digitalmars.D - Value-based function overloading - Going Fourth
- Nod (93/93) May 20 2005 Replies are scarce. How shall I interpret this silence? That the techniq...
- Matthias Becker (1/4) May 21 2005 You are talking about what is generally known as predicate methods.
- Nod (5/9) May 21 2005 Nice, now I don't have to invent new names for it all the time :)
- pragma (32/45) May 21 2005 I can only speak for myself: I was skeptical at first, but now I see so...
- Nod (39/86) May 21 2005 Music to my ears! Ehm.. eyes. :)
- TechnoZeus (31/124) May 21 2005 Nothing positive to add... so simply observing.
- Nod (21/51) May 21 2005 You have understood things very well. And your example is a nice one too...
- TechnoZeus (13/27) May 22 2005 About the same to me at this time, but I could deal with it,
- Burton Radons (40/43) May 21 2005 All right, you asked for it.
- Nod (40/85) May 21 2005 You are missing the extra factors in the upper block of code. There's th...
- Hasan Aljudy (3/7) May 21 2005 I was enthusiastic about the idea, but I got lost in the second take.
- Nod (4/11) May 21 2005 Argh. Maybe I was getting too wound up in the nitty-gritty details.
Replies are scarce. How shall I interpret this silence? That the technique is
perfect, and requires no further improvement? That the technique is useless, and
not worth discussing? That no one really understands what the hell I'm on about?
Who knows. We push on. Mixins and range support in this draft part four.
Here are the first three drafts for reference:
1) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23796
2) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23897
3) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23966
/****************************************
* Mixins
***/
Mixins act identically to templates(see draft 3), except for, of course, their
being mixins.
Extending the debugging example of the last draft, we get the following savoury
snippet.
Example:
:template debugFoo(R)
:{
: override void foo(char c in R)
: {
: printf("Foo called with: %c\n", c);
: .foo(c);
: }
:}
:void foo(char c) { ... }
:int main()
:{
: mixin debugFoo!(" "); // catch only the character in the main scope
: void bar()
: {
: /+ ... foo invocations ... +/
: }
: /+ ... foo invocations ... +/
:}
This enables swift, one-liner, drop-in monitoring of function calls, in a
specific scope.
/****************************************
* Range support
***/
With value types, range or array literals are used to specify ranges: 0..10,
'a'..'z', 0.0..1.0, ["foo","bar","baz"]. Literal ranges are inclusive, as
suggested in a response from Thomas Kuehne[1].
Container objects can act as ranges. They have to overload the opIn method:
:class Foo { ... } // FooContainer contains Foos
:class FooContainer { opIn(Foo obj) { return (...) ? true : false; } }
:int main()
:{
: void foo(Foo f in myFooCont) { ... }
: // you get the picture :)
:}
Support for boolean operations on ranges, such as and/or/xor could become
available once array operations are implemented.
Example:
:const int[] a = 0..5;
:const int[] b = 15..20;
:
:void foo(int i) { ... }
:void foo(int i in a & b) { ... } // joins ranges
:int main()
:{
: foo(2); // calls second
: foo(10); // calls first
: foo(17); // calls second
:}
This isn't very efficient though, as for ints/chars/floats, we only need to know
the endpoints of the range. A simple template will solve that, but built-in
compiler support for these operations on ranges would be more efficient.
Example of built-in support:
:void foo(int i) { ... }
:void foo(int i in 0..5 & 15..20) { ... } // joins ranges
:// results as above
Ranges can also be used in switch-case/if-else statements and for initializing
arrays, upping code clarity a notch or two.
Example:
:switch (i)
:{
:case 0..3: ... break;
:case 4..6: ... break;
:default:
:}
Example:
:if (i in 128..256) { ... }
Example:
:char[] str = 'a'..'z'; // abcdef...xyz
/****************************************
* Wee!
***/
I'm really getting more and more excited about this technique. It seems to fit
very well into D, and *I* have *actually* found some useful idioms. Who would
have thought that? :)
-Nod-
[1] http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23839
May 20 2005
Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?You are talking about what is generally known as predicate methods.
