23 replies to this topic

[BS-er]

In Ada and probably numerous other languages, string representations of each enum value name are available in an elegant manner, and the language has built-in functions for converting between the actual value and the string representation, and vice versa. The language implementation is able to do this in an optimized manner as well. I suspect that it can't be as easy as that in C++, given its lower-level nature. If that's the case it's unfortunate, since it seems to be a recurring pattern for me to need to convert between enum values and string representations of those values. I was wondering if there is anything out there maybe in the form of a template that makes it easier for us to convert enum values to string representations and vice versa. Hopefully something compact as well. I thought I spotted something similar in some proposed Boost features, but there wasn't enough information on it, and I had hoped to avoid the headaches and confusion of installing a large template library for a few features. Thanks for any advice.

[Aardvajk]

You are correct that you can't do this in a built in way with C++.

When I needed something like this a couple of years ago, I wrote a little program that took a text file as input, for example say we had fruits.txt:

apples
oranges
pears

and generated a .h file like:

#ifndef fruits_H
#define fruits_H

enum fruits { apples,oranges,pears };
const char *fruits_str[]={ "apples","oranges","pears" };

#endif

Obviously this was a pretty trivial program to implement.

[DrEvil]

I usually keep synchronized my enum and string arrays manually, with a static assert to help me catch at compile time to ensure they are the same size, and help to prevent forgetting to update one after the other.

[Simian Man]

Another approach is to overload operator<<

enum fruits
{
   apples,
   oranges,
   pears
};

std::ostream& operator<<( std::ostream& os, const fruits& fruit )
{
   switch( fruit )
   {
      case apples: os << "apples"; break;
      case oranges: os << "oranges"; break;
      case pears: os << "pears"; break;
   }
}

I suppose you could write a program to generate this as well.

[Muhammad Haggag]

This is rather clever, even though a bit ugly.

Alternatives exist at the IDE-level. Using vim, Emacs, or Visual Studio you can use a script/macro that generates a string table from an enumeration automatically. Vim already has such a script here. It should be trivial to write one for Visual Studio using C#/VB.NET.

[scgames]

Although I'm sure there are more sophisticated approaches one could take, I generally use a std::map, e.g.:

std::map<Uint16,std::string> formats = boost::assign::map_list_of
    (AUDIO_U8,     "AUDIO_U8"    )
    (AUDIO_S8,     "AUDIO_S8"    )
    (AUDIO_U16LSB, "AUDIO_U16LSB")
    (AUDIO_S16LSB, "AUDIO_S16LSB")
    (AUDIO_U16MSB, "AUDIO_U16MSB")
    (AUDIO_S16MSB, "AUDIO_S16MSB")
    (AUDIO_U16,    "AUDIO_U16"   )
    (AUDIO_S16,    "AUDIO_S16"   )
    (AUDIO_U16SYS, "AUDIO_U16SYS")
    (AUDIO_S16SYS, "AUDIO_S16SYS");

Advantages of this approach include:

1. Easier to maintain than a switch statement
2. More flexible than overloading operator<<()
3. Enumerants can have arbitrary values (as opposed to the array method)
4. 'Invalid' enumerants won't lead to UB (as opposed to the array method)

[Telastyn]

I've used EasilyConfused's pattern in the past (though often with a "MAX_FRUIT" enum at the end which helps in looping over the group and similar constraint needs).

jyk's suggestion though is clearly superior in the face of enums which aren't simply 0->MAX.

[MaulingMonkey]

Quote:

Original post by EasilyConfused You are correct that you can't do this in a built in way with C++. When I needed something like this a couple of years ago, I wrote a little program that took a text file as input, for example say we had fruits.txt: apples oranges pears and generated a .h file like: #ifndef fruits_H #define fruits_H enum fruits { apples,oranges,pears }; const char *fruits_str[]={ "apples","oranges","pears" }; #endif Obviously this was a pretty trivial program to implement.

