recordDecl() AST matcher

[Aaron Ballman via cfe-commits]

On Sat, Sep 12, 2015 at 11:06 PM, Manuel Klimek <klimek at google.com> wrote:
>
>
> On Sat, Sep 12, 2015, 9:25 PM Aaron Ballman <aaron at aaronballman.com> wrote:
>>
>> On Sat, Sep 12, 2015 at 8:22 AM, Manuel Klimek <klimek at google.com> wrote:
>> >
>> >
>> > On Fri, Sep 11, 2015 at 10:39 PM Aaron Ballman <aaron at aaronballman.com>
>> > wrote:
>> >>
>> >> On Fri, Sep 11, 2015 at 4:30 PM, Richard Smith <richard at metafoo.co.uk>
>> >> wrote:
>> >> > I don't think CXXRecordDecl is an anachronism, so much as an
>> >> > implementation
>> >> > detail; it makes sense to use a smaller class when in C mode, as we
>> >> > don't
>> >> > need most of the features and complexity that CXXRecordDecl brings
>> >> > with
>> >> > it.
>> >> > But... as a user of clang matchers, I don't think I'd want to care
>> >> > about
>> >> > the
>> >> > difference, and it'd be more convenient if I could nest (say) a
>> >> > hasMethod
>> >> > matcher within a recordDecl matcher, since it's completely obvious
>> >> > what
>> >> > that
>> >> > should mean. If I have a matcher that says:
>> >> >
>> >> >   recordDecl(or(hasMethod(...), hasField(...)))
>> >> >
>> >> > I would expect that to work in both C and C++ (and the only way it
>> >> > could
>> >> > match in C would be on a record with the specified field, since the
>> >> > hasMethod matcher would always fail).
>> >>
>> >> Okay, so then it sounds like we want recordDecl to *mean* RecordDecl,
>> >> but we want the traversal and narrowing matchers that currently take a
>> >> CXXRecordDecl to instead take a RecordDecl and handle the CXX part
>> >> transparently? This means we would not need to add a cxxRecordDecl()
>> >> matcher, but could still access CXX-only functionality (like access
>> >> control, base classes, etc) through recordDecl()?
>> >
>> >
>> > I'm against that proposal. I think we have tried to make the matchers
>> > more
>> > "user friendly" in the past, and all those attempts have failed
>> > miserably;
>> > in the end, users will do ast-dump to see what they want to match, and
>> > then
>> > be confused when the matchers do follow the AST 99% of the time, but try
>> > to
>> > be smart 1% of the time.
>> > I think given that we want to keep CXXRecordDecl, the right solution is
>> > to
>> > have a cxxRecordDecl() matcher.
>>
>> Personally, I think this makes the most sense, at least to me. The
>> recommendation I've always heard (and given) is to use -ast-dump and
>> write matchers from there. (Consequently, the more I work with type
>> traversal matchers, the more I wish we had -ast-dump-types to give
>> even *more* information for writing matchers.)
>>
>> But the question still remains, what do we do with recordDecl? Right
>> now, it means CXXRecordDecl instead of RecordDecl. If we change it to
>> mean RecordDecl instead, there's a chance we'll break existing,
>> reasonable code. Are we okay with that risk? If we're at least
>> conceptually okay with it, I could make the change locally and see
>> just how much of our own code breaks, and report back. But if that
>> turns out to be problematic, do we want to deprecate recordDecl and
>> replace it with structDecl as our fallback position? Or is there a
>> better solution?
>>
>> Basically, I see a few ways to solve this (and there may be other ways
>> I'm not thinking about yet):
>>
>> 1) Undocument/deprecate recordDecl, add structDecl, unionDecl, and
>> cxxRecordDecl. This does not match the AST because we have no
>> StructDecl or UnionDecl types. The more I think about this option, the
>> less I like it. It's easy to implement, but seems hard to relate to
>> the AST.
>> 2) Make recordDecl match RecordDecl, don't touch other matchers. Add
>> way to distinguish unions from structs (e.g., isUnion(), isStruct()),
>> add cxxRecordDecl. This matches the AST most closely, but may break
>> code. I think that I prefer this approach, but it depends heavily on
>> what "may break code" looks like in practice.
>> 3) Make recordDecl match RecordDecl, fix other matchers that currently
>> take CXXRecordDecl to instead take RecordDecl and handle sensibly when
>> possible. Add a way to distinguish unions from structs, add
>> cxxRecordDecl. This doesn't match the AST because we will have
>> matchers taking a RecordDecl when the AST would require a
>> CXXRecordDecl, but is likely to break less code.
>
>
> That sums it up. My preferences are 2, 3, 1 in that order :)

