I wrote a cpp source with pari.h header file included:
#include<string>
#include<vector>
#include<algorithm>
#include<cmath>
#include<stdlib.h>
#include<time.h>
#include<iterator> // for ostream_iterator
#include<strings.h>
#include<string.h>
#include<sstream>
#include <pari/pari.h> // for PARI/GP library
#include<Rcpp.h> // for sourceCpp to work, this line must be uncommented
// Enable C++11 via this plugin (Rcpp 0.10.3 or later)
// [[Rcpp::plugins(cpp11)]]
using namespace std;
using namespace Rcpp; // for sourceCpp to work, this line must be uncommented
// [[Rcpp::export]]
int main() {
long maxp = 1000000; // max value
pari_init(500000,2); // initiate pari
size_t rsize = 500000; // set stack size variables
size_t vsize = 100000000;
void paristack_setsize(size_t rsize, size_t vsize); // declare stack function
paristack_setsize(rsize, vsize); // set stack size
gp_allocatemem(stoi(100000000)); // allocate memory
GEN p1; // declare PARI variable
p1 = cgetg(maxp, t_VEC); // make the PARI variable a vector
long j; // declare the variable for the number to be checked. one above the vector iterator
for (long i = 0; i <= maxp; ++i) { // iterate over PARI vector
j = i + 1; // decrement index for number
gel(p1, i) = sumdiv(stoi(j)); // calculate the sum of divisors and update the vector
}
vector<long> p2(maxp); // empty vector of native type
GEN x; // declare a PARI variable to subset PARI vector
for (long i = 0; i < maxp; i++) { // for2, across vector indices
x = gel(p1, i); // subset one item of vector
p2[i] = gtolong(x); // convert PARI to native long integer and update long vector item
} // close for2
for (long z = 0; z < maxp; z++) { // for3, to iterate for stdout
cout << p2[z] << '\n'; // return the result. the vector items are printed separately
} // close for3
} // close function
(Note that, there may be unnecessary headers, I usually copy all of them across sources, but that's not an issue). Similar source files without pari.h header compile well with Rcpp with the necessary parts included (such as header, namespace, export line, etc).
The source, when Rcpp related references are commented, compiles well and works with no problem when compiled directly with g++ with the following flags:
g++ -lpari -fpermissive -Wall -Wextra -lm -fno-strict-aliasing -fomit-frame-pointer -o sumdivisors.o sumdivisors.cpp
I imported those flags to R also:
Sys.setenv("PKG_CXXFLAGS"="-lpari -fpermissive -Wall -Wextra -lm -fno-strict-aliasing -fomit-frame-pointer")
I also created a symlink at /usr/local/lib64/R/library/Rcpp/include/ to pari directory under /usr/include.
However the output from sourceCpp command is as such:
> sourceCpp("sumdivisors.cpp")
In file included from /usr/local/lib64/R/library/Rcpp/include/Rcpp/r/headers.h:48:0,
from /usr/local/lib64/R/library/Rcpp/include/RcppCommon.h:29,
from /usr/local/lib64/R/library/Rcpp/include/Rcpp.h:27,
from sumdivisors.cpp:15:
/usr/local/lib64/R/library/Rcpp/include/Rcpp/platform/compiler.h:47:0: warning: "GCC_VERSION" redefined
#define GCC_VERSION (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__)
In file included from /usr/local/lib64/R/library/Rcpp/include/pari/pari.h:16:0,
from sumdivisors.cpp:14:
/usr/local/lib64/R/library/Rcpp/include/pari/paricfg.h:19:0: note: this is the location of the previous definition
#define GCC_VERSION "gcc version 6.2.1 20160830 (GCC)"
Error in dyn.load("/tmp/Rtmpc9edZe/sourceCpp-x86_64-pc-linux-gnu-0.12.8/sourcecpp_188e46b44088/sourceCpp_2.so") :
unable to load shared object '/tmp/Rtmpc9edZe/sourceCpp-x86_64-pc-linux-gnu-0.12.8/sourcecpp_188e46b44088/sourceCpp_2.so':
/tmp/Rtmpc9edZe/sourceCpp-x86_64-pc-linux-gnu-0.12.8/sourcecpp_188e46b44088/sourceCpp_2.so: undefined symbol: pari_mainstack
I replicated the steps with or without including the C++11 enable line, nothing changes. I also change gcc flags, with no result. It seems that there is a problem with gcc version definition and the definition of pari_mainstack.