May 21 2005
Nod <Nod_member pathlink.com> In article <d6n9bf$2v7v$1 digitaldaemon.com>, Matthias Becker says...Nice, now I don't have to invent new names for it all the time :) Have you got any resources about this, or examples of languages in which it is implemented? -Nod-Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?You are talking about what is generally known as predicate methods.
May 21 2005
In article <d6mkfr$2cb9$1 digitaldaemon.com>, Nod says...Replies are scarce. How shall I interpret this silence?I can only speak for myself: I was skeptical at first, but now I see some merit in this feature. I also dismissed things as being more applicable to the in{} block with asserts, until I realized that you're attempting to override based on constant values. I'm also a big supporter of this approach, so please, continue refining your work here in the newsgroup. :) If I may, I have a few critiques and suggestions. - How does this specification affect the D calling convention and linkages? I suspect you'd rely on the compiler to find the predicates in the D source code, and then use the symbol for that particular method. In that case, you'd need to find a way to use D's name mangling to somehow tell a set of similar methods apart. I suspect this would need a revision of the name-mangling spec, including how to reliably map specific predicates (read: not dependent on the order they were parsed/compiled) to D-style symbols. - From your specification: ">:void foo(int i in a & b) { ... } // joins ranges" Wouldn't the predicate expression be better off defined as a typical D statement, provided that it operates strictly on constants? If a and b are both arrays, wouldn't '~' be more appropriate? If that causes trouble, then why not use the typical boolean and operator '&&'. - One of the things on the table for D is that templates matches be deduced, so the template call syntax (via '!') would only be needed to solve ambiguities (like in C++). Should such a feature become a part of D, wouldn't these predicate expressions be better suited for templates instead?template foo(A : int in [3,4,5]){ foo(A a){ /* specialized foo } } template foo(A : int){ foo(A a){ /* generalized foo } } void main(){ foo(4); // matches specialized template for foo foo(1); // matched generalized template for foo foo(new Object()); // error: no match for foo(Object); }Not to turn your spec completely on its head, but it would mesh well with the current separation of responsibility between templates and functions. Templates already do some crude constant matching. It also makes sure that the compiler can supply the appropriate symbols for the linker as this already works with explicit template calls. Then again, if predicates were extended to templates (like above), then your syntax would simply be a shortcut for function templates. - EricAnderton at yahoo
May 21 2005
In article <d6nhtr$51a$1 digitaldaemon.com>, pragma says...In article <d6mkfr$2cb9$1 digitaldaemon.com>, Nod says...Music to my ears! Ehm.. eyes. :) A small clarification though: The technique is not limited to constant values. This a side-effect the fact that any overloads may be perfectly rewritten as an if-else or switch-case construct which can be executed at run-time (incurring a small run-time overhead on each function call). For example: :int a = 0, b = 10; // globals :void foo(int i in a..b) { ... } :void foo(int i) { ... } :int main() :{ : foo(2); // calls first : a = 5; : foo(2); // calls second :}Replies are scarce. How shall I interpret this silence?I can only speak for myself: I was skeptical at first, but now I see some merit in this feature. I also dismissed things as being more applicable to the in{} block with asserts, until I realized that you're attempting to override based on constant values. I'm also a big supporter of this approach, so please, continue refining your work here in the newsgroup. :)If I may, I have a few critiques and suggestions.Critiques/suggestions are highly welcomed.- How does this specification affect the D calling convention and linkages? I suspect you'd rely on the compiler to find the predicates in the D source code, and then use the symbol for that particular method. In that case, you'd need to find a way to use D's name mangling to somehow tell a set of similar methods apart. I suspect this would need a revision of the name-mangling spec, including how to reliably map specific predicates (read: not dependent on the order they were parsed/compiled) to D-style symbols.I admit I have not yet given this much thought, but on first glance, what you're saying seems to be correct. Constant ranges would have to be uniquely mangled. Though, the order in which they were parsed/compiled doesn't matter currently either, as required for forward referencing, so that shouldn't be a problem. Functions with dynamic ranges should only need one symbol, as the by-value overload logic gets folded into the function. Hmm, this may not always be the best way to do it though, as the compiler could inline the overload logic for performance. I'm not yet sure how to solve this case. Tricky stuff, mangling.