[PATCH] D57464: Generalize method overloading on addr spaces to C++

Bevin Hansson via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Fri Feb 1 08:09:40 PST 2019


ebevhan added inline comments.


================
Comment at: lib/Parse/ParseDeclCXX.cpp:2313
+    }
+  }
 
----------------
Anastasia wrote:
> ebevhan wrote:
> > Anastasia wrote:
> > > Anastasia wrote:
> > > > ebevhan wrote:
> > > > > Is there a reason that the attributes are parsed here and not in `ParseFunctionDeclarator` like the rest of the ref-qualifiers (and the OpenCL++ ASes)?
> > > > > 
> > > > > That is possibly why the parser is getting confused with the trailing return.
> > > > Good question! I have a feeling that this was done to unify parsing of all CXX members, not just methods. For collecting the method qualifiers it would certainly help making flow more coherent if this was moved to `ParseFunctionDeclarator`. I will try to experiment with this a bit more. What I found strange that the attributes here are attached to the function type itself instead of its  qualifiers. May be @rjmccall can shed some more light on the overall flow...
> > > I looked at this a bit more and it seems that all the GNU attributes other than addr spaces belong to the function (not to be applied to `this`) for example https://blog.timac.org/2016/0716-constructor-and-destructor-attributes/. It seems correct to parse them here and apply to the function type. Although we could separate parsing of addr space attribute only and move into `ParseFunctionDeclarator`.  However this will only be needed for the function type not for the data members. So not sure whether it will make the flow any cleaner.
> > > I looked at this a bit more and it seems that all the GNU attributes other than addr spaces belong to the function (not to be applied to this) 
> > 
> > Hm, I know what you mean. It's why Clang usually complains when you try placing an AS attribute after a function declarator, because it's trying to apply it to the function rather than the method qualifiers.
> > 
> > > However this will only be needed for the function type not for the data members. 
> > 
> > But we aren't really interested in data members, are we? Like struct fields, non-static data members cannot be qualified with ASes (they should 'inherit' the AS from the object type), and static ones should probably just accept ASes as part of the member type itself.
> > 
> > These patches are to enable AS-qualifiers on methods, so it only stands to reason that those ASes would be parsed along with the other method qualifiers.
> > 
> > ParseFunctionDeclarator calls ParseTypeQualifierListOpt to get the cv-qualifier-seq for the method qualifiers. Currently it's set to `AR_NoAttributesParsed`. If we enable parsing attributes there, then the qualifier parsing might 'eat' attributes that were possibly meant for the function.
> > 
> > This is just a guess, but what I think you could do is include attributes in the cv-qualifier parsing with `AR_GNUAttributesParsed`, look for any AS attributes, take their AS and mark them invalid, then move the attributes (somehow) from the qualifier-DeclSpec to the `FnAttrs` function attribute list.
> > 
> > (The reason I'm concerned about where/how the qualifiers are parsed is because this approach only works for custom ASes applied with the GNU attribute directly. It won't work if custom address spaces are given with a custom keyword qualifier, like in OpenCL. If the ASes are parsed into the DeclSpec used for the other ref-qualifiers, then both of these cases should 'just work'.)
> > But we aren't really interested in data members, are we? Like struct fields, non-static data members cannot be qualified with ASes (they should 'inherit' the AS from the object type), and static ones should probably just accept ASes as part of the member type itself.
>  
> Pointer data members can be qualified by and address space, but this should be parsed before.
> 
> 
> > This is just a guess, but what I think you could do is include attributes in the cv-qualifier parsing with AR_GNUAttributesParsed, look for any AS attributes, take their AS and mark them invalid, then move the attributes (somehow) from the qualifier-DeclSpec to the FnAttrs function attribute list.
> 
> Ok, I will try that. Thanks for sharing your ideas!
> 
> Pointer data members can be qualified by and address space, but this should be parsed before.

Right, pointer-to-member is a thing. I always forget about those.


