!standard 8.6(9) 12-11-29 AI12-0040-1/01 !class binding interpretation 12-11-29 !status work item 12-11-29 !status received 12-11-28 !priority Low !difficulty Easy !subject Resolving the selecting_expression of a case_expression !summary The selecting_expression of a case_expression resolves in the same way as the selecting_expression of a case_statement; in particular, it is a complete context. !question The selecting_expression of a case_statement is a complete context by 8.6(9); the selecting_expression of a case_expression is not. In unusual cases, this can make case_expressions legal where the corresponding case_statement would be illegal. Consider: procedure Complete_Contexts is type E1 is (Aa, Bb, Cc); type E2 is (Bb, Cc, Dd); function F return E1 is begin return E1'First; end F; function F return E2 is begin return E2'First; end F; N : Natural; begin N := (case F is -- Legal? (No.) when Aa => 123, when Bb => 456, when Cc => 789); case F is -- Illegal. when Aa => null; when Bb => null; when Cc => null; end case; end; Should this be fixed? (Yes.) !recommendation (See summary.) !wording Modify 8.6(9): The {*selecting_*}expression of a case_statement {or case_expression}. !discussion We want case statements and case expressions resolve the same way; consistency is important. It's a bit odd to have a complete context in the middle of an expression, but as the wording of 8.6(4) makes it clear that complete contexts can be nested, there is no semantic problem. Moreover, other constructs like the operand of a type conversion and conditions of an if expression act like complete contexts (their resolution has no effect on whether the rest of the expression can be resolved). So this isn't really new. A selecting_expression has to be a complete context in order to prevent the case choices from affecting the resolution. In cases that aren't complete contexts (like the target of an assignment), all of the possibilities are considered together (thus the source expression of an assignment can help the resolutiuon of the target). !ACATS test An ACATS B-Test like the example in the question should be created to verify that this is detected. It's low priority, though, as it is hard to imagine that this would be done wrong. !appendix From: Steve Baird Sent: Wednesday, November 28, 2012 5:43 PM We've got (in 8.6): Each of the following constructs is a complete context: ... The expression of a case_statement. Ramification: This means that the expression is resolved without looking at the choices. Shouldn't the expression of a case_expression also be on this list? I believe the following example was intended to be illegal, but is legal because of this omission: procedure Complete_Contexts is type E1 is (Aa, Bb, Cc); type E2 is (Bb, Cc, Dd); function F return E1 is begin return E1'First; end F; function F return E2 is begin return E2'First; end F; N : Natural; begin N := (case F is when Aa => 123, when Bb => 456, when Cc => 789); -- case F is -- when Aa => null; -- when Bb => null; -- when Cc => null; -- end case; end; Note that the commented out case statement, if uncommented, is certainly illegal. I believe that we want case statements and case expressions to be treated consistently in this respect. What do folk think about replacing (in 8.6) "The expression of a case_statement." with "The expression of a case_statement or case_expression." ? **************************************************************** From: Edmond Schonberg Sent: Wednesday, November 28, 2012 9:17 PM It's certainly simple to state and consistent with case statements, but it find it a bit bizarre to say that a portion of an expression is a complete context. After all the case expression appears within some larger context that imposes a type on it, and maybe this larger context is used to resolve the expression. This would be consistent with overload resolution everywhere else. But your proposed rule is simple to present and trivial to implement, obviously a big plus. **************************************************************** From: Randy Brukardt Sent: Wednesday, November 28, 2012 9:49 PM The selecting_expression of a case statement (like a condition of an if expression) doesn't contribute in any way to the final type of the expression. It is "used up". Its resolution and legality is separate from any other part of the expression. I'll call such a (sub)expression an "expression sink" for the lack of a better term. The point here is that a selecting_expression is an "expression sink" without a defined type. We don't want to be using information from elsewhere (specifically from the case choices) to resolve this expression. Most other "expression sinks" (like the condition of an if statement or expression) have a defined type, so they don't need to be complete contexts. I agree that it is weird to consider a selecting_expression of a case_expression a complete context, but since it doesn't have any effect on the resolution of the rest of the expression, I don't quite know what else it could be. BTW, we gave this thing a name ("selecting_expression"), so it would be better to use that name in the wording above: "The selecting_expression of a case_statement or case_expression." (thus preventing any confusion with the choice_expressions). **************************************************************** From: Tucker Taft Sent: Wednesday, November 28, 2012 11:31 PM I agree, we should make the selecting expression a complete context. Note that the operand of a type conversion is also effectively a complete context. This is essentially equivalent to saying the expected type is "any type." **************************************************************** From: Randy Brukardt Sent: Thursday, November 29, 2012 12:20 AM Humm. The expression of a type conversion is *not* defined to be a complete context, but it has an expected type of "any type". If that's good enough for type conversions to make them "effectively