[cfe-commits] [Patch] Update Sema Diagnostics to use %diff

[Richard Smith]

On Thu, Jul 12, 2012 at 4:53 PM, Richard Trieu <rtrieu at google.com> wrote:

> On Thu, Jul 12, 2012 at 3:22 PM, Richard Smith <richard at metafoo.co.uk>wrote:
>
>> On Thu, Jul 12, 2012 at 2:07 PM, Richard Trieu <rtrieu at google.com> wrote:
>>
>>> Change diagnostic messages in DiagnosticsSemaKinds.td to use %diff when
>>> two QualType's are required.  This will enable template type diffing for
>>> all Sema messages.  The default text of the diff matches the old message so
>>> no test cases need to be updated.  Some tests have been added for the
>>> modified messages.
>>>
>>> Patch attached and also available at:
>>> http://llvm-reviews.chandlerc.com/D6
>>>
>>
>> This looks awesome! I have a few questions:
>>
>> 1) A few of the diagnostics contain multiple %diff{...}s, for instance,
>> err_init_conversion_failed. How well does that work in tree mode?
>>
> The current diagnostics can't print two trees.  For the
> err_init_conversion_failed, the first %diff prints if args 1 & 3 are
> templates.  The other %diff's only print if 1 & 3 are function pointer
> types.  Skimming the code, it appears that it does allow two trees to be
> printed, but problems may arise if only one tree is requested.  I think
> that printing the first possible tree, then ignoring any further trees
> should be the way to go.
>

Sounds good to me.


> 2) How should we handle diagnostics which refer only to built-in types?
>> (Examples: the warn_impcast_* and most of the err_typecheck_* diagnostics.)
>>
>> Is the diff valuable here? Are there cases where this will cause worse
>> diagnostics? More generally, how hesitant should we be before adding %diff
>> to a diagnostic, and are there any rules we should be following?
>>
> The diff will almost always fall back to standard printing for some of the
> diagnostics.  I applied the changes as broadly as possible in case anybody
> else wanted to improve the %diff for other cases.
>

OK, this seems reasonable, and I suppose a diff for
IdentityAliasTemplate<int> vs IdentityAliasTemplate<long> might be nice.


> I'm not overly fond of the "between types" wording in the tree case for
>> the warn_impcast_* set (it seems redundant to say that a cast is between
>> types, although I see that you need some way to introduce the type tree). I
>> wonder if something like this would read more naturally:
>>
>> 'implicit conversion changes value from %2 to %3; types are:
>> <tree>'
>>
> That is certainly possible to do, but also means that the diagnostic will
> be longer since the full types are printed.
>

In the above example, %2 and %3 there aren't the types in question, and I'm
only suggesting changing the tree form of the diagnostic. If my counting is
right, the tweak should make that diagnostic 1 character shorter.


> 3) How should we handle diagnostics are just naming two types which
>> weren't expected to be the same, rather than comparing them to each other?
>> (Examples: most of the cases where the diagnostic concerns an explicit
>> cast.)
>>
>> In such a case, I would expect the way that the types differ would be
>> uninteresting. My concern in such cases is mainly that adding a diff might
>> make the diagnostic less clear, in the case where the diff either performs
>> some desugaring or produces a tree.
>>
>> This also applies to err_ovl_ambiguous_oper_binary
>> and err_typecheck_invalid_operands. Those are a little more interesting,
>> because I suppose it will frequently be the case that the operands to a
>> binary operator were supposed to be the same. However, I think it could be
>> confusing to produce a type tree for these, since the diagnostic is
>> fundamentally trying to list two types, not point out the difference
>> between two types.
>>
> Like I said, I applied the diff to as many places as I could.  I don't
> mind changing the wording or removing %diff's if that makes things better.
>

I think it would be useful to have a general policy for this. For that, I
suggest: "%diff should be used when a diagnostic contains two types which
the user might reasonably have expected to be the same. It should not be
used when a diagnostic contains two types which were likely to have been
intentionally different." -- unless we have a motivating example for
wanting a diff in the latter case.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20120712/30760e83/attachment.html>