I've attached a patch that implements #2, but it comes with ~85 errors
from C++ matchers that use recordDecl to mean cxxRecordDecl.

http://pastebin.com/bxkRcqBV

If this is an acceptable failure rate, I can also update the failing
matchers to use cxxRecordDecl instead of recordDecl where applicable.
Doing some spot-checking of the failing code, the failures are ones we
anticipated, such as:

constructorDecl(ofClass(recordDecl(
hasDeclaration(recordDecl(hasMethod(hasName("begin")),
hasMethod(hasName("end"))))
etc

~Aaron

>
>>
>> ~Aaron
>>
>>
>> > Richard: if CXXRecordDecl was really an implementation detail, it would
>> > be
>> > hidden behind a RecordDecl class, as an implementation detail. The
>> > reasons
>> > why we don't want it to be an implementation detail in the code
>> > (performance, data structure size) don't matter - in the end, it's in
>> > the
>> > AST API.
>> >
>> >>
>> >>
>> >> ~Aaron
>> >>
>> >> >
>> >> > On Fri, Sep 11, 2015 at 6:30 AM, Manuel Klimek <klimek at google.com>
>> >> > wrote:
>> >> >>
>> >> >> Richard! We need an informed opinion :D
>> >> >>
>> >> >> On Fri, Sep 11, 2015 at 3:07 PM Aaron Ballman
>> >> >> <aaron at aaronballman.com>
>> >> >> wrote:
>> >> >>>
>> >> >>> Ping?
>> >> >>>
>> >> >>> On Tue, Sep 8, 2015 at 9:26 AM, Manuel Klimek <klimek at google.com>
>> >> >>> wrote:
>> >> >>> > On Tue, Sep 8, 2015 at 3:23 PM Aaron Ballman
>> >> >>> > <aaron at aaronballman.com>
>> >> >>> > wrote:
>> >> >>> >>
>> >> >>> >> On Tue, Sep 8, 2015 at 9:18 AM, Manuel Klimek
>> >> >>> >> <klimek at google.com>
>> >> >>> >> wrote:
>> >> >>> >> > On Tue, Sep 8, 2015 at 2:23 PM Aaron Ballman
>> >> >>> >> > <aaron at aaronballman.com>
>> >> >>> >> > wrote:
>> >> >>> >> >>
>> >> >>> >> >> On Tue, Sep 8, 2015 at 5:40 AM, Manuel Klimek
>> >> >>> >> >> <klimek at google.com>
>> >> >>> >> >> wrote:
>> >> >>> >> >> > Yea, we had that discussion a few times, and I can never
>> >> >>> >> >> > remember
>> >> >>> >> >> > why
>> >> >>> >> >> > we
>> >> >>> >> >> > ended up in the state we're in.
>> >> >>> >> >> > We definitely had a time where we switched to just using
>> >> >>> >> >> > the
>> >> >>> >> >> > exact
>> >> >>> >> >> > same
>> >> >>> >> >> > name
>> >> >>> >> >> > as the node's class name for the matchers.
>> >> >>> >> >> > I *think* we didn't do it for cxxRecordDecl, because
>> >> >>> >> >> > Richard
>> >> >>> >> >> > said
>> >> >>> >> >> > that's
>> >> >>> >> >> > a
>> >> >>> >> >> > relic we should get rid of anyway, but I'm not sure.
>> >> >>> >> >>
>> >> >>> >> >> FWIW, I think the state we're in is the worst of all worlds.
>> >> >>> >> >> It's
>> >> >>> >> >> not
>> >> >>> >> >> intuitive that recordDecl() doesn't match a struct in C mode,
>> >> >>> >> >> and
>> >> >>> >> >> as
>> >> >>> >> >> it stands, there is no way to match a struct or union
>> >> >>> >> >> declaration
>> >> >>> >> >> in C
>> >> >>> >> >> at all.
>> >> >>> >> >
>> >> >>> >> >
>> >> >>> >> > Agreed. Best intentions. Worst possible outcome. That's
>> >> >>> >> > software
>> >> >>> >> > development
>> >> >>> >> > :)
>> >> >>> >> >
>> >> >>> >> >> >
>> >> >>> >> >> > On Fri, Sep 4, 2015 at 8:32 PM Aaron Ballman
>> >> >>> >> >> > <aaron at aaronballman.com>
>> >> >>> >> >> > wrote:
>> >> >>> >> >> >>
>> >> >>> >> >> >> It turns out that the recordDecl() AST matcher doesn't
>> >> >>> >> >> >> match
>> >> >>> >> >> >> RecordDecl objects; instead, it matches CXXRecordDecl
