The proposal is to extend template constraint from:

if(<expr>)
 to
if(<expr>; <message-expr>)

Where the <message-expr> is CTFE-able expression with type string.

The message-expr yields a hint string for compiler to show when instantiation failed instead of full expression that failed. It has to be CTFE-able even if template instantiation never actually fails.

The original full expression verbose behavior can be restored if desired with "verbose" compiler switch. "Verbose template instantiation failure switch" can be later be further enhanced to detect which sub-expressions are false for each candidate template.


Now how it should look like in messages.

Taking a motivating example from pull
https://github.com/D-Programming-Language/phobos/pull/1047
(though all of you have certainly seen far, far worse):

bug.d(11): Error: template std.range.chain does not match any function template declaration. Candidates are:
std/range.d(2018):        std.range.chain(Ranges...)(Ranges rs) if (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual, Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual, Ranges))) == void))
bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs) if (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual, Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual, Ranges))) == void)) cannot deduce template function from argument types !()(MapResult!(__lambda2, Foo[]),string)

First there is another sub-enhancement to remove constraints from the last message
Error: template XYZ <constraint *again* > cannot deduce ... from

as listing of candidates already shown them all.

Second and the actual point applied (message can be improved but in any case it should accurately & concisely describe the constraint):

std/range.d(2018):        std.range.chain(Ranges...)(Ranges rs) accepts: Any number of Input Ranges all having a common type among types of their elements.
bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs)
 cannot deduce template function from argument types !()(MapResult!(__lambda2, Foo[]),string)

where "accepts:" is compiler generated prefix and "Any number..." is a hint message.

What do you think?

-- 
Dmitry Olshansky

On Thursday, January 03, 2013 01:25:17 Dmitry Olshansky wrote:
> What do you think?

Hmmm. Typically the thing that you really want to know is which portions were true and which were false, and this doesn't help with that at all. What it does is make the errors more understandable for newbies, which is good. But for those of use who know more, they'd actually be annoying I think, because what I really want to know is what the actual constraint was so that I can figure out what's wrong. But I'm likely to look at the source for that anyway, since it's far more legible there than it is the error message.

So, this could be a good idea in that it could make the errors more understandable for newbies, but it still really doesn't help tell you what's wrong, which is what you really want to know. It just tries to translate the constraint into English. So, I don't know.

I'm not opposed to the idea, but I'm not exactly sold on it either. Certainly, I don't think that it would benefit me, personally, at all, but it may be worth it for newbies.

- Jonathan M Davis

On Wednesday, 2 January 2013 at 21:43:31 UTC, Jonathan M Davis wrote:
> On Thursday, January 03, 2013 01:25:17 Dmitry Olshansky wrote:
>> What do you think?
>
> Hmmm. Typically the thing that you really want to know is which portions were
> true and which were false, and this doesn't help with that at all. What it
> does is make the errors more understandable for newbies, which is good. But
> for those of use who know more, they'd actually be annoying I think, because
> what I really want to know is what the actual constraint was so that I can
> figure out what's wrong. But I'm likely to look at the source for that anyway,
> since it's far more legible there than it is the error message.
>
> So, this could be a good idea in that it could make the errors more
> understandable for newbies, but it still really doesn't help tell you what's
> wrong, which is what you really want to know. It just tries to translate the
> constraint into English. So, I don't know.
>
> I'm not opposed to the idea, but I'm not exactly sold on it either. Certainly,
> I don't think that it would benefit me, personally, at all, but it may be worth
> it for newbies.
>
> - Jonathan M Davis

Well, to be honest, some of our overloads are really *complicated*. I swear that for some of them, even understanding what the constraint are and why is a huge headache. Having some english in there can't hurt, IMO. That said, judicious use of traits (even private traits) can achieve the same goal, and reduce the complexity.

----
That said, I think another field of improvement we should concentrate in phobos is to have a single public entry point, and then privately forward to private specialized overloads.

Users really shouldn't have to deal with functions whose restraints are things such as "input range, but not random access", or "random access with length, but not sliceable" or some of our other *6* liners. It gets especially bad when the compiler lists all 7 of our (implementation defined) overloads. Nobody needs to see that.

As a end user, I'd really want to see a single function with "isForwardRange!R". Period. The fact that the implementation differs depending on all that *crap* really isn't my problem and shouldn't show up in my interface >:(

On Wednesday, January 02, 2013 23:02:12 monarch_dodra wrote:
> That said, I think another field of improvement we should concentrate in phobos is to have a single public entry point, and then privately forward to private specialized overloads.
> 
> Users really shouldn't have to deal with functions whose restraints are things such as "input range, but not random access", or "random access with length, but not sliceable" or some of our other *6* liners. It gets especially bad when the compiler lists all 7 of our (implementation defined) overloads. Nobody needs to see that.
> 
> As a end user, I'd really want to see a single function with "isForwardRange!R". Period. The fact that the implementation differs depending on all that *crap* really isn't my problem and shouldn't show up in my interface >:(

I think that this is sensible. I think that I haven't been doing it primarily because of how some of the functions are laid out in the source file (particularly with regards to the unit tests). Some of them will still need to be separate though for documentation purposes (or need version(StdDdoc) blocks which provide documentation if they're merged into a single function or template). find would be a prime example of this.

