Type-sensitive Control-flow Analysis

John Reppy
University of Chicago
[email protected]

Abstract analysis (CFA) algorithms to enable more advanced optimiza-

Higher-order typed languages, such as ML, provide strong sup- tions [Shi88]. One weakness of CFA is that one must be conser-
port for data and type abstraction. While such abstraction is of- vative about values that escape from the unit of analysis.2 In this
ten viewed as costing performance, there are situations where it paper, we present a modification of Serrano’s CFA [Ser95], that
may provide opportunities for more aggressive program optimiza- exploits type abstraction to improve the accuracy of analysis (and
tion. Specifically, we can exploit the fact that type abstraction guar- thus enable more optimizations). While our approach is presented
antees representation independence, which allows the compiler to in the context of a specific analysis algorithm, we believe that it
specialize data representations. This paper describes a first step in is a general technique that can integrated into other flavors of flow
supporting such optimizations; namely a control-flow analysis that analysis for higher-order languages. Unlike many program analy-
uses the program’s type information to compute more precise re- ses, which are whole-program analyses, our analysis is modular by
sults. We present our algorithm as an extension of Serrano’s version design.
of 0-CFA and we show that it respects types. We also discuss appli- Our analysis is based on the observation that if a type is abstract
cations of the analysis with examples of optimizations enabled by outside the scope of a module, then any value of that type must have
the analysis that would not be possible using normal CFA. been created inside the module at one of a statically known set of
program points. Furthermore, the value can only be deconstructed
Categories and Subject Descriptors D.3.4 [Processors]: Com- at statically known program points in the module. We exploit this
pilers, Optimization; D.3.2 [Language Classifications]: Applica- fact by computing for each abstract type an approximation of the
tive (functional) languages; D.3.3 [Language Constructs and Fea- values that escape the module into the wild, and thus might return.
tures]: Abstract data types, Modules, packages We use this approximation when confronted with an unknown
value of the given type. To extend this mechanism to aggragate
General Terms Languages values and higher-order functions, we also use a more refined
Keywords ML, Control-flow analysis representation of top values that is indexed by type. For example,
the result of calling an unknown function that returns a value of
an abstract type can be approximated by the escaping values of
1. Introduction that type. While these changes have the potential to significantly
One of distinguishing characteristics of the ML family of lan- improve the quality of information provided by an analysis, they
guages is the module system, which provides strong support for do not significantly add to the complexity of the implementation.
data and type abstraction, as well as for modular programming. The remainder of this paper is organized as follows. In the next
While such abstraction is often been considered as a source of over- section, we briefly review Serrano’s version of CFA, which serves
head that reduces program performance,1 it has the important ben- as the basis of our algorithm, and give an example of how exploiting
efit of isolating clients of interfaces from implementation details, knowledge of type abstraction can produce a better result. Then,
such as data representation. What is somewhat surprising is that, in Section 3, we present our type-sensitive analysis algorithm. We
to the best of our knowledge, existing implementations of ML-like follow this with a small, but detailed, example of how the analysis
languages do not take advantage of abstractions in the source pro- works. In Section 5, we discuss the correctness of the analysis. We
gram to optimize data representations, etc. We are interested in such present our algorithm using a stripped down version of Core SML.
optimizations, particularly in the case of CML-style concurrency In Section 6, we discuss how the algorithm can be extended to cover
abstractions [Rep91, Rep99]. a greater fraction of ML features. Our motivation for this research
Control-flow information is not syntactically visible in higher- is the optimization of CML-style concurrency mechanisms and we
order languages, such as ML, so compilers must use control-flow discuss that application as well as others in Section 7. We review
related work in Section 8 and conclude in Section 9.
1 For example, Andrew Appel reports that switching to a concrete repre-
sentation of graphs in the Tiger compiler improved performance of liveness
analysis by almost a factor of about ten [App98]. 2. Background
In this section, we give a brief overview of Serrano’s version of
CFA [Ser95], which we use as the basis of our algorithm3 and we
motivate our improvements.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation 2 One
on the first page. To copy otherwise, to republish, to post on servers or to redistribute solution is whole-program analysis, but that does not scale well to
to lists, requires prior specific permission and/or a fee. large code bases.
ML’06 September 16, 2006, Portland, Oregon, USA. 3 Our approach does not depend on the specifics of Serrano’s algorithm and
Copyright c 2006 ACM 1-59593-483-9/06/0009. . . $5.00. should be applicable to other versions of CFA.
 Serrano’s analysis computes an imperative mapping A that
p ::= e
maps variables to their approximate values. The approximate val-
| d; p
ues are ⊥, which denotes an approximation not yet computed, >,
which denotes an unknown value, and finite sets of function names. d ::= abstype T = C of τ with fb 1 · · · fb n end
It is straightforward to extend this set of values with basic constants
(e.g., booleans and integers) and data constructors (e.g., tuples). fb ::= fun f (x) = e
A function is said to escape if it can be called at unknown sites.
There are several ways that a function might escape: it may be e ::= x
exported by the module, it may be passed as an argument to an | let x = e1 in e2
unknown function, or it may be returned as the result of an escaping | fun f (x) = e1 in e2
function. If a function f escapes, then A must be defined to map | e1 e2
f ’s parameters to >, since we have no way of approximating them. | Ce
But in a program that involves abstract types, we can exploit the | let C x = e1 in e2
type abstraction to compute better approximations. For example, | he1 , e2 i
consider the following SML code: | fst(e)
signature SIG =
| snd(e)
sig
type t τ ::= T
val new : int -> t | τ1 → τ2
val pick : t -> int | τ1 × τ2
end
Figure 1. A simple language
structure M :> SIG =
struct
datatype t = C of int list
fun new x = C[x]
fun pick (C(y::_)) = y plication and deconstruction,5 and pair construction and projection.
end Types include abstract types (T), function types, and pair types.
