Bug 9289 – [Regression 2.061] Had been deprecated language features are enabled again in default

This regression is introduced by the "Make deprecations as warnings the default. From my comment: https://github.com/D-Programming-Language/dmd/pull/1287#issuecomment-12080557 --- I found a big breaking introduced by this, in 2.061. In current, all deprecated language features until 2.060 are accepted in default. For example: void main() { typedef int X = 10; // !! X x; assert(x == 10); } typedef was deprecated from 2.058, then we couldn't compile the code in default. But, from 2.061, we can compile the code and run in default again. More than worse, all deprecations which reported in semantic phase or later will break code meaning silently. import std.stdio; void main() { int[] arr; static if (is(typeof({ auto ptr = *arr; }))) writeln("*arr is not deprecated"); else writeln("*arr is deprecated, instead use arr.ptr"); } The meaning of is(typeof({ auto ptr = *arr; })) is changed in default. Therefore this is a huge breaking against the language improvement process in the past few years.

(In reply to comment #0) > This regression is introduced by the "Make deprecations as warnings the > default. > From my comment: > https://github.com/D-Programming-Language/dmd/pull/1287#issuecomment-12080557 [...] > Therefore this is a huge breaking against the language improvement process in > the past few years. How the default deprecation handling is changed compared to using -d in 2.060? I'm not entirely convinced is really a regression if is the same. I understand there is a problem, but I wonder if your suggested solution is really needed: > To fix the problem, we need to split deprecated language features into the two. > > Truly deprecated features > This group contains all language features which already deprecated in 2.060 and earlier. > Each of them prints deprecated message without -d switch. > Each of them raises an error without -d switch. (Important!!) > > Softly deprecated features > This group will print deprecation message for the using but not make an error in default, and may contain newly deprecated features from 2.061. > Each of them prints deprecated message without -d switch. > Each of them raises an error without -de switch. (Important!!) Shouldn't be just plain errors the things that you don't want to allow them at all? Or are you suggesting just a migration path to the new way of presenting deprecations?

(In reply to comment #1) I brushed up my idea: The -dw/-de switches affect to the use of deprecated "symbols", but doesn't to the use of deprecated language features. You can see my implementation. https://github.com/D-Programming-Language/dmd/pull/1459

> I brushed up my idea: The -dw/-de switches affect to the use of deprecated > "symbols", but doesn't to the use of deprecated language features. That's probably the right way to handle it. Deprecated language features already effectively have the separation between soft and hard deprecation, because they generally get normal warnings before being deprecated, whereas deprecated symbols aren't part of that, since they aren't hard-coded into the compiler. We can just go back to treating deprecated language features the way that we have and take advantage of the recent changes to deprecated for deprecated symbols.

I still don't understand why is it a good idea to treat deprecated language features differently from deprecated symbols. Deprecated language features get a warning but you only get a warning if you use -w, which implies having warnings for lots of other stuff that have nothing to do with deprecations.

Well I see the second example from Kenji doesn't print any warnings, that's certainly a problem. But the first example works just fine, it compiles issuing a warning. That is what it's supposed to do, that was the original idea. Is not a bug, is a feature :)

I'm not sold on the idea of adding additional complexity to the behavior of the deprecation-related switches. If you feel that some language features should be more than just deprecated by now, why don't chose the obvious option of just turning them into an error?

(In reply to comment #6) > I'm not sold on the idea of adding additional complexity to the behavior of the > deprecation-related switches. > > If you feel that some language features should be more than just deprecated by > now, why don't chose the obvious option of just turning them into an error? I agree, I think we are overengineering this.

