Kyle Simpson | You dont know Javascript (Episode 20)
If you are new to programming or JS, be aware that these books are not intended as a gentle intro to Javascript.
· 24 phút đọc.
By now you should have a decent grasp of the nesting of scopes, from the global scope downward—called a program’s scope chain.
But just knowing which scope a variable comes from is only part of the story. If a variable declaration appears past the first statement of a scope, how will any references to that identifier before the declaration behave? What happens if you try to declare the same variable twice in a scope?
JS’s particular flavor of lexical scope is rich with nuance in how and when variables come into existence and become available to the program.
When Can I Use a Variable?
At what point does a variable become available to use within its scope? There may seem to be an obvious answer: after the variable has been declared/created. Right? Not quite.
Consider:
greeting();
// Hello!
function greeting() {
console.log("Hello!");
}
This code works fine. You may have seen or even written code like it before. But did you ever wonder how or why it works? Specifically, why can you access the identifier greeting
from line 1 (to retrieve and execute a function reference), even though the greeting()
function declaration doesn’t occur until line 4?
Recall Chapter 1 points out that all identifiers are registered to their respective scopes during compile time. Moreover, every identifier is created at the beginning of the scope it belongs to, every time that scope is entered.
The term most commonly used for a variable being visible from the beginning of its enclosing scope, even though its declaration may appear further down in the scope, is called hoisting.
But hoisting alone doesn’t fully answer the question. We can see an identifier called greeting
from the beginning of the scope, but why can we call the greeting()
function before it’s been declared?
In other words, how does the variable greeting
have any value (the function reference) assigned to it, from the moment the scope starts running? The answer is a special characteristic of formal function
declarations, called function hoisting. When a function
declaration’s name identifier is registered at the top of its scope, it’s additionally auto-initialized to that function’s reference. That’s why the function can be called throughout the entire scope!
One key detail is that both function hoisting and var
-flavored variable hoisting attach their name identifiers to the nearest enclosing function scope (or, if none, the global scope), not a block scope.
| NOTE: | | : – - | | Declarations with let
and const
still hoist (see the TDZ discussion later in this chapter). But these two declaration forms attach to their enclosing block rather than just an enclosing function as with var
and function
declarations. See “Scoping with Blocks” in Chapter 6 for more information. |
Hoisting: Declaration vs. Expression
Function hoisting only applies to formal function
declarations (specifically those which appear outside of blocks—see “FiB” in Chapter 6), not to function
expression assignments. Consider:
greeting();
// TypeError
var greeting = function greeting() {
console.log("Hello!");
};
Line 1 (greeting();
) throws an error. But the kind of error thrown is very important to notice. A TypeError
means we’re trying to do something with a value that is not allowed. Depending on your JS environment, the error message would say something like, “‘undefined’ is not a function,” or more helpfully, “‘greeting’ is not a function.”
Notice that the error is not a ReferenceError
. JS isn’t telling us that it couldn’t find greeting
as an identifier in the scope. It’s telling us that greeting
was found but doesn’t hold a function reference at that moment. Only functions can be invoked, so attempting to invoke some non-function value results in an error.
But what does greeting
hold, if not the function reference?
In addition to being hoisted, variables declared with var
are also automatically initialized to undefined
at the beginning of their scope—again, the nearest enclosing function, or the global. Once initialized, they’re available to be used (assigned to, retrieved from, etc.) throughout the whole scope.
So on that first line, greeting
exists, but it holds only the default undefined
value. It’s not until line 4 that greeting
gets assigned the function reference.
Pay close attention to the distinction here. A function
declaration is hoisted and initialized to its function value (again, called function hoisting). A var
variable is also hoisted, and then auto-initialized to undefined
. Any subsequent function
expression assignments to that variable don’t happen until that assignment is processed during runtime execution.
In both cases, the name of the identifier is hoisted. But the function reference association isn’t handled at initialization time (beginning of the scope) unless the identifier was created in a formal function
declaration.
Variable Hoisting
Let’s look at another example of variable hoisting:
greeting = "Hello!";
console.log(greeting);
// Hello!
var greeting = "Howdy!";
Though greeting
isn’t declared until line 5, it’s available to be assigned to as early as line 1. Why?