Abstract types are either predefined types (e.g., unit, int, bool,
The programmer knows that the pick function can never raise a etc.) or are defined by an abstype definition.
Match exception, even though it is defined using a non-exhaustive We assume that variables, abstract-type names, and data-constructor
match. Without a whole-program analysis, existing compilers will names are globally unique. We also assume that variables and con-
not recognize this fact. Our analysis, however, uses the fact that structors are annotated with their type. We omit this type informa-
type t is abstract, and will detect the fact that any value passed to tion most of the time for the sake of brevity, but, when necessary,
pick as an argument must have been created by new. we write it as a superscript (e.g., xτ ). One should think of this lan-
guage as a compiler’s intermediate representation produced by the
typechecking phase.
3. Type-sensitive CFA
We are now ready to present our type-sensitive CFA algorithm. Our 3.2 Approximate values
main modifications to Serrano’s algorithm [Ser95] are:
The analysis computes a mapping from variables to approximate
• we use a more refined representation of approximate values values, which are given by the following grammar:
• we compute an additional approximation T that tracks escaping
values of abstract type. vb ::= ⊥
| C(b v)
• our algorithm is presented as a functional program without side
| hb
v 1 , vb2 i
effects. | F
| Tb
3.1 Preliminaries | τ1\ → τ2
We present our algorithm in the context of a small typed language. | >
This language is a monomorphic subset of Core SML [MTHM97]
with explicit types4 . Standard ML and other ML-like languages use where C is a data constructor, F is a finite set of function names,
modules to organize code and signature ascription to define abstrac- and T is an abstract type. We use ⊥ to denote undefined or not yet
tion. For this paper, we use the abstype definition to define abstrac- computed values, C(b v ) for an approximate value constructed by
tions in lieu of modules. We further simplify this definition form to applying C to vb, hb
v 1 , vb2 i for an approximate pair, and F for a set of
only have a single data constructor. Figure 1 gives the abstract syn- known functions. Our analysis will only compute sets of functions
tax for this simple language. The top-level abstype definition is the F where all the members have the same type (see Section 5 for a
unit of analysis (a program consists of a collection of these). Each proof of this property) and so we extend our type annotation syntax
abstype definition defines a new abstract type (T ) and correspond- to include such sets. In addition to the single top value found in
ing data constructor (C) and a collection of functions (fb i ). Outside most presentations of CFA, we have a family of top values (b τ)
the abstype definition, the type T is abstract (i.e., the data construc- indexed by type. The value τb represents an unknown value of type
tor C is not in scope). The expression forms include let-bindings, τ (where τ is either a function or abstract type). The auxiliary
nested function bindings, function application, data-constructor ap- function U : T YPE → VALUE maps types to their corresponding

4 We discuss handling polymorphic types and user-defined type constructors 5 Ina language with sum types, deconstruction would be replaced by a case
in Section 6. expression.
unknown value: “e” is an object-language syntactic form and V[x 7→ v] to denote
the functional update of an approximation (likewise for R and T ).
U(T) = Tb
Our unit of analysis is a single abstype definition. We use LVAR
U(τ1 → τ2 ) = τ1\→ τ2 to denote the set of variables defined in the definition, GVAR to de-
U (hτ1 , τ2 i) = hU(τ1 ), U (τn )i note variables defined elsewhere, and VAR = LVAR ∪ GVAR for
Note that for unknown pair values, we preserve the fact that they all variables defined or mentioned in the definition being analyzed.
have a pair structure. Lastly, the > value is used to cutoff expansion We denote the known function identifiers by F UN I D ⊂ LVAR (i.e.,
of recursive types as described below. those variables that are defined by function bindings). These in-
We define the join of two approximate values as follows: clude the top-level function bindings in the definition, as well as
any nested function definitions. Our algorithm analyses the func-
⊥ ∨ vb = vb tion definitions in the declaration repeatedly until a fixed-point is
vb ∨ ⊥ = vb reached. The initial approximations map local variables, function
C(b v 1 ) ∨ C(b v2 ) = C(b v 1 ∨ vb2 ) results, and abstract types to ⊥, and map global variables and ex-
v 1 , vb2 i ∨ hvb0 1 , vb0 2 i = hb
hb v 1 ∨ vb0 1 , vb2 ∨ vb0 2 i ternal types to unknown values.