Honestly, I think that it would be great in general to be able to distinguish between soft deprecation (warn about it) and hard deprecation (make it an error), but I can also understand not wanting to add that sort of complication to the feature, especially since it's a feature that should be needed relatively rarely. And if you only get one of the two, soft deprecations are far less disruptive, and I think that we were right to make deprecated warn by default, but it does pose a bit of a problem for the language features. For that, I think that there are two basic approaches which make sense: 1. Make deprecated work as it has been for language features and use normal warnings when we want to "soft" deprecate a language feature, which is exactly what we have been doing. But we keep the new behavior (warn by default) for the deprecated keyword. 2. Make at least some of the currently deprecated language features into errors rather than have them be deprecated. Then have deprecation for language features be exactly the same as it is for the deprecated keyword, so some features end up being errors which can't be disabled, and some become warnings rather than errors by default. If we go that route, we should probably also think about making some of the current warnings into deprecation warnings (e.g. when skipping the use of override). The second approach would probably be the cleaner of the two, but it does mean that we lose the ability to give an error for language deprecation by default and still allow for the feature to be used. On the other hand, using deprecation warnings instead of normal warnings for deprecating features would be great, because then people don't have to use -w or -wi to see them. And that would make doing stuff like deprecating delete much smoother, since it would ensure that anyone using it got bugged about it rather than just those folks who compile with warnings enabled. It also keeps its behavior in line with the deprecated keyword, which would probably be a good idea if we can reasonably do so.

I don't see why language features should be treated differently. The only problem I see right now is deprecated features are not being warn when gagging (I think that's the term for when errors are silent to evaluate something at compile time that is OK to fail, like tratis(compile) or static if's). I think usage of deprecated features there should trigger a warning too. If the idea behind some of this construct is just to test if an old feature is still working, maybe that should be fixed and "calculated" through the compiler's version or something like that. If we keep deprecated language features as errors by default, we will end up with the exact same problem we had before making deprecations as warnings the default. Why would we want to do that?

(In reply to comment #9) > I don't see why language features should be treated differently. The only > problem I see right now is deprecated features are not being warn when gagging > (I think that's the term for when errors are silent to evaluate something at > compile time that is OK to fail, like tratis(compile) or static if's). I think > usage of deprecated features there should trigger a warning too. If the idea > behind some of this construct is just to test if an old feature is still > working, maybe that should be fixed and "calculated" through the compiler's > version or something like that. > > If we keep deprecated language features as errors by default, we will end up > with the exact same problem we had before making deprecations as warnings the > default. Why would we want to do that? OK. I was convinced that it is the right thing to some extent. To make things keep simple, deprecations between language feature and user-defined symbols should be treated by a same way. I withdraw this regression, by lowering priority to "major". > The only problem I see right now is deprecated features are not being warn when gagging I'd try to implement it in experiment.

(In reply to comment #10) > (In reply to comment #9) > > If we keep deprecated language features as errors by default, we will end up > > with the exact same problem we had before making deprecations as warnings the > > default. Why would we want to do that? > > OK. I was convinced that it is the right thing to some extent. > To make things keep simple, deprecations between language feature and > user-defined symbols should be treated by a same way. > I withdraw this regression, by lowering priority to "major". I changed the title too, feel free to edit if you think is not entirely correct. > > The only problem I see right now is deprecated features are not being warn when gagging > > I'd try to implement it in experiment. Great! Thanks. Unless I'm missing something (I probably am :), this should be enough: --- diff --git a/src/mars.c b/src/mars.c index 612b205..588ca40 100644 --- a/src/mars.c +++ b/src/mars.c @@ -269,7 +269,7 @@ void vdeprecation(Loc loc, const char *format, va_list ap, static const char *header = "Deprecation: "; if (global.params.useDeprecated == 0) verror(loc, format, ap, p1, p2, header); - else if (global.params.useDeprecated == 2 && !global.gag) + else if (global.params.useDeprecated == 2) verrorPrint(loc, header, format, ap, p1, p2); } --- I don't have time now to make testcases and verify this would work though.

