Optimizing delegates

[Andrei Alexandrescu]

On 12/19/10 11:13 AM, Ary Borenszweig wrote:
> On 12/19/2010 01:44 PM, Andrei Alexandrescu wrote:
>> On 12/19/10 10:35 AM, Ary Borenszweig wrote:
>>> On 12/19/2010 01:21 PM, Andrei Alexandrescu wrote:
>>>> On 12/19/10 9:32 AM, Ary Borenszweig wrote:
>>>>> I have this code:
>>>>>
>>>>> ---
>>>>> import std.stdio;
>>>>>
>>>>> int foobar(int delegate(int) f) {
>>>>> return f(1);
>>>>> }
>>>>>
>>>>> int foobar2(string s)() {
>>>>> int x = 1;
>>>>> mixin("return " ~ s ~ ";");
>>>>> }
>>>>>
>>>>> void main() {
>>>>> writefln("%d", foobar((int x) { return 2*x; }));
>>>>> writefln("%d", foobar2!("9876*x"));
>>>>> }
>>>>> ---
>>>>>
>>>>> When I compile it with -O -inline I can see with obj2asm that for the
>>>>> first writefln the delegate is being called. However, for the second
>>>>> it just passes
>>>>> 9876 to writefln.
>>>>>
>>>>> From this I can say many things:
>>>>> - It seems that if I want hyper-high performance in my code I must use
>>>>> string mixins because delegate calls, even if they are very simple and
>>>>> the
>>>>> functions that uses them are also very simple, are not inlined. This
>>>>> has the drawback that each call to foobar2 with a different string
>>>>> will generate a
>>>>> different method in the object file.
>>>>
>>>> You forgot:
>>>>
>>>> writefln("%d", foobar2!((x) { return 2*x; })());
>>>>
>>>> That's a real delegate, not a string, but it will be inlined.
>>>>
>>>>
>>>> Andrei
>>>
>>> Sorry, I don't understand. I tried these:
>>>
>>> 1.
>>> int foobar3(int delegate(int) f)() {
>>> return f(1);
>>> }
>>>
>>> writefln("%d", foobar3!((int x) { return 2*x; })());
>>>
>>> => foo.d(12): Error: arithmetic/string type expected for
>>> value-parameter, not int delegate(int)
>>>
>>> 2.
>>> int foobar3()(int delegate(int) f) {
>>> return f(1);
>>> }
>>>
>>> writefln("%d", foobar3!()((int x) { return 2*x; }));
>>>
>>> => Works, but it doesn't get inlined.
>>>
>>> And I tried that "(x) { ... }" syntax and it doesn't work.
>>>
>>> Sorry, it must be my fault I'm doing something wrong. What's the correct
>>> way of writing optimized code in D, code that I'm sure the compiler will
>>> know how to optimize?
>>
>> void foobar3(alias fun)() {
>> return fun(1);
>> }
>>
>>
>> Andrei
>
> This of course has the following problem:
>
> int foobar2(int delegate(int x) f) {
> }
>
> foobar2((int x, int y) { ... });
>
> Error: function foobar2 (int delegate(int) f) is not callable using
> argument types (int delegate(int x, int y))
>
> ---
>
> int foobar3(alias f)() {
> f(1);
> }
>
> foobar3((x, y) { ... });
>
> foo.d(8): Error: template foo.main.__dgliteral1(__T2,__T3) does not
> match any function template declaration
> foo.d(8): Error: template foo.main.__dgliteral1(__T2,__T3) cannot deduce
> template function from argument types !()(int)
> foo.d(12): Error: template instance foo.main.foobar3!(__dgliteral1)
> error instantiating
>
> So I have to go to foo.d(8) to see what the problem is, understand what
> is being invoked (in this case it was easy but it get can harder), or
> otherwise say "Hey, the one that implemented foo, please do a static
> assert msg if f is not what you expect". Basically "Implement the error
> message that the compiler would have given you for free if you didn't
> use a template".

Template constraints are meant to assuage that problem.

Inlining delegates is technically much more difficult than inlining 
aliases. This is because a different function will be generated for each 
alias argument, whereas only one function would be used for all 
delegates. There are techniques to address that in the compiler, but 
they are rather complex.


Andrei