Authors Koushik Sen Michael Pradel
License CC-BY-3.0
The Good, the Bad, and the Ugly: An Empirical
Study of Implicit Type Conversions in JavaScript
Michael Pradel1 and Koushik Sen2
1 TU Darmstadt
Department of Computer Science
Germany
michael@binaervarianz.de
2 University of California, Berkeley
EECS Department
USA
ksen@berkeley.edu
Abstract
Most popular programming languages support situations where a value of one type is converted
into a value of another type without any explicit cast. Such implicit type conversions, or type
coercions, are a highly controversial language feature. Proponents argue that type coercions
enable writing concise code. Opponents argue that type coercions are error-prone and that
they reduce the understandability of programs. This paper studies the use of type coercions in
JavaScript, a language notorious for its widespread use of coercions. We dynamically analyze
hundreds of programs, including real-world web applications and popular benchmark programs.
We find that coercions are widely used (in 80.42% of all function executions) and that most
coercions are likely to be harmless (98.85%). Furthermore, we identify a set of rarely occurring
and potentially harmful coercions that safer subsets of JavaScript or future language designs
may want to disallow. Our results suggest that type coercions are significantly less evil than
commonly assumed and that analyses targeted at real-world JavaScript programs must consider
coercions.
1998 ACM Subject Classification D.3.3 Language Constructs and Features, F.3.2 Semantics of
Programming Languages, D.2.8 Metrics
Keywords and phrases Types, Type coercions, JavaScript, Dynamically typed languages
Digital Object Identifier 10.4230/LIPIcs.ECOOP.2015.999
1 Introduction
In most popular programming languages, a value has a type that determines how to interpret
the data represented by the value. To change the way a value is interpreted, programmers can
convert a value of one type into a value of another type, called type conversion. In addition
to explicit type conversions, or casts, many languages perform implicit type conversions,
called type coercions. Type coercions are applied by the compiler, e.g., to transform an
otherwise type-incorrect program into a type-correct program, or by the runtime system,
e.g., to evaluate expressions that involve operands of multiple types. Both statically typed
and dynamically typed languages use type coercions. For example, C and Python coerce
numeric values from integer types to floating point types, and vice versa, and Java coerces
instances of a subtype into an instance of a supertype. Some dynamically typed languages,
such as JavaScript, make even wider use of type coercions.
© Michael Pradel and Koushik Sen;
licensed under Creative Commons License CC-BY
29th European Conference on Object-Oriented Programming (ECOOP’15).
Editor: John Tang Boyland; pp. 999–1021
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
1000 Implicit Type Conversions in JavaScript
Type coercions are a highly controversial language feature. On the one hand, they enable
programmers to write concise code. On the other hand, type coercions can obfuscate a
program, because the source code does not clarify which types of values a piece of code
operates on, and even hide errors, because an unintended operation or an unintended loss of
information may remain unnoticed.
This paper focuses on JavaScript, a language notorious for its heavy use of type coercions.
Many operations that are considered illegal in languages with a stricter type system are
legal in JavaScript. The reason is that JavaScript avoids throwing runtime type mismatch
exceptions as much as possible so that programs can tolerate errors as much as possible.
Although the rules for coercing types are well defined [1], even expert JavaScript developers
struggle to fully comprehend the behavior of some code. For illustration, consider the
following rather surprising examples:
"false" == false evaluates to false, but "0" == "false" evaluates to true.
new String("a") == "a" and "a" == new String("a") evaluate to true, but new String("a")
== new String("a") evaluates to false.
[] << "2" evaluates to 0, [1] << "2" evaluates to 4, and [1,2] << "2" evaluates to 0.
Because type coercions can lead to such surprising behavior in JavaScript, it is common
to assume that coercions are error-prone and therefore rarely used in practice. Based on
these assumptions, existing static type inference and checking approaches for dynamically
typed languages do not support type coercions at all [25, 3] or allow only a limited set of
coercions [11, 5]. For example, Chugh et al. prohibit implicit coercions since they “often lead
to subtle programming errors” [3].
While such assumptions about coercions are common, little is currently known about
how type coercions are used in practice. This paper addresses this problem and asks the
following research questions:
RQ 1: How prevalent are type coercions in JavaScript, and what are they used for?
RQ 2: Are type coercions in JavaScript error-prone?
RQ 3: Do type coercions in JavaScript harm code understandability?
Answering these questions provides insights that enable informed decision making in at
least three situations. First, developers of static and dynamic analyses benefit from the
answers when deciding about how to handle coercions. Second, developers of safer subsets of
JavaScript, such as strict mode [1], can use the answers to estimate the number of changes
that a particular subset of JavaScript imposes on existing programs. Third, designers of
future languages can benefit from the answers to decide whether and how to include type
coercions.
To address these questions, this paper presents an empirical study of type coercions in
real-world JavaScript web applications and in popular benchmark suites. As a result of the
study, we can substantiate and refute several pieces of conventional wisdom, respectively.
Our results include the following findings:
Type coercions are widely used: 80.42% of all function executions perform at least one
coercion and 17.74% of all operations that may apply a coercion do apply a coercion, on
average over all programs. (RQ 1)
In contrast to coercions, explicit type conversions are significantly less prevalent. For
each explicit type conversion that occurs at runtime, there are 269 coercions. (RQ 1)
We classify coercions into harmless and potentially harmful coercions, and we find that
98.85% of all coercions are harmless and likely to not introduce any misbehavior. (RQ 1)
A small but non-negligible percentage (1.15%) of all type coercions are potentially harmful.
Future language designs or restricted versions of JavaScript may want to forbid them.
M. Pradel and K. Sen 1001
For today’s JavaScript programs, a runtime analysis can report these potentially harmful
coercions. (RQ 2)
Out of 30 manually inspected potentially harmful code locations, 22 are, to the best of
our knowledge, correct, and only one is a clear bug. These results suggest that the overall
percentage of erroneous coercions is very small. (RQ 2)
Most code locations with coercions are monomorphic (86.13%), i.e., they always convert
the same type into the same other type, suggesting that these locations could be refactored
into explicit type conversions for improved code understandability. (RQ 3)
Most polymorphic code locations (93.79%), i.e., locations that convert multiple different
types, are conditionals where either some defined value or the undefined value is coerced
to a boolean. (RQ 3)
JavaScript’s strict and non-strict equality checks are mostly used interchangeably, suggest-
ing that refactoring non-strict equality checks into strict equality checks can significantly
improve code understandability. (RQ 3)
All data gathered for this study and the full details of our results are available for download:
http://mp.binaervarianz.de/ecoop2015
In summary, this paper contributes the following:
The first in-depth study of type coercions in real-world JavaScript programs.
Empirical evidence that contradicts the common assumption that type coercions are
rarely used and error-prone. Instead, we find that coercions are highly prevalent and
mostly harmless.
A classification of coercions into harmless and potentially harmful that may guide the
development of safer subsets of JavaScript and future language designs.
2 Methodology and Subject Programs
To answer the research questions from the introduction, we perform a study of type coercions
in real-world JavaScript programs. We dynamically analyze an execution of each program to
record all runtime events that may cause a coercion (Section 2.1). Based on these data, we
perform a set of offline analyses that summarize the runtime data into representations that
allow for answering the research questions.
2.1 Dynamic Analysis
To gather runtime information about coercions, we instrument the program to intercept all
runtime events that may cause a coercion or an explicit type conversion:
Unary and binary operations. For unary and binary operations, the analysis records
the source code location, the operator, the types of the input(s) and the output of the
operation, and an abstraction (defined below) of the runtime values of the input(s) and
the output of the operation.
