[PATCH] IR: Add specialized debug info metadata nodes

[Duncan P. N. Exon Smith]

> On 2015-Feb-06, at 13:30, Frédéric Riss <friss at apple.com> wrote:
> 
>> 
>> On Feb 6, 2015, at 1:24 PM, David Blaikie <dblaikie at gmail.com> wrote:
>> 
>> 
>> 
>> On Fri, Feb 6, 2015 at 1:20 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>> 
>> > On 2015-Feb-06, at 10:36, David Blaikie <dblaikie at gmail.com> wrote:
>> >
>> >
>> >
>> > On Wed, Feb 4, 2015 at 4:17 PM, Adrian Prantl <aprantl at apple.com> wrote:
>> >
>> >> On Feb 4, 2015, at 3:53 PM, Frédéric Riss <friss at apple.com> wrote:
>> >>
>> >>
>> >>> On Feb 4, 2015, at 3:46 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
>> >>>
>> >>>
>> >>>> On 2015-Feb-04, at 15:37, Frédéric Riss <friss at apple.com> wrote:
>> >>>>
>> >>>>>>
>> >>>>>> - The word 'context' is overloaded: `MDNode::getContext()` already
>> >>>>>> exists, and returns an `LLVMContext&`; `DIDescriptor` uses 'context'
>> >>>>>> to mean "the node that this one is defined inside".  I chose the
>> >>>>>> word 'parent' instead of 'context' here.  Is this word okay?  If
>> >>>>>> not, what about 'scope'?  This will be reflected in the assembly
>> >>>>>> changes to come (I'd like the C++ names to match the assembly names,
>> >>>>>> although technically it's not necessary).
>> >>>>>>
>> >>>>>> I'd /probably/ go with scope (we already have scope in the MDLocations, so that seems consistent), but fairly on-the-fence.
>> >>>>>>
>> >>>>>
>> >>>>> Weirdly, I didn't even notice that :).  In that case I like 'scope'
>> >>>>> better too.  I'll update to that before commit.
>> >>>>
>> >>>> Seems most natural. Can the futur getScope() return something that doesn’t derive from MDScope?
>> >>>
>> >>> I don't think so.
>> >>
>> >> That was my impression also, and it makes it even more appropriate IMHO.
>> >>
>> >
>> > Just a few general remarks to throw into the discussion:
>> >
>> > - Would it make sense to use something like tablegen to generate the repetitive parts? I’m slightly worried about copy&paste bugs, moderately worried about refactoring it in the future.
>> >
>> > <classllvm_1_1DIScope__inherit__graph.png>
>> >
>> > As for scopes, there are several things that bug me about the current class hierarchy that we could fix now:
>> > - DIFile should not be a scope (the concept of files is IMO orthogonal to scoping and there is always something more appropriate to put a node into: compile unit, module, namespace)
>> >
>> > IIRC we do this when # line directives change the file-name part-way through. But I could be wrong... only vague idea.
>> >
>> > - DIBasicType should not be scope
>> >
>> > Part of this is the ability to treat all types the same - multiple inheritance might be an alternative, but without that we want to treat all types equally in some codepaths, but then only some types are valid scopes, etc. - so there's a few different axes on which we want to refer to these things. It's tricky. But certainly worth thinking about.
>> 
>> LLVM-style RTTI doesn't handle multiple inheritance, which restricts
>> us somewhat here.
>> 
>> *nod* I seemed to recall something like that
>>  
>> I imagine having a hierarchy like:
>> 
>>   DIScopeBase
>>    -> DIScope
>>        -> DINamespace
>>        -> DISubprogram
>>    -> DIType
>>        -> DIBasicType
>>        -> DIDerivedType
>>        -> DITypeScope
>>            -> DICompositeType
>> 
>> would get us most of the way there.
>> 
>> Pointers can be `DIScopeBase`.  Call-sites that care about scopes
>> can check:
>> 
>>     assert(isa<DIScope>(N) || isa<DITypeScope>(N));
>> 
>> Yep, if we're willing to pay the cost of having scopes not have a common ancestor to use. That might be the right tradeoff, I'm not sure.
>>  
>> (We can add `DIScopeBase::isValidScope()` or something to do this.)
>> 
>> Or maybe we don't even need a `DIScopeBase`.
>> 
>> Yeah, if we go that way I don't think DIScopeBase would add anything, but not 100% sure.
> 
> When I first discovered the type hierarchy, the every-type-is-a-scope thing got me a bit perplexed. I really think we should get rid of it. Would it be possible to have the Composite types point to a real scope instead of being one?

I was just thinking the same thing through on my whiteboard :).

I think separating the "type" and "scope" aspect of struct/class/etc. nodes makes a lot of sense.