complete contexts" (and I can't find any specific rules for type conversions), then there is nothing that we need to do for selecting_expressions of case expressions. That's because they're defined to resolve to "any discrete type" (5.4(4/3)). Steve is arguing that the types of the choice expressions could be used to disambiguate a selecting_expression. But I'm not sure why we would expect that to work differently than the "any type" of a type conversion - you don't get to use the target type to disambiguate the operand expression of a type conversion. So I conclude that there is no problem here - at least until someone points out some explicit language about type conversions not using the target type. (In which case we need similar language for *all* selecting_expressions.) BTW, I'm not sure why we need special language to make the selecting_expression of a case statement a "complete context", while we don't have such language for the condition of an if_statement -- these both are data and resolution "sinks". If the only point is to enforce the separation of the selecting_expression from the choice_expressions, that probably should be explicit in the resolution rules, rather than implicit in something irrelevant like "complete context". (Even if there is no problem, we still need a ramification AI to point that out.) **************************************************************** From: Steve Baird Sent: Thursday, November 29, 2012 11:23 AM > Humm. The expression of a type conversion is *not* defined to be a > complete context, but it has an expected type of "any type". If that's > good enough for type conversions to make them "effectively complete > contexts" (and I can't find any specific rules for type conversions), > then there is nothing that we need to do for selecting_expressions of case > expressions. Randy - You make a good argument that there is no need to define the selecting_expressions of a case_expression to be a complete context. If that argument is valid (and it sounds right to me), then doesn't it also apply to case statements? You make the same point: > BTW, I'm not sure why we need special language to make the > selecting_expression of a case statement a "complete context", while > we don't have such language for the condition of an if_statement -- > these both are data and resolution "sinks". If the only point is to > enforce the separation of the selecting_expression from the > choice_expressions, that probably should be explicit in the resolution > rules, rather than implicit in something irrelevant like "complete context". I understand the "if it ain't broke" argument (and the "insufficiently broken" argument), but I still think that case_statements and case_expressions should be treated consistently - selecting_expressions of either both or neither should be defined to be complete contexts. It sounds like you are arguing for "neither", which sounds good to me. So there *is* a need for a wording change. **************************************************************** From: Bob Duff Sent: Thursday, November 29, 2012 12:00 PM > So there *is* a need for a wording change. I agree with you and Randy that there's no need to say "complete context" in either case, because there is an expected type defined (and that expected type is not defined in terms of the when's). But this falls into the category of language changes that have no effect: No compiler writer or programmer will change their behavior. I'm opposed to making such purely-cosmetic changes. **************************************************************** From: Steve Baird Sent: Thursday, November 29, 2012 12:24 PM > But this falls into the category of language changes that have no > effect: No compiler writer or programmer will change their behavior. > I'm opposed to making such purely-cosmetic changes. You are right about the impact on compilers, but the current RM wording is misleading (it suggests a difference in the treatment of the two "case" constructs when there is none) and the AARM note 8.6(9.a) is simply wrong. The real problem is that the current wording is a source of confusion; it certainly confused me. Given that all of this can be cleaned up with one simple deletion (delete 8.6(9) and its accompanying AARM note), I think this is worth doing. On the other hand, I would agree that this is one is a close call precisely because of the issues you raise. **************************************************************** From: Randy Brukardt Sent: Thursday, November 29, 2012 1:21 PM > > So there *is* a need for a wording change. > > I agree with you and Randy that there's no need to say "complete > context" in either case, because there is an expected type defined > (and that expected type is not defined in terms of the when's). I'm not so sure I agree with myself this morning. :-) The problem is that the expected type of the choices are defined in terms of the type of the selecting_expression. In other cases in the language (such as assignments), we interpret that to mean that all of these things are resolved together. Here, we don't want that to happen. Compare the wording for assignments [5.2(4/2)]: The variable_name of an assignment_statement is expected to be of any type. The expected type for the expression is the type of the target. to the wording for case statements (and case expressions) [5.4(4/3]: The selecting_expression is expected to be of any discrete type. The expected type for each discrete_choice is the type of the selecting_expression. These are almost identical, but we want different things to happen. That won't work! I was mislead by the lack of a similar rule for type conversions. But that's not needed, because the rules on the target type are all Legality Rules, and thus aren't considered during resolution. So I now think Steve's original suggestion was correct, and we *do* need a wording change in 8.6(9): The {selecting_}expression of a case_statement {or