Conditionals. For each conditional evaluated during the execution, the analysis records
the code location, the type and the abstracted runtime value of the evaluated expression,
and the boolean value into which the expression gets coerced.
Function calls. For each function call, the analysis checks whether the called function is
any of the built-in functions that explicitly convert values from one type to another, i.e.,
Boolean, Number, and String. In this case, the analysis records the code location, as well
as the types and the abstracted values of the function argument and the return value.
ECOOP’15
1002 Implicit Type Conversions in JavaScript
We abstract runtime values as follows: For booleans, undefined, null, and NaN (not a number),
the analysis stores the value. For other numbers, it stores whether the value is zero or non-
zero. For strings, it stores whether the string is empty or non-empty. For objects, including
arrays, it stores whether the object has any own properties, i.e., whether it is an empty object
or array. Furthermore, the analysis records whether an object has valueOf and toString
methods that differ from the default implementation inherited from Object and, if an object
has such a valueOf method, the type of valueOf’s return value. This extra type information
allows us to identify values that explicitly define how to coerce them into another type.
Our implementation builds upon Jalangi [22], which instruments the JavaScript source
code so that the analysis can intercept all necessary runtime events. For programs that
execute on Node.js, we instrument the program on the file system. For web applications,
we modify the Firefox browser so that it intercepts and instruments all JavaScript code
before executing the code. The dynamic analysis creates a file that summarizes all recorded
information for a program, and we analyze this file offline, i.e., after the execution.
For the purpose of this study, dynamic analysis has several benefits over a comparable
static analysis. First, statically determining whether a code location coerces a type is
impossible in general due to the highly dynamic nature of JavaScript. We believe that
statically overapproximating potential coercions may skew the study results. Second, dynamic
analysis enables us to reason about concrete runtime values, which is important for the
qualitative part of our study, where we manually inspect coercions to determine whether
coercions are harmful. Finally, dynamic analysis enables us to quantify how often coercions
occur at runtime. On the downside, dynamic analysis misses coercions on paths that are not
executed. We discuss implications of this limitation in Section 5.
2.2 Subject Programs
Our study considers three kinds of programs: the SunSpider benchmarks, the Octane
benchmarks, and the top 100 most popular web sites according to Alexa1 . For each benchmark,
we analyze an execution of the benchmark in its default setup. For each web site, we load
the start page and analyze all JavaScript code that gets executed, including code loaded
from third-party libraries. In total, the study considers 138,979,028 runtime events from 132
programs. These events are generated by 321,711 unique source code locations.
In addition to the main research questions of this paper, this setup also allows us
to address the question how accurately the benchmark suites, which are widely used to
evaluate JavaScript engines, represent the type coercion-related behavior of real-world web
applications.
3 Classification of Type Coercions
In this section, we propose a classification of all type coercions that may occur in JavaScript
into likely harmless and potentially harmful coercions. This classification may serve three
purposes. First, we use it to approximate the harmfulness of coercions in practice by classifying
the coercions we observe in real-world programs. Second, the classification may provide a
basis for defining a safer subset of JavaScript that forbids or warns about potentially harmful
coercions. Such a subset may be enforced through runtime checks, similar to JavaScript’s
strict mode [1]. Third, the classification may guide future language designs that want to
1
http://www.alexa.com/topsites, accessed on July 16, 2014
M. Pradel and K. Sen 1003
allow harmless coercions for conciseness but disallow potentially harmful coercions. The
results of our study will help approaches of the second and third direction by providing
empirical data about how often different kinds of coercions occur, i.e., how much code one
would break by disallowing them.
Since developers may purposefully exploit the behavior of any type coercion, there is no
clear-cut definition of when a coercion constitutes an error. The proposed classification is
based on our own experience with JavaScript, on reports of the experience of others, e.g., in
web forums, and on a comparison with other programming languages. We classify a coercion
as potentially harmful if its semantics deviates from what is common in other, more strongly
typed languages, such as C, Java, Python, or Ruby, if the operation that triggers the coercion
has no intuitive meaning, or if the rules that determine which coercion to apply are very
complex. We classify all other type coercions as harmless. The remainder of this section
presents and illustrates our classification, which is summarized in Table 1.
3.1 Terminology
JavaScript has six basic types: The three primitive types boolean, number, and string, the
special, single-value types undefined and null, and the object type, which includes arrays
and functions. To simplify our classification, we use the following additional terms.
A quasi-number is a primitive number or an object that defines a valueOf method that
returns a primitive number. Developers can use valueOf to specify how to coerce an
object, and the execution environment calls the method whenever the language requires
a coercion.
A quasi-string is a primitive string or an object that defines a valueOf method that
returns a primitive string.
A wrapped primitive is an object created with one of the built-in wrappers, new Boolean(),
new Number(), and new String().
An empty object is an object without any own properties, e.g., the result of evaluating
the object literal expression {}.
An empty array is an array with length zero, e.g., the result of evaluating the array literal
expression [].
A defined value is every value except for undefined and null.
3.2 Conditional-related Coercions
In JavaScript, values of all types may be used as conditions. Furthermore, all types may
occur as the operand of the logical negation operator ! and as the operands of the binary
logical operators && and ||. The semantics are straightforward: All objects, all strings
except the empty string, and all numbers except zero and NaN coerce to true. Note that all
objects include objects whose valueOf returns false, empty arrays, and empty objects. All
other values, including undefined and null, coerce to false. Because of these rather simple
semantics and because using arbitrary types in conditionals is very common in JavaScript,
we consider these coercions as harmless.
As an exception to the above classification, we classify wrapped primitives used as
conditions or as operands of !, &&, and || as potentially harmful. Because all objects, including
wrapped primitives, coerce to true, the semantics of wrapped primitives in conditionals
differs from their primitive counterparts, which may surprise developers. For example, the
wrapped primitive new Boolean(false) coerces to true. Popular guidelines [4] suggest to
avoid wrapped primitives altogether.
ECOOP’15
1004 Implicit Type Conversions in JavaScript
Operation Type of operands Example Comment Class.
Conditional:
- Wrapped primitives if (new Boolean(false)) Behavior differs 6
from primitives
- All other types if (someNumber) Coerced to boolean 4
Unary operations:
+, -, ˜ All except quasi-numbers -"abc" Coerced to number 6
! Wrapped primitives !(new Boolean(false)) Behavior differs 6
from primitives
! All except wrapped prim- !someNumber Coerced to boolean 4
itives
Binary operations:
-, *, /, % All except quasi-numbers "abc" * false Meaningful only for 6
numbers
<<, >>, >>> All except quasi-numbers {} << 23 Meaningful only for 6
numbers
+ Quasi-string and var x; x+="abc" Result contains 6
undefined, or quasi- "undefined" or
string and null "null"
+ Quasi-string and a de- "Names: " + Common to con- 4
fined value arrayOfNames struct strings
+ Two non-quasi-strings false + [2,3] Confusing semantics 6
<, >, <=, >= All except two quasi- [1,2] < function f() {} Meaningful only for 6
numbers and two quasi- numbers and strings
strings
==, != All types, unless both 0 == "false" Confusing semantics 6
types are the same, or
one is undefined or null
==, != One value is undefined someObject != null Common in condi- 4
or null tionals
|, || Left type is undefined, x = (x | 0) + 1 Common pattern to 4
right value is 0 initialize counters
&, |, ˆ All types, unless both [1,2] & "abc" Meaningful only for 6
are quasi-numbers, or numbers
counter initialization pat-
tern from above
&&, || At least one wrapped new Number(0) || new Behavior differs 6
primitive Boolean(false) from primitives
&&, || All except wrapped prim- someNumber && Common in condi- 4
itives someBoolean tionals
Table 1 Classification of type coercions into likely harmless (4) and potentially harmful (6).