F ∨F0 = F ∪ F0
> ∨ vb = > fun cfa [[abstype T = C of τ with fb 1 · · · fb n end]] =
τb ∨ vb = τb let
vb ∨ τb = τb fun iterate A0 = let
val A1 = cfaFB (A0 , fb 1 )
vb ∨ > = > ···
Note that this operation is not total, but it is defined for any two val An = cfaFB (An−1 , fb n )
approximate values of the same type and we show in Section 5 that in
if (A0 6= An )
it preserves types.
then iterate An
One technical complication is that we need to keep our approx- else A0
imate values finite. For example, consider the following pathologi- end
cal example: let V = {x 7→ ⊥ | x ∈ LVAR \ F UN I D}
∪ {f 7→ {f } | f ∈ F UN I D}
abstype T = C of T with fun f (x) = C x end ∪ {x 7→ U(τ ) | xτ ∈ GVAR}
If we are not careful, our analysis might diverge computing ever let R = {f 7→ ⊥ | f ∈ F UN I D}
larger approximations of C ∞ (⊥) as the result of f . To avoid this let T = {T 7→ ⊥} ∪ {S 7→ S b | S ∈ (A BSTY \ {T })}
in
problem, we define a limit on the depth of approximations for
iterate (V, R, T )
recursive types as follows: end
d⊥eC = ⊥

> if C ∈ C The cfaFB function analyses a function binding in the abstype
τ →T
˚ ˇ
C (b
v) C = C(dbv eC∪{C} ) if C 6∈ C definition by “applying” the function to the unknown value of
the function’s argument type. The result of the application is then
dhb
v1 , vb2 ieC = h db
v1 eC , db
v2 eC i
recorded as escaping.
dF eC = F
db
τ eC = τb fun cfaFB (A, [[fun f (xτ ) = e]]) = let
val (A, vb) = applyFun ({}, A, f , U(τ ))
where C ⊂ DATAC ON is a set of constructors. We write db v e for in
db
v e∅ . We use > to cutoff the expansion of approximate values in- escape ({}, A, v b)
stead of Tb, because the approximation of escaping values of type end
T may not be an accurate approximation of the nested values. This
definition does not allow nested applications of the same construc- The applyFun function analyses the application of a known
tor. For example, the analysis will be forced to approximate the function f to an approximate value vb. The first argument to
escaping values of type T by C(>) in the above example. applyFun is a set M ∈ 2F UN I D of known functions that are
currently being analysed; if f is in this set, then we use the ap-
3.3 CFA proximation R instead of recursively analysing the f ’s body. This
mechanism is necessary to guarantee termination when analysing
Our analysis algorithm computes a triple of approximations: A = recursive functions. We assume the existence of a helper function
(V, R, T ), where bindingOf that maps known function names to their bindings
in the source. Once we have computed the approximate result (r)
V ∈ VAR → V\ ALUE variable approximation
of evaluating the function’s body, we add that information to the
R ∈ F UN I D → V\
ALUE function-result approximation result approximation.
T ∈ A BSTY → V\ ALUE escaping abstract-value
approximation fun applyFun (M, A as (V,R,T ), f , v) =
if f ∈ M
Our V approximation corresponds to Serrano’s A. The R ap- then (A, R(f ))
proximation records an approximation of function results for each else let
known function; this approximation is used in lieu of analysing a val [[fun f (x) = e]] = bindingOf (f )
function’s body when the function is already being analysed and is val V = V[x 7→ dV(x) ∨ ve]
needed to guarantee termination. We use the T approximation to val ((V, R, T ), r) =
cfaExp (M ∪ {f }, (V,R,T ), [[e]])
interpret abstract values of the form Tb. val R = R[f 7→ dR(f ) ∨ re]
We present the analysis algorithm using SML syntax extended in
with mathematical notation such as set operations, and the ∨ oper- ((V,R,T ), r)
ation on approximate values. We use the notation [[e]] to denote that end
 fun cfaExp (M, A as (V,R,T ), [[x]]) = (A, V(x)) tions, an approximation triple, and an syntactic expression as ar-
| cfaExp (M, A, [[let x = e1 in e2 ]]) = let guments and returns updated approximations and a value that ap-
val ((V,R,T ), v b) = cfaExp (M, A, [[e1 ]])
proximates the result of the expression. For function applications,
val V = V[x 7→ dV(x) ∨ v be]
in
we use the apply helper function (discussed below) and for value
cfaExp (M, (V, R, T ), [[e2 ]]) deconstruction, we use the decon helper function, which handles
end the deconstruction of approximate values and their binding to vari-
| cfaExp (M, A, [[fun f (x) = e1 in e2 ]]) = ables. When the value is unknown (i.e., Tb), then we use the T ap-
cfaExp (M, A, [[e2 ]]) proximation to determine the value being deconstructed.
| cfaExp (M, A, [[e1 e2 ]]) = let
val (A, v b1 ) = cfaExp (M, A, [[e1 ]]) fun decon (V, T , [[C(x)]], C(b v )) = V[x 7→ dV(x) ∨ v
be]]
val (A, v b2 ) = cfaExp (M, A, [[e2 ]]) | decon (V, T , [[C τ →T (x)]], Tb) = (case T (T )
in of Tb => V[x 7→ dV(x) ∨ U (τ )e]
apply (M, A, v b1 , v
b2 ) b => decon(V, T , [[C(x)]], v
| v b)
end (* end case *))
| cfaExp (M, A, [[C e]]) = let | decon (V, T , [[C(x)]], ⊥) = V
val (A, v b) = cfaExp (M, A, [[e]]) | decon (V, T , [[C(x)]], >) = V[x 7→ >]
in
(A, C(bv ))
end
The apply function records the fact that an approximate func-
| cfaExp (M, A, [[let C x = e1 in e2 ]]) = let tion value is being applied to a approximate argument. When the
val ((V,R,T ), v b) = cfaExp (M, A, [[e1 ]]) approximation is a set of known functions, then we apply each
val V = decon (V, T , [[C x]], v b) function in the set to the argument compute the join of the results.
in When the function is unknown (i.e., a top value), then the argu-
cfaExp (M, (V,R,T ), [[e2 ]]) ment is marked as escaping and the result is the top value for the
end function’s range.
