Re: [PATCH 2/5] Make sure that static data member constexpr isn't optimized away in test.

[Jason Merrill <jason@redhat.com>]

On 9/1/20 2:46 PM, Jason Merrill wrote:
> On 8/25/20 5:19 AM, Mark Wielaard wrote:
>> On Mon, 2020-08-24 at 17:38 -0400, Jason Merrill wrote:
>>>> This looks incorrect to me, that is a workaround for a real GCC bug.
>>>>
>>>> Shouldn't we instead do something like (untested) following patch?
>>>> I mean, for DWARF < 5 the static data members were using DW_TAG_member,
>>>> which has been always marked by the function, so IMHO we should also 
>>>> always
>>>> mark the DW_TAG_variables at the class scope that replaced those.
>>>
>>> The earlier behavior seems like an accident, that happened because we
>>> always need to emit information about non-static data members.  I don't
>>> think we should take it as guidance.
>>
>> Maybe the reason they got emitted this way was an accident on the GCC
>> side. But I don't think it is an accident on the GDB side. At least
>> looking at the various gdb testcases, the intention is to show a
>> struct/class type as defined to the user, which includes both the
>> static and non-static data members of a class.
> 
> That would make sense.

So, GDB prefers no pruning of members...

>>> In this case one reason we don't emit debug info is because (before
>>> C++17) there's no definition of 'b'.  After C++17 the in-class
>>> declaration of 'b' is a definition, but we don't have to give it a
>>> symbol, so there's still nothing for the debug info to describe.
>>
>> But don't we describe other parts of a type that don't have a symbol as
>> part of the debug info?
> 
> It seems that currently we describe unused/undefined functions, but not 
> unused nested types/typedefs.

> On 8/25/20 8:41 AM, Jakub Jelinek wrote:
>>
>> On Mon, Aug 24, 2020 at 05:38:28PM -0400, Jason Merrill via 
>> Gcc-patches wrote:
>>>
>>> This issue doesn't seem specific to class members; it also affects
>>> namespace-scope C++17 inline variables:
>>>
>>> inline const int var = 42;
>>> int main() { return var; }
>>>
>>> Compiling this testcase with -g doesn't emit any debug info for 'var' 
>>> even
>>> though it's used.
>>>
>>> Should we assume that if a variable with DW_AT_const_value is 
>>> TREE_USED, we
>>> need to write out debug info for it?
>>
>> I guess it is a question of how exactly the (default on)
>> -feliminate-unused-debug-symbols should behave with different kind of
>> entities.
>>
>> One thing are the non-inline static data members without const/constexpr or
>> without initializer in the class.  Those need a definition if they are ever
>> used, so it is probably fine to only emit them in the class in the TU where
>> they are defined.  But note that e.g. with -fdebug-types-section we still
>> force them to be output in class and do no pruning (and the pruning actually
>> makes dwz less efficient unless dwz is tought to not only merge the DIEs
>> with the same children and attributes, but also reconstruct more complete
>> DIEs out of several less complete ones for the same class).
> 
> Right, this gets at Mark's point above.  How much pruning do we want to 
> do of class bodies?  We currently do some, but how much benefit does 
> that actually give us?  Is it worth the cost?

...and our deduplication mechanisms prefer no pruning of members.

>> Another case is non-inline static const data member with initializer,
>> do we track non-odr uses e.g. during constant evaluation and if so, should
>> that result in the static data member appearing?  Because if the static
>> const data member with initializer is never odr used, it doesn't have to be
>> ever defined and so it might never appear in the debug info.
>>
>> Another case are inline vars, shall we treat as being used just that they
>> were used in some constant expression, or do only odr uses count?
> 
> If the goal of debug info is to be able to evaluate the same expressions 
> that appear in the source, constant uses need to count, too.  I wonder 
> how we could associate the uses with their context so pruning works 
> properly.

For GCC 11, I think let's fix the regression with your (Jakub) earlier 
patch, maybe only for DIEs with DW_AT_const_value.

For GCC 12, maybe we want to stop pruning any class members by default.

Jason