M. Pradel and K. Sen 1005
3.3 The Plus Operator
The semantics of the + operator is defined for numbers, where it means addition, and for
strings, where it means concatenation. If any value that is neither a number nor a string is
given to +, JavaScript coerces the value either into a primitive number or into a primitive
string. Then, it applies string concatenation if at least one operand is a string and addition
otherwise. The rules for deciding whether a value gets coerced into a number or a string are
somewhat intricate. For most but not all values, JavaScript attempts to coerce the value
into a number by calling valueOf and falls back on coercing into a string by calling toString.
We refer to [1] for a full description of the rules.
Given that + has easy to understand semantics when being applied to two quasi-numbers
or to two quasi-strings, we classify coercions that happen in these cases as harmless. We
also classify coercions as harmless if they result from combining a quasi-string with any
defined value because developers commonly use this pattern to concatenate strings. For
example, "Names: " + arrayOfNames causes a harmless coercion of an array into a string
representation of the array. In contrast to these harmless coercions, we classify all other
coercions triggered by + as potentially harmful because their semantics differ from more
strongly typed languages. For example, the statements var x; x+="abc";, which yield the
string "undefinedabc" instead of the probably expected "abc", and the expression false+{},
which yields "false[object Object]", are classified as potentially harmful.
3.4 Arithmetic and Bitwise Operators
The unary + and - operators coerce their operand to a number. The arithmetic operators -,
*, /, %, and the bitwise shift operators <<, >>, and >>> are defined for numbers, and applying
them to any types except for two numbers triggers a coercion to number. Similar to the
arithmetic operators, the unary bitwise ˜ operator and the binary bitwise operators &, |, and
ˆ are defined for 32-bit integers. Applying these operators to any other values leads to a
coercion of the operands into 32-bit integers.
Because all these operations are meaningful only for numbers and values that coerce
into a number, we consider them as harmless when being applied to quasi-numbers, and
as potentially harmful otherwise. For example, we consider the following operations as
potentially harmful:
-"abc", which results in NaN
[1,2] & "abc", which yields 0
{} << 23, which yields 0
The above classification considers coercing a string to a number in an arithmetic operation
as potentially harmful. The reason is that one arithmetic operator, +, has a different semantics
than the other arithmetic operators, which may easily confuse developers: 23 - "5" is
interpreted arithmetically and yields 18, but 23 + "5" is interpreted as string concatenation
and yields "235".
As an exception to the above classification, we consider a common code idiom for
initializing counter variables, e.g., x = (x | 0) + 1. This idiom initializes x to zero when
the code is executed for the first time, and it increments x otherwise. That is, at the first
execution, the idiom coerces undefined to 0. Because this idiom has clear semantics and is
commonly used, we classify coercions caused by this pattern as harmless.
ECOOP’15
1006 Implicit Type Conversions in JavaScript
3.5 Relational Operators
The semantics of the relational operators <, >, <=, and >= is defined for numbers (numeric
comparison) and for strings (lexicographic order). JavaScript coerces any pairs of values that
contain a non-number or a non-string to either a pair of numbers or a pair of strings, and
then applies the respective operation. The rules for such coercions are the following: At first,
JavaScript coerces any non-primitive into a primitive, where a valueOf method that returns
a number is preferred over a toString that returns a string. Then, JavaScript coerces any
remaining non-numbers into numbers, unless both primitives are strings, in which case the
lexicographic order is computed.
Because relational operators have an intuitive semantics for pairs of numbers and for
pairs of strings, we classify coercions that occur when combining two quasi-numbers or two
quasi-strings as harmless. In contrast, we classify all other coercions as potentially harmful,
because relational operations have no meaningful semantics for other types. For example, we
consider [1,2] < function foo() {}, which yields true, as potentially harmful.
3.6 Equality Operators
JavaScript provides two kinds of (in)equality operators: the strict operators === and !==,
and the non-strict operators == and !=. The strict operators never apply any type coercions;
instead, they consider any two values of different types as unequal. In contrast, their non-
strict counterparts coerce operands to evaluate whether they may be considered equal despite
having different types. Because the coercion rules for non-strict equality operations are
rather complex and sometimes unintuitive (see [1] for full details), guidelines [4] recommend
to avoid non-strict equality operations. Examples for the somewhat surprising behavior of
non-strict equality operations include:
"" == 0 yields true but "false" == 0 yields false because comparing a string and a
number leads to coercing the string into a number, where "" is coerced into 0 and "false"
is coerced into NaN.
true == "true" yields false but true == "1" yields true because comparing a boolean
and any non-boolean leads to coercing the boolean into a number, which coerces true to
1.
Both new Number(5) == 5 and 5 == new Number(5) yield true but new Number(5) == new
Number(5) yields false because non-strict equality is non-transitive.
Because of their rather confusing semantics, we consider type coercions caused by non-
strict equality operations as potentially harmful, including all the examples given above.
The only exception to this classification are comparisons of arbitrary values with undefined
or null. Because undefined and null essentially mean the same in a non-strict comparison,
such comparisons are commonly used to check for defined values. For example, we classify
the following check as harmless: someObject != null.
4 Type Coercions in the Wild
4.1 RQ 1: Prevalence of Type Coercions
In the following, we address the question how prevalent type coercions are in real-world
JavaScript programs.
M. Pradel and K. Sen 1007
Function executions
100
with coercion
80
60
40
20
0
da da co bit cry bit str str re ma Ot Go Go Go Go de ra Xh Ch Ne str
te- te- ntr op pt op ing ing ge th he og og og og lta ytr am in ob ing
for for olf s-3 o-s s-b -fa -ta xp -sp rs le le le le blu ace st ad ux -u
ma ma low bi ha its st gc ec .ru .es .co .co e er aily .co np
t-x t-t -re t-bi 1.j -in- a.js lou tra .jp m. .co .c m ac
pa oft cu ts- s by d l hk m om k
rb e.js rsi in- te. .js - n or .cn -cod
.js ve by j s m .js e.j
.js te. s
js
Figure 1 Percentage of function executions with at least one type coercion (top 10 and bottom
10 programs only).
4.1.1 Function Executions With At Least One Coercion
As a measure of the prevalence of type coercions, we assess during how many function
executions at least one coercion occurs:
I Definition 1 (Function executions with coercion, F EC). The percentage F EC of function
executions with a coercion is the number of function executions where at least one type
coercion occurs between entering the function and exiting the function, excluding the
execution of the function’s callees, divided by the total number of function executions.
For example, executing the following program yields a F EC of 50% because f() does not
perform any coercion, but calling g() triggers the coercion in line 6.
1 function f() {
2 g();
3 return 5 + 1;
4 }
5 function g() {
6 return 5 + true;
7 }
8 f();
Over all programs we analyze, the F EC is 80.42%. This number indicates that coercions
are a highly prevalent phenomenon that cannot be ignored when analyzing real-world
JavaScript programs.