| cfaExp (M, A, [[he1 , e2 i]]) = let
val (A, v b1 ) = cfaExp (M, A, [[e1 ]]) fun apply (M, A, F , arg) = let
val (A, v b2 ) = cfaExp (M, A, [[e2 ]]) fun applyf (f , (A, res)) = let
in val (A, v b) = applyFun (M, A, f , arg)
(A, hb b2 i)
v1 , v in
end (A, res ∨ v
b)
| cfaExp (M, A, [[fst(e)]]) = ( end
case cfaExp (M, A, [[e]]) in
of (A, hb b2 i) => (A, v
v1 , v b2 ) fold applyf (V, T ) F
| (A, vb) => (A, v b) end
(* end case *)) | apply (M, A, τ1\ → τ2 , v
b) = let
| cfaExp (M, A, [[snd(e)]]) = ( val A = escape(M, A, v b)
case cfaExp (M, A, [[e]]) in
of (A, hb b2 i) => (A, v
v1 , v b1 ) (A, τb2 )
| (A, vb) => (A, v b) end
(* end case *)) | apply (M, A, ⊥, v b) = (A, ⊥)
| apply (M, A, >, v b) = let
Figure 2. CFA for expressions val A = escape(M, A, v b)
in
(A, >)
end
The escape function records the fact that a value escapes into
the wild. If the value has an abstract type, then it is added to the
approximation of wild values for the type; if it is a set of known
functions, then we apply each function in the set to the appropriate
top value; and if it is a pair, we record that its subcomponents 4. An example
are escaping. The escape function also takes the set of currently To illustrate the analysis, we step through its application to the
active functions as its first argument, since it needs to pass this following small example:6
value to the applyFun function.
abstype t = C of (int * int)
fun escape (_, (V,R,T ), C(b v )) = with
(V, R, T [T 7→ dT (T ) ∨ C(b v )e]) fun new x = C(1, x)
| escape (M, A, F ) = let fun pick y = let C(z) = y in fst(z)
fun esc (f τ1 → τ2 , A) = let end
val (A, vb) = applyFun(M, A, f , U(τ1 ))
in A end
This code has two functions and three other local variables:
in
fold esc A F
end
F UN I D = {new, pick}
| escape (M, A, hb b2 i) = let
v1 , , v LVAR = {x, y, z} ∪ F UN I D
val A1 = escape (M, A, v b1 ) GVAR = {}
val A2 = escape (M, A1 , v b2 )
in A2 end
| escape (_, A, v b) = A
6 This example is a variation of the one given in Section 2. We have also
Expressions are analysed by the cfaExp function, whose code taken the liberty of adding integer constants to our language and to the
is given in Figure 2. This function takes the set of active func- representation of approximate values.
The initial approximations are so the final result is:
A0 = (V0 , R0 , T0 ) V(x) = int
d
V0 = {new 7→ {new}, pick 7→ {pick}} V(y) = C(1, int)
d
∪ {x 7→ ⊥, y 7→ ⊥, z 7→ ⊥} V(z) = (1, int)
d
R0 = {f 7→ ⊥| f ∈ F UN I D} R(new) = C(1, int)
d
T0 = {t 7→ ⊥} ∪ {S 7→ Sb | S 6= t} R(pick) = 1
The cfa function will apply cfaFB to each of the bindings. We T (t) = C(1, int)
d
start with new and compute
b1 ) = applyFun({}, A0 , new, int)
(A1 , v d
5. Correctness of the analysis
Computing this application of applyFun involves computing The correctness of our analysis can be judged on several dimen-
cfaExp ({new}, (V1 , R0 , T0 ), [[C(1, x)]]) sions. First, there is the question of safety: does the analysis com-
pute an approximation of the actual computation? Second is the
where V1 = V0 [x 7→ int].d The result of analyzing the body of question of whether the algorithm terminates? The third correct-
new will be ((V1 , R0 , T0 ), C(1, int)),
d thus we have ness issue is the question of whether the approximations computed
by the analysis are faithful to the type of the program. This ques-
A1 = (V1 , R1 , T0 ) tion is important, since our analysis is guided by type information
in a number of situations. We discuss the first and third of these
R1 = {new 7→ C(1, int),
d pick 7→ ⊥}
questions in the remainder of this section.
vb1 = C(1, int)
d
5.1 Safety
The last step in analyzing new is to compute We postulate that our analysis is safe, i.e., that if it computes a
escape ({}, A1 , v
b1 ) value approximation V, then for any variable x ∈ dom(V) and any
execution of the program, the values taken on by x will be covered
which results in an enriched approximation that records the escap- by the approximate value V(x). One can formalize this statement
ing value of type t. in terms of a collecting semantics [CC77, Shi91] and prove it using
standard techniques, but we will make a less formal argument here.
