Representation is the essence of programming...
The more precise meaning of Everything is a string is every value is a string. This is one of the central features of Tcl. In most other languages, statements and expressions are comprised of some mix of operators, functions, various types of data, and syntactic contructs for expressing structure or operation. Tcl is a more simple kind of language comprised of only commands and substitutions into the words that form those commands.
This page lays out the implications of the everything is a string design, as well as the details of working with this design when using the Tcl C API, where the string value observed at the script level is part of a structure that contains both the string representation and a typed interpretation of that string. If this typed interpretation isn't of the appropriate type for the current context, it is discarded and replaced with a new interpretation. A script can avoid this shimmering by consistently using a value such that its interpretation remains the same.
A string is a sequence of code points. Generally, each code point signifies a character, but there also exist code points that do not correspond to a character, and may serve some other purpose within a string. In Tcl version 8.6 and prior, strings were limited to code points in the Unicode basic multilingual plane. In future releases, any code point may occur in a string, including some code points that can not be encoded into any of the standard UTF encoding forms. The vast majority of the time, strings are comprised only of the Unicode scalar values, which is all the Unicode code points except the surrogate code points.
In a Tcl script, everything is a string, and Tcl assigns no meaning to any string, making it a typeless language: Since a string has no particular value, it also has no particular type. The rules for Tcl state that a script is composed of commands, and a command is composed of words, the first of which is the name which is used to locate the procedure used to execute the command, and the remainder of which are the arguments that are passed to that procedure.
A script is composed entirely of commands. Before evaluating a command, Tcl forms the words of the command by performing any needed substitutions: variable substitution, command substitution, and backslash substitution. Since variable values, command evaluation results, and interpreted backslash sequences can be substituted into words, it follows that the result of each substitution is a string. The phrases everything is a string, everything is a script, everything is a command, and everything is a word allude to this design, which is fundamental to Tcl.
Since no type syntax is necessary in Tcl, characters like double quote or braces may be put to other uses. Sometimes double quotes are used such that they appear to denote strings, but they actually don't. Braces can be used used in such a way that they appear to denote a block of code, but they actually don't. In Tcl, double quotes and braces merely change the context for determining which characters should be interpreted as substitutions or word delimiters.
A string has no intrinsic meaning. The only fundamental feature of a string is that it is unique among the set of possible strings. This feature is sufficient to differentiate procedure names. The strings that Tcl passes to a procedure are not in themselves identifiers, numbers, functions, variables, lists, dictionaries, blocks of code, or special values like null, although a procedure may choose to assign one of those meanings to a string that is passed to it. That is to say, it is the individual procedures in Tcl that assign meaning to a string, not Tcl itself.
A string is not an object, so it has no configurable properties. The identity of indiscernibles applies to strings: Each identical string has the same meaning. Any extension that looks at some internal data Tcl keeps about the value to determine the meaning of a string is ill-behaved. Any such internal data is merely an implementation detail. On the other hand, it is perfectly fair for each separate routine to interpret the same string representation in its own way, or even for one routine to interpret a single string in multiple ways, but it makes its decisions based on the context of its operation and upon the meaning it can derive from the string itself.
This design makes Tcl rather unique. In other languages each value has a particular type, for example, "8" may be a string, while 8 is a number. In Tcl there is just 8, which one procedure may use as a string, and another may use as a number. In other dynamic languages where each value is an object, and may therefore come with various behaviours, duck typing means that if the object "walks like a duck, and quacks like a duck, it's a duck. A simple string has no such behaviours attached to it, so duck typing in Tcl is even more simple: If it looks like a duck, then it's a duck, although it may also be a witch.
Because everything is a string, there are no operators, no keywords, and no code blocks. If there were then values would have different types: strings, operators, and keywords. This unification of all things into strings opens the door for some peculiar modes of operation. In other languages, the behaviour of an operator may be overloaded based on the types of its operands. In Tcl a procedure maybe be overloaded by moving it out of the way and replacing it with another in its place. Since everything is a string, including things that look like code blocks, and things like loops are implemented as procedures rather than keywords, even fundamental things like conditional execution and loops may be customized by replacing built-in procedures. Tcl's everything is a string design makes such radical language modification low-friction.
