[petters]

I can see how `semi::map` could be useful, but `semi::static_map` just seems to be an alternate syntax for global variables.

[hogliux]

Yes, in a way you are right: the compiler really does boil it down to global variables. However, `semi::static_map` gives you some additional features, which may be useful for a lot of situations

1) `semi::static_map` gives you the optional run-time look-up with the same API (`semi::static_map::get`). Looking up a global variable with a runtime key isn't straight-forward. This makes it particularly nice to use an API where a function might take either a compile-time or run-time key. I'm thinking a Font API with a `getFont` method, for example. The way you use the method is the same regardless if you are using a compile-time key or not.

2) `semi::static_map` also gives you control over the lifetime of the objects. For example, it's easy to delete all the values in the map at once. Again, this is not straight-forward to do with a bunch of globals.

3) the keys are also values themselves and it's straight-forward to evaluate them at run-time. Printing the name of a global variable, for example, requires all sorts of hackery.

4) you do not need to know all your keys in advance. Simply using getFont with a unique constexpr key will create a new global variable.

Edit: added point (4)

[jstimpfle]

Re: 3), here's a clean and technically humble way that I typically use.

    enum {
        KEY_FOO,
        KEY_BAR,
        KEY_BLAH,
        NUM_KEYS,
    };

    struct KeyInfo {
        const char *name;
        int info1;
        float info2;
    };

    const static struct KeyInfo keyInfo[NUM_KEYS] = {
    #define MAKE(x, y, z) [x] = { #x, y, z }
        MAKE( KEY_FOO,   1,   1.0 ),
        MAKE( KEY_BAR,   7,   3.5 ),
        MAKE( KEY_BLAH, 42, 127.2 ),
    #undef MAKE
    };

    void print_key(int key)
    {
        printf("%d's name is %s\n", key, keyInfo[key].name);
    }

It's both low-tech and maintainable. It compiles super quick. If one doesn't like that one has to type KEY_FOO twice (in the enum and in the array definition), one can use X-macros or code generation. But personally I don't care.

[hogliux]

Yes, that's a nice approach. However, this approach requires you to list all your keys in advance (in the enum).

The `semi::static_map` does not require this.This is especially useful if you are writing library code: imagine you are programming a `getFont` method - you don't know with which constexpr keys your method will be called, so there is no way for you to list all these keys in an enum.

[kazinator]

If you don't like pointlessly repeating occurrences of identifiers, you should hardly be favoring C++.

   class foo {
   public:
     foo();
     ~foo();
   };
 
   foo::foo() { }

   foo::~foo() { }

Ad nauseum.

Oh, but C++ can eliminate redundancy in this one cool case I'm thinking of!

[IshKebab]

C++ doesn't support designated initialisers.

[petters]

Thanks for this clarification! I agree it can be useful.

[Nalta]

True, but I think this could really be useful for cases similar to google flags (https://github.com/gflags/gflags) wherein each file in a library adds to the global configuration. Then your main can more easy parse out these options to be shared across the entire codebase.

[nightfly]

But somehow slower! "In fact, when using semi::static_map and looking up a value with C++ literal as a key, then the value lookup is nearly as efficient as looking up a global variable."

[hogliux]

Yes with the optional run-time lookup there is an extra two machine instructions (a cmp and jne) needed to check if this is the first time accessing the value. It is exactly as fast as a global variable if you don't need the optional run-time lookup.

As stated in the talk, the main use case for a map like this is to cache objects which are expensive to load/compute (again something like `getFont` comes to mind). In these use-cases you would likely also need a check like this anyway: a global pointer object which you need to check if it's a nullptr or not before using it.

[quotemstr]

Not exactly as fast. Without -Bsymbolic, a global variabile with global symbol visibility gets its own GOT entry, meaning access to your global goes through an indirection table that supports ELF symbol interposition. (Which, IMHO, was a bad idea, but you can't unbreak an egg.) With a data structure like this (or with a conventional "struct globals"), there's one symbol at the ELF level to look up instead of one per variable, so your code might end up doing fewer GOT lookups.

The last paragraph was just arcana though. You should be passing -fvisibility=hidden and explicitly exporting from your DSO only the symbols you might legitimately expect some other DSO to use, obviating the issue.

[hogliux]

Thanks, I'll have a look at this!

[jstimpfle]

Look at the approach from my other comment. Requires only 1 symbol.

[hogliux]

Both your approach and the `semi::static_map` will generate the same type of load instruction. In both cases, the compiler knows the offset of the values in memory at compile-time. The number of symbols is not really important here.