(In reply to comment #11) > I don't have time now to make testcases and verify this would work though. Woops! I made some time :D

Comment from don to the pull request: > Although this idea is appealing in some circumstances (specifically, > when you have just upgraded from one compiler version to another), in > others it is NOT what you want, and it breaks existing code. Using > your test case, suppose you have the code: > > --- > auto foo(P)(P p){ > static if ( is(typeof( *p)) { > return *p; > } else static if (is(typeof(p[0])) { > return p[0]; > } else static assert(0, "unsupported"); > } > --- > > Currently, if -d is not specified, passing an array to this will work > correctly: the second return will be used. If -d is specified, it will > also work correctly: the first return will be used. With your patch, > the code will not work at all. If I have a codebase where I never > even use -d, this pull request is suddenly making that obscure, > stupid, long-deprecated feature have an impact on my code! > > I think we just have to accept that having deprecated features still > accepted in the language is a major annoyance. I'm not sure that there > is a perfect solution. We should really try to get rid of them. The first one will work correctly *BUT* you'll get the warning, so you can inspect the code and fix it. I think this solution is not worse than the current situation where what previously had fail, now passes silently. If you are using deprecated features in your code, your code is broken (or at least outdated) and you should fix it (or update it). Now those obscure changes on behaviour are there, and you'll never find out. This pull request gives visibility to this issues and allows you to update your code. I really don't see the problem with this solution. I don't think is a big annoyance. The default behaviour is to warn you when you are using something that's deprecated so you can update it.

We should probably seriously consider making some of the currently deprecated features an outright error rather than leaving them as deprecated (especially if they've been deprecated for quite a while). That would reduce any problems caused by deprecated features suddenly compiling with warning messages by default.

(In reply to comment #14) > If you are using deprecated features in your code, your code is broken (or at > least outdated) and you should fix it (or update it). Yes, but this change would affect code which is NOT using deprecated features. You have completely misunderstood my point. > Now those obscure changes > on behaviour are there, and you'll never find out. This pull request gives > visibility to this issues and allows you to update your code. No, it has the reverse behaviour. You would need to update your code to specifically recognize deprecated features! The existing behaviour is that, if you are not using -d, deprecated language features act as if they do not exist. This change would make deprecated features have an effect on ALL code. Even when not compiling with -d. I am completely opposed to this change. It creates a special case which introduces new problems that are at least as bad as the old ones.

(In reply to comment #16) > This change would make deprecated features have an effect on ALL code. Even > when not compiling with -d. That change is already in Git master. I'm pretty convinced that the best way of fixing any related issues is just to finally remove any deprecated language features. > I am completely opposed to this change. It creates a special case which > introduces new problems that are at least as bad as the old ones. What exactly do you regard as the special case here?

(In reply to comment #17) > (In reply to comment #16) > > This change would make deprecated features have an effect on ALL code. Even > > when not compiling with -d. > > That change is already in Git master. I'm pretty convinced that the best way of > fixing any related issues is just to finally remove any deprecated language > features. Not only master, is in the last release. > > I am completely opposed to this change. It creates a special case which > > introduces new problems that are at least as bad as the old ones. > > What exactly do you regard as the special case here? Yeah, I think it was pretty clear that making deprecated features as warnings was, at some point, a breaking change, unless you use -de to achieve the old behaviour. I still have to review Don's example, because evidently I missed his point.

Currently, is(typeof( xxxx )) NEVER, EVER produces an error message, no matter what xxxx is. You can put any kind of nonsense operation in there, and it will still compile, and return false. This change would create a horrific special case: is(typeof( xxxx )) never produces an error message, no matter what xxxx is, unless xxxx happened to be legal in an ancient version of DMD but is not legal in current DMD. This is a killer for generic code.

I agree with David on this one. If there's stuff here that really shouldn't be compiling, then lets finally remove it. Code shouldn't need to be checking whether deprecated features work or not. It shouldn't be trying to use them in the first place. And many of the currently deprecated language features have been deprecated for some time ande really should be outright removed. In general, warning for deprecations just works a lot better, so I'd hate to lose that.