- From your specification: ">:void foo(int i in a & b) { ... } // joins ranges" Wouldn't the predicate expression be better off defined as a typical D statement, provided that it operates strictly on constants? If a and b are both arrays, wouldn't '~' be more appropriate? If that causes trouble, then why not use the typical boolean and operator '&&'.The a & b example assumes array operations has been implemented, they currently are not, but the docs say they are planned. I assume that they would work on arrays in the same way as bit operations work on bits, allowing not only merging but also subtraction |, difference ^, and negation !. Using ~ gives an odd condition when using overlapping ranges. One could of course special-case this to be allowed in this context. :0..4 ~ 2..6 // 0123423456 :0..4 & 2..6 // 0123456- One of the things on the table for D is that templates matches be deduced, so the template call syntax (via '!') would only be needed to solve ambiguities (like in C++). Should such a feature become a part of D, wouldn't these predicate expressions be better suited for templates instead?Yes, this should be possible, but in addition to the "shortcut syntax" (see below).template foo(A : int in [3,4,5]){ foo(A a){ /* specialized foo } } template foo(A : int){ foo(A a){ /* generalized foo } } void main(){ foo(4); // matches specialized template for foo foo(1); // matched generalized template for foo foo(new Object()); // error: no match for foo(Object); }Not to turn your spec completely on its head, but it would mesh well with the current separation of responsibility between templates and functions. Templates already do some crude constant matching.It also makes sure that the compiler can supply the appropriate symbols for the linker as this already works with explicit template calls.I don't fully understand this. Could you elaborate, or give an example?Then again, if predicates were extended to templates (like above), then your syntax would simply be a shortcut for function templates. - EricAnderton at yahooThe shortcut syntax is, in itself, valuable. It allows easy extension/splitting of functions without having to touch the syntax of the original function. That means less risk for error, and less work for your fingers :) -Nod-
May 21 2005
Sure. The D spec already has a way of name-mangling templates, so that all the possible instances of templates are mapped to unique symbols. Piggybacking or extending templates as we understand them, via these predicate expessions, would be easier to implement than another form of function overloading altogether, IMO. - EricAnderton at yahooIt also makes sure that the compiler can supply the appropriate symbols for the linker as this already works with explicit template calls.I don't fully understand this. Could you elaborate, or give an example?
May 22 2005
In article <d6q5mg$21vb$1 digitaldaemon.com>, pragma says...Ah, yes indeed, from that perspective you are correct. Though I won't give up the more complex version just yet. It may be harder to implement, but it also has its share of additional benefits. Less typing and more integration with the language as a whole to name a few. -Nod-Sure. The D spec already has a way of name-mangling templates, so that all the possible instances of templates are mapped to unique symbols. Piggybacking or extending templates as we understand them, via these predicate expessions, would be easier to implement than another form of function overloading altogether, IMO. - EricAnderton at yahooIt also makes sure that the compiler can supply the appropriate symbols for the linker as this already works with explicit template calls.I don't fully understand this. Could you elaborate, or give an example?
May 22 2005
"Nod" <Nod_member pathlink.com> wrote in message
news:d6mkfr$2cb9$1 digitaldaemon.com...
Replies are scarce. How shall I interpret this silence? That the technique is
perfect, and requires no further improvement? That the technique is useless,
and
not worth discussing? That no one really understands what the hell I'm on
about?
Who knows. We push on. Mixins and range support in this draft part four.
Here are the first three drafts for reference:
1) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23796
2) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23897
3) http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23966
/****************************************
* Mixins
***/
Mixins act identically to templates(see draft 3), except for, of course, their
being mixins.
Extending the debugging example of the last draft, we get the following savoury
snippet.