Figure 1 shows the F EC for a subset of all programs. The figure includes the ten
programs with the highest percentage, the ten programs with the lowest percentage, and
the average percentage of all other programs (“Others”). These results show that even
programs with relatively few coercions still perform a non-negligible number of coercions.
The figure excludes programs with less than 100 function executions because the F EC is
not representative for these programs. In particular, the figure excludes several SunSpider
benchmarks that perform less than 10 function executions and for which the F EC is 100%.
4.1.2 Coercions versus Non-coercions
As another measure of how prevalent coercions are, we compute how often an operation that
could lead to a coercion does coerce one or more values:
ECOOP’15
1008 Implicit Type Conversions in JavaScript
50 Quartiles Quartiles
50
Coercions among operations
Coercions among operations
40 40
30 30
20 20
10 10
0 0
we oc su we oc su
bs tan ns bsit tan ns
ite e pid es e pid
s er er
(a) Dynamic occurrences. (b) Static occurrences.
Figure 2 Prevalence of type coercions as percentage over all operations where type coercions
may occur.
I Definition 2 (Coercions among operations, CAO). The percentage CAO of coercions among
operations is the number of operations that perform a coercion divided by the total number
of executed unary and binary operations and evaluated conditionals that could perform a
coercion.
Figure 2 shows the CAO for the three groups of programs that we analyze. For each
group, the figure shows the mean CAO over all programs from the group, the upper and
lower quartiles, and the minimal and maximal CAO. We show two variants of the CAO
measure. Figure 2a is based on the dynamic frequency of operations, i.e., each dynamic
occurrence of an operation counts. In contrast, Figure 2b is based on static code locations,
i.e., each static code location counts at most once.
The results reveal two interesting properties. First, type coercions occur in a non-negligible
fraction of all operations that may cause coercions. For web sites, 36.25% of all code locations
that may coerce values indeed do it in the analyzed executions. Second, type coercions are
significantly more prevalent in web sites than in the SunSpider and Octane benchmarks.
These benchmark suites have been criticized to be unrepresentative for real-world JavaScript
programs [21, 19], and our study confirms this observation for type coercions.
Figure 3 shows the CAO for the top 10 and bottom 10 programs. On many popular
web sites, such as Yahoo.com and Apple.com, more than a third of all operations that could
perform a coercion do perform a coercion. The 10 programs with the smallest CAO are all
from the SunSpider or Octane benchmarks, which again shows that these benchmarks fail to
accurately represent real-world web applications. The figure excludes programs with less
than 100 observations because measuring their CAO does not provide representative results.
4.1.3 Kinds of Type Coercions
To better understand why coercions are so prevalent in real-world JavaScript programs, we
analyze what kinds of coercions occur. Figure 4 shows the most and the least prevalent
kinds of type coercions. The horizontal axis clusters similar kinds of coercions. For example,
“number in conditional” means that a value of type number is coerced into a boolean because
it occurs in a conditional, and “+-˜ null” means that one of the arithmetic operators +, -, or
˜ is applied to null, which leads to a coercion into the number zero.
M. Pradel and K. Sen 1009
Coercions among operations~ ~
45
40
35
30
25
20
15
10
5
0
Ya Ap Ra Dr Go Eb Yo Ha Xv Vk Ot pd cry cry gb cry na str 3d ac ac
ho ple ku op .c ay utu o1 ide .co he fjs pt pt em pt vie ing -cu ce ce
o.c .c ten bo om .co be 23 os m rs o- o- u o r- -b b ss ss
md sh str as e.j -n -fa
om om .co x.c m .c .co .co
om m m 5.j a1. ok e6 s bo nn
.jp om s js es 4. dy ku
js .js ch
.js
(a) Dynamic occurrences.
Coercions among operations~ ~
45
40
35
30
25
20
15
10
5
0
Yo Go Go Go Go Go Go Go Go Go Ot da re cry 3d cry str cry ac ac 3d
utu og og og og og og og og og he te- gex pt -ra pt ing pt ce ce -cu
be le. le. le. le. le. le. le. le. le. rs form p o-s ytra o-m -ba o-a ss-n ss-f be
.co com ca co. com de it fr com co. at- ha c d se es b an .js
m .a in .m .tr id xp 1.j e.j 5.js 64 .js od nk
u x ar s s .js y.j uc
b.j s h.j
s s
(b) Static occurrences.
Figure 3 Programs with the highest and lowest prevalence of type coercions (measured as
percentage over all operations where type coercions may occur).
The figure shows that conditionals and logical negations, which are typically used in
conditionals, are the most prevalent kinds of coercion. Overall, coercions that result from
conditionals or from operations that are typically used in conditionals (!, &&, and ||) account
for 93.01% of all coercions. This result suggests that analyses of JavaScript, such as type
inference and checking approaches, should at least consider these kinds of coercions because
they occur frequently in practice.
4.1.4 Implicit versus Explicit Type Conversions
As an alternative to coercions, JavaScript developers can explicitly convert values from
arbitrary types into booleans, numbers, and strings using the built-in functions Boolean,
Number, and String. We measure how prevalent such explicit type conversions are and
compare their prevalence to coercions. In total, we observe 20,407 explicit type conversions
during the execution of all programs, and 5,497,545 coercions. That is, for every explicit type
conversion that occurs, there are 269 implicit type conversions. We conclude that explicit
conversions are used significantly less frequently than coercions. A possible explanation
is that developers prefer the conciseness of coercions over the potentially increased code
understandability through explicit conversions.
ECOOP’15
1010 Implicit Type Conversions in JavaScript
25
Percentage 20
15
10
5
0
nu un str ! n ob ! o ar nu fun nu Ot nu nu bo bo +- nu fun +- ar ! [o
mb de ing um jec bje ray ll i ct ll E he ll R mb ole ole ~ n mb ct ~ b ray bj
er fine in be t in ct in n c ion Q rs EL er an an ull er ion oo RE ect
in d co r co co on in ob str RE RE + f BI BI lea L Bo
co in nd nd nd dit co jec ing L o L un T T n nu o
str un
nd co iti i t itio ion nd t n c
bje um tio ing de mb lea
itio nd on ion na al itio er n]
na itio al al l na ct be n fin
l n al l r ed
(a) Dynamic occurrences.
20
Percentage
15
10
5
0
un ! n ob str nu fun nu ! u ar ! o Ot ! [o nu bo bo nu nu fun +- ar un
de um jec ing mb ct ll i nd ray bje he bj mb ole ole ll R mb ct ~ n ray de
fin be t i in er ion n c efi in ct rs ec er an an E er ion ul R fin
ed r n c c in in on ne co tB R R + Ls B B l EL ed
in on on co co dit d nd oo EL EL fu tri IT IT nu BIT
co dit diti nd nd ion itio lea o n nc ng str un
b mb u
nd ion on iti iti al n n ] j e um tio ing de er nd
itio al al ona ona al ct be n fin efi
na l l r ed ne
l d
(b) Static occurrences.
Figure 4 Prevalence of different kinds of type coercions as percentage over all type coercions (top
10 and bottom 10 only). Horizontal axes: BIT, EQ, and REL mean bitwise, equality, and relational
operations, respectively.
4.2 RQ 2: Harmfulness of Type Coercions
We address the question whether type coercions are error-prone in two ways. First, we
measure how many type coercions are harmless and potentially harmful according to the
classification from Section 3. Second, we manually inspect a sample of the potentially harmful
type coercions to assess whether their behavior is intended or erroneous.