Tcl's typeless approach puts it in a somewhat exclusive category of programming languages. The other languages in this category are forth, BCPL, and assembly language. Almost all other languages practice a form of tight syntactic coupling of data and metadata where the pattern of characters that compose a value is juxtaposed with other patterns of characters that communicate some additional information about the value. In the most common case this additional information provides a type for the value. To Tcl, each pattern is just a message to pass along. Tcl doesn't read the message. It simply passes it to a routine. It is the routine that ascribes some meaning to the pattern and that returns another pattern that, by contract, also has a certain meaning. The programmer must understand the contract and rely on it to weave together a sequence of routines that make some sense and accomplish some objective.
A string may be endowed with a rich and structured meaning. Like a command, a list is represented as a string that is a sequence of words separated by whitespace. Each of the words in a list may itself be another list, in which case the string represents a recursive data structure. A procedure that understands the format and syntax of a list can perform operations on the list. For example, it might return one of the words in the list, or one of the words contained in that word. A word in a list might represent some other structrue as well, like a dictionary. This concept of using strings to represent more complex data strucures is one of the distinguishing concepts of Tcl. In Tcl, every operation on a data structure is also, strictly speaking, a string operation. This design, however, only represents the script interface of Tcl. Internally, and also in the C API, Tcl features mechanisms for using more performant in-memory data structures to carry out these operations. The more performant implementations must make sure, however, to maintain the string semantics that the script interface provides.
In assembly language, bits/bytes are paired with instructions, which manipulate those bits/bytes. The assembler doesn't concern itself with any notion of type for those bits/bytes. Instructions are issued, and each instruction knows how to treat the bits/bytes handed to it. Tcl operates in the same way, except that the fundamental transactional unit between instructions is the string rather than bits/bytes. This string veneer makes it easy to stretch Tcl over existing projects, providing a connecting layer between components. Each command is an entry point into a system of arbitrary complexity, and the string is the universal currency. Each new routine can extended Tcl into a new domain, and allow operation between that domain and the other that have a window in Tcl.
Here are some examples of the interpretations routines give to the values passed to them:
If one desires more types, numerous systems exist that provide them and that allow the user to define their own. See object orientation for a list of such systems. The stringineess of Tcl makes it as flexible as a language can be, and the implementation has focused on providing the primitives on top of which virtually any programming paradigm can be implemented.
lists and dictionaries are strings that conform to a particular format. Other values are used as handles for data structures or resources that are not directly accessible at the script level. The following resources, accessed by name, are examples of handles:
Typically it is sufficient for caller to understand how a routine interprets each argument, and then provide arguments that have suitable interpretations. When a routine needs structured information, it can specify that an argument is a dictionary, a list, or even a command. A commmand passed to a routine can be used by that routine. If the command represents an object, it can provide a set of subcommands that that implement some duck type.
In expr each operator provides a context, and expr tries to interpret each term as a type that best fits that context. Where an operator is flexible, expr prefers a numeric interpretation.
If a value can have multiple interpretations, a routine that uses it must implement some strategy for determining which interpretation to use.
tcl chatroom 2013-04-30:
DGP: I find it useful to think of "types" in Tcl as being subsets of the value universe. So it doesn't make sense to ask what type a value is. Instead, you can identify those types where a value is a member, and where the value is not a member.
CMcC: Right, subsets, not partitions
DGP: "Everything is a String" is just the trivial observation that all values are in the same value universe.
In the implementation of Tcl the structure that holds the string representation for a value may also hold one typed interpretation of the value. Modifying the string representation clears the typed interpretation, and vice versa. For example, a list can be modified either by changing its string representation or by using a command like lappend, which works directly with typed interpretation if it is of list type.
After either the string representation or the typed interpretation is cleared it is only generated again when needed. This means for example that a value that has a typed interpretation but no string representation can be passed between routines that use the same typed interpretation without incurring the expense of generating a string representation. Needless to say, modifying a list by modifying its typed interpretation is much faster than modifying the string representation.
The typed interpretation of a value is just an implementation detail. It is not exposed at the script level, and does not have any semantic impact on the language. The dual-representation format is used only at the implementation level, and only as an optimization. Two objects with the same string representation are the same value, whether or not the string representation has been generated. At the implementation level, there may well be two Tcl_Obj structures with the same string representation, but with different typed representions, and any function that accepts a Tcl_Obj as an argument must interpret them as having the same value. A user of Tcl's C API will gain an appreciation for the way Tcl values are handled at the C level, working with either the string representation or the typed interpretation as is expedient.
EIAS is one of the grand unifying concepts of Tcl. As Edsger Dijkstra noted in On the cruelty of teaching computer science , a program can be viewed as a formula that must be derived by the programmer, and the only known reliable way of doing that is by symbol manipulation. Hence, we construct mechanical symbol manipulators by means of human symbol manipulation. EIAS facilitates such a mathematical style of programming by merging the concepts of code and data more completely than even Lisp, as a Tcl script itself morphs to become its own result.
