r350920 - [Sema] Make canPassInRegisters return true if the CXXRecordDecl passed

[Akira Hatanaka via cfe-commits]

> On Jan 16, 2019, at 4:20 PM, John McCall via cfe-commits <cfe-commits at lists.llvm.org> wrote:
> 
> 
> 
> On 16 Jan 2019, at 18:32, Richard Smith wrote:
> 
>> On Wed, 16 Jan 2019 at 09:10, John McCall via cfe-commits <
>> cfe-commits at lists.llvm.org> wrote:
>> 
>>> On 16 Jan 2019, at 9:13, Aaron Ballman wrote:
>>> 
>>>> On Wed, Jan 16, 2019 at 1:57 AM Akira Hatanaka <ahatanaka at apple.com>
>>>> wrote:
>>>>> 
>>>>> Yes, the behavior of the compiler doesn’t match what’s explained
>>>>> in the documentation anymore.
>>>>> 
>>>>> Please take a look at the attached patch, which updates the
>>>>> documentation.
>>>> 
>>>> Patch mostly LGTM, but I did have one wording suggestion.
>>>> 
>>>>> diff --git a/include/clang/Basic/AttrDocs.td
>>>>> b/include/clang/Basic/AttrDocs.td
>>>>> index 5773a92c9c..ca3cfcf9b2 100644
>>>>> --- a/include/clang/Basic/AttrDocs.td
>>>>> +++ b/include/clang/Basic/AttrDocs.td
>>>>> @@ -2478,15 +2478,20 @@ def TrivialABIDocs : Documentation {
>>>>>   let Category = DocCatVariable;
>>>>>   let Content = [{
>>>>> The ``trivial_abi`` attribute can be applied to a C++ class, struct,
>>>>> or union.
>>>>> -It instructs the compiler to pass and return the type using the C
>>>>> ABI for the
>>>>> +``trivial_abi`` has the following effects:
>>>>> +
>>>>> +- It instructs the compiler to pass and return the type using the C
>>>>> ABI for the
>>>>> underlying type when the type would otherwise be considered
>>>>> non-trivial for the
>>>>> purpose of calls.
>>>>> -A class annotated with `trivial_abi` can have non-trivial
>>>>> destructors or copy/move constructors without automatically becoming
>>>>> non-trivial for the purposes of calls. For example:
>>>>> +- It makes the destructor and copy and move constructors of the
>>>>> class trivial
>>>>> +that would otherwise be considered non-trivial under the C++ ABI
>>>>> rules.
>>>> 
>>>> How about: It makes the destructor, copy constructors, and move
>>>> constructors of the class trivial even if they would otherwise be
>>>> non-trivial under the C++ ABI rules.
>>> 
>>> Let's not say that it makes them trivial, because it doesn't.  It causes
>>> their immediate non-triviality to be ignored for the purposes of
>>> deciding
>>> whether the type can be passed in registers.
>>> 
>> 
>> Given the attribute now forces the type to be passed in registers, I think
>> it'd be more to the point to say that it makes the triviality of those
>> special members be ignored when deciding whether to pass a type with a
>> subobject of this type in registers.
> 
> Wait, it forces it to be passed in registers?  I thought the design
> was that it didn't override the non-trivial-abi-ness of subobjects;
> see all the discussion of trivially_relocatable.
> 

As stated in the documentation, trivial_abi on a class has no effect when one of its subobject is non-trivial for the purpose of calls. In the following example, clang (Sema::checkIllFormedTrivialABIStruct) issues a warning and drops the attribute on D since B is non-trivial:

struct B {
  B(const B &);
};

struct __attribute__((trivial_abi)) D {  // warning: 'trivial_abi' cannot be applied to ‘D' [-Wignored-attributes]
  B b;
};

If the attribute on the class survives the check by Sema::checkIllFormedTrivialABIStruct, the class is treated as being trivial for the purposes of calls.

> John.
> _______________________________________________
> cfe-commits mailing list
> cfe-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20190116/56b4956d/attachment.html>