case_expression}. [Might as well use the new prefix in this wording.] Note to Ed: It is expected that complete contexts can be nested, as 8.6(4) mentions "not counting inner complete contexts". So there isn't any real problem with part of an expression being a complete context. Note to self: resolution "sinks" like conditions and type conversions probably ought to have been "complete contexts", but as no resolution rules depend on their resolved type (hopefully), it's not required to get the correct effect. And probably the 9x team decided to keep this as simple as possible, leaving out such sinks. Note to Bob: I agree that no compiler is going to implement this wrong (it would be a lot of extra work), but some pedant like Adam will ask the question if we don't handle it now. That's the task of the ARG, after all, answering boring questions about unimportant holes in the Standard. The fun stuff like designing new kinds of contracts is more of the ARG's hobby (especially now, between revisions). **************************************************************** From: Bob Duff Sent: Thursday, November 29, 2012 1:40 PM > The problem is that the expected type of the choices are defined in > terms of the type of the selecting_expression. In other cases in the > language (such as assignments), we interpret that to mean that all of > these things are resolved together. Here, we don't want that to happen. Good point. > Note to Bob: I agree that no compiler is going to implement this wrong > (it would be a lot of extra work), but some pedant like Adam will ask > the question if we don't handle it now. That's the task of the ARG, > after all, answering boring questions about unimportant holes in the > Standard. The fun stuff like designing new kinds of contracts is more > of the ARG's hobby (especially now, between revisions). I think we should be more willing to vote "No Action" on the kinds of things Adam comes up with. If he's honestly confused, as to nitpicking the wording, we one of us can tell him on ada-comment@ that case exps work like case stms, or whatever. **************************************************************** From: Tucker Taft Sent: Thursday, November 29, 2012 2:33 PM I disagree with Tuck version 1, and agree with Randy version 2 and Steve version 1. We *should* say "complete context" to prevent it from factoring in the name resolution rules that apply to the various choice expressions. And I agree with Steve versions 1 and 2 that these two should be consistent. And I agree with Bob (all versions) that this is very low priority. But if we do some fixups, might as well fix this one up. I trust that was crystal clear... **************************************************************** From: Bob Duff Sent: Thursday, November 29, 2012 2:54 PM > I trust that was crystal clear... I think this discussion illustrates the point that when we meddle with RM wording, we're about as likely to break things as to fix things -- we almost decided to break case statements rather than fixing case expressions, until Randy came to the rescue. So we should, in general be reluctant to change things. I don't feel strongly about whether this one should be fixed now that we all (think we) understand it. Just keep in mind that if we keep fiddling with the RM forever, we will never reach perfection, nor even asymptotically approach it. But we can spend a lot of energy discussing it. ;-) **************************************************************** From: Bob Duff Sent: Thursday, November 29, 2012 3:16 PM ... > I think this discussion illustrates the point that when we meddle with > RM wording, we're about as likely to break things as to fix things -- > we almost decided to break case statements rather than fixing case > expressions, until Randy came to the rescue. So we should, in general > be reluctant to change things. ... > Just keep in mind that if we keep fiddling with the RM forever, we > will never reach perfection, nor even asymptotically approach it. But > we can spend a lot of energy discussing it. ;-) I think you once pointed out that the RM doesn't have a regression test suite. Most of us are used to fiddling with our compilers in an attempt to reach perfection. But we have the ACATS (and usually in-house tests) to keep us on course if we screw up. (And in my experience, I end up breaking something fully a quarter of the time I "fix" our compiler.) We don't have that sort of backup for the RM. (Other than for basic editing issues; Pascal built a tool years ago to compare the wording in the Amendment document with the wording in the RM; that found a lot of glitches in both documents, a few which were significant. Ada 2020 probably won't have that, as I doubt that we'll do an Amendment-style document [to be discussed].) OTOH, a lot of these changes are straightforward. And quite a few of them indicate real language problems (for instance, the ability to use a generic unit to convert from a tagged type to an extension of it -- without having to mess with those pesky extension components -- that's the Adam AI I wrote up last night) -- even if those are obscure, they're often the sort of holes that we can't tolerate). Cases like the one at hand (where no one would do it wrong whether or not the RM is changed) are in fact rather rare. And remember the lesson of the Sofcheck employee who dutifully implemented exactly what the Standard said, no matter how nonsensical. [That also happened to me, with the Ada 95 alignment rules, which I implemented exactly as written in the Standard, although I later found out that no one intended them to work that way. I would have liked to find that out a lot sooner!] Takeaway: we should fix as much as we can, but we must never stop being vigilant. But it probably isn't worth it to fix ticky-tack wording that is slightly confusing to someone - just stick in an AARM note and go. ****************************************************************