BTW, Don, I see what you mean now, but that problem existed before making deprecations emit warnings by default (if you used -d). The default was only changed. I know that introduces some kind of regression in a sense because code will behave differently unless you use -de, but I think this is only a transitional problem, as again, only the default changed. If some deprecated feature gets too problematic, I agree with David that the best solution might be to just remove them, specially if they have been deprecated long ago...

(In reply to comment #21) > BTW, Don, I see what you mean now, but that problem existed before making > deprecations emit warnings by default (if you used -d). The default was only > changed. It is true that there's always been a difference in is(typeof()) between when compiled with -d, and when not. However, historically that hasn't mattered so much -- it has been crazy to compile with -d. This bug report gives the -d behaviour a higher priority than the no-flags behaviour. You're assuming that when a deprecated feature appears while errors are gagged, it's because it used to compile. That assumption is false. Such code might, or might not be a use of a deprecated feature. There is no way to tell. The patch moves it from "assume it's never a deprecated feature" to "assume it's always a deprecated feature". Effectively, you're changing an accepts-invalid bug to a rejects-valid bug. It's not an improvement. Here's what I think should happen: -dXX should give a warning (not an error, not silence), whenever a deprecated feature is used when errors are gagged. In other cases, the existing behaviour (silence) should be preserved. Basically I think that with any form of -d option, deprecated features should never compile without generating a warning (even when errors are gagged). When -d is not used, it should be as if those features don't exist -- there should be no difference between something that is deprecated, vs any other sort of error.

(In reply to comment #22) > (In reply to comment #21) > > BTW, Don, I see what you mean now, but that problem existed before making > > deprecations emit warnings by default (if you used -d). The default was only > > changed. > > It is true that there's always been a difference in is(typeof()) between when > compiled with -d, and when not. However, historically that hasn't mattered so > much -- it has been crazy to compile with -d. Agreed. > This bug report gives the -d behaviour a higher priority than > the no-flags behaviour. You're assuming that when a deprecated feature appears > while errors are gagged, it's because it used to compile. That assumption is > false. Yes, as I said I understand what you mean and agree too. I'll roll back the title to the original one to avoid spreading this confusion. > Such code might, or might not be a use of a deprecated feature. There is no way > to tell. The patch moves it from "assume it's never a deprecated feature" to > "assume it's always a deprecated feature". > Effectively, you're changing an accepts-invalid bug to a rejects-valid bug. > It's not an improvement. > > Here's what I think should happen: > > -dXX should give a warning (not an error, not silence), whenever a deprecated > feature is used when errors are gagged. > In other cases, the existing behaviour (silence) should be preserved. > > Basically I think that with any form of -d option, deprecated features should > never compile without generating a warning (even when errors are gagged). When > -d is not used, it should be as if those features don't exist -- there should > be no difference between something that is deprecated, vs any other sort of > error.

Woops! I wanted to commit only the title change... (In reply to comment #22) > Such code might, or might not be a use of a deprecated feature. There is no way > to tell. The patch moves it from "assume it's never a deprecated feature" to > "assume it's always a deprecated feature". > Effectively, you're changing an accepts-invalid bug to a rejects-valid bug. > It's not an improvement. Again, agreed. > Here's what I think should happen: > > -dXX should give a warning (not an error, not silence), whenever a deprecated > feature is used when errors are gagged. You're talking about a new flag I guess... > In other cases, the existing behaviour (silence) should be preserved. Only when gagged I assume. > Basically I think that with any form of -d option, deprecated features should > never compile without generating a warning (even when errors are gagged). Except for plain -d I guess. That's the original meaning of -d, silence everything. > When -d is not used, it should be as if those features don't exist -- there should > be no difference between something that is deprecated, vs any other sort of > error. Only when errors are gagged I guess. Otherwise is going back to where we started.