Example:
:template debugFoo(R)
:{
: override void foo(char c in R)
: {
: printf("Foo called with: %c\n", c);
: .foo(c);
: }
:}
:void foo(char c) { ... }
:int main()
:{
: mixin debugFoo!(" "); // catch only the character in the main scope
: void bar()
: {
: /+ ... foo invocations ... +/
: }
: /+ ... foo invocations ... +/
:}
This enables swift, one-liner, drop-in monitoring of function calls, in a
specific scope.
/****************************************
* Range support
***/
With value types, range or array literals are used to specify ranges: 0..10,
'a'..'z', 0.0..1.0, ["foo","bar","baz"]. Literal ranges are inclusive, as
suggested in a response from Thomas Kuehne[1].
Container objects can act as ranges. They have to overload the opIn method:
:class Foo { ... } // FooContainer contains Foos
:class FooContainer { opIn(Foo obj) { return (...) ? true : false; } }
:int main()
:{
: void foo(Foo f in myFooCont) { ... }
: // you get the picture :)
:}
Support for boolean operations on ranges, such as and/or/xor could become
available once array operations are implemented.
Example:
:const int[] a = 0..5;
:const int[] b = 15..20;
:
:void foo(int i) { ... }
:void foo(int i in a & b) { ... } // joins ranges
:int main()
:{
: foo(2); // calls second
: foo(10); // calls first
: foo(17); // calls second
:}
This isn't very efficient though, as for ints/chars/floats, we only need to
know
the endpoints of the range. A simple template will solve that, but built-in
compiler support for these operations on ranges would be more efficient.
Example of built-in support:
:void foo(int i) { ... }
:void foo(int i in 0..5 & 15..20) { ... } // joins ranges
:// results as above
Ranges can also be used in switch-case/if-else statements and for initializing
arrays, upping code clarity a notch or two.
Example:
:switch (i)
:{
:case 0..3: ... break;
:case 4..6: ... break;
:default:
:}
Example:
:if (i in 128..256) { ... }
Example:
:char[] str = 'a'..'z'; // abcdef...xyz
/****************************************
* Wee!
***/
I'm really getting more and more excited about this technique. It seems to fit
very well into D, and *I* have *actually* found some useful idioms. Who would
have thought that? :)
-Nod-
[1] http://www.digitalmars.com/drn-bin/wwwnews?digitalmars.D/23839
Nothing positive to add... so simply observing.
Observations so far indicate that it's a viable idea that represents a
different way of writing something like...
void main()
{
int v;
/* assign a value to v ... */
printf("%d",a(v));
}
int a(int v)
{
int a_lm5(int v){ /*...*/ }
int a_m5t5(int v){ /*...*/ }
int a_g5(int v){ /*...*/ }
if (v < -5) return a_lm5(v);
if (v <= -5 && v >= 5) return a_m5t5(v);
if (v > 5) return a_g5(v);
}
/.....
and whether or not that is of value to a person would depend, I would think, on
that person's point of view...
so if I have misunderstood, or missed something... please let me know.
In the mean time, this reason and the lack of sufficient time to keep up as
well as I would like to are why I have been silent on this subject.
I would like to say though, that I have noticed your use of structures like 0
.. 9 such as...
if (v in -5 .. 5) return a_m5t5(v);
...and I have yet to find documentation on this type of structure in the D
language, although it would be nice to see.
In other languages, I have seen the equivilent as [0 .. 9] which in D
represents (to my knowledge) at array slice,
although I see no reason why it couldn't also be used in something like...
if (v in [-5 .. 5]) return a_m5t5(v);
...except that in D array slices, [-5 .. 5] means from -5 up to but not
including 5, while the range [-5 .. 5] as I have seen elsewhere
would mean from and including -5, to and including 5.
