Oh brother...

What about: lambdas, type inference, tuples, robust templates, compile time constants and code execution, move semantics, expanded PoD, initializer lists, range based for loops, user literals, return type deduction... I could go on and on. All features I use daily. And I'm not even getting into C++17.

Half-broken, bolted-on garbage that is found in a much nicer form in high level languages which have those things in their core vision from the beginning, or the kind of extensibility which can cleanly support new ideas.

> And I'm not even getting into C++17.

I'm with you there, in a way.

I don't find auto, lambda, and variadic templates, move semantics and smart pointers to be bolted on. You can write Fortran in any language.

You think rvalue references are a natural and elegant construct? That if C++ were designed with move semantics in mind, this is the road they would have travelled down?

That's the biggest reason that I like Rust more than C++, even more than lifetime/safety. I've been working on a C++14 application, and "std::move" shows up a lot more than implicit copies, which not only makes the code more verbose, but analyzing performance more difficult. (I so wish I could be using Rust, but, unfortunately, the only FOSS DNP3 implementation is in C++.)

Please show me the "nice" and "well thought" and "elegant" and "modern" language with the sheer amount of resources, documentation and momentum that C++ has? It's not going to go away for the foreseeable future, even if slightly better options arise (notice the slightly, no language has substantially improved on C++ in a meaningful enough way to make the benefits of migrating greater than the headaches).

Here is another thing: each time they add this or that cruft to C++, the syntax for that tidbit has to go into the hard-coded grammar. For the past 30 years, they have been missing opportunity after opportunity to somehow make the language internally extensible, so that new rules don't have to be added any more to a hard-coded parser (but, say, defined by C++ programs themselves). So now, check this out, GNU C++ has a hand-written recursive-descent parser ... that is 1195625 characters of C. It's almost 40K lines just to scan some tokens and get an abstract syntax tree.

https://gcc.gnu.org/git/?p=gcc.git;a=tree;f=gcc/cp

At this point, is there any parser out there that could deal with C++'s grammar?

What you're suggesting is tantamount to throwing out C++'s current syntax, which will never happen, or taking the existing parser and making it even more complicated so that it supports extensions.

IIRC g++ had historically a machine generated parser (bison maybe?) But it got replaced around 3.0 time because the new hand written parser was faster and more maintainable.

I think, you cannot beat beats shift-reduce LALR(1) with recursive descent on raw performance on a reasonable syntax. However, parsing C++ with LALR(1) will require unmaintainable hacks in the grammar. Recursive descent accomodates ad-hoc fudging much better than parser generation. In any given rule, you can look ahead as much as you want and invoke any Turing computation to decide the parse. Also, recursive descent approaches support backtracking fairly naturally: try it this way, as far ahead as necessary, or else start again, falling back on this. Let's see ...

  28374 /* Commit to the currently active tentative parse.  */
  28375 
  28376 static void
  28377 cp_parser_commit_to_tentative_parse (cp_parser* parser)
  28378 {
  28379   cp_parser_context *context;
  28380   cp_lexer *lexer;
  28381 
  28382   /* Mark all of the levels as committed.  */
  28383   lexer = parser->lexer;
  28384   for (context = parser->context; context->next; context = context->next)
  28385     {
  28386       if (context->status == CP_PARSER_STATUS_KIND_COMMITTED)
  28387         break;
  28388       context->status = CP_PARSER_STATUS_KIND_COMMITTED;
  28389       while (!cp_lexer_saving_tokens (lexer))
  28390         lexer = lexer->next;
  28391       cp_lexer_commit_tokens (lexer);
  28392     }
  28393 }

Bison supports GLR ("generalized LR") via some technique of forking the parser and trying alternatives in parallel, which seems related to this.

The problem with C++ grammar is that inside templates, things are too flexible at times. For example:

    template<bool> struct a_t;

    template<> struct a_t<true> {
        template<int> struct b {};
    };

    template<> struct a_t<false> {
        enum { b };
    };

    typedef a_t<sizeof(void*)==sizeof(int)> a;

    enum { c, d };
    int main() {
        a::b<c>d; // declaration or expression?
    }

It's a fun C++ IDE torture test, which few pass. One that does is Visual Studio - if you create a C++ project and set up both 32-bit and 64-bit build targets, you can flip between them, and see the syntax highlighting change on the line with the comment.

Make it more complicated once (bite the proverbial bullet) and then stop, versus keep on creeping in complexity.