(In reply to comment #24) > Woops! I wanted to commit only the title change... > > (In reply to comment #22) > > Such code might, or might not be a use of a deprecated feature. There is no way > > to tell. The patch moves it from "assume it's never a deprecated feature" to > > "assume it's always a deprecated feature". > > Effectively, you're changing an accepts-invalid bug to a rejects-valid bug. > > It's not an improvement. > > Again, agreed. > > > Here's what I think should happen: > > > > -dXX should give a warning (not an error, not silence), whenever a deprecated > > feature is used when errors are gagged. > > You're talking about a new flag I guess... Not necessarily. It could be done with -dw for example. > > In other cases, the existing behaviour (silence) should be preserved. > > Only when gagged I assume. Yes, otherwise it's an error. > > > Basically I think that with any form of -d option, deprecated features should > > never compile without generating a warning (even when errors are gagged). > > Except for plain -d I guess. That's the original meaning of -d, silence > everything. Possibly. We should just kill that, it's ridiculous. > > When -d is not used, it should be as if those features don't exist -- there should > > be no difference between something that is deprecated, vs any other sort of > > error. > > Only when errors are gagged I guess. Otherwise is going back to where we > started. When errors aren't gagged, using a deprecated feature is just an error. (The wording may mention that it's deprecated, but otherwise it's a normal error). When errors are gagged, it's again just an error, but nothing is printed. When compiled with -dw, whenever anything deprecated is encountered, display a warning message. Display this even if errors are gagged. ie, anything deprecated is either a normal error, or noisy accepted, and this is not affected by gagging.

(In reply to comment #25) > > > -dXX should give a warning (not an error, not silence), whenever a deprecated > > > feature is used when errors are gagged. > > > > You're talking about a new flag I guess... > > Not necessarily. It could be done with -dw for example. But then make -dw mean something different than <no flag at all>, right? Just trying to understand your suggestion. > > > In other cases, the existing behaviour (silence) should be preserved. > > > > Only when gagged I assume. > > Yes, otherwise it's an error. Mmm, an error? Using -de I guess, if that's the default this is like reverting the "make deprecations issue warnings by default". > > > Basically I think that with any form of -d option, deprecated features should > > > never compile without generating a warning (even when errors are gagged). > > > > Except for plain -d I guess. That's the original meaning of -d, silence > > everything. > > Possibly. We should just kill that, it's ridiculous. Maybe. > > Only when errors are gagged I guess. Otherwise is going back to where we > > started. > > When errors aren't gagged, using a deprecated feature is just an error. (The > wording may mention that it's deprecated, but otherwise it's a normal error). > When errors are gagged, it's again just an error, but nothing is printed. Why should it be an error to use a deprecated feature? I still don't understand why is this desirable when it was the whole point of making "deprecation as warnings by default". With this, you upgrade to a new DMD version and your code gets broken if you happened to be using a recently deprecated feature instead of just getting a deprecation message. Deprecating something means "it will be removed in the future" not "it already have been removed". And if you used some deprecated feature where errors are gagged, you're application will change its behaviour without any warnings. > When compiled with -dw, whenever anything deprecated is encountered, display a > warning message. Display this even if errors are gagged. > > ie, anything deprecated is either a normal error, or noisy accepted, and this > is not affected by gagging. I think this really makes the feature useless. If you don't use -dw, your code won't compile. If you use it, you'll potentially get a lot of spurious warnings about gagged code that is using some feature that used to have a different meaning a while ago. I think that's far from ideal... If we leave things are they are now, the only problem I see is you don't know if some piece of error-gagged code uses a deprecated feature or not, and you'll only find out when the feature is completely removed, if you are lucky enough to notice the program changed its behaviour (if we keep deprecated features as errors, this will happen as soon as the feature gets deprecated instead of removed). Maybe we can add a flag (-dv for verbose maybe?) to print warnings for deprecated errors even when errors are gagged.

This issue is no longer valid. I cannot reproduce any of the reported failings.