================
Comment at: test/SemaCXX/address-space-method-overloading.cpp:28
+   //inas4.bar();
+   noas.bar(); // expected-error{{call to member function 'bar' is ambiguous}}
+}
----------------
Anastasia wrote:
> ebevhan wrote:
> > Just as an aside (I think I mentioned this on one of the earlier patches as well); treating a non-AS-qualified object as being in a 'generic' AS even for normal C++ isn't a great idea IMO. The object initialization code should be checking if the ASes of the actual object and the desired object type are compatible, and only if so, permit the overload.
> I think changing this would require changing AS compatibility rules in general, not just for overloading but for example for conversions too. This seems like a wider scope change that might affect backwards compatibility. We might need to start an RFC on this first. John has suggested adding a target setting  for this on the other review. I think that should work.  Another idea could be to add some compiler directives to specify what generic address space is (if any).
> 
> Using default (no address space) as generic has a lot of advantages since it doesn't need many parser changes. In OpenCL we weren't lucky that initial implementation was added that used default for private memory and therefore when generic was introduced we had to map it to a new lang addr space value. That required a lot of changes in the parser. But once we fix those actually, anyone should be able  to map generic to anything. I initially thought it has no other use cases than in OpenCL but now I feel there might be a value to map default case (no address space) for something specific and then use generic to specify a placeholder address space on pointers or references. We could just expose generic address space generically and also have a mode with no generic address space. The latter means that conversions between different address spaces is never valid. Would this make sense?
The big problem with the address space support in Clang is that it isn't really very well formalized unless you're on OpenCL. There's no real way for a target to express the relations between its address spaces; most of the relations that exist are hard-coded.

The support should probably be generalized for both targets and for LangASes like OpenCL's. Maybe:

* the ASes would be defined in a TableGen file which specifies which ASes exist, which ones are compatible/subsets/supersets, etc,
* or, just have a target hook that lets a target express that a conversion from AS A to AS B is prohibited/permitted explicitly/permitted implicitly.

Just some loose ideas; an RFC for this is possibly the right way forward.

The reason I bring all of this up is primarily because in our target, the 'default' AS is disjoint from our other ones, so there's no 'generic' AS to be had. Both implicit and explicit conversion between these ASes is prohibited.  Since Clang currently doesn't allow you to express that ASes are truly incompatible, we have a flag that blocks such conversions unconditionally. Ideally it would be target-expressible.

-----

> I think changing this would require changing AS compatibility rules in general, not just for overloading but for example for conversions too. This seems like a wider scope change that might affect backwards compatibility. We might need to start an RFC on this first. John has suggested adding a target setting for this on the other review. I think that should work. Another idea could be to add some compiler directives to specify what generic address space is (if any).

I'm unsure whether any changes to the current semantics really need to be done, at least for the overloading problem.

For example, explicit conversion from a non-address space qualified pointer type to an address space qualified pointer type (and vice versa) is permitted, but implicit conversion is an error in both C and C++: https://godbolt.org/z/UhOC0g

For an overload candidate to be viable, there has to be an implicit conversion sequence for the implicit object argument to the implicit object parameter type. But no such implicit conversion sequence exists for types in different ASes, even today, or the implicit example in the godbolt would pass. So, an overload candidate with a different AS qualification than the object should not be viable.

This is clearly not compatible with OpenCL's notion of the `__generic` AS, but OpenCL already has logic for `__generic` in Qualifiers to handle that case (`isAddressSpaceSupersetOf`). Arguably, that behavior should probably not be hard coded, but that's how it is today.

(Also, just because an AS is a superset of another AS does not necessarily mean that the language should permit an implicit conversion between them, but it's a reasonable behavior, I guess.)

-----

Ultimately, whether or not a conversion of a pointer/reference from address space A to address space B is permitted should both be a function of what the target and the language allows, but that support doesn't exist. I also don't think that exposing a 'generic' AS is the right approach. There are ASes, and some conversions from ASes to other ASes are legal, while others are not. There's also the question of the difference between implicit and explicit conversions; blocking all implicit conversions between ASes would not work for OpenCL, so there must be a way to express that as well.

-----

These are all just loose thoughts from my head. I'm not terribly familiar with OpenCL except for what I've seen around the Clang code, so there might be details there that I'm not aware of.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D57464/new/

https://reviews.llvm.org/D57464





More information about the cfe-commits mailing list