>> >> >>> >> >> >> objects.
>> >> >>> >> >> >> This
>> >> >>> >> >> >> is... unfortunate... as it makes writing AST matchers more
>> >> >>> >> >> >> complicated
>> >> >>> >> >> >> because of having to translate between
>> >> >>> >> >> >> recordDecl()/CXXRecordDecl.
>> >> >>> >> >> >> It
>> >> >>> >> >> >> also makes it impossible to match a struct or union
>> >> >>> >> >> >> declaration
>> >> >>> >> >> >> in C
>> >> >>> >> >> >> or ObjC. However, given how prevalent recordDecl()'s use
>> >> >>> >> >> >> is
>> >> >>> >> >> >> in
>> >> >>> >> >> >> the
>> >> >>> >> >> >> wild (I'm guessing), changing it at this point would be a
>> >> >>> >> >> >> Bad
>> >> >>> >> >> >> Thing.
>> >> >>> >> >> >>
>> >> >>> >> >> >> For people trying to write AST matchers for languages like
>> >> >>> >> >> >> C
>> >> >>> >> >> >> or
>> >> >>> >> >> >> ObjC,
>> >> >>> >> >> >> I would like to propose adding:
>> >> >>> >> >> >>
>> >> >>> >> >> >> structDecl()
>> >> >>> >> >> >> unionDecl()
>> >> >>> >> >> >> tagDecl()
>> >> >>> >> >> >>
>> >> >>> >> >> >> These will match nicely with the existing enumDecl() AST
>> >> >>> >> >> >> matcher.
>> >> >>> >> >> >>
>> >> >>> >> >> >> Additionally, I would like to add cxxRecordDecl() to match
>> >> >>> >> >> >> CXXRecordDecl objects. While it duplicates the
>> >> >>> >> >> >> functionality
>> >> >>> >> >> >> exposed
>> >> >>> >> >> >> by recordDecl(), it more clearly matches the intention of
>> >> >>> >> >> >> which
>> >> >>> >> >> >> AST
>> >> >>> >> >> >> node it corresponds to.
>> >> >>> >> >> >>
>> >> >>> >> >> >> Finally, I would like to undocument recordDecl() and
>> >> >>> >> >> >> change
>> >> >>> >> >> >> our
>> >> >>> >> >> >> existing documentation and AST matcher uses to use
>> >> >>> >> >> >> cxxRecordDecl/structDecl() instead. Maybe someday we can
>> >> >>> >> >> >> deprecate
>> >> >>> >> >> >> recordDecl() more officially.
>> >> >>> >> >> >>
>> >> >>> >> >> >> I'm open to other ideas if there are better ways to move
>> >> >>> >> >> >> forward. If
>> >> >>> >> >> >> you think changing the meaning of recordDecl() is
>> >> >>> >> >> >> acceptable,
>> >> >>> >> >> >> I
>> >> >>> >> >> >> can
>> >> >>> >> >> >> also go that route (though I would still propose adding
>> >> >>> >> >> >> unionDecl()
>> >> >>> >> >> >> and cxxRecordDecl() in that case).
>> >> >>> >> >> >
>> >> >>> >> >> >
>> >> >>> >> >> > I think changing recordDecl is acceptable. I believe very
>> >> >>> >> >> > few
>> >> >>> >> >> > tools
>> >> >>> >> >> > will
>> >> >>> >> >> > actually start doing wrong things because of it. I'd like
>> >> >>> >> >> > more
>> >> >>> >> >> > opinions
>> >> >>> >> >> > first, though :)
>> >> >>> >> >>
>> >> >>> >> >> I was giving this more thought over the long weekend, and I
>> >> >>> >> >> think
>> >> >>> >> >> you
>> >> >>> >> >> may be right. I think changing recordDecl() to mean
>> >> >>> >> >> RecordDecl
>> >> >>> >> >> will
>> >> >>> >> >> fix more code than it breaks, so long as we take a holistic
>> >> >>> >> >> approach
>> >> >>> >> >> to the change and see which narrowing and traversal matchers
>> >> >>> >> >> we
>> >> >>> >> >> need
>> >> >>> >> >> to fix up at the same time. When I tried to think of AST
>> >> >>> >> >> matchers
>> >> >>> >> >> that
>> >> >>> >> >> mean CXXRecordDecl but *not* RecordDecl, they were horribly
>> >> >>> >> >> contrived
>> >> >>> >> >> because you usually are matching on additional selection