A2 = (V1 , R1 , T1 ) The core of our analysis is the well known 0-CFA, which has been
T1 = {t 7→ C(1, int)}
d proven correct in a number of papers, but we have extended this
analysis with the T approximation for tracking abstract values that
Now we are ready to analyse the pick function binding, which escape to the wild. For a given abstract-type definition
means computing
abstype T = C of τ in ... end
b3 ) = applyFun({}, A2 , pick, t
(A3 , v b)
the ML type system restricts the scope of C to the “...”; thus,
Let V2 = V1 [y 7→ tb], then we evaluate the body of pick with the values of type T can only be constructed/deconstructed in the body
initial approximations (V2 , R1 , T1 ). The interesting case is when of the definition. Therefore, we claim that if the T approximation
cfaExp gets to the deconstruction of the value bound to y. In this computed by our analysis is “safe,” then our analysis is correct.
case, we must compute There are two aspects to the safety of T : first, does it correctly
approximate the values that escape and second, does the analysis
decon (V2 , T1 , [[C(z)]], t
b) correctly identify all possible places where escaping values could
reenter the definition? There are only two ways for values to escape
This case is handled by the second clause of the decon function, the definition: they can be returned in the result of one of the
which applies T1 to t, producing C(1, int),
d which results in the operations or they can be passed as an argument to an external
recursive call of decon or unknown function.7 The first of these cases is covered by the
call to escape in cfaFB, while the second is covered by the
decon (V2 , T1 , [[C(z)]], C(1, int))
d
call to escape in in apply. Thus, we claim that any escaping
The recursive call is handled by the first clause of decon, which abstract value in any possible execution will be covered by the
returns the augmented value approximation T approximation computed by our analysis. There are also only
two ways for escaping values to enter the definition: they can be
V3 = V2 [z 7→ (1, int)]
d passed in an argument to one of the definition’s operations or they
can be returned by a call to an external or unknown function. In
This approximation will be used in the analysis of [[fst(z)]], both of these situations, we use the approximate value U(τ ) to
which will produce 1 as its approximate result. Thus, the result represent unknown values of type τ . If such a value propagates to a
of applyFun on pick is deconstruction site, then we use the T approximation to determine
the values bound to the pattern variables.
A3 = (V2 , T1 , R2 )
R2 = R1 [pick 7→ 1] 5.2 Termination
vb3 = 1 The question of termination has been addressed by Serrano [Ser95]
(the bounding of the sizes of abstract values is crucial to the termi-
The escape function will not change the approximations in this
case, so A3 is the result of the first iteration over the function 7 Ifour language had references, then assignment would be another way for
bindings. It is also the fixedpoint of the analysis for this example, values to escape.
 ` fb 1 : Ok · · · ` fb n : Ok ` τ : Type
` abstype T = C τ →T of τ with fb 1 · · · fb n end : Ok ` ⊥: τ

` e : τ2 ` vb : τ
` fun f τ1 →τ2 (xτ1 ) = e : Ok ` C τ →T (b
v) : T
` vb1 : τ1 ` vb2 : τ2
` xτ : τ ` hbv1 , vb2 i : τ1 × τ2
` e1 : τ1 ` e2 : τ2 ` f : τ → τ 0 for all f ∈ F
` let xτ1 = e1 in e2 : τ2 ` F : τ → τ0
` fun f (x) = e1 : Ok ` e2 : τ
` fun f (x) = e1 in e2 : τ ` τb : τ

` e1 : τ2 → τ ` e2 : τ2 ` τ : Type
` e1 e2 : τ ` > : τ

` e : τ Figure 4. Typing rules for approximate values

` C τ →T e : T
` e1 : T ` e2 : τ2 fragment, x ∈ dom(V) and ` V(x) : τ . A function-result ap-
proximation R is type consistent with respect to a fragment if for
` let C τ →T xτ = e1 in e2 : τ2 all functions f τ1 →τ2 defined in the fragment, f ∈ dom(R) and
` e1 : τ1 ` e2 : τ2 ` R(f ) : τ2 . Likewise, an approximation T is type consistent
with respect to a fragment if for all type names defined or men-
` he1 , e2 i : τ1 × τ2
tioned in the fragment, T ∈ dom(T ) and ` T (T ) : T . An ap-
` e : τ1 × τ2 proximation triple A = (V, R, T ) is type consistent if its compo-
` fst(e) : τ1 nents are type consistent.
Our main result is that cfaExp computes type correct approxi-
` e : τ1 × τ2 mations (or respects types), but we need the following lemma first,
` snd(e) : τ2 which says that the ∨ operator preserves types.
L EMMA 1. If ` vb1 : τ and ` vb2 : τ , then ` vb1 ∨ vb2 : τ .
Figure 3. Typing rules for declarations, function bindings, and
expressions Proof. The proof two parts. First, we can prove that if ` vb1 : τ
and ` vb2 : τ , then vb1 ∨ vb2 is defined (recall that ∨ is partial) by
case analysis of the shape of τ and the typing rules. Then, a simple
nation of the fixed-point iteration) and his argument applies directly inductive argument shows that ` vb1 ∨ vb2 : τ
to our version of his algorithm.
A related question is does the analysis avoid uncaught runtime T HEOREM 1. Given an expression e, with ` e : τ , and a type
exceptions? The two helper functions decon and apply are both consistent approximation triple A, if cfaExp({}, A, e) returns
partial, so the possibility of a Match exception exists (for the the result (A0 , vb), then A0 is type consistent and ` vb : τ .
fourth argument of decon and the third argument of apply). We
Proof. This theorem is proven by induction on the structure of the
claim that such exceptions will not occur. Because the program
expressions. The induction is well-founded because the analysis
being analyzed is type correct and based on the type correctness
never examines a function’s body if it is already being analysed.
of the analysis (discussed in the next section), we see that any type-
correct approximate value that might be passed in as an argument Since the initial approximations in cfa are type consistent, the
is covered by the functions’ cases. above theorem guarantees that the resulting approximation will be
type consistent.
