Lua: Good, bad, and ugly parts - ZeroBrane

I have been programming in Lua for about 9 months and it seemed like a good time to pause and reflect on my experience with it. I have used various languages over the years -- Perl (soaplite.com, and other projects, including my current consulting work), C (DHCPLite and a ping-pong juggling robot), JavaScript (experiments with Google Maps and canvas), MATLAB (ping-pong juggling robot), and others, from Turbo Pascal to F# -- and it was interesting to see how Lua compares to the other languages I've worked with. I have done different types of projects in Lua: a remote debugger (MobDebug), extending a Lua IDE (ZeroBrane Studio), a mobile app (LuaRemote), several educational scripts (EduPack), and a demo of drawing on browser canvas with Lua.

I have come across several detailed lists that mention good and not-so-good parts of Lua (for example, Lua benefits, why Lua, why Lua is not more widely used, advantages of Lua, Lua good/bad, Lua vs. JavaScript, and Lua Gotchas), but I found that some of the features that tripped me or that I cared about were not listed, so I put together my own list. It is far from being comprehensive and some aspects of the language are not covered (for example, math and string libraries), but it captures the gist of my experience with the language.

Good

• Small: 20000 lines of C code that can be built into a 182K executable interpreter (under Linux).
• Portable: builds on any platform with an ANSI C compiler. You can see it running on almost anything from microcontrollers and Lego Minstorms NXT, to mobile platforms, to game consoles, to a browser (translated to JavaScript).
• Embedded and extensible language that provides a straightforward interface to/from C/C++.
• Sufficiently fast: performs well comparing to other languages and has a JIT compiler that noticeably improves performance on many tasks; those who may still be not satisfied with the performance, can implement critical parts in C and, given the ease of integration with C, still benefit from other good aspects. [Updated 3/9/2013] replaced shootout results that are no longer available with benchmarksgame.
• Well documented: reference manual, book, wiki, 6-page short reference and more.
• Friendly and enthusiastic community. Between the excellent documentation, the wiki, the mailing list, and StackOverflow, I didn't have any issues finding answers to my questions.
• Clean and simple syntax suitable for beginners and accessible to non-programmers. Lua has borrowed most of its control syntax from Modula, the descendent of Pascal, which was widely used in education as an introductory language. I still remember using early versions of Philippe Kahn's fast and elegant Turbo PascalIDE.
• Integrated interpreter: just run lua from the command line.
• Native support for coroutines to implement iterators and non-preemptive multi-threading.
• Incremental garbage collector that has low latency, no additional memory cost, little implementation complexity, and support for weak tables.
• Powerful heterogeneous tables that store values of any type (except nil) and can be indexed by values of any type (except nil): [1, 2, 5, foo = "bar", [func] = "something", ["some spaces"] = value()].
• Lexical scoping.
• Functional programming with first class functions and closures.
• Tail calls: return functioncall().
• Recursive functions don't need to be pre-declared: local function foo() ... foo() ... end; note this doesn't work with local foo = function() ... foo() ... end.
• Functions return multiple values: return 1, 2, 3. The caller can expect any number of values returned: if less than three is expected, the rest is discarded and if more than three is expected, the rest is nil-initialized.
• Functions allow variable number of parameters with function foo(...) local args = {...}; bar(param, ...) end.
• Tables can be "unpacked" into a list of parameters with unpack (or table.unpack in Lua 5.2): print(unpack({1, 2, 3})) prints 1 2 3.
• Manipulating environments (getfenv and setfenv in Lua 5.1 and _ENV manipulation in Lua 5.2), which allows building sandboxes among other things.
• Assignment to a list of variables: local a, b, c = 1, 2, x, y = y, x, or a, b = foo().
• Multiline strings (using [[...]]; can be enclosed with [[...[=[...]=]...]]) and comments (--[[...]]).
• Semicolon as a statement separator is optional (mostly used to resolve ambiguous cases as in a = f; (g).x(a)).
• Overloading using metatables.
• Metaprogramming to do things from getting and modifying an abstract syntax tree to creating a new syntax for your DSL.
• The for statement has two forms: generic (loops over iterators: for a in iter() do ... end) and numeric (loops over numbers: for a = 1, 5, 0.1 do ... end); the numeric one supports all numeric values for steps (not just integers).
• Syntactic sugar for function calls (f'string', f"string", f[[string]], and f{table}) and method calls (obj:m()).
• Simple yet powerful debug library.

Different

• Tables and strings are indexed from 1 rather than 0.
• Assigning nil as a value removes the element from a table. This is consistent with returning nil for non-existing element, so it makes no difference whether the element does not exist or exists with a value of nil. a = {b = nil} produces an empty table.
• No integers as a separate numeric type; the number type represent real numbers.
• No classes; object-orientation is implemented using tables and functions; inheritance is implemented using the metatable mechanism.
• Method calls are implemented using object:method(args) notation, which is the same as object.method(object, args) notation, but with object evaluated only once.
• nil and false are the only false values; 0, 0.0, "0" and all other values evaluate as true.
• Non-equality operator is ~= (for example, if a ~= 1 then ... end).
• not, or, and keywords used for logical operators.
• Assignments are statements, which means there is no a=b=1 or if (a=1) then ... end.
• No a+=1, a++, or similar shorthand forms.
• No continue statement, although there is an explanation and a number of alternatives, like using repeat break until true inside the loop to break out of or a goto statement introduced in Lua 5.2.
• No switch statement.
• Brackets may be required in some contexts; for example, a = {}; a.field works, but {}.field doesn't; the latter needs to be specified as ({}).field.
• A control variable in a loop is localized by default and is not available after the loop.
• Limit and step values in the numeric for loop are cached; this means that in for i = init(), limit(), step() do ... end all three functions init, limit, and step are called once before the loop is executed.
• Conditionals and other control structures require no brackets.
• Strings and numbers are automatically converted (if string is used where a number is expected and vice versa), but not in equality/inequality comparisons: 0 == "0" is false, {} ~= 1 is true, and foo["0"] and foo[0] refer to two different keys in the table; other relational operators generate errors on comparing values of different types.
• Both commas and semicolons can be used in table literals; both can be placed before the closing curly bracket as an optional separator: a = {a = 1, b = 2, }.
• Smaller than expected number of components that are available "out of the box"; some people see this as "batteries not included". This is the other side of having a compact and portable core and is well compensated by having LuaRocks and libraries like Penlight.