4.2.1 Harmless versus Potentially Harmful Coercions
Given the classification from Section 3, Figure 5 shows how many of all observed type
coercions are potentially harmful. For most programs, a relatively small percentage of
coercions is potentially harmful. For example, on average over all websites, only 1.25% of all
dynamic occurrences of coercions are potentially harmful. The average over the coercions
of all programs is 1.15%. Some of the SunSpider benchmarks are outliers in Figure 5, with
up to 99.98% potentially harmful coercions. By manually inspecting these benchmarks we
find that they indeed perform various potentially harmful coercions, which lead to a high
percentage because these benchmark programs are relatively small.
Figure 6 shows the ten applications that have the highest and lowest percentage of poten-
M. Pradel and K. Sen 1011
100 Quartiles 90 Quartiles
80
Pot. harmful coerc. (%)
Pot. harmful coerc. (%)
80 70
60
60
50
40
40
30
20 20
10
0 0
we oc su we oc su
bs tan ns bs tan ns
ite e pid ite e pid
s er s er
(a) Dynamic occurrences. (b) Static occurrences.
Figure 5 Percentage of potentially harmful coercions over all coercions.
tially harmful coercions, respectively. Apart from the outliers in the SunSpider benchmarks
discussed above, all programs have less than 12% potentially harmful coercions (for both
dynamic and static occurrences).
4.2.2 Kinds of Potentially Harmful Coercions
Since the number of potentially harmful coercions is non-negligible, we further analyze what
kinds of potentially harmful coercions occur. Figure 7 shows the 10 most prevalent and the 10
least prevalent kinds of potentially harmful coercions. The most prevalent potentially harmful
coercion (by number of static occurrences) are non-strict (in)equality checks that compare
two objects of different types, which is a common source of confusion. Several prevalent
kinds of potentially harmful coercions involve undefined, such as concatenating undefined
with a string, which yields a string that contains "undefined", and relative operators applied
to undefined and a number, which always yields false. We speculate that most of these
coercions are caused by an undefined value that accidentally propagates through the program.
Our results suggest that developing analyses that warn programmers about potentially
harmful coercions is a promising line of future work, e.g., along the lines of [11]. Our study
provides empirical data on how many warnings such analyses may yield, so that developers
of such analyses can focus on rarely occurring, potentially harmful coercions. Another
direction for future work are techniques that prevent the undefined value from propagating
in undesired ways.
4.2.3 Binary Plus Operations
One of the most debated and potentially error-prone operators in JavaScript is the binary
+ operator. For example, a popular guideline [4] lists the operator as “problematic” and
mentions that it is “a common source of bugs”. To better understand how dangerous + is in
practice, we analyze all occurrences of this operator.
Figure 8 shows what kinds of types + is applied on, including types that do not lead to
any coercion. The labels on the horizontal axis also indicate whether a coercion occurs and
if yes, whether we classify the coercion as potentially harmful. Most + operations apply to
either two strings or to two numbers, i.e., they do not coerce any types. The most prevalent
ECOOP’15
1012 Implicit Type Conversions in JavaScript
Pot. harmful coerc. (%) 100
80
60
40
20
0
cry cry 3d Vk Ra Yo Bl Bl Si pd Ot Tm Ta ric ra sp re str da str cry
pto pto -ra .co ku utu ogs ogg na. fjs her al ob har ytra lay gex ing te- ing pto
-m -sh ytra m ten be po er. com s l.co ao ds ce p -un for -ba -a
d5 a1 ce .co .co t.c co .c m .co pa ma se es.
.js .js .js .jp m om m n m ck t-x 64 js
-co pa .js
de rb.j
.js s
(a) Dynamic occurrences.
Pot. harmful coerc. (%)
90
80
70
60
50
40
30
20
10
0
cry cry 3d pd Bl Bl gb Vk Bi Ch Ot Cr Ki Hu Ab Go Od cry da str str
pto pto -ra fjs ogs ogg em .co ng. ina her aig cka ffin ou og no pto te- ing ing
-sh -m ytra po er u m com da s slis ss gt t.c le. kla -a for -ba -un
a1 d5 ce t.c .co
om m
ily
.co t.o .to onp om co. ss es. ma se pa
.js .js .js rg os uk nik js t-x 64 ck
m. t.c i.ru pa .js -c
cn om rb od
.js e.j
s
(b) Static occurrences.
Figure 6 Programs with the highest and lowest percentage of potentially harmful coercions.
occurrence of + that coerces operands, number + string, is harmless. In total, only a very
small percentage of all dynamic occurrences of + lead to a potentially harmful coercion.
We conclude from these results that the + operator is less dangerous than commonly
expected. Programmers are disciplined enough to apply + (mostly) in situations where the
operation does not cause any type coercion or where it applies a harmless coercion that has
obvious semantics. That said, reconsidering the semantics of + in future language designs to
reduce its complexity seems to be a good idea. To deal with today’s JavaScript, checking
for the rarely occurring potentially harmful usages of + is a promising endeavor for static or
dynamic analyses.
4.2.4 Manual Inspection of Potentially Harmful Coercions
To gain further insights into the harmfulness of coercions, we manually inspect a random
sample of all potentially harmful coercions. Out of the total of 1,329 unique code locations
that perform at least one potentially harmful coercion, we inspect a random sample of size 30.
Inspecting these coercions is non-trivial because most web sites obfuscate and minify their
JavaScript code, and because we are not familiar with the implementation of the studied
programs. Therefore, we cannot provide a clear-cut classification of all inspected locations.
To the best of our knowledge, out of the 30 inspected locations:
M. Pradel and K. Sen 1013
45
40
35
Percentage
30
25
20
15
10
5
0
nu nu bo pr ob str ob nu ob +- Ot un +- bo bo +- nu fun nu ar ! [o
mb mb ole imi jec ing jec mb jec ~ s he de ~ b ole ole ~ n mb ct ll R ray bj
er er an tive t E + t E er t E trin rs fine oo an an ull er ion EL RE ect
RE BI BI E Q un Q AR Q g d B lea RE + f BI BI s L Bo
T T trin nu o
L u T u T u Q s obj def prim IT stri IT n L n unc str un g m lea
n n t e i n
nd de d rin ct e iti M g H n un u t ing de be n]
e
efi fin fin g d v de m i o
ne ed e e str
f be n fin r
d d i ng i ne r ed
d
(a) Dynamic occurrences.
20
Percentage
15
10
5
0
ob nu str nu pr nu bo +- +- nu Ot +- bo bo nu ! [o nu fun nu ar un
jec mb ing mb imi mb ole ~ u ~ s mb he ~ b ole ole ll R bj mb ct mb ray de
t E er + er tive er an nd trin er rs oo an an EL ect er ion er R fin
Q AR un RE E RE B efi g BI lea + R s B B B + EL ed
ob IT de L Q L IT ne T
un n fun EL trin ool IT s IT u obj nu BIT
jec H fin un str st nu d de cti nu g ean trin nd ect mb u
t M s ed de ing ring mb fin on m ] g efi er nd
trin fin er ed be ne efi
g ed r d ne
d
(b) Static occurrences.
Figure 7 Kinds of potentially harmful type coercions (top 10 and bottom 10 kinds only).
Horizontal axes: ARITHM, BIT, EQ, and REL mean arithmetic, bitwise, equality, and relational
operations, respectively.
22 are probably correct,
1 is a clear bug,
3 may be buggy, but we cannot confirm any visible misbehavior, and
4 are unclear.