There’s two necessary parts to the explanation:
the identifier is hoisted, and it’s automatically initialized to the value undefined
from the top of the scope.
| NOTE: | | : – - | | Using variable hoisting of this sort probably feels unnatural, and many readers might rightly want to avoid relying on it in their programs. But should all hoisting (including function hoisting) be avoided? We’ll explore these different perspectives on hoisting in more detail in Appendix A. |
Hoisting: Yet Another Metaphor
Chapter 2 was full of metaphors (to illustrate scope), but here we are faced with yet another: hoisting itself. Rather than hoisting being a concrete execution step the JS engine performs, it’s more useful to think of hoisting as a visualization of various actions JS takes in setting up the program before execution.
The typical assertion of what hoisting means: lifting—like lifting a heavy weight upward—any identifiers all the way to the top of a scope. The explanation often asserted is that the JS engine will actually rewrite that program before execution, so that it looks more like this:
var greeting; // hoisted declaration
greeting = "Hello!"; // the original line 1
console.log(greeting); // Hello!
greeting = "Howdy!"; // `var` is gone!
The hoisting (metaphor) proposes that JS pre-processes the original program and re-arranges it a bit, so that all the declarations have been moved to the top of their respective scopes, before execution. Moreover, the hoisting metaphor asserts that function
declarations are, in their entirety, hoisted to the top of each scope. Consider:
studentName = "Suzy";
greeting();
// Hello Suzy!
function greeting() {
console.log(`Hello ${ studentName }!`);
}
var studentName;
The “rule” of the hoisting metaphor is that function declarations are hoisted first, then variables are hoisted immediately after all the functions. Thus, the hoisting story suggests that program is re-arranged by the JS engine to look like this:
function greeting() {
console.log(`Hello ${ studentName }!`);
}
var studentName;
studentName = "Suzy";
greeting();
// Hello Suzy!
This hoisting metaphor is convenient. Its benefit is allowing us to hand wave over the magical look-ahead pre-processing necessary to find all these declarations buried deep in scopes and somehow move (hoist) them to the top; we can just think about the program as if it’s executed by the JS engine in a single pass, top-down.
Single-pass definitely seems more straightforward than Chapter 1’s assertion of a two-phase processing.
Hoisting as a mechanism for re-ordering code may be an attractive simplification, but it’s not accurate. The JS engine doesn’t actually re-arrange the code. It can’t magically look ahead and find declarations; the only way to accurately find them, as well as all the scope boundaries in the program, would be to fully parse the code.
Guess what parsing is? The first phase of the two-phase processing! There’s no magical mental gymnastics that gets around that fact.
So if the hoisting metaphor is (at best) inaccurate, what should we do with the term? I think it’s still useful—indeed, even members of TC39 regularly use it!—but I don’t think we should claim it’s an actual re-arrangement of source code.
| WARNING: | | : – - | | Incorrect or incomplete mental models often still seem sufficient because they can occasionally lead to accidental right answers. But in the long run it’s harder to accurately analyze and predict outcomes if your thinking isn’t particularly aligned with how the JS engine works. |
I assert that hoisting should be used to refer to the compile-time operation of generating runtime instructions for the automatic registration of a variable at the beginning of its scope, each time that scope is entered.
That’s a subtle but important shift, from hoisting as a runtime behavior to its proper place among compile-time tasks.
Re-declaration?
What do you think happens when a variable is declared more than once in the same scope? Consider:
var studentName = "Frank";
console.log(studentName);
// Frank
var studentName;
console.log(studentName); // ?
What do you expect to be printed for that second message? Many believe the second var studentName
has re-declared the variable (and thus “reset” it), so they expect undefined
to be printed.
But is there such a thing as a variable being “re-declared” in the same scope? No.