Here's a way to avoid putting more tools into your build chain (but it uses boost, and worse than templates --- it uses macros):

#include <boost/preprocessor.hpp>

#define SEQ (apples)(oranges)(pears)
#define TO_STR(unused,data,elem) BOOST_PP_STRINGIZE(elem) ,

enum fruits { BOOST_PP_SEQ_ENUM(SEQ) };
const char * fruit_strings[]={ BOOST_PP_SEQ_FOR_EACH(TO_STR,~,SEQ) };

#undef SEQ
#undef CAT

EDIT: Enterprisey-fied:

#include <boost/preprocessor.hpp>

#define PROJECT_PREFIX_DO_EVIL_TO_STR(unused,data,elem) BOOST_PP_STRINGIZE(elem) ,
#define PROJECT_PREFIX_DO_EVIL(enum_,strings,elements)     \ 
enum enum_ { BOOST_PP_SEQ(elements) };                     \ 
const char * strings[]={ BOOST_PP_SEQ_FOR_EACH(PROJECT_PREFIX_DO_EVIL_TO_STR,~,SEQ) };

//-----------------------

PROJECT_PREFIX_DO_EVIL(fruit,fruit_strs,(apple)(pear)(pineapple))
PROJECT_PREFIX_DO_EVIL(animal,animal_strs,(dog)(cat)(panda)(monkey))

(edit/note: trailing whitespaces to preserve \s inserted)

[kalmiya]

Quote:

Original post by Simian Man Another approach is to overload operator<< enum fruits { apples, oranges, pears }; std::ostream& operator<<( std::ostream& os, const fruits& fruit ) { switch( fruit ) { case apples: os << "apples"; break; case oranges: os << "oranges"; break; case pears: os << "pears"; break; } } I suppose you could write a program to generate this as well.

Try that and include this header from several different cpp's... Now
take a hex-editor and search the generated exe for "apples" - the
entire lookup table will be there as often as you included it.

Regards

[JohnBolton]

Quote:

Original post by Muhammad Haggag This is rather clever, even though a bit ugly.

Here is a simplified version:

    #define DAYS_OF_THE_WEEK         \ 
        ENUM_OR_STRING( Sunday ),    \ 
        ENUM_OR_STRING( Monday ),    \ 
        ENUM_OR_STRING( Tuesday ),   \ 
        ENUM_OR_STRING( Wednesday ), \ 
        ENUM_OR_STRING( Thursday ),  \ 
        ENUM_OR_STRING( Friday ),    \ 
        ENUM_OR_STRING( Saturday )
    
    // Enum
    
    #undef ENUM_OR_STRING
    #define ENUM_OR_STRING( x ) x
    
    enum DaysOfTheWeek
    {
        DAYS_OF_THE_WEEK
    };
    
    // Strings
    
    #undef ENUM_OR_STRING
    #define ENUM_OR_STRING( x ) #x
    
    char * DaysOfTheWeekStrings[] =
    {
        DAYS_OF_THE_WEEK
    }; 

You could also put the elements in an include file instead of a macro:

---- DaysOfTheWeek.h
    
    ENUM_OR_STRING( Sunday ),
    ENUM_OR_STRING( Monday ),
    ENUM_OR_STRING( Tuesday ),
    ENUM_OR_STRING( Wednesday ),
    ENUM_OR_STRING( Thursday ),
    ENUM_OR_STRING( Friday ),
    ENUM_OR_STRING( Saturday )
    
---- source file
       
    // Enum
    
    #undef ENUM_OR_STRING
    #define ENUM_OR_STRING( x ) x
    
    enum DaysOfTheWeek
    {
        #include "DaysOfTheWeek.h"
    };
    
    // Strings
    
    #undef ENUM_OR_STRING
    #define ENUM_OR_STRING( x ) #x
    
    char * DaysOfTheWeekStrings[] =
    {
        #include "DaysOfTheWeek.h"
    }; 

[Aardvajk]

Quote:

Original post by Kitt3n Try that and include this header from several different cpp's... Now take a hex-editor and search the generated exe for "apples" - the entire lookup table will be there as often as you included it.

Thinking about it, that would be equally true of the example I provided. I must have used that approach back in the early days when I used to just preprocess my entire program into one translation unit. [smile]

I guess to use my autogenerated sourcefiles approach properly, the program would have to generate a .h and .cpp file. Given the trouble and complexity this starts creating, I'd probably go more with something like jyk or JohnBolton or others have suggested.