The clear bug was part of Sina.com.cn, a popular Chinese web portal. The page contained
code that encodes various environment settings, such as the current timezone or the version
number of the installed Flash player, into a single number. To access the environment
settings, the page provides helper functions that are supposed to compare the number, m,
against particular bit patterns:
1 timezone: function() {
2 return (m & 16384 == 16384) ? (new Date().getTimezoneOffset() / 60) : ""
3 },
4 flashVer: function() {
5 if (m & 8192 != 8192) {
6 return ""
7 }
8 ..
9 }
ECOOP’15
1014 Implicit Type Conversions in JavaScript
100
80
Percentage
60
40
20
0
nu s n [ s o a n b b b n f a u n b
mb tring umb obje tring bjec rray ull + oole oole oole umb unct rray nde umb oole
er + s er ct S + u t + + s str an an an er ion + a fine er an
+ n tr + s tri n s tr in + + + + + r d + +
um ing trin ng] def tring ing g (h stri num boo und stri ray + u obje fun
be (no g ( + s ine (h (ha arm ng b e lea efi ng (ha nde ct ctio
r ( ne ha tri d ( arm rm f. (ha r ( n n e (h rm f (h n
no ) rm ng ha l. ) rm ha (ha d (h arm f.) ined arm (ha
ne l.) (h rm l.) ) l.) rm rm ar l.) (h f.) rm
) ar f.) f.) f.) m ar f.)
ml f.) mf
.) .)
(a) Dynamic occurrences.
60
50
Percentage
40
30
20
10
0
str n n s b o a [ f n n a b u b n b
ing umb umb tring oole bjec rray obje unct umb ull + rray oole nde oole umb oole
+ s er er + u an t + + a ct S ion er str + s an fine an er an
trin + n + s nd + s stri rra trin + s + u ing trin + n d + + b + o + f
g ( um trin efin trin ng y (h g] trin nde (ha g ( um un oo bje unc
no be g ( ed g ( (ha ar + s g ( fin rm ha be de lea ct ( tio
ne r (n ha h r m tr h e f rml r (h fine n (h ha n (h
) on rml (har arm ml.) f.) ing arm d (h .) .) ar d rm
e) .) mf l.) (h l.) ar
ar mf mf (h arm f.) arm
.) ml . .) arm f.) f.)
.) ) f.)
(b) Static occurrences.
Figure 8 Kinds of plus operations.
Unfortunately, these helper functions always return the same value because equality operators
have precedence over the bitwise & operator. For example, in function flashVer, 8192 !=
8192 always yields false and therefore m & 8192 != 8192 always yields zero, which is coerced
to false. This operation is classified as potentially harmful because it combines a number
and undefined with the & operator. When we tried to reproduce the problem three days after
gathering the data for this study, the corresponding code had been removed from the site.
The three locations that we classify as maybe buggy include two locations that produce
NaN because arithmetic operations are applied to undefined. The other location produces a
string for debugging purposes but concatenates a string with null into "Source-type: null",
which may or may not help with debugging.
The 22 probably correct locations include several recurring patterns:
Ten locations apply a relational or arithmetic operator to a number and a string, where
the string coerces into a number. Representing numbers as strings is not recommended
in general because + does not mean addition. However, the inspected code is correct
because it does not use the + operator.
Three locations concatenate a string to undefined, and then check whether the result
matches a regular expression. This check could be implemented more efficiently, but it is
correct.
Two locations use non-strict equality checks even though they should only match if the
M. Pradel and K. Sen 1015
operands have equal types. These locations are correct but would be easier to understand
if strict equality was used instead.
Our results suggest that even among the potentially harmful coercions, most coercions do
not cause incorrect behavior. Instead, some developers use JavaScript’s coercion semantics
in unusual yet correct ways.
4.2.5 Manual Inspection of Harmless Coercions
To validate the classification from Section 3, we also inspect a random sample of size 30 of all
harmless coercions. We find that all inspected coercions are indeed harmless. As expected
from Figure 4, most of the coercions (26 of 30) are related to conditionals. Moreover, we
identify the following recurring patterns:
Ten coercions are from conditionals that check if a value is defined before using it. Seven
of them check if an object exists before accessing its properties, three check if a function
exists before calling it.
Three coercions check if an optional function argument is defined. JavaScript supports
variadic functions and optional arguments are commonly used.
Three coercions are instances of the initialization pattern discussed in Section 3.4.
Four coercions are due to minified code that uses !0 and !1 as a concise way to express
true and false, respectively.
Overall, we conclude from these results that most coercions that occur in practice are
harmless, which contradicts the common assumption that coercions are error-prone. We draw
this conclusion for two reasons. First, even though our classification cannot rule out that
some coercions classified as harmless cause errors, we believe that most JavaScript developers
are aware of the semantics of the coercions that we classify as harmless. The results of
manually inspecting coercions supports this assumption. Second, under the assumption that
most of the analyzed programs perform the expected behavior, most of the coercions are
likely to be correct, simply because coercions are very prevalent.
4.3 RQ 3: Influence on Understandability
To address the question whether and how type coercions influence the understandability
of code, we analyze two particularly confusing coercion-related properties of code. First,
we analyze the degree of polymorphism of code locations that apply coercions. Second, we
present a detailed analysis of strict and non-strict equality checks, whose semantics often
confuse developers and therefore may harm code understandability. Both analyses are proxy
metrics that estimate to what degree coercions influence understandability. We leave a more
detailed analysis of this question, e.g., through controlled experiments with developers [8],
for future work.
4.3.1 Degree of Polymorphism
Do code locations with coercions always apply the same kind of coercion or do the types
coerced at a particular location differ over time? Figure 9a shows the number of different
types that are coerced at locations where we observed at least one type coercion. Most
locations (86.13%) that apply a coercion always coerce values of the same types into each
other, and very few locations (2.67%) apply three or more different kinds of coercions.
ECOOP’15
1016 Implicit Type Conversions in JavaScript
Operation Coerced types Locs.
Percentage of locations
90
80
70 conditional [object Object], undefined 2011
60 conditional string, undefined 1726
50
40 conditional number, undefined 1034
30
20 conditional function, undefined 746
10
0 conditional array, undefined 685
1 2 3- 6- >1
5 10 0 conditional array, null 410
Number of different types coerced ! [object Object], undefined 399
(a) Percentage of code locations that conditional null, string 301
coerce a particular number of different
! number, undefined 277
types at different times the location
is reached (only locations with coer- conditional number, string 240
cions).
(b) Kinds of locations with polymorphic coercions
(top 10).
Figure 9 Polymorphism of code locations that perform coercions.
100
Locations with equality check (%)
always same
90 mixed
80 always different
70
60
50
40
30
20
10
0
!= !==
or or
== ==
=
Figure 10 Percentage of code locations with (in)equality checks where all observed values have
always the same type, always a different type, or both, respectively.
To understand why polymorphic code locations apply multiple kinds of coercions, Table 9b
list the most prevalent kinds of polymorphic coercion locations and the types that they
coerce. The table shows that the main reason for polymorphic coercions are conditionals
that check whether some non-boolean value is defined. This pattern is common in JavaScript
and a programmer that checks whether some value is defined will expect that the check may
be applied to both a defined or an undefined value.
These results suggest that most coercions do not significantly harm code understandability,
at least not because of polymorphic code locations. A more detailed study on how coercions
influence human understanding of code, e.g., through a human study in the style of [8]
remains for future work.