>> >> >>> >> >> criteria
>> >> >>> >> >> that
>> >> >>> >> >> is specific to C++ (such as hasMethod() or isDerivedFrom())
>> >> >>> >> >> which
>> >> >>> >> >> would cause the match to continue to fail, as expected. Code
>> >> >>> >> >> that
>> >> >>> >> >> uses
>> >> >>> >> >> recordDecl() to mean RecordDecl will suddenly start to match
>> >> >>> >> >> in
>> >> >>> >> >> more
>> >> >>> >> >> cases, but that's likely to be a bug fix more than a breaking
>> >> >>> >> >> change.
>> >> >>> >> >> To the best of my understanding, the only breaking cases
>> >> >>> >> >> would
>> >> >>> >> >> be
>> >> >>> >> >> where you wrote recordDecl(), meant CXXRecordDecl, had no
>> >> >>> >> >> further
>> >> >>> >> >> narrowing or traversal matchers, and were compiling in C
>> >> >>> >> >> mode;
>> >> >>> >> >> with
>> >> >>> >> >> the result being additional unexpected matches.
>> >> >>> >> >
>> >> >>> >> >
>> >> >>> >> > Ah, there's one thing that can break: the compile can break:
>> >> >>> >> > recordDecl(hasMethod(...)) will *not* compile (it'll work in
>> >> >>> >> > the
>> >> >>> >> > dynamic
>> >> >>> >> > matchers and fail as you suggest, but the in-C++ DSL does more
>> >> >>> >> > static
>> >> >>> >> > type
>> >> >>> >> > checking).
>> >> >>> >> > I don't think that's super bad though.
>> >> >>> >> >
>> >> >>> >> >>
>> >> >>> >> >> So perhaps it would make sense to:
>> >> >>> >> >>
>> >> >>> >> >> 1) Make recordDecl() mean RecordDecl
>> >> >>> >> >> 2) Do a comprehensive review of matchers that take a
>> >> >>> >> >> CXXRecordDecl
>> >> >>> >> >> and
>> >> >>> >> >> see if they should instead take a RecordDecl
>> >> >>> >> >> 3) Add unionDecl() as a node matcher (or should we add
>> >> >>> >> >> isUnion()
>> >> >>> >> >> and
>> >> >>> >> >> isStruct() as narrowing matchers?)
>> >> >>> >> >> 4) Add tagDecl() as a node matcher, but not add
>> >> >>> >> >> cxxRecordDecl()
>> >> >>> >> >
>> >> >>> >> >
>> >> >>> >> > Why not add cxxRecordDecl()? I think we need it if we want
>> >> >>> >> > narrowing
>> >> >>> >> > matchers on CXXRecordDecl?
>> >> >>> >>
>> >> >>> >> If Richard thinks CXXRecordDecl is an anachronism, I figured we
>> >> >>> >> didn't
>> >> >>> >> want to expose it. Instead, we could make hasMethod (et al)
>> >> >>> >> accept
>> >> >>> >> a
>> >> >>> >> RecordDecl and do the type checking for the caller. Then
>> >> >>> >> recordDecl(hasMethod(...)) continues to compile and work, and
>> >> >>> >> when
>> >> >>> >> hasMethod is given a RecordDecl instead of a CXXRecordDecl, it
>> >> >>> >> simply
>> >> >>> >> matches nothing. But you bring up a good point about the C++ DSL
>> >> >>> >> being
>> >> >>> >> a problem still, I hadn't considered that.
>> >> >>> >
>> >> >>> >
>> >> >>> > First I want Richard to confirm that. I have a very bad memory,
>> >> >>> > so I
>> >> >>> > might
>> >> >>> > as well misremember :)
>> >> >>> >
>> >> >>> >>
>> >> >>> >>
>> >> >>> >> ~Aaron
>> >> >>> >>
>> >> >>> >> >
>> >> >>> >> >>
>> >> >>> >> >>
>> >> >>> >> >> ~Aaron
>> >> >>> >> >>
>> >> >>> >> >> >
>> >> >>> >> >> >>
>> >> >>> >> >> >>
>> >> >>> >> >> >> Thanks!
>> >> >>> >> >> >>
>> >> >>> >> >> >> ~Aaron
>> >> >
>> >> >
-------------- next part --------------
A non-text attachment was scrubbed...
Name: recordDecl.patch
Type: application/octet-stream
Size: 16075 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150914/95dd8ac3/attachment-0001.obj>