std::snprintf not found with GCC 4.8, 4.6

[mboisson]

Hi,
For some reason, GCC 4.8 and 4.6 tell me that snprintf is not part of std. I tried virtually every -std options, and nothing works. However, replacing std::snprintf by just snprintf in arma_boost.hpp solved the problem as it falls back on snprintf from C99.

While I consider this to be a compiler bug, would it be possible to just change it in your code to use snprintf instead of std::snprintf ?

[kevinushey]

Thanks. In fact, snprintf being in std is a C++11 feature (http://en.cppreference.com/w/cpp/io/c/fprintf).

I imagine RcppArmadillo must also assume C++98 support, and since that also precludes C99 features, I am not sure what the best way to handle this is. But I guess not specifying std won't hurt anyone, since the 'main' compilers will silently borrow C99 features as needed...

Thoughts, @eddelbuettel?

[eddelbuettel]

For changes to Armadillo, you need to talk Armadillo's author -- which is Conrad.

[eddelbuettel]

Ok, on second look the file in question, ie arma_boost.hpp, has this carefully #ifdef'ed (as is common with Conrad's code).

But it turns out that we do define #define ARMA_HAVE_STD_SNPRINTF -- and that line is from Aug 2010 and has not bitten anybody yet. Oh, and I even overlooked that we then undefine it in some cases:

/* TODO: we might need to undef this on other platforms as well */
#if defined(__GNUC__) && defined(_WIN64) || defined(__FreeBSD__)
#undef ARMA_HAVE_STD_SNPRINTF                                                                                                                                                                                                               
#endif

Could you be a bit more specific in how this hurt you? I guess your code used snprintf and the compiler choked?

We could of course flip the definition, but I'd have to do a wee bit more research about this.

[mboisson]

Here is what I get :
/software6/compilers/gcc/4.8/bin/g++ -I/software6/bioinfo/apps/mugqic_space//software/R_3.0.2_gcc4.8/R-3.0.2/lib64/R/include -DNDEBUG -I/usr/local/include -I"/software6/bioinfo/apps/mugqic_space/software/R_3.0.2_gcc4.8/R-3.0.2/lib64/R/library/Rcpp/include" -I../inst/include -fpic -g -O2 -c RcppArmadillo.cpp -o RcppArmadillo.o
In file included from ../inst/include/armadillo:97:0,
from ../inst/include/RcppArmadilloForward.h:37,
from ../inst/include/RcppArmadillo.h:30,
from RcppArmadillo.cpp:22:
../inst/include/armadillo_bits/arma_boost.hpp: In function ‘std::string arma::arma_boost::str(const arma::arma_boost::basic_format<arma::arma_boost::format, T2>&)’:
../inst/include/armadillo_bits/arma_boost.hpp:18:27: erreur: ‘snprintf’ is not a member of ‘std’
#define arma_snprintf std::snprintf

I tried manual compilation after installing through R failed with that error, to try and follow what is going on.

[kevinushey]

You should be able to avoid this locally by enforcing -std=c++11 or -std=c++0x.

[mboisson]

I did try with std=c++11, std=c++0x, std=gnu++0x, none of them changed the result. As I said, I do recognize that this might be a compiler or libstdc++ bug, but it's probably easier to go around it than to get this fixed in a timely manner.

[eddelbuettel]

I would edit RcppArmadilloConfig.h for now.

Note that the problem has not bitten anybody else in years it seems, so is there any chance it could be local to you setup?

[kevinushey]

I don't see -std=c++11 in any of the compilation examples you showed either -- are you sure that you are setting it correctly? E.g., in a file ~/.R/Makevars with PKG_CXXFLAGS=-g -O3 -Wall -std=c++11?

[eddelbuettel]

Moreover, showing an assert is a real big NO for CRAN, and we would know. You are running R from a seemingly large and complex lab installation and I fear that you may have a #define somewhere that alters the flow of includes etc pp -- because nobody else sees this. Yet Rcpp et al are fairly widely tested.

[mboisson]

Here is the example with -std=c++11. I put CXXFLAGS in ~/.R/Makevars, and not PKG_CXXFLAGS, but it was used.

[mboisson@colosse3 src]$ /software6/compilers/gcc/4.8/bin/g++ -I/software6/bioinfo/apps/mugqic_space//software/R_3.0.2_gcc4.8/R-3.0.2/lib64/R/include -DNDEBUG -I/usr/local/include -I"/software6/bioinfo/apps/mugqic_space/software/R_3.0.2_gcc4.8/R-3.0.2/lib64/R/library/Rcpp/include" -I../inst/include -fpic -g -O2 -c RcppArmadillo.cpp -o RcppArmadillo.o -std=c++11
In file included from ../inst/include/armadillo:97:0,
from ../inst/include/RcppArmadilloForward.h:37,
from ../inst/include/RcppArmadillo.h:30,
from RcppArmadillo.cpp:22:
../inst/include/armadillo_bits/arma_boost.hpp: In function ‘std::string arma::arma_boost::str(const arma::arma_boost::basic_format<arma::arma_boost::format, T2>&)’:
../inst/include/armadillo_bits/arma_boost.hpp:18:27: erreur: ‘snprintf’ is not a member of ‘std’
#define arma_snprintf std::snprintf

It is possible that it's something with our installation, as we compile everything from source (including GCC). However, I am now at over a hundred of packages that all compiled fine with our GCC 4.8 installation.

Have you tried compiling R and Rcpparmadillo with GCC 4.8 ?

[mboisson]

Here is what I get with a sample code to just test std::snprintf with GCC 4.8 :

#include
int main()
{
char * test = new char[10];
std::snprintf(test, 10, "%f", 2.);
return 0;
}

g++ test.cpp -std=c++11
test.cpp: In function ‘int main()’:
test.cpp:5:2: erreur: ‘snprintf’ is not a member of ‘std’
std::snprintf(test, 10, "%f", 2.);
^
test.cpp:5:2: note: suggested alternative:
In file included from /software6/compilers/gcc/4.8.2/include/c++/4.8.2/cstdio:42:0,
from test.cpp:1:
/usr/include/stdio.h:385:12: note: ‘snprintf’
extern int snprintf (char *__restrict __s, size_t __maxlen,
^

[mboisson]

That #include was of cstdio

[mboisson]

By the way, I opened a bug report on GCC bug tracker, for libstdc++ : http://gcc.gnu.org/bugzilla/show_bug.cgi?id=61101

[eddelbuettel]

Yes, gcc 4.8 is my default on Ubuntu, and has been used a gazillion times.

[mboisson]

Hum, I don't know then, maybe it has to do with the glibc from CentOS 6... My question remains though, why force std::snprintf rather than just letting the compiler find whatever it can ? Does std::snprintf have different behavior than snprintf ?

[eddelbuettel]

I see no issue here so I am inclined to close this.

edd@max:/tmp$ cat mboisson.cpp 
#include <cstdio>
#include <iostream>

int main() {
  char * test = new char[10];
  std::snprintf(test, 10, "%f", 2.);
  std::cout << test << std::endl;
  return 0;
}
edd@max:/tmp$ g++ -Wall -o mboisson mboisson.cpp 
edd@max:/tmp$ ./mboisson
2.000000
edd@max:/tmp$ 

[eddelbuettel]

I am closing this as a local / non-reproducible issue.