[PATCH] D43248: [Attr] Fix printing of parameter indices in attributes

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added inline comments.


================
Comment at: include/clang/Basic/Attr.td:182
+  // it would always be false.
+  string DisallowImplicitThisParamName = disallowImplicitThisParamName;
+}
----------------
jdenny wrote:
> aaron.ballman wrote:
> > jdenny wrote:
> > > jdenny wrote:
> > > > aaron.ballman wrote:
> > > > > Is there much benefit to forcing the attribute author to pick a name for this? It seems like this is more of a Boolean value: either implicit this is allowed or not. It would be really nice if we could hide these mechanics as implementation details so that the user only needs to write `VariadicParamIndexArgument<"name", DisallowImplicitThis>` (or something similar) when defining the attribute, and ideally not have to do any extra work in SemaDeclAttr.cpp by taking care of it in ClangAttrEmitter.cpp if possible.
> > > > Thanks for the review.  I'll work on that.
> > > > Is there much benefit to forcing the attribute author to pick a name for this? It seems like this is more of a Boolean value: either implicit this is allowed or not.
> > > 
> > > If the attribute author picks the name, then the attribute author can ensure there's only one of these per attribute.  I could rewrite it to have one per VariadicParamIndexArgument and one per ParamIndexArgument if you like.  In that case, ArgumentWithTypeTagAttr would end up with two of these, and future attributes could potentially have more, but they should all have the same value within a single attribute.  I didn't investigate how that redundancy would actually impact memory usage.  What do you think?
> > > 
> > > >  It would be really nice if we could hide these mechanics as implementation details so that the user only needs to write VariadicParamIndexArgument<"name", DisallowImplicitThis> (or something similar) when defining the attribute, and ideally not have to do any extra work in SemaDeclAttr.cpp by taking care of it in ClangAttrEmitter.cpp if possible.
> > > 
> > > So far, I haven't found a good way to accomplish that, or maybe I've misunderstood you....
> > > 
> > > The logic of checkFunctionOrMethodParameterIndex in SemaDeclAttr.cpp seems pretty tightly coupled with its users' logic.  For example, handleNonNullAttr uses the indices as adjusted by checkFunctionOrMethodParameterIndex to determine which indices belong in the array to be passed to the NonNullAttr constructor.  We could try to have NonNullAttr (in the constructor, presumably) perform the adjustment of indices so that SemaDeclAttr.cpp doesn't need that logic, but then it would be too late to for handleNonNullAttr to filter them.
> > > 
> > > The only extra work this patch adds to SemaDeclAttr.cpp beyond what's already there is to reuse the DisallowImplicitThis that is essentially already computed in checkFunctionOrMethodParameterIndex.
> > > 
> > > Another possibility is to have SemaDeclAttr.cpp fully encapsulate the index adjustment logic and pass an index offset to attribute constructors, so ClangAttrEmitter.cpp would just know it has to print indices with some given offset.  But a general index offset is wider than the single bool being stored now.  Again, I haven't investigated the actual impact on memory usage.
> > > 
> > > Do you see a better way to achieve the encapsulation you're looking for?
> > > 
> > > 
> > > In that case, ArgumentWithTypeTagAttr would end up with two of these, and future attributes could potentially have more, but they should all have the same value within a single attribute. I didn't investigate how that redundancy would actually impact memory usage. What do you think?
> > 
> > That redundancy could probably be worked around in ClangAttrEmitter.cpp by inspecting the attribute argument list and noticing duplicate `[Variadic]ParamIndexArgument` before generating the code for the arguments, perhaps?
> > 
> > > Another possibility is to have SemaDeclAttr.cpp fully encapsulate the index adjustment logic and pass an index offset to attribute constructors, so ClangAttrEmitter.cpp would just know it has to print indices with some given offset.
> > 
> > This was more along the lines of what I was thinking of.
> > 
> > Basically, it seems like the declarative part in Attr.td should be able to specify an attribute argument is intended to be a function parameter positional index, and whether implicit this needs special handling or not. Past that, the semantics of the attribute really shouldn't matter -- when asking for the index, it should be automatically adjusted so that users of the attribute don't have to do anything, and when printing the attribute, the index should be appropriately adjusted back. I'd like to avoid needing collusion between the declarative part in Attr.td and the semantic part in SemaDeclAttr.cpp because that collusion can get out of sync. It feels like we should be able to hide that collusion in ClangAttrEmitter.cpp by using some extra flags on the attribute that get set (automatically) during construction. The way you have things set up right now is very close to this, except the information has to be passed to the constructor manually in SemaDeclAttr after being calculated by checkFunctionOrMethodParameterIndex().
> > 
> > I'm not too worried about the memory usage at this point; if it looks problematic, we can likely address it.
> > That redundancy could probably be worked around in ClangAttrEmitter.cpp by inspecting the attribute argument list and noticing duplicate [Variadic]ParamIndexArgument before generating the code for the arguments, perhaps?
> 
> Seems reasonable.
> 
> There's also the question of the constructor interface: does it accept the same information multiple times or one time?
> 
> For now, the memory optimization doesn't seem too important, so I won't work on detecting duplicates.  If we implement that later, we should be able to easily adjust callers for the resulting interface change.
> 
> >> Another possibility is to have SemaDeclAttr.cpp fully encapsulate the index adjustment logic and pass an index offset to attribute constructors, so ClangAttrEmitter.cpp would just know it has to print indices with some given offset.
> > 
> > This was more along the lines of what I was thinking of.
> 
> Sorry, does "This" refer to the alternative proposal I just described above or your following paragraph?
> 
> > Basically, it seems like the declarative part in Attr.td should be able to specify an attribute argument is intended to be a function parameter positional index, and whether implicit this needs special handling or not.
> 
> Attr.td cannot specify declaratively whether implicit this needs special handling (I've been calling this DisallowImplicitThisParam, but let's call it AdjustForImplicitThis for clarity) because part of that decision is per function:  whether the function actually has an implicit this (HasImplicitThis).  Attr.td might always be able to specify whether implicit this is allowed (AllowImplicitThis) and thus, when there actually is an implicit this (HasImplicitThis), whether an extra increment should be added for printing (AdjustForImplicitThis = HasImplicitThis && !AllowImplicitThis).  ClangAttrEmitter.cpp would then have to learn how to compute AdjustForImplicitThis from the other two Booleans, but SemaDeclAttr.cpp already does that.
> 
> In my current patch, only SemaDeclAttr.cpp understands how to compute AdjustForImplicitThis, but ClangAttrEmitter.cpp documents it too and knows there's always another increment (to make indices one-origin) for printing.  In my alternative proposal above, ClangAttrEmitter.cpp would not document or understand any of this.  Instead, ClangAttrEmitter.cpp would only know that indices must be adjusted by some arbitrary value for the sake of printing, and SemaDeclAttr.cpp would tell it how much.  That seems like the cleanest separation.
> 
> > Past that, the semantics of the attribute really shouldn't matter -- when asking for the index, it should be automatically adjusted so that users of the attribute don't have to do anything, and when printing the attribute, the index should be appropriately adjusted back.
> 
> Agreed.  Both my current patch and my alternative proposal above offer that.
> 
> > I'd like to avoid needing collusion between the declarative part in Attr.td and the semantic part in SemaDeclAttr.cpp because that collusion can get out of sync.
> 
> The only collusion I see in my current patch (besides some perhaps unnecessary documentation in ClangAttrEmitter.cpp) is the understanding that the Boolean indicates that one increment beyond the usual single increment is needed when printing.  I believe my alternative proposal eliminates even that: then there's just some arbitrary increment when printing, and ClangAttrEmitter.cpp doesn't know or care why.
> 
> > It feels like we should be able to hide that collusion in ClangAttrEmitter.cpp by using some extra flags on the attribute that get set (automatically) during construction. The way you have things set up right now is very close to this, except the information has to be passed to the constructor manually in SemaDeclAttr after being calculated by checkFunctionOrMethodParameterIndex().
> 
> In my last response, I was explaining why checkFunctionOrMethodParameterIndex probably needs to calculate that information anyway for the sake of its users, and it seems that having both SemaDeclAttr and ClangAttrEmitter know how to calculate it is worse for maintenance than having SemaDeclAttr pass its result to ClangAttrEmitter.  My alternative proposal above does not require ClangAttrEmitter to calculate the adjustment at all.  It just receives the net adjustment from SemaDeclAttr.
> 
> I feel like my alternative proposal would satisfy your desire for a better separation of concerns.  The difference from what I think you're describing is that my alternative proposal keeps the index adjustment amount computation in SemaDeclAttr rather than trying to move any of that computation to ClangAttrEmitter.  Does that seem reasonable to you?
> Sorry, does "This" refer to the alternative proposal I just described above or your following paragraph?

