prev

next

out of 32

View

212Download

0

Embed Size (px)

Improvement Theory and itsApplications

David Sands Chalmers University and the University of GoteborgAbstract

An improvement theory is a variant of the standard theories of observa-tional approximation (or equivalence) in which the basic observations madeof a functional programs execution include some intensional information about,for example, the programs computational cost. One program is an improve-ment of another if its execution is more efficient in any program context. Inthis article we give an overview of our work on the theory and applications ofimprovement. Applications include reasoning about time properties of func-tional programs, and proving the correctness of program transformation meth-ods. We also introduce a new application, in the form of some bisimulation-like proof techniques for equivalence, with something of the flavour of San-giorgis bisimulation up-to expansion and context.

1 Introduction

An improvement theory is a variant of the standard theories of observational ap-proximation (or equivalence) in which the basic observations made of a programsexecution include some intensional information about, for example, the programscomputational cost. One program is an improvement of another if its execution ismore efficient in any program context.

In this article we survey a number of applications of a particular improvementtheory developed for a small untyped functional language.

The initial motivation for considering such relations was to support a simple cal-culus for reasoning about time behaviour of non-strict functional programs (Sands1993). From this work it was a natural step to consider a more general class ofimprovement theories for functional languages, providing contextually robust the-ories of optimisation (Sands 1991).

However, the real pay-off for this study is not in reasoning about efficiency prop-erties per se, but in reasoning about equivalence: The Improvement Theorem (Sands1995b) provides a condition for the total correctness of transformations on recur-sive programs. Roughly speaking, the Improvement Theorem says that if the localsteps of a transformation are contained in a particular improvement theory, thencorrectness of the transformation follows. This result has furnished:

To appear: Eds. A. Gordon and A Pitts, Higher-Order Operational Techniques in Semantics,Cambridge University Press, 1997.

1

2 Sands

an equivalence-preserving variant of the classic Unfold-Fold transformationin a higher-order functional languages (Sands 1995b), and the first correctness proofs for a number of well-known transformation meth-ods(Sands 1996a) (see also (Sands 1995a)).

New work reported in this article is the development of some proof techniques forequivalence based on the improvement theory, with something of the flavour ofSangiorgis bisimulation up-to expansion and context.

Overview

In the next section we introduce the syntax and semantics of the language used inthe remainder of the article. Section 3 introduces a simple improvement theory forthe language based on observing the number of recursive calls made during evalu-ation. We introduce proof techniques for establishing improvement, and introducethe tick algebra which facilitates the calculation of simple improvement properties.Section 4 considers the application which originally motivated the study of im-provement relations, namely the problem of time analysis of functional programs.Section 5 describes a correctness problem in program transformation. We presentThe Improvement Theorem, which shows, in principle, how the problem can besolved by using improvement. Section 6 illustrates this application by consider-ing a particular program transformation method. Finally, Section 7 introduces andillustrates some proof techniques for the equivalence of programs which can be de-rived from the Improvement Theorem.

2 Preliminaries

We summarise some of the notation used in specifying the language and its oper-ational semantics. The subject of this study will be an untyped higher-order non-strict functional language with lazy data-constructors.

We assume a flat collection of mutually recursive definitions of constants, eachof the form

, where is a closed expression. Symbols , , , range overthe constants, over variables and , , over expressions. Thesyntax of expressions is as follows: (Variable; Recursive constant; Application) (Lambda-abstraction) !"$# &% ')(* + '(*-, + ,(Case expressions)/.10213 (Constructor expressions and constants)54023 (Strict primitive functions)')(* .1062 3 (Patterns)

Improvement Theory and its Applications 3

Somewhat improperly we will refer to the constants, , etc., as function names,

or function calls. In general they need not be functions however.We assume that each constructor . and each primitive function 4 has a fixed ar-

ity, and that the constructors include constants (i.e. constructors of arity zero). Con-stants will be written as . rather than .10 3 . The primitives and constructors are notcurried they cannot be written without their full complement of operands. Weassume that the primitive functions map constants to constants.

