While [] + [] is an empty string, [] + {} is "[object Object]", and {} + [] is 0. Why is {} + {} NaN?
> {} + {}
NaN
My question isn't why ({} + {}).toString() is "[object Object][object Object]" while NaN.toString() is "NaN", this part has an answer here already.
My question is why does this happen only on the client side? On the server side (Node.js) {} + {} is "[object Object][object Object]".
> {} + {}
'[object Object][object Object]'
Summarizing:
On the client side:
[] + [] // Returns ""
[] + {} // Returns "[object Object]"
{} + [] // Returns 0
{} + {} // Returns NaN
NaN.toString() // Returns "NaN"
({} + {}).toString() // Returns "[object Object][object Object]"
var a = {} + {}; // 'a' will be "[object Object][object Object]"
In Node.js:
[] + [] // Returns "" (like on the client)
[] + {} // Returns "[object Object]" (like on the client)
{} + [] // Returns "[object Object]" (not like on the client)
{} + {} // Returns "[object Object][object Object]" (not like on the client)
Updated note: this has been fixed in Chrome 49.
Very interesting question! Let's dig in.
The root cause
The root of the difference is in how Node.js evaluates these statements vs. how the Chrome development tools do.
What Node.js does
Node.js uses the repl module for this.
From the Node.js REPL source code:
self.eval(
'(' + evalCmd + ')',
self.context,
'repl',
function (e, ret) {
if (e && !isSyntaxError(e))
return finish(e);
if (typeof ret === 'function' && /^[\r\n\s]*function/.test(evalCmd) || e) {
// Now as statement without parens.
self.eval(evalCmd, self.context, 'repl', finish);
}
else {
finish(null, ret);
}
}
);
This acts just like running ({}+{}) in the Chrome developer tools, which also produces "[object Object][object Object]" as you'd expect.
What the chrome developer tools do
On the other hand Chrome dveloper tools does the following:
try {
if (injectCommandLineAPI && inspectedWindow.console) {
inspectedWindow.console._commandLineAPI = new CommandLineAPI(this._commandLineAPIImpl, isEvalOnCallFrame ? object : null);
expression = "with ((window && window.console && window.console._commandLineAPI) || {}) {\n" + expression + "\n}";
}
var result = evalFunction.call(object, expression);
if (objectGroup === "console")
this._lastResult = result;
return result;
}
finally {
if (injectCommandLineAPI && inspectedWindow.console)
delete inspectedWindow.console._commandLineAPI;
}
So basically, it performs a call on the object with the expression. The expression being:
with ((window && window.console && window.console._commandLineAPI) || {}) {
{}+{};// <-- This is your code
}
So, as you can see, the expression is being evaluted directly, without the wrapping parenthesis.
Why Node.js acts differently
Node.js's source justifies this:
// This catches '{a : 1}' properly.
Node did not always act like this. Here is the actual commit that changed it. Ryan left the following comment on the change: "Improve how REPL commands are evaled" with an example of the difference.
Rhino
Update - OP was interested in how Rhino behaves (and why it behaves like the Chrome devtools and unlike nodejs).
Rhino uses a completely different JS engine unlike the Chrome developer tools and Node.js's REPL which both use V8.
Here is the basic pipe line of what happens when you eval a JavaScript command with Rhino in the Rhino shell.
The shell runs org.mozilla.javascript.tools.shell.main.
In turn, it calls this new IProxy(IProxy.EVAL_INLINE_SCRIPT); for example, if the code was passed directly with the inline switch -e.
This hits IProxy's run method.
It invokes evalInlineScript (src). This simply compiles the string and evals it.
Basically:
Script script = cx.compileString(scriptText, "<command>", 1, null);
if (script != null) {
script.exec(cx, getShellScope()); // <- just an eval
}
Out of the three, Rhino's shell is the one that does the closest thing to an actual eval without any wrapping. Rhino's is the closest to an actual eval() statement and you can expect it to behave exactly like eval would.