When everything is a string, every kind of data is readily accessible: When some new data type is introduced in a language like C or Java, it usually has to come with its own library for printing values, doing I/O, initialising variables, and often even for copying values. In Tcl all that is immediately available, since it can be done with strings and the new data type is represented using strings. This common ground eases the burden of the programmer.
Strings are general. The standard computing models are all readily expressible in terms of strings. The tape of a Turing machine contains a finite string of symbols. Lambda calculus is manipulation of Post production systems, which model computability by replacing parts of strings with other strings.
NEM 2010-12-15: One aspect of EIAS that is worth consideration is how it has kept Tcl "pure" in some sense. Part of EIAS that is little mentioned is that Tcl's strings are immutable. This means that Tcl's value space is purely functional, in the Haskell sense. All side-effects are confined to the shadow world of commands and variables and other second-class entities. What this means is that Tcl now possesses some very powerful purely functional data-structures that are somewhat better than those available in other languages. For instance, I cannot think of another popular language that supplies O(1) purely functional dictionaries and lists (arrays) out of the box (or even in the library). Not to mention efficient Unicode and binary strings.
DKF: See also tcl::unsupported::representation, which can peek behind this veil. If you use this, feel dirty!
AMG: In testing and debugging high-performance applications I use this to confirm that I'm avoiding shimmering.
Programmers more familiar with other language sometimes criticize Tcl's EIAS design, usually because they assume that complex algorithms requiring data structures are not possible in Tcl. What they might be missing is that although they can't directly translate some of their idioms into Tcl, equally powerful Tcl idioms exist and are waiting to be discovered. By sticking to EIAS, Tcl elegantly disposes of problematic "features" of other language, such as C features that make aliasing in a possiblity. The data structures that others may think Tcl is missing are simply expressed in another way, but that is difficult to see at the outset.
However LV would like to point out that the true philosophy of Tcl says Do all that you can in Tcl - but then, do the rest in C/assembly/whatever and create glue and handles to it for Tcl.
Donald Porter remarked in the Tcl chatroom: More precisely, every value has a string representation. Tcl arrays are not values; they are special types of variables.
lvirden: I guess there are other things that fit into the same category as arrays - created items like procs, and in tk all sorts of widgets, etc.
aku: But most have a way to serialize them into a value, and back (array set|get, proc|info body|arg|default)
kennykb: And the ones that don't have natural serialization generally are managing external resources (channel handles are the most obvious example)
Shin The Gin: If everything was a string, then one could easily save the whole runtime environment to a file and restore it later.
RS: likes the ditty "I'm not afraid of anything, if everything is a string". In fact, the Tcl mantra often relieves fears of complexity: anything that can be brought to the prototype "string in, string out", can be nicely done in Tcl. Arabic, A little Korean editor? Of course, everything is a Unicode string! Geographic mapping? Just give me a string with the latitudes, longitudes, and whatever other data, and presto - Tclworld. Images can in many ways be rendered as strings (XBM, PNM...); one pretty intuitive way is in strimj - string image routines.
Todd Coram: Data typing is an illusion. Everything is a sequence of bytes. Call 'em ints, floats, symbols, strings, whatever. Tcl exposes both code and data to the user as sequences of bytes (called strings). This is Tcl's choice of abstraction. And its quite a powerful choice IMHO.
BR: Hm, isn't it actually like that a string is a sequence of characters, and bytes (in Tcl) are just characters with the values 0 - 255? I think that's the model of binary data in Tcl. IOW bytes are not fundamental in Tcl, but characters and strings are.
Except that characters could be Unicode instead of ASCII.
2003-05-13: Recently, Bruce Eckel in Strong Typing vs. Strong Testing , and Robert C. Martin in Are Dynamic Languages Going to Replace Static Languages? talk about weak typing and dynamic languages.
CL thinks these two make mistakes, but hasn't time now to explain more. In any case, yes, these are good noteworthy references.
escargo 2003-05-13: Another way that everything is a string can be an issue is where a string representation can only be an approximation of what is being represented. The main instances of this that come to mind are floating point numbers (for which there are already some existing wiki pages). There may be other examples as well.