5.3 Type correctness of the analysis
Since we assume that programs have already been typechecked and 6. Extensions
that variables and data-constructors are annotated with their types,
the typing rules for our language do not require a context. We have We presented our algorithm as a modification of Serrano’s 0-CFA,
three judgment forms: but we see no reason why our techniques will not extend to similar
program analyses, such as sub-zero CFA [AD98] or 1-CFA [Shi91].
` d : Ok the declaration d is well-typed We have also used a greatly reduced subset of Core SML to
` fb : Ok the function binding fb is well-typed present our approach. In the remainder of this section, we discuss
` e : τ the expression e is well-typed and has type τ extensions of the analysis to other ML features.
The typing rules for our language are straightforward and are given
in Figure 3. 6.1 Modules
We also define typing rules for approximate values (again anno- Our example language uses the abstype declaration as its unit
tations take the place of context). These rules are given in Figure 4. of modularity and analysis. This feature has been depreciated in
Note that the ⊥ and > values can have any well-formed type. SML in favor of using modules and signature ascription, so to
We say that a variable approximation V is type consistent with extend our technique to the full ML language, we need to have
respect to a code fragment (i.e., declaration, function binding, or a different way of identifying abstract types. The basic goal is to
expression), if for all variables xτ defined or mentioned in the have the T approximation return U (T ) when T is not abstract in
the module’s signature, since clients of the module can construct We also extend the domain of the T approximation to include
values of type T using the constructor C. We see two ways to abstract-type constructor applications. We believe that tracking
achieve this behavior. The first is to make the initialization of the T each distinct application of an abstract-type constructor will be
approximation map depend on the module’s signature. If a datatype useful in optimizing modules that use phantom types in their inter-
has the specification faces.
datatype T = C of τ
Polymorphism is also a form of type abstraction, which we
should exploit when analyzing calls to external or unknown func-
in the signature, then we initialize T (T ) to U(T ), instead of the ⊥. tions. For example, the map function on lists has the type
The other approach would be to view the constructors of T as
val map : (’a -> ’b) -> ’a list -> ’b list
escaping functions, which has the same effect of T . This approach
also handles the situation where a constructor is passed as a value From this type, we know that the map function can only manipulate
to an external function. the items in the argument and result lists by using its first argument.
A more difficult issue is that abstract types can be defined Thus, a call to map like the one for the union function in the
without data constructors using opaque signature matching. For previous section
example,
List.map (fn ABS s => s) ss
structure S :> sig
type set does not actually expose the representation of the abstract values.
val singleton : int -> set
val items : set -> int list 6.3 Datatypes
val member : (set * int) -> bool The abstype declarations in our language are limited to single
...
end = struct
constructors. Extending the language with ML datatypes suggests
type set = int list a corresponding extension of approximate values to include sets of
fun singleton x = [x] tagged values. The analysis then takes on some of the aspects of
fun items s = s shape analysis [NNH99, Section 2.6] and one might want a cutoff
fun member (s, x) = on recursive values that is more generous.
List.exists (fn y => (x = y)) s
...
end 7. Applications
In this situation there are no syntactic signposts to mark the type Serrano’s paper [Ser95] describes a number of applications of CFA
abstractions and some uses of the type “int list” are abstract in a S CHEME compiler. Many of these applications, such as iden-
and some are concrete. One possible solution is to use the module’s tifying known functions, are also useful in ML compilers. Our al-
signature as a guide to add annotations to the module’s code that gorithm will provide as good, or better information without a sig-
mark the abstractions. This process is similar to Leroy’s type- nificant increase in implementation complexity. Furthermore, there
directed unboxing [Ler92] and Grossman, et al.’s syntatic type are situations where we believe our techniques are necessary to
abstraction [GMZ00]. For example, the above code would become achieve the desired optimizations. These situations can be divided
into roughly two kinds. The first are situations where we want to
fun singleton x = ABS[x] identify all uses of a value, which can allow us to specialize the
fun member (ABS s, x) = value’s representation. For example, data structures might be flat-
List.exists (fn y => (x = y)) s
tened, which reduces space and improves access time, and fields
where ABS marks the abstraction barriers. As is the case with might be packed (e.g., storing booleans in a byte instead of a word),
Leroy’s technique, aggregates of abstract values are problematic. which reduces space. One can also reorganize the order of fields to
For example, if we had improve cache performance [CDL99]. The second case is when we
want to identify the possible definitions of an argument to an oper-
val union : set list -> set
ation, which can allow us to specialize the operation. For example,
in the signature with the implementation we may be able to eliminate conditionals or use more efficient call-
ing conventions. In both situations, if the value is embedded in an
fun union ss = List.concat ss
abstraction, we can track its uses, even if it escapes its defining
then we would need to map the projection of ABS over the argu- module. We give an example of each of these situations below.
ment list
7.1 Optimizing CML
fun union ss = let
val ss’ = List.map (fn ABS s => s) ss Our original motivation for this algorithm was the problem of op-
in timizing Concurrent ML (CML) [Rep99]. Our goal is to statically
ABS(List.concat ss’) determine the usage patterns of communication primitives (we call
end these patterns the communication topology), which will allow us to
which introduces unnecessary computation and reduces the quality specialize them to more efficient versions. A known channel is a
of the analysis. More thought on this problem is definitely required. channel for which we know all of its use sites (i.e., all of the send
and recv operations). Our analysis can compute a static approxi-
6.2 Polymorphism and type constructors mation of the communication topology of known channels, but to
be effective, we need the type-sensitive CFA described in this paper
We presented our analysis for a language without polymorphism
or user-defined type constructors. The main impact of extending it to determine the known channels.8
to a richer type system is the representation of top values. The U For example, consider the simple service implemented in Fig-
function gets extended as follows: ure 5. The new function creates a new service, which is represented

U (α) = > 8 Another analysis technique might also work, but our type-sensitive CFA
U (→
−
τ T) = →
−
τ T
d fits well with common CML programming idioms.
