A rINam es),
lupapp(FnExp.Ar.Exp,FnNames.Ar.Names.EJ:p.Names). L
lup(EnY,lam bdaUd.E:l.p).E:lpr,Nam es) :­
blnd(Eny)d,Nar1.,EnYI).
lup(Enyl.&p.EBodyO.NBodyO).
lookup(Envl,Id,LeY,Nu••).
lam bodyCL ey.EBod,O,NBodyO.EBodJ,NBodJ),
luplam(lam bdl(Id.E:lp).EBodJ.NBody.E:lpr,Names,NarIJ). l

Jupapp(cxpr(FL.FM.FE,FS),cxpr(AL,AM.AE.AS).FNm,,ANms.
expr(L.M,. pp 1,(FE.AE),.pp Iy(FS,ASJJ,N fa,) :­
(FL-o~L-O;Fl;"AL), m....(FL,AL.L).
Ippend(FM.AM,M), append(FNmJ,ANmt,Nmt).
lupapp(cxpr(FL,FM.FE,FS),cxpr(AL,AM.AE.AS),FNmt,ANmt.
apr(AL.txpr(FL,FM,FE,FS).AM..pp1l( ar ,(NN),AE),lpply(FS,AS)),
name(NN,FNm.).ANml) :­
FL<AL. FL \-0.
I·Jp.pp(cxpr(FL.FM.FE,FS),upr(AL.AM.AE,AS),FNmt,ANIa'.
apr(FL,apr(AL,AM,AE,AS).FM..pply(FE.ul(NN)),lpply(FS,AS)),
namc(NN,ANm.).FNm.) :­
FL> AL, A.L \-0.

laU(A.B.B) :- A<B.
laU(A,.B,A) :- A)-a

Ilrnbod,(Lev.apr(L.M,E.S),N,cxpr(L.M.E,5).N) :- Lty-L; L=O.

lambody(LeY.apr(L Jd,E"S).N,a pr( Lty.[a pr (L,M.E.S)llf I(NI),S).[nam c( NI,N)]) :­
Ln\-L. L \-<1.

I1lplam(E:l.p.apr(BL,.BM.BE,BS),NBodJ,Expr,N. . a.N u ls) :­

IOrtlaCc:s(BM,NBodJ.BMI,NBodJI},
opllllCc:s(BMI,NBodll.BM1,NBodJl).
lalr:lP(cxpr(Olltupu(Nu Is,BEJ,& p),BM2.,NBody1,.&pr,Namrs.NIrII).

OplmCallUJ,lIJD.
opt.Ca(M.Ms,N.Ns.Jo(.MsI,N.Nsl} :. not(membet(N.Mt)). optmfc:s(Ms.Ns.Msl,Nsl}.
optIlllCa(M.Ms"N.Ns.Ns1.NII) :- demUll(Ms......). a1e::.ent(NI),N), optlafc:s(Ms.Ns,MaI,HsI).
..'mreo(lllUW).
10' Ull res«( apr(L .M,E-SlIMs],N,MsI,Nsl)
Iplilm res(L In pr(L,M..E.Sl[Msl,N,sm M,sm N,EqM,Eq N,Bi M,Bi N).
IOfIDlres(SmM'sn:lN,smMI,smNI). ~ortmrc.o(BiM.BiN.BiMl,.BiNI).
append(BIMJ.EqM.BiEqM). appeod(BiNI.EqN.BiEqN),
appcnd(BiEqM,5mMl.Msl). append(BiEqN,smNI.Nsl).

_,hm reo(L lUl.IUJlU WI)).

sp;itlll rt.3(LO.a.pr(L,M.E,s).Ms.N.Ns,upr(L,M.E,S)SmM.N5m N.EqM.EqN.8iM.BiN)
L<tO.lplitmrt.3(LO,Ms.Ns.SnIM,smN,EqM.•EqN,BiM,BiN).
• pI till res(LO.ClIpr(LO,M,E,S).Ms,N.Ns,5m M,Sm N,tl pr(LO.M,E,S).EqM.N.Eq N,Bi M.Bi N)
apliln:1res(LO,Ms,Na,SmM,5mN,EqM,EqN,BiM,BiN).
spltm res(L O,u:pr(L ,M,E.S).Ms,N.Ns,Sm M,Sm N,EqM.EqN,u pr(L ,M.E,S).BiM,N.Bi N)
L> LO, splltDl res(LO.Ms,Ns,.sm M,5m N.EqM.EqN.BiM,BiN).

mlap(Com blJn.Com bll.l).

m~.p( Com b,A, •. M.nam e(NUIII.A r INam es ).N,& pr .N am cs.N it r.s) :­
lft klp(Com b,N,N,E.lprl,NI'" esl,N ar Id),
.UPlPP(EJr. pr).A r I,N1m csl,A r I N am es,EApr,Nlm cs),
NUl' it N1rl.I+L
 I.
~
t
o
"~
ffi i
A.S The interpreter
In order to ensure that our measurements of the erriciency of SKI
combinator! and super-combinators were comparable, we used the same
interpreter in each casc. Our interpreter Wa! written in BCPt, and was
very similar to the one described in rTurne,19~]. The program being
cJ:eculed WlS represented in memory as a graph of application cells.
The interpreter reduced an expression by. first of all, following
function pointers until it reached a non-application. This would be the
function to be applied. The nodes passed were placed on a
"Iefl-ancestor stack", from which the arguments of the function could
be (oun~. The result of the function application was computed, and
the original expression over-written with an "indirection: node" referring
to this result. These indirection nodes were removed by the garbage
collector. This process continued until either an atomic result was
computed, or a function was found without enough arguments. At this
point. the original eJl'pression was fully reduced.