Bad

• Limited error handling support (using pcall and xpcall), although some may argue that it is sufficient and just needs some syntactic sugar and more feature support (like deterministic finalizers). The combination of pcall and error is quite powerful, especially given that error can return anything (for example, a table) rather than just a string, but having catch ... finally constructs may be cleaner and simpler to read in many cases.
• Global scoping by default (this has been partially addressed in Lua 5.2, which has no globals). There is a strict module that requires all global variables to be initialized. I have not had many issues caused by uninitialized globals, but still put this one into the "bad" category as I once made a mistake of calling a variable "next" and not localizing it, which caused an issue with an iterator in a completely different module as it overwrote the next function used with iterators.
• No Unicode support (at the very least you don't get string.len and pattern matching functions to recognize Unicode characters); there is a binding to ICU library that implements full Unicode support. See also this message and follow-ups for a good summary of what is already supported and what modifications may be required for string.* functions.
• Limited pattern-matching support, although the included one is still quite powerful. After using Perl for over 15 years, I miss some of the regexp features (mostly look-aheads, optional groups (group )?, and groups inside groups), but nothing that warrants the additional complexity in the implementation. Those who need more power in their regexps can use LPeg and its re module.
• No ternary operator; several alternatives are available. I usually end up using foo = test and value1 or value2 form with the caveat that value2 can be assigned if both test and value1 end up being false.
• No POSIX functions built-in. There is the luaposix module, but it requires compilation, which is not always an option. I didn't miss this much, but I did come across a case where I needed to get/set an environment variable, and having access to getenv and setenv would be convenient [Updated 6/1/2012] As miko noted in the comments, there is os.getenv, but no corresponding os.setenv.
• No class/object finalizers. Lua provides finalizer functionality through the __gc metamethod, but it is available only for userdata types (and not tables) and doesn't match the functionality provided by other languages, for example, DESTROY and END methods in Perl. [Updated 05/27/2012] There is an undocumented newproxy feature in Lua 5.1 that allows implementation of finalizers on tables; Lua 5.2 removed that feature as it added support for __gc metamethod on tables.
• No yielding between Lua and C code: coroutine.yield call across Lua/C boundary fails with attempt to yield across metamethod/C-call boundary. I happened to come across this error several times as I was doing async programming with luasocket and coroutines, but solved it using the copas module. This has been addressed in Lua 5.2.

Ugly

• Number of elements in a table is not easy to get and the result depends on how you do this (or what you mean by "length"). This is probably not surprising, given how powerful tables are in Lua and the fact that they support flexible indexing (by numbers and any other Lua type except nil). Tables in Lua have two parts: an "array/vector" part (generated with t = {1, 2, 3}) and a "hash" part (generated with t = {a = "foo", ["b"] = 2}); the two can be flexibly combined. #table returns the length of the shortest "array/vector" part (without any gaps) and table.maxn(t) returns the longest "array/vector" part (this function is removed in Lua 5.2). The "hash" part doesn't have a defined length. Both parts can be iterated over using the pairs method, which allows you to count the number of elements in them. However, print(#{1, 2, nil, 3}) prints 4 and not 2 as one may expect, whereas print(#{1, 2, nil, 3, nil}) prints 2. I'm sure there is a good reasonable explanation for this, but for now it is in the "ugly" bucket.[Updated 11/17/2012] As FireFly noted in the comments, in Lua 5.2 the length operator is only defined for tables that don't have holes in them.
• return statement can't be used if it's not the last statement in a block; in other words, function foo() print(1); return; print(2) end will trigger an error 'end' expected... or unexpected symbol near <whatever statement you have after 'return'> (depending on whether you have semicolon after return or not). Not that anyone would want use this for anything other than debugging, but I got bitten by it couple of times. I would have put this in the "different" category, but I find it inconsistent that I can't use return, but can use do return end in exactly the same place.[Updated 5/19/2012] This also applies to break statement, although in Lua 5.2 break is no longer required to be the last statement in a block.
• Only one value is returned from a function if it's not the last one in a list; for example:

   function f123() return 1, 2, 3 end
   function f456() return 4, 5, 6 end
   print(f123(), f456()) -- prints 1, 4, 5, 6
   print(f456(), f123()) -- prints 4, 1, 2, 3

  The related behavior of return is also affected by this rule: return f456() returns three values, but return (f456()) returns only one value (note the extra pair of parentheses). This matches the overall simplicity of the language and is well documented, but I still find it to be a bit ugly (although ugliness as beauty is in the eye of the beholder).

Overall, I have so far enjoyed the simplicity and consistency of the language, although there are few things I wish were done a bit differently. My eight-year-old son also picked Lua syntax quickly, which reminded me a lot about my experience with Turbo Pascal decades ago.