[Rcpp-devel] Typecasting bug in Rcpp::List (generic_name_proxy) when using clang

[Christian Authmann]

I've removed the bool cast from last week's git, reinstalled the 
package, ran the test suite with --allTests, recompiled my own codes, 
everything seems to work.

If the bool cast is missing, the code should just fall back to the 
templated operator T(), with T=bool. Unless RCPP_DEBUG_LEVEL is enabled, 
those are identical.


RUNIT TEST PROTOCOL -- Tue Jul 29 13:20:29 2014
***********************************************
Number of test functions: 439
Number of errors: 0
Number of failures: 0


Sorry, due to a restrictive firewall on my dev computer, I cannot access 
github from it. A mail will have to suffice for now.


Ciao,
Christian

On 29.07.2014 10:44, Romain Francois wrote:
> This is likely to be a bug in Rcpp, don't go bother clang about it:)
> Perhaps you can try to remove the bool cast and run the test suite. If everything works, then it either means that there are no tests for the particular case or that the overload is useless.
>
> You probably have to consider what happens with :
>
> list["a"] = false ;
> bool a = list["a"] ;
>
> i.e. does it work when the overload is removed, etc ...
>
> Also, it will be more useful if the example was not using RInside, as this has nothing to do with RInside really. You can easily demonstrate the problem with a cppFunction or a sourceCpp without having to go through RInside.
>
> I'd recommend to go here for bug submission:
> https://github.com/RcppCore/Rcpp/issues
>
> Romain
>
> Le 29 juil. 2014 à 10:36, Christian Authmann <authmanc at Mathematik.Uni-Marburg.de> a écrit :
>
>> Hi,
>>
>> the following testcase exposes a difference in gcc and clang
>>
>> -------------------- snip --------------------
>> #include <Rcpp.h>
>> #include <RInside.h>
>>
>> int main(int argc, char *argv[]) {
>> 	// We need some kind of R environment to create a List.
>> 	RInside R;
>>
>> 	Rcpp::List list;
>> 	list["a"] = 42;
>>
>> 	int a_auto = list["a"];
>> 	int a_cast = Rcpp::as<int>(list["a"]);
>>
>> 	std::cout << "list[\"a\"]                = " << a_auto << std::endl;
>> 	std::cout << "Rcpp::as<int>(list[\"a\"]) = " << a_cast << std::endl;
>> }
>> -------------------- snip --------------------
>>
>> Results:
>>
>> gcc version 4.8.2 (Ubuntu 4.8.2-19ubuntu1)
>> list["a"]                = 42
>> Rcpp::as<int>(list["a"]) = 42
>>
>> Ubuntu clang version 3.5-1ubuntu1 (trunk) (based on LLVM 3.5)
>> list["a"]                = 1
>> Rcpp::as<int>(list["a"]) = 42
>>
>>
>> list["a"] returns a Rcpp::internal::generic_name_proxy<19>, which has the following typecasting operators: (rcpp/vector/proxy.h)
>>
>> -------------------- snip --------------------
>> template <typename T>
>> operator T() const {
>> 	#if RCPP_DEBUG_LEVEL > 0
>> 	SEXP res = get() ;
>> 	RCPP_DEBUG_1( "generic_name_proxy::get() = <%p> ", res ) ;
>> 	return ::Rcpp::as<T>( res ) ;
>> 	#else
>> 	return ::Rcpp::as<T>( get() ) ;
>> 	#endif
>> }
>>
>> operator bool() const{
>>     return ::Rcpp::as<bool>(get());
>> }
>> -------------------- snip --------------------
>>
>> A few sprinkled printf's reveal that clang will call the bool typecast, while gcc calls the template as intended.
>>
>>
>> Now that I know, I can easily work around it by using Rcpp::as where needed (and I'm posting here so maybe others can find the information), but maybe it should be fixed. ;)
>>
>>
>> I don't know all the intricacies of the C++ standard. Is this a bug in clang? Or is clang free to use any conversion operator it likes, and Rcpp is at fault for relying on implementation details of gcc?
>>
>> If the former, I'll create a reduced testcase and file a bug with clang.
>>
>> If the latter, I noticed that removing the bool typecast altogether will cause clang to correctly return 42. Unless there is a good reason for its existence (say, to suppress logging for bool casts), we could just remove it.
>> Grep'ping all the includes suggest that this is the only class with redundant typecast operators.
>>
>>
>> Ciao,
>> Christian

-- 
Christian Authmann
Philipps-Universität Marburg
Fachbereich Mathematik und Informatik
AG Datenbanksysteme
Hans-Meerwein-Straße
D-35032 Marburg