Don wrote:
Andrei Alexandrescu wrote:
Bill Baxter wrote:
On Sat, Dec 27, 2008 at 9:42 AM, The Anh Tran <trthe...@gmail.com>
wrote:
aarti_pl wrote:
Andrei Alexandrescu pisze:
> We're trying to make that work. D is due for an operator overhaul.
>
> Andrei
Is there any chance that we get possibility to overload "raw
operators",
like in C++? I think that they may coexist with currently defined
operator
overloads with simple semantic rules, which will not allow them to
work
together at the same time.
..........
BR
Marcin Kuszczak
(aarti_pl)
Me also have a dream :D
<Daydream mode>
class Foo
{
auto op(++)(); // bar++
auto op(++)(int); // ++bar
op(cast)(uint); // cast(uint)bar // opCast
auto op(())(int, float); // Foo(123, 123.456) // opCall
auto op(+)(Foo rhs); // bar1 + bar2
auto op(+=)(int); // bar += 1234;
auto op(.)(); // bar.xyz // opDot
Foo op([][][])(int, char, float); // bar[123]['x'][123.456]
auto op([..])(); // i = bar2[] // opSlide
auto op([..])(int, int); // bar[1..10]
auto op([..]=)(float); // bar[] = 12.3 //opSlideAssign
auto op([..]=)(int, int, float); // bar[1..3] = 123.4
}
</Dream>
When I suggested this kind of thing long ago, Walter said that it
encourages operator overload abuse, because it suggests that + is
just a generic symbolic operator rather than something that
specifically means "addition". That's why D uses "opAdd" instead.
It's supposed to encourage only creating overloads that follow the
original meaning of the operator closely. That way when you see a+b
you can be reasonably sure that it means addition or something quite
like it.
I think that argument is rather weak and ought to be revisited. It's
weak to start with as if writing "+" in a D program hardly evokes
anything else but "plus". What the notation effectively achieved was
put more burden on the programmer to memorize some names for the
already-known symbols. I think the entire operator overloading
business, which started from a legitimate desire to improve on C++'s,
ended up worse off.
I feel quite strongly that C++'s operator overloading was a failed
experiment. The original intention (AFAIK) was to allow creation of
mathematical entities which could use natural syntax. The classic
example was complex numbers, and it works reasonably well for that,
although it requires you to create an absurd number of repetitive
functions.
But for anything much more complicated, such as matrices, tensors, big
integer arithmetic, etc -- it's an abject failure. It's clumsy, and
creates masses of temporary objects, which kills performance so
completely that it's unusable. But the whole point of operator
overloading was to allow nice notation in a performace-oriented
language! Expression templates are basically a hack to restore
performance in most cases, but it comes at a massive cost in simplicity.
And the performance even then is not always optimal.
I think that Walter's idea, in tightening the semantics of overloaded
operators, is the right approach. Unfortunately, it doesn't go far
enough, so we get the worst of both worlds: the C++ freedom is
curtailed, but there isn't enough power to replace it.
Very well put.
Ultimately, I think that the problem is that ideally, '+' is not simply
a call to a function called 'plus()'. What you'd like an operator to
compile to, depends on the expression in which it is embedded. For
maximum performance, an expression needs to be digested before it is
converted into elementary functions.
In my 'operator overloading without temporaries' proposal in Bugzilla,
I showed that DEFINING a -= b as being identical to a = a - b, and then
creating a symmetric operation for a = b - a allows optimal code
generation in a great many cases. It's not a complete solution, though.
In particular, irreducible temporaries need more thought. Ideally, in
something like a += b * c + d, b*c would be created in a memory pool,
and deleted at the end of the expression.
(By contrast, a = b*c+d, would translate to a=b*c; a+=d; so no temporary
is required).
That's an awesome proposal. I'd like to expand it to comprehend fusion
as well. Consider:
A = B + C - D;
where the operands are matrices. The best hand-written implementation
would loop once through the three matrices and assign to the destination
element-wise A[i, j] = B[i, j] + C[i, j] - D[i, j]. However, with an
approach that has only one operator application as its horizon, it is
impossible to achieve that optimization. So I wonder what abstraction
could be devised that makes it easy and natural to support such fusion.
Expression templates achieve that by saving the right-hand expression
tree as a type and then using it during the assignment. This requires a
considerable effort and has some drawbacks.
There are other, less serious problems which also need to be addressed.
Defining ++a as a+=1 is probably a mistake. It raises lots of nasty issues.
* If a is a complex number, a = a + 1 makes perfect sense. But it's not
obvious that ++a is sensible.
* What type is '1'? Is it an int, a uint, a long, ... You don't have
that issue with increment.
Great points!
As I see it, there are two possible strategies:
(1) Pursuing optimal performance, which requires semantic tightening,
and reduced flexibility, or
(2) Pursure simplicity and semantic flexibility, sacrificing performance.
I think those two possibilities are mutually exclusive.
I tend to be more optimistic, but if asked to choose, I'd go for (1).
One important lesson learned from C++'s operator overloading is that
freedom was almost always badly used. Tellingly, whenever operator
overloading is taught or talked about, the first caveat mentioned is
that defining inconsistent batteries of operators is exceedingly easy.
Andrei