Your proposal, which I thought was solving what I described in the following paragraph.

>  Attr.td might always be able to specify whether implicit this is allowed (AllowImplicitThis) and thus, when there actually is an implicit this (HasImplicitThis), whether an extra increment should be added for printing (AdjustForImplicitThis = HasImplicitThis && !AllowImplicitThis). ClangAttrEmitter.cpp would then have to learn how to compute AdjustForImplicitThis from the other two Booleans, but SemaDeclAttr.cpp already does that.

Okay, I think I see where we may not quite be on the same page. The Attr.td bit knows whether the attribute needs to adjust for implicit this or not, but only SemaDeclAttr.cpp knows about the actual function the attribute is trying to be applied to. What I was envisioning was to move the logic from SemaDeclAttr.cpp into something accessible from the code generated by ClangAttrEmitter.cpp, but that may not be sufficient because the semantic attribute needs that information from the specific Decl, which by definition requires something additional to be passed in from SemaDeclAttr.cpp.

Thanks for bearing with me. :-)

> I feel like my alternative proposal would satisfy your desire for a better separation of concerns. The difference from what I think you're describing is that my alternative proposal keeps the index adjustment amount computation in SemaDeclAttr rather than trying to move any of that computation to ClangAttrEmitter. Does that seem reasonable to you?

I think it's reasonable, depending on how it's surfaced. Would it be reasonable to make it look something like this?
```
// Attr.td
// ...
  let Args = [VariadicParamIndexArgument<"Args", AdjustsImplicitThis>];
// ...

// SemaDeclAttr.cpp
// ...
ImplicitThisAdjustment AdjustsImplicitThis;
if (!checkFunctionOrMethodParameterIndex(..., ..., AdjustsImplicitThis))
  return;

// ...
D->addAttr(::new (S.Context) NonNullAttr(..., ..., AdjustsImplicitThis));
```
(I'm using enumerations in both files rather than Boolean values so it's more clear what's happening).

This implies that any attribute accepting a parameter index will have a constructor that requires you to specify whether that instance of the attribute adjusts implicit this or not.


https://reviews.llvm.org/D43248