Re: [RFC] c++: Print function template parms when relevant (was: [PATCH v4] c++: Add gnu::diagnose_as attribute)

[Jason Merrill <jason@redhat.com>]

On 11/17/21 17:51, Matthias Kretz wrote:
> On Wednesday, 17 November 2021 19:25:46 CET Jason Merrill wrote:
>> On 11/17/21 04:04, Matthias Kretz wrote:
>>> On Wednesday, 17 November 2021 07:09:18 CET Jason Merrill wrote:
>>>>> -  if (CHECKING_P)
>>>>> -    SET_NON_DEFAULT_TEMPLATE_ARGS_COUNT (a, TREE_VEC_LENGTH (a));
>>>>> +  SET_NON_DEFAULT_TEMPLATE_ARGS_COUNT (a, nondefault);
>>>>
>>>> should have been
>>>>
>>>> if (CHECKING_P || nondefault != TREE_VEC_LENGTH (a))
>>>>
>>>>      SET_NON_DEFAULT_TEMPLATE_ARGS_COUNT (a, nondefault);
>>>
>>> TBH, I don't understand the purpose of CHECKING_P here, or rather it makes
>>> me nervous because AFAIU I'm only testing with CHECKING_P enabled. Why
>>> make behavior dependent on CHECKING_P? I expected CHECKING_P to basically
>>> only add more assertions.
>>
>> The idea when NON_DEFAULT_TEMPLATE_ARGS_COUNT was added years back was
>> to leave the TREE_CHAIN null when !CHECKING_P and treat that as
>> equivalent to TREE_VEC_LENGTH (args).  But perhaps you're right that
>> it's not a savings worth the complexity.
> 
> Thanks, now I understand.
> 
>>>>> (copy_template_args): Jason?
>>>>
>>>> Only copy the non-default template args count on TREE_VECs that should
>>>> have it.
>>>
>>> Why not simply set the count on all args? Is it a performance concern? The
>>> INTEGER_CST the TREE_CHAIN has to point to exists anyway, so it's not
>>> wasting any memory, right?
>>
>> In this case the TREE_VEC we're excluding is the one wrapping multiple
>> levels of template args; it doesn't contain args directly, so setting
>> NON_DEFAULT_ARGS_COUNT on it doesn't make sense.
> 
> Right, I had already added a `gcc_assert (!TMPL_ARGS_HAVE_MULTIPLE_LEVELS
> (args))` to my new set_non_default_template_args_count function and found cp/
> constraint.cc:2896 (get_mapped_args), which calls
> SET_NON_DEFAULT_TEMPLATE_ARGS_COUNT on the outer TREE_VEC. Was this supposed
> to apply to all inner TREE_VECs? Or is deleting the line the correct fix?

That should apply to the inner TREE_VECs (and probably use list.length)

>>>>> +  /* Pretty print only template instantiations. Don't pretty print
>>>>> explicit
>>>>> +     specializations like 'template <> void fun<int> (int)'.
>>>>
>>>> This seems like a significant change of behavior unrelated to printing
>>>> default template arguments.  What's the rationale for handling
>>>> specializations differently from instantiations?
>>>
>>> Right, this is about "The general idea of this change is to print template
>>> parms wherever they would appear in the source code as well".
>>>
>>> Initially, the change to print function template arguments/parameters only
>>> if the args were explicitly specified lead to printing 'void fun (T)
>>> [with T = ...]' or 'template <> void fun (int)'. Both are not telling the
>>> full story, even if the former is how the function would be called.
>>
>> and the latter is how I expect the specialization to be declared, not
>> with the deducible template argument made explicit.
> 
> You're right. From my tests:
> 
> template <class a>
>    [[deprecated]] void f4(a);
> 
> template <>
>    [[deprecated]] void f4<int>(int);
> 
> template <>
>    [[deprecated]] void f4(float);
> 
>    f4(1.);          // { dg-warning "'void f4\\(a\\) .with a = double.'" }
>    f4(1);           // { dg-warning "'void f4<int>\\(int\\)'" }
>    f4(1.f);         // { dg-warning "'void f4\\(float\\)'" }
> 
> So how it's printed depends on how the specialization is declared. It just
> falls out that way and it didn't seem awfully wrong... ;)
> 
>>> But if the reader
>>> should quickly recognize what code is getting called, it is helpful to see
>>> right away that a function template specialization is called. (It might
>>> also reveal an implementation detail of a library, so it's not 100%
>>> obvious how to choose here.) Also, saying 'T = int' is kind of wrong.
>>> Yes, 'int' was deduced. But there's no T in fun<int>:
>>>
>>> template <class T> void fun (T);
>>> template <> void fun<int> (int);
>>
>> There's a T in the template, and as you said above, that's how it's
>> called (and mangled).
>>
>>> __FUNCTION__ was 'fun<int>' all the time, but __PRETTY_FUNCTION__ was
>>> 'void
>>> fun(T) [with T = int]'.
>>
>> Isn't that true for instantiations, as well?
> 
> No, instantiations don't have template args/parms in __FUNCTION__.