4.3.2 (In)equality Checks
A particularly intricate situation where JavaScript applies type coercions are (in)equality
checks with == and !=. Guidelines [4] suggest to avoid == and != altogether and to instead
use their “non-evil” twins === and !==. Yet, we find that both kinds of equality checks
are used in practice: In total, we observe 2,026,782 strict equality checks and 3,143,592
M. Pradel and K. Sen 1017
non-strict equality checks during the execution of all programs. To better understand to
what degree coercions at equality checks influence code understandability, we measure how
often a particular code location compares values of the same type (i.e., no coercion) and
values of different types (i.e., coercion). “Same type” here means two types that can be
trivially compared: (i) exactly the same type, (ii) in a == or != check, any type compared to
undefined or null, (iii) in a === or !== check, any type and undefined, or (iv) in a === or !==
check, any non-primitive type and undefined or null.
Figure 10 gives the results for non-strict and strict comparisons. The figure shows that
for almost all locations that compare values with == or !=, the values that occur at runtime
are always of the same type. The results are very similar for comparisons with the strict
operators === and !==.
We draw two conclusions from these results. First, developers seem to use non-strict and
strict equality interchangeably. Under the assumptions that programmers fully understand
the semantics of strict and non-strict equality and that they follow the advice to use strict
equality when comparing values of the same type, one would expect that non-strict equality is
mostly applied to different types. However, the results show that the percentage of locations
that always compare the same types is almost equally high for non-strict and strict equality.
Second, many non-strict equality checks could likely be refactored into strict checks. Since
our study may not consider all paths, we cannot say with certainty that particular non-strict
checks always compare values of the same type. Nevertheless, the results suggest that many
code locations use == and != without any need, and that these location could use === and
!== instead.
5 Threats to Validity
The validity of the conclusions drawn from our results are subject to several threats. First,
since our study is based on dynamic analysis, it ignores coercions triggered on paths not
executed during the analysis. In particular, the results on polymorphic code locations
(Figure 9a) potentially underestimate the number of coercions that occur at a location, and
the results on equality checks (Figure 10) may classify a location as “always same” even
though it may compare values of different types. Second, the classification of coercions
into harmless and potentially harmful may be biased by the subjective experiences of the
authors. We try to minimize this bias by considering informal experience reports of other
JavaScript developers, e.g., in web forums and by comparing the behavior of JavaScript to
other languages. Third, the subject programs of the study may not be representative for a
larger population of programs. By focusing on the most popular web sites, code shared by
multiple popular domains or popular third-party libraries may be overrepresented. Moreover,
some of the benchmark programs contain generated JavaScript code, which may not be
representative for human-written code. Fourth, the results of manually inspecting code that
performs coercions are influenced by our limited ability to understand this code. To reduce
this bias, we use a deobfuscation technique [18]2 to inspect minified and obfuscated code,
and we interactively debug the inspected code locations. Finally, this study is limited to
JavaScript, whose approach to type coercions occupies an extreme spot in the language
spectrum. Our conclusions are for JavaScript and may not extend to other languages.
2
http://jsnice.org
ECOOP’15
1018 Implicit Type Conversions in JavaScript
6 Related Work
Studying how programming languages are used in practice has a long history (for computer
science standards), e.g., going back to a more than 40 years old study of Fortran programs by
D. Knuth [12]. More recently, Richards et al. investigate the dynamic behavior of JavaScript
programs and show that several dynamic features are widely used [21]. A study of Python
programs draws similar conclusions and shows that the assumption that Python programmers
only rarely use dynamic language features is false [10]. Another study [20] provides a detailed
analysis of JavaScript’s notorious eval function. In contrast to our work, none of these
studies investigates type coercions. Nikiforakis et al. describe a large-scale study on how
JavaScript-based web applications include code from third parties, and how these inclusions
influence security [15]. Our work shares with [20] and [15] the idea to analyze in-depth how
a particular language feature is used in the wild. Callau et al. describe a study of dynamic
language features in Smalltalk [2]. In contrast to the above approaches and our work, they
use static analysis. Their work focuses on reflection-related language features and finds that
these features are used infrequently (in 1.76% of all methods).
There are various studies of how Java programmers use Java’s language features. For
example, Tempero et al. study the use of inheritance [24] and of fields [23]. Other work
proposes an infrastructure for querying facts extracted from the source code of a corpus of
Java programs, and shows how to use this infrastructure to answer various questions on how
“typical” Java code is written [6]. Malayeri and Aldrich investigate whether Java programs
could benefit from structural subtyping through a mixture of static and manual analysis [13].
Instead of analyzing how a feature could be used if it existed, we analyze how an already
existing feature is used. To understand how much multiple dispatch is used and could be
used, Muschevici et al. study programs written in six languages that support this feature and
in Java, respectively [14]. All these approaches are based on static analysis of the subject
programs, whereas we use dynamic analysis.
Complementary to studying source code and its execution are studies on how human
subjects react to particular languages or language features. Hanenberg presents such a study
on whether a static type system reduces development time [8]. Our work raises several
questions that could be addressed in similar studies, e.g., how type coercions influence program
understandability, or whether the conciseness of code written with coercions outweighs the
potential error-proneness of coercions during development. Ocariza et al. perform studies of
JavaScript errors [17] and their root causes [16]. They do not identify coercions as a particular
cause of errors, which matches our finding that most coercions do not cause misbehavior.
Several approaches for inferring and checking types in JavaScript programs have been pro-
posed, some of which raise errors on type coercions. A type system for a subset of JavaScript
by Thiemann [25] reports all coercions as errors, presumably under the assumption that
coercions are generally erroneous. A statically typed dialect of JavaScript, called “Depen-
dent JavaScript”, prohibits type coercions because they “often lead to subtle programming
errors” [3]. Our work contradicts this assumption based on empirical evidence showing
that most coercions are harmless. A type analysis by Jensen et al. warns developers about
particular kinds of coercions [11]. We also classify some of them as potentially harmful, e.g.,
coercing undefined to a number. Our results could help to reduce the number of warnings of
their analysis by focusing on kinds of coercions that occur rarely and that are potentially
harmful. Other type inference and checking approaches for JavaScript [9, 7] do not explicitly
discuss if and how they handle coercions. Furr et al. propose a profile-guided static typing
approach for Ruby [5]. They report that some coercions cause type errors that forced Furr
M. Pradel and K. Sen 1019
et al. to refactor code to avoid coercions. Supporting type coercions in a static analyses is
non-trivial and researchers need guidance on whether and how to support coercions. Our
study provides empirical evidence that supporting type coercions is vital, along with guidance
on which coercions to address first.
JavaScript’s strict mode disallows some language features that are generally considered
as dangerous. However, strict mode does not change the semantics of type coercions. A
complementary approach to strict mode, called restrict mode 3 , forbids several potentially
harmful type coercions. At the time of this writing, the coercions allowed by restrict mode
and the coercions that we classify as harmless overlap partly. For example, restrict mode does
not warn about wrapped primitives in conditionals, whereas we classify them as potentially
harmful, but it forbids concatenating strings and arrays, whereas we classify this operation
as harmless. We believe that an empirical study of coercions provides a good base for an
informed decision about which coercions to allow or disallow in a safer JavaScript subset.
7 Conclusion
This paper presents the first in-depth analysis of implicit type conversions in real-world
JavaScript programs. In reference to the title of this paper, we show that most coercions
are “good”, few coercions are “bad”, and some coercions are “ugly” but nevertheless correct.
