[cfe-commits] r168521 - in /cfe/trunk: include/clang/AST/ExprCXX.h lib/Sema/SemaOverload.cpp test/SemaTemplate/default-expr-arguments.cpp

[Richard Smith]

On Fri, Nov 23, 2012 at 11:47 AM, Benjamin Kramer <benny.kra at gmail.com> wrote:
>
> On 23.11.2012, at 18:04, Benjamin Kramer <benny.kra at googlemail.com> wrote:
>
>> Author: d0k
>> Date: Fri Nov 23 11:04:52 2012
>> New Revision: 168521
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=168521&view=rev
>> Log:
>> Sema: Provide a valid source location when instantiating templates based on a CXXDefaultArgExpr.
>>
>> Fixes PR13758.
>>
>> Modified:
>>    cfe/trunk/include/clang/AST/ExprCXX.h
>>    cfe/trunk/lib/Sema/SemaOverload.cpp
>>    cfe/trunk/test/SemaTemplate/default-expr-arguments.cpp
>>
>> Modified: cfe/trunk/include/clang/AST/ExprCXX.h
>> URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/include/clang/AST/ExprCXX.h?rev=168521&r1=168520&r2=168521&view=diff
>> ==============================================================================
>> --- cfe/trunk/include/clang/AST/ExprCXX.h (original)
>> +++ cfe/trunk/include/clang/AST/ExprCXX.h Fri Nov 23 11:04:52 2012
>> @@ -795,6 +795,8 @@
>>     return SourceRange();
>>   }
>>
>> +  SourceLocation getExprLoc() const LLVM_READONLY { return Loc; }
>> +
>>   static bool classof(const Stmt *T) {
>>     return T->getStmtClass() == CXXDefaultArgExprClass;
>>   }
>>
>> Modified: cfe/trunk/lib/Sema/SemaOverload.cpp
>> URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Sema/SemaOverload.cpp?rev=168521&r1=168520&r2=168521&view=diff
>> ==============================================================================
>> --- cfe/trunk/lib/Sema/SemaOverload.cpp (original)
>> +++ cfe/trunk/lib/Sema/SemaOverload.cpp Fri Nov 23 11:04:52 2012
>> @@ -2980,7 +2980,7 @@
>>          S.IsDerivedFrom(From->getType(), ToType)))
>>       ConstructorsOnly = true;
>>
>> -    S.RequireCompleteType(From->getLocStart(), ToType, 0);
>> +    S.RequireCompleteType(From->getExprLoc(), ToType, 0);

Hmm, does getLocStart on a CXXDefaultArgExpr not work?

>>     // RequireCompleteType may have returned true due to some invalid decl
>>     // during template instantiation, but ToType may be complete enough now
>>     // to try to recover.
>>
>> Modified: cfe/trunk/test/SemaTemplate/default-expr-arguments.cpp
>> URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/test/SemaTemplate/default-expr-arguments.cpp?rev=168521&r1=168520&r2=168521&view=diff
>> ==============================================================================
>> --- cfe/trunk/test/SemaTemplate/default-expr-arguments.cpp (original)
>> +++ cfe/trunk/test/SemaTemplate/default-expr-arguments.cpp Fri Nov 23 11:04:52 2012
>> @@ -303,3 +303,21 @@
>>   {
>>   }
>> }
>> +
>> +namespace PR13758 {
>> +  template <typename T> struct move_from {
>> +    T invalid; // expected-error {{field has incomplete type 'void'}}
>> +  };
>> +  template <class K>
>> +  struct unordered_map {
>> +    explicit unordered_map(int n = 42);
>> +    unordered_map(move_from<K> other);
>> +  };
>> +  template<typename T>
>> +  void StripedHashTable() {
>> +    new unordered_map<void>(); // expected-note {{in instantiation of template class 'PR13758::move_from<void>' requested here}}
>> +  }
>> +  void tt() {
>> +    StripedHashTable<int>(); // expected-note {{in instantiation of function template specialization 'PR13758::StripedHashTable<int>' requested here}}
>> +  }
>> +}
>
> Looking closer at the test case, I don't see why it instantiates move_from<void> at that point. GCC and EDG accept the code, maybe there is another clang bug here. Any ideas?

Yes, we shouldn't be instantiating move_from<void> here. llvm.org/demo
accepts this code, so it looks like this is a regression.

Looking just at what you needed to change in your patch, my wild guess
is that we might be trying to initialize an unordered_map object from
the default argument expression '42', rather than trying to initialize
the 'n' parameter of the explicit constructor from it. That would
explain why we instantiate move_from<K>: we'd be trying to use the
second constructor to convert '42' to an unordered_map, but would be a
very surprising bug to have...