Hmm, that inconsistency seems like a bug, though I'm not sure whether it 
should have the template args or not; I lean toward not.  The standard 
says that the value of __func__ is implementation-defined, the GCC 
manual says it's "the unadorned name of the function".

>>> It's more consistent that __PRETTY_FUNCTION__ contains __FUNCTION__, IMHO
>>
>> I suppose, but I don't see that as a strong enough motivation to mix
>> this up.
> 
> What about
> 
> template <class T> void f();
> template <> void f<int>();
> 
> With -fpretty-templates shouldn't it print as 'void f<T>() [with T = float]'
> and 'void f<int>()'? Yes, it's probably too subtle for most users to notice
> the difference. But I find it's more consistent this way.

I disagree; the function signature is the same whether a particular 
function is an explicit specialization or an instantiation.

>>> so it would have to be at least 'void fun<int>(T) [with T
>>> = int]'. But that's strange: How it uses T and int for the same type. So I
>>> settled on 'void fun<int>(int)'.
>>>
>>>> I also don't understand the purpose of TFF_AS_PRIMARY.
>>>
>>> dump_function_decl generalizes the TEMPLATE_DECL (if flag_pretty_templates
>>> is true) and, before this change, passes the generalized TEMPLATE_DECL to
>>> dump_type (... DECL_CONTEXT (t) ...) and dump_function_name (... t ...).
>>> That's why the whole template is printed as primary template (i.e. with
>>> template parms instead of template args, as is needed for
>>> flag_pretty_templates). But this drops the count of non-default template
>>> args.
>> Ah, you're trying to omit defaulted parms from the <list>?  I'm not sure
>> that's necessary, leaving them out of the [with ...] list should be
>> sufficient.
> 
> I was thinking about all the std::allocator defaults in the standard library.
> I don't want to see them. E.g. vector<int>::clear() on const object:
> 
> error: passing 'const std::vector<int>' as 'this' argument discards qualifiers
> [...]/stl_vector.h:1498:7: note:   in call to 'void std::vector<_Tp,
> _Alloc>::clear() [with _Tp = int; _Alloc = std::allocator<int>]'
> 
> With my patch the last line becomes
> [...]/stl_vector.h:1498:7: note:   in call to 'void std::vector<_Tp>::clear()
> [with _Tp = int]'
> 
> 
> Another case I didn't consider before:
> 
> template <class T, class U = int> struct A {
>    [[deprecated]] void f(U);
> };
> 
> A<float> a; a.f(1);
> 
> With my patch it prints 'void A<T>::f(U) [with T = float]', with your
> suggestion 'void A<T, U>::f(U) [with T = float]'. Both are missing important
> information in the substitution list, IMHO. Would 'void A<T, U = int>::f(U)
> [with T = float]' be an improvement? Or should find_typenames (in cp/error.c)
> find defaulted template parms and add them to its list? IIUC find_typenames
> would find all template parms and couldn't know whether they're defaulted.

That sounds good: omit defaulted parms only if they don't appear in the 
signature (other than as another default template argument).

>>>>> +   4. either
>>>>> +      - flags requests to show no function arguments, in which case
>>>>> deduced +        types could be hidden, or
>>>>> +      - at least one function template argument was given explicitly,
>>>>> or
>>>>> +      - we're printing a DWARF name,
>>>>
>>>> Though, why do we want
>>>> different behavior here when printing a DWARF name?
>>>
>>> Sorry, I should have asked ... I also added the same issue as an open
>>> question on the diagnose_as patch. When I ran the whole testsuite I had
>>> failures in the DWARF tests. This change resolved them. I don't know
>>> enough about how those strings are used and whether they may change
>>> between GCC versions. Anyway, the DWARF strings in that test had only the
>>> function name and template argument list (i.e. no function arguments).
>>
>> If they only have arguments, isn't TFF_NO_FUNCTION_ARGUMENTS set?
> 
> No, it isn't set. I could try to set it for DWARF names. It sounds like the
> right solution here.

I agree.

Jason