If you consider this program from the perspective of the hoisting metaphor, the code would be re-arranged like this for execution purposes:
var studentName;
var studentName; // clearly a pointless no-op!
studentName = "Frank";
console.log(studentName);
// Frank
console.log(studentName);
// Frank
Since hoisting is actually about registering a variable at the beginning of a scope, there’s nothing to be done in the middle of the scope where the original program actually had the second var studentName
statement. It’s just a no-op(eration), a pointless statement.
| TIP: | | : – - | | In the style of the conversation narrative from Chapter 2, Compiler would find the second var
declaration statement and ask the Scope Manager if it had already seen a studentName
identifier; since it had, there wouldn’t be anything else to do. |
It’s also important to point out that var studentName;
doesn’t mean var studentName = undefined;
, as most assume. Let’s prove they’re different by considering this variation of the program:
var studentName = "Frank";
console.log(studentName); // Frank
var studentName;
console.log(studentName); // Frank < – - still!
// let's add the initialization explicitly
var studentName = undefined;
console.log(studentName); // undefined < – - see?
See how the explicit = undefined
initialization produces a different outcome than assuming it happens implicitly when omitted? In the next section, we’ll revisit this topic of initialization of variables from their declarations.
A repeated var
declaration of the same identifier name in a scope is effectively a do-nothing operation. Here’s another illustration, this time across a function of the same name:
var greeting;
function greeting() {
console.log("Hello!");
}
// basically, a no-op
var greeting;
typeof greeting; // "function"
var greeting = "Hello!";
typeof greeting; // "string"
The first greeting
declaration registers the identifier to the scope, and because it’s a var
the auto-initialization will be undefined
. The function
declaration doesn’t need to re-register the identifier, but because of function hoisting it overrides the auto-initialization to use the function reference. The second var greeting
by itself doesn’t do anything since greeting
is already an identifier and function hoisting already took precedence for the auto-initialization.
Actually assigning "Hello!"
to greeting
changes its value from the initial function greeting()
to the string; var
itself doesn’t have any effect.
What about repeating a declaration within a scope using let
or const
?
let studentName = "Frank";
console.log(studentName);
let studentName = "Suzy";
This program will not execute, but instead immediately throw a SyntaxError
. Depending on your JS environment, the error message will indicate something like: “studentName has already been declared.” In other words, this is a case where attempted “re-declaration” is explicitly not allowed!
It’s not just that two declarations involving let
will throw this error. If either declaration uses let
, the other can be either let
or var
, and the error will still occur, as illustrated with these two variations:
var studentName = "Frank";
let studentName = "Suzy";
and:
let studentName = "Frank";
var studentName = "Suzy";
In both cases, a SyntaxError
is thrown on the second declaration. In other words, the only way to “re-declare” a variable is to use var
for all (two or more) of its declarations.
But why disallow it? The reason for the error is not technical per se, as var
“re-declaration” has always been allowed; clearly, the same allowance could have been made for let
.
It’s really more of a “social engineering” issue. “Re-declaration” of variables is seen by some, including many on the TC39 body, as a bad habit that can lead to program bugs. So when ES6 introduced let
, they decided to prevent “re-declaration” with an error.
| NOTE: | | : – - | | This is of course a stylistic opinion, not really a technical argument. Many developers agree with the position, and that’s probably in part why TC39 included the error (as well as let
conforming to const
). But a reasonable case could have been made that staying consistent with var
’s precedent was more prudent, and that such opinion-enforcement was best left to opt-in tooling like linters. In Appendix A, we’ll explore whether var
(and its associated behavior, like “re-declaration”) can still be useful in modern JS. |
When Compiler asks Scope Manager about a declaration, if that identifier has already been declared, and if either/both declarations were made with let
, an error is thrown. The intended signal to the developer is “Stop relying on sloppy re-declaration!”
Constants?
The const
keyword is more constrained than let
. Like let
, const
cannot be repeated with the same identifier in the same scope. But there’s actually an overriding technical reason why that sort of “re-declaration” is disallowed, unlike let
which disallows “re-declaration” mostly for stylistic reasons.