[scgames]

Although I stand by my suggestion of using a std::map for associating strings with enumerants, in all fairness I have to acknowledge that there was a mistake in my posted example:

std::map<Uint16,std::string> formats = boost::assign::map_list_of
    (AUDIO_U8,     "AUDIO_U8"    )
    (AUDIO_S8,     "AUDIO_S8"    )
    (AUDIO_U16LSB, "AUDIO_U16LSB")
    (AUDIO_S16LSB, "AUDIO_S16LSB")
    (AUDIO_U16MSB, "AUDIO_U16MSB")
    (AUDIO_S16MSB, "AUDIO_S16MSB")
    (AUDIO_U16,    "AUDIO_U16"   )
    (AUDIO_S16,    "AUDIO_S16"   )
    (AUDIO_U16SYS, "AUDIO_U16SYS")
    (AUDIO_S16SYS, "AUDIO_S16SYS");

Oops! Some of these enumerants are aliases for each other. The *SYS variants are aliases for the *SB variants as determined by the endianness of the system, while *16 aliases to *16LSB.

Furthermore, I didn't notice the error until I compared this version of the code with a version that inserted the map elements manually. Because map_list_of inserts the elements in a different order than the corresponding 'manual' code, the end effect of the aliasing was different in each case. Subtle :-|

Now, unless I'm missing something else obvious, this doesn't negate any of the advantages of this method (over arrays or switch statements) mentioned earlier. However, although using a map means you don't have to worry about the particular values of the enumerants, you do have to know if any of them alias each other. This is particularly important when dealing with a third-party API rather than your own code (I had to check the corresponding SDL header to know that *16 aliases *16LSB, which isn't necessarily intuitively obvious).

[Zahlman]

Quote:

Original post by jyk However, although using a map means you don't have to worry about the particular values of the enumerants, you do have to know if any of them alias each other.

Which is why I would generate code rather than using tools like boost::assign. :)

[scgames]

Quote:

Original post by Zahlman Which is why I would generate code rather than using tools like boost::assign. :)

Sure, although I'd be interested to know how you'd apply this solution in the particular case that I presented.

Unless I've missed something, in both of the 'generated code' examples presented earlier in the thread the enums themselves are generated, not just the associated strings. Therefore the values are known, and furthermore the values are known to be sequential and zero-based. The corresponding strings are then stored in an array.

In the case I presented the enumerants come from a third-party library. Although one can of course examine the appropriate header file, let's say for the sake of argument that the values are not known. How would one automate the generation of associated strings in this case?

Replacing map with multimap in my previous example solves the problem and makes the solution robust in the presence of arbitrary values for the enumerants, including those that are aliases of each other. However, this comes at the cost of added complexity elsewhere in the code.

So although I'm sure you're right, I'm not quite clear on how your suggestion to use code generation rather than 'naive' application of a map specifically addresses the problem I presented earlier. I'd certainly be interested in seeing a concrete example - you always seem to have interesting tricks up your sleeve :-)

[ToohrVyk]

A generation tool can read the enumeration description from a file and output the correct display function. For instance:

// some defines in an "enum-helper" file
#define GEN_OUTPUT(X)
#define GEN_ALT_OUTPUT(X)
#define GEN_IGNORE

// ======================================================
// An enumeration definition, which is tagged 
// to generate an output function
GEN_OUTPUT(generated.foo.hpp) 
enum foo {

  // use the default name and value
  apples,

  // set the numeric value ourselves
  bananas = 35,

  // set the displayed text ourselves
  GEN_ALT_OUTPUT(green lemon) lime = 34,

  // we don't want this value to be used
  GEN_IGNORE ignored = 35
};

#include "generated.foo.hpp"

// ======================================================
// An example of generated file:
const char _foo_apples[] = "apples";
const char _foo_bananas[] = "bananas";
const char _foo_lime[] = "green lemon";

STATIC_ASSERT(apples != bananas);
STATIC_ASSERT(bananas != lime);
STATIC_ASSERT(apples != lime);

std::ostream& operator<<(std::ostream& out, const foo& f)
{
  switch(f)
  {
    case apples: return out << _foo_apples;
    case bananas: return out << _foo_bananas;
    case lime: return out << _foo_lime;
    default: assert(false);
  }
}

Optimizations would include transforming the switch statement into an if-else tree based on the likelihood of each enumeration value.

If enumeration values overlap, the generated operator would fail to compile. It's also possible to generate a set of static assertions, as illustrated above, which would cause failure in a cleaner way.

[scgames]

Quote:

Original post by ToohrVyk A generation tool can read the enumeration description from a file and output the correct display function. For instance: *** Source Snippet Removed *** Optimizations would include transforming the switch statement into an if-else tree based on the likelihood of each enumeration value. If enumeration values overlap, the generated operator would fail to compile. It's also possible to generate a set of static assertions, as illustrated above, which would cause failure in a cleaner way.