The inlerpreter contained a number of hand-coded functions for

performillg basic operations, and hand-coded definitions of all the SKI
combinalOrs. When used to eJl'ecute super-combinator code, the
compiled super-combinators were loaded along with the interpreter and
called directly by it.

The inlerpreter made a. number of measurements during eJl'ecuLion.

These were:

Tbe number of reductions performed. We expect the

super-combinator interpreter to do significantly fewer reduclions
sillce each super-combinator corresponds to several SKI
ccmbinators. The ratio of SKI reductions to super-combinator
uduction. is an indication of the number of SKI combinators
that each super-combinator corresponds to.

Trae total number of cells claimed. This measure was used by

Turner in his comparison of SKI combinators with the SEeD
machine. It is probably the best machine-independent measurement
.of efficiency.

Trae number of garbage collections and the maximum number of

cells in use simultaneously, measured at each garbage collection.
This figure WI.' intended to reveal the amount of store actually
required by each program. Unfortunately. our compiler
restricted us to such small programs that garbage collections a't
fIIn-time were rare, and no meaningful figures were obtained.

The run-time in seconds.

The relulu of these measuremenu, and other measure menu performed

by tbe compiler. are summarised in the next section.
A.6 Experimental result.
Table I. Purpose and size or program source. rncuured in list cells.

Program liil&. fl!llllls.

I 26 caU "twice" [Turner19a]
1 36 Ackerman's function (curried)
3 49 towers or hanoi

,
4
"
'5
Ackerman's funclion (non-curried)
faclorial
6 93 append
7 106 20 primes
& II' eratosthenes' sieve
9 307 unirication algorithm
10 317 e to 20 decimal places

Table II. Compile-time in seconds.

Prplram liil&. SKI ~ ~

I 26 124 177 -43
2 36 166 206 -24
3 49 225 261 -16
4 .$1 199 243 -22
5 15 230 342 -49
6 93 321 372 -16
7 106 422 429 -2
& m 463 46& -I
9 307 1591 1341 +16
10 317 2216 1265 +43

Table III. Code size in cells.

Prl!&Wll ~ SKI ~ ~
1 26 22 30 -36
1 36 48 43 +10
3 49 70 70 0
,4 .$1 76
91
62
99
+5
-9
6
7
"
93
106
130
160
11&
14'
+9
+9
8
9
II'
307
176
479
153 +13
+7
44'
10 317 639 435 +31
T'ble IV. Number of reductions.

Program ~ SKI Sl;;. ~

I 26 120 104 +13
2 36 782 410 +48
3 49 2430 1423 +42
4 51 1566 913 +47
5 75 1145 692 +36
6 93 215 126 +42
7 106 7463 5429 +23
8 1I5 11919 8707 +30
9 307 2713 1675 +39
10 317 257'90 J031St +60

Table V. Total ceLIs claimed.

fro,ram ~ W Sl;;. ~
I 26 65 90 -39
! 36 851 1235 -46
I 49 3300 3626 -10
4 51 1446 1897 -32
1 75 887 967 -9
6 93 199 168 +16
1 106 7463 6108 +19
8 1/5 8337 7728 +8
9 307 2787 3363 -21
IG 317 27220' 177712 +35

Table Vl Run-time in seconds.

Pr!!i.Wll Si.l&. W Sl;;. ~

1 26 0 0 0
2 36 2 2 0
l 49 8 7 +12
4 51 2 2 0
5 75 3 3 +3
6 93 1 I 0
7 106 14 II +21
S 1I5 18 16 +11
9 307 6 5 +17
10 317 544 299 +45
A.1 References
(HLllhc.s80] R.J.M.Hulhes. The DClil" and Implemenu.tion or an Applicathe L;anluale.
Cambrldlt Uni'clsily Diplollli Dluc.rtation.

[Landin66) p.J.Landin. The N~t 700 Prolnmmlnl Lanlul.et. Cornmunlcatioru of the "eM,
9(l). 157-16t. March 1966.

[Turnu16] D.A.Turner, SASl. Llnlullc Manual, 5t.Andrew'. Unhcnity.

[Turner79al D.A.Tu'ncr. A New Jmph:mentation Tecbnique for Applil:atin LanlUllts. Sohware:

practice and EJ.perieftl:.t, Vol. 9.

[Turner79b] DA.Turntr, Another Allorithm for Brackd Ab'trlcUon. Journal of Symbolic LOlk,
44(2). June 1979.

[PAL rd]