What? There is no reason a string can't fully represent a floating point number. And Kevin Kenny has a TIP in the works to ensure that Tcl does always indeed achieve exactness in this case - Roy Terry, But really, it seems a waste of time to make fine points about "everything is a string" which is merely a programmer's cliche and doesn't begin to express the power of Tcl.
escargo 2003-05-14: Sorry; there was a slip of the finger there. I said "floating point" and what I meant was "real".
Lars H 2003-05-15: Real numbers are beyond what is computable. The number of possible outputs from a Turing machine (and thus the set of real numbers which one can specify in any way whatsoever) is merely countable, whereas the set of all real numbers is uncountable. But this view does provide an answer to why "Everything is a string" is such a powerful idea. Many languages (most notably C) take the approach that "Everything is a number (with native machine representation) or some fixed aggregation of such numbers", but all such representations are limited. In order to support general strings, it is necessary to venture into some scheme of dynamic memory allocation and pointers to allocated objects. The string, on the other hand, achieves the maximal generality of a Turing machine (the tape always has an obvious representation as a string) and thus if something wouldn't be representable as a string, it wouldn't be computable either.
escargo 2003-05-16: What would it take to make Tk widgets serializable? I was thinking about xml2gui and wondering what it would take to make a widget produce an XML description of itself. Further, what would it take to have widgets that contain other widgets produce XML of themselves? This would seem to me to be one useful goal.
Another goal would be the converse, what XML would need to be used to create all the Tk widgets (and pack them the right way, etc.)? (This would be a suitable storage format for GUI Building Tools.)
jcw 2003-05-17: There already is a serialized form of Tk, able to cope with any complexity of widget hierarchies: the Tcl script that creates them.
jmn: Yes, but is there a canonical form for it?
escargo: I am reminded of one-way hashes. You can have a function that given an input can produce a hash value that cannot be used to derive the original input. Just because I have a widget does not make it clear to me that I can derive in an algorithmic way Tk and Tcl code to recreate the widget. Perhaps this is something for the Tk 9.0 WishList, but I would certainly like to see whatever changes would be necessary to allow this (if it's practical at all).
jcw 2003-05-17: While EIAS is indeed a wonderfully powerful and flexible abstraction, I'd like to point out that LISP'ers and Scheme'rs have a very similar set of self-contained mechanisms at their disposal, based on "everything is made up of cons cells" (it's more of a mouthful, though...). IMO, "strings" as convention to represent data in a certain way is not inherently different from other representation choices - one could even use neurons and synapses if that were practical. What EIAS does imply is "code is data" and "data can be used as code", which is why one can play so many tricks in Tcl (and in LISP).
NEM 2005-07-25: replying to this a couple of years too late... The difference with Lisp is that cons cells aren't universal; as I understand it, some basic data types like numbers are not represented as cons cells. You could build up everything from cons cells, in a similar way to building everything from set theory, but Lisp doesn't, and so you can't treat an integer as a list. In Tcl, though, the string is the universal medium of representation, so I can treat an integer as a list (of one element).
FW: Come to think of it, what are some other typeless languages in the "everything is a string" sense - RS has already submitted and documented thoroughly the antique TRAC in Playing TRAC.
JE: MUMPS , or M, is another EIAS language. Forth and BCPL are also typeless, but there the fundamental type is a "cell" or native machine word instead of strings. (BCPL seems to be extinct, but Forth and MUMPS are still around.
If "everything is a string," then how can you tell what's an object?
escargo 2005-07-23: That's what I woke up to this morning. I was thinking that Tcl lacks what I have seen called a "meta-object protocol," something that allows some object-oriented languages (like Smalltalk) to do some useful operations on objects and classes. I like Snit because of what it allows me to do to compose objects using delegation. However, if I'm operating in Tcl (or in Tk) and I have an identifier, how can I tell if its value represents an object from an object system like Snit (or any of the other object systems added onto Tcl). And if it is an object, how can I tell which object system it is an object in, so that I can guess what behavior is has (which functions it understands or implements)?
The only way I can see something like this working is if there were some agreed-upon standard for names (or references) such that a classifier (say [string is object ...]) could return a yes or no answer.
Even better would be one that could tell which object system implemented the object (say [string is objectsystem ...]).
This might be possible in a system like Jim if the Jim References encoded the object system and whether something was an object.
Even without add-on object systems, it would be nice to be able to determine if there could be [string is command ...], but that in some respects defeats the purpose of unknown. (I'm still fuzzy from sleep, so maybe there is something that does this already, otherwise how would unknown get called?)