The const
keyword requires a variable to be initialized, so omitting an assignment from the declaration results in a SyntaxError
:
const empty; // SyntaxError
const
declarations create variables that cannot be re-assigned:
const studentName = "Frank";
console.log(studentName);
// Frank
studentName = "Suzy"; // TypeError
The studentName
variable cannot be re-assigned because it’s declared with a const
.
| WARNING: | | : – - | | The error thrown when re-assigning studentName
is a TypeError
, not a SyntaxError
. The subtle distinction here is actually pretty important, but unfortunately far too easy to miss. Syntax errors represent faults in the program that stop it from even starting execution. Type errors represent faults that arise during program execution. In the preceding snippet, "Frank"
is printed out before we process the re-assignment of studentName
, which then throws the error. |
So if const
declarations cannot be re-assigned, and const
declarations always require assignments, then we have a clear technical reason why const
must disallow any “re-declarations”: any const
“re-declaration” would also necessarily be a const
re-assignment, which can’t be allowed!
const studentName = "Frank";
// obviously this must be an error
const studentName = "Suzy";
Since const
“re-declaration” must be disallowed (on those technical grounds), TC39 essentially felt that let
“re-declaration” should be disallowed as well, for consistency. It’s debatable if this was the best choice, but at least we have the reasoning behind the decision.
Loops
So it’s clear from our previous discussion that JS doesn’t really want us to “re-declare” our variables within the same scope. That probably seems like a straightforward admonition, until you consider what it means for repeated execution of declaration statements in loops. Consider:
var keepGoing = true;
while (keepGoing) {
let value = Math.random();
if (value > 0.5) {
keepGoing = false;
}
}
Is value
being “re-declared” repeatedly in this program? Will we get errors thrown? No.
All the rules of scope (including “re-declaration” of let
-created variables) are applied per scope instance. In other words, each time a scope is entered during execution, everything resets.
Each loop iteration is its own new scope instance, and within each scope instance, value
is only being declared once. So there’s no attempted “re-declaration,” and thus no error. Before we consider other loop forms, what if the value
declaration in the previous snippet were changed to a var
?
var keepGoing = true;
while (keepGoing) {
var value = Math.random();
if (value > 0.5) {
keepGoing = false;
}
}
Is value
being “re-declared” here, especially since we know var
allows it? No. Because var
is not treated as a block-scoping declaration (see Chapter 6), it attaches itself to the global scope. So there’s just one value
variable, in the same scope as keepGoing
(global scope, in this case). No “re-declaration” here, either!
One way to keep this all straight is to remember that var
, let
, and const
keywords are effectively removed from the code by the time it starts to execute. They’re handled entirely by the compiler.
If you mentally erase the declarator keywords and then try to process the code, it should help you decide if and when (re-)declarations might occur.
What about “re-declaration” with other loop forms, like for
-loops?
for (let i = 0; i < 3; i++) {
let value = i 10;
console.log(`${ i }: ${ value }`);
}
// 0: 0
// 1: 10
// 2: 20
It should be clear that there’s only one value
declared per scope instance. But what about i
? Is it being “re-declared”?
To answer that, consider what scope i
is in. It might seem like it would be in the outer (in this case, global) scope, but it’s not. It’s in the scope of for
-loop body, just like value
is. In fact, you could sorta think about that loop in this more verbose equivalent form:
{
// a fictional variable for illustration
let $$i = 0;
for ( / nothing /; $$i < 3; $$i++) {
// here's our actual loop `i`!
let i = $$i;
let value = i 10;
console.log(`${ i }: ${ value }`);
}
// 0: 0
// 1: 10
// 2: 20
}
Now it should be clear: the i
and value
variables are both declared exactly once per scope instance. No “re-declaration” here.
What about other for
-loop forms?
for (let index in students) {
// this is fine
}
for (let student of students) {
// so is this
}
Same thing with for…in
and for…of
loops: the declared variable is treated as inside the loop body, and thus is handled per iteration (aka, per scope instance). No “re-declaration.”
OK, I know you’re thinking that I sound like a broken record at this point. But let’s explore how const
impacts these looping constructs. Consider:
var keepGoing = true;
while (keepGoing) {
// ooo, a shiny constant!
const value = Math.random();
if (value > 0.5) {
keepGoing = false;
}
}
Just like the let
variant of this program we saw earlier, const
is being run exactly once within each loop iteration, so it’s safe from “re-declaration” troubles. But things get more complicated when we talk about for
-loops.
for…in
and for…of
are fine to use with const
:
for (const index in students) {
// this is fine
}
for (const student of students) {
// this is also fine
}
But not the general for
-loop:
for (const i = 0; i < 3; i++) {
// oops, this is going to fail with
// a Type Error after the first iteration
}
What’s wrong here? We could use let
just fine in this construct, and we asserted that it creates a new i
for each loop iteration scope, so it doesn’t even seem to be a “re-declaration.”