Let me describe more clearly the particular case from which my example was drawn.

SDL uses a number of macros to identify various audio sample formats. Some are aliases for others, and some are switched based on the endianness of the platform. I'll 'paraphase' the relevant portion of the header file here:

#define AUDIO_U8        < unique integer value > 
#define AUDIO_S8        < unique integer value > 
#define AUDIO_U16LSB    < unique integer value > 
#define AUDIO_S16LSB    < unique integer value > 
#define AUDIO_U16MSB    < unique integer value > 
#define AUDIO_S16MSB    < unique integer value > 
#define AUDIO_U16	AUDIO_U16LSB
#define AUDIO_S16	AUDIO_S16LSB
    #if ENDIANNESS == LITTLE_ENDIAN
#define AUDIO_U16SYS	AUDIO_U16LSB
#define AUDIO_S16SYS	AUDIO_S16LSB
    #else
#define AUDIO_U16SYS	AUDIO_U16MSB
#define AUDIO_S16SYS	AUDIO_S16MSB
    #endif
#define DEFAULT_FORMAT  AUDIO_S16SYS

The purpose of the code I posted is relatively simple: to associate with these values human-readable strings for output to a logging system. The log records, among other things, whether the requested specs match the queried specs, and what the requested format translated to on the system in question (e.g. DEFAULT_FORMAT becomes AUDIO_S16MSB on a PowerPC Mac).

The question then is how best to automate the generation of string representations for these values, and whether it's worth the trouble. It seems to me that none of the examples presented thus far, including the example you posted above, are directly applicable in this case without significant modification (the fact that the enumerants are macros is incidental - the same would be true were they elements of an enum).

I'm happy to concede the point based solely on your and Zahlman's considerable expertise, but I would still be interested in seeing how the proposed solutions could be applied here without undue difficulty.

[Zahlman]

(EDIT: Fixed stupidly wide code block; reformatted the ending comments; added a comment about serendipitous alias resolution. Also fixed some variable names - obviously I can't call an instance of an enumeration 'enum' [razz])

OK, let's say we have a header file with an enum (I'd rather not touch the problem of converting stupid #define usage into enumeration usage [wink] ):

#ifndef AUDIO_H
#include AUDIO_H

enum AUDIO {
  AUDIO_U8 = < unique integer value >,
  AUDIO_S8 = < unique integer value >,
  AUDIO_U16LSB = < unique integer value >,
  AUDIO_S16LSB = < unique integer value >,
  AUDIO_U16MSB = < unique integer value >, 
  AUDIO_S16MSB = < unique integer value >,
  AUDIO_U16 = AUDIO_U16LSB,
  AUDIO_S16 = AUDIO_S16LSB,
#if ENDIANNESS == LITTLE_ENDIAN
  AUDIO_U16SYS = AUDIO_U16LSB,
  AUDIO_S16SYS = AUDIO_S16LSB,
#else
  AUDIO_U16SYS = AUDIO_U16MSB,
  AUDIO_S16SYS = AUDIO_S16MSB,
#endif
  DEFAULT_FORMAT = AUDIO_S16SYS
}
#endif

Now, the aliasing problem is one that *can't* be resolved perfectly, for the simple reason that information is lost - a value of type AUDIO with value AUDIO_U16 is identical to a value of type AUDIO with value AUDIO_U16LSB, so there is no way to know which symbol was used in the source code. (After all, we can also create a variable of type AUDIO by reading in an int from a file and doing an explicit cast). However, let's say arbitrarily that we will resolve these problems by always stringizing an enum value according to the *first* enumerant in the enum with the appropriate value. Thus, in effect, the way to deal with aliased values is simply to *ignore* them ;)

(That is, a multimap doesn't help: there is no way to determine which value to select. So pragmatically, we have to just pick one, which takes us back to using a plain map. Incidentally, the scheme I propose automatically causes 'DEFAULT_FORMAT to become "AUDIO_S16MSB" on a PowerPC Mac', since that's the first enumerant with that value, so that's what will be used for the stringization. I suspect this simple heuristic will be best, really.)