Based on the results of this study, we draw the following conclusions about real-world
JavaScript programs and analyses that target them:
Type coercions are widely used and analyses that target realistic programs should take
them into account.
Most coercions are not erroneous, and even among those kinds of coercions that seem
error-prone, most coercions do not cause any misbehavior.
The previous two conclusions lead to the third one: Research on static analyses for
JavaScript faces the challenge of checking programs with various harmless coercions.
Most coercions occur in conditions and conditional-related operations, where some value
is coerced into a boolean. Static analyses that address these coercions will handle a large
part of all coercions.
A very small subset of all coercions are potentially harmful because their semantics may
surprise developers. Restricted variants of JavaScript and future language designs can
prohibit these coercions while still supporting most existing code.
Most code locations that coerce types are monomorphic, and most polymorphic locations
are unsurprising because they are related to conditionals that check for undefined val-
ues. That is, polymorphism caused by coercions degrades code understandability only
marginally.
Developers use non-strict and strict equality almost interchangeably. Automated refac-
toring approaches that soundly transform non-strict into strict equality operations seem
a promising direction for future work.
We confirm earlier results showing that the SunSpider and Octane benchmarks do not
accurately represent real-world web sites [21, 19], and we extend the scope of this finding
to type coercions.
Given the increasing importance of JavaScript as a language for web, mobile, desktop, and
server applications, we believe that our work is an important step toward understanding how
3
http://restrictmode.org
ECOOP’15
1020 Implicit Type Conversions in JavaScript
developers use JavaScript and toward aligning future research activities with the real-world
usage of the language.
Acknowledgments This research is supported by the German Federal Ministry of Education and
Research (BMBF) within EC SPRIDE, by the German Research Foundation (DFG) within the
Emmy Noether Project “ConcSys”, by NSF Grants CCF-1423645 and CCF-1409872, by a gift from
Mozilla, and by a Sloan Foundation Fellowship. Thanks to Liang Gong for numerous discussions on
type coercions and JavaScript, and to the anonymous reviewers for their valuable feedback.
References
1 ECMAScript language specification, 5.1 edition, June 2011.
2 Oscar Callaú, Romain Robbes, Éric Tanter, and David Röthlisberger. How (and why)
developers use the dynamic features of programming languages: the case of smalltalk.
Empirical Software Engineering, 18(6):1156–1194, 2013.
3 Ravi Chugh, David Herman, and Ranjit Jhala. Dependent types for JavaScript. In Confer-
ence on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA),
pages 587–606, 2012.
4 Douglas Crockford. JavaScript: The Good Parts. O’Reilly, 2008.
5 Michael Furr, Jong-hoon (David) An, and Jeffrey S. Foster. Profile-guided static typing
for dynamic scripting languages. In Conference on Object-Oriented Programming, Systems,
Languages, and Applications (OOPSLA), pages 283–300. ACM, 2009.
6 Mark Grechanik, Collin McMillan, Luca DeFerrari, Marco Comi, Stefano Crespi-Reghizzi,
Denys Poshyvanyk, Chen Fu, Qing Xie, and Carlo Ghezzi. An empirical investigation
into a large-scale Java open source code repository. In Symposium on Empirical Software
Engineering and Measurement (ESEM), 2010.
7 Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi. Typing local control and state
using flow analysis. In European Symposium on Programming (ESOP), pages 256–275,
2011.
8 Stefan Hanenberg. An experiment about static and dynamic type systems: doubts about
the positive impact of static type systems on development time. In Conference on Object-
Oriented Programming Systems, Languages, and Applications (OOPSLA), pages 22–35,
2010.
9 Phillip Heidegger and Peter Thiemann. Recency types for analyzing scripting languages.
In European Conference on Object-Oriented Programming (ECOOP), pages 200–224, 2010.
10 Alex Holkner and James Harland. Evaluating the dynamic behaviour of Python applica-
tions. In Australasian Computer Science Conference (ACSC), pages 17–25, 2009.
11 Simon Holm Jensen, Anders Møller, and Peter Thiemann. Type analysis for JavaScript.
In Symposium on Static Analysis (SAS), pages 238–255. Springer, 2009.
12 Donald E. Knuth. An empirical study of FORTRAN programs. Software Practice and
Experience, pages 105–133, 1971.
13 Donna Malayeri and Jonathan Aldrich. Is structural subtyping useful? An empirical study.
In European Symposium on Programming (ESOP), pages 95–111, 2009.
14 Radu Muschevici, Alex Potanin, Ewan D. Tempero, and James Noble. Multiple dispatch
in practice. In Conference on Object-Oriented Programming, Systems, Languages, and
Applications (OOPSLA), pages 563–582, 2008.
M. Pradel and K. Sen 1021
15 Nick Nikiforakis, Luca Invernizzi, Alexandros Kapravelos, Steven Van Acker, Wouter
Joosen, Christopher Kruegel, Frank Piessens, and Giovanni Vigna. You are what you
include: large-scale evaluation of remote JavaScript inclusions. In CCS, pages 736–747,
2012.
16 Frolin S. Ocariza Jr., Kartik Bajaj, Karthik Pattabiraman, and Ali Mesbah. An empirical
study of client-side JavaScript bugs. In Symposium on Empirical Software Engineering and
Measurement (ESEM), pages 55–64, 2013.
17 Frolin S. Ocariza Jr., Karthik Pattabiraman, and Benjamin G. Zorn. JavaScript errors in
the wild: An empirical study. In International Symposium on Software Reliability Engi-
neering (ISSRE), pages 100–109, 2011.
18 Veselin Raychev, Martin T. Vechev, and Andreas Krause. Predicting program properties
from "big code". In Principles of Programming Languages (POPL), pages 111–124, 2015.
19 Gregor Richards, Andreas Gal, Brendan Eich, and Jan Vitek. Automated construction of
JavaScript benchmarks. In Conference on Object-Oriented Programming, Systems, Lan-
guages, and Applications (OOPSLA), pages 677–694, 2011.
20 Gregor Richards, Christian Hammer, Brian Burg, and Jan Vitek. The eval that men do -
a large-scale study of the use of eval in JavaScript applications. In European Conference
on Object-Oriented Programming (ECOOP), pages 52–78, 2011.
21 Gregor Richards, Sylvain Lebresne, Brian Burg, and Jan Vitek. An analysis of the dynamic
behavior of JavaScript programs. In Conference on Programming Language Design and
Implementation (PLDI), pages 1–12, 2010.
22 Koushik Sen, Swaroop Kalasapur, Tasneem Brutch, and Simon Gibbs. Jalangi: A selec-
tive record-replay and dynamic analysis framework for JavaScript. In European Software
Engineering Conference and Symposium on the Foundations of Software Engineering (ES-
EC/FSE), pages 488–498, 2013.
23 Ewan D. Tempero. How fields are used in Java: An empirical study. In Australian Software
Engineering Conference (ASWEC), pages 91–100, 2009.
24 Ewan D. Tempero, James Noble, and Hayden Melton. How do Java programs use inher-
itance? An empirical study of inheritance in Java software. In European Conference on
Object-Oriented Programming (ECOOP), pages 667–691. Springer, 2008.
25 Peter Thiemann. Towards a type system for analyzing JavaScript programs. In European
Symposium on Programming (ESOP), pages 408–422, 2005.
ECOOP’15