Let’s mentally “expand” that loop like we did earlier:
{
// a fictional variable for illustration
const $$i = 0;
for ( ; $$i < 3; $$i++) {
// here's our actual loop `i`!
const i = $$i;
//…
}
}
Do you spot the problem? Our i
is indeed just created once inside the loop. That’s not the problem. The problem is the conceptual $$i
that must be incremented each time with the $$i++
expression. That’s re-assignment (not “re-declaration”), which isn’t allowed for constants.
Remember, this “expanded” form is only a conceptual model to help you intuit the source of the problem. You might wonder if JS could have effectively made the const $$i = 0
instead into let $ii = 0
, which would then allow const
to work with our classic for
-loop? It’s possible, but then it could have introduced potentially surprising exceptions to for
-loop semantics.
For example, it would have been a rather arbitrary (and likely confusing) nuanced exception to allow i++
in the for
-loop header to skirt strictness of the const
assignment, but not allow other re-assignments of i
inside the loop iteration, as is sometimes useful.
The straightforward answer is: const
can’t be used with the classic for
-loop form because of the required re-assignment.
Interestingly, if you don’t do re-assignment, then it’s valid:
var keepGoing = true;
for (const i = 0; keepGoing; / nothing here / ) {
keepGoing = (Math.random() > 0.5);
//…
}
That works, but it’s pointless. There’s no reason to declare i
in that position with a const
, since the whole point of such a variable in that position is to be used for counting iterations. Just use a different loop form, like a while
loop, or use a let
!
Uninitialized Variables (aka, TDZ)
With var
declarations, the variable is “hoisted” to the top of its scope. But it’s also automatically initialized to the undefined
value, so that the variable can be used throughout the entire scope.
However, let
and const
declarations are not quite the same in this respect.
Consider:
console.log(studentName);
// ReferenceError
let studentName = "Suzy";
The result of this program is that a ReferenceError
is thrown on the first line. Depending on your JS environment, the error message may say something like: “Cannot access studentName before initialization.”
| NOTE: | | : – - | | The error message as seen here used to be much more vague or misleading. Thankfully, several of us in the community were successfully able to lobby for JS engines to improve this error message so it more accurately tells you what’s wrong! |
That error message is quite indicative of what’s wrong: studentName
exists on line 1, but it’s not been initialized, so it cannot be used yet. Let’s try this:
studentName = "Suzy"; // let's try to initialize it!
// ReferenceError
console.log(studentName);
let studentName;
Oops. We still get the ReferenceError
, but now on the first line where we’re trying to assign to (aka, initialize!) this so-called “uninitialized” variable studentName
. What’s the deal?
The real question is, how do we initialize an uninitialized variable? For let
/const
, the only way to do so is with an assignment attached to a declaration statement. An assignment by itself is insufficient! Consider:
let studentName = "Suzy";
console.log(studentName); // Suzy
Here, we are initializing the studentName
(in this case, to "Suzy"
instead of undefined
) by way of the let
declaration statement form that’s coupled with an assignment.
Alternatively:
//…
let studentName;
// or:
// let studentName = undefined;
//…
studentName = "Suzy";
console.log(studentName);
// Suzy
| NOTE: | | : – - | | That’s interesting! Recall from earlier, we said that var studentName;
is not the same as var studentName = undefined;
, but here with let
, they behave the same. The difference comes down to the fact that var studentName
automatically initializes at the top of the scope, where let studentName
does not. |
Remember that we’ve asserted a few times so far that Compiler ends up removing any var
/let
/const
declarators, replacing them with the instructions at the top of each scope to register the appropriate identifiers.
So if we analyze what’s going on here, we see that an additional nuance is that Compiler is also adding an instruction in the middle of the program, at the point where the variable studentName
was declared, to handle that declaration’s auto-initialization. We cannot use the variable at any point prior to that initialization occurring. The same goes for const
as it does for let
.