TZ
May 21 2005
In article <d6nncg$8ps$1 digitaldaemon.com>, TechnoZeus says...You have understood things very well. And your example is a nice one too. The idea is much about making things more clear and maintainable. Wouldn't you agree that the following is easier to read and understand? void main() { int v; /* assign a value to v ... */ printf("%d",a(v)); } int a(int v in int.min..-5) { /*...*/ } int a(int v in -5..5) { /*...*/ } int a(int v in 5..int.max) { /*...*/ }<snip>Nothing positive to add... so simply observing. Observations so far indicate that it's a viable idea that represents a different way of writing something like... void main() { int v; /* assign a value to v ... */ printf("%d",a(v)); } int a(int v) { int a_lm5(int v){ /*...*/ } int a_m5t5(int v){ /*...*/ } int a_g5(int v){ /*...*/ } if (v < -5) return a_lm5(v); if (v <= -5 && v >= 5) return a_m5t5(v); if (v > 5) return a_g5(v); /..... and whether or not that is of value to a person would depend, I would think, on that person's point of view... so if I have misunderstood, or missed something... please let me know.In the mean time, this reason and the lack of sufficient time to keep up as well as I would like to are why I have been silent on this subject.Sorry, I'm too impatient I guess :)I would like to say though, that I have noticed your use of structures like 0 .. 9 such as... if (v in -5 .. 5) return a_m5t5(v); ...and I have yet to find documentation on this type of structure in the D language, although it would be nice to see.The addition of ranges like that is something of a "side-proposal". It's not strictly necessary for value-based overloading, but it makes things more clear. By using it in the drafts, I'm doing a bit of advertisement for that :)In other languages, I have seen the equivilent as [0 .. 9] which in D represents (to my knowledge) at array slice, although I see no reason why it couldn't also be used in something like... if (v in [-5 .. 5]) return a_m5t5(v); ...except that in D array slices, [-5 .. 5] means from -5 up to but not including 5, while the range [-5 .. 5] as I have seen elsewhere would mean from and including -5, to and including 5. TZIn my proposal, slices are always inclusive at both ends. I feel that adding support for exclusive ranges would be more trouble than it's worth in terms of syntactical clutter vs. usefulness. -Nod-
May 21 2005
"Nod" <Nod_member pathlink.com> wrote in message news:d6oten$1255$1 digitaldaemon.com...In article <d6nncg$8ps$1 digitaldaemon.com>, TechnoZeus says... You have understood things very well. And your example is a nice one too. The idea is much about making things more clear and maintainable. Wouldn't you agree that the following is easier to read and understand? void main() { int v; /* assign a value to v ... */ printf("%d",a(v)); } int a(int v in int.min..-5) { /*...*/ } int a(int v in -5..5) { /*...*/ } int a(int v in 5..int.max) { /*...*/ }*snip*-Nod-About the same to me at this time, but I could deal with it, and if it's easier or more clear to someone else, then I'm all for it. In fact, with a small amount of actual usage, I could probably learn to prefer it... although I have a feeling it would be difficult to implement. The only things left to determine, in my opinion, are... a set of exact details for a proposed standard, and whether or not the amount of good it would do is worth the time and effort needed to implement the proposed standard. TZ
May 22 2005
Nod wrote:Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?All right, you asked for it. At the core, writing one function this way: void foo (int a) { if (a < 5) ... else if (a < 10) ... } And another function this way: void foo (int a in int.min .. 4) ... void foo (int a in 5 .. 9) ... Doesn't matter, they're conceptually equivalent. However, there are four extra factors in the second example. There's the repeating declaration (names, return type, any additional parameters), the need to understand exactly how the first example works, the ability to separate the declarations so that they're no longer related, and the fact that the range doesn't behave like a regular D range. So it has higher complexity, not lower: real control flow is more concise, more flexible, and easier to deal with. This leads me to believe that this would simply cause confusion because it makes the feature hard to use properly. How would I document these functions? How do I share common processing between them? How do I override these methods - without looking at the specification to be certain I'm doing this right? This also screws up function overloading, making it a lot more complex and hazardous - particularly inner functions and overrides and whatever else you had going on in there. That entirely misses the spirit of D's function overloading. Now, templates COULD have benefited from this, but that was before static if. Now that there is control flow in the templates, there's no reason to have it in specialization. I won't speak to your mixins and overrides and the other dozen things you have going on here because then we'll be here for weeks. Where's the dire need - or ANY need? This is all just syntax sugar, and in this case syntax sugar that increases complexity! I don't see any reason at all why I would spend a few weeks implementing this feature; I would be in the exact same place as before, only with a much more complex language.
