[cfe-dev] AST Matcher for Decl subclass CilkSpawnDecl

[Richard Smith]

On Thu, Nov 20, 2014 at 4:46 AM, Manuel Klimek <klimek at google.com> wrote:

> On Thu Nov 20 2014 at 1:43:48 PM Georg Altmann <
> georg.altmann at informatik.stud.uni-erlangen.de> wrote:
>
>> On 20.11.2014 10:25, Manuel Klimek wrote:
>> > On Wed Nov 19 2014 at 3:38:44 PM Georg Altmann
>> > <georg.altmann at informatik.stud.uni-erlangen.de
>> > <mailto:georg.altmann at informatik.stud.uni-erlangen.de>> wrote:
>> >
>> >     Thank you for the quick reply.
>> >
>> >     I found a partial solution to my problem. I dug down into the AST
>> >     matching code and noticed RecursiveASTVisitor did not recurse into
>> >     CapturedStmt in CilkSpawnDecl. So I did the patch
>> >
>> >
>> >     diff --git a/include/clang/AST/__RecursiveASTVisitor.h
>> >     b/include/clang/AST/__RecursiveASTVisitor.h
>> >     index 35eec69..a16e772 100755
>> >     --- a/include/clang/AST/__RecursiveASTVisitor.h
>> >     +++ b/include/clang/AST/__RecursiveASTVisitor.h
>> >     @@ -1305,6 +1305,7 @@ DEF_TRAVERSE_DECL(__CapturedDecl, {
>> >        })
>> >
>> >      DEF_TRAVERSE_DECL(__CilkSpawnDecl, {
>> >     +    TRY_TO(TraverseStmt(D->__getCapturedStmt()));
>> >        })
>> >
>> >      DEF_TRAVERSE_DECL(EmptyDecl, { })
>> >
>> >
>> >     Again, this is what I want to match:
>> >     The AST:
>> >     |-CompoundStmt
>> >     | `-CilkSpawnExpr
>> >     |   |-CilkSpawnDecl
>> >     |   | `-CapturedStmt
>> >     |   |   |-CapturedDecl
>> >     |   |   | `-ImplicitParamDecl
>> >     |   |   |-ExprWithCleanups
>> >     |   |   | `-CallExpr
>> >     --------------^ this guy
>> >
>> >
>> >     After the patch, these match now:
>> >
>> >     // 1) matches
>> >     compoundStmt(
>> >       has(cilkSpawnExpr(
>> >        has(cilkSpawnDecl(
>> >          has(capturedStmt()))))))
>> >
>> >     // 2) matches and binds
>> >     compoundStmt(
>> >       has(cilkSpawnExpr(
>> >         has(cilkSpawnDecl(
>> >           has(capturedStmt(
>> >             hasDescendant(__exprWithCleanups(
>> >               has(callExpr().bind("callExpr"__))))))))))
>> >
>> >
>> >     However this doesn't match, while it should IMHO.
>> >
>> >     // 3) does not match
>> >     compoundStmt(
>> >       has(cilkSpawnExpr(
>> >         has(cilkSpawnDecl(
>> >             has(capturedStmt(
>> >                 has(exprWithCleanups())))))))
>> >
>> >
>> > Which one in this chain is the one that doesn't match? (remove inner
>> > ones until it matches, send result).
>>
>>
>> These match:
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt()))))))
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         has(decl()))))))))
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         // for some reason has() does not work here
>>         hasDescendant(exprWithCleanups(
>>           has(callExpr().bind("callExpr")))))))))))
>>
>>
>> None of these match:
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         has(exprWithCleanups()))))))))
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         has(expr()))))))))
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         has(stmt()))))))))
>>
>> compoundStmt(
>>   has(cilkSpawnExpr(
>>     has(cilkSpawnDecl(
>>       has(capturedStmt(
>>         has(decl()))))))))
>>
>> To me this looks like either
>> a) the ast-dump is incorrect and ExprWithCleanups is not a direct child
>> of CapturedStmt
>> or
>> b) there is a bug with CapturedStmt in the ast matching code /
>> RecursiveASTVisitor which makes ExprWithCleanups being found as
>> descendant but not as direct child.
>>
>> Maybe a test-case against vanilla clang would be helpful? However I
>> don't know how to reproduce the AST without the cilk extensions.
>>
>
> The problem is that the RecursiveASTVisitor and the -ast-dump logic use
> completely different code paths... I still think this is probably a bug in
> one of the two.
> +richardsmith, who knows more about this...
>

-ast-dump has no custom logic for CapturedStmt, so I'm surprised you're
getting the dump you provide above. I don't see that with clang trunk.
Perhaps this is something that the cilk folks changed.


> Regards,
>> Georg
>>
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20141120/4130c511/attachment.html>