We then write our script as follows:

- Invoke the preprocessor on the header file, i.e. ask the compiler what the
  system endianness is ;)
- From the preprocessor's output, parse out the enum declaration.
- Initialize an empty associative array from integer values to strings.
- For each enumerant:
  - Determine the int value.
  - If it is not found in the associative array, add it, associating it with 
    the stringized version of the enumerant.
- Output code which initializes a std::map<int, const char*> with the
  contents of our associative array, by iterating over our AA's keys and
  generating corresponding map.insert() statements. (Better yet, write code
  which wraps the whole thing up in a class. We can use a single class and
  create a global static instance of it for each enum, and do the initialization
  by clever use of operator chaining.)

Since implicit conversion does happen from the enumeration *to* an int, we can use enumeration values to look up the name in the std::map just fine.

The class might look something like this - all completely off the top of my head at 4:30 AM, but damned if it doesn't look good to me right now ;)

template <typename E>
class Enumeration {
  // Map values must always be string literals! We will never do any memory
  // management here; we freely copy pointers, and at the end, the static
  // section of the executable data is cleaned up in one go, and there are no
  // leaks or double-deletes.
  typedef const char* symbol;
  typedef std::map<int, symbol> table_t;
  table_t table;
  static symbol nil;
  static Enumeration instance;
  int nextkey;

  public:
  // We will use operator chaining in order to initialize a single static
  // instance that's accessible to preprocessor magics. :) No need to make this
  // a Singleton; this class only gets instantiated by our auto-generated code.
  // Callers don't need to know it exists to use it ;)
  Enumeration(symbol name, int value = 0) : nextkey(0) { (*this)(name, value); }

  Enumeration& operator()(symbol name, int value = nextkey) {
    assert(value >= nextkey);
    nextkey = value + 1;
    table[value] = name;
    return *this;
  }

  symbol operator()(E value) {
    table_t::iterator it = table.find(value);
    return (it == table.end()) ? nil : it->second;
  }
};

template <typename E>
Enumeration<E>::symbol Enumeration<E>::nil = "";
// If we don't trust the compiler to share that string constant, we could force
// that by making a separate "" constant instead and having nil alias it...

// Now, the templating so far just looks like it gives us type-safety for the
// operator()(E). But actually it will let us implement some obscene syntactic
// sugar as well, if I'm thinking clearly ;)
template <typename E>
ostream& operator<<(ostream& os, const E& e) {
  return os << Enumeration<E>::instance(e);
}
// And thus we accomplish the claimed goal of users not needing to know about
// the Enumeration class. For non-enumeration types, this should not cause any
// interference, due to SFINAE - again, if I'm thinking clearly... if not, I'm
// sure there's a way we can make it work... ;)

And we just write that code once; our auto-generated code just has to initialize Enumeration<E>::instance for each typename E that is appropriate (i.e., each enumeration in the program). We just emit something like:

Enumeration<AUDIO> 
Enumeration<AUDIO>::instance("AUDIO_U8")("AUDIO_S8")("AUDIO_U16LSB")
("AUDIO_S16LSB")("AUDIO_U16MSB")("AUDIO_S16MSB");

And that should work even at top level, because we're just initializing the variable; no procedural code here, nope, no sir ;) Even if that doesn't work, though, I'm fairly sure that "Enumeration<AUDIO> Enumeration<AUDIO>::instance = Enumeration<AUDIO>(etc....)" will.

[SunDog]

Err, well what I wanted to say that the above was generally good programming practice. One of the problems with C++ enums is that they are integers, so there is no type information when passing paramters. If you get paramter order fouled up or pass an illegal value, this error will be detected only at run time.

You could also avoid passing the ordinal in the constructor with a static ordinal field that gets incremented every time the constuctor is called. This way you won't have to pass an ordinal value to the constuctor. Also, you could implement next() and prev() functions or even use the STL's iterator mechanism.

[gdunbar]

Quote:

Original post by JohnBolton Quote: Original post by Muhammad Haggag This is rather clever, even though a bit ugly. Simplified cleverness removed.

Belmont, CA, huh? Did you by any chance work at Oracle at some point? I did, and their code was full of this stuff. I believe there was one header file that got included 7 times into a source file doing cleverness like the above. Ultimately I think this path ends up being more confusing than helpful; at least it confused the hell out of me! Having experienced that, I'd probably be more likely to go with matching defines and strings, and some sort of asserts to be sure things stay in sync.

By no means am I denigrating what you've suggested; it's probably the most "correct" way to do something like like in C/C++. I just want to issue this cautionary tale.

Thanks,
Geoff