[cxx-abi-dev] mangling of vendor extension, expression on function (not function type)

[John McCall]

On Jan 24, 2014, at 7:30 PM, Nick Lewycky <nlewycky at google.com> wrote:
> On 24 January 2014 17:54, John McCall <rjmccall at apple.com> wrote:
> On Jan 24, 2014, at 5:36 PM, Nick Lewycky <nlewycky at google.com> wrote:
>> On 23 January 2014 11:52, John McCall <rjmccall at apple.com> wrote:
>> On Jan 21, 2014, at 2:01 PM, Richard Smith <richardsmith at google.com> wrote:
>>> On 21 January 2014 09:36, John McCall <rjmccall at apple.com> wrote:
>>> On Jan 20, 2014, at 6:13 PM, Nick Lewycky <nlewycky at google.com> wrote:
>>> > I'm trying to mangle a vendor extension that encodes an expression which applies to function overload candidates, with the goal that a user running the demangler would see their expression demangled. While there are various vendor extension points, none of them allow me to go into encoding an expression, unless I stick a stray "decltype" in there, or similar (expression in a template argument that doesn't actually exist, etc.).
>>> >
>>> > The vendor extension is described in full here: http://clang.llvm.org/docs/LanguageExtensions.html#controlling-overload-resolution .
>>> >
>>> > I don't think I can do it without an extension to the mangling rules. As a strawman proposal, I suggest this:
>>> >
>>> > <type> ::= Ue <expression> E # vendor extended type qualifier
>>> 
>>> I think you mean
>>> 
>>>   <type> ::= Ue <source-name> <expression> E <type>
>>> 
>>> And this would be valuable for mangling e.g. dependent address space qualifiers, if anybody ever wants to do those.
>> 
>> Yep, that's what I meant. Thanks!
>>> We could generalize this slightly to
>>> 
>>>   <type> ::= U <source-name> <template-args> <type>
>>> 
>>> to allow multiple arguments (as types or expressions), dependent pack expansions, and so on.
>> 
>> That’s a bit more future-proof, I suppose, although I think it might stretch the definition of a type-qualifier to embed anything other than a value, and value pack-expansions are already part of the <expression> grammar.  You’re really asking for a “allow an arbitrarily complex type to be manufactured here” mangling.
>> 
>>> However, it feels really forced to add your feature, specifically, to <type>, since it can’t appear in an arbitrary position; it’s much closer to a qualified method name.  But the ref/cv-qualifiers area is only allowed in a <nested-name>, and you need to be able to do this on a top-level function, so I suppose extending <type> in a known-useful direction and then abusing <type> might be the best thing.
>>> 
>>> On the other hand, isn’t this a proposed direction for standardization?  I would have no problem with giving this a proper, non-vendor mangling just in case.
>>> 
>>> It's not proposed for standardization with this syntax, and it's likely that the final semantics will differ from the Clang extension in some ways (the proposed partial ordering rules for constraints are rather more complex, for instance), but it's close enough that it's an option worth considering.
>> 
>> Unless the feature is likely to diverge so much that it won’t even be an expression anymore, I don’t think this poses any problem for the ABI.
>> 
>> Vendor hat on, I reserve the right to make my extension behave differently from anything that's been standardized. As long as I can slip a vendor extension particle into the mangled name I'll be happy to use otherwise normal mangling. If it turns out I don't have to, all the better, but I'm not banking on it.
> 
> I completely agree that this is acceptable vendor-hat behavior and that the fake-qualifier idea isn’t a bad approach for it.
> 
>> Do you want me to try to prepare a patch for template constraints? I think it would look similar to my strawman proposal, except that I'd pick some other available letter?
> 
> Yes, except that grammatically you should make it part of the function <encoding> instead of adding it to <type>.  It works out to the same basic position.
> 
> Okay, first attempt at a patch attached. Please review.
> 
> A couple things. I chose 'Q', short for 'requires' to indicate a constraint. I put the new part on all encodings, not just functions, because you may need to mangle a static data member inside a class that has a concept applied, and similarly for its vtable and special members.

I’m confused about what you mean by ‘concept’ here.  If it's just jargon for the enable_if feature, it seems completely counterproductive.

We do not need to mangle enable_if conditions on class template patterns because of the ODR (and related rules in the templates chapter).

enable_if has to be mangled for a function template because it’s part of the template signature.  You can have two function template declarations in the scope scope with the same name and with identical template parameter signatures and type signatures, but if they have different enable_if conditions, then they’re considered different templates.  In typical use, this isn’t actually important: two overlapping function templates are probably defined in the same header file, and it’s likely that every call site will see both of them, and so any given list of template arguments that doesn’t lead to an ambiguity will probably pick the same template every time.  But that’s not actually *required*: it’s totally fine to have two different function templates in two different translation units that both accept the same template arguments, and we need to distinguish them.

But two class temploids with the same name in the same scope always declare the same entity for the ODR’s purposes, and if a particular list of template arguments would pick an explicit specialization ([temp.expl.spec]p6) or partial specialization ([temp.class.spec]p1), that specialization is required to be defined in every translation unit that would use it.  So while you can guard a template pattern with an enable_if clause, it just controls whether that pattern is selected for instantiation, and the decision to select that pattern doesn’t need to be mangled in for the same basic reason that we don’t have to mangle in the partial-specialization signature we choose: ultimately, you can never legally simultaneously instantiate foo<int,char*> from two different patterns.

Now, there are concepts proposals that interfere with this along multiple axes.  There are proposals which allow you to explicitly choose between ambiguous concept maps for a template argument, and there are proposals which allow you to define types in a concept map that would otherwise conflict in name with each other.  Both proposals affect mangling: the former suggests that we need to mangle a reference to the concept map as part of the template arguments, and the latter suggests that we need to mangle the concept map as part of the context.  But in neither case does a random expression in the <encoding> help.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sourcerytools.com/pipermail/cxx-abi-dev/attachments/20140127/5fd63a9b/attachment.html>