 abstype serv = S of (int * int chan) chan abstype noise = N of {
in perm : Word8Array.array,
fun new () = let grad : real array
val reqCh = channel() }
fun server v = let in
val (req, replCh) = recv reqCh fun mkNoise () = let
in val perm = Word8Array.array(256, 0w0)
send(replCh, v); val grad = Array(256, 0.0)
server req ...
end in
in N{perm = perm, grad = grad}
spawn (server 0); end
S reqCh fun noise (N{perm, grad}, x) = let
end val t = x + 512.0
fun call (S ch, v) = let val b0 = (floor t) mod 256
val replCh = channel() val b1 = (b0 + 1) mod 256
in val r0 = t - Real.realFloor t
send (ch, (v, replCh)); val r1 = r0 - 1.0
recv replCh val sX = sCurve r0
end fun get i =
end Array.sub(grad,
Word8.toInt(
Figure 5. A simple service with an abstract client-server protocol Word8Array.sub(perm, i)))
val u = r0 * get b0
val v = r1 * get b1
in
lerp (sx, u, v)
end
by an abstract type, and the call function allows clients to re- end
quest the service. We have developed an analysis that can detect
that the server’s request channel (reqCh) is used by potentially
many different senders, but by only on receiver, and that the re- Figure 7. 1D Perlin Noise
ply channel allocated for a given request (replCh) is used only
once [Xia05, RX06]. These properties allow the optimizer to re-
checking,9 so array-bounds-check elimination is an important opti-
place the general channel operations with more specialized ones.
mization for this application. To make this example more concrete,
Continuing with the example, Figure 6 illustrates the flow of the
Figure 7 sketches the code for 1D Perlin noise. The key point about
server’s request channel from its allocation site in the new function
this code is that the array subscript operations used to compute u
to its two use sites (the recv in new and the send in call).
and v can be statically eliminated as long as we know that the perm
Note that while the unknown clients of the service may store a a
and grad arrays have 256 elements. Using our type-sensitive CFA,
serv value in data structures, etc.., they may not access the inter-
we can map the arrays used in noise back to the allocation sites
nal representation and perform operations directly on the request
in mkNoise and thus enable array-bounds-check elimination.
channel. Because the server’s request channel is embedded in an
abstract value, our analysis is able to determine its use sites and
we can classify it as a known channel. The analysis will also iden- 8. Related work
tify replCh as a known channel, which allows the communication The application of control-flow analysis for higher-order func-
topology of this example to be accurately approximated. tional languages dates back to Shivers’ seminal work on control-
flow analysis for S CHEME [Shi88, Shi91]. Many variations of this
approach have been published including Serrano’s algorithm on
7.2 Perlin noise
which we base the presentation in this paper [Ser95].
Procedural generation of geometry and textures in computer graph- There is a significant body of work that falls into the intersection
ics applications often uses noise functions to introduce variability of type systems and program analysis. Some researchers have used
and produce more realistic images. One popular technique, which control-flow analysis to compute type information for untyped lan-
was developed by Ken Perlin, defines a mapping from <n to the guages [Shi91], while others have used type systems for program
unit interval, such that values that are close in <n will be mapped analysis [Pal01, Jen02].
to values that are close (ı.e., the noise has local similarity, but global Perhaps the most closely related work has been on using type
randomness) [PH89, Per02]. To compute the noise function, we information to guide analyses. For example, Jagannathan et al. de-
first divide up a unit n-cube into equal sized cells and pre-compute vised a flow analysis for a typed intermediate language as one
a random gradient vector in <n for each of the cell corners. Then might find in an ML compiler. Their analysis uses type information
the noise function is computed by mapping a point in <n to a cell to control polyvariance in the analysis and they prove that the anal-
interpolating between the 2n gradient vectors at the corners of the ysis respects the type system [WJ98]. Saha et al. used type infor-
cell. In a typical application, the noise function is sampled fre- mation to improve the performance of a demand-driven implemen-
quently and so its efficiency is paramount. The standard implemen- tation of CFA [SHO98] in the SML/NJ compiler. Lastly, Diwan et
tation of Perlin’s noise function uses a pair of precomputed arrays al. used type information to improve alias analysis for M ODULA-3
to allow fast calculation of the noise function. One of these arrays programs [DMM98]. We are not aware of any existing algorithms
is an array of 256 random gradient vectors in <n ; the second is a that use type abstraction to track values leaving and re-entering the
random permutation array of byte-sized indices. The permutation unit of analysis as we do.
array is used to map cell corners to indices in the gradient-vector
array. In a language like SML, the performance of the noise func- 9 For example, in the two-dimensional case, the noise function does 10 array
tion can be significantly reduced by the overhead of array bounds subscript operations.
 fun new () = let
val reqCh = channel()
fun server v = let
val (req, replCh) = recv reqCh
in
send(replCh, v);
server req
end
in
spawn (server 0);
S reqCh
Unknown end
clients fun call ( S ch , v) = let
val replCh = channel()
in
send ( ch , (v, replCh));
recv replCh
end

Figure 6. Data-flow of the server’s request channel

9. Conclusion [DMM98] Diwan, A., K. S. McKinley, and J. E. B. Moss. Type-

based alias analysis. In Proceedings of the SIGPLAN’98
We have presented an extension of control-flow analysis for ML Conference on Programming Language Design and Imple-
that exploits type abstraction present in the source program to im- mentation, New York, NY, June 1998. ACM, pp. 106–117.