I believe the problem is not about how the source is written. Below two example in which, the above cpp code is converted to a one returning a vector and the function is not main. A similar and simple code which compiles well with Rcpp is also presented:
#include<stdio.h>
#include<numeric> // for "accumulate"
#include<iostream>
#include<string>
#include<vector>
#include<algorithm>
#include<cmath>
#include<stdlib.h>
#include<time.h>
#include<iterator> // for ostream_iterator
#include<strings.h>
#include<string.h>
#include<sstream>
#include <pari/pari.h> // for PARI/GP library
#include<Rcpp.h> // for sourceCpp to work, this line must be uncommented
// Enable C++11 via this plugin (Rcpp 0.10.3 or later)
// [[Rcpp::plugins(cpp11)]]
using namespace std;
using namespace Rcpp; // for sourceCpp to work, this line must be uncommented
// [[Rcpp::export]]
vector<long> sumdivisors() {
long maxp = 1000000; // max value
pari_init(500000,2); // initiate pari
size_t rsize = 500000; // set stack size variables
size_t vsize = 100000000;
void paristack_setsize(size_t rsize, size_t vsize); // declare stack function
paristack_setsize(rsize, vsize); // set stack size
gp_allocatemem(stoi(100000000)); // allocate memory
GEN p1; // declare PARI variable
p1 = cgetg(maxp, t_VEC); // make the PARI variable a vector
long j; // declare the variable for the number to be checked. one above the vector iterator
for (long i = 0; i <= maxp; ++i) { // iterate over PARI vector
j = i + 1; // decrement index for number
gel(p1, i) = sumdiv(stoi(j)); // calculate the sum of divisors and update the vector
}
vector<long> p2(maxp); // empty vector of native type
GEN x; // declare a PARI variable to subset PARI vector
for (long i = 0; i < maxp; i++) { // for2, across vector indices
x = gel(p1, i); // subset one item of vector
p2[i] = gtolong(x); // convert PARI to native long integer and update long vector item
} // close for2
return(p2);
/*
for (long z = 0; z < maxp; z++) { // for3, to iterate for stdout
cout << p2[z] << '\n'; // return the result. the vector items are printed separately
} // close for3
*/
} // close function
.
#include<stdio.h>
#include<iostream>
#include<string>
#include<vector>
#include<algorithm>
#include<cmath>
#include<math.h>
#include<time.h>
#include<Rcpp.h>
using namespace std;
using namespace Rcpp;
//#include "std_lib_facilities.h"
// [[Rcpp::export]]
int pe001Cpp(int x) { // define a function pe001 with one in$teger input
int sum35 = 0; // define a scalar for the sum. start value is 0
for (int i=1; i<x; ++i) { // for 1 loop for counting up to x
if (i % 3 == 0 || i % 5 == 0) { // if 1, divisible by 3 or 5
sum35 += i; // update sum
} // close if 1
} // close for 1
return sum35; // return the final value
} // close function
// [[Rcpp::export]]
int pe001Cppb(int x) { // efficient method
int sumdivisible(int x, int y); // declare the below function in this scope
return sumdivisible(x, 3) + sumdivisible(x, 5) - sumdivisible(x, 15); // return the total sum
} // close function pe001Cppb
int sumdivisible(int x, int y) { // sum of terms divisibile by y
int ny = floor ((x-1) / y); // number of terms less than x and divisible by y
return ny * (ny + 1) / 2 * y; // return the sum
} // close function sumdivisible
The filtered strace output from the execution of the directly compiled binary is as follows:
open("/usr/lib/libpari-gmp-tls.so.5", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libstdc++.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libm.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libgcc_s.so.1", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libgmp.so.10", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libpthread.so.0", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_compat.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnsl.so.1", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_nis.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_files.so.2", O_RDONLY|O_CLOEXEC) = 3
As we see from here https://github.com/rstats-db/RPostgres/issues/80, the problem may be a wrong version of linked library which can be solved by symlinking. So I have to know which library files Rcpp tries to link with.
