Deprecation is a nice feature.  There doesn't seem to be any doubt about it.

But it's by no means perfect, compiler bugs aside.  I have identified these ideals of deprecation:

(a) Code that compiles without -d compiles and behaves the same with -d.

(b) Code that compiles without -d compiles and behaves the same if all entities declared as deprecated are removed from the code.

(c) Code that compiles with or without -d compiles and behaves the same if all deprecated entities are de-deprecated.

(d) Code that compiles without -d either compiles and behaves the same, or fails to compile at all, if any entities are newly deprecated (i.e. does not silently change its behaviour).

(e) Validity checking through compile-time reflection is always consistent with whether the compiler actually accepts the code.

(OK, so the switch used to enable deprecated entities might vary between brands of compiler, but I've used "-d" here for simplicity.)


But these ideals cannot all be satisfied at once.

The main reason for this seems to be compile-time reflection.  My experiment (under DMD 1.071 and 2.057) shows that an IsExpression evaluates to false if the compiler would reject it because of deprecation.  This satisfies ideals (b) and (e), but violates (a), (c) and (d).

If we changed it to evaluate to true even if compiling without -d, then it would satisfy (a), (c) and (d), but violate (b) and (e).  Moreover, when/if we get -di
http://d.puremagic.com/issues/show_bug.cgi?id=7041
what would it do if we have deprecation-dependent behaviour?

But deprecated entities aren't always treated as though they're not there if compiling without -d.  For example, deprecated functions take part in overload resolution.  This is necessary in order to satisfy (a), (c) and (d).  Though it doesn't seem to make any difference to (b) or (e).


Maybe DDB is useful at times.  The use case that comes to mind is suppressing the handling of a deprecated property as part of a non-deprecated method.  But still, is accidentally programming DDB in something that needs to be guarded against?  Or do we just accept DDB as a natural, unpreventable consequence of reflection?

I'm not sure whether all ideals would be satisfied in the absence of CTR, or if there are other aspects of D that prevent it from being so.

I suppose the overall point is: Is the current compiler behaviour the best that can be done?  Which ideals of deprecation do we really need to follow, and which can we do without?


Stewart.

On 02/12/2012 03:21 AM, Stewart Gordon wrote:
> Deprecation is a nice feature. There doesn't seem to be any doubt about it.
>
> But it's by no means perfect, compiler bugs aside. I have identified
> these ideals of deprecation:
>
> (a) Code that compiles without -d compiles and behaves the same with -d.
>
> (b) Code that compiles without -d compiles and behaves the same if all
> entities declared as deprecated are removed from the code.
>
> (c) Code that compiles with or without -d compiles and behaves the same
> if all deprecated entities are de-deprecated.
>
> (d) Code that compiles without -d either compiles and behaves the same,
> or fails to compile at all, if any entities are newly deprecated (i.e.
> does not silently change its behaviour).
>
> (e) Validity checking through compile-time reflection is always
> consistent with whether the compiler actually accepts the code.
>
> (OK, so the switch used to enable deprecated entities might vary between
> brands of compiler, but I've used "-d" here for simplicity.)
>
>
> But these ideals cannot all be satisfied at once.
>
> The main reason for this seems to be compile-time reflection. My
> experiment (under DMD 1.071 and 2.057) shows that an IsExpression
> evaluates to false if the compiler would reject it because of
> deprecation. This satisfies ideals (b) and (e), but violates (a), (c)
> and (d).
>
> If we changed it to evaluate to true even if compiling without -d, then
> it would satisfy (a), (c) and (d), but violate (b) and (e). Moreover,
> when/if we get -di
> http://d.puremagic.com/issues/show_bug.cgi?id=7041
> what would it do if we have deprecation-dependent behaviour?
>
> But deprecated entities aren't always treated as though they're not
> there if compiling without -d. For example, deprecated functions take
> part in overload resolution. This is necessary in order to satisfy (a),
> (c) and (d). Though it doesn't seem to make any difference to (b) or (e).
>
>
> Maybe DDB is useful at times. The use case that comes to mind is
> suppressing the handling of a deprecated property as part of a
> non-deprecated method. But still, is accidentally programming DDB in
> something that needs to be guarded against? Or do we just accept DDB as
> a natural, unpreventable consequence of reflection?
>
> I'm not sure whether all ideals would be satisfied in the absence of
> CTR, or if there are other aspects of D that prevent it from being so.
>
> I suppose the overall point is: Is the current compiler behaviour the
> best that can be done? Which ideals of deprecation do we really need to
> follow, and which can we do without?
>
>
> Stewart.

I think (e) is not an ideal of deprecation. If a feature is deprecated, then checking whether it compiles should be deprecated too.

On 12/02/2012 14:44, Timon Gehr wrote:
<snip excessive quote>
> I think (e) is not an ideal of deprecation. If a feature is deprecated, then checking
> whether it compiles should be deprecated too.

Oh yes, that's another possibility.  Make an IsExpression on something deprecated emit a deprecation error instead of returning either true or false.  This satisfies (a), (b), (c) and (d).

And it satisfies a variant of (e): validity checking through compile-time reflection is always either consistent with whether the compiler accepts the code or an error in itself.

This already covers the cases where the body of the IsExpression isn't syntactically valid.  Though there's a difference there in that the error happens at the parsing stage, rather than the semantic analysis stage.  But there's also a difference here in that syntax doesn't depend on code elsewhere in the source file or in other source files, and so there isn't much of a use case for syntactic validity checking as a form of reflection.

Stewart.