prove the quality of the analysis. Our technique is based on the [GMZ00] Grossman, D., G. Morrisett, and S. Zdancewic. Syntactic
combination of a more refined representation of unknown values type abstraction. ACM Transactions on Programming
and computing an approximation of abstract values that escape into Languages and Systems, 22(6), 2000, pp. 1037–1080.
the wild. We argued that our analysis is safe and faithful to the lan- [Jen02] Jensen, T. Types in program analysis. In The essence
guage’s type system, and we gave two examples of optimizations of computation: complexity, analysis, transformation, vol.
that require the information provided by our analysis. 2566 of Lecture Notes in Computer Science, pp. 204–222.
The key idea of tracking abstract values that escape could also Springer-Verlag, New York, NY, 2002.
be applied to other analysis frameworks, such as first-order data- [Ler92] Leroy, X. Unboxed objects and polymorphic typing. In
flow analysis or more aggressive control-flow analyses (e.g., 1- Conference Record of the 19th Annual ACM Symposium
CFA [Shi91] or ∆-CFA [MS06]). We presented our analysis for on Principles of Programming Languages, New York, NY,
a monomorphic subset of SML This analysis is being used as part January 1992. ACM, pp. 177–188.
of an optimizer for CML and it will need to be extended to the [MS06] Might, M. and O. Shivers. Environment analysis via ∆CFA.
full SML language. We described some ideas for how to extend In Conference Record of the 33rd Annual ACM Symposium
the analysis, but more research is needed to fully deal with opaque on Principles of Programming Languages, New York, NY,
signature ascription and polymorphism. January 2006. ACM, pp. 127–140.
[MTHM97] Milner, R., M. Tofte, R. Harper, and D. MacQueen. The
Definition of Standard ML (Revised). The MIT Press,
References Cambridge, MA, 1997.
[NNH99] Nielson, F., H. R. Nielson, and C. Hankin. Principles of
[AD98] Ashley, J. M. and R. K. Dybvig. A practical and flexible flow
Program Analysis. Springer-Verlag, New York, NY, 1999.
analysis for higher-order languages. ACM Transactions on
Programming Languages and Systems, 20(4), July 1998, pp. [Pal01] Palsberg, J. Type-based analysis and applications. In
845–868. PASTE’01, June 2001, pp. 20–27.
[App98] Appel, A. An alternate graph representation for the [Per02] Perlin, K. Improving noise. In SIGGRAPH ’02: Proceedings
tiger compiler. http://www.cs.princeton.edu/ of the 29th annual conference on Computer graphics and
~appel/modern/ml/altgraph.html, May 1998. interactive techniques, New York, NY, 2002. ACM, pp.
681–682.
[CC77] Cousot, P. and R. Cousot. Abstract interpretation: a unified
lattice model for static analysis of programs by construction [PH89] Perlin, K. and E. M. Hoffert. Hypertexture. In SIGGRAPH
or approximation of fixpoints. In POPL ’77: Proceedings of ’89: Proceedings of the 16th annual conference on Computer
the 4th ACM SIGACT-SIGPLAN symposium on Principles graphics and interactive techniques, New York, NY, 1989.
of programming languages, New York, NY, January 1977. ACM, pp. 253–262.
ACM, pp. 238–252. [Rep91] Reppy, J. H. CML: A higher-order concurrent language. In
[CDL99] Chilimbi, T. M., B. Davidson, and J. R. Larus. Cache- Proceedings of the SIGPLAN’91 Conference on Program-
conscious structure definition. In Proceedings of the ming Language Design and Implementation, New York, NY,
SIGPLAN’99 Conference on Programming Language June 1991. ACM, pp. 293–305.
Design and Implementation, New York, NY, May 1999. [Rep99] Reppy, J. H. Concurrent Programming in ML. Cambridge
ACM, pp. 13–24. University Press, Cambridge, England, 1999.
[RX06] Reppy, J. and Y. Xiao. Specialization of CML message- [SHO98] Saha, B., N. Heintze, and D. Oliva. Subtransitive CFA
passing primitives. Submitted for publication., 2006. using types. Technical Report YALEU/DCS/TR-1166, Yale
[Ser95] Serrano, M. Control flow analysis: a functional languages University, Department of Computer Science, New Haven,
compilation paradigm. In SAC ’95: Proceedings of the 1995 CT, October 1998.
ACM symposium on Applied Computing, New York, NY, [WJ98] Wright, A. K. and S. Jagannathan. Polymorphic splitting:
1995. ACM, pp. 118–122. an effective polyvariant flow analysis. ACM Transactions
[Shi88] Shivers, O. Control flow analysis in scheme. In Proceedings on Programming Languages and Systems, 20(1), 1998, pp.
of the SIGPLAN’88 Conference on Programming Language 166–207.
Design and Implementation, New York, NY, June 1988. [Xia05] Xiao, Y. Toward optimization of Concurrent ML. Master’s
ACM, pp. 164–174. dissertation, University of Chicago, December 2005.
[Shi91] Shivers, O. Control-flow analysis of higher-order languages.
Ph.D. dissertation, School of Computer Science, Carnegie
Mellon University, Pittsburgh, PA, May 1991.