Update:
Scanelf output shows that the problematic symbol is in /usr/lib/libpari-gmp-tls.so.2.9.1 .
[s@SS ~]$ scanelf -l -s pari_mainstack | grep pari_mainstack
ET_DYN pari_mainstack /usr/lib/libpari-gmp-tls.so.2.9.1
strace output of the g++ compiled file shows that the executable is linked to /usr/lib/libpari-gmp-tls.so.5 which itself a symlink to 2.9.1 version:
[s@SS library]$ strace ./sumdivisors3.o |& grep so | grep -v "No such file"
execve("./sumdivisors3.o", ["./sumdivisors3.o"], [/* 79 vars */]) = 0
open("/usr/lib/libpari-gmp-tls.so.5", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libstdc++.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libm.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libgcc_s.so.1", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libc.so.6", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libgmp.so.10", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libdl.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libpthread.so.0", O_RDONLY|O_CLOEXEC) = 3
socket(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0) = 3
socket(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0) = 3
open("/usr/lib/libnss_compat.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnsl.so.1", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_nis.so.2", O_RDONLY|O_CLOEXEC) = 3
open("/usr/lib/libnss_files.so.2", O_RDONLY|O_CLOEXEC) = 3
The ldd output from the sourceCpp_4.so file created by sourceCpp command is as follows:
[s@SS library]$ ldd /tmp/Rtmpau9YqY/sourceCpp-x86_64-pc-linux-gnu-0.12.8/sourcecpp_3a105ad2bdba/sourceCpp_4.so
linux-vdso.so.1 (0x00007ffc28f9d000)
libR.so => not found
libstdc++.so.6 => /usr/lib/libstdc++.so.6 (0x00007f5077111000)
libm.so.6 => /usr/lib/libm.so.6 (0x00007f5076e0d000)
libgcc_s.so.1 => /usr/lib/libgcc_s.so.1 (0x00007f5076bf6000)
libc.so.6 => /usr/lib/libc.so.6 (0x00007f5076858000)
/usr/lib64/ld-linux-x86-64.so.2 (0x0000564489276000)
I tracked all those files with ldd and there is no link to /usr/lib/libpari-gmp-tls.so.2.9.1 or /usr/lib/libpari-gmp-tls.so.5 libraries. So the question is why can't sourceCpp link to those files given the necessary headers included (and while g++ can)?
Update:
Verbose output of sourceCpp shows the following command:
g++ -I/usr/local/lib64/R/include -DNDEBUG -I/usr/local/include -I"/usr/local/lib64/R/library/Rcpp/include" -I"/home/s/codes/cpp/projecteuler/library" -lpari -fpic -g -O2 -c sumdivisors2.cpp -o sumdivisors2.o
g++ -shared -L/usr/local/lib64/R/lib -L/usr/local/lib64 -o sourceCpp_5.so sumdivisors2.o -L/usr/local/lib64/R/lib -lR
I set the flags with (in fact -lpari suffices:
Sys.setenv("PKG_CXXFLAGS"="-lpari")
According to gp2c output, -lpari flag should also be included in the linking stage but here the linking command does not have it. Can it be the source of the problem? Or before that, why sourceCpp_5.so file is not linked to the necessary pari library?
And the finale:
The dependent libraries for the linking should also be explicitly declared via:
Sys.setenv("PKG_LIBS"="-lm -lpari -lc")
The library flags are given by the gp2c output. By the way to come over the gcc version issue, instead of creating a symlink to original pari headers directory, I created a copy inside R library path and commented out the line:
//#define GCC_VERSION "gcc version 6.2.1 20160830 (GCC)"
Now the compilation is successfull, a R can enjoy PARI/GP speed in number theoretic computations in R, thanks to Rcpp!