May 21 2005
In article <d6o7sg$kdh$1 digitaldaemon.com>, Burton Radons says...Nod wrote:Sure did. *provocative grin*Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?All right, you asked for it.At the core, writing one function this way: void foo (int a) { if (a < 5) ... else if (a < 10) ... } And another function this way: void foo (int a in int.min .. 4) ... void foo (int a in 5 .. 9) ... Doesn't matter, they're conceptually equivalent. However, there are four extra factors in the second example. There's the repeating declaration (names, return type, any additional parameters), the need to understand exactly how the first example works, the ability to separate the declarations so that they're no longer related, and the fact that the range doesn't behave like a regular D range. So it has higher complexity, not lower: real control flow is more concise, more flexible, and easier to deal with. This leads me to believe that this would simply cause confusion because it makes the feature hard to use properly.You are missing the extra factors in the upper block of code. There's the extra level of indentation, the dependency on block order, the inability to override blocks in specific scopes, and the denseness of code that results from cramming all the cases into one function. So it turns out about equal. But if you also take into account that the repeating declarations - and more - is needed also in your example when branching out into single functions, that you equally need to understand all of the code in both examples, that separating the declarations can equally well be beneficial, and that the range issue is just a pet peeve of mine, and can be resolved... then my version wins <g>. But scorn as I might, I agree with you. It does not really matter which one you choose. What matters is that you have a choice. Because sometimes one of them is the better choice.How would I document these functions?You would document them as you would any other overload. Explain that there are overloads, and what the differences are. As you will probably be writing for an audience that knows D, they'll understand what you mean.How do I share common processing between them?Try mixins. Or functions. Or just don't use them when there's a lot of common processing.How do I override these methods - without looking at the specification to be certain I'm doing this right?By actually *learning* the specification?!This also screws up function overloading, making it a lot more complex and hazardous - particularly inner functions and overrides and whatever else you had going on in there. That entirely misses the spirit of D's function overloading.If you would actually *read* the drafts you'll see that the feature integrates perfectly with inner functions, and screws up no old code whatsoever. It does make the compiler more complex, I agree with that, but not as much as you'd think. Most features are simple extensions of already existing features. As for making things more hazardous, that's only true if you misuse it; like anything else. And what exactly *is* the spirit of D function overloading, that I'm so totally missing? I thought D was a practical language, for practical programmers, and aims to make programming less error-prone. By modularizing the code, this feature fits.Now, templates COULD have benefited from this, but that was before static if. Now that there is control flow in the templates, there's no reason to have it in specialization. I won't speak to your mixins and overrides and the other dozen things you have going on here because then we'll be here for weeks.You have been here for years already. What's the rush?Where's the dire need - or ANY need? This is all just syntax sugar, and in this case syntax sugar that increases complexity! I don't see any reason at all why I would spend a few weeks implementing this feature; I would be in the exact same place as before, only with a much more complex language.It is syntactic sugar. Like being able to overload functions based on type is syntactic sugar. It is not required, but it makes life easier. You are wrong when you make the generalized statement that it makes code more complex. There are cases in which using this feature leads to clearer code and increased maintainability. A few of those cases are described in the drafts, but there are certainly more. In my opinion, being able to choose the best method for the job in these cases gives this feature all the mojo it needs. -Nod-
May 21 2005
Nod wrote:Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?I was enthusiastic about the idea, but I got lost in the second take. Too complicated for me to follow. (I'm just lazy).
May 21 2005
In article <d6obsn$mlc$1 digitaldaemon.com>, Hasan Aljudy says...Nod wrote:Argh. Maybe I was getting too wound up in the nitty-gritty details. Point taken. -Nod-Replies are scarce. How shall I interpret this silence? That the technique is perfect, and requires no further improvement? That the technique is useless, and not worth discussing? That no one really understands what the hell I'm on about?I was enthusiastic about the idea, but I got lost in the second take. Too complicated for me to follow. (I'm just lazy).
May 21 2005