There's also the question of how to combine funcitons. There are 3 ways that I can think of off the top of my head:

1. Put them in a single function with many static ifs (not terribly clean if the functions were big enough that you split them apart in the first place).

2. Use a public function which calls a private impl function, which potentially creates unnecessary overhead (especially without -inline) but allows you to keep the separate functions more or less as they are now internally.

3. Use a template with the simplified template constraint on it with inner templated functions for each of the overloads.

I've generally though of doing #3 (as I think that we've done that in some cases already), but it forces you to place the unit tests farther from the functions, which I don't like. You seem to be suggesting #2, and that would probably be a better approach because of that, though it does bug me to create a useless outer function just for a template constraint.

- Jonathan M Davis

I'm reposting message from Don Clugston. Somehow it's not on the list and reply-to email address looks inaccessible.

On 02/01/13 22:25, Dmitry Olshansky wrote:
> The proposal is to extend template constraint from:
>
> if(<expr>)
>   to
> if(<expr>; <message-expr>)
>
> Where the <message-expr> is CTFE-able expression with type string.
>
> The message-expr yields a hint string for compiler to show when
> instantiation failed instead of full expression that failed. It has to
> be CTFE-able even if template instantiation never actually fails.
>
> The original full expression verbose behavior can be restored if desired
> with "verbose" compiler switch. "Verbose template instantiation failure
> switch" can be later be further enhanced to detect which sub-expressions
> are false for each candidate template.

I think we should just introduce an "else" or "default" constraint. For example:

void chain(Ranges...)(Ranges rs) default
{
  static assert("std.range.chain requires: Any number of Input Ranges all having a common type among types of their elements.");
}


ie, the default is used only if all template constraints fail.

/end of message

Which could be a more coherent way to solve it. The problem I see is that messages are disconnected from signatures of failed templates (that can change later and then message is easily out sync).

Either way what about just removing the constraints on the second error message here:

bug.d(11): Error: template std.range.chain does not match any function template declaration. Candidates are:
std/range.d(2018):        std.range.chain(Ranges...)(Ranges rs) if (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual, Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual, Ranges))) == void))

That was the first, it list candidates. It's relatively new thing, nice to have it.

bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs) if (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual, Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual, Ranges))) == void)) cannot deduce template function from argument types !()(MapResult!(__lambda2, Foo[]),string)

The second need not to realist constraints here, they are obvious from above, what needed is the last line - template arguments. So just make it:

bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs)  cannot deduce template function from argument types !()(MapResult!(__lambda2, Foo[]),string)

Bugzilla worthy?

-- 
Dmitry Olshansky

On Friday, January 04, 2013 18:16:44 Dmitry Olshansky wrote:
> I think we should just introduce an "else" or "default" constraint. For example:
> 
> void chain(Ranges...)(Ranges rs) default
> {
> static assert("std.range.chain requires: Any number of Input Ranges
> all having a common type among types of their elements.");
> }
> 
> 
> ie, the default is used only if all template constraints fail.
> 
> /end of message
> 
> Which could be a more coherent way to solve it. The problem I see is that messages are disconnected from signatures of failed templates (that can change later and then message is easily out sync).

Doing that wouldn't be all that different from using your suggestion with a single point of entry (e.g. using an outer template or wrapper function with a simplified template constraint with the separate functions having the more detailed constraints required to separate them). And with your suggestion, you could give each candidate an informative message if you wanted multiple points of entry (though I'd favor listing the message in addition to the constraint rather than instead of like you proposed).

> Either way what about just removing the constraints on the second error message here:
> 
> bug.d(11): Error: template std.range.chain does not match any function
> template declaration. Candidates are:
> std/range.d(2018): std.range.chain(Ranges...)(Ranges rs) if
> (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual,
> Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual,
> Ranges))) == void))
> 
> That was the first, it list candidates. It's relatively new thing, nice to have it.
> 
> bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs) if
> (Ranges.length > 0 && allSatisfy!(isInputRange, staticMap!(Unqual,
> Ranges)) && !is(CommonType!(staticMap!(ElementType, staticMap!(Unqual,
> Ranges))) == void)) cannot deduce template function from argument types
> !()(MapResult!(__lambda2, Foo[]),string)
> 
> The second need not to realist constraints here, they are obvious from above, what needed is the last line - template arguments. So just make it:
> 
> bug.d(11): Error: template std.range.chain(Ranges...)(Ranges rs) cannot
> deduce template function from argument types !()(MapResult!(__lambda2,
> Foo[]),string)
> 
> Bugzilla worthy?

Duplicate errors are definitely bugzilla worthy, and duplicated information within errors is bugzilla worthy - _especially_ in template error messages. We want them to be informative, but let's not make them longer than they need to be. It's the error messages that generally scare people away from templates.

- Jonathan M Davis