On the richness of C++

[Kevin Bealer]

Sean Kelly Wrote:

> == Quote from Sean Kelly (sean at invisibleduck.org)'s article
> > == Quote from Janice Caron (caron800 at googlemail.com)'s article
> > > On 16/04/2008, Walter Bright <newshound1 at digitalmars.com> wrote:
> > > >  The only way to merge function pointers with delegates is to have the
> > > > compiler generate trampolines. This isn't very runtime efficient - the
> > > > template approach will be faster.
> > > I think that all that is being suggested is that function (that which
> > > you get when you take the address of a static or global function)
> > > should implicitly cast to delegate (that which you get when you take
> > > the address of a local or member function).
> > I think the problem is that functions and delegates are called differently.
> > That is, the context pointer is passed as an invisible first parameter to
> > a delegate, while functions have no such invisible argument.  Interestingly,
> > because the first parameter of a D function is supposed to be passed in a
> > register (EAX) on x86, I think you're right that it should be possible to
> > allow such efficient conversion so long as the D calling convention
> > accounts for it properly (and the platform supports such things).
> 
> Er... forget that.  A D function would obviously expect its first argument to
> be in EAX regardless of whether it's a delegate.  Some conversion would be
> necessary unless I'm missing something.
> 
> 
> Sean

I'm not much of an ASM guru, but I would think that this could be done by
"multiple entry point" functions, i.e. prefixes that move registers around.

Non-delegate users would not pay for this code, but register-to-register
"MOV" should be cheap anyway, since this part of the code only touches
fields that must be touched soon anyway, and no branching, jumping, or
register saving/unsaving cycles is needed, so it should be pipelinable.

Pseudo ASM code:

method_start:
  debug { assert(! this); }
  mov arg1, arg2
  mov arg2, arg3
  mov arg3, [non-register argument]
  (adjust stack fields to indicate 1 fewer arguments?)

function_start:
  (code expecting a1, a2, a3)
  return ...;

I guess this could only apply to "D linkage" functions, since we probably
don't know the parentage of C linkage functions, and some won't have the
special code.

Kevin