The term coined by TC39 to refer to this period of time from the entering of a scope to where the auto-initialization of the variable occurs is: Temporal Dead Zone (TDZ).
The TDZ is the time window where a variable exists but is still uninitialized, and therefore cannot be accessed in any way. Only the execution of the instructions left by Compiler at the point of the original declaration can do that initialization. After that moment, the TDZ is done, and the variable is free to be used for the rest of the scope.
A var
also technically has a TDZ, but it’s zero in length and thus unobservable to our programs! Only let
and const
have an observable TDZ.
By the way, “temporal” in TDZ does indeed refer to time not position in code. Consider:
askQuestion();
// ReferenceError
let studentName = "Suzy";
function askQuestion() {
console.log(`${ studentName }, do you know?`);
}
Even though positionally the console.log(…)
referencing studentName
comes after the let studentName
declaration, timing wise the askQuestion()
function is invoked before the let
statement is encountered, while studentName
is still in its TDZ! Hence the error.
There’s a common misconception that TDZ means let
and const
do not hoist. This is an inaccurate, or at least slightly misleading, claim. They definitely hoist.
The actual difference is that let
/const
declarations do not automatically initialize at the beginning of the scope, the way var
does. The debate then is if the auto-initialization is part of hoisting, or not? I think auto-registration of a variable at the top of the scope (i.e., what I call “hoisting”) and auto-initialization at the top of the scope (to undefined
) are distinct operations and shouldn’t be lumped together under the single term “hoisting.”
We’ve already seen that let
and const
don’t auto-initialize at the top of the scope. But let’s prove that let
and const
do hoist (auto-register at the top of the scope), courtesy of our friend shadowing (see “Shadowing” in Chapter 3):
var studentName = "Kyle";
{
console.log(studentName);
// ?
//…
let studentName = "Suzy";
console.log(studentName);
// Suzy
}
What’s going to happen with the first console.log(…)
statement? If let studentName
didn’t hoist to the top of the scope, then the first console.log(…)
should print "Kyle"
, right? At that moment, it would seem, only the outer studentName
exists, so that’s the variable console.log(…)
should access and print.
But instead, the first console.log(…)
throws a TDZ error, because in fact, the inner scope’s studentName
was hoisted (auto-registered at the top of the scope). What didn’t happen (yet!) was the auto-initialization of that inner studentName
; it’s still uninitialized at that moment, hence the TDZ violation!
So to summarize, TDZ errors occur because let
/const
declarations do hoist their declarations to the top of their scopes, but unlike var
, they defer the auto-initialization of their variables until the moment in the code’s sequencing where the original declaration appeared. This window of time (hint: temporal), whatever its length, is the TDZ.
How can you avoid TDZ errors?
My advice: always put your let
and const
declarations at the top of any scope. Shrink the TDZ window to zero (or near zero) length, and then it’ll be moot.
But why is TDZ even a thing? Why didn’t TC39 dictate that let
/const
auto-initialize the way var
does? Just be patient, we’ll come back to explore the why of TDZ in Appendix A.
Finally Initialized
Working with variables has much more nuance than it seems at first glance. Hoisting, (re)declaration, and the TDZ are common sources of confusion for developers, especially those who have worked in other languages before coming to JS. Before moving on, make sure your mental model is fully grounded on these aspects of JS scope and variables.
Hoisting is generally cited as an explicit mechanism of the JS engine, but it’s really more a metaphor to describe the various ways JS handles variable declarations during compilation. But even as a metaphor, hoisting offers useful structure for thinking about the life-cycle of a variable—when it’s created, when it’s available to use, when it goes away.
Declaration and re-declaration of variables tend to cause confusion when thought of as runtime operations. But if you shift to compile-time thinking for these operations, the quirks and shadows diminish.
The TDZ (temporal dead zone) error is strange and frustrating when encountered. Fortunately, TDZ is relatively straightforward to avoid if you’re always careful to place let
/const
declarations at the top of any scope.
As you successfully navigate these twists and turns of variable scope, the next chapter will lay out the factors that guide our decisions to place our declarations in various scopes, especially nested blocks.