Lars H 2005-07-24: I think the best way of pointing out how your analysis here is wrong is to point out that
indeed, it follows from everything is a string that there cannot be such a command (at least not one that doesn't just return "string" or whatever regardless of input), since nothing but the string may define the value. If you want to type-tag values, then you must include that tag in the value itself.
Put another way: Values in Tcl are what you make of them. Values don't "know"[*] that they're command names, integers, or variable names--they become whatever you decide to treat them as (or cause an error to be thrown when they cannot be interpreted the way you claimed).
A consequence of this is that one shouldn't write programs that just throws all data in a huge bowl and lets unknown (or whatever) sort it out later, one should write programs so that there at every point in the program is clear what type of data is going to be passed around. It is sometimes useful to let the type of some argument be "either an A or a B", but then one must also have sorted out whether it may happen, and if so what it means, when data comes along that is both A and B.
So what has this to do with objects, then? Everything, since whatever one uses to identify an object is just another string (even though few eyebrows are raised these days when people request magical behaviour from objects--for some reason it seems politically correct to regard objects as nobler than ordinary data). You can ask an object system whether it recognises a particular string as identifying one its objects (but this assumes the system is implemented in such a way that this is possible), and you could start an object system registry that goes around asking all known object systems whether they recognise a particular object as theirs, but that's about it, and I doubt it would be of much use outside debugging.
In a sense, the proper response to "how can you tell what's an object?" should be:
[*] Techincally, on the C level, most Tcl values kind of know from their Tcl_Obj internal representation whether they are command names, integers, variable names, etc., but it is more accurate to describe this information as if I'm a command name/integer/whatever, then I'm the name of that command/integer/whatever since this type information is shimmered away whenever the Tcl_Obj is used in a different sense.
escargo 2005-07-25: This is closer to the problem that I felt I was dealing with. In a unified object system (or alternatively, where only one object system is possible), you don't have to speculate about what kind of behavior an arbitrary object might exhibit.
In Tcl (especially if you are using Snit to delegate to some arbitrary objects), you don't know (and as you pointed out, perhaps cannot know) what object behavior a particular object (for which you have a string to use as a reference) might exhibit.
If I have a string, I can use winfo exists to see if it's a Tk window.
I can use info procs to see if it is a proc.
If it were an object, then I expect that it has some behavior (otherwise what's the point of it being an object). But without knowing more about it, it's not safe to try different probes into its behavior to see what it can do. The irony is that at least some object systems for Tcl provide some kind of introspection, but I doubt that they provide it the same way, so you can't just use it to find out more about the object.
Why does this matter? - The reason I feel that it matters is that there has to be somewhere where knowledge of the type of object has to be carried around so that you can write your programs correctly. (The type in this sense being the add-on object system that implements the object.) If you can't determine the type of object from the object itself, then you have to code that information into comments or else invent some other means of doing it.
It's not that this can't be done, but it's a wish I have that the answer were within the language itself, either by implementation (e.g., you could deconstruct a reference to determine the object system) or convention (all object systems implemented an info command that all objects responded to that could, as one of the items that might be returned, respond with the name, and maybe revision level) of the implementing object system.
I realize that's not going to happen, but if enough people agreed with the need, then progress could be made in that direction.
NEM 2005-07-25: This all boils down to fundamental philosophical beliefs about the nature of values and types. What really marks Tcl out from most other languages, and what is at the heart of this debate is not that strings are such wonderful things that they should be used for everything, but rather a recognition that the notion of a "type" of a value is extrinsic to the value itself. In other words, a type is an indication of some interpretation of a value. Any representation of a value can have multiple different interpretations, and so to talk of the type of a value without reference to the particular system doing the interpretation is difficult. Conversely, any abstract type can have multiple possible representations (the key idea of abstraction/encapsulation). So, the connection between values and types is a many-to-many connection. Most languages assume a 1-to-many connection, so each value has a single type which is associated with it by the language, with less categorisation left up to individual commands/functions (although it is not true to say that no choice is left; every function performs an interpretation of its arguments to some degree). Tcl, however, is different in that it performs almost no interpretation of the values it is passed. It does basic tokenization and grouping, but leaves values as they are found: as strings. The only further bit of interpretation that Tcl does is to treat the first word of each line (talking loosely) as the name of a command. (Well, there is also variable substitution and other items in Tcl.n, but we'll ignore those for now). It is then the individual commands which take care of any further interpretation. You can think of this as a form of extreme lazy evaluation: even parsing is left to the last possible moment.