We can assume that the case expressions are defined for any subset of patterns7 '(* ')(* ,98 such that the constructors of the patterns are distinct. A variablecan occur at most once in a given pattern; the number of variables must match thearity of the constructor, and these variables are considered to be bound in the cor-responding branch of the case-expression.

A list of zero or more expressions , will often be denoted 2 . Application,as is usual, associates to the left, so 0$0;:::0

4 Sands

The operational semantics is call-by-name, and H is defined in terms of a one-stepevaluation relation using the notion of a reduction context (Felleisen, Friedman, andKohlbecker 1987). If HNI for some closed expression then we say that evaluatesto I . We say that converges, and sometimes write H if there exists a I such that HNI . Otherwise we say that diverges. We make no finer distinctions betweendivergent expressions, so that run-time errors and infinite loops are identified.

Reduction contexts, ranged over by O P , are contexts containing a single holewhich is used to identify the next expression to be evaluated (reduced).

Definition 1 A reduction context O P is given inductively by the following grammarO P DE QO P R !";#>O P % ')(* + 1 ')(* , + , 4S0;2.6O PJ123The reduction context for primitive functions forces left-to-right evaluation of thearguments. This is just a matter of convenience to make the one-step evaluationrelation deterministic.

Now we define the one step reduction relation. We assume that each primitivefunction 4 is given meaning by a partial function D D 4 E E from vectors of constants (ac-cording to the arity of 4 ) to the constants (nullary constructors). We do not needto specify the exact set of primitive functions; it will suffice to note that they arestrictall operands must evaluate to constants before the result of an application,if any, can be returned and are only defined over constants, not over arbitraryweak head normal forms.

Definition 2 One-step reduction TU is the least relation on closed expressions sat-isfying the rules given in Figure 1.

O P D E TU O P D E (fun)0 if is defined by V3O P D 0V 13)6? E TU O P D 7 J= W? 8 E ( X )O P D !";#Y.Z-023N% .Z-062[Z 3 + ZQ E TU O P D Z 7 2ZQ= 2 8 E (case)O P D 40;2. 3 E TU O P D . ? E (prim)0 if D D 4 E E 2. . ? 3Figure 1: One-step reduction rules

In each rule of the form O P D E TU O P D 6? E in Figure 1, the expression is referred toas a redex. The one step evaluation relation is deterministic; this relies on the factthat if 1 TU then can be uniquely factored into a reduction context O P and aredex W? such that 1 O P D W? E . Let TU]\ denote the transitive reflexive closure of TU .

Improvement Theory and its Applications 5

Definition 3 Closed expression converges to weak head normal form I , HNI , ifand only if TU]\I .Using this notion of convergence we now define the standard notions of operationalapproximation and equivalence. We use is the standard Morris-style contextual or-dering, or observational approximation see e.g. (Plotkin 1975).The notion of ob-servation we take is just the fact of convergence, as in the lazy lambda calculus(Abramsky 1990). Operational equivalence equates two expressions if and only ifin all closing contexts they give rise to the same observation i.e. either they bothconverge, or they both diverge.

Definition 4 1. operationally approximates ^? , G_` W? , if for all contexts Csuch that C D E , C D 6? E are closed, if C D E H then C D ? E H .

2. is operationally equivalent to ? , ` W? , if >_` W? and 6?_` .Choosing to observe, say, only computations which produce constants would giverise to slightly weaker versions of operational approximation and equivalence - butthe above versions would still be sound for reasoning about the weaker variants ofthe relation.

3 A Theory of Improvement

In this section we introduce a theory of improvement, as a refinement of the the-ory of operational approximation. Roughly speaking, improvement is a refinementof operational approximation which expression is improved by ? if, in all clos-ing contexts, computation using ^? is no less efficient than when using , measuredin terms of the number of function calls (

, , etc.) made. From the point of view

of applications of the theory to program transformation and analysis, the impor-tant property of improvement is that it is a contextual congruencean expressioncan be improved by improving a sub-expression. For reasoning about improvementa more tractable