C++ compiler vs D compiler

[Timon Gehr via Digitalmars-d]

On 10/03/2015 12:45 PM, Abdulhaq wrote:
> Perhaps the answer to this is obvious, but what's harder to write from
> scratch - a C++ compiler or a D compiler? :-)
>
> We know Walter wrote a C++ compiler single handedly, does anyone else
> recall the C++ Grandmaster qualification, the free course where
> participants get to write a complete C++ compiler from scratch? I think
> it's dead now but I can't find any real info about that despite a
> serious google. What's the chances of anyone single-handedly writing a D
> compiler from scratch in the future? I know deadalnix is writing SDC in
> D - sounds interesting.

I have also started a similar project four years ago (it's currently 
just an incomplete compiler front-end), but I have not been able to work 
much on it during the last year. (It only compiles with DMD 2.060 
though, because of issues similar to what I discuss below.)

> Is the D language well enough documented /
> specified for a complete D implementation to be even possible (as things
> stand now)?
>
>

Well, not really. The main impediment to a fully formal specification is 
the interplay of forward references and the meta-programming system. 
(DMD just does a best-effort kind of thing, where the common case 
scenarios for which there were bug reports work, but in general the 
semantics of the resulting code can depend on things like the order that 
modules are passed on the command line, or perfectly valid code is 
rejected with a "forward reference error".) The documentation just 
specifies that everything works, but there is no consistent way to 
interpret it.

E.g.:

static if(!is(typeof(x))) enum y=2;
static if(!is(typeof(y))) enum x=2;

Arbitrarily abstruse examples can be constructed, e.g. this is from my 
test suite:

struct TestInvalidInheritance{
     class A{ int string; } // error
     template Mixin(string s){
         mixin("alias "~s~" Mixin;");
     }
     class D: Mixin!({D d = new E; return d.foo();}()){
         int foo(int x){ return 2;}
         string foo(){ return "X"; }
     }
     class E: D{
         override int foo(int x){ return super.foo(x); }
         override string foo(){ return "A"; }
     }
}

(I currently accept this code if the declaration of int string in class 
A is removed. Otherwise the code is analyzed until the point when it is 
clear that A is in D's superclass chain, and it is also clear that the 
symbol 'string' which was necessary to resolve in order to discover this 
fact was resolved incorrectly. The compiler then gives up and prints an 
error message:

example.d:3:18: error: declaration of 'string' is invalid
     class A{ int string; } // error
                  ^─────
example.d:9:9: note: this lookup on subclass 'D' should have resolved to it
         string foo(){ return "X"; })
         ^─────

I and SDC have different ways to deal with those kinds of examples, and 
I think the SDC way does not work (unless things have changed since I 
have looked at it). It assumes that declarations can be ordered and that 
it is fine to depend on the order of declarations.

My implementation is designed to be independent of declaration order and 
to reject the cases where there is no single obvious and consistent 
interpretation of the program. The drawbacks currently are:

- It is overly conservative in some cases, especially when string mixins 
are involved. E.g., the following code has only one consistent 
interpretation, but it is rejected (as is any reordering of those 
declarations):

enum x = "enum xx = q{int y = 0;};";

struct SS{
     mixin(xx);
     mixin(x);
}

- The current implementation is somewhat slow. IIRC, N-fold recursive 
template instantiation currently runs in Ω(N²). It's clear that this 
needs to be improved. If this is to be adopted as the official solution, 
it should not make the compiler any slower, at least in the common case.


There are also some other, more minor issues. For example, when the 
language specification speaks about "memory safety", it is really 
unclear what this means, as the language designers seem to think it that 
it is fine to have undefined behaviour in a section of code that is 
"verified memory safe".