So, what are the trade-offs? On the negative side, the fact that Tcl does less interpretation for you means that it makes fewer guarantees (e.g., it's hard to do garbage collection of references if you can't guarantee that X is a reference and Y isn't). Another difficulty, is that it is possible to break abstractions in Tcl: you can always drop down to the level of strings and manipulate the representation of a value, rather than use any higher-level interface. I actually think this is one of Tcl's strengths, but it is a longer argument. You can also get around this by using opaque handles, which hide the representation behind a layer of indirection, that may or may not be introspectable. On the plus side, the fact that Tcl has an ultimate fear of commitment, means that commands have more free reign in deciding how they will interpret the values. This, I suggest, is the heart of what makes Tcl a good glue language: by not committing to a single interpretation of a value it allows multiple components to make their own, possibly conflicting, interpretations. (As an aside, an interesting parallel can be drawn here with Daniel Dennett's Multiple Drafts theory of cognition/consciousness). Another way to look at this is to say that by providing a common representation medium you reduce the number of explicit conversions that have to be done. If you have N distinct types, then in order to convert between them you potentially need N!/(N-2)! different conversion functions (i.e. N 2-way permutations, e.g. int2double, double2int, int2string, string2int, etc). If you have a common representational medium, then you can use that as an intermediate, thus reducing the number of conversion functions needed to just 2(N-1), and just two functions are needed for each type: toString and fromString (the string type itself obviously doesn't need these).
Can we combine the benefits of both approaches? I think we can. TOOT was about doing just this, and Interpreting TOOT has my earlier thoughts on the subject. I've been thinking about this some more since, and will hopefully soon have time to write some more code and an essay detailing my further thoughts. For now, I will point at Monadic TOOT, which contains some clues to a possible way forward. Those who know about monads will know that they are useful for confining effects and enforcing abstraction boundaries. I think we can use the same techniques in Tcl to create packets in which abstractions can be enforced and guarantees can be made, if needed. The other side of this process is partial evaluation, a TOOT bundle of
type: value
can be partially evaluated (or partially applied), to yield a new function specialised for that interpretation of that value. This can be optimised and can enforce a type abstraction.
Lars H: Well put. The part about late "commitment" puts a name on something I think is very important in understanding the strengths of Tcl. I'll see if I can find a good place to put this idea for easy access.
DKF: Actually, in 8.6 there is tcl::unsupported::representation, which includes type cache information in its result. Don't use it for anything other debugging. Or if you do, feel very naughty. It is very bad style to write code that depends on types (albeit inevitable for solving certain types of problem in the support of Java and JSON correctly, alas).
SYStems 2005-07-23: Those are not very complete thoughts, but. I think to really answer and understand the idiom everything is a string, we need to identify the context, or perspective.
A Tcl script is a series, a sequence of statements, each statement receive input
perform action on this input and then
Every statement input and/or output is a string, only side effects (and maybe input events) can be NOT A STRING, but all input and output must have a string representation.
Each input and ouput, can have a different in-memory representation, or on disk representation. But inside a script it must have a string, or I prefer to say, textual representation.
Since everything written in a Tcl script, is textual. A script can be the input of a Tcl command, for example control structure commands.
Every input, must have an inline textual representation, this is why an command must have a textual ouput, so that when its substituted it produces a string, the only thing that is good as an inline input.
All input and output must have inline textual representation (this is the part I an hesistant about, I am not really sure this is correct, I am using the word inline loosely here I propably mean infile!!)
[[set] is the command used to manage a Tcl script memory, all [[set] variables must have a string value.
This may sound weird, but I write this hoping that Tcl doesn't lose its primary principles. For example, I see many people talking about Tcl variables, Tcl doesn't have variables. [[set] is a command that has a Tcl interface, gives a Tcl script the notion of variables by associating a name with a string value.
Depending on the value-string-representation, set will store it differently in-memory. set not Tcl.
[[set] for example, doesn't store the variables on disk. A good Tcl'er might create a command that give a Tcl script the notion of persistant data, data stored on disk!
Tcl helps guide thinking by recognizing the syntax $name, and treating that as [set name]
Anyway, back to the fact that set can only associate a name with a string. set is used to store another tcl command's output, and pass it later as an input.
So we can say that everything inline- a tcl script, anything that can be passed around, a tcl script memory, a tcl script internal environment, must be a string. Or in other words, we can say, that Tcl introduces a new in-tcl context, where everything must have a textual representation.
Anything outside a Tcl script, outside the in-tcl context, for example, a command side effect, or an external environment, can be not a string.