Re: [PATCH] c++/modules: depending local enums [PR104919, PR106009]

[Jason Merrill <jason@redhat.com>]

On 3/1/24 10:32, Jason Merrill wrote:
> On 3/1/24 10:00, Patrick Palka wrote:
>> On Fri, 1 Mar 2024, Jason Merrill wrote:
>>
>>> On 2/29/24 15:56, Patrick Palka wrote:
>>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look
>>>> OK for trunk?
>>>>
>>>> -- >8 --
>>>>
>>>> For local enums defined in a non-template function or a function 
>>>> template
>>>> instantiation it seems we neglect to make the function depend on the 
>>>> enum
>>>> definition, which ultimately causes streaming to fail due to the enum
>>>> definition not being streamed before uses of its enumerators are 
>>>> streamed,
>>>> as far as I can tell.
>>>
>>> I would think that the function doesn't need to depend on the local 
>>> enum in
>>> order for the local enum to be streamed before the use of the 
>>> enumerator,
>>> which comes after the definition of the enum in the function body?
>>>
>>> Why isn't streaming the body of the function outputting the enum 
>>> definition
>>> before the use of the enumerator?
>>
>> IIUC (based on observing the behavior for local classes) streaming the
>> definition of a local class/enum as part of the function definition is
>> what we want to avoid; we want to treat a local type definition as a
>> logically separate definition and stream it separately (similar
>> to class defns vs member defns I guess).  And by not registering a 
>> dependency
>> between the function and the local enum, we end up never streaming out
>> the local enum definition separately and instead stream it out as part
>> of the function definition (accidentally) which we then can't stream in
>> properly.
>>
>> Perhaps the motivation for treating local type definitions as logically
>> separate from the function definition is because they can leak out of a
>> function with a deduced return type:
>>
>>    auto f() {
>>      struct A { };
>>      return A();
>>    }
>>
>>    using type = decltype(f()); // refers directly to f()::A
> 
> Yes, I believe that's what modules.cc refers to as a "voldemort".
> 
> But for non-voldemort local types, the declaration of the function 
> doesn't depend on them, only the definition.  Why doesn't streaming them 
> in the definition work